kernel/fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 static const struct inode_operations btrfs_dir_inode_operations;
  70 static const struct inode_operations btrfs_symlink_inode_operations;
  71 static const struct inode_operations btrfs_dir_ro_inode_operations;
  72 static const struct inode_operations btrfs_special_inode_operations;
  73 static const struct inode_operations btrfs_file_inode_operations;
  74 static const struct address_space_operations btrfs_aops;
  75 static const struct address_space_operations btrfs_symlink_aops;
  76 static const struct file_operations btrfs_dir_file_operations;
  77 static struct extent_io_ops btrfs_extent_io_ops;
  78
  79 static struct kmem_cache *btrfs_inode_cachep;
  80 static struct kmem_cache *btrfs_delalloc_work_cachep;
  81 struct kmem_cache *btrfs_trans_handle_cachep;
  82 struct kmem_cache *btrfs_transaction_cachep;
  83 struct kmem_cache *btrfs_path_cachep;
  84 struct kmem_cache *btrfs_free_space_cachep;
  85
  86 #define S_SHIFT 12
  87 static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  88         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  89         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  90         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  91         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  92         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  93         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  94         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  95 };
  96
  97 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  98 static int btrfs_truncate(struct inode *inode);
  99 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 100 static noinline int cow_file_range(struct inode *inode,
 101                                    struct page *locked_page,
 102                                    u64 start, u64 end, int *page_started,
 103                                    unsigned long *nr_written, int unlock);
 104 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 105                                            u64 len, u64 orig_start,
 106                                            u64 block_start, u64 block_len,
 107                                            u64 orig_block_len, u64 ram_bytes,
 108                                            int type);
 109
 110 static int btrfs_dirty_inode(struct inode *inode);
 111
 112 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 113 void btrfs_test_inode_set_ops(struct inode *inode)
 114 {
 115         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 116 }
 117 #endif
 118
 119 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 120                                      struct inode *inode,  struct inode *dir,
 121                                      const struct qstr *qstr)
 122 {
 123         int err;
 124
 125         err = btrfs_init_acl(trans, inode, dir);
 126         if (!err)
 127                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 128         return err;
 129 }
 130
 131 /*
 132  * this does all the hard work for inserting an inline extent into
 133  * the btree.  The caller should have done a btrfs_drop_extents so that
 134  * no overlapping inline items exist in the btree
 135  */
 136 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 137                                 struct btrfs_path *path, int extent_inserted,
 138                                 struct btrfs_root *root, struct inode *inode,
 139                                 u64 start, size_t size, size_t compressed_size,
 140                                 int compress_type,
 141                                 struct page **compressed_pages)
 142 {
 143         struct extent_buffer *leaf;
 144         struct page *page = NULL;
 145         char *kaddr;
 146         unsigned long ptr;
 147         struct btrfs_file_extent_item *ei;
 148         int err = 0;
 149         int ret;
 150         size_t cur_size = size;
 151         unsigned long offset;
 152
 153         if (compressed_size && compressed_pages)
 154                 cur_size = compressed_size;
 155
 156         inode_add_bytes(inode, size);
 157
 158         if (!extent_inserted) {
 159                 struct btrfs_key key;
 160                 size_t datasize;
 161
 162                 key.objectid = btrfs_ino(inode);
 163                 key.offset = start;
 164                 key.type = BTRFS_EXTENT_DATA_KEY;
 165
 166                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 167                 path->leave_spinning = 1;
 168                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 169                                               datasize);
 170                 if (ret) {
 171                         err = ret;
 172                         goto fail;
 173                 }
 174         }
 175         leaf = path->nodes[0];
 176         ei = btrfs_item_ptr(leaf, path->slots[0],
 177                             struct btrfs_file_extent_item);
 178         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 179         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 180         btrfs_set_file_extent_encryption(leaf, ei, 0);
 181         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 182         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 183         ptr = btrfs_file_extent_inline_start(ei);
 184
 185         if (compress_type != BTRFS_COMPRESS_NONE) {
 186                 struct page *cpage;
 187                 int i = 0;
 188                 while (compressed_size > 0) {
 189                         cpage = compressed_pages[i];
 190                         cur_size = min_t(unsigned long, compressed_size,
 191                                        PAGE_CACHE_SIZE);
 192
 193                         kaddr = kmap_atomic(cpage);
 194                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 195                         kunmap_atomic(kaddr);
 196
 197                         i++;
 198                         ptr += cur_size;
 199                         compressed_size -= cur_size;
 200                 }
 201                 btrfs_set_file_extent_compression(leaf, ei,
 202                                                   compress_type);
 203         } else {
 204                 page = find_get_page(inode->i_mapping,
 205                                      start >> PAGE_CACHE_SHIFT);
 206                 btrfs_set_file_extent_compression(leaf, ei, 0);
 207                 kaddr = kmap_atomic(page);
 208                 offset = start & (PAGE_CACHE_SIZE - 1);
 209                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 210                 kunmap_atomic(kaddr);
 211                 page_cache_release(page);
 212         }
 213         btrfs_mark_buffer_dirty(leaf);
 214         btrfs_release_path(path);
 215
 216         /*
 217          * we're an inline extent, so nobody can
 218          * extend the file past i_size without locking
 219          * a page we already have locked.
 220          *
 221          * We must do any isize and inode updates
 222          * before we unlock the pages.  Otherwise we
 223          * could end up racing with unlink.
 224          */
 225         BTRFS_I(inode)->disk_i_size = inode->i_size;
 226         ret = btrfs_update_inode(trans, root, inode);
 227
 228         return ret;
 229 fail:
 230         return err;
 231 }
 232
 233
 234 /*
 235  * conditionally insert an inline extent into the file.  This
 236  * does the checks required to make sure the data is small enough
 237  * to fit as an inline extent.
 238  */
 239 static noinline int cow_file_range_inline(struct btrfs_root *root,
 240                                           struct inode *inode, u64 start,
 241                                           u64 end, size_t compressed_size,
 242                                           int compress_type,
 243                                           struct page **compressed_pages)
 244 {
 245         struct btrfs_trans_handle *trans;
 246         u64 isize = i_size_read(inode);
 247         u64 actual_end = min(end + 1, isize);
 248         u64 inline_len = actual_end - start;
 249         u64 aligned_end = ALIGN(end, root->sectorsize);
 250         u64 data_len = inline_len;
 251         int ret;
 252         struct btrfs_path *path;
 253         int extent_inserted = 0;
 254         u32 extent_item_size;
 255
 256         if (compressed_size)
 257                 data_len = compressed_size;
 258
 259         if (start > 0 ||
 260             actual_end > PAGE_CACHE_SIZE ||
 261             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 262             (!compressed_size &&
 263             (actual_end & (root->sectorsize - 1)) == 0) ||
 264             end + 1 < isize ||
 265             data_len > root->fs_info->max_inline) {
 266                 return 1;
 267         }
 268
 269         path = btrfs_alloc_path();
 270         if (!path)
 271                 return -ENOMEM;
 272
 273         trans = btrfs_join_transaction(root);
 274         if (IS_ERR(trans)) {
 275                 btrfs_free_path(path);
 276                 return PTR_ERR(trans);
 277         }
 278         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 279
 280         if (compressed_size && compressed_pages)
 281                 extent_item_size = btrfs_file_extent_calc_inline_size(
 282                    compressed_size);
 283         else
 284                 extent_item_size = btrfs_file_extent_calc_inline_size(
 285                     inline_len);
 286
 287         ret = __btrfs_drop_extents(trans, root, inode, path,
 288                                    start, aligned_end, NULL,
 289                                    1, 1, extent_item_size, &extent_inserted);
 290         if (ret) {
 291                 btrfs_abort_transaction(trans, root, ret);
 292                 goto out;
 293         }
 294
 295         if (isize > actual_end)
 296                 inline_len = min_t(u64, isize, actual_end);
 297         ret = insert_inline_extent(trans, path, extent_inserted,
 298                                    root, inode, start,
 299                                    inline_len, compressed_size,
 300                                    compress_type, compressed_pages);
 301         if (ret && ret != -ENOSPC) {
 302                 btrfs_abort_transaction(trans, root, ret);
 303                 goto out;
 304         } else if (ret == -ENOSPC) {
 305                 ret = 1;
 306                 goto out;
 307         }
 308
 309         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 310         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 311         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 312 out:
 313         /*
 314          * Don't forget to free the reserved space, as for inlined extent
 315          * it won't count as data extent, free them directly here.
 316          * And at reserve time, it's always aligned to page size, so
 317          * just free one page here.
 318          */
 319         btrfs_qgroup_free_data(inode, 0, PAGE_CACHE_SIZE);
 320         btrfs_free_path(path);
 321         btrfs_end_transaction(trans, root);
 322         return ret;
 323 }
 324
 325 struct async_extent {
 326         u64 start;
 327         u64 ram_size;
 328         u64 compressed_size;
 329         struct page **pages;
 330         unsigned long nr_pages;
 331         int compress_type;
 332         struct list_head list;
 333 };
 334
 335 struct async_cow {
 336         struct inode *inode;
 337         struct btrfs_root *root;
 338         struct page *locked_page;
 339         u64 start;
 340         u64 end;
 341         struct list_head extents;
 342         struct btrfs_work work;
 343 };
 344
 345 static noinline int add_async_extent(struct async_cow *cow,
 346                                      u64 start, u64 ram_size,
 347                                      u64 compressed_size,
 348                                      struct page **pages,
 349                                      unsigned long nr_pages,
 350                                      int compress_type)
 351 {
 352         struct async_extent *async_extent;
 353
 354         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 355         BUG_ON(!async_extent); /* -ENOMEM */
 356         async_extent->start = start;
 357         async_extent->ram_size = ram_size;
 358         async_extent->compressed_size = compressed_size;
 359         async_extent->pages = pages;
 360         async_extent->nr_pages = nr_pages;
 361         async_extent->compress_type = compress_type;
 362         list_add_tail(&async_extent->list, &cow->extents);
 363         return 0;
 364 }
 365
 366 static inline int inode_need_compress(struct inode *inode)
 367 {
 368         struct btrfs_root *root = BTRFS_I(inode)->root;
 369
 370         /* force compress */
 371         if (btrfs_test_opt(root, FORCE_COMPRESS))
 372                 return 1;
 373         /* bad compression ratios */
 374         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 375                 return 0;
 376         if (btrfs_test_opt(root, COMPRESS) ||
 377             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 378             BTRFS_I(inode)->force_compress)
 379                 return 1;
 380         return 0;
 381 }
 382
 383 /*
 384  * we create compressed extents in two phases.  The first
 385  * phase compresses a range of pages that have already been
 386  * locked (both pages and state bits are locked).
 387  *
 388  * This is done inside an ordered work queue, and the compression
 389  * is spread across many cpus.  The actual IO submission is step
 390  * two, and the ordered work queue takes care of making sure that
 391  * happens in the same order things were put onto the queue by
 392  * writepages and friends.
 393  *
 394  * If this code finds it can't get good compression, it puts an
 395  * entry onto the work queue to write the uncompressed bytes.  This
 396  * makes sure that both compressed inodes and uncompressed inodes
 397  * are written in the same order that the flusher thread sent them
 398  * down.
 399  */
 400 static noinline void compress_file_range(struct inode *inode,
 401                                         struct page *locked_page,
 402                                         u64 start, u64 end,
 403                                         struct async_cow *async_cow,
 404                                         int *num_added)
 405 {
 406         struct btrfs_root *root = BTRFS_I(inode)->root;
 407         u64 num_bytes;
 408         u64 blocksize = root->sectorsize;
 409         u64 actual_end;
 410         u64 isize = i_size_read(inode);
 411         int ret = 0;
 412         struct page **pages = NULL;
 413         unsigned long nr_pages;
 414         unsigned long nr_pages_ret = 0;
 415         unsigned long total_compressed = 0;
 416         unsigned long total_in = 0;
 417         unsigned long max_compressed = 128 * 1024;
 418         unsigned long max_uncompressed = 128 * 1024;
 419         int i;
 420         int will_compress;
 421         int compress_type = root->fs_info->compress_type;
 422         int redirty = 0;
 423
 424         /* if this is a small write inside eof, kick off a defrag */
 425         if ((end - start + 1) < 16 * 1024 &&
 426             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 427                 btrfs_add_inode_defrag(NULL, inode);
 428
 429         actual_end = min_t(u64, isize, end + 1);
 430 again:
 431         will_compress = 0;
 432         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 433         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
 434
 435         /*
 436          * we don't want to send crud past the end of i_size through
 437          * compression, that's just a waste of CPU time.  So, if the
 438          * end of the file is before the start of our current
 439          * requested range of bytes, we bail out to the uncompressed
 440          * cleanup code that can deal with all of this.
 441          *
 442          * It isn't really the fastest way to fix things, but this is a
 443          * very uncommon corner.
 444          */
 445         if (actual_end <= start)
 446                 goto cleanup_and_bail_uncompressed;
 447
 448         total_compressed = actual_end - start;
 449
 450         /*
 451          * skip compression for a small file range(<=blocksize) that
 452          * isn't an inline extent, since it dosen't save disk space at all.
 453          */
 454         if (total_compressed <= blocksize &&
 455            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 456                 goto cleanup_and_bail_uncompressed;
 457
 458         /* we want to make sure that amount of ram required to uncompress
 459          * an extent is reasonable, so we limit the total size in ram
 460          * of a compressed extent to 128k.  This is a crucial number
 461          * because it also controls how easily we can spread reads across
 462          * cpus for decompression.
 463          *
 464          * We also want to make sure the amount of IO required to do
 465          * a random read is reasonably small, so we limit the size of
 466          * a compressed extent to 128k.
 467          */
 468         total_compressed = min(total_compressed, max_uncompressed);
 469         num_bytes = ALIGN(end - start + 1, blocksize);
 470         num_bytes = max(blocksize,  num_bytes);
 471         total_in = 0;
 472         ret = 0;
 473
 474         /*
 475          * we do compression for mount -o compress and when the
 476          * inode has not been flagged as nocompress.  This flag can
 477          * change at any time if we discover bad compression ratios.
 478          */
 479         if (inode_need_compress(inode)) {
 480                 WARN_ON(pages);
 481                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 482                 if (!pages) {
 483                         /* just bail out to the uncompressed code */
 484                         goto cont;
 485                 }
 486
 487                 if (BTRFS_I(inode)->force_compress)
 488                         compress_type = BTRFS_I(inode)->force_compress;
 489
 490                 /*
 491                  * we need to call clear_page_dirty_for_io on each
 492                  * page in the range.  Otherwise applications with the file
 493                  * mmap'd can wander in and change the page contents while
 494                  * we are compressing them.
 495                  *
 496                  * If the compression fails for any reason, we set the pages
 497                  * dirty again later on.
 498                  */
 499                 extent_range_clear_dirty_for_io(inode, start, end);
 500                 redirty = 1;
 501                 ret = btrfs_compress_pages(compress_type,
 502                                            inode->i_mapping, start,
 503                                            total_compressed, pages,
 504                                            nr_pages, &nr_pages_ret,
 505                                            &total_in,
 506                                            &total_compressed,
 507                                            max_compressed);
 508
 509                 if (!ret) {
 510                         unsigned long offset = total_compressed &
 511                                 (PAGE_CACHE_SIZE - 1);
 512                         struct page *page = pages[nr_pages_ret - 1];
 513                         char *kaddr;
 514
 515                         /* zero the tail end of the last page, we might be
 516                          * sending it down to disk
 517                          */
 518                         if (offset) {
 519                                 kaddr = kmap_atomic(page);
 520                                 memset(kaddr + offset, 0,
 521                                        PAGE_CACHE_SIZE - offset);
 522                                 kunmap_atomic(kaddr);
 523                         }
 524                         will_compress = 1;
 525                 }
 526         }
 527 cont:
 528         if (start == 0) {
 529                 /* lets try to make an inline extent */
 530                 if (ret || total_in < (actual_end - start)) {
 531                         /* we didn't compress the entire range, try
 532                          * to make an uncompressed inline extent.
 533                          */
 534                         ret = cow_file_range_inline(root, inode, start, end,
 535                                                     0, 0, NULL);
 536                 } else {
 537                         /* try making a compressed inline extent */
 538                         ret = cow_file_range_inline(root, inode, start, end,
 539                                                     total_compressed,
 540                                                     compress_type, pages);
 541                 }
 542                 if (ret <= 0) {
 543                         unsigned long clear_flags = EXTENT_DELALLOC |
 544                                 EXTENT_DEFRAG;
 545                         unsigned long page_error_op;
 546
 547                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 548                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 549
 550                         /*
 551                          * inline extent creation worked or returned error,
 552                          * we don't need to create any more async work items.
 553                          * Unlock and free up our temp pages.
 554                          */
 555                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 556                                                      clear_flags, PAGE_UNLOCK |
 557                                                      PAGE_CLEAR_DIRTY |
 558                                                      PAGE_SET_WRITEBACK |
 559                                                      page_error_op |
 560                                                      PAGE_END_WRITEBACK);
 561                         goto free_pages_out;
 562                 }
 563         }
 564
 565         if (will_compress) {
 566                 /*
 567                  * we aren't doing an inline extent round the compressed size
 568                  * up to a block size boundary so the allocator does sane
 569                  * things
 570                  */
 571                 total_compressed = ALIGN(total_compressed, blocksize);
 572
 573                 /*
 574                  * one last check to make sure the compression is really a
 575                  * win, compare the page count read with the blocks on disk
 576                  */
 577                 total_in = ALIGN(total_in, PAGE_CACHE_SIZE);
 578                 if (total_compressed >= total_in) {
 579                         will_compress = 0;
 580                 } else {
 581                         num_bytes = total_in;
 582                 }
 583         }
 584         if (!will_compress && pages) {
 585                 /*
 586                  * the compression code ran but failed to make things smaller,
 587                  * free any pages it allocated and our page pointer array
 588                  */
 589                 for (i = 0; i < nr_pages_ret; i++) {
 590                         WARN_ON(pages[i]->mapping);
 591                         page_cache_release(pages[i]);
 592                 }
 593                 kfree(pages);
 594                 pages = NULL;
 595                 total_compressed = 0;
 596                 nr_pages_ret = 0;
 597
 598                 /* flag the file so we don't compress in the future */
 599                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 600                     !(BTRFS_I(inode)->force_compress)) {
 601                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 602                 }
 603         }
 604         if (will_compress) {
 605                 *num_added += 1;
 606
 607                 /* the async work queues will take care of doing actual
 608                  * allocation on disk for these compressed pages,
 609                  * and will submit them to the elevator.
 610                  */
 611                 add_async_extent(async_cow, start, num_bytes,
 612                                  total_compressed, pages, nr_pages_ret,
 613                                  compress_type);
 614
 615                 if (start + num_bytes < end) {
 616                         start += num_bytes;
 617                         pages = NULL;
 618                         cond_resched();
 619                         goto again;
 620                 }
 621         } else {
 622 cleanup_and_bail_uncompressed:
 623                 /*
 624                  * No compression, but we still need to write the pages in
 625                  * the file we've been given so far.  redirty the locked
 626                  * page if it corresponds to our extent and set things up
 627                  * for the async work queue to run cow_file_range to do
 628                  * the normal delalloc dance
 629                  */
 630                 if (page_offset(locked_page) >= start &&
 631                     page_offset(locked_page) <= end) {
 632                         __set_page_dirty_nobuffers(locked_page);
 633                         /* unlocked later on in the async handlers */
 634                 }
 635                 if (redirty)
 636                         extent_range_redirty_for_io(inode, start, end);
 637                 add_async_extent(async_cow, start, end - start + 1,
 638                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 639                 *num_added += 1;
 640         }
 641
 642         return;
 643
 644 free_pages_out:
 645         for (i = 0; i < nr_pages_ret; i++) {
 646                 WARN_ON(pages[i]->mapping);
 647                 page_cache_release(pages[i]);
 648         }
 649         kfree(pages);
 650 }
 651
 652 static void free_async_extent_pages(struct async_extent *async_extent)
 653 {
 654         int i;
 655
 656         if (!async_extent->pages)
 657                 return;
 658
 659         for (i = 0; i < async_extent->nr_pages; i++) {
 660                 WARN_ON(async_extent->pages[i]->mapping);
 661                 page_cache_release(async_extent->pages[i]);
 662         }
 663         kfree(async_extent->pages);
 664         async_extent->nr_pages = 0;
 665         async_extent->pages = NULL;
 666 }
 667
 668 /*
 669  * phase two of compressed writeback.  This is the ordered portion
 670  * of the code, which only gets called in the order the work was
 671  * queued.  We walk all the async extents created by compress_file_range
 672  * and send them down to the disk.
 673  */
 674 static noinline void submit_compressed_extents(struct inode *inode,
 675                                               struct async_cow *async_cow)
 676 {
 677         struct async_extent *async_extent;
 678         u64 alloc_hint = 0;
 679         struct btrfs_key ins;
 680         struct extent_map *em;
 681         struct btrfs_root *root = BTRFS_I(inode)->root;
 682         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 683         struct extent_io_tree *io_tree;
 684         int ret = 0;
 685
 686 again:
 687         while (!list_empty(&async_cow->extents)) {
 688                 async_extent = list_entry(async_cow->extents.next,
 689                                           struct async_extent, list);
 690                 list_del(&async_extent->list);
 691
 692                 io_tree = &BTRFS_I(inode)->io_tree;
 693
 694 retry:
 695                 /* did the compression code fall back to uncompressed IO? */
 696                 if (!async_extent->pages) {
 697                         int page_started = 0;
 698                         unsigned long nr_written = 0;
 699
 700                         lock_extent(io_tree, async_extent->start,
 701                                          async_extent->start +
 702                                          async_extent->ram_size - 1);
 703
 704                         /* allocate blocks */
 705                         ret = cow_file_range(inode, async_cow->locked_page,
 706                                              async_extent->start,
 707                                              async_extent->start +
 708                                              async_extent->ram_size - 1,
 709                                              &page_started, &nr_written, 0);
 710
 711                         /* JDM XXX */
 712
 713                         /*
 714                          * if page_started, cow_file_range inserted an
 715                          * inline extent and took care of all the unlocking
 716                          * and IO for us.  Otherwise, we need to submit
 717                          * all those pages down to the drive.
 718                          */
 719                         if (!page_started && !ret)
 720                                 extent_write_locked_range(io_tree,
 721                                                   inode, async_extent->start,
 722                                                   async_extent->start +
 723                                                   async_extent->ram_size - 1,
 724                                                   btrfs_get_extent,
 725                                                   WB_SYNC_ALL);
 726                         else if (ret)
 727                                 unlock_page(async_cow->locked_page);
 728                         kfree(async_extent);
 729                         cond_resched();
 730                         continue;
 731                 }
 732
 733                 lock_extent(io_tree, async_extent->start,
 734                             async_extent->start + async_extent->ram_size - 1);
 735
 736                 ret = btrfs_reserve_extent(root,
 737                                            async_extent->compressed_size,
 738                                            async_extent->compressed_size,
 739                                            0, alloc_hint, &ins, 1, 1);
 740                 if (ret) {
 741                         free_async_extent_pages(async_extent);
 742
 743                         if (ret == -ENOSPC) {
 744                                 unlock_extent(io_tree, async_extent->start,
 745                                               async_extent->start +
 746                                               async_extent->ram_size - 1);
 747
 748                                 /*
 749                                  * we need to redirty the pages if we decide to
 750                                  * fallback to uncompressed IO, otherwise we
 751                                  * will not submit these pages down to lower
 752                                  * layers.
 753                                  */
 754                                 extent_range_redirty_for_io(inode,
 755                                                 async_extent->start,
 756                                                 async_extent->start +
 757                                                 async_extent->ram_size - 1);
 758
 759                                 goto retry;
 760                         }
 761                         goto out_free;
 762                 }
 763                 /*
 764                  * here we're doing allocation and writeback of the
 765                  * compressed pages
 766                  */
 767                 btrfs_drop_extent_cache(inode, async_extent->start,
 768                                         async_extent->start +
 769                                         async_extent->ram_size - 1, 0);
 770
 771                 em = alloc_extent_map();
 772                 if (!em) {
 773                         ret = -ENOMEM;
 774                         goto out_free_reserve;
 775                 }
 776                 em->start = async_extent->start;
 777                 em->len = async_extent->ram_size;
 778                 em->orig_start = em->start;
 779                 em->mod_start = em->start;
 780                 em->mod_len = em->len;
 781
 782                 em->block_start = ins.objectid;
 783                 em->block_len = ins.offset;
 784                 em->orig_block_len = ins.offset;
 785                 em->ram_bytes = async_extent->ram_size;
 786                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 787                 em->compress_type = async_extent->compress_type;
 788                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 789                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 790                 em->generation = -1;
 791
 792                 while (1) {
 793                         write_lock(&em_tree->lock);
 794                         ret = add_extent_mapping(em_tree, em, 1);
 795                         write_unlock(&em_tree->lock);
 796                         if (ret != -EEXIST) {
 797                                 free_extent_map(em);
 798                                 break;
 799                         }
 800                         btrfs_drop_extent_cache(inode, async_extent->start,
 801                                                 async_extent->start +
 802                                                 async_extent->ram_size - 1, 0);
 803                 }
 804
 805                 if (ret)
 806                         goto out_free_reserve;
 807
 808                 ret = btrfs_add_ordered_extent_compress(inode,
 809                                                 async_extent->start,
 810                                                 ins.objectid,
 811                                                 async_extent->ram_size,
 812                                                 ins.offset,
 813                                                 BTRFS_ORDERED_COMPRESSED,
 814                                                 async_extent->compress_type);
 815                 if (ret) {
 816                         btrfs_drop_extent_cache(inode, async_extent->start,
 817                                                 async_extent->start +
 818                                                 async_extent->ram_size - 1, 0);
 819                         goto out_free_reserve;
 820                 }
 821
 822                 /*
 823                  * clear dirty, set writeback and unlock the pages.
 824                  */
 825                 extent_clear_unlock_delalloc(inode, async_extent->start,
 826                                 async_extent->start +
 827                                 async_extent->ram_size - 1,
 828                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 829                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 830                                 PAGE_SET_WRITEBACK);
 831                 ret = btrfs_submit_compressed_write(inode,
 832                                     async_extent->start,
 833                                     async_extent->ram_size,
 834                                     ins.objectid,
 835                                     ins.offset, async_extent->pages,
 836                                     async_extent->nr_pages);
 837                 if (ret) {
 838                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 839                         struct page *p = async_extent->pages[0];
 840                         const u64 start = async_extent->start;
 841                         const u64 end = start + async_extent->ram_size - 1;
 842
 843                         p->mapping = inode->i_mapping;
 844                         tree->ops->writepage_end_io_hook(p, start, end,
 845                                                          NULL, 0);
 846                         p->mapping = NULL;
 847                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 848                                                      PAGE_END_WRITEBACK |
 849                                                      PAGE_SET_ERROR);
 850                         free_async_extent_pages(async_extent);
 851                 }
 852                 alloc_hint = ins.objectid + ins.offset;
 853                 kfree(async_extent);
 854                 cond_resched();
 855         }
 856         return;
 857 out_free_reserve:
 858         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 859 out_free:
 860         extent_clear_unlock_delalloc(inode, async_extent->start,
 861                                      async_extent->start +
 862                                      async_extent->ram_size - 1,
 863                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 864                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 865                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 866                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 867                                      PAGE_SET_ERROR);
 868         free_async_extent_pages(async_extent);
 869         kfree(async_extent);
 870         goto again;
 871 }
 872
 873 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 874                                       u64 num_bytes)
 875 {
 876         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 877         struct extent_map *em;
 878         u64 alloc_hint = 0;
 879
 880         read_lock(&em_tree->lock);
 881         em = search_extent_mapping(em_tree, start, num_bytes);
 882         if (em) {
 883                 /*
 884                  * if block start isn't an actual block number then find the
 885                  * first block in this inode and use that as a hint.  If that
 886                  * block is also bogus then just don't worry about it.
 887                  */
 888                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 889                         free_extent_map(em);
 890                         em = search_extent_mapping(em_tree, 0, 0);
 891                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 892                                 alloc_hint = em->block_start;
 893                         if (em)
 894                                 free_extent_map(em);
 895                 } else {
 896                         alloc_hint = em->block_start;
 897                         free_extent_map(em);
 898                 }
 899         }
 900         read_unlock(&em_tree->lock);
 901
 902         return alloc_hint;
 903 }
 904
 905 /*
 906  * when extent_io.c finds a delayed allocation range in the file,
 907  * the call backs end up in this code.  The basic idea is to
 908  * allocate extents on disk for the range, and create ordered data structs
 909  * in ram to track those extents.
 910  *
 911  * locked_page is the page that writepage had locked already.  We use
 912  * it to make sure we don't do extra locks or unlocks.
 913  *
 914  * *page_started is set to one if we unlock locked_page and do everything
 915  * required to start IO on it.  It may be clean and already done with
 916  * IO when we return.
 917  */
 918 static noinline int cow_file_range(struct inode *inode,
 919                                    struct page *locked_page,
 920                                    u64 start, u64 end, int *page_started,
 921                                    unsigned long *nr_written,
 922                                    int unlock)
 923 {
 924         struct btrfs_root *root = BTRFS_I(inode)->root;
 925         u64 alloc_hint = 0;
 926         u64 num_bytes;
 927         unsigned long ram_size;
 928         u64 disk_num_bytes;
 929         u64 cur_alloc_size;
 930         u64 blocksize = root->sectorsize;
 931         struct btrfs_key ins;
 932         struct extent_map *em;
 933         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 934         int ret = 0;
 935
 936         if (btrfs_is_free_space_inode(inode)) {
 937                 WARN_ON_ONCE(1);
 938                 ret = -EINVAL;
 939                 goto out_unlock;
 940         }
 941
 942         num_bytes = ALIGN(end - start + 1, blocksize);
 943         num_bytes = max(blocksize,  num_bytes);
 944         disk_num_bytes = num_bytes;
 945
 946         /* if this is a small write inside eof, kick off defrag */
 947         if (num_bytes < 64 * 1024 &&
 948             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 949                 btrfs_add_inode_defrag(NULL, inode);
 950
 951         if (start == 0) {
 952                 /* lets try to make an inline extent */
 953                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 954                                             NULL);
 955                 if (ret == 0) {
 956                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 957                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 958                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 959                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 960                                      PAGE_END_WRITEBACK);
 961
 962                         *nr_written = *nr_written +
 963                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 964                         *page_started = 1;
 965                         goto out;
 966                 } else if (ret < 0) {
 967                         goto out_unlock;
 968                 }
 969         }
 970
 971         BUG_ON(disk_num_bytes >
 972                btrfs_super_total_bytes(root->fs_info->super_copy));
 973
 974         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 975         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 976
 977         while (disk_num_bytes > 0) {
 978                 unsigned long op;
 979
 980                 cur_alloc_size = disk_num_bytes;
 981                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 982                                            root->sectorsize, 0, alloc_hint,
 983                                            &ins, 1, 1);
 984                 if (ret < 0)
 985                         goto out_unlock;
 986
 987                 em = alloc_extent_map();
 988                 if (!em) {
 989                         ret = -ENOMEM;
 990                         goto out_reserve;
 991                 }
 992                 em->start = start;
 993                 em->orig_start = em->start;
 994                 ram_size = ins.offset;
 995                 em->len = ins.offset;
 996                 em->mod_start = em->start;
 997                 em->mod_len = em->len;
 998
 999                 em->block_start = ins.objectid;
1000                 em->block_len = ins.offset;
1001                 em->orig_block_len = ins.offset;
1002                 em->ram_bytes = ram_size;
1003                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1004                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1005                 em->generation = -1;
1006
1007                 while (1) {
1008                         write_lock(&em_tree->lock);
1009                         ret = add_extent_mapping(em_tree, em, 1);
1010                         write_unlock(&em_tree->lock);
1011                         if (ret != -EEXIST) {
1012                                 free_extent_map(em);
1013                                 break;
1014                         }
1015                         btrfs_drop_extent_cache(inode, start,
1016                                                 start + ram_size - 1, 0);
1017                 }
1018                 if (ret)
1019                         goto out_reserve;
1020
1021                 cur_alloc_size = ins.offset;
1022                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1023                                                ram_size, cur_alloc_size, 0);
1024                 if (ret)
1025                         goto out_drop_extent_cache;
1026
1027                 if (root->root_key.objectid ==
1028                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1029                         ret = btrfs_reloc_clone_csums(inode, start,
1030                                                       cur_alloc_size);
1031                         if (ret)
1032                                 goto out_drop_extent_cache;
1033                 }
1034
1035                 if (disk_num_bytes < cur_alloc_size)
1036                         break;
1037
1038                 /* we're not doing compressed IO, don't unlock the first
1039                  * page (which the caller expects to stay locked), don't
1040                  * clear any dirty bits and don't set any writeback bits
1041                  *
1042                  * Do set the Private2 bit so we know this page was properly
1043                  * setup for writepage
1044                  */
1045                 op = unlock ? PAGE_UNLOCK : 0;
1046                 op |= PAGE_SET_PRIVATE2;
1047
1048                 extent_clear_unlock_delalloc(inode, start,
1049                                              start + ram_size - 1, locked_page,
1050                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1051                                              op);
1052                 disk_num_bytes -= cur_alloc_size;
1053                 num_bytes -= cur_alloc_size;
1054                 alloc_hint = ins.objectid + ins.offset;
1055                 start += cur_alloc_size;
1056         }
1057 out:
1058         return ret;
1059
1060 out_drop_extent_cache:
1061         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1062 out_reserve:
1063         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1064 out_unlock:
1065         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1066                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1067                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1068                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1069                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1070         goto out;
1071 }
1072
1073 /*
1074  * work queue call back to started compression on a file and pages
1075  */
1076 static noinline void async_cow_start(struct btrfs_work *work)
1077 {
1078         struct async_cow *async_cow;
1079         int num_added = 0;
1080         async_cow = container_of(work, struct async_cow, work);
1081
1082         compress_file_range(async_cow->inode, async_cow->locked_page,
1083                             async_cow->start, async_cow->end, async_cow,
1084                             &num_added);
1085         if (num_added == 0) {
1086                 btrfs_add_delayed_iput(async_cow->inode);
1087                 async_cow->inode = NULL;
1088         }
1089 }
1090
1091 /*
1092  * work queue call back to submit previously compressed pages
1093  */
1094 static noinline void async_cow_submit(struct btrfs_work *work)
1095 {
1096         struct async_cow *async_cow;
1097         struct btrfs_root *root;
1098         unsigned long nr_pages;
1099
1100         async_cow = container_of(work, struct async_cow, work);
1101
1102         root = async_cow->root;
1103         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
1104                 PAGE_CACHE_SHIFT;
1105
1106         /*
1107          * atomic_sub_return implies a barrier for waitqueue_active
1108          */
1109         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1110             5 * 1024 * 1024 &&
1111             waitqueue_active(&root->fs_info->async_submit_wait))
1112                 wake_up(&root->fs_info->async_submit_wait);
1113
1114         if (async_cow->inode)
1115                 submit_compressed_extents(async_cow->inode, async_cow);
1116 }
1117
1118 static noinline void async_cow_free(struct btrfs_work *work)
1119 {
1120         struct async_cow *async_cow;
1121         async_cow = container_of(work, struct async_cow, work);
1122         if (async_cow->inode)
1123                 btrfs_add_delayed_iput(async_cow->inode);
1124         kfree(async_cow);
1125 }
1126
1127 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1128                                 u64 start, u64 end, int *page_started,
1129                                 unsigned long *nr_written)
1130 {
1131         struct async_cow *async_cow;
1132         struct btrfs_root *root = BTRFS_I(inode)->root;
1133         unsigned long nr_pages;
1134         u64 cur_end;
1135         int limit = 10 * 1024 * 1024;
1136
1137         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1138                          1, 0, NULL, GFP_NOFS);
1139         while (start < end) {
1140                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1141                 BUG_ON(!async_cow); /* -ENOMEM */
1142                 async_cow->inode = igrab(inode);
1143                 async_cow->root = root;
1144                 async_cow->locked_page = locked_page;
1145                 async_cow->start = start;
1146
1147                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1148                     !btrfs_test_opt(root, FORCE_COMPRESS))
1149                         cur_end = end;
1150                 else
1151                         cur_end = min(end, start + 512 * 1024 - 1);
1152
1153                 async_cow->end = cur_end;
1154                 INIT_LIST_HEAD(&async_cow->extents);
1155
1156                 btrfs_init_work(&async_cow->work,
1157                                 btrfs_delalloc_helper,
1158                                 async_cow_start, async_cow_submit,
1159                                 async_cow_free);
1160
1161                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
1162                         PAGE_CACHE_SHIFT;
1163                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1164
1165                 btrfs_queue_work(root->fs_info->delalloc_workers,
1166                                  &async_cow->work);
1167
1168                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1169                         wait_event(root->fs_info->async_submit_wait,
1170                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1171                             limit));
1172                 }
1173
1174                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1175                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1176                         wait_event(root->fs_info->async_submit_wait,
1177                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1178                            0));
1179                 }
1180
1181                 *nr_written += nr_pages;
1182                 start = cur_end + 1;
1183         }
1184         *page_started = 1;
1185         return 0;
1186 }
1187
1188 static noinline int csum_exist_in_range(struct btrfs_root *root,
1189                                         u64 bytenr, u64 num_bytes)
1190 {
1191         int ret;
1192         struct btrfs_ordered_sum *sums;
1193         LIST_HEAD(list);
1194
1195         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1196                                        bytenr + num_bytes - 1, &list, 0);
1197         if (ret == 0 && list_empty(&list))
1198                 return 0;
1199
1200         while (!list_empty(&list)) {
1201                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1202                 list_del(&sums->list);
1203                 kfree(sums);
1204         }
1205         return 1;
1206 }
1207
1208 /*
1209  * when nowcow writeback call back.  This checks for snapshots or COW copies
1210  * of the extents that exist in the file, and COWs the file as required.
1211  *
1212  * If no cow copies or snapshots exist, we write directly to the existing
1213  * blocks on disk
1214  */
1215 static noinline int run_delalloc_nocow(struct inode *inode,
1216                                        struct page *locked_page,
1217                               u64 start, u64 end, int *page_started, int force,
1218                               unsigned long *nr_written)
1219 {
1220         struct btrfs_root *root = BTRFS_I(inode)->root;
1221         struct btrfs_trans_handle *trans;
1222         struct extent_buffer *leaf;
1223         struct btrfs_path *path;
1224         struct btrfs_file_extent_item *fi;
1225         struct btrfs_key found_key;
1226         u64 cow_start;
1227         u64 cur_offset;
1228         u64 extent_end;
1229         u64 extent_offset;
1230         u64 disk_bytenr;
1231         u64 num_bytes;
1232         u64 disk_num_bytes;
1233         u64 ram_bytes;
1234         int extent_type;
1235         int ret, err;
1236         int type;
1237         int nocow;
1238         int check_prev = 1;
1239         bool nolock;
1240         u64 ino = btrfs_ino(inode);
1241
1242         path = btrfs_alloc_path();
1243         if (!path) {
1244                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1245                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1246                                              EXTENT_DO_ACCOUNTING |
1247                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1248                                              PAGE_CLEAR_DIRTY |
1249                                              PAGE_SET_WRITEBACK |
1250                                              PAGE_END_WRITEBACK);
1251                 return -ENOMEM;
1252         }
1253
1254         nolock = btrfs_is_free_space_inode(inode);
1255
1256         if (nolock)
1257                 trans = btrfs_join_transaction_nolock(root);
1258         else
1259                 trans = btrfs_join_transaction(root);
1260
1261         if (IS_ERR(trans)) {
1262                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1263                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1264                                              EXTENT_DO_ACCOUNTING |
1265                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1266                                              PAGE_CLEAR_DIRTY |
1267                                              PAGE_SET_WRITEBACK |
1268                                              PAGE_END_WRITEBACK);
1269                 btrfs_free_path(path);
1270                 return PTR_ERR(trans);
1271         }
1272
1273         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1274
1275         cow_start = (u64)-1;
1276         cur_offset = start;
1277         while (1) {
1278                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1279                                                cur_offset, 0);
1280                 if (ret < 0)
1281                         goto error;
1282                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1283                         leaf = path->nodes[0];
1284                         btrfs_item_key_to_cpu(leaf, &found_key,
1285                                               path->slots[0] - 1);
1286                         if (found_key.objectid == ino &&
1287                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1288                                 path->slots[0]--;
1289                 }
1290                 check_prev = 0;
1291 next_slot:
1292                 leaf = path->nodes[0];
1293                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1294                         ret = btrfs_next_leaf(root, path);
1295                         if (ret < 0)
1296                                 goto error;
1297                         if (ret > 0)
1298                                 break;
1299                         leaf = path->nodes[0];
1300                 }
1301
1302                 nocow = 0;
1303                 disk_bytenr = 0;
1304                 num_bytes = 0;
1305                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1306
1307                 if (found_key.objectid > ino)
1308                         break;
1309                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1310                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1311                         path->slots[0]++;
1312                         goto next_slot;
1313                 }
1314                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1315                     found_key.offset > end)
1316                         break;
1317
1318                 if (found_key.offset > cur_offset) {
1319                         extent_end = found_key.offset;
1320                         extent_type = 0;
1321                         goto out_check;
1322                 }
1323
1324                 fi = btrfs_item_ptr(leaf, path->slots[0],
1325                                     struct btrfs_file_extent_item);
1326                 extent_type = btrfs_file_extent_type(leaf, fi);
1327
1328                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1329                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1330                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1331                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1332                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1333                         extent_end = found_key.offset +
1334                                 btrfs_file_extent_num_bytes(leaf, fi);
1335                         disk_num_bytes =
1336                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1337                         if (extent_end <= start) {
1338                                 path->slots[0]++;
1339                                 goto next_slot;
1340                         }
1341                         if (disk_bytenr == 0)
1342                                 goto out_check;
1343                         if (btrfs_file_extent_compression(leaf, fi) ||
1344                             btrfs_file_extent_encryption(leaf, fi) ||
1345                             btrfs_file_extent_other_encoding(leaf, fi))
1346                                 goto out_check;
1347                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1348                                 goto out_check;
1349                         if (btrfs_extent_readonly(root, disk_bytenr))
1350                                 goto out_check;
1351                         if (btrfs_cross_ref_exist(trans, root, ino,
1352                                                   found_key.offset -
1353                                                   extent_offset, disk_bytenr))
1354                                 goto out_check;
1355                         disk_bytenr += extent_offset;
1356                         disk_bytenr += cur_offset - found_key.offset;
1357                         num_bytes = min(end + 1, extent_end) - cur_offset;
1358                         /*
1359                          * if there are pending snapshots for this root,
1360                          * we fall into common COW way.
1361                          */
1362                         if (!nolock) {
1363                                 err = btrfs_start_write_no_snapshoting(root);
1364                                 if (!err)
1365                                         goto out_check;
1366                         }
1367                         /*
1368                          * force cow if csum exists in the range.
1369                          * this ensure that csum for a given extent are
1370                          * either valid or do not exist.
1371                          */
1372                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1373                                 goto out_check;
1374                         nocow = 1;
1375                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1376                         extent_end = found_key.offset +
1377                                 btrfs_file_extent_inline_len(leaf,
1378                                                      path->slots[0], fi);
1379                         extent_end = ALIGN(extent_end, root->sectorsize);
1380                 } else {
1381                         BUG_ON(1);
1382                 }
1383 out_check:
1384                 if (extent_end <= start) {
1385                         path->slots[0]++;
1386                         if (!nolock && nocow)
1387                                 btrfs_end_write_no_snapshoting(root);
1388                         goto next_slot;
1389                 }
1390                 if (!nocow) {
1391                         if (cow_start == (u64)-1)
1392                                 cow_start = cur_offset;
1393                         cur_offset = extent_end;
1394                         if (cur_offset > end)
1395                                 break;
1396                         path->slots[0]++;
1397                         goto next_slot;
1398                 }
1399
1400                 btrfs_release_path(path);
1401                 if (cow_start != (u64)-1) {
1402                         ret = cow_file_range(inode, locked_page,
1403                                              cow_start, found_key.offset - 1,
1404                                              page_started, nr_written, 1);
1405                         if (ret) {
1406                                 if (!nolock && nocow)
1407                                         btrfs_end_write_no_snapshoting(root);
1408                                 goto error;
1409                         }
1410                         cow_start = (u64)-1;
1411                 }
1412
1413                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1414                         struct extent_map *em;
1415                         struct extent_map_tree *em_tree;
1416                         em_tree = &BTRFS_I(inode)->extent_tree;
1417                         em = alloc_extent_map();
1418                         BUG_ON(!em); /* -ENOMEM */
1419                         em->start = cur_offset;
1420                         em->orig_start = found_key.offset - extent_offset;
1421                         em->len = num_bytes;
1422                         em->block_len = num_bytes;
1423                         em->block_start = disk_bytenr;
1424                         em->orig_block_len = disk_num_bytes;
1425                         em->ram_bytes = ram_bytes;
1426                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1427                         em->mod_start = em->start;
1428                         em->mod_len = em->len;
1429                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1430                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1431                         em->generation = -1;
1432                         while (1) {
1433                                 write_lock(&em_tree->lock);
1434                                 ret = add_extent_mapping(em_tree, em, 1);
1435                                 write_unlock(&em_tree->lock);
1436                                 if (ret != -EEXIST) {
1437                                         free_extent_map(em);
1438                                         break;
1439                                 }
1440                                 btrfs_drop_extent_cache(inode, em->start,
1441                                                 em->start + em->len - 1, 0);
1442                         }
1443                         type = BTRFS_ORDERED_PREALLOC;
1444                 } else {
1445                         type = BTRFS_ORDERED_NOCOW;
1446                 }
1447
1448                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1449                                                num_bytes, num_bytes, type);
1450                 BUG_ON(ret); /* -ENOMEM */
1451
1452                 if (root->root_key.objectid ==
1453                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1454                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1455                                                       num_bytes);
1456                         if (ret) {
1457                                 if (!nolock && nocow)
1458                                         btrfs_end_write_no_snapshoting(root);
1459                                 goto error;
1460                         }
1461                 }
1462
1463                 extent_clear_unlock_delalloc(inode, cur_offset,
1464                                              cur_offset + num_bytes - 1,
1465                                              locked_page, EXTENT_LOCKED |
1466                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1467                                              PAGE_SET_PRIVATE2);
1468                 if (!nolock && nocow)
1469                         btrfs_end_write_no_snapshoting(root);
1470                 cur_offset = extent_end;
1471                 if (cur_offset > end)
1472                         break;
1473         }
1474         btrfs_release_path(path);
1475
1476         if (cur_offset <= end && cow_start == (u64)-1) {
1477                 cow_start = cur_offset;
1478                 cur_offset = end;
1479         }
1480
1481         if (cow_start != (u64)-1) {
1482                 ret = cow_file_range(inode, locked_page, cow_start, end,
1483                                      page_started, nr_written, 1);
1484                 if (ret)
1485                         goto error;
1486         }
1487
1488 error:
1489         err = btrfs_end_transaction(trans, root);
1490         if (!ret)
1491                 ret = err;
1492
1493         if (ret && cur_offset < end)
1494                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1495                                              locked_page, EXTENT_LOCKED |
1496                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1497                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1498                                              PAGE_CLEAR_DIRTY |
1499                                              PAGE_SET_WRITEBACK |
1500                                              PAGE_END_WRITEBACK);
1501         btrfs_free_path(path);
1502         return ret;
1503 }
1504
1505 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1506 {
1507
1508         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1509             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1510                 return 0;
1511
1512         /*
1513          * @defrag_bytes is a hint value, no spinlock held here,
1514          * if is not zero, it means the file is defragging.
1515          * Force cow if given extent needs to be defragged.
1516          */
1517         if (BTRFS_I(inode)->defrag_bytes &&
1518             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1519                            EXTENT_DEFRAG, 0, NULL))
1520                 return 1;
1521
1522         return 0;
1523 }
1524
1525 /*
1526  * extent_io.c call back to do delayed allocation processing
1527  */
1528 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1529                               u64 start, u64 end, int *page_started,
1530                               unsigned long *nr_written)
1531 {
1532         int ret;
1533         int force_cow = need_force_cow(inode, start, end);
1534
1535         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1536                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1537                                          page_started, 1, nr_written);
1538         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1539                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1540                                          page_started, 0, nr_written);
1541         } else if (!inode_need_compress(inode)) {
1542                 ret = cow_file_range(inode, locked_page, start, end,
1543                                       page_started, nr_written, 1);
1544         } else {
1545                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1546                         &BTRFS_I(inode)->runtime_flags);
1547                 ret = cow_file_range_async(inode, locked_page, start, end,
1548                                            page_started, nr_written);
1549         }
1550         return ret;
1551 }
1552
1553 static void btrfs_split_extent_hook(struct inode *inode,
1554                                     struct extent_state *orig, u64 split)
1555 {
1556         u64 size;
1557
1558         /* not delalloc, ignore it */
1559         if (!(orig->state & EXTENT_DELALLOC))
1560                 return;
1561
1562         size = orig->end - orig->start + 1;
1563         if (size > BTRFS_MAX_EXTENT_SIZE) {
1564                 u64 num_extents;
1565                 u64 new_size;
1566
1567                 /*
1568                  * See the explanation in btrfs_merge_extent_hook, the same
1569                  * applies here, just in reverse.
1570                  */
1571                 new_size = orig->end - split + 1;
1572                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1573                                         BTRFS_MAX_EXTENT_SIZE);
1574                 new_size = split - orig->start;
1575                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1576                                         BTRFS_MAX_EXTENT_SIZE);
1577                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1578                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1579                         return;
1580         }
1581
1582         spin_lock(&BTRFS_I(inode)->lock);
1583         BTRFS_I(inode)->outstanding_extents++;
1584         spin_unlock(&BTRFS_I(inode)->lock);
1585 }
1586
1587 /*
1588  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1589  * extents so we can keep track of new extents that are just merged onto old
1590  * extents, such as when we are doing sequential writes, so we can properly
1591  * account for the metadata space we'll need.
1592  */
1593 static void btrfs_merge_extent_hook(struct inode *inode,
1594                                     struct extent_state *new,
1595                                     struct extent_state *other)
1596 {
1597         u64 new_size, old_size;
1598         u64 num_extents;
1599
1600         /* not delalloc, ignore it */
1601         if (!(other->state & EXTENT_DELALLOC))
1602                 return;
1603
1604         if (new->start > other->start)
1605                 new_size = new->end - other->start + 1;
1606         else
1607                 new_size = other->end - new->start + 1;
1608
1609         /* we're not bigger than the max, unreserve the space and go */
1610         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1611                 spin_lock(&BTRFS_I(inode)->lock);
1612                 BTRFS_I(inode)->outstanding_extents--;
1613                 spin_unlock(&BTRFS_I(inode)->lock);
1614                 return;
1615         }
1616
1617         /*
1618          * We have to add up either side to figure out how many extents were
1619          * accounted for before we merged into one big extent.  If the number of
1620          * extents we accounted for is <= the amount we need for the new range
1621          * then we can return, otherwise drop.  Think of it like this
1622          *
1623          * [ 4k][MAX_SIZE]
1624          *
1625          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1626          * need 2 outstanding extents, on one side we have 1 and the other side
1627          * we have 1 so they are == and we can return.  But in this case
1628          *
1629          * [MAX_SIZE+4k][MAX_SIZE+4k]
1630          *
1631          * Each range on their own accounts for 2 extents, but merged together
1632          * they are only 3 extents worth of accounting, so we need to drop in
1633          * this case.
1634          */
1635         old_size = other->end - other->start + 1;
1636         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1637                                 BTRFS_MAX_EXTENT_SIZE);
1638         old_size = new->end - new->start + 1;
1639         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1640                                  BTRFS_MAX_EXTENT_SIZE);
1641
1642         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1643                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1644                 return;
1645
1646         spin_lock(&BTRFS_I(inode)->lock);
1647         BTRFS_I(inode)->outstanding_extents--;
1648         spin_unlock(&BTRFS_I(inode)->lock);
1649 }
1650
1651 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1652                                       struct inode *inode)
1653 {
1654         spin_lock(&root->delalloc_lock);
1655         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1656                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1657                               &root->delalloc_inodes);
1658                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1659                         &BTRFS_I(inode)->runtime_flags);
1660                 root->nr_delalloc_inodes++;
1661                 if (root->nr_delalloc_inodes == 1) {
1662                         spin_lock(&root->fs_info->delalloc_root_lock);
1663                         BUG_ON(!list_empty(&root->delalloc_root));
1664                         list_add_tail(&root->delalloc_root,
1665                                       &root->fs_info->delalloc_roots);
1666                         spin_unlock(&root->fs_info->delalloc_root_lock);
1667                 }
1668         }
1669         spin_unlock(&root->delalloc_lock);
1670 }
1671
1672 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1673                                      struct inode *inode)
1674 {
1675         spin_lock(&root->delalloc_lock);
1676         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1677                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1678                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1679                           &BTRFS_I(inode)->runtime_flags);
1680                 root->nr_delalloc_inodes--;
1681                 if (!root->nr_delalloc_inodes) {
1682                         spin_lock(&root->fs_info->delalloc_root_lock);
1683                         BUG_ON(list_empty(&root->delalloc_root));
1684                         list_del_init(&root->delalloc_root);
1685                         spin_unlock(&root->fs_info->delalloc_root_lock);
1686                 }
1687         }
1688         spin_unlock(&root->delalloc_lock);
1689 }
1690
1691 /*
1692  * extent_io.c set_bit_hook, used to track delayed allocation
1693  * bytes in this file, and to maintain the list of inodes that
1694  * have pending delalloc work to be done.
1695  */
1696 static void btrfs_set_bit_hook(struct inode *inode,
1697                                struct extent_state *state, unsigned *bits)
1698 {
1699
1700         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1701                 WARN_ON(1);
1702         /*
1703          * set_bit and clear bit hooks normally require _irqsave/restore
1704          * but in this case, we are only testing for the DELALLOC
1705          * bit, which is only set or cleared with irqs on
1706          */
1707         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1708                 struct btrfs_root *root = BTRFS_I(inode)->root;
1709                 u64 len = state->end + 1 - state->start;
1710                 bool do_list = !btrfs_is_free_space_inode(inode);
1711
1712                 if (*bits & EXTENT_FIRST_DELALLOC) {
1713                         *bits &= ~EXTENT_FIRST_DELALLOC;
1714                 } else {
1715                         spin_lock(&BTRFS_I(inode)->lock);
1716                         BTRFS_I(inode)->outstanding_extents++;
1717                         spin_unlock(&BTRFS_I(inode)->lock);
1718                 }
1719
1720                 /* For sanity tests */
1721                 if (btrfs_test_is_dummy_root(root))
1722                         return;
1723
1724                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1725                                      root->fs_info->delalloc_batch);
1726                 spin_lock(&BTRFS_I(inode)->lock);
1727                 BTRFS_I(inode)->delalloc_bytes += len;
1728                 if (*bits & EXTENT_DEFRAG)
1729                         BTRFS_I(inode)->defrag_bytes += len;
1730                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1731                                          &BTRFS_I(inode)->runtime_flags))
1732                         btrfs_add_delalloc_inodes(root, inode);
1733                 spin_unlock(&BTRFS_I(inode)->lock);
1734         }
1735 }
1736
1737 /*
1738  * extent_io.c clear_bit_hook, see set_bit_hook for why
1739  */
1740 static void btrfs_clear_bit_hook(struct inode *inode,
1741                                  struct extent_state *state,
1742                                  unsigned *bits)
1743 {
1744         u64 len = state->end + 1 - state->start;
1745         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1746                                     BTRFS_MAX_EXTENT_SIZE);
1747
1748         spin_lock(&BTRFS_I(inode)->lock);
1749         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1750                 BTRFS_I(inode)->defrag_bytes -= len;
1751         spin_unlock(&BTRFS_I(inode)->lock);
1752
1753         /*
1754          * set_bit and clear bit hooks normally require _irqsave/restore
1755          * but in this case, we are only testing for the DELALLOC
1756          * bit, which is only set or cleared with irqs on
1757          */
1758         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1759                 struct btrfs_root *root = BTRFS_I(inode)->root;
1760                 bool do_list = !btrfs_is_free_space_inode(inode);
1761
1762                 if (*bits & EXTENT_FIRST_DELALLOC) {
1763                         *bits &= ~EXTENT_FIRST_DELALLOC;
1764                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1765                         spin_lock(&BTRFS_I(inode)->lock);
1766                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1767                         spin_unlock(&BTRFS_I(inode)->lock);
1768                 }
1769
1770                 /*
1771                  * We don't reserve metadata space for space cache inodes so we
1772                  * don't need to call dellalloc_release_metadata if there is an
1773                  * error.
1774                  */
1775                 if (*bits & EXTENT_DO_ACCOUNTING &&
1776                     root != root->fs_info->tree_root)
1777                         btrfs_delalloc_release_metadata(inode, len);
1778
1779                 /* For sanity tests. */
1780                 if (btrfs_test_is_dummy_root(root))
1781                         return;
1782
1783                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1784                     && do_list && !(state->state & EXTENT_NORESERVE))
1785                         btrfs_free_reserved_data_space_noquota(inode,
1786                                         state->start, len);
1787
1788                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1789                                      root->fs_info->delalloc_batch);
1790                 spin_lock(&BTRFS_I(inode)->lock);
1791                 BTRFS_I(inode)->delalloc_bytes -= len;
1792                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1793                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1794                              &BTRFS_I(inode)->runtime_flags))
1795                         btrfs_del_delalloc_inode(root, inode);
1796                 spin_unlock(&BTRFS_I(inode)->lock);
1797         }
1798 }
1799
1800 /*
1801  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1802  * we don't create bios that span stripes or chunks
1803  */
1804 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1805                          size_t size, struct bio *bio,
1806                          unsigned long bio_flags)
1807 {
1808         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1809         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1810         u64 length = 0;
1811         u64 map_length;
1812         int ret;
1813
1814         if (bio_flags & EXTENT_BIO_COMPRESSED)
1815                 return 0;
1816
1817         length = bio->bi_iter.bi_size;
1818         map_length = length;
1819         ret = btrfs_map_block(root->fs_info, rw, logical,
1820                               &map_length, NULL, 0);
1821         /* Will always return 0 with map_multi == NULL */
1822         BUG_ON(ret < 0);
1823         if (map_length < length + size)
1824                 return 1;
1825         return 0;
1826 }
1827
1828 /*
1829  * in order to insert checksums into the metadata in large chunks,
1830  * we wait until bio submission time.   All the pages in the bio are
1831  * checksummed and sums are attached onto the ordered extent record.
1832  *
1833  * At IO completion time the cums attached on the ordered extent record
1834  * are inserted into the btree
1835  */
1836 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1837                                     struct bio *bio, int mirror_num,
1838                                     unsigned long bio_flags,
1839                                     u64 bio_offset)
1840 {
1841         struct btrfs_root *root = BTRFS_I(inode)->root;
1842         int ret = 0;
1843
1844         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1845         BUG_ON(ret); /* -ENOMEM */
1846         return 0;
1847 }
1848
1849 /*
1850  * in order to insert checksums into the metadata in large chunks,
1851  * we wait until bio submission time.   All the pages in the bio are
1852  * checksummed and sums are attached onto the ordered extent record.
1853  *
1854  * At IO completion time the cums attached on the ordered extent record
1855  * are inserted into the btree
1856  */
1857 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1858                           int mirror_num, unsigned long bio_flags,
1859                           u64 bio_offset)
1860 {
1861         struct btrfs_root *root = BTRFS_I(inode)->root;
1862         int ret;
1863
1864         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1865         if (ret) {
1866                 bio->bi_error = ret;
1867                 bio_endio(bio);
1868         }
1869         return ret;
1870 }
1871
1872 /*
1873  * extent_io.c submission hook. This does the right thing for csum calculation
1874  * on write, or reading the csums from the tree before a read
1875  */
1876 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1877                           int mirror_num, unsigned long bio_flags,
1878                           u64 bio_offset)
1879 {
1880         struct btrfs_root *root = BTRFS_I(inode)->root;
1881         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1882         int ret = 0;
1883         int skip_sum;
1884         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1885
1886         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1887
1888         if (btrfs_is_free_space_inode(inode))
1889                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1890
1891         if (!(rw & REQ_WRITE)) {
1892                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1893                 if (ret)
1894                         goto out;
1895
1896                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1897                         ret = btrfs_submit_compressed_read(inode, bio,
1898                                                            mirror_num,
1899                                                            bio_flags);
1900                         goto out;
1901                 } else if (!skip_sum) {
1902                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1903                         if (ret)
1904                                 goto out;
1905                 }
1906                 goto mapit;
1907         } else if (async && !skip_sum) {
1908                 /* csum items have already been cloned */
1909                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1910                         goto mapit;
1911                 /* we're doing a write, do the async checksumming */
1912                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1913                                    inode, rw, bio, mirror_num,
1914                                    bio_flags, bio_offset,
1915                                    __btrfs_submit_bio_start,
1916                                    __btrfs_submit_bio_done);
1917                 goto out;
1918         } else if (!skip_sum) {
1919                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1920                 if (ret)
1921                         goto out;
1922         }
1923
1924 mapit:
1925         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1926
1927 out:
1928         if (ret < 0) {
1929                 bio->bi_error = ret;
1930                 bio_endio(bio);
1931         }
1932         return ret;
1933 }
1934
1935 /*
1936  * given a list of ordered sums record them in the inode.  This happens
1937  * at IO completion time based on sums calculated at bio submission time.
1938  */
1939 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1940                              struct inode *inode, u64 file_offset,
1941                              struct list_head *list)
1942 {
1943         struct btrfs_ordered_sum *sum;
1944
1945         list_for_each_entry(sum, list, list) {
1946                 trans->adding_csums = 1;
1947                 btrfs_csum_file_blocks(trans,
1948                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1949                 trans->adding_csums = 0;
1950         }
1951         return 0;
1952 }
1953
1954 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1955                               struct extent_state **cached_state)
1956 {
1957         WARN_ON((end & (PAGE_CACHE_SIZE - 1)) == 0);
1958         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1959                                    cached_state, GFP_NOFS);
1960 }
1961
1962 /* see btrfs_writepage_start_hook for details on why this is required */
1963 struct btrfs_writepage_fixup {
1964         struct page *page;
1965         struct btrfs_work work;
1966 };
1967
1968 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1969 {
1970         struct btrfs_writepage_fixup *fixup;
1971         struct btrfs_ordered_extent *ordered;
1972         struct extent_state *cached_state = NULL;
1973         struct page *page;
1974         struct inode *inode;
1975         u64 page_start;
1976         u64 page_end;
1977         int ret;
1978
1979         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1980         page = fixup->page;
1981 again:
1982         lock_page(page);
1983         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1984                 ClearPageChecked(page);
1985                 goto out_page;
1986         }
1987
1988         inode = page->mapping->host;
1989         page_start = page_offset(page);
1990         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1991
1992         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end, 0,
1993                          &cached_state);
1994
1995         /* already ordered? We're done */
1996         if (PagePrivate2(page))
1997                 goto out;
1998
1999         ordered = btrfs_lookup_ordered_extent(inode, page_start);
2000         if (ordered) {
2001                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2002                                      page_end, &cached_state, GFP_NOFS);
2003                 unlock_page(page);
2004                 btrfs_start_ordered_extent(inode, ordered, 1);
2005                 btrfs_put_ordered_extent(ordered);
2006                 goto again;
2007         }
2008
2009         ret = btrfs_delalloc_reserve_space(inode, page_start,
2010                                            PAGE_CACHE_SIZE);
2011         if (ret) {
2012                 mapping_set_error(page->mapping, ret);
2013                 end_extent_writepage(page, ret, page_start, page_end);
2014                 ClearPageChecked(page);
2015                 goto out;
2016          }
2017
2018         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2019         ClearPageChecked(page);
2020         set_page_dirty(page);
2021 out:
2022         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2023                              &cached_state, GFP_NOFS);
2024 out_page:
2025         unlock_page(page);
2026         page_cache_release(page);
2027         kfree(fixup);
2028 }
2029
2030 /*
2031  * There are a few paths in the higher layers of the kernel that directly
2032  * set the page dirty bit without asking the filesystem if it is a
2033  * good idea.  This causes problems because we want to make sure COW
2034  * properly happens and the data=ordered rules are followed.
2035  *
2036  * In our case any range that doesn't have the ORDERED bit set
2037  * hasn't been properly setup for IO.  We kick off an async process
2038  * to fix it up.  The async helper will wait for ordered extents, set
2039  * the delalloc bit and make it safe to write the page.
2040  */
2041 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2042 {
2043         struct inode *inode = page->mapping->host;
2044         struct btrfs_writepage_fixup *fixup;
2045         struct btrfs_root *root = BTRFS_I(inode)->root;
2046
2047         /* this page is properly in the ordered list */
2048         if (TestClearPagePrivate2(page))
2049                 return 0;
2050
2051         if (PageChecked(page))
2052                 return -EAGAIN;
2053
2054         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2055         if (!fixup)
2056                 return -EAGAIN;
2057
2058         SetPageChecked(page);
2059         page_cache_get(page);
2060         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2061                         btrfs_writepage_fixup_worker, NULL, NULL);
2062         fixup->page = page;
2063         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2064         return -EBUSY;
2065 }
2066
2067 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2068                                        struct inode *inode, u64 file_pos,
2069                                        u64 disk_bytenr, u64 disk_num_bytes,
2070                                        u64 num_bytes, u64 ram_bytes,
2071                                        u8 compression, u8 encryption,
2072                                        u16 other_encoding, int extent_type)
2073 {
2074         struct btrfs_root *root = BTRFS_I(inode)->root;
2075         struct btrfs_file_extent_item *fi;
2076         struct btrfs_path *path;
2077         struct extent_buffer *leaf;
2078         struct btrfs_key ins;
2079         int extent_inserted = 0;
2080         int ret;
2081
2082         path = btrfs_alloc_path();
2083         if (!path)
2084                 return -ENOMEM;
2085
2086         /*
2087          * we may be replacing one extent in the tree with another.
2088          * The new extent is pinned in the extent map, and we don't want
2089          * to drop it from the cache until it is completely in the btree.
2090          *
2091          * So, tell btrfs_drop_extents to leave this extent in the cache.
2092          * the caller is expected to unpin it and allow it to be merged
2093          * with the others.
2094          */
2095         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2096                                    file_pos + num_bytes, NULL, 0,
2097                                    1, sizeof(*fi), &extent_inserted);
2098         if (ret)
2099                 goto out;
2100
2101         if (!extent_inserted) {
2102                 ins.objectid = btrfs_ino(inode);
2103                 ins.offset = file_pos;
2104                 ins.type = BTRFS_EXTENT_DATA_KEY;
2105
2106                 path->leave_spinning = 1;
2107                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2108                                               sizeof(*fi));
2109                 if (ret)
2110                         goto out;
2111         }
2112         leaf = path->nodes[0];
2113         fi = btrfs_item_ptr(leaf, path->slots[0],
2114                             struct btrfs_file_extent_item);
2115         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2116         btrfs_set_file_extent_type(leaf, fi, extent_type);
2117         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2118         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2119         btrfs_set_file_extent_offset(leaf, fi, 0);
2120         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2121         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2122         btrfs_set_file_extent_compression(leaf, fi, compression);
2123         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2124         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2125
2126         btrfs_mark_buffer_dirty(leaf);
2127         btrfs_release_path(path);
2128
2129         inode_add_bytes(inode, num_bytes);
2130
2131         ins.objectid = disk_bytenr;
2132         ins.offset = disk_num_bytes;
2133         ins.type = BTRFS_EXTENT_ITEM_KEY;
2134         ret = btrfs_alloc_reserved_file_extent(trans, root,
2135                                         root->root_key.objectid,
2136                                         btrfs_ino(inode), file_pos,
2137                                         ram_bytes, &ins);
2138         /*
2139          * Release the reserved range from inode dirty range map, as it is
2140          * already moved into delayed_ref_head
2141          */
2142         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2143 out:
2144         btrfs_free_path(path);
2145
2146         return ret;
2147 }
2148
2149 /* snapshot-aware defrag */
2150 struct sa_defrag_extent_backref {
2151         struct rb_node node;
2152         struct old_sa_defrag_extent *old;
2153         u64 root_id;
2154         u64 inum;
2155         u64 file_pos;
2156         u64 extent_offset;
2157         u64 num_bytes;
2158         u64 generation;
2159 };
2160
2161 struct old_sa_defrag_extent {
2162         struct list_head list;
2163         struct new_sa_defrag_extent *new;
2164
2165         u64 extent_offset;
2166         u64 bytenr;
2167         u64 offset;
2168         u64 len;
2169         int count;
2170 };
2171
2172 struct new_sa_defrag_extent {
2173         struct rb_root root;
2174         struct list_head head;
2175         struct btrfs_path *path;
2176         struct inode *inode;
2177         u64 file_pos;
2178         u64 len;
2179         u64 bytenr;
2180         u64 disk_len;
2181         u8 compress_type;
2182 };
2183
2184 static int backref_comp(struct sa_defrag_extent_backref *b1,
2185                         struct sa_defrag_extent_backref *b2)
2186 {
2187         if (b1->root_id < b2->root_id)
2188                 return -1;
2189         else if (b1->root_id > b2->root_id)
2190                 return 1;
2191
2192         if (b1->inum < b2->inum)
2193                 return -1;
2194         else if (b1->inum > b2->inum)
2195                 return 1;
2196
2197         if (b1->file_pos < b2->file_pos)
2198                 return -1;
2199         else if (b1->file_pos > b2->file_pos)
2200                 return 1;
2201
2202         /*
2203          * [------------------------------] ===> (a range of space)
2204          *     |<--->|   |<---->| =============> (fs/file tree A)
2205          * |<---------------------------->| ===> (fs/file tree B)
2206          *
2207          * A range of space can refer to two file extents in one tree while
2208          * refer to only one file extent in another tree.
2209          *
2210          * So we may process a disk offset more than one time(two extents in A)
2211          * and locate at the same extent(one extent in B), then insert two same
2212          * backrefs(both refer to the extent in B).
2213          */
2214         return 0;
2215 }
2216
2217 static void backref_insert(struct rb_root *root,
2218                            struct sa_defrag_extent_backref *backref)
2219 {
2220         struct rb_node **p = &root->rb_node;
2221         struct rb_node *parent = NULL;
2222         struct sa_defrag_extent_backref *entry;
2223         int ret;
2224
2225         while (*p) {
2226                 parent = *p;
2227                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2228
2229                 ret = backref_comp(backref, entry);
2230                 if (ret < 0)
2231                         p = &(*p)->rb_left;
2232                 else
2233                         p = &(*p)->rb_right;
2234         }
2235
2236         rb_link_node(&backref->node, parent, p);
2237         rb_insert_color(&backref->node, root);
2238 }
2239
2240 /*
2241  * Note the backref might has changed, and in this case we just return 0.
2242  */
2243 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2244                                        void *ctx)
2245 {
2246         struct btrfs_file_extent_item *extent;
2247         struct btrfs_fs_info *fs_info;
2248         struct old_sa_defrag_extent *old = ctx;
2249         struct new_sa_defrag_extent *new = old->new;
2250         struct btrfs_path *path = new->path;
2251         struct btrfs_key key;
2252         struct btrfs_root *root;
2253         struct sa_defrag_extent_backref *backref;
2254         struct extent_buffer *leaf;
2255         struct inode *inode = new->inode;
2256         int slot;
2257         int ret;
2258         u64 extent_offset;
2259         u64 num_bytes;
2260
2261         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2262             inum == btrfs_ino(inode))
2263                 return 0;
2264
2265         key.objectid = root_id;
2266         key.type = BTRFS_ROOT_ITEM_KEY;
2267         key.offset = (u64)-1;
2268
2269         fs_info = BTRFS_I(inode)->root->fs_info;
2270         root = btrfs_read_fs_root_no_name(fs_info, &key);
2271         if (IS_ERR(root)) {
2272                 if (PTR_ERR(root) == -ENOENT)
2273                         return 0;
2274                 WARN_ON(1);
2275                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2276                          inum, offset, root_id);
2277                 return PTR_ERR(root);
2278         }
2279
2280         key.objectid = inum;
2281         key.type = BTRFS_EXTENT_DATA_KEY;
2282         if (offset > (u64)-1 << 32)
2283                 key.offset = 0;
2284         else
2285                 key.offset = offset;
2286
2287         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2288         if (WARN_ON(ret < 0))
2289                 return ret;
2290         ret = 0;
2291
2292         while (1) {
2293                 cond_resched();
2294
2295                 leaf = path->nodes[0];
2296                 slot = path->slots[0];
2297
2298                 if (slot >= btrfs_header_nritems(leaf)) {
2299                         ret = btrfs_next_leaf(root, path);
2300                         if (ret < 0) {
2301                                 goto out;
2302                         } else if (ret > 0) {
2303                                 ret = 0;
2304                                 goto out;
2305                         }
2306                         continue;
2307                 }
2308
2309                 path->slots[0]++;
2310
2311                 btrfs_item_key_to_cpu(leaf, &key, slot);
2312
2313                 if (key.objectid > inum)
2314                         goto out;
2315
2316                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2317                         continue;
2318
2319                 extent = btrfs_item_ptr(leaf, slot,
2320                                         struct btrfs_file_extent_item);
2321
2322                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2323                         continue;
2324
2325                 /*
2326                  * 'offset' refers to the exact key.offset,
2327                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2328                  * (key.offset - extent_offset).
2329                  */
2330                 if (key.offset != offset)
2331                         continue;
2332
2333                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2334                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2335
2336                 if (extent_offset >= old->extent_offset + old->offset +
2337                     old->len || extent_offset + num_bytes <=
2338                     old->extent_offset + old->offset)
2339                         continue;
2340                 break;
2341         }
2342
2343         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2344         if (!backref) {
2345                 ret = -ENOENT;
2346                 goto out;
2347         }
2348
2349         backref->root_id = root_id;
2350         backref->inum = inum;
2351         backref->file_pos = offset;
2352         backref->num_bytes = num_bytes;
2353         backref->extent_offset = extent_offset;
2354         backref->generation = btrfs_file_extent_generation(leaf, extent);
2355         backref->old = old;
2356         backref_insert(&new->root, backref);
2357         old->count++;
2358 out:
2359         btrfs_release_path(path);
2360         WARN_ON(ret);
2361         return ret;
2362 }
2363
2364 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2365                                    struct new_sa_defrag_extent *new)
2366 {
2367         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2368         struct old_sa_defrag_extent *old, *tmp;
2369         int ret;
2370
2371         new->path = path;
2372
2373         list_for_each_entry_safe(old, tmp, &new->head, list) {
2374                 ret = iterate_inodes_from_logical(old->bytenr +
2375                                                   old->extent_offset, fs_info,
2376                                                   path, record_one_backref,
2377                                                   old);
2378                 if (ret < 0 && ret != -ENOENT)
2379                         return false;
2380
2381                 /* no backref to be processed for this extent */
2382                 if (!old->count) {
2383                         list_del(&old->list);
2384                         kfree(old);
2385                 }
2386         }
2387
2388         if (list_empty(&new->head))
2389                 return false;
2390
2391         return true;
2392 }
2393
2394 static int relink_is_mergable(struct extent_buffer *leaf,
2395                               struct btrfs_file_extent_item *fi,
2396                               struct new_sa_defrag_extent *new)
2397 {
2398         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2399                 return 0;
2400
2401         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2402                 return 0;
2403
2404         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2405                 return 0;
2406
2407         if (btrfs_file_extent_encryption(leaf, fi) ||
2408             btrfs_file_extent_other_encoding(leaf, fi))
2409                 return 0;
2410
2411         return 1;
2412 }
2413
2414 /*
2415  * Note the backref might has changed, and in this case we just return 0.
2416  */
2417 static noinline int relink_extent_backref(struct btrfs_path *path,
2418                                  struct sa_defrag_extent_backref *prev,
2419                                  struct sa_defrag_extent_backref *backref)
2420 {
2421         struct btrfs_file_extent_item *extent;
2422         struct btrfs_file_extent_item *item;
2423         struct btrfs_ordered_extent *ordered;
2424         struct btrfs_trans_handle *trans;
2425         struct btrfs_fs_info *fs_info;
2426         struct btrfs_root *root;
2427         struct btrfs_key key;
2428         struct extent_buffer *leaf;
2429         struct old_sa_defrag_extent *old = backref->old;
2430         struct new_sa_defrag_extent *new = old->new;
2431         struct inode *src_inode = new->inode;
2432         struct inode *inode;
2433         struct extent_state *cached = NULL;
2434         int ret = 0;
2435         u64 start;
2436         u64 len;
2437         u64 lock_start;
2438         u64 lock_end;
2439         bool merge = false;
2440         int index;
2441
2442         if (prev && prev->root_id == backref->root_id &&
2443             prev->inum == backref->inum &&
2444             prev->file_pos + prev->num_bytes == backref->file_pos)
2445                 merge = true;
2446
2447         /* step 1: get root */
2448         key.objectid = backref->root_id;
2449         key.type = BTRFS_ROOT_ITEM_KEY;
2450         key.offset = (u64)-1;
2451
2452         fs_info = BTRFS_I(src_inode)->root->fs_info;
2453         index = srcu_read_lock(&fs_info->subvol_srcu);
2454
2455         root = btrfs_read_fs_root_no_name(fs_info, &key);
2456         if (IS_ERR(root)) {
2457                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2458                 if (PTR_ERR(root) == -ENOENT)
2459                         return 0;
2460                 return PTR_ERR(root);
2461         }
2462
2463         if (btrfs_root_readonly(root)) {
2464                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2465                 return 0;
2466         }
2467
2468         /* step 2: get inode */
2469         key.objectid = backref->inum;
2470         key.type = BTRFS_INODE_ITEM_KEY;
2471         key.offset = 0;
2472
2473         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2474         if (IS_ERR(inode)) {
2475                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2476                 return 0;
2477         }
2478
2479         srcu_read_unlock(&fs_info->subvol_srcu, index);
2480
2481         /* step 3: relink backref */
2482         lock_start = backref->file_pos;
2483         lock_end = backref->file_pos + backref->num_bytes - 1;
2484         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2485                          0, &cached);
2486
2487         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2488         if (ordered) {
2489                 btrfs_put_ordered_extent(ordered);
2490                 goto out_unlock;
2491         }
2492
2493         trans = btrfs_join_transaction(root);
2494         if (IS_ERR(trans)) {
2495                 ret = PTR_ERR(trans);
2496                 goto out_unlock;
2497         }
2498
2499         key.objectid = backref->inum;
2500         key.type = BTRFS_EXTENT_DATA_KEY;
2501         key.offset = backref->file_pos;
2502
2503         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2504         if (ret < 0) {
2505                 goto out_free_path;
2506         } else if (ret > 0) {
2507                 ret = 0;
2508                 goto out_free_path;
2509         }
2510
2511         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2512                                 struct btrfs_file_extent_item);
2513
2514         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2515             backref->generation)
2516                 goto out_free_path;
2517
2518         btrfs_release_path(path);
2519
2520         start = backref->file_pos;
2521         if (backref->extent_offset < old->extent_offset + old->offset)
2522                 start += old->extent_offset + old->offset -
2523                          backref->extent_offset;
2524
2525         len = min(backref->extent_offset + backref->num_bytes,
2526                   old->extent_offset + old->offset + old->len);
2527         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2528
2529         ret = btrfs_drop_extents(trans, root, inode, start,
2530                                  start + len, 1);
2531         if (ret)
2532                 goto out_free_path;
2533 again:
2534         key.objectid = btrfs_ino(inode);
2535         key.type = BTRFS_EXTENT_DATA_KEY;
2536         key.offset = start;
2537
2538         path->leave_spinning = 1;
2539         if (merge) {
2540                 struct btrfs_file_extent_item *fi;
2541                 u64 extent_len;
2542                 struct btrfs_key found_key;
2543
2544                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2545                 if (ret < 0)
2546                         goto out_free_path;
2547
2548                 path->slots[0]--;
2549                 leaf = path->nodes[0];
2550                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2551
2552                 fi = btrfs_item_ptr(leaf, path->slots[0],
2553                                     struct btrfs_file_extent_item);
2554                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2555
2556                 if (extent_len + found_key.offset == start &&
2557                     relink_is_mergable(leaf, fi, new)) {
2558                         btrfs_set_file_extent_num_bytes(leaf, fi,
2559                                                         extent_len + len);
2560                         btrfs_mark_buffer_dirty(leaf);
2561                         inode_add_bytes(inode, len);
2562
2563                         ret = 1;
2564                         goto out_free_path;
2565                 } else {
2566                         merge = false;
2567                         btrfs_release_path(path);
2568                         goto again;
2569                 }
2570         }
2571
2572         ret = btrfs_insert_empty_item(trans, root, path, &key,
2573                                         sizeof(*extent));
2574         if (ret) {
2575                 btrfs_abort_transaction(trans, root, ret);
2576                 goto out_free_path;
2577         }
2578
2579         leaf = path->nodes[0];
2580         item = btrfs_item_ptr(leaf, path->slots[0],
2581                                 struct btrfs_file_extent_item);
2582         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2583         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2584         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2585         btrfs_set_file_extent_num_bytes(leaf, item, len);
2586         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2587         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2588         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2589         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2590         btrfs_set_file_extent_encryption(leaf, item, 0);
2591         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2592
2593         btrfs_mark_buffer_dirty(leaf);
2594         inode_add_bytes(inode, len);
2595         btrfs_release_path(path);
2596
2597         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2598                         new->disk_len, 0,
2599                         backref->root_id, backref->inum,
2600                         new->file_pos); /* start - extent_offset */
2601         if (ret) {
2602                 btrfs_abort_transaction(trans, root, ret);
2603                 goto out_free_path;
2604         }
2605
2606         ret = 1;
2607 out_free_path:
2608         btrfs_release_path(path);
2609         path->leave_spinning = 0;
2610         btrfs_end_transaction(trans, root);
2611 out_unlock:
2612         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2613                              &cached, GFP_NOFS);
2614         iput(inode);
2615         return ret;
2616 }
2617
2618 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2619 {
2620         struct old_sa_defrag_extent *old, *tmp;
2621
2622         if (!new)
2623                 return;
2624
2625         list_for_each_entry_safe(old, tmp, &new->head, list) {
2626                 kfree(old);
2627         }
2628         kfree(new);
2629 }
2630
2631 static void relink_file_extents(struct new_sa_defrag_extent *new)
2632 {
2633         struct btrfs_path *path;
2634         struct sa_defrag_extent_backref *backref;
2635         struct sa_defrag_extent_backref *prev = NULL;
2636         struct inode *inode;
2637         struct btrfs_root *root;
2638         struct rb_node *node;
2639         int ret;
2640
2641         inode = new->inode;
2642         root = BTRFS_I(inode)->root;
2643
2644         path = btrfs_alloc_path();
2645         if (!path)
2646                 return;
2647
2648         if (!record_extent_backrefs(path, new)) {
2649                 btrfs_free_path(path);
2650                 goto out;
2651         }
2652         btrfs_release_path(path);
2653
2654         while (1) {
2655                 node = rb_first(&new->root);
2656                 if (!node)
2657                         break;
2658                 rb_erase(node, &new->root);
2659
2660                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2661
2662                 ret = relink_extent_backref(path, prev, backref);
2663                 WARN_ON(ret < 0);
2664
2665                 kfree(prev);
2666
2667                 if (ret == 1)
2668                         prev = backref;
2669                 else
2670                         prev = NULL;
2671                 cond_resched();
2672         }
2673         kfree(prev);
2674
2675         btrfs_free_path(path);
2676 out:
2677         free_sa_defrag_extent(new);
2678
2679         atomic_dec(&root->fs_info->defrag_running);
2680         wake_up(&root->fs_info->transaction_wait);
2681 }
2682
2683 static struct new_sa_defrag_extent *
2684 record_old_file_extents(struct inode *inode,
2685                         struct btrfs_ordered_extent *ordered)
2686 {
2687         struct btrfs_root *root = BTRFS_I(inode)->root;
2688         struct btrfs_path *path;
2689         struct btrfs_key key;
2690         struct old_sa_defrag_extent *old;
2691         struct new_sa_defrag_extent *new;
2692         int ret;
2693
2694         new = kmalloc(sizeof(*new), GFP_NOFS);
2695         if (!new)
2696                 return NULL;
2697
2698         new->inode = inode;
2699         new->file_pos = ordered->file_offset;
2700         new->len = ordered->len;
2701         new->bytenr = ordered->start;
2702         new->disk_len = ordered->disk_len;
2703         new->compress_type = ordered->compress_type;
2704         new->root = RB_ROOT;
2705         INIT_LIST_HEAD(&new->head);
2706
2707         path = btrfs_alloc_path();
2708         if (!path)
2709                 goto out_kfree;
2710
2711         key.objectid = btrfs_ino(inode);
2712         key.type = BTRFS_EXTENT_DATA_KEY;
2713         key.offset = new->file_pos;
2714
2715         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2716         if (ret < 0)
2717                 goto out_free_path;
2718         if (ret > 0 && path->slots[0] > 0)
2719                 path->slots[0]--;
2720
2721         /* find out all the old extents for the file range */
2722         while (1) {
2723                 struct btrfs_file_extent_item *extent;
2724                 struct extent_buffer *l;
2725                 int slot;
2726                 u64 num_bytes;
2727                 u64 offset;
2728                 u64 end;
2729                 u64 disk_bytenr;
2730                 u64 extent_offset;
2731
2732                 l = path->nodes[0];
2733                 slot = path->slots[0];
2734
2735                 if (slot >= btrfs_header_nritems(l)) {
2736                         ret = btrfs_next_leaf(root, path);
2737                         if (ret < 0)
2738                                 goto out_free_path;
2739                         else if (ret > 0)
2740                                 break;
2741                         continue;
2742                 }
2743
2744                 btrfs_item_key_to_cpu(l, &key, slot);
2745
2746                 if (key.objectid != btrfs_ino(inode))
2747                         break;
2748                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2749                         break;
2750                 if (key.offset >= new->file_pos + new->len)
2751                         break;
2752
2753                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2754
2755                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2756                 if (key.offset + num_bytes < new->file_pos)
2757                         goto next;
2758
2759                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2760                 if (!disk_bytenr)
2761                         goto next;
2762
2763                 extent_offset = btrfs_file_extent_offset(l, extent);
2764
2765                 old = kmalloc(sizeof(*old), GFP_NOFS);
2766                 if (!old)
2767                         goto out_free_path;
2768
2769                 offset = max(new->file_pos, key.offset);
2770                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2771
2772                 old->bytenr = disk_bytenr;
2773                 old->extent_offset = extent_offset;
2774                 old->offset = offset - key.offset;
2775                 old->len = end - offset;
2776                 old->new = new;
2777                 old->count = 0;
2778                 list_add_tail(&old->list, &new->head);
2779 next:
2780                 path->slots[0]++;
2781                 cond_resched();
2782         }
2783
2784         btrfs_free_path(path);
2785         atomic_inc(&root->fs_info->defrag_running);
2786
2787         return new;
2788
2789 out_free_path:
2790         btrfs_free_path(path);
2791 out_kfree:
2792         free_sa_defrag_extent(new);
2793         return NULL;
2794 }
2795
2796 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2797                                          u64 start, u64 len)
2798 {
2799         struct btrfs_block_group_cache *cache;
2800
2801         cache = btrfs_lookup_block_group(root->fs_info, start);
2802         ASSERT(cache);
2803
2804         spin_lock(&cache->lock);
2805         cache->delalloc_bytes -= len;
2806         spin_unlock(&cache->lock);
2807
2808         btrfs_put_block_group(cache);
2809 }
2810
2811 /* as ordered data IO finishes, this gets called so we can finish
2812  * an ordered extent if the range of bytes in the file it covers are
2813  * fully written.
2814  */
2815 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2816 {
2817         struct inode *inode = ordered_extent->inode;
2818         struct btrfs_root *root = BTRFS_I(inode)->root;
2819         struct btrfs_trans_handle *trans = NULL;
2820         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2821         struct extent_state *cached_state = NULL;
2822         struct new_sa_defrag_extent *new = NULL;
2823         int compress_type = 0;
2824         int ret = 0;
2825         u64 logical_len = ordered_extent->len;
2826         bool nolock;
2827         bool truncated = false;
2828
2829         nolock = btrfs_is_free_space_inode(inode);
2830
2831         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2832                 ret = -EIO;
2833                 goto out;
2834         }
2835
2836         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2837                                      ordered_extent->file_offset +
2838                                      ordered_extent->len - 1);
2839
2840         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2841                 truncated = true;
2842                 logical_len = ordered_extent->truncated_len;
2843                 /* Truncated the entire extent, don't bother adding */
2844                 if (!logical_len)
2845                         goto out;
2846         }
2847
2848         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2849                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2850
2851                 /*
2852                  * For mwrite(mmap + memset to write) case, we still reserve
2853                  * space for NOCOW range.
2854                  * As NOCOW won't cause a new delayed ref, just free the space
2855                  */
2856                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2857                                        ordered_extent->len);
2858                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2859                 if (nolock)
2860                         trans = btrfs_join_transaction_nolock(root);
2861                 else
2862                         trans = btrfs_join_transaction(root);
2863                 if (IS_ERR(trans)) {
2864                         ret = PTR_ERR(trans);
2865                         trans = NULL;
2866                         goto out;
2867                 }
2868                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2869                 ret = btrfs_update_inode_fallback(trans, root, inode);
2870                 if (ret) /* -ENOMEM or corruption */
2871                         btrfs_abort_transaction(trans, root, ret);
2872                 goto out;
2873         }
2874
2875         lock_extent_bits(io_tree, ordered_extent->file_offset,
2876                          ordered_extent->file_offset + ordered_extent->len - 1,
2877                          0, &cached_state);
2878
2879         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2880                         ordered_extent->file_offset + ordered_extent->len - 1,
2881                         EXTENT_DEFRAG, 1, cached_state);
2882         if (ret) {
2883                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2884                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2885                         /* the inode is shared */
2886                         new = record_old_file_extents(inode, ordered_extent);
2887
2888                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2889                         ordered_extent->file_offset + ordered_extent->len - 1,
2890                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2891         }
2892
2893         if (nolock)
2894                 trans = btrfs_join_transaction_nolock(root);
2895         else
2896                 trans = btrfs_join_transaction(root);
2897         if (IS_ERR(trans)) {
2898                 ret = PTR_ERR(trans);
2899                 trans = NULL;
2900                 goto out_unlock;
2901         }
2902
2903         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2904
2905         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2906                 compress_type = ordered_extent->compress_type;
2907         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2908                 BUG_ON(compress_type);
2909                 ret = btrfs_mark_extent_written(trans, inode,
2910                                                 ordered_extent->file_offset,
2911                                                 ordered_extent->file_offset +
2912                                                 logical_len);
2913         } else {
2914                 BUG_ON(root == root->fs_info->tree_root);
2915                 ret = insert_reserved_file_extent(trans, inode,
2916                                                 ordered_extent->file_offset,
2917                                                 ordered_extent->start,
2918                                                 ordered_extent->disk_len,
2919                                                 logical_len, logical_len,
2920                                                 compress_type, 0, 0,
2921                                                 BTRFS_FILE_EXTENT_REG);
2922                 if (!ret)
2923                         btrfs_release_delalloc_bytes(root,
2924                                                      ordered_extent->start,
2925                                                      ordered_extent->disk_len);
2926         }
2927         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2928                            ordered_extent->file_offset, ordered_extent->len,
2929                            trans->transid);
2930         if (ret < 0) {
2931                 btrfs_abort_transaction(trans, root, ret);
2932                 goto out_unlock;
2933         }
2934
2935         add_pending_csums(trans, inode, ordered_extent->file_offset,
2936                           &ordered_extent->list);
2937
2938         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2939         ret = btrfs_update_inode_fallback(trans, root, inode);
2940         if (ret) { /* -ENOMEM or corruption */
2941                 btrfs_abort_transaction(trans, root, ret);
2942                 goto out_unlock;
2943         }
2944         ret = 0;
2945 out_unlock:
2946         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2947                              ordered_extent->file_offset +
2948                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2949 out:
2950         if (root != root->fs_info->tree_root)
2951                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2952         if (trans)
2953                 btrfs_end_transaction(trans, root);
2954
2955         if (ret || truncated) {
2956                 u64 start, end;
2957
2958                 if (truncated)
2959                         start = ordered_extent->file_offset + logical_len;
2960                 else
2961                         start = ordered_extent->file_offset;
2962                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2963                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2964
2965                 /* Drop the cache for the part of the extent we didn't write. */
2966                 btrfs_drop_extent_cache(inode, start, end, 0);
2967
2968                 /*
2969                  * If the ordered extent had an IOERR or something else went
2970                  * wrong we need to return the space for this ordered extent
2971                  * back to the allocator.  We only free the extent in the
2972                  * truncated case if we didn't write out the extent at all.
2973                  */
2974                 if ((ret || !logical_len) &&
2975                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2976                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2977                         btrfs_free_reserved_extent(root, ordered_extent->start,
2978                                                    ordered_extent->disk_len, 1);
2979         }
2980
2981
2982         /*
2983          * This needs to be done to make sure anybody waiting knows we are done
2984          * updating everything for this ordered extent.
2985          */
2986         btrfs_remove_ordered_extent(inode, ordered_extent);
2987
2988         /* for snapshot-aware defrag */
2989         if (new) {
2990                 if (ret) {
2991                         free_sa_defrag_extent(new);
2992                         atomic_dec(&root->fs_info->defrag_running);
2993                 } else {
2994                         relink_file_extents(new);
2995                 }
2996         }
2997
2998         /* once for us */
2999         btrfs_put_ordered_extent(ordered_extent);
3000         /* once for the tree */
3001         btrfs_put_ordered_extent(ordered_extent);
3002
3003         return ret;
3004 }
3005
3006 static void finish_ordered_fn(struct btrfs_work *work)
3007 {
3008         struct btrfs_ordered_extent *ordered_extent;
3009         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3010         btrfs_finish_ordered_io(ordered_extent);
3011 }
3012
3013 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3014                                 struct extent_state *state, int uptodate)
3015 {
3016         struct inode *inode = page->mapping->host;
3017         struct btrfs_root *root = BTRFS_I(inode)->root;
3018         struct btrfs_ordered_extent *ordered_extent = NULL;
3019         struct btrfs_workqueue *wq;
3020         btrfs_work_func_t func;
3021
3022         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3023
3024         ClearPagePrivate2(page);
3025         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3026                                             end - start + 1, uptodate))
3027                 return 0;
3028
3029         if (btrfs_is_free_space_inode(inode)) {
3030                 wq = root->fs_info->endio_freespace_worker;
3031                 func = btrfs_freespace_write_helper;
3032         } else {
3033                 wq = root->fs_info->endio_write_workers;
3034                 func = btrfs_endio_write_helper;
3035         }
3036
3037         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3038                         NULL);
3039         btrfs_queue_work(wq, &ordered_extent->work);
3040
3041         return 0;
3042 }
3043
3044 static int __readpage_endio_check(struct inode *inode,
3045                                   struct btrfs_io_bio *io_bio,
3046                                   int icsum, struct page *page,
3047                                   int pgoff, u64 start, size_t len)
3048 {
3049         char *kaddr;
3050         u32 csum_expected;
3051         u32 csum = ~(u32)0;
3052
3053         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3054
3055         kaddr = kmap_atomic(page);
3056         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3057         btrfs_csum_final(csum, (char *)&csum);
3058         if (csum != csum_expected)
3059                 goto zeroit;
3060
3061         kunmap_atomic(kaddr);
3062         return 0;
3063 zeroit:
3064         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3065                 "csum failed ino %llu off %llu csum %u expected csum %u",
3066                            btrfs_ino(inode), start, csum, csum_expected);
3067         memset(kaddr + pgoff, 1, len);
3068         flush_dcache_page(page);
3069         kunmap_atomic(kaddr);
3070         if (csum_expected == 0)
3071                 return 0;
3072         return -EIO;
3073 }
3074
3075 /*
3076  * when reads are done, we need to check csums to verify the data is correct
3077  * if there's a match, we allow the bio to finish.  If not, the code in
3078  * extent_io.c will try to find good copies for us.
3079  */
3080 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3081                                       u64 phy_offset, struct page *page,
3082                                       u64 start, u64 end, int mirror)
3083 {
3084         size_t offset = start - page_offset(page);
3085         struct inode *inode = page->mapping->host;
3086         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3087         struct btrfs_root *root = BTRFS_I(inode)->root;
3088
3089         if (PageChecked(page)) {
3090                 ClearPageChecked(page);
3091                 return 0;
3092         }
3093
3094         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3095                 return 0;
3096
3097         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3098             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3099                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
3100                                   GFP_NOFS);
3101                 return 0;
3102         }
3103
3104         phy_offset >>= inode->i_sb->s_blocksize_bits;
3105         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3106                                       start, (size_t)(end - start + 1));
3107 }
3108
3109 struct delayed_iput {
3110         struct list_head list;
3111         struct inode *inode;
3112 };
3113
3114 /* JDM: If this is fs-wide, why can't we add a pointer to
3115  * btrfs_inode instead and avoid the allocation? */
3116 void btrfs_add_delayed_iput(struct inode *inode)
3117 {
3118         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3119         struct delayed_iput *delayed;
3120
3121         if (atomic_add_unless(&inode->i_count, -1, 1))
3122                 return;
3123
3124         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
3125         delayed->inode = inode;
3126
3127         spin_lock(&fs_info->delayed_iput_lock);
3128         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
3129         spin_unlock(&fs_info->delayed_iput_lock);
3130 }
3131
3132 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3133 {
3134         LIST_HEAD(list);
3135         struct btrfs_fs_info *fs_info = root->fs_info;
3136         struct delayed_iput *delayed;
3137         int empty;
3138
3139         spin_lock(&fs_info->delayed_iput_lock);
3140         empty = list_empty(&fs_info->delayed_iputs);
3141         spin_unlock(&fs_info->delayed_iput_lock);
3142         if (empty)
3143                 return;
3144
3145         spin_lock(&fs_info->delayed_iput_lock);
3146         list_splice_init(&fs_info->delayed_iputs, &list);
3147         spin_unlock(&fs_info->delayed_iput_lock);
3148
3149         while (!list_empty(&list)) {
3150                 delayed = list_entry(list.next, struct delayed_iput, list);
3151                 list_del(&delayed->list);
3152                 iput(delayed->inode);
3153                 kfree(delayed);
3154         }
3155 }
3156
3157 /*
3158  * This is called in transaction commit time. If there are no orphan
3159  * files in the subvolume, it removes orphan item and frees block_rsv
3160  * structure.
3161  */
3162 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3163                               struct btrfs_root *root)
3164 {
3165         struct btrfs_block_rsv *block_rsv;
3166         int ret;
3167
3168         if (atomic_read(&root->orphan_inodes) ||
3169             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3170                 return;
3171
3172         spin_lock(&root->orphan_lock);
3173         if (atomic_read(&root->orphan_inodes)) {
3174                 spin_unlock(&root->orphan_lock);
3175                 return;
3176         }
3177
3178         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3179                 spin_unlock(&root->orphan_lock);
3180                 return;
3181         }
3182
3183         block_rsv = root->orphan_block_rsv;
3184         root->orphan_block_rsv = NULL;
3185         spin_unlock(&root->orphan_lock);
3186
3187         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3188             btrfs_root_refs(&root->root_item) > 0) {
3189                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3190                                             root->root_key.objectid);
3191                 if (ret)
3192                         btrfs_abort_transaction(trans, root, ret);
3193                 else
3194                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3195                                   &root->state);
3196         }
3197
3198         if (block_rsv) {
3199                 WARN_ON(block_rsv->size > 0);
3200                 btrfs_free_block_rsv(root, block_rsv);
3201         }
3202 }
3203
3204 /*
3205  * This creates an orphan entry for the given inode in case something goes
3206  * wrong in the middle of an unlink/truncate.
3207  *
3208  * NOTE: caller of this function should reserve 5 units of metadata for
3209  *       this function.
3210  */
3211 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3212 {
3213         struct btrfs_root *root = BTRFS_I(inode)->root;
3214         struct btrfs_block_rsv *block_rsv = NULL;
3215         int reserve = 0;
3216         int insert = 0;
3217         int ret;
3218
3219         if (!root->orphan_block_rsv) {
3220                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3221                 if (!block_rsv)
3222                         return -ENOMEM;
3223         }
3224
3225         spin_lock(&root->orphan_lock);
3226         if (!root->orphan_block_rsv) {
3227                 root->orphan_block_rsv = block_rsv;
3228         } else if (block_rsv) {
3229                 btrfs_free_block_rsv(root, block_rsv);
3230                 block_rsv = NULL;
3231         }
3232
3233         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3234                               &BTRFS_I(inode)->runtime_flags)) {
3235 #if 0
3236                 /*
3237                  * For proper ENOSPC handling, we should do orphan
3238                  * cleanup when mounting. But this introduces backward
3239                  * compatibility issue.
3240                  */
3241                 if (!xchg(&root->orphan_item_inserted, 1))
3242                         insert = 2;
3243                 else
3244                         insert = 1;
3245 #endif
3246                 insert = 1;
3247                 atomic_inc(&root->orphan_inodes);
3248         }
3249
3250         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3251                               &BTRFS_I(inode)->runtime_flags))
3252                 reserve = 1;
3253         spin_unlock(&root->orphan_lock);
3254
3255         /* grab metadata reservation from transaction handle */
3256         if (reserve) {
3257                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3258                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3259         }
3260
3261         /* insert an orphan item to track this unlinked/truncated file */
3262         if (insert >= 1) {
3263                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3264                 if (ret) {
3265                         atomic_dec(&root->orphan_inodes);
3266                         if (reserve) {
3267                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3268                                           &BTRFS_I(inode)->runtime_flags);
3269                                 btrfs_orphan_release_metadata(inode);
3270                         }
3271                         if (ret != -EEXIST) {
3272                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3273                                           &BTRFS_I(inode)->runtime_flags);
3274                                 btrfs_abort_transaction(trans, root, ret);
3275                                 return ret;
3276                         }
3277                 }
3278                 ret = 0;
3279         }
3280
3281         /* insert an orphan item to track subvolume contains orphan files */
3282         if (insert >= 2) {
3283                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3284                                                root->root_key.objectid);
3285                 if (ret && ret != -EEXIST) {
3286                         btrfs_abort_transaction(trans, root, ret);
3287                         return ret;
3288                 }
3289         }
3290         return 0;
3291 }
3292
3293 /*
3294  * We have done the truncate/delete so we can go ahead and remove the orphan
3295  * item for this particular inode.
3296  */
3297 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3298                             struct inode *inode)
3299 {
3300         struct btrfs_root *root = BTRFS_I(inode)->root;
3301         int delete_item = 0;
3302         int release_rsv = 0;
3303         int ret = 0;
3304
3305         spin_lock(&root->orphan_lock);
3306         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3307                                &BTRFS_I(inode)->runtime_flags))
3308                 delete_item = 1;
3309
3310         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3311                                &BTRFS_I(inode)->runtime_flags))
3312                 release_rsv = 1;
3313         spin_unlock(&root->orphan_lock);
3314
3315         if (delete_item) {
3316                 atomic_dec(&root->orphan_inodes);
3317                 if (trans)
3318                         ret = btrfs_del_orphan_item(trans, root,
3319                                                     btrfs_ino(inode));
3320         }
3321
3322         if (release_rsv)
3323                 btrfs_orphan_release_metadata(inode);
3324
3325         return ret;
3326 }
3327
3328 /*
3329  * this cleans up any orphans that may be left on the list from the last use
3330  * of this root.
3331  */
3332 int btrfs_orphan_cleanup(struct btrfs_root *root)
3333 {
3334         struct btrfs_path *path;
3335         struct extent_buffer *leaf;
3336         struct btrfs_key key, found_key;
3337         struct btrfs_trans_handle *trans;
3338         struct inode *inode;
3339         u64 last_objectid = 0;
3340         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3341
3342         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3343                 return 0;
3344
3345         path = btrfs_alloc_path();
3346         if (!path) {
3347                 ret = -ENOMEM;
3348                 goto out;
3349         }
3350         path->reada = -1;
3351
3352         key.objectid = BTRFS_ORPHAN_OBJECTID;
3353         key.type = BTRFS_ORPHAN_ITEM_KEY;
3354         key.offset = (u64)-1;
3355
3356         while (1) {
3357                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3358                 if (ret < 0)
3359                         goto out;
3360
3361                 /*
3362                  * if ret == 0 means we found what we were searching for, which
3363                  * is weird, but possible, so only screw with path if we didn't
3364                  * find the key and see if we have stuff that matches
3365                  */
3366                 if (ret > 0) {
3367                         ret = 0;
3368                         if (path->slots[0] == 0)
3369                                 break;
3370                         path->slots[0]--;
3371                 }
3372
3373                 /* pull out the item */
3374                 leaf = path->nodes[0];
3375                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3376
3377                 /* make sure the item matches what we want */
3378                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3379                         break;
3380                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3381                         break;
3382
3383                 /* release the path since we're done with it */
3384                 btrfs_release_path(path);
3385
3386                 /*
3387                  * this is where we are basically btrfs_lookup, without the
3388                  * crossing root thing.  we store the inode number in the
3389                  * offset of the orphan item.
3390                  */
3391
3392                 if (found_key.offset == last_objectid) {
3393                         btrfs_err(root->fs_info,
3394                                 "Error removing orphan entry, stopping orphan cleanup");
3395                         ret = -EINVAL;
3396                         goto out;
3397                 }
3398
3399                 last_objectid = found_key.offset;
3400
3401                 found_key.objectid = found_key.offset;
3402                 found_key.type = BTRFS_INODE_ITEM_KEY;
3403                 found_key.offset = 0;
3404                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3405                 ret = PTR_ERR_OR_ZERO(inode);
3406                 if (ret && ret != -ESTALE)
3407                         goto out;
3408
3409                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3410                         struct btrfs_root *dead_root;
3411                         struct btrfs_fs_info *fs_info = root->fs_info;
3412                         int is_dead_root = 0;
3413
3414                         /*
3415                          * this is an orphan in the tree root. Currently these
3416                          * could come from 2 sources:
3417                          *  a) a snapshot deletion in progress
3418                          *  b) a free space cache inode
3419                          * We need to distinguish those two, as the snapshot
3420                          * orphan must not get deleted.
3421                          * find_dead_roots already ran before us, so if this
3422                          * is a snapshot deletion, we should find the root
3423                          * in the dead_roots list
3424                          */
3425                         spin_lock(&fs_info->trans_lock);
3426                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3427                                             root_list) {
3428                                 if (dead_root->root_key.objectid ==
3429                                     found_key.objectid) {
3430                                         is_dead_root = 1;
3431                                         break;
3432                                 }
3433                         }
3434                         spin_unlock(&fs_info->trans_lock);
3435                         if (is_dead_root) {
3436                                 /* prevent this orphan from being found again */
3437                                 key.offset = found_key.objectid - 1;
3438                                 continue;
3439                         }
3440                 }
3441                 /*
3442                  * Inode is already gone but the orphan item is still there,
3443                  * kill the orphan item.
3444                  */
3445                 if (ret == -ESTALE) {
3446                         trans = btrfs_start_transaction(root, 1);
3447                         if (IS_ERR(trans)) {
3448                                 ret = PTR_ERR(trans);
3449                                 goto out;
3450                         }
3451                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3452                                 found_key.objectid);
3453                         ret = btrfs_del_orphan_item(trans, root,
3454                                                     found_key.objectid);
3455                         btrfs_end_transaction(trans, root);
3456                         if (ret)
3457                                 goto out;
3458                         continue;
3459                 }
3460
3461                 /*
3462                  * add this inode to the orphan list so btrfs_orphan_del does
3463                  * the proper thing when we hit it
3464                  */
3465                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3466                         &BTRFS_I(inode)->runtime_flags);
3467                 atomic_inc(&root->orphan_inodes);
3468
3469                 /* if we have links, this was a truncate, lets do that */
3470                 if (inode->i_nlink) {
3471                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3472                                 iput(inode);
3473                                 continue;
3474                         }
3475                         nr_truncate++;
3476
3477                         /* 1 for the orphan item deletion. */
3478                         trans = btrfs_start_transaction(root, 1);
3479                         if (IS_ERR(trans)) {
3480                                 iput(inode);
3481                                 ret = PTR_ERR(trans);
3482                                 goto out;
3483                         }
3484                         ret = btrfs_orphan_add(trans, inode);
3485                         btrfs_end_transaction(trans, root);
3486                         if (ret) {
3487                                 iput(inode);
3488                                 goto out;
3489                         }
3490
3491                         ret = btrfs_truncate(inode);
3492                         if (ret)
3493                                 btrfs_orphan_del(NULL, inode);
3494                 } else {
3495                         nr_unlink++;
3496                 }
3497
3498                 /* this will do delete_inode and everything for us */
3499                 iput(inode);
3500                 if (ret)
3501                         goto out;
3502         }
3503         /* release the path since we're done with it */
3504         btrfs_release_path(path);
3505
3506         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3507
3508         if (root->orphan_block_rsv)
3509                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3510                                         (u64)-1);
3511
3512         if (root->orphan_block_rsv ||
3513             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3514                 trans = btrfs_join_transaction(root);
3515                 if (!IS_ERR(trans))
3516                         btrfs_end_transaction(trans, root);
3517         }
3518
3519         if (nr_unlink)
3520                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3521         if (nr_truncate)
3522                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3523
3524 out:
3525         if (ret)
3526                 btrfs_err(root->fs_info,
3527                         "could not do orphan cleanup %d", ret);
3528         btrfs_free_path(path);
3529         return ret;
3530 }
3531
3532 /*
3533  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3534  * don't find any xattrs, we know there can't be any acls.
3535  *
3536  * slot is the slot the inode is in, objectid is the objectid of the inode
3537  */
3538 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3539                                           int slot, u64 objectid,
3540                                           int *first_xattr_slot)
3541 {
3542         u32 nritems = btrfs_header_nritems(leaf);
3543         struct btrfs_key found_key;
3544         static u64 xattr_access = 0;
3545         static u64 xattr_default = 0;
3546         int scanned = 0;
3547
3548         if (!xattr_access) {
3549                 xattr_access = btrfs_name_hash(POSIX_ACL_XATTR_ACCESS,
3550                                         strlen(POSIX_ACL_XATTR_ACCESS));
3551                 xattr_default = btrfs_name_hash(POSIX_ACL_XATTR_DEFAULT,
3552                                         strlen(POSIX_ACL_XATTR_DEFAULT));
3553         }
3554
3555         slot++;
3556         *first_xattr_slot = -1;
3557         while (slot < nritems) {
3558                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3559
3560                 /* we found a different objectid, there must not be acls */
3561                 if (found_key.objectid != objectid)
3562                         return 0;
3563
3564                 /* we found an xattr, assume we've got an acl */
3565                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3566                         if (*first_xattr_slot == -1)
3567                                 *first_xattr_slot = slot;
3568                         if (found_key.offset == xattr_access ||
3569                             found_key.offset == xattr_default)
3570                                 return 1;
3571                 }
3572
3573                 /*
3574                  * we found a key greater than an xattr key, there can't
3575                  * be any acls later on
3576                  */
3577                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3578                         return 0;
3579
3580                 slot++;
3581                 scanned++;
3582
3583                 /*
3584                  * it goes inode, inode backrefs, xattrs, extents,
3585                  * so if there are a ton of hard links to an inode there can
3586                  * be a lot of backrefs.  Don't waste time searching too hard,
3587                  * this is just an optimization
3588                  */
3589                 if (scanned >= 8)
3590                         break;
3591         }
3592         /* we hit the end of the leaf before we found an xattr or
3593          * something larger than an xattr.  We have to assume the inode
3594          * has acls
3595          */
3596         if (*first_xattr_slot == -1)
3597                 *first_xattr_slot = slot;
3598         return 1;
3599 }
3600
3601 /*
3602  * read an inode from the btree into the in-memory inode
3603  */
3604 static void btrfs_read_locked_inode(struct inode *inode)
3605 {
3606         struct btrfs_path *path;
3607         struct extent_buffer *leaf;
3608         struct btrfs_inode_item *inode_item;
3609         struct btrfs_root *root = BTRFS_I(inode)->root;
3610         struct btrfs_key location;
3611         unsigned long ptr;
3612         int maybe_acls;
3613         u32 rdev;
3614         int ret;
3615         bool filled = false;
3616         int first_xattr_slot;
3617
3618         ret = btrfs_fill_inode(inode, &rdev);
3619         if (!ret)
3620                 filled = true;
3621
3622         path = btrfs_alloc_path();
3623         if (!path)
3624                 goto make_bad;
3625
3626         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3627
3628         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3629         if (ret)
3630                 goto make_bad;
3631
3632         leaf = path->nodes[0];
3633
3634         if (filled)
3635                 goto cache_index;
3636
3637         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3638                                     struct btrfs_inode_item);
3639         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3640         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3641         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3642         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3643         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3644
3645         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3646         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3647
3648         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3649         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3650
3651         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3652         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3653
3654         BTRFS_I(inode)->i_otime.tv_sec =
3655                 btrfs_timespec_sec(leaf, &inode_item->otime);
3656         BTRFS_I(inode)->i_otime.tv_nsec =
3657                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3658
3659         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3660         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3661         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3662
3663         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3664         inode->i_generation = BTRFS_I(inode)->generation;
3665         inode->i_rdev = 0;
3666         rdev = btrfs_inode_rdev(leaf, inode_item);
3667
3668         BTRFS_I(inode)->index_cnt = (u64)-1;
3669         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3670
3671 cache_index:
3672         /*
3673          * If we were modified in the current generation and evicted from memory
3674          * and then re-read we need to do a full sync since we don't have any
3675          * idea about which extents were modified before we were evicted from
3676          * cache.
3677          *
3678          * This is required for both inode re-read from disk and delayed inode
3679          * in delayed_nodes_tree.
3680          */
3681         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3682                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3683                         &BTRFS_I(inode)->runtime_flags);
3684
3685         /*
3686          * We don't persist the id of the transaction where an unlink operation
3687          * against the inode was last made. So here we assume the inode might
3688          * have been evicted, and therefore the exact value of last_unlink_trans
3689          * lost, and set it to last_trans to avoid metadata inconsistencies
3690          * between the inode and its parent if the inode is fsync'ed and the log
3691          * replayed. For example, in the scenario:
3692          *
3693          * touch mydir/foo
3694          * ln mydir/foo mydir/bar
3695          * sync
3696          * unlink mydir/bar
3697          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3698          * xfs_io -c fsync mydir/foo
3699          * <power failure>
3700          * mount fs, triggers fsync log replay
3701          *
3702          * We must make sure that when we fsync our inode foo we also log its
3703          * parent inode, otherwise after log replay the parent still has the
3704          * dentry with the "bar" name but our inode foo has a link count of 1
3705          * and doesn't have an inode ref with the name "bar" anymore.
3706          *
3707          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3708          * but it guarantees correctness at the expense of ocassional full
3709          * transaction commits on fsync if our inode is a directory, or if our
3710          * inode is not a directory, logging its parent unnecessarily.
3711          */
3712         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3713
3714         path->slots[0]++;
3715         if (inode->i_nlink != 1 ||
3716             path->slots[0] >= btrfs_header_nritems(leaf))
3717                 goto cache_acl;
3718
3719         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3720         if (location.objectid != btrfs_ino(inode))
3721                 goto cache_acl;
3722
3723         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3724         if (location.type == BTRFS_INODE_REF_KEY) {
3725                 struct btrfs_inode_ref *ref;
3726
3727                 ref = (struct btrfs_inode_ref *)ptr;
3728                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3729         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3730                 struct btrfs_inode_extref *extref;
3731
3732                 extref = (struct btrfs_inode_extref *)ptr;
3733                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3734                                                                      extref);
3735         }
3736 cache_acl:
3737         /*
3738          * try to precache a NULL acl entry for files that don't have
3739          * any xattrs or acls
3740          */
3741         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3742                                            btrfs_ino(inode), &first_xattr_slot);
3743         if (first_xattr_slot != -1) {
3744                 path->slots[0] = first_xattr_slot;
3745                 ret = btrfs_load_inode_props(inode, path);
3746                 if (ret)
3747                         btrfs_err(root->fs_info,
3748                                   "error loading props for ino %llu (root %llu): %d",
3749                                   btrfs_ino(inode),
3750                                   root->root_key.objectid, ret);
3751         }
3752         btrfs_free_path(path);
3753
3754         if (!maybe_acls)
3755                 cache_no_acl(inode);
3756
3757         switch (inode->i_mode & S_IFMT) {
3758         case S_IFREG:
3759                 inode->i_mapping->a_ops = &btrfs_aops;
3760                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3761                 inode->i_fop = &btrfs_file_operations;
3762                 inode->i_op = &btrfs_file_inode_operations;
3763                 break;
3764         case S_IFDIR:
3765                 inode->i_fop = &btrfs_dir_file_operations;
3766                 if (root == root->fs_info->tree_root)
3767                         inode->i_op = &btrfs_dir_ro_inode_operations;
3768                 else
3769                         inode->i_op = &btrfs_dir_inode_operations;
3770                 break;
3771         case S_IFLNK:
3772                 inode->i_op = &btrfs_symlink_inode_operations;
3773                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3774                 break;
3775         default:
3776                 inode->i_op = &btrfs_special_inode_operations;
3777                 init_special_inode(inode, inode->i_mode, rdev);
3778                 break;
3779         }
3780
3781         btrfs_update_iflags(inode);
3782         return;
3783
3784 make_bad:
3785         btrfs_free_path(path);
3786         make_bad_inode(inode);
3787 }
3788
3789 /*
3790  * given a leaf and an inode, copy the inode fields into the leaf
3791  */
3792 static void fill_inode_item(struct btrfs_trans_handle *trans,
3793                             struct extent_buffer *leaf,
3794                             struct btrfs_inode_item *item,
3795                             struct inode *inode)
3796 {
3797         struct btrfs_map_token token;
3798
3799         btrfs_init_map_token(&token);
3800
3801         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3802         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3803         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3804                                    &token);
3805         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3806         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3807
3808         btrfs_set_token_timespec_sec(leaf, &item->atime,
3809                                      inode->i_atime.tv_sec, &token);
3810         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3811                                       inode->i_atime.tv_nsec, &token);
3812
3813         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3814                                      inode->i_mtime.tv_sec, &token);
3815         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3816                                       inode->i_mtime.tv_nsec, &token);
3817
3818         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3819                                      inode->i_ctime.tv_sec, &token);
3820         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3821                                       inode->i_ctime.tv_nsec, &token);
3822
3823         btrfs_set_token_timespec_sec(leaf, &item->otime,
3824                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3825         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3826                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3827
3828         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3829                                      &token);
3830         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3831                                          &token);
3832         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3833         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3834         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3835         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3836         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3837 }
3838
3839 /*
3840  * copy everything in the in-memory inode into the btree.
3841  */
3842 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3843                                 struct btrfs_root *root, struct inode *inode)
3844 {
3845         struct btrfs_inode_item *inode_item;
3846         struct btrfs_path *path;
3847         struct extent_buffer *leaf;
3848         int ret;
3849
3850         path = btrfs_alloc_path();
3851         if (!path)
3852                 return -ENOMEM;
3853
3854         path->leave_spinning = 1;
3855         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3856                                  1);
3857         if (ret) {
3858                 if (ret > 0)
3859                         ret = -ENOENT;
3860                 goto failed;
3861         }
3862
3863         leaf = path->nodes[0];
3864         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3865                                     struct btrfs_inode_item);
3866
3867         fill_inode_item(trans, leaf, inode_item, inode);
3868         btrfs_mark_buffer_dirty(leaf);
3869         btrfs_set_inode_last_trans(trans, inode);
3870         ret = 0;
3871 failed:
3872         btrfs_free_path(path);
3873         return ret;
3874 }
3875
3876 /*
3877  * copy everything in the in-memory inode into the btree.
3878  */
3879 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3880                                 struct btrfs_root *root, struct inode *inode)
3881 {
3882         int ret;
3883
3884         /*
3885          * If the inode is a free space inode, we can deadlock during commit
3886          * if we put it into the delayed code.
3887          *
3888          * The data relocation inode should also be directly updated
3889          * without delay
3890          */
3891         if (!btrfs_is_free_space_inode(inode)
3892             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3893             && !root->fs_info->log_root_recovering) {
3894                 btrfs_update_root_times(trans, root);
3895
3896                 ret = btrfs_delayed_update_inode(trans, root, inode);
3897                 if (!ret)
3898                         btrfs_set_inode_last_trans(trans, inode);
3899                 return ret;
3900         }
3901
3902         return btrfs_update_inode_item(trans, root, inode);
3903 }
3904
3905 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3906                                          struct btrfs_root *root,
3907                                          struct inode *inode)
3908 {
3909         int ret;
3910
3911         ret = btrfs_update_inode(trans, root, inode);
3912         if (ret == -ENOSPC)
3913                 return btrfs_update_inode_item(trans, root, inode);
3914         return ret;
3915 }
3916
3917 /*
3918  * unlink helper that gets used here in inode.c and in the tree logging
3919  * recovery code.  It remove a link in a directory with a given name, and
3920  * also drops the back refs in the inode to the directory
3921  */
3922 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3923                                 struct btrfs_root *root,
3924                                 struct inode *dir, struct inode *inode,
3925                                 const char *name, int name_len)
3926 {
3927         struct btrfs_path *path;
3928         int ret = 0;
3929         struct extent_buffer *leaf;
3930         struct btrfs_dir_item *di;
3931         struct btrfs_key key;
3932         u64 index;
3933         u64 ino = btrfs_ino(inode);
3934         u64 dir_ino = btrfs_ino(dir);
3935
3936         path = btrfs_alloc_path();
3937         if (!path) {
3938                 ret = -ENOMEM;
3939                 goto out;
3940         }
3941
3942         path->leave_spinning = 1;
3943         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3944                                     name, name_len, -1);
3945         if (IS_ERR(di)) {
3946                 ret = PTR_ERR(di);
3947                 goto err;
3948         }
3949         if (!di) {
3950                 ret = -ENOENT;
3951                 goto err;
3952         }
3953         leaf = path->nodes[0];
3954         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3955         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3956         if (ret)
3957                 goto err;
3958         btrfs_release_path(path);
3959
3960         /*
3961          * If we don't have dir index, we have to get it by looking up
3962          * the inode ref, since we get the inode ref, remove it directly,
3963          * it is unnecessary to do delayed deletion.
3964          *
3965          * But if we have dir index, needn't search inode ref to get it.
3966          * Since the inode ref is close to the inode item, it is better
3967          * that we delay to delete it, and just do this deletion when
3968          * we update the inode item.
3969          */
3970         if (BTRFS_I(inode)->dir_index) {
3971                 ret = btrfs_delayed_delete_inode_ref(inode);
3972                 if (!ret) {
3973                         index = BTRFS_I(inode)->dir_index;
3974                         goto skip_backref;
3975                 }
3976         }
3977
3978         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3979                                   dir_ino, &index);
3980         if (ret) {
3981                 btrfs_info(root->fs_info,
3982                         "failed to delete reference to %.*s, inode %llu parent %llu",
3983                         name_len, name, ino, dir_ino);
3984                 btrfs_abort_transaction(trans, root, ret);
3985                 goto err;
3986         }
3987 skip_backref:
3988         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3989         if (ret) {
3990                 btrfs_abort_transaction(trans, root, ret);
3991                 goto err;
3992         }
3993
3994         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
3995                                          inode, dir_ino);
3996         if (ret != 0 && ret != -ENOENT) {
3997                 btrfs_abort_transaction(trans, root, ret);
3998                 goto err;
3999         }
4000
4001         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4002                                            dir, index);
4003         if (ret == -ENOENT)
4004                 ret = 0;
4005         else if (ret)
4006                 btrfs_abort_transaction(trans, root, ret);
4007 err:
4008         btrfs_free_path(path);
4009         if (ret)
4010                 goto out;
4011
4012         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4013         inode_inc_iversion(inode);
4014         inode_inc_iversion(dir);
4015         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4016         ret = btrfs_update_inode(trans, root, dir);
4017 out:
4018         return ret;
4019 }
4020
4021 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4022                        struct btrfs_root *root,
4023                        struct inode *dir, struct inode *inode,
4024                        const char *name, int name_len)
4025 {
4026         int ret;
4027         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4028         if (!ret) {
4029                 drop_nlink(inode);
4030                 ret = btrfs_update_inode(trans, root, inode);
4031         }
4032         return ret;
4033 }
4034
4035 /*
4036  * helper to start transaction for unlink and rmdir.
4037  *
4038  * unlink and rmdir are special in btrfs, they do not always free space, so
4039  * if we cannot make our reservations the normal way try and see if there is
4040  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4041  * allow the unlink to occur.
4042  */
4043 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4044 {
4045         struct btrfs_root *root = BTRFS_I(dir)->root;
4046
4047         /*
4048          * 1 for the possible orphan item
4049          * 1 for the dir item
4050          * 1 for the dir index
4051          * 1 for the inode ref
4052          * 1 for the inode
4053          */
4054         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4055 }
4056
4057 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4058 {
4059         struct btrfs_root *root = BTRFS_I(dir)->root;
4060         struct btrfs_trans_handle *trans;
4061         struct inode *inode = d_inode(dentry);
4062         int ret;
4063
4064         trans = __unlink_start_trans(dir);
4065         if (IS_ERR(trans))
4066                 return PTR_ERR(trans);
4067
4068         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4069
4070         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4071                                  dentry->d_name.name, dentry->d_name.len);
4072         if (ret)
4073                 goto out;
4074
4075         if (inode->i_nlink == 0) {
4076                 ret = btrfs_orphan_add(trans, inode);
4077                 if (ret)
4078                         goto out;
4079         }
4080
4081 out:
4082         btrfs_end_transaction(trans, root);
4083         btrfs_btree_balance_dirty(root);
4084         return ret;
4085 }
4086
4087 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4088                         struct btrfs_root *root,
4089                         struct inode *dir, u64 objectid,
4090                         const char *name, int name_len)
4091 {
4092         struct btrfs_path *path;
4093         struct extent_buffer *leaf;
4094         struct btrfs_dir_item *di;
4095         struct btrfs_key key;
4096         u64 index;
4097         int ret;
4098         u64 dir_ino = btrfs_ino(dir);
4099
4100         path = btrfs_alloc_path();
4101         if (!path)
4102                 return -ENOMEM;
4103
4104         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4105                                    name, name_len, -1);
4106         if (IS_ERR_OR_NULL(di)) {
4107                 if (!di)
4108                         ret = -ENOENT;
4109                 else
4110                         ret = PTR_ERR(di);
4111                 goto out;
4112         }
4113
4114         leaf = path->nodes[0];
4115         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4116         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4117         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4118         if (ret) {
4119                 btrfs_abort_transaction(trans, root, ret);
4120                 goto out;
4121         }
4122         btrfs_release_path(path);
4123
4124         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4125                                  objectid, root->root_key.objectid,
4126                                  dir_ino, &index, name, name_len);
4127         if (ret < 0) {
4128                 if (ret != -ENOENT) {
4129                         btrfs_abort_transaction(trans, root, ret);
4130                         goto out;
4131                 }
4132                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4133                                                  name, name_len);
4134                 if (IS_ERR_OR_NULL(di)) {
4135                         if (!di)
4136                                 ret = -ENOENT;
4137                         else
4138                                 ret = PTR_ERR(di);
4139                         btrfs_abort_transaction(trans, root, ret);
4140                         goto out;
4141                 }
4142
4143                 leaf = path->nodes[0];
4144                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4145                 btrfs_release_path(path);
4146                 index = key.offset;
4147         }
4148         btrfs_release_path(path);
4149
4150         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4151         if (ret) {
4152                 btrfs_abort_transaction(trans, root, ret);
4153                 goto out;
4154         }
4155
4156         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4157         inode_inc_iversion(dir);
4158         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4159         ret = btrfs_update_inode_fallback(trans, root, dir);
4160         if (ret)
4161                 btrfs_abort_transaction(trans, root, ret);
4162 out:
4163         btrfs_free_path(path);
4164         return ret;
4165 }
4166
4167 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4168 {
4169         struct inode *inode = d_inode(dentry);
4170         int err = 0;
4171         struct btrfs_root *root = BTRFS_I(dir)->root;
4172         struct btrfs_trans_handle *trans;
4173
4174         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4175                 return -ENOTEMPTY;
4176         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4177                 return -EPERM;
4178
4179         trans = __unlink_start_trans(dir);
4180         if (IS_ERR(trans))
4181                 return PTR_ERR(trans);
4182
4183         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4184                 err = btrfs_unlink_subvol(trans, root, dir,
4185                                           BTRFS_I(inode)->location.objectid,
4186                                           dentry->d_name.name,
4187                                           dentry->d_name.len);
4188                 goto out;
4189         }
4190
4191         err = btrfs_orphan_add(trans, inode);
4192         if (err)
4193                 goto out;
4194
4195         /* now the directory is empty */
4196         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4197                                  dentry->d_name.name, dentry->d_name.len);
4198         if (!err)
4199                 btrfs_i_size_write(inode, 0);
4200 out:
4201         btrfs_end_transaction(trans, root);
4202         btrfs_btree_balance_dirty(root);
4203
4204         return err;
4205 }
4206
4207 static int truncate_space_check(struct btrfs_trans_handle *trans,
4208                                 struct btrfs_root *root,
4209                                 u64 bytes_deleted)
4210 {
4211         int ret;
4212
4213         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4214         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4215                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4216         if (!ret)
4217                 trans->bytes_reserved += bytes_deleted;
4218         return ret;
4219
4220 }
4221
4222 static int truncate_inline_extent(struct inode *inode,
4223                                   struct btrfs_path *path,
4224                                   struct btrfs_key *found_key,
4225                                   const u64 item_end,
4226                                   const u64 new_size)
4227 {
4228         struct extent_buffer *leaf = path->nodes[0];
4229         int slot = path->slots[0];
4230         struct btrfs_file_extent_item *fi;
4231         u32 size = (u32)(new_size - found_key->offset);
4232         struct btrfs_root *root = BTRFS_I(inode)->root;
4233
4234         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4235
4236         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4237                 loff_t offset = new_size;
4238                 loff_t page_end = ALIGN(offset, PAGE_CACHE_SIZE);
4239
4240                 /*
4241                  * Zero out the remaining of the last page of our inline extent,
4242                  * instead of directly truncating our inline extent here - that
4243                  * would be much more complex (decompressing all the data, then
4244                  * compressing the truncated data, which might be bigger than
4245                  * the size of the inline extent, resize the extent, etc).
4246                  * We release the path because to get the page we might need to
4247                  * read the extent item from disk (data not in the page cache).
4248                  */
4249                 btrfs_release_path(path);
4250                 return btrfs_truncate_page(inode, offset, page_end - offset, 0);
4251         }
4252
4253         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4254         size = btrfs_file_extent_calc_inline_size(size);
4255         btrfs_truncate_item(root, path, size, 1);
4256
4257         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4258                 inode_sub_bytes(inode, item_end + 1 - new_size);
4259
4260         return 0;
4261 }
4262
4263 /*
4264  * this can truncate away extent items, csum items and directory items.
4265  * It starts at a high offset and removes keys until it can't find
4266  * any higher than new_size
4267  *
4268  * csum items that cross the new i_size are truncated to the new size
4269  * as well.
4270  *
4271  * min_type is the minimum key type to truncate down to.  If set to 0, this
4272  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4273  */
4274 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4275                                struct btrfs_root *root,
4276                                struct inode *inode,
4277                                u64 new_size, u32 min_type)
4278 {
4279         struct btrfs_path *path;
4280         struct extent_buffer *leaf;
4281         struct btrfs_file_extent_item *fi;
4282         struct btrfs_key key;
4283         struct btrfs_key found_key;
4284         u64 extent_start = 0;
4285         u64 extent_num_bytes = 0;
4286         u64 extent_offset = 0;
4287         u64 item_end = 0;
4288         u64 last_size = new_size;
4289         u32 found_type = (u8)-1;
4290         int found_extent;
4291         int del_item;
4292         int pending_del_nr = 0;
4293         int pending_del_slot = 0;
4294         int extent_type = -1;
4295         int ret;
4296         int err = 0;
4297         u64 ino = btrfs_ino(inode);
4298         u64 bytes_deleted = 0;
4299         bool be_nice = 0;
4300         bool should_throttle = 0;
4301         bool should_end = 0;
4302
4303         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4304
4305         /*
4306          * for non-free space inodes and ref cows, we want to back off from
4307          * time to time
4308          */
4309         if (!btrfs_is_free_space_inode(inode) &&
4310             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4311                 be_nice = 1;
4312
4313         path = btrfs_alloc_path();
4314         if (!path)
4315                 return -ENOMEM;
4316         path->reada = -1;
4317
4318         /*
4319          * We want to drop from the next block forward in case this new size is
4320          * not block aligned since we will be keeping the last block of the
4321          * extent just the way it is.
4322          */
4323         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4324             root == root->fs_info->tree_root)
4325                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4326                                         root->sectorsize), (u64)-1, 0);
4327
4328         /*
4329          * This function is also used to drop the items in the log tree before
4330          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4331          * it is used to drop the loged items. So we shouldn't kill the delayed
4332          * items.
4333          */
4334         if (min_type == 0 && root == BTRFS_I(inode)->root)
4335                 btrfs_kill_delayed_inode_items(inode);
4336
4337         key.objectid = ino;
4338         key.offset = (u64)-1;
4339         key.type = (u8)-1;
4340
4341 search_again:
4342         /*
4343          * with a 16K leaf size and 128MB extents, you can actually queue
4344          * up a huge file in a single leaf.  Most of the time that
4345          * bytes_deleted is > 0, it will be huge by the time we get here
4346          */
4347         if (be_nice && bytes_deleted > 32 * 1024 * 1024) {
4348                 if (btrfs_should_end_transaction(trans, root)) {
4349                         err = -EAGAIN;
4350                         goto error;
4351                 }
4352         }
4353
4354
4355         path->leave_spinning = 1;
4356         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4357         if (ret < 0) {
4358                 err = ret;
4359                 goto out;
4360         }
4361
4362         if (ret > 0) {
4363                 /* there are no items in the tree for us to truncate, we're
4364                  * done
4365                  */
4366                 if (path->slots[0] == 0)
4367                         goto out;
4368                 path->slots[0]--;
4369         }
4370
4371         while (1) {
4372                 fi = NULL;
4373                 leaf = path->nodes[0];
4374                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4375                 found_type = found_key.type;
4376
4377                 if (found_key.objectid != ino)
4378                         break;
4379
4380                 if (found_type < min_type)
4381                         break;
4382
4383                 item_end = found_key.offset;
4384                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4385                         fi = btrfs_item_ptr(leaf, path->slots[0],
4386                                             struct btrfs_file_extent_item);
4387                         extent_type = btrfs_file_extent_type(leaf, fi);
4388                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4389                                 item_end +=
4390                                     btrfs_file_extent_num_bytes(leaf, fi);
4391                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4392                                 item_end += btrfs_file_extent_inline_len(leaf,
4393                                                          path->slots[0], fi);
4394                         }
4395                         item_end--;
4396                 }
4397                 if (found_type > min_type) {
4398                         del_item = 1;
4399                 } else {
4400                         if (item_end < new_size)
4401                                 break;
4402                         if (found_key.offset >= new_size)
4403                                 del_item = 1;
4404                         else
4405                                 del_item = 0;
4406                 }
4407                 found_extent = 0;
4408                 /* FIXME, shrink the extent if the ref count is only 1 */
4409                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4410                         goto delete;
4411
4412                 if (del_item)
4413                         last_size = found_key.offset;
4414                 else
4415                         last_size = new_size;
4416
4417                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4418                         u64 num_dec;
4419                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4420                         if (!del_item) {
4421                                 u64 orig_num_bytes =
4422                                         btrfs_file_extent_num_bytes(leaf, fi);
4423                                 extent_num_bytes = ALIGN(new_size -
4424                                                 found_key.offset,
4425                                                 root->sectorsize);
4426                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4427                                                          extent_num_bytes);
4428                                 num_dec = (orig_num_bytes -
4429                                            extent_num_bytes);
4430                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4431                                              &root->state) &&
4432                                     extent_start != 0)
4433                                         inode_sub_bytes(inode, num_dec);
4434                                 btrfs_mark_buffer_dirty(leaf);
4435                         } else {
4436                                 extent_num_bytes =
4437                                         btrfs_file_extent_disk_num_bytes(leaf,
4438                                                                          fi);
4439                                 extent_offset = found_key.offset -
4440                                         btrfs_file_extent_offset(leaf, fi);
4441
4442                                 /* FIXME blocksize != 4096 */
4443                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4444                                 if (extent_start != 0) {
4445                                         found_extent = 1;
4446                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4447                                                      &root->state))
4448                                                 inode_sub_bytes(inode, num_dec);
4449                                 }
4450                         }
4451                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4452                         /*
4453                          * we can't truncate inline items that have had
4454                          * special encodings
4455                          */
4456                         if (!del_item &&
4457                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4458                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4459
4460                                 /*
4461                                  * Need to release path in order to truncate a
4462                                  * compressed extent. So delete any accumulated
4463                                  * extent items so far.
4464                                  */
4465                                 if (btrfs_file_extent_compression(leaf, fi) !=
4466                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4467                                         err = btrfs_del_items(trans, root, path,
4468                                                               pending_del_slot,
4469                                                               pending_del_nr);
4470                                         if (err) {
4471                                                 btrfs_abort_transaction(trans,
4472                                                                         root,
4473                                                                         err);
4474                                                 goto error;
4475                                         }
4476                                         pending_del_nr = 0;
4477                                 }
4478
4479                                 err = truncate_inline_extent(inode, path,
4480                                                              &found_key,
4481                                                              item_end,
4482                                                              new_size);
4483                                 if (err) {
4484                                         btrfs_abort_transaction(trans,
4485                                                                 root, err);
4486                                         goto error;
4487                                 }
4488                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4489                                             &root->state)) {
4490                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4491                         }
4492                 }
4493 delete:
4494                 if (del_item) {
4495                         if (!pending_del_nr) {
4496                                 /* no pending yet, add ourselves */
4497                                 pending_del_slot = path->slots[0];
4498                                 pending_del_nr = 1;
4499                         } else if (pending_del_nr &&
4500                                    path->slots[0] + 1 == pending_del_slot) {
4501                                 /* hop on the pending chunk */
4502                                 pending_del_nr++;
4503                                 pending_del_slot = path->slots[0];
4504                         } else {
4505                                 BUG();
4506                         }
4507                 } else {
4508                         break;
4509                 }
4510                 should_throttle = 0;
4511
4512                 if (found_extent &&
4513                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4514                      root == root->fs_info->tree_root)) {
4515                         btrfs_set_path_blocking(path);
4516                         bytes_deleted += extent_num_bytes;
4517                         ret = btrfs_free_extent(trans, root, extent_start,
4518                                                 extent_num_bytes, 0,
4519                                                 btrfs_header_owner(leaf),
4520                                                 ino, extent_offset);
4521                         BUG_ON(ret);
4522                         if (btrfs_should_throttle_delayed_refs(trans, root))
4523                                 btrfs_async_run_delayed_refs(root,
4524                                         trans->delayed_ref_updates * 2, 0);
4525                         if (be_nice) {
4526                                 if (truncate_space_check(trans, root,
4527                                                          extent_num_bytes)) {
4528                                         should_end = 1;
4529                                 }
4530                                 if (btrfs_should_throttle_delayed_refs(trans,
4531                                                                        root)) {
4532                                         should_throttle = 1;
4533                                 }
4534                         }
4535                 }
4536
4537                 if (found_type == BTRFS_INODE_ITEM_KEY)
4538                         break;
4539
4540                 if (path->slots[0] == 0 ||
4541                     path->slots[0] != pending_del_slot ||
4542                     should_throttle || should_end) {
4543                         if (pending_del_nr) {
4544                                 ret = btrfs_del_items(trans, root, path,
4545                                                 pending_del_slot,
4546                                                 pending_del_nr);
4547                                 if (ret) {
4548                                         btrfs_abort_transaction(trans,
4549                                                                 root, ret);
4550                                         goto error;
4551                                 }
4552                                 pending_del_nr = 0;
4553                         }
4554                         btrfs_release_path(path);
4555                         if (should_throttle) {
4556                                 unsigned long updates = trans->delayed_ref_updates;
4557                                 if (updates) {
4558                                         trans->delayed_ref_updates = 0;
4559                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4560                                         if (ret && !err)
4561                                                 err = ret;
4562                                 }
4563                         }
4564                         /*
4565                          * if we failed to refill our space rsv, bail out
4566                          * and let the transaction restart
4567                          */
4568                         if (should_end) {
4569                                 err = -EAGAIN;
4570                                 goto error;
4571                         }
4572                         goto search_again;
4573                 } else {
4574                         path->slots[0]--;
4575                 }
4576         }
4577 out:
4578         if (pending_del_nr) {
4579                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4580                                       pending_del_nr);
4581                 if (ret)
4582                         btrfs_abort_transaction(trans, root, ret);
4583         }
4584 error:
4585         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4586                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4587
4588         btrfs_free_path(path);
4589
4590         if (be_nice && bytes_deleted > 32 * 1024 * 1024) {
4591                 unsigned long updates = trans->delayed_ref_updates;
4592                 if (updates) {
4593                         trans->delayed_ref_updates = 0;
4594                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4595                         if (ret && !err)
4596                                 err = ret;
4597                 }
4598         }
4599         return err;
4600 }
4601
4602 /*
4603  * btrfs_truncate_page - read, zero a chunk and write a page
4604  * @inode - inode that we're zeroing
4605  * @from - the offset to start zeroing
4606  * @len - the length to zero, 0 to zero the entire range respective to the
4607  *      offset
4608  * @front - zero up to the offset instead of from the offset on
4609  *
4610  * This will find the page for the "from" offset and cow the page and zero the
4611  * part we want to zero.  This is used with truncate and hole punching.
4612  */
4613 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4614                         int front)
4615 {
4616         struct address_space *mapping = inode->i_mapping;
4617         struct btrfs_root *root = BTRFS_I(inode)->root;
4618         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4619         struct btrfs_ordered_extent *ordered;
4620         struct extent_state *cached_state = NULL;
4621         char *kaddr;
4622         u32 blocksize = root->sectorsize;
4623         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4624         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4625         struct page *page;
4626         gfp_t mask = btrfs_alloc_write_mask(mapping);
4627         int ret = 0;
4628         u64 page_start;
4629         u64 page_end;
4630
4631         if ((offset & (blocksize - 1)) == 0 &&
4632             (!len || ((len & (blocksize - 1)) == 0)))
4633                 goto out;
4634         ret = btrfs_delalloc_reserve_space(inode,
4635                         round_down(from, PAGE_CACHE_SIZE), PAGE_CACHE_SIZE);
4636         if (ret)
4637                 goto out;
4638
4639 again:
4640         page = find_or_create_page(mapping, index, mask);
4641         if (!page) {
4642                 btrfs_delalloc_release_space(inode,
4643                                 round_down(from, PAGE_CACHE_SIZE),
4644                                 PAGE_CACHE_SIZE);
4645                 ret = -ENOMEM;
4646                 goto out;
4647         }
4648
4649         page_start = page_offset(page);
4650         page_end = page_start + PAGE_CACHE_SIZE - 1;
4651
4652         if (!PageUptodate(page)) {
4653                 ret = btrfs_readpage(NULL, page);
4654                 lock_page(page);
4655                 if (page->mapping != mapping) {
4656                         unlock_page(page);
4657                         page_cache_release(page);
4658                         goto again;
4659                 }
4660                 if (!PageUptodate(page)) {
4661                         ret = -EIO;
4662                         goto out_unlock;
4663                 }
4664         }
4665         wait_on_page_writeback(page);
4666
4667         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state);
4668         set_page_extent_mapped(page);
4669
4670         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4671         if (ordered) {
4672                 unlock_extent_cached(io_tree, page_start, page_end,
4673                                      &cached_state, GFP_NOFS);
4674                 unlock_page(page);
4675                 page_cache_release(page);
4676                 btrfs_start_ordered_extent(inode, ordered, 1);
4677                 btrfs_put_ordered_extent(ordered);
4678                 goto again;
4679         }
4680
4681         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4682                           EXTENT_DIRTY | EXTENT_DELALLOC |
4683                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4684                           0, 0, &cached_state, GFP_NOFS);
4685
4686         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4687                                         &cached_state);
4688         if (ret) {
4689                 unlock_extent_cached(io_tree, page_start, page_end,
4690                                      &cached_state, GFP_NOFS);
4691                 goto out_unlock;
4692         }
4693
4694         if (offset != PAGE_CACHE_SIZE) {
4695                 if (!len)
4696                         len = PAGE_CACHE_SIZE - offset;
4697                 kaddr = kmap(page);
4698                 if (front)
4699                         memset(kaddr, 0, offset);
4700                 else
4701                         memset(kaddr + offset, 0, len);
4702                 flush_dcache_page(page);
4703                 kunmap(page);
4704         }
4705         ClearPageChecked(page);
4706         set_page_dirty(page);
4707         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4708                              GFP_NOFS);
4709
4710 out_unlock:
4711         if (ret)
4712                 btrfs_delalloc_release_space(inode, page_start,
4713                                              PAGE_CACHE_SIZE);
4714         unlock_page(page);
4715         page_cache_release(page);
4716 out:
4717         return ret;
4718 }
4719
4720 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4721                              u64 offset, u64 len)
4722 {
4723         struct btrfs_trans_handle *trans;
4724         int ret;
4725
4726         /*
4727          * Still need to make sure the inode looks like it's been updated so
4728          * that any holes get logged if we fsync.
4729          */
4730         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4731                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4732                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4733                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4734                 return 0;
4735         }
4736
4737         /*
4738          * 1 - for the one we're dropping
4739          * 1 - for the one we're adding
4740          * 1 - for updating the inode.
4741          */
4742         trans = btrfs_start_transaction(root, 3);
4743         if (IS_ERR(trans))
4744                 return PTR_ERR(trans);
4745
4746         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4747         if (ret) {
4748                 btrfs_abort_transaction(trans, root, ret);
4749                 btrfs_end_transaction(trans, root);
4750                 return ret;
4751         }
4752
4753         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4754                                        0, 0, len, 0, len, 0, 0, 0);
4755         if (ret)
4756                 btrfs_abort_transaction(trans, root, ret);
4757         else
4758                 btrfs_update_inode(trans, root, inode);
4759         btrfs_end_transaction(trans, root);
4760         return ret;
4761 }
4762
4763 /*
4764  * This function puts in dummy file extents for the area we're creating a hole
4765  * for.  So if we are truncating this file to a larger size we need to insert
4766  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4767  * the range between oldsize and size
4768  */
4769 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4770 {
4771         struct btrfs_root *root = BTRFS_I(inode)->root;
4772         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4773         struct extent_map *em = NULL;
4774         struct extent_state *cached_state = NULL;
4775         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4776         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4777         u64 block_end = ALIGN(size, root->sectorsize);
4778         u64 last_byte;
4779         u64 cur_offset;
4780         u64 hole_size;
4781         int err = 0;
4782
4783         /*
4784          * If our size started in the middle of a page we need to zero out the
4785          * rest of the page before we expand the i_size, otherwise we could
4786          * expose stale data.
4787          */
4788         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4789         if (err)
4790                 return err;
4791
4792         if (size <= hole_start)
4793                 return 0;
4794
4795         while (1) {
4796                 struct btrfs_ordered_extent *ordered;
4797
4798                 lock_extent_bits(io_tree, hole_start, block_end - 1, 0,
4799                                  &cached_state);
4800                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4801                                                      block_end - hole_start);
4802                 if (!ordered)
4803                         break;
4804                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4805                                      &cached_state, GFP_NOFS);
4806                 btrfs_start_ordered_extent(inode, ordered, 1);
4807                 btrfs_put_ordered_extent(ordered);
4808         }
4809
4810         cur_offset = hole_start;
4811         while (1) {
4812                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4813                                 block_end - cur_offset, 0);
4814                 if (IS_ERR(em)) {
4815                         err = PTR_ERR(em);
4816                         em = NULL;
4817                         break;
4818                 }
4819                 last_byte = min(extent_map_end(em), block_end);
4820                 last_byte = ALIGN(last_byte , root->sectorsize);
4821                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4822                         struct extent_map *hole_em;
4823                         hole_size = last_byte - cur_offset;
4824
4825                         err = maybe_insert_hole(root, inode, cur_offset,
4826                                                 hole_size);
4827                         if (err)
4828                                 break;
4829                         btrfs_drop_extent_cache(inode, cur_offset,
4830                                                 cur_offset + hole_size - 1, 0);
4831                         hole_em = alloc_extent_map();
4832                         if (!hole_em) {
4833                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4834                                         &BTRFS_I(inode)->runtime_flags);
4835                                 goto next;
4836                         }
4837                         hole_em->start = cur_offset;
4838                         hole_em->len = hole_size;
4839                         hole_em->orig_start = cur_offset;
4840
4841                         hole_em->block_start = EXTENT_MAP_HOLE;
4842                         hole_em->block_len = 0;
4843                         hole_em->orig_block_len = 0;
4844                         hole_em->ram_bytes = hole_size;
4845                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4846                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4847                         hole_em->generation = root->fs_info->generation;
4848
4849                         while (1) {
4850                                 write_lock(&em_tree->lock);
4851                                 err = add_extent_mapping(em_tree, hole_em, 1);
4852                                 write_unlock(&em_tree->lock);
4853                                 if (err != -EEXIST)
4854                                         break;
4855                                 btrfs_drop_extent_cache(inode, cur_offset,
4856                                                         cur_offset +
4857                                                         hole_size - 1, 0);
4858                         }
4859                         free_extent_map(hole_em);
4860                 }
4861 next:
4862                 free_extent_map(em);
4863                 em = NULL;
4864                 cur_offset = last_byte;
4865                 if (cur_offset >= block_end)
4866                         break;
4867         }
4868         free_extent_map(em);
4869         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4870                              GFP_NOFS);
4871         return err;
4872 }
4873
4874 static int wait_snapshoting_atomic_t(atomic_t *a)
4875 {
4876         schedule();
4877         return 0;
4878 }
4879
4880 static void wait_for_snapshot_creation(struct btrfs_root *root)
4881 {
4882         while (true) {
4883                 int ret;
4884
4885                 ret = btrfs_start_write_no_snapshoting(root);
4886                 if (ret)
4887                         break;
4888                 wait_on_atomic_t(&root->will_be_snapshoted,
4889                                  wait_snapshoting_atomic_t,
4890                                  TASK_UNINTERRUPTIBLE);
4891         }
4892 }
4893
4894 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4895 {
4896         struct btrfs_root *root = BTRFS_I(inode)->root;
4897         struct btrfs_trans_handle *trans;
4898         loff_t oldsize = i_size_read(inode);
4899         loff_t newsize = attr->ia_size;
4900         int mask = attr->ia_valid;
4901         int ret;
4902
4903         /*
4904          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4905          * special case where we need to update the times despite not having
4906          * these flags set.  For all other operations the VFS set these flags
4907          * explicitly if it wants a timestamp update.
4908          */
4909         if (newsize != oldsize) {
4910                 inode_inc_iversion(inode);
4911                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4912                         inode->i_ctime = inode->i_mtime =
4913                                 current_fs_time(inode->i_sb);
4914         }
4915
4916         if (newsize > oldsize) {
4917                 truncate_pagecache(inode, newsize);
4918                 /*
4919                  * Don't do an expanding truncate while snapshoting is ongoing.
4920                  * This is to ensure the snapshot captures a fully consistent
4921                  * state of this file - if the snapshot captures this expanding
4922                  * truncation, it must capture all writes that happened before
4923                  * this truncation.
4924                  */
4925                 wait_for_snapshot_creation(root);
4926                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4927                 if (ret) {
4928                         btrfs_end_write_no_snapshoting(root);
4929                         return ret;
4930                 }
4931
4932                 trans = btrfs_start_transaction(root, 1);
4933                 if (IS_ERR(trans)) {
4934                         btrfs_end_write_no_snapshoting(root);
4935                         return PTR_ERR(trans);
4936                 }
4937
4938                 i_size_write(inode, newsize);
4939                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4940                 ret = btrfs_update_inode(trans, root, inode);
4941                 btrfs_end_write_no_snapshoting(root);
4942                 btrfs_end_transaction(trans, root);
4943         } else {
4944
4945                 /*
4946                  * We're truncating a file that used to have good data down to
4947                  * zero. Make sure it gets into the ordered flush list so that
4948                  * any new writes get down to disk quickly.
4949                  */
4950                 if (newsize == 0)
4951                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4952                                 &BTRFS_I(inode)->runtime_flags);
4953
4954                 /*
4955                  * 1 for the orphan item we're going to add
4956                  * 1 for the orphan item deletion.
4957                  */
4958                 trans = btrfs_start_transaction(root, 2);
4959                 if (IS_ERR(trans))
4960                         return PTR_ERR(trans);
4961
4962                 /*
4963                  * We need to do this in case we fail at _any_ point during the
4964                  * actual truncate.  Once we do the truncate_setsize we could
4965                  * invalidate pages which forces any outstanding ordered io to
4966                  * be instantly completed which will give us extents that need
4967                  * to be truncated.  If we fail to get an orphan inode down we
4968                  * could have left over extents that were never meant to live,
4969                  * so we need to garuntee from this point on that everything
4970                  * will be consistent.
4971                  */
4972                 ret = btrfs_orphan_add(trans, inode);
4973                 btrfs_end_transaction(trans, root);
4974                 if (ret)
4975                         return ret;
4976
4977                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4978                 truncate_setsize(inode, newsize);
4979
4980                 /* Disable nonlocked read DIO to avoid the end less truncate */
4981                 btrfs_inode_block_unlocked_dio(inode);
4982                 inode_dio_wait(inode);
4983                 btrfs_inode_resume_unlocked_dio(inode);
4984
4985                 ret = btrfs_truncate(inode);
4986                 if (ret && inode->i_nlink) {
4987                         int err;
4988
4989                         /*
4990                          * failed to truncate, disk_i_size is only adjusted down
4991                          * as we remove extents, so it should represent the true
4992                          * size of the inode, so reset the in memory size and
4993                          * delete our orphan entry.
4994                          */
4995                         trans = btrfs_join_transaction(root);
4996                         if (IS_ERR(trans)) {
4997                                 btrfs_orphan_del(NULL, inode);
4998                                 return ret;
4999                         }
5000                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5001                         err = btrfs_orphan_del(trans, inode);
5002                         if (err)
5003                                 btrfs_abort_transaction(trans, root, err);
5004                         btrfs_end_transaction(trans, root);
5005                 }
5006         }
5007
5008         return ret;
5009 }
5010
5011 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5012 {
5013         struct inode *inode = d_inode(dentry);
5014         struct btrfs_root *root = BTRFS_I(inode)->root;
5015         int err;
5016
5017         if (btrfs_root_readonly(root))
5018                 return -EROFS;
5019
5020         err = inode_change_ok(inode, attr);
5021         if (err)
5022                 return err;
5023
5024         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5025                 err = btrfs_setsize(inode, attr);
5026                 if (err)
5027                         return err;
5028         }
5029
5030         if (attr->ia_valid) {
5031                 setattr_copy(inode, attr);
5032                 inode_inc_iversion(inode);
5033                 err = btrfs_dirty_inode(inode);
5034
5035                 if (!err && attr->ia_valid & ATTR_MODE)
5036                         err = posix_acl_chmod(inode, inode->i_mode);
5037         }
5038
5039         return err;
5040 }
5041
5042 /*
5043  * While truncating the inode pages during eviction, we get the VFS calling
5044  * btrfs_invalidatepage() against each page of the inode. This is slow because
5045  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5046  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5047  * extent_state structures over and over, wasting lots of time.
5048  *
5049  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5050  * those expensive operations on a per page basis and do only the ordered io
5051  * finishing, while we release here the extent_map and extent_state structures,
5052  * without the excessive merging and splitting.
5053  */
5054 static void evict_inode_truncate_pages(struct inode *inode)
5055 {
5056         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5057         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5058         struct rb_node *node;
5059
5060         ASSERT(inode->i_state & I_FREEING);
5061         truncate_inode_pages_final(&inode->i_data);
5062
5063         write_lock(&map_tree->lock);
5064         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5065                 struct extent_map *em;
5066
5067                 node = rb_first(&map_tree->map);
5068                 em = rb_entry(node, struct extent_map, rb_node);
5069                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5070                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5071                 remove_extent_mapping(map_tree, em);
5072                 free_extent_map(em);
5073                 if (need_resched()) {
5074                         write_unlock(&map_tree->lock);
5075                         cond_resched();
5076                         write_lock(&map_tree->lock);
5077                 }
5078         }
5079         write_unlock(&map_tree->lock);
5080
5081         /*
5082          * Keep looping until we have no more ranges in the io tree.
5083          * We can have ongoing bios started by readpages (called from readahead)
5084          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5085          * still in progress (unlocked the pages in the bio but did not yet
5086          * unlocked the ranges in the io tree). Therefore this means some
5087          * ranges can still be locked and eviction started because before
5088          * submitting those bios, which are executed by a separate task (work
5089          * queue kthread), inode references (inode->i_count) were not taken
5090          * (which would be dropped in the end io callback of each bio).
5091          * Therefore here we effectively end up waiting for those bios and
5092          * anyone else holding locked ranges without having bumped the inode's
5093          * reference count - if we don't do it, when they access the inode's
5094          * io_tree to unlock a range it may be too late, leading to an
5095          * use-after-free issue.
5096          */
5097         spin_lock(&io_tree->lock);
5098         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5099                 struct extent_state *state;
5100                 struct extent_state *cached_state = NULL;
5101                 u64 start;
5102                 u64 end;
5103
5104                 node = rb_first(&io_tree->state);
5105                 state = rb_entry(node, struct extent_state, rb_node);
5106                 start = state->start;
5107                 end = state->end;
5108                 spin_unlock(&io_tree->lock);
5109
5110                 lock_extent_bits(io_tree, start, end, 0, &cached_state);
5111
5112                 /*
5113                  * If still has DELALLOC flag, the extent didn't reach disk,
5114                  * and its reserved space won't be freed by delayed_ref.
5115                  * So we need to free its reserved space here.
5116                  * (Refer to comment in btrfs_invalidatepage, case 2)
5117                  *
5118                  * Note, end is the bytenr of last byte, so we need + 1 here.
5119                  */
5120                 if (state->state & EXTENT_DELALLOC)
5121                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5122
5123                 clear_extent_bit(io_tree, start, end,
5124                                  EXTENT_LOCKED | EXTENT_DIRTY |
5125                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5126                                  EXTENT_DEFRAG, 1, 1,
5127                                  &cached_state, GFP_NOFS);
5128
5129                 cond_resched();
5130                 spin_lock(&io_tree->lock);
5131         }
5132         spin_unlock(&io_tree->lock);
5133 }
5134
5135 void btrfs_evict_inode(struct inode *inode)
5136 {
5137         struct btrfs_trans_handle *trans;
5138         struct btrfs_root *root = BTRFS_I(inode)->root;
5139         struct btrfs_block_rsv *rsv, *global_rsv;
5140         int steal_from_global = 0;
5141         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5142         int ret;
5143
5144         trace_btrfs_inode_evict(inode);
5145
5146         evict_inode_truncate_pages(inode);
5147
5148         if (inode->i_nlink &&
5149             ((btrfs_root_refs(&root->root_item) != 0 &&
5150               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5151              btrfs_is_free_space_inode(inode)))
5152                 goto no_delete;
5153
5154         if (is_bad_inode(inode)) {
5155                 btrfs_orphan_del(NULL, inode);
5156                 goto no_delete;
5157         }
5158         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5159         if (!special_file(inode->i_mode))
5160                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5161
5162         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5163
5164         if (root->fs_info->log_root_recovering) {
5165                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5166                                  &BTRFS_I(inode)->runtime_flags));
5167                 goto no_delete;
5168         }
5169
5170         if (inode->i_nlink > 0) {
5171                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5172                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5173                 goto no_delete;
5174         }
5175
5176         ret = btrfs_commit_inode_delayed_inode(inode);
5177         if (ret) {
5178                 btrfs_orphan_del(NULL, inode);
5179                 goto no_delete;
5180         }
5181
5182         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5183         if (!rsv) {
5184                 btrfs_orphan_del(NULL, inode);
5185                 goto no_delete;
5186         }
5187         rsv->size = min_size;
5188         rsv->failfast = 1;
5189         global_rsv = &root->fs_info->global_block_rsv;
5190
5191         btrfs_i_size_write(inode, 0);
5192
5193         /*
5194          * This is a bit simpler than btrfs_truncate since we've already
5195          * reserved our space for our orphan item in the unlink, so we just
5196          * need to reserve some slack space in case we add bytes and update
5197          * inode item when doing the truncate.
5198          */
5199         while (1) {
5200                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5201                                              BTRFS_RESERVE_FLUSH_LIMIT);
5202
5203                 /*
5204                  * Try and steal from the global reserve since we will
5205                  * likely not use this space anyway, we want to try as
5206                  * hard as possible to get this to work.
5207                  */
5208                 if (ret)
5209                         steal_from_global++;
5210                 else
5211                         steal_from_global = 0;
5212                 ret = 0;
5213
5214                 /*
5215                  * steal_from_global == 0: we reserved stuff, hooray!
5216                  * steal_from_global == 1: we didn't reserve stuff, boo!
5217                  * steal_from_global == 2: we've committed, still not a lot of
5218                  * room but maybe we'll have room in the global reserve this
5219                  * time.
5220                  * steal_from_global == 3: abandon all hope!
5221                  */
5222                 if (steal_from_global > 2) {
5223                         btrfs_warn(root->fs_info,
5224                                 "Could not get space for a delete, will truncate on mount %d",
5225                                 ret);
5226                         btrfs_orphan_del(NULL, inode);
5227                         btrfs_free_block_rsv(root, rsv);
5228                         goto no_delete;
5229                 }
5230
5231                 trans = btrfs_join_transaction(root);
5232                 if (IS_ERR(trans)) {
5233                         btrfs_orphan_del(NULL, inode);
5234                         btrfs_free_block_rsv(root, rsv);
5235                         goto no_delete;
5236                 }
5237
5238                 /*
5239                  * We can't just steal from the global reserve, we need tomake
5240                  * sure there is room to do it, if not we need to commit and try
5241                  * again.
5242                  */
5243                 if (steal_from_global) {
5244                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5245                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5246                                                               min_size);
5247                         else
5248                                 ret = -ENOSPC;
5249                 }
5250
5251                 /*
5252                  * Couldn't steal from the global reserve, we have too much
5253                  * pending stuff built up, commit the transaction and try it
5254                  * again.
5255                  */
5256                 if (ret) {
5257                         ret = btrfs_commit_transaction(trans, root);
5258                         if (ret) {
5259                                 btrfs_orphan_del(NULL, inode);
5260                                 btrfs_free_block_rsv(root, rsv);
5261                                 goto no_delete;
5262                         }
5263                         continue;
5264                 } else {
5265                         steal_from_global = 0;
5266                 }
5267
5268                 trans->block_rsv = rsv;
5269
5270                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5271                 if (ret != -ENOSPC && ret != -EAGAIN)
5272                         break;
5273
5274                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5275                 btrfs_end_transaction(trans, root);
5276                 trans = NULL;
5277                 btrfs_btree_balance_dirty(root);
5278         }
5279
5280         btrfs_free_block_rsv(root, rsv);
5281
5282         /*
5283          * Errors here aren't a big deal, it just means we leave orphan items
5284          * in the tree.  They will be cleaned up on the next mount.
5285          */
5286         if (ret == 0) {
5287                 trans->block_rsv = root->orphan_block_rsv;
5288                 btrfs_orphan_del(trans, inode);
5289         } else {
5290                 btrfs_orphan_del(NULL, inode);
5291         }
5292
5293         trans->block_rsv = &root->fs_info->trans_block_rsv;
5294         if (!(root == root->fs_info->tree_root ||
5295               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5296                 btrfs_return_ino(root, btrfs_ino(inode));
5297
5298         btrfs_end_transaction(trans, root);
5299         btrfs_btree_balance_dirty(root);
5300 no_delete:
5301         btrfs_remove_delayed_node(inode);
5302         clear_inode(inode);
5303         return;
5304 }
5305
5306 /*
5307  * this returns the key found in the dir entry in the location pointer.
5308  * If no dir entries were found, location->objectid is 0.
5309  */
5310 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5311                                struct btrfs_key *location)
5312 {
5313         const char *name = dentry->d_name.name;
5314         int namelen = dentry->d_name.len;
5315         struct btrfs_dir_item *di;
5316         struct btrfs_path *path;
5317         struct btrfs_root *root = BTRFS_I(dir)->root;
5318         int ret = 0;
5319
5320         path = btrfs_alloc_path();
5321         if (!path)
5322                 return -ENOMEM;
5323
5324         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5325                                     namelen, 0);
5326         if (IS_ERR(di))
5327                 ret = PTR_ERR(di);
5328
5329         if (IS_ERR_OR_NULL(di))
5330                 goto out_err;
5331
5332         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5333 out:
5334         btrfs_free_path(path);
5335         return ret;
5336 out_err:
5337         location->objectid = 0;
5338         goto out;
5339 }
5340
5341 /*
5342  * when we hit a tree root in a directory, the btrfs part of the inode
5343  * needs to be changed to reflect the root directory of the tree root.  This
5344  * is kind of like crossing a mount point.
5345  */
5346 static int fixup_tree_root_location(struct btrfs_root *root,
5347                                     struct inode *dir,
5348                                     struct dentry *dentry,
5349                                     struct btrfs_key *location,
5350                                     struct btrfs_root **sub_root)
5351 {
5352         struct btrfs_path *path;
5353         struct btrfs_root *new_root;
5354         struct btrfs_root_ref *ref;
5355         struct extent_buffer *leaf;
5356         struct btrfs_key key;
5357         int ret;
5358         int err = 0;
5359
5360         path = btrfs_alloc_path();
5361         if (!path) {
5362                 err = -ENOMEM;
5363                 goto out;
5364         }
5365
5366         err = -ENOENT;
5367         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5368         key.type = BTRFS_ROOT_REF_KEY;
5369         key.offset = location->objectid;
5370
5371         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5372                                 0, 0);
5373         if (ret) {
5374                 if (ret < 0)
5375                         err = ret;
5376                 goto out;
5377         }
5378
5379         leaf = path->nodes[0];
5380         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5381         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5382             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5383                 goto out;
5384
5385         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5386                                    (unsigned long)(ref + 1),
5387                                    dentry->d_name.len);
5388         if (ret)
5389                 goto out;
5390
5391         btrfs_release_path(path);
5392
5393         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5394         if (IS_ERR(new_root)) {
5395                 err = PTR_ERR(new_root);
5396                 goto out;
5397         }
5398
5399         *sub_root = new_root;
5400         location->objectid = btrfs_root_dirid(&new_root->root_item);
5401         location->type = BTRFS_INODE_ITEM_KEY;
5402         location->offset = 0;
5403         err = 0;
5404 out:
5405         btrfs_free_path(path);
5406         return err;
5407 }
5408
5409 static void inode_tree_add(struct inode *inode)
5410 {
5411         struct btrfs_root *root = BTRFS_I(inode)->root;
5412         struct btrfs_inode *entry;
5413         struct rb_node **p;
5414         struct rb_node *parent;
5415         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5416         u64 ino = btrfs_ino(inode);
5417
5418         if (inode_unhashed(inode))
5419                 return;
5420         parent = NULL;
5421         spin_lock(&root->inode_lock);
5422         p = &root->inode_tree.rb_node;
5423         while (*p) {
5424                 parent = *p;
5425                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5426
5427                 if (ino < btrfs_ino(&entry->vfs_inode))
5428                         p = &parent->rb_left;
5429                 else if (ino > btrfs_ino(&entry->vfs_inode))
5430                         p = &parent->rb_right;
5431                 else {
5432                         WARN_ON(!(entry->vfs_inode.i_state &
5433                                   (I_WILL_FREE | I_FREEING)));
5434                         rb_replace_node(parent, new, &root->inode_tree);
5435                         RB_CLEAR_NODE(parent);
5436                         spin_unlock(&root->inode_lock);
5437                         return;
5438                 }
5439         }
5440         rb_link_node(new, parent, p);
5441         rb_insert_color(new, &root->inode_tree);
5442         spin_unlock(&root->inode_lock);
5443 }
5444
5445 static void inode_tree_del(struct inode *inode)
5446 {
5447         struct btrfs_root *root = BTRFS_I(inode)->root;
5448         int empty = 0;
5449
5450         spin_lock(&root->inode_lock);
5451         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5452                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5453                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5454                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5455         }
5456         spin_unlock(&root->inode_lock);
5457
5458         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5459                 synchronize_srcu(&root->fs_info->subvol_srcu);
5460                 spin_lock(&root->inode_lock);
5461                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5462                 spin_unlock(&root->inode_lock);
5463                 if (empty)
5464                         btrfs_add_dead_root(root);
5465         }
5466 }
5467
5468 void btrfs_invalidate_inodes(struct btrfs_root *root)
5469 {
5470         struct rb_node *node;
5471         struct rb_node *prev;
5472         struct btrfs_inode *entry;
5473         struct inode *inode;
5474         u64 objectid = 0;
5475
5476         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5477                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5478
5479         spin_lock(&root->inode_lock);
5480 again:
5481         node = root->inode_tree.rb_node;
5482         prev = NULL;
5483         while (node) {
5484                 prev = node;
5485                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5486
5487                 if (objectid < btrfs_ino(&entry->vfs_inode))
5488                         node = node->rb_left;
5489                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5490                         node = node->rb_right;
5491                 else
5492                         break;
5493         }
5494         if (!node) {
5495                 while (prev) {
5496                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5497                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5498                                 node = prev;
5499                                 break;
5500                         }
5501                         prev = rb_next(prev);
5502                 }
5503         }
5504         while (node) {
5505                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5506                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5507                 inode = igrab(&entry->vfs_inode);
5508                 if (inode) {
5509                         spin_unlock(&root->inode_lock);
5510                         if (atomic_read(&inode->i_count) > 1)
5511                                 d_prune_aliases(inode);
5512                         /*
5513                          * btrfs_drop_inode will have it removed from
5514                          * the inode cache when its usage count
5515                          * hits zero.
5516                          */
5517                         iput(inode);
5518                         cond_resched();
5519                         spin_lock(&root->inode_lock);
5520                         goto again;
5521                 }
5522
5523                 if (cond_resched_lock(&root->inode_lock))
5524                         goto again;
5525
5526                 node = rb_next(node);
5527         }
5528         spin_unlock(&root->inode_lock);
5529 }
5530
5531 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5532 {
5533         struct btrfs_iget_args *args = p;
5534         inode->i_ino = args->location->objectid;
5535         memcpy(&BTRFS_I(inode)->location, args->location,
5536                sizeof(*args->location));
5537         BTRFS_I(inode)->root = args->root;
5538         return 0;
5539 }
5540
5541 static int btrfs_find_actor(struct inode *inode, void *opaque)
5542 {
5543         struct btrfs_iget_args *args = opaque;
5544         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5545                 args->root == BTRFS_I(inode)->root;
5546 }
5547
5548 static struct inode *btrfs_iget_locked(struct super_block *s,
5549                                        struct btrfs_key *location,
5550                                        struct btrfs_root *root)
5551 {
5552         struct inode *inode;
5553         struct btrfs_iget_args args;
5554         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5555
5556         args.location = location;
5557         args.root = root;
5558
5559         inode = iget5_locked(s, hashval, btrfs_find_actor,
5560                              btrfs_init_locked_inode,
5561                              (void *)&args);
5562         return inode;
5563 }
5564
5565 /* Get an inode object given its location and corresponding root.
5566  * Returns in *is_new if the inode was read from disk
5567  */
5568 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5569                          struct btrfs_root *root, int *new)
5570 {
5571         struct inode *inode;
5572
5573         inode = btrfs_iget_locked(s, location, root);
5574         if (!inode)
5575                 return ERR_PTR(-ENOMEM);
5576
5577         if (inode->i_state & I_NEW) {
5578                 btrfs_read_locked_inode(inode);
5579                 if (!is_bad_inode(inode)) {
5580                         inode_tree_add(inode);
5581                         unlock_new_inode(inode);
5582                         if (new)
5583                                 *new = 1;
5584                 } else {
5585                         unlock_new_inode(inode);
5586                         iput(inode);
5587                         inode = ERR_PTR(-ESTALE);
5588                 }
5589         }
5590
5591         return inode;
5592 }
5593
5594 static struct inode *new_simple_dir(struct super_block *s,
5595                                     struct btrfs_key *key,
5596                                     struct btrfs_root *root)
5597 {
5598         struct inode *inode = new_inode(s);
5599
5600         if (!inode)
5601                 return ERR_PTR(-ENOMEM);
5602
5603         BTRFS_I(inode)->root = root;
5604         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5605         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5606
5607         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5608         inode->i_op = &btrfs_dir_ro_inode_operations;
5609         inode->i_fop = &simple_dir_operations;
5610         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5611         inode->i_mtime = CURRENT_TIME;
5612         inode->i_atime = inode->i_mtime;
5613         inode->i_ctime = inode->i_mtime;
5614         BTRFS_I(inode)->i_otime = inode->i_mtime;
5615
5616         return inode;
5617 }
5618
5619 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5620 {
5621         struct inode *inode;
5622         struct btrfs_root *root = BTRFS_I(dir)->root;
5623         struct btrfs_root *sub_root = root;
5624         struct btrfs_key location;
5625         int index;
5626         int ret = 0;
5627
5628         if (dentry->d_name.len > BTRFS_NAME_LEN)
5629                 return ERR_PTR(-ENAMETOOLONG);
5630
5631         ret = btrfs_inode_by_name(dir, dentry, &location);
5632         if (ret < 0)
5633                 return ERR_PTR(ret);
5634
5635         if (location.objectid == 0)
5636                 return ERR_PTR(-ENOENT);
5637
5638         if (location.type == BTRFS_INODE_ITEM_KEY) {
5639                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5640                 return inode;
5641         }
5642
5643         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5644
5645         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5646         ret = fixup_tree_root_location(root, dir, dentry,
5647                                        &location, &sub_root);
5648         if (ret < 0) {
5649                 if (ret != -ENOENT)
5650                         inode = ERR_PTR(ret);
5651                 else
5652                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5653         } else {
5654                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5655         }
5656         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5657
5658         if (!IS_ERR(inode) && root != sub_root) {
5659                 down_read(&root->fs_info->cleanup_work_sem);
5660                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5661                         ret = btrfs_orphan_cleanup(sub_root);
5662                 up_read(&root->fs_info->cleanup_work_sem);
5663                 if (ret) {
5664                         iput(inode);
5665                         inode = ERR_PTR(ret);
5666                 }
5667         }
5668
5669         return inode;
5670 }
5671
5672 static int btrfs_dentry_delete(const struct dentry *dentry)
5673 {
5674         struct btrfs_root *root;
5675         struct inode *inode = d_inode(dentry);
5676
5677         if (!inode && !IS_ROOT(dentry))
5678                 inode = d_inode(dentry->d_parent);
5679
5680         if (inode) {
5681                 root = BTRFS_I(inode)->root;
5682                 if (btrfs_root_refs(&root->root_item) == 0)
5683                         return 1;
5684
5685                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5686                         return 1;
5687         }
5688         return 0;
5689 }
5690
5691 static void btrfs_dentry_release(struct dentry *dentry)
5692 {
5693         kfree(dentry->d_fsdata);
5694 }
5695
5696 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5697                                    unsigned int flags)
5698 {
5699         struct inode *inode;
5700
5701         inode = btrfs_lookup_dentry(dir, dentry);
5702         if (IS_ERR(inode)) {
5703                 if (PTR_ERR(inode) == -ENOENT)
5704                         inode = NULL;
5705                 else
5706                         return ERR_CAST(inode);
5707         }
5708
5709         return d_splice_alias(inode, dentry);
5710 }
5711
5712 unsigned char btrfs_filetype_table[] = {
5713         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5714 };
5715
5716 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5717 {
5718         struct inode *inode = file_inode(file);
5719         struct btrfs_root *root = BTRFS_I(inode)->root;
5720         struct btrfs_item *item;
5721         struct btrfs_dir_item *di;
5722         struct btrfs_key key;
5723         struct btrfs_key found_key;
5724         struct btrfs_path *path;
5725         struct list_head ins_list;
5726         struct list_head del_list;
5727         int ret;
5728         struct extent_buffer *leaf;
5729         int slot;
5730         unsigned char d_type;
5731         int over = 0;
5732         u32 di_cur;
5733         u32 di_total;
5734         u32 di_len;
5735         int key_type = BTRFS_DIR_INDEX_KEY;
5736         char tmp_name[32];
5737         char *name_ptr;
5738         int name_len;
5739         int is_curr = 0;        /* ctx->pos points to the current index? */
5740         bool emitted;
5741
5742         /* FIXME, use a real flag for deciding about the key type */
5743         if (root->fs_info->tree_root == root)
5744                 key_type = BTRFS_DIR_ITEM_KEY;
5745
5746         if (!dir_emit_dots(file, ctx))
5747                 return 0;
5748
5749         path = btrfs_alloc_path();
5750         if (!path)
5751                 return -ENOMEM;
5752
5753         path->reada = 1;
5754
5755         if (key_type == BTRFS_DIR_INDEX_KEY) {
5756                 INIT_LIST_HEAD(&ins_list);
5757                 INIT_LIST_HEAD(&del_list);
5758                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5759         }
5760
5761         key.type = key_type;
5762         key.offset = ctx->pos;
5763         key.objectid = btrfs_ino(inode);
5764
5765         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5766         if (ret < 0)
5767                 goto err;
5768
5769         emitted = false;
5770         while (1) {
5771                 leaf = path->nodes[0];
5772                 slot = path->slots[0];
5773                 if (slot >= btrfs_header_nritems(leaf)) {
5774                         ret = btrfs_next_leaf(root, path);
5775                         if (ret < 0)
5776                                 goto err;
5777                         else if (ret > 0)
5778                                 break;
5779                         continue;
5780                 }
5781
5782                 item = btrfs_item_nr(slot);
5783                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5784
5785                 if (found_key.objectid != key.objectid)
5786                         break;
5787                 if (found_key.type != key_type)
5788                         break;
5789                 if (found_key.offset < ctx->pos)
5790                         goto next;
5791                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5792                     btrfs_should_delete_dir_index(&del_list,
5793                                                   found_key.offset))
5794                         goto next;
5795
5796                 ctx->pos = found_key.offset;
5797                 is_curr = 1;
5798
5799                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5800                 di_cur = 0;
5801                 di_total = btrfs_item_size(leaf, item);
5802
5803                 while (di_cur < di_total) {
5804                         struct btrfs_key location;
5805
5806                         if (verify_dir_item(root, leaf, di))
5807                                 break;
5808
5809                         name_len = btrfs_dir_name_len(leaf, di);
5810                         if (name_len <= sizeof(tmp_name)) {
5811                                 name_ptr = tmp_name;
5812                         } else {
5813                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5814                                 if (!name_ptr) {
5815                                         ret = -ENOMEM;
5816                                         goto err;
5817                                 }
5818                         }
5819                         read_extent_buffer(leaf, name_ptr,
5820                                            (unsigned long)(di + 1), name_len);
5821
5822                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5823                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5824
5825
5826                         /* is this a reference to our own snapshot? If so
5827                          * skip it.
5828                          *
5829                          * In contrast to old kernels, we insert the snapshot's
5830                          * dir item and dir index after it has been created, so
5831                          * we won't find a reference to our own snapshot. We
5832                          * still keep the following code for backward
5833                          * compatibility.
5834                          */
5835                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5836                             location.objectid == root->root_key.objectid) {
5837                                 over = 0;
5838                                 goto skip;
5839                         }
5840                         over = !dir_emit(ctx, name_ptr, name_len,
5841                                        location.objectid, d_type);
5842
5843 skip:
5844                         if (name_ptr != tmp_name)
5845                                 kfree(name_ptr);
5846
5847                         if (over)
5848                                 goto nopos;
5849                         emitted = true;
5850                         di_len = btrfs_dir_name_len(leaf, di) +
5851                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5852                         di_cur += di_len;
5853                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5854                 }
5855 next:
5856                 path->slots[0]++;
5857         }
5858
5859         if (key_type == BTRFS_DIR_INDEX_KEY) {
5860                 if (is_curr)
5861                         ctx->pos++;
5862                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5863                 if (ret)
5864                         goto nopos;
5865         }
5866
5867         /*
5868          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5869          * it was was set to the termination value in previous call. We assume
5870          * that "." and ".." were emitted if we reach this point and set the
5871          * termination value as well for an empty directory.
5872          */
5873         if (ctx->pos > 2 && !emitted)
5874                 goto nopos;
5875
5876         /* Reached end of directory/root. Bump pos past the last item. */
5877         ctx->pos++;
5878
5879         /*
5880          * Stop new entries from being returned after we return the last
5881          * entry.
5882          *
5883          * New directory entries are assigned a strictly increasing
5884          * offset.  This means that new entries created during readdir
5885          * are *guaranteed* to be seen in the future by that readdir.
5886          * This has broken buggy programs which operate on names as
5887          * they're returned by readdir.  Until we re-use freed offsets
5888          * we have this hack to stop new entries from being returned
5889          * under the assumption that they'll never reach this huge
5890          * offset.
5891          *
5892          * This is being careful not to overflow 32bit loff_t unless the
5893          * last entry requires it because doing so has broken 32bit apps
5894          * in the past.
5895          */
5896         if (key_type == BTRFS_DIR_INDEX_KEY) {
5897                 if (ctx->pos >= INT_MAX)
5898                         ctx->pos = LLONG_MAX;
5899                 else
5900                         ctx->pos = INT_MAX;
5901         }
5902 nopos:
5903         ret = 0;
5904 err:
5905         if (key_type == BTRFS_DIR_INDEX_KEY)
5906                 btrfs_put_delayed_items(&ins_list, &del_list);
5907         btrfs_free_path(path);
5908         return ret;
5909 }
5910
5911 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5912 {
5913         struct btrfs_root *root = BTRFS_I(inode)->root;
5914         struct btrfs_trans_handle *trans;
5915         int ret = 0;
5916         bool nolock = false;
5917
5918         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5919                 return 0;
5920
5921         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5922                 nolock = true;
5923
5924         if (wbc->sync_mode == WB_SYNC_ALL) {
5925                 if (nolock)
5926                         trans = btrfs_join_transaction_nolock(root);
5927                 else
5928                         trans = btrfs_join_transaction(root);
5929                 if (IS_ERR(trans))
5930                         return PTR_ERR(trans);
5931                 ret = btrfs_commit_transaction(trans, root);
5932         }
5933         return ret;
5934 }
5935
5936 /*
5937  * This is somewhat expensive, updating the tree every time the
5938  * inode changes.  But, it is most likely to find the inode in cache.
5939  * FIXME, needs more benchmarking...there are no reasons other than performance
5940  * to keep or drop this code.
5941  */
5942 static int btrfs_dirty_inode(struct inode *inode)
5943 {
5944         struct btrfs_root *root = BTRFS_I(inode)->root;
5945         struct btrfs_trans_handle *trans;
5946         int ret;
5947
5948         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5949                 return 0;
5950
5951         trans = btrfs_join_transaction(root);
5952         if (IS_ERR(trans))
5953                 return PTR_ERR(trans);
5954
5955         ret = btrfs_update_inode(trans, root, inode);
5956         if (ret && ret == -ENOSPC) {
5957                 /* whoops, lets try again with the full transaction */
5958                 btrfs_end_transaction(trans, root);
5959                 trans = btrfs_start_transaction(root, 1);
5960                 if (IS_ERR(trans))
5961                         return PTR_ERR(trans);
5962
5963                 ret = btrfs_update_inode(trans, root, inode);
5964         }
5965         btrfs_end_transaction(trans, root);
5966         if (BTRFS_I(inode)->delayed_node)
5967                 btrfs_balance_delayed_items(root);
5968
5969         return ret;
5970 }
5971
5972 /*
5973  * This is a copy of file_update_time.  We need this so we can return error on
5974  * ENOSPC for updating the inode in the case of file write and mmap writes.
5975  */
5976 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5977                              int flags)
5978 {
5979         struct btrfs_root *root = BTRFS_I(inode)->root;
5980
5981         if (btrfs_root_readonly(root))
5982                 return -EROFS;
5983
5984         if (flags & S_VERSION)
5985                 inode_inc_iversion(inode);
5986         if (flags & S_CTIME)
5987                 inode->i_ctime = *now;
5988         if (flags & S_MTIME)
5989                 inode->i_mtime = *now;
5990         if (flags & S_ATIME)
5991                 inode->i_atime = *now;
5992         return btrfs_dirty_inode(inode);
5993 }
5994
5995 /*
5996  * find the highest existing sequence number in a directory
5997  * and then set the in-memory index_cnt variable to reflect
5998  * free sequence numbers
5999  */
6000 static int btrfs_set_inode_index_count(struct inode *inode)
6001 {
6002         struct btrfs_root *root = BTRFS_I(inode)->root;
6003         struct btrfs_key key, found_key;
6004         struct btrfs_path *path;
6005         struct extent_buffer *leaf;
6006         int ret;
6007
6008         key.objectid = btrfs_ino(inode);
6009         key.type = BTRFS_DIR_INDEX_KEY;
6010         key.offset = (u64)-1;
6011
6012         path = btrfs_alloc_path();
6013         if (!path)
6014                 return -ENOMEM;
6015
6016         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6017         if (ret < 0)
6018                 goto out;
6019         /* FIXME: we should be able to handle this */
6020         if (ret == 0)
6021                 goto out;
6022         ret = 0;
6023
6024         /*
6025          * MAGIC NUMBER EXPLANATION:
6026          * since we search a directory based on f_pos we have to start at 2
6027          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6028          * else has to start at 2
6029          */
6030         if (path->slots[0] == 0) {
6031                 BTRFS_I(inode)->index_cnt = 2;
6032                 goto out;
6033         }
6034
6035         path->slots[0]--;
6036
6037         leaf = path->nodes[0];
6038         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6039
6040         if (found_key.objectid != btrfs_ino(inode) ||
6041             found_key.type != BTRFS_DIR_INDEX_KEY) {
6042                 BTRFS_I(inode)->index_cnt = 2;
6043                 goto out;
6044         }
6045
6046         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6047 out:
6048         btrfs_free_path(path);
6049         return ret;
6050 }
6051
6052 /*
6053  * helper to find a free sequence number in a given directory.  This current
6054  * code is very simple, later versions will do smarter things in the btree
6055  */
6056 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6057 {
6058         int ret = 0;
6059
6060         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6061                 ret = btrfs_inode_delayed_dir_index_count(dir);
6062                 if (ret) {
6063                         ret = btrfs_set_inode_index_count(dir);
6064                         if (ret)
6065                                 return ret;
6066                 }
6067         }
6068
6069         *index = BTRFS_I(dir)->index_cnt;
6070         BTRFS_I(dir)->index_cnt++;
6071
6072         return ret;
6073 }
6074
6075 static int btrfs_insert_inode_locked(struct inode *inode)
6076 {
6077         struct btrfs_iget_args args;
6078         args.location = &BTRFS_I(inode)->location;
6079         args.root = BTRFS_I(inode)->root;
6080
6081         return insert_inode_locked4(inode,
6082                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6083                    btrfs_find_actor, &args);
6084 }
6085
6086 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6087                                      struct btrfs_root *root,
6088                                      struct inode *dir,
6089                                      const char *name, int name_len,
6090                                      u64 ref_objectid, u64 objectid,
6091                                      umode_t mode, u64 *index)
6092 {
6093         struct inode *inode;
6094         struct btrfs_inode_item *inode_item;
6095         struct btrfs_key *location;
6096         struct btrfs_path *path;
6097         struct btrfs_inode_ref *ref;
6098         struct btrfs_key key[2];
6099         u32 sizes[2];
6100         int nitems = name ? 2 : 1;
6101         unsigned long ptr;
6102         int ret;
6103
6104         path = btrfs_alloc_path();
6105         if (!path)
6106                 return ERR_PTR(-ENOMEM);
6107
6108         inode = new_inode(root->fs_info->sb);
6109         if (!inode) {
6110                 btrfs_free_path(path);
6111                 return ERR_PTR(-ENOMEM);
6112         }
6113
6114         /*
6115          * O_TMPFILE, set link count to 0, so that after this point,
6116          * we fill in an inode item with the correct link count.
6117          */
6118         if (!name)
6119                 set_nlink(inode, 0);
6120
6121         /*
6122          * we have to initialize this early, so we can reclaim the inode
6123          * number if we fail afterwards in this function.
6124          */
6125         inode->i_ino = objectid;
6126
6127         if (dir && name) {
6128                 trace_btrfs_inode_request(dir);
6129
6130                 ret = btrfs_set_inode_index(dir, index);
6131                 if (ret) {
6132                         btrfs_free_path(path);
6133                         iput(inode);
6134                         return ERR_PTR(ret);
6135                 }
6136         } else if (dir) {
6137                 *index = 0;
6138         }
6139         /*
6140          * index_cnt is ignored for everything but a dir,
6141          * btrfs_get_inode_index_count has an explanation for the magic
6142          * number
6143          */
6144         BTRFS_I(inode)->index_cnt = 2;
6145         BTRFS_I(inode)->dir_index = *index;
6146         BTRFS_I(inode)->root = root;
6147         BTRFS_I(inode)->generation = trans->transid;
6148         inode->i_generation = BTRFS_I(inode)->generation;
6149
6150         /*
6151          * We could have gotten an inode number from somebody who was fsynced
6152          * and then removed in this same transaction, so let's just set full
6153          * sync since it will be a full sync anyway and this will blow away the
6154          * old info in the log.
6155          */
6156         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6157
6158         key[0].objectid = objectid;
6159         key[0].type = BTRFS_INODE_ITEM_KEY;
6160         key[0].offset = 0;
6161
6162         sizes[0] = sizeof(struct btrfs_inode_item);
6163
6164         if (name) {
6165                 /*
6166                  * Start new inodes with an inode_ref. This is slightly more
6167                  * efficient for small numbers of hard links since they will
6168                  * be packed into one item. Extended refs will kick in if we
6169                  * add more hard links than can fit in the ref item.
6170                  */
6171                 key[1].objectid = objectid;
6172                 key[1].type = BTRFS_INODE_REF_KEY;
6173                 key[1].offset = ref_objectid;
6174
6175                 sizes[1] = name_len + sizeof(*ref);
6176         }
6177
6178         location = &BTRFS_I(inode)->location;
6179         location->objectid = objectid;
6180         location->offset = 0;
6181         location->type = BTRFS_INODE_ITEM_KEY;
6182
6183         ret = btrfs_insert_inode_locked(inode);
6184         if (ret < 0)
6185                 goto fail;
6186
6187         path->leave_spinning = 1;
6188         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6189         if (ret != 0)
6190                 goto fail_unlock;
6191
6192         inode_init_owner(inode, dir, mode);
6193         inode_set_bytes(inode, 0);
6194
6195         inode->i_mtime = CURRENT_TIME;
6196         inode->i_atime = inode->i_mtime;
6197         inode->i_ctime = inode->i_mtime;
6198         BTRFS_I(inode)->i_otime = inode->i_mtime;
6199
6200         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6201                                   struct btrfs_inode_item);
6202         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6203                              sizeof(*inode_item));
6204         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6205
6206         if (name) {
6207                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6208                                      struct btrfs_inode_ref);
6209                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6210                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6211                 ptr = (unsigned long)(ref + 1);
6212                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6213         }
6214
6215         btrfs_mark_buffer_dirty(path->nodes[0]);
6216         btrfs_free_path(path);
6217
6218         btrfs_inherit_iflags(inode, dir);
6219
6220         if (S_ISREG(mode)) {
6221                 if (btrfs_test_opt(root, NODATASUM))
6222                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6223                 if (btrfs_test_opt(root, NODATACOW))
6224                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6225                                 BTRFS_INODE_NODATASUM;
6226         }
6227
6228         inode_tree_add(inode);
6229
6230         trace_btrfs_inode_new(inode);
6231         btrfs_set_inode_last_trans(trans, inode);
6232
6233         btrfs_update_root_times(trans, root);
6234
6235         ret = btrfs_inode_inherit_props(trans, inode, dir);
6236         if (ret)
6237                 btrfs_err(root->fs_info,
6238                           "error inheriting props for ino %llu (root %llu): %d",
6239                           btrfs_ino(inode), root->root_key.objectid, ret);
6240
6241         return inode;
6242
6243 fail_unlock:
6244         unlock_new_inode(inode);
6245 fail:
6246         if (dir && name)
6247                 BTRFS_I(dir)->index_cnt--;
6248         btrfs_free_path(path);
6249         iput(inode);
6250         return ERR_PTR(ret);
6251 }
6252
6253 static inline u8 btrfs_inode_type(struct inode *inode)
6254 {
6255         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6256 }
6257
6258 /*
6259  * utility function to add 'inode' into 'parent_inode' with
6260  * a give name and a given sequence number.
6261  * if 'add_backref' is true, also insert a backref from the
6262  * inode to the parent directory.
6263  */
6264 int btrfs_add_link(struct btrfs_trans_handle *trans,
6265                    struct inode *parent_inode, struct inode *inode,
6266                    const char *name, int name_len, int add_backref, u64 index)
6267 {
6268         int ret = 0;
6269         struct btrfs_key key;
6270         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6271         u64 ino = btrfs_ino(inode);
6272         u64 parent_ino = btrfs_ino(parent_inode);
6273
6274         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6275                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6276         } else {
6277                 key.objectid = ino;
6278                 key.type = BTRFS_INODE_ITEM_KEY;
6279                 key.offset = 0;
6280         }
6281
6282         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6283                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6284                                          key.objectid, root->root_key.objectid,
6285                                          parent_ino, index, name, name_len);
6286         } else if (add_backref) {
6287                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6288                                              parent_ino, index);
6289         }
6290
6291         /* Nothing to clean up yet */
6292         if (ret)
6293                 return ret;
6294
6295         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6296                                     parent_inode, &key,
6297                                     btrfs_inode_type(inode), index);
6298         if (ret == -EEXIST || ret == -EOVERFLOW)
6299                 goto fail_dir_item;
6300         else if (ret) {
6301                 btrfs_abort_transaction(trans, root, ret);
6302                 return ret;
6303         }
6304
6305         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6306                            name_len * 2);
6307         inode_inc_iversion(parent_inode);
6308         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
6309         ret = btrfs_update_inode(trans, root, parent_inode);
6310         if (ret)
6311                 btrfs_abort_transaction(trans, root, ret);
6312         return ret;
6313
6314 fail_dir_item:
6315         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6316                 u64 local_index;
6317                 int err;
6318                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6319                                  key.objectid, root->root_key.objectid,
6320                                  parent_ino, &local_index, name, name_len);
6321
6322         } else if (add_backref) {
6323                 u64 local_index;
6324                 int err;
6325
6326                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6327                                           ino, parent_ino, &local_index);
6328         }
6329         return ret;
6330 }
6331
6332 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6333                             struct inode *dir, struct dentry *dentry,
6334                             struct inode *inode, int backref, u64 index)
6335 {
6336         int err = btrfs_add_link(trans, dir, inode,
6337                                  dentry->d_name.name, dentry->d_name.len,
6338                                  backref, index);
6339         if (err > 0)
6340                 err = -EEXIST;
6341         return err;
6342 }
6343
6344 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6345                         umode_t mode, dev_t rdev)
6346 {
6347         struct btrfs_trans_handle *trans;
6348         struct btrfs_root *root = BTRFS_I(dir)->root;
6349         struct inode *inode = NULL;
6350         int err;
6351         int drop_inode = 0;
6352         u64 objectid;
6353         u64 index = 0;
6354
6355         /*
6356          * 2 for inode item and ref
6357          * 2 for dir items
6358          * 1 for xattr if selinux is on
6359          */
6360         trans = btrfs_start_transaction(root, 5);
6361         if (IS_ERR(trans))
6362                 return PTR_ERR(trans);
6363
6364         err = btrfs_find_free_ino(root, &objectid);
6365         if (err)
6366                 goto out_unlock;
6367
6368         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6369                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6370                                 mode, &index);
6371         if (IS_ERR(inode)) {
6372                 err = PTR_ERR(inode);
6373                 goto out_unlock;
6374         }
6375
6376         /*
6377         * If the active LSM wants to access the inode during
6378         * d_instantiate it needs these. Smack checks to see
6379         * if the filesystem supports xattrs by looking at the
6380         * ops vector.
6381         */
6382         inode->i_op = &btrfs_special_inode_operations;
6383         init_special_inode(inode, inode->i_mode, rdev);
6384
6385         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6386         if (err)
6387                 goto out_unlock_inode;
6388
6389         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6390         if (err) {
6391                 goto out_unlock_inode;
6392         } else {
6393                 btrfs_update_inode(trans, root, inode);
6394                 unlock_new_inode(inode);
6395                 d_instantiate(dentry, inode);
6396         }
6397
6398 out_unlock:
6399         btrfs_end_transaction(trans, root);
6400         btrfs_balance_delayed_items(root);
6401         btrfs_btree_balance_dirty(root);
6402         if (drop_inode) {
6403                 inode_dec_link_count(inode);
6404                 iput(inode);
6405         }
6406         return err;
6407
6408 out_unlock_inode:
6409         drop_inode = 1;
6410         unlock_new_inode(inode);
6411         goto out_unlock;
6412
6413 }
6414
6415 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6416                         umode_t mode, bool excl)
6417 {
6418         struct btrfs_trans_handle *trans;
6419         struct btrfs_root *root = BTRFS_I(dir)->root;
6420         struct inode *inode = NULL;
6421         int drop_inode_on_err = 0;
6422         int err;
6423         u64 objectid;
6424         u64 index = 0;
6425
6426         /*
6427          * 2 for inode item and ref
6428          * 2 for dir items
6429          * 1 for xattr if selinux is on
6430          */
6431         trans = btrfs_start_transaction(root, 5);
6432         if (IS_ERR(trans))
6433                 return PTR_ERR(trans);
6434
6435         err = btrfs_find_free_ino(root, &objectid);
6436         if (err)
6437                 goto out_unlock;
6438
6439         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6440                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6441                                 mode, &index);
6442         if (IS_ERR(inode)) {
6443                 err = PTR_ERR(inode);
6444                 goto out_unlock;
6445         }
6446         drop_inode_on_err = 1;
6447         /*
6448         * If the active LSM wants to access the inode during
6449         * d_instantiate it needs these. Smack checks to see
6450         * if the filesystem supports xattrs by looking at the
6451         * ops vector.
6452         */
6453         inode->i_fop = &btrfs_file_operations;
6454         inode->i_op = &btrfs_file_inode_operations;
6455         inode->i_mapping->a_ops = &btrfs_aops;
6456
6457         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6458         if (err)
6459                 goto out_unlock_inode;
6460
6461         err = btrfs_update_inode(trans, root, inode);
6462         if (err)
6463                 goto out_unlock_inode;
6464
6465         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6466         if (err)
6467                 goto out_unlock_inode;
6468
6469         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6470         unlock_new_inode(inode);
6471         d_instantiate(dentry, inode);
6472
6473 out_unlock:
6474         btrfs_end_transaction(trans, root);
6475         if (err && drop_inode_on_err) {
6476                 inode_dec_link_count(inode);
6477                 iput(inode);
6478         }
6479         btrfs_balance_delayed_items(root);
6480         btrfs_btree_balance_dirty(root);
6481         return err;
6482
6483 out_unlock_inode:
6484         unlock_new_inode(inode);
6485         goto out_unlock;
6486
6487 }
6488
6489 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6490                       struct dentry *dentry)
6491 {
6492         struct btrfs_trans_handle *trans = NULL;
6493         struct btrfs_root *root = BTRFS_I(dir)->root;
6494         struct inode *inode = d_inode(old_dentry);
6495         u64 index;
6496         int err;
6497         int drop_inode = 0;
6498
6499         /* do not allow sys_link's with other subvols of the same device */
6500         if (root->objectid != BTRFS_I(inode)->root->objectid)
6501                 return -EXDEV;
6502
6503         if (inode->i_nlink >= BTRFS_LINK_MAX)
6504                 return -EMLINK;
6505
6506         err = btrfs_set_inode_index(dir, &index);
6507         if (err)
6508                 goto fail;
6509
6510         /*
6511          * 2 items for inode and inode ref
6512          * 2 items for dir items
6513          * 1 item for parent inode
6514          */
6515         trans = btrfs_start_transaction(root, 5);
6516         if (IS_ERR(trans)) {
6517                 err = PTR_ERR(trans);
6518                 trans = NULL;
6519                 goto fail;
6520         }
6521
6522         /* There are several dir indexes for this inode, clear the cache. */
6523         BTRFS_I(inode)->dir_index = 0ULL;
6524         inc_nlink(inode);
6525         inode_inc_iversion(inode);
6526         inode->i_ctime = CURRENT_TIME;
6527         ihold(inode);
6528         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6529
6530         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6531
6532         if (err) {
6533                 drop_inode = 1;
6534         } else {
6535                 struct dentry *parent = dentry->d_parent;
6536                 err = btrfs_update_inode(trans, root, inode);
6537                 if (err)
6538                         goto fail;
6539                 if (inode->i_nlink == 1) {
6540                         /*
6541                          * If new hard link count is 1, it's a file created
6542                          * with open(2) O_TMPFILE flag.
6543                          */
6544                         err = btrfs_orphan_del(trans, inode);
6545                         if (err)
6546                                 goto fail;
6547                 }
6548                 d_instantiate(dentry, inode);
6549                 btrfs_log_new_name(trans, inode, NULL, parent);
6550         }
6551
6552         btrfs_balance_delayed_items(root);
6553 fail:
6554         if (trans)
6555                 btrfs_end_transaction(trans, root);
6556         if (drop_inode) {
6557                 inode_dec_link_count(inode);
6558                 iput(inode);
6559         }
6560         btrfs_btree_balance_dirty(root);
6561         return err;
6562 }
6563
6564 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6565 {
6566         struct inode *inode = NULL;
6567         struct btrfs_trans_handle *trans;
6568         struct btrfs_root *root = BTRFS_I(dir)->root;
6569         int err = 0;
6570         int drop_on_err = 0;
6571         u64 objectid = 0;
6572         u64 index = 0;
6573
6574         /*
6575          * 2 items for inode and ref
6576          * 2 items for dir items
6577          * 1 for xattr if selinux is on
6578          */
6579         trans = btrfs_start_transaction(root, 5);
6580         if (IS_ERR(trans))
6581                 return PTR_ERR(trans);
6582
6583         err = btrfs_find_free_ino(root, &objectid);
6584         if (err)
6585                 goto out_fail;
6586
6587         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6588                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6589                                 S_IFDIR | mode, &index);
6590         if (IS_ERR(inode)) {
6591                 err = PTR_ERR(inode);
6592                 goto out_fail;
6593         }
6594
6595         drop_on_err = 1;
6596         /* these must be set before we unlock the inode */
6597         inode->i_op = &btrfs_dir_inode_operations;
6598         inode->i_fop = &btrfs_dir_file_operations;
6599
6600         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6601         if (err)
6602                 goto out_fail_inode;
6603
6604         btrfs_i_size_write(inode, 0);
6605         err = btrfs_update_inode(trans, root, inode);
6606         if (err)
6607                 goto out_fail_inode;
6608
6609         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6610                              dentry->d_name.len, 0, index);
6611         if (err)
6612                 goto out_fail_inode;
6613
6614         d_instantiate(dentry, inode);
6615         /*
6616          * mkdir is special.  We're unlocking after we call d_instantiate
6617          * to avoid a race with nfsd calling d_instantiate.
6618          */
6619         unlock_new_inode(inode);
6620         drop_on_err = 0;
6621
6622 out_fail:
6623         btrfs_end_transaction(trans, root);
6624         if (drop_on_err) {
6625                 inode_dec_link_count(inode);
6626                 iput(inode);
6627         }
6628         btrfs_balance_delayed_items(root);
6629         btrfs_btree_balance_dirty(root);
6630         return err;
6631
6632 out_fail_inode:
6633         unlock_new_inode(inode);
6634         goto out_fail;
6635 }
6636
6637 /* Find next extent map of a given extent map, caller needs to ensure locks */
6638 static struct extent_map *next_extent_map(struct extent_map *em)
6639 {
6640         struct rb_node *next;
6641
6642         next = rb_next(&em->rb_node);
6643         if (!next)
6644                 return NULL;
6645         return container_of(next, struct extent_map, rb_node);
6646 }
6647
6648 static struct extent_map *prev_extent_map(struct extent_map *em)
6649 {
6650         struct rb_node *prev;
6651
6652         prev = rb_prev(&em->rb_node);
6653         if (!prev)
6654                 return NULL;
6655         return container_of(prev, struct extent_map, rb_node);
6656 }
6657
6658 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6659  * the existing extent is the nearest extent to map_start,
6660  * and an extent that you want to insert, deal with overlap and insert
6661  * the best fitted new extent into the tree.
6662  */
6663 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6664                                 struct extent_map *existing,
6665                                 struct extent_map *em,
6666                                 u64 map_start)
6667 {
6668         struct extent_map *prev;
6669         struct extent_map *next;
6670         u64 start;
6671         u64 end;
6672         u64 start_diff;
6673
6674         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6675
6676         if (existing->start > map_start) {
6677                 next = existing;
6678                 prev = prev_extent_map(next);
6679         } else {
6680                 prev = existing;
6681                 next = next_extent_map(prev);
6682         }
6683
6684         start = prev ? extent_map_end(prev) : em->start;
6685         start = max_t(u64, start, em->start);
6686         end = next ? next->start : extent_map_end(em);
6687         end = min_t(u64, end, extent_map_end(em));
6688         start_diff = start - em->start;
6689         em->start = start;
6690         em->len = end - start;
6691         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6692             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6693                 em->block_start += start_diff;
6694                 em->block_len -= start_diff;
6695         }
6696         return add_extent_mapping(em_tree, em, 0);
6697 }
6698
6699 static noinline int uncompress_inline(struct btrfs_path *path,
6700                                       struct inode *inode, struct page *page,
6701                                       size_t pg_offset, u64 extent_offset,
6702                                       struct btrfs_file_extent_item *item)
6703 {
6704         int ret;
6705         struct extent_buffer *leaf = path->nodes[0];
6706         char *tmp;
6707         size_t max_size;
6708         unsigned long inline_size;
6709         unsigned long ptr;
6710         int compress_type;
6711
6712         WARN_ON(pg_offset != 0);
6713         compress_type = btrfs_file_extent_compression(leaf, item);
6714         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6715         inline_size = btrfs_file_extent_inline_item_len(leaf,
6716                                         btrfs_item_nr(path->slots[0]));
6717         tmp = kmalloc(inline_size, GFP_NOFS);
6718         if (!tmp)
6719                 return -ENOMEM;
6720         ptr = btrfs_file_extent_inline_start(item);
6721
6722         read_extent_buffer(leaf, tmp, ptr, inline_size);
6723
6724         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
6725         ret = btrfs_decompress(compress_type, tmp, page,
6726                                extent_offset, inline_size, max_size);
6727         kfree(tmp);
6728         return ret;
6729 }
6730
6731 /*
6732  * a bit scary, this does extent mapping from logical file offset to the disk.
6733  * the ugly parts come from merging extents from the disk with the in-ram
6734  * representation.  This gets more complex because of the data=ordered code,
6735  * where the in-ram extents might be locked pending data=ordered completion.
6736  *
6737  * This also copies inline extents directly into the page.
6738  */
6739
6740 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6741                                     size_t pg_offset, u64 start, u64 len,
6742                                     int create)
6743 {
6744         int ret;
6745         int err = 0;
6746         u64 extent_start = 0;
6747         u64 extent_end = 0;
6748         u64 objectid = btrfs_ino(inode);
6749         u32 found_type;
6750         struct btrfs_path *path = NULL;
6751         struct btrfs_root *root = BTRFS_I(inode)->root;
6752         struct btrfs_file_extent_item *item;
6753         struct extent_buffer *leaf;
6754         struct btrfs_key found_key;
6755         struct extent_map *em = NULL;
6756         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6757         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6758         struct btrfs_trans_handle *trans = NULL;
6759         const bool new_inline = !page || create;
6760
6761 again:
6762         read_lock(&em_tree->lock);
6763         em = lookup_extent_mapping(em_tree, start, len);
6764         if (em)
6765                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6766         read_unlock(&em_tree->lock);
6767
6768         if (em) {
6769                 if (em->start > start || em->start + em->len <= start)
6770                         free_extent_map(em);
6771                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6772                         free_extent_map(em);
6773                 else
6774                         goto out;
6775         }
6776         em = alloc_extent_map();
6777         if (!em) {
6778                 err = -ENOMEM;
6779                 goto out;
6780         }
6781         em->bdev = root->fs_info->fs_devices->latest_bdev;
6782         em->start = EXTENT_MAP_HOLE;
6783         em->orig_start = EXTENT_MAP_HOLE;
6784         em->len = (u64)-1;
6785         em->block_len = (u64)-1;
6786
6787         if (!path) {
6788                 path = btrfs_alloc_path();
6789                 if (!path) {
6790                         err = -ENOMEM;
6791                         goto out;
6792                 }
6793                 /*
6794                  * Chances are we'll be called again, so go ahead and do
6795                  * readahead
6796                  */
6797                 path->reada = 1;
6798         }
6799
6800         ret = btrfs_lookup_file_extent(trans, root, path,
6801                                        objectid, start, trans != NULL);
6802         if (ret < 0) {
6803                 err = ret;
6804                 goto out;
6805         }
6806
6807         if (ret != 0) {
6808                 if (path->slots[0] == 0)
6809                         goto not_found;
6810                 path->slots[0]--;
6811         }
6812
6813         leaf = path->nodes[0];
6814         item = btrfs_item_ptr(leaf, path->slots[0],
6815                               struct btrfs_file_extent_item);
6816         /* are we inside the extent that was found? */
6817         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6818         found_type = found_key.type;
6819         if (found_key.objectid != objectid ||
6820             found_type != BTRFS_EXTENT_DATA_KEY) {
6821                 /*
6822                  * If we backup past the first extent we want to move forward
6823                  * and see if there is an extent in front of us, otherwise we'll
6824                  * say there is a hole for our whole search range which can
6825                  * cause problems.
6826                  */
6827                 extent_end = start;
6828                 goto next;
6829         }
6830
6831         found_type = btrfs_file_extent_type(leaf, item);
6832         extent_start = found_key.offset;
6833         if (found_type == BTRFS_FILE_EXTENT_REG ||
6834             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6835                 extent_end = extent_start +
6836                        btrfs_file_extent_num_bytes(leaf, item);
6837         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6838                 size_t size;
6839                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6840                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6841         }
6842 next:
6843         if (start >= extent_end) {
6844                 path->slots[0]++;
6845                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6846                         ret = btrfs_next_leaf(root, path);
6847                         if (ret < 0) {
6848                                 err = ret;
6849                                 goto out;
6850                         }
6851                         if (ret > 0)
6852                                 goto not_found;
6853                         leaf = path->nodes[0];
6854                 }
6855                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6856                 if (found_key.objectid != objectid ||
6857                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6858                         goto not_found;
6859                 if (start + len <= found_key.offset)
6860                         goto not_found;
6861                 if (start > found_key.offset)
6862                         goto next;
6863                 em->start = start;
6864                 em->orig_start = start;
6865                 em->len = found_key.offset - start;
6866                 goto not_found_em;
6867         }
6868
6869         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6870
6871         if (found_type == BTRFS_FILE_EXTENT_REG ||
6872             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6873                 goto insert;
6874         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6875                 unsigned long ptr;
6876                 char *map;
6877                 size_t size;
6878                 size_t extent_offset;
6879                 size_t copy_size;
6880
6881                 if (new_inline)
6882                         goto out;
6883
6884                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6885                 extent_offset = page_offset(page) + pg_offset - extent_start;
6886                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6887                                 size - extent_offset);
6888                 em->start = extent_start + extent_offset;
6889                 em->len = ALIGN(copy_size, root->sectorsize);
6890                 em->orig_block_len = em->len;
6891                 em->orig_start = em->start;
6892                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6893                 if (create == 0 && !PageUptodate(page)) {
6894                         if (btrfs_file_extent_compression(leaf, item) !=
6895                             BTRFS_COMPRESS_NONE) {
6896                                 ret = uncompress_inline(path, inode, page,
6897                                                         pg_offset,
6898                                                         extent_offset, item);
6899                                 if (ret) {
6900                                         err = ret;
6901                                         goto out;
6902                                 }
6903                         } else {
6904                                 map = kmap(page);
6905                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6906                                                    copy_size);
6907                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6908                                         memset(map + pg_offset + copy_size, 0,
6909                                                PAGE_CACHE_SIZE - pg_offset -
6910                                                copy_size);
6911                                 }
6912                                 kunmap(page);
6913                         }
6914                         flush_dcache_page(page);
6915                 } else if (create && PageUptodate(page)) {
6916                         BUG();
6917                         if (!trans) {
6918                                 kunmap(page);
6919                                 free_extent_map(em);
6920                                 em = NULL;
6921
6922                                 btrfs_release_path(path);
6923                                 trans = btrfs_join_transaction(root);
6924
6925                                 if (IS_ERR(trans))
6926                                         return ERR_CAST(trans);
6927                                 goto again;
6928                         }
6929                         map = kmap(page);
6930                         write_extent_buffer(leaf, map + pg_offset, ptr,
6931                                             copy_size);
6932                         kunmap(page);
6933                         btrfs_mark_buffer_dirty(leaf);
6934                 }
6935                 set_extent_uptodate(io_tree, em->start,
6936                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6937                 goto insert;
6938         }
6939 not_found:
6940         em->start = start;
6941         em->orig_start = start;
6942         em->len = len;
6943 not_found_em:
6944         em->block_start = EXTENT_MAP_HOLE;
6945         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6946 insert:
6947         btrfs_release_path(path);
6948         if (em->start > start || extent_map_end(em) <= start) {
6949                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6950                         em->start, em->len, start, len);
6951                 err = -EIO;
6952                 goto out;
6953         }
6954
6955         err = 0;
6956         write_lock(&em_tree->lock);
6957         ret = add_extent_mapping(em_tree, em, 0);
6958         /* it is possible that someone inserted the extent into the tree
6959          * while we had the lock dropped.  It is also possible that
6960          * an overlapping map exists in the tree
6961          */
6962         if (ret == -EEXIST) {
6963                 struct extent_map *existing;
6964
6965                 ret = 0;
6966
6967                 existing = search_extent_mapping(em_tree, start, len);
6968                 /*
6969                  * existing will always be non-NULL, since there must be
6970                  * extent causing the -EEXIST.
6971                  */
6972                 if (start >= extent_map_end(existing) ||
6973                     start <= existing->start) {
6974                         /*
6975                          * The existing extent map is the one nearest to
6976                          * the [start, start + len) range which overlaps
6977                          */
6978                         err = merge_extent_mapping(em_tree, existing,
6979                                                    em, start);
6980                         free_extent_map(existing);
6981                         if (err) {
6982                                 free_extent_map(em);
6983                                 em = NULL;
6984                         }
6985                 } else {
6986                         free_extent_map(em);
6987                         em = existing;
6988                         err = 0;
6989                 }
6990         }
6991         write_unlock(&em_tree->lock);
6992 out:
6993
6994         trace_btrfs_get_extent(root, em);
6995
6996         btrfs_free_path(path);
6997         if (trans) {
6998                 ret = btrfs_end_transaction(trans, root);
6999                 if (!err)
7000                         err = ret;
7001         }
7002         if (err) {
7003                 free_extent_map(em);
7004                 return ERR_PTR(err);
7005         }
7006         BUG_ON(!em); /* Error is always set */
7007         return em;
7008 }
7009
7010 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7011                                            size_t pg_offset, u64 start, u64 len,
7012                                            int create)
7013 {
7014         struct extent_map *em;
7015         struct extent_map *hole_em = NULL;
7016         u64 range_start = start;
7017         u64 end;
7018         u64 found;
7019         u64 found_end;
7020         int err = 0;
7021
7022         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7023         if (IS_ERR(em))
7024                 return em;
7025         if (em) {
7026                 /*
7027                  * if our em maps to
7028                  * -  a hole or
7029                  * -  a pre-alloc extent,
7030                  * there might actually be delalloc bytes behind it.
7031                  */
7032                 if (em->block_start != EXTENT_MAP_HOLE &&
7033                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7034                         return em;
7035                 else
7036                         hole_em = em;
7037         }
7038
7039         /* check to see if we've wrapped (len == -1 or similar) */
7040         end = start + len;
7041         if (end < start)
7042                 end = (u64)-1;
7043         else
7044                 end -= 1;
7045
7046         em = NULL;
7047
7048         /* ok, we didn't find anything, lets look for delalloc */
7049         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7050                                  end, len, EXTENT_DELALLOC, 1);
7051         found_end = range_start + found;
7052         if (found_end < range_start)
7053                 found_end = (u64)-1;
7054
7055         /*
7056          * we didn't find anything useful, return
7057          * the original results from get_extent()
7058          */
7059         if (range_start > end || found_end <= start) {
7060                 em = hole_em;
7061                 hole_em = NULL;
7062                 goto out;
7063         }
7064
7065         /* adjust the range_start to make sure it doesn't
7066          * go backwards from the start they passed in
7067          */
7068         range_start = max(start, range_start);
7069         found = found_end - range_start;
7070
7071         if (found > 0) {
7072                 u64 hole_start = start;
7073                 u64 hole_len = len;
7074
7075                 em = alloc_extent_map();
7076                 if (!em) {
7077                         err = -ENOMEM;
7078                         goto out;
7079                 }
7080                 /*
7081                  * when btrfs_get_extent can't find anything it
7082                  * returns one huge hole
7083                  *
7084                  * make sure what it found really fits our range, and
7085                  * adjust to make sure it is based on the start from
7086                  * the caller
7087                  */
7088                 if (hole_em) {
7089                         u64 calc_end = extent_map_end(hole_em);
7090
7091                         if (calc_end <= start || (hole_em->start > end)) {
7092                                 free_extent_map(hole_em);
7093                                 hole_em = NULL;
7094                         } else {
7095                                 hole_start = max(hole_em->start, start);
7096                                 hole_len = calc_end - hole_start;
7097                         }
7098                 }
7099                 em->bdev = NULL;
7100                 if (hole_em && range_start > hole_start) {
7101                         /* our hole starts before our delalloc, so we
7102                          * have to return just the parts of the hole
7103                          * that go until  the delalloc starts
7104                          */
7105                         em->len = min(hole_len,
7106                                       range_start - hole_start);
7107                         em->start = hole_start;
7108                         em->orig_start = hole_start;
7109                         /*
7110                          * don't adjust block start at all,
7111                          * it is fixed at EXTENT_MAP_HOLE
7112                          */
7113                         em->block_start = hole_em->block_start;
7114                         em->block_len = hole_len;
7115                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7116                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7117                 } else {
7118                         em->start = range_start;
7119                         em->len = found;
7120                         em->orig_start = range_start;
7121                         em->block_start = EXTENT_MAP_DELALLOC;
7122                         em->block_len = found;
7123                 }
7124         } else if (hole_em) {
7125                 return hole_em;
7126         }
7127 out:
7128
7129         free_extent_map(hole_em);
7130         if (err) {
7131                 free_extent_map(em);
7132                 return ERR_PTR(err);
7133         }
7134         return em;
7135 }
7136
7137 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7138                                                   u64 start, u64 len)
7139 {
7140         struct btrfs_root *root = BTRFS_I(inode)->root;
7141         struct extent_map *em;
7142         struct btrfs_key ins;
7143         u64 alloc_hint;
7144         int ret;
7145
7146         alloc_hint = get_extent_allocation_hint(inode, start, len);
7147         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7148                                    alloc_hint, &ins, 1, 1);
7149         if (ret)
7150                 return ERR_PTR(ret);
7151
7152         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
7153                               ins.offset, ins.offset, ins.offset, 0);
7154         if (IS_ERR(em)) {
7155                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7156                 return em;
7157         }
7158
7159         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
7160                                            ins.offset, ins.offset, 0);
7161         if (ret) {
7162                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7163                 free_extent_map(em);
7164                 return ERR_PTR(ret);
7165         }
7166
7167         return em;
7168 }
7169
7170 /*
7171  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7172  * block must be cow'd
7173  */
7174 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7175                               u64 *orig_start, u64 *orig_block_len,
7176                               u64 *ram_bytes)
7177 {
7178         struct btrfs_trans_handle *trans;
7179         struct btrfs_path *path;
7180         int ret;
7181         struct extent_buffer *leaf;
7182         struct btrfs_root *root = BTRFS_I(inode)->root;
7183         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7184         struct btrfs_file_extent_item *fi;
7185         struct btrfs_key key;
7186         u64 disk_bytenr;
7187         u64 backref_offset;
7188         u64 extent_end;
7189         u64 num_bytes;
7190         int slot;
7191         int found_type;
7192         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7193
7194         path = btrfs_alloc_path();
7195         if (!path)
7196                 return -ENOMEM;
7197
7198         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7199                                        offset, 0);
7200         if (ret < 0)
7201                 goto out;
7202
7203         slot = path->slots[0];
7204         if (ret == 1) {
7205                 if (slot == 0) {
7206                         /* can't find the item, must cow */
7207                         ret = 0;
7208                         goto out;
7209                 }
7210                 slot--;
7211         }
7212         ret = 0;
7213         leaf = path->nodes[0];
7214         btrfs_item_key_to_cpu(leaf, &key, slot);
7215         if (key.objectid != btrfs_ino(inode) ||
7216             key.type != BTRFS_EXTENT_DATA_KEY) {
7217                 /* not our file or wrong item type, must cow */
7218                 goto out;
7219         }
7220
7221         if (key.offset > offset) {
7222                 /* Wrong offset, must cow */
7223                 goto out;
7224         }
7225
7226         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7227         found_type = btrfs_file_extent_type(leaf, fi);
7228         if (found_type != BTRFS_FILE_EXTENT_REG &&
7229             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7230                 /* not a regular extent, must cow */
7231                 goto out;
7232         }
7233
7234         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7235                 goto out;
7236
7237         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7238         if (extent_end <= offset)
7239                 goto out;
7240
7241         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7242         if (disk_bytenr == 0)
7243                 goto out;
7244
7245         if (btrfs_file_extent_compression(leaf, fi) ||
7246             btrfs_file_extent_encryption(leaf, fi) ||
7247             btrfs_file_extent_other_encoding(leaf, fi))
7248                 goto out;
7249
7250         backref_offset = btrfs_file_extent_offset(leaf, fi);
7251
7252         if (orig_start) {
7253                 *orig_start = key.offset - backref_offset;
7254                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7255                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7256         }
7257
7258         if (btrfs_extent_readonly(root, disk_bytenr))
7259                 goto out;
7260
7261         num_bytes = min(offset + *len, extent_end) - offset;
7262         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7263                 u64 range_end;
7264
7265                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7266                 ret = test_range_bit(io_tree, offset, range_end,
7267                                      EXTENT_DELALLOC, 0, NULL);
7268                 if (ret) {
7269                         ret = -EAGAIN;
7270                         goto out;
7271                 }
7272         }
7273
7274         btrfs_release_path(path);
7275
7276         /*
7277          * look for other files referencing this extent, if we
7278          * find any we must cow
7279          */
7280         trans = btrfs_join_transaction(root);
7281         if (IS_ERR(trans)) {
7282                 ret = 0;
7283                 goto out;
7284         }
7285
7286         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7287                                     key.offset - backref_offset, disk_bytenr);
7288         btrfs_end_transaction(trans, root);
7289         if (ret) {
7290                 ret = 0;
7291                 goto out;
7292         }
7293
7294         /*
7295          * adjust disk_bytenr and num_bytes to cover just the bytes
7296          * in this extent we are about to write.  If there
7297          * are any csums in that range we have to cow in order
7298          * to keep the csums correct
7299          */
7300         disk_bytenr += backref_offset;
7301         disk_bytenr += offset - key.offset;
7302         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7303                                 goto out;
7304         /*
7305          * all of the above have passed, it is safe to overwrite this extent
7306          * without cow
7307          */
7308         *len = num_bytes;
7309         ret = 1;
7310 out:
7311         btrfs_free_path(path);
7312         return ret;
7313 }
7314
7315 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7316 {
7317         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7318         int found = false;
7319         void **pagep = NULL;
7320         struct page *page = NULL;
7321         int start_idx;
7322         int end_idx;
7323
7324         start_idx = start >> PAGE_CACHE_SHIFT;
7325
7326         /*
7327          * end is the last byte in the last page.  end == start is legal
7328          */
7329         end_idx = end >> PAGE_CACHE_SHIFT;
7330
7331         rcu_read_lock();
7332
7333         /* Most of the code in this while loop is lifted from
7334          * find_get_page.  It's been modified to begin searching from a
7335          * page and return just the first page found in that range.  If the
7336          * found idx is less than or equal to the end idx then we know that
7337          * a page exists.  If no pages are found or if those pages are
7338          * outside of the range then we're fine (yay!) */
7339         while (page == NULL &&
7340                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7341                 page = radix_tree_deref_slot(pagep);
7342                 if (unlikely(!page))
7343                         break;
7344
7345                 if (radix_tree_exception(page)) {
7346                         if (radix_tree_deref_retry(page)) {
7347                                 page = NULL;
7348                                 continue;
7349                         }
7350                         /*
7351                          * Otherwise, shmem/tmpfs must be storing a swap entry
7352                          * here as an exceptional entry: so return it without
7353                          * attempting to raise page count.
7354                          */
7355                         page = NULL;
7356                         break; /* TODO: Is this relevant for this use case? */
7357                 }
7358
7359                 if (!page_cache_get_speculative(page)) {
7360                         page = NULL;
7361                         continue;
7362                 }
7363
7364                 /*
7365                  * Has the page moved?
7366                  * This is part of the lockless pagecache protocol. See
7367                  * include/linux/pagemap.h for details.
7368                  */
7369                 if (unlikely(page != *pagep)) {
7370                         page_cache_release(page);
7371                         page = NULL;
7372                 }
7373         }
7374
7375         if (page) {
7376                 if (page->index <= end_idx)
7377                         found = true;
7378                 page_cache_release(page);
7379         }
7380
7381         rcu_read_unlock();
7382         return found;
7383 }
7384
7385 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7386                               struct extent_state **cached_state, int writing)
7387 {
7388         struct btrfs_ordered_extent *ordered;
7389         int ret = 0;
7390
7391         while (1) {
7392                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7393                                  0, cached_state);
7394                 /*
7395                  * We're concerned with the entire range that we're going to be
7396                  * doing DIO to, so we need to make sure theres no ordered
7397                  * extents in this range.
7398                  */
7399                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7400                                                      lockend - lockstart + 1);
7401
7402                 /*
7403                  * We need to make sure there are no buffered pages in this
7404                  * range either, we could have raced between the invalidate in
7405                  * generic_file_direct_write and locking the extent.  The
7406                  * invalidate needs to happen so that reads after a write do not
7407                  * get stale data.
7408                  */
7409                 if (!ordered &&
7410                     (!writing ||
7411                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7412                         break;
7413
7414                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7415                                      cached_state, GFP_NOFS);
7416
7417                 if (ordered) {
7418                         btrfs_start_ordered_extent(inode, ordered, 1);
7419                         btrfs_put_ordered_extent(ordered);
7420                 } else {
7421                         /* Screw you mmap */
7422                         ret = btrfs_fdatawrite_range(inode, lockstart, lockend);
7423                         if (ret)
7424                                 break;
7425                         ret = filemap_fdatawait_range(inode->i_mapping,
7426                                                       lockstart,
7427                                                       lockend);
7428                         if (ret)
7429                                 break;
7430
7431                         /*
7432                          * If we found a page that couldn't be invalidated just
7433                          * fall back to buffered.
7434                          */
7435                         ret = invalidate_inode_pages2_range(inode->i_mapping,
7436                                         lockstart >> PAGE_CACHE_SHIFT,
7437                                         lockend >> PAGE_CACHE_SHIFT);
7438                         if (ret)
7439                                 break;
7440                 }
7441
7442                 cond_resched();
7443         }
7444
7445         return ret;
7446 }
7447
7448 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7449                                            u64 len, u64 orig_start,
7450                                            u64 block_start, u64 block_len,
7451                                            u64 orig_block_len, u64 ram_bytes,
7452                                            int type)
7453 {
7454         struct extent_map_tree *em_tree;
7455         struct extent_map *em;
7456         struct btrfs_root *root = BTRFS_I(inode)->root;
7457         int ret;
7458
7459         em_tree = &BTRFS_I(inode)->extent_tree;
7460         em = alloc_extent_map();
7461         if (!em)
7462                 return ERR_PTR(-ENOMEM);
7463
7464         em->start = start;
7465         em->orig_start = orig_start;
7466         em->mod_start = start;
7467         em->mod_len = len;
7468         em->len = len;
7469         em->block_len = block_len;
7470         em->block_start = block_start;
7471         em->bdev = root->fs_info->fs_devices->latest_bdev;
7472         em->orig_block_len = orig_block_len;
7473         em->ram_bytes = ram_bytes;
7474         em->generation = -1;
7475         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7476         if (type == BTRFS_ORDERED_PREALLOC)
7477                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7478
7479         do {
7480                 btrfs_drop_extent_cache(inode, em->start,
7481                                 em->start + em->len - 1, 0);
7482                 write_lock(&em_tree->lock);
7483                 ret = add_extent_mapping(em_tree, em, 1);
7484                 write_unlock(&em_tree->lock);
7485         } while (ret == -EEXIST);
7486
7487         if (ret) {
7488                 free_extent_map(em);
7489                 return ERR_PTR(ret);
7490         }
7491
7492         return em;
7493 }
7494
7495 struct btrfs_dio_data {
7496         u64 outstanding_extents;
7497         u64 reserve;
7498 };
7499
7500 static void adjust_dio_outstanding_extents(struct inode *inode,
7501                                            struct btrfs_dio_data *dio_data,
7502                                            const u64 len)
7503 {
7504         unsigned num_extents;
7505
7506         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7507                                            BTRFS_MAX_EXTENT_SIZE);
7508         /*
7509          * If we have an outstanding_extents count still set then we're
7510          * within our reservation, otherwise we need to adjust our inode
7511          * counter appropriately.
7512          */
7513         if (dio_data->outstanding_extents) {
7514                 dio_data->outstanding_extents -= num_extents;
7515         } else {
7516                 spin_lock(&BTRFS_I(inode)->lock);
7517                 BTRFS_I(inode)->outstanding_extents += num_extents;
7518                 spin_unlock(&BTRFS_I(inode)->lock);
7519         }
7520 }
7521
7522 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7523                                    struct buffer_head *bh_result, int create)
7524 {
7525         struct extent_map *em;
7526         struct btrfs_root *root = BTRFS_I(inode)->root;
7527         struct extent_state *cached_state = NULL;
7528         struct btrfs_dio_data *dio_data = NULL;
7529         u64 start = iblock << inode->i_blkbits;
7530         u64 lockstart, lockend;
7531         u64 len = bh_result->b_size;
7532         int unlock_bits = EXTENT_LOCKED;
7533         int ret = 0;
7534
7535         if (create)
7536                 unlock_bits |= EXTENT_DIRTY;
7537         else
7538                 len = min_t(u64, len, root->sectorsize);
7539
7540         lockstart = start;
7541         lockend = start + len - 1;
7542
7543         if (current->journal_info) {
7544                 /*
7545                  * Need to pull our outstanding extents and set journal_info to NULL so
7546                  * that anything that needs to check if there's a transction doesn't get
7547                  * confused.
7548                  */
7549                 dio_data = current->journal_info;
7550                 current->journal_info = NULL;
7551         }
7552
7553         /*
7554          * If this errors out it's because we couldn't invalidate pagecache for
7555          * this range and we need to fallback to buffered.
7556          */
7557         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7558                                create)) {
7559                 ret = -ENOTBLK;
7560                 goto err;
7561         }
7562
7563         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7564         if (IS_ERR(em)) {
7565                 ret = PTR_ERR(em);
7566                 goto unlock_err;
7567         }
7568
7569         /*
7570          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7571          * io.  INLINE is special, and we could probably kludge it in here, but
7572          * it's still buffered so for safety lets just fall back to the generic
7573          * buffered path.
7574          *
7575          * For COMPRESSED we _have_ to read the entire extent in so we can
7576          * decompress it, so there will be buffering required no matter what we
7577          * do, so go ahead and fallback to buffered.
7578          *
7579          * We return -ENOTBLK because thats what makes DIO go ahead and go back
7580          * to buffered IO.  Don't blame me, this is the price we pay for using
7581          * the generic code.
7582          */
7583         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7584             em->block_start == EXTENT_MAP_INLINE) {
7585                 free_extent_map(em);
7586                 ret = -ENOTBLK;
7587                 goto unlock_err;
7588         }
7589
7590         /* Just a good old fashioned hole, return */
7591         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7592                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7593                 free_extent_map(em);
7594                 goto unlock_err;
7595         }
7596
7597         /*
7598          * We don't allocate a new extent in the following cases
7599          *
7600          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7601          * existing extent.
7602          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7603          * just use the extent.
7604          *
7605          */
7606         if (!create) {
7607                 len = min(len, em->len - (start - em->start));
7608                 lockstart = start + len;
7609                 goto unlock;
7610         }
7611
7612         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7613             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7614              em->block_start != EXTENT_MAP_HOLE)) {
7615                 int type;
7616                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7617
7618                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7619                         type = BTRFS_ORDERED_PREALLOC;
7620                 else
7621                         type = BTRFS_ORDERED_NOCOW;
7622                 len = min(len, em->len - (start - em->start));
7623                 block_start = em->block_start + (start - em->start);
7624
7625                 if (can_nocow_extent(inode, start, &len, &orig_start,
7626                                      &orig_block_len, &ram_bytes) == 1) {
7627                         if (type == BTRFS_ORDERED_PREALLOC) {
7628                                 free_extent_map(em);
7629                                 em = create_pinned_em(inode, start, len,
7630                                                        orig_start,
7631                                                        block_start, len,
7632                                                        orig_block_len,
7633                                                        ram_bytes, type);
7634                                 if (IS_ERR(em)) {
7635                                         ret = PTR_ERR(em);
7636                                         goto unlock_err;
7637                                 }
7638                         }
7639
7640                         ret = btrfs_add_ordered_extent_dio(inode, start,
7641                                            block_start, len, len, type);
7642                         if (ret) {
7643                                 free_extent_map(em);
7644                                 goto unlock_err;
7645                         }
7646                         goto unlock;
7647                 }
7648         }
7649
7650         /*
7651          * this will cow the extent, reset the len in case we changed
7652          * it above
7653          */
7654         len = bh_result->b_size;
7655         free_extent_map(em);
7656         em = btrfs_new_extent_direct(inode, start, len);
7657         if (IS_ERR(em)) {
7658                 ret = PTR_ERR(em);
7659                 goto unlock_err;
7660         }
7661         len = min(len, em->len - (start - em->start));
7662 unlock:
7663         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7664                 inode->i_blkbits;
7665         bh_result->b_size = len;
7666         bh_result->b_bdev = em->bdev;
7667         set_buffer_mapped(bh_result);
7668         if (create) {
7669                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7670                         set_buffer_new(bh_result);
7671
7672                 /*
7673                  * Need to update the i_size under the extent lock so buffered
7674                  * readers will get the updated i_size when we unlock.
7675                  */
7676                 if (start + len > i_size_read(inode))
7677                         i_size_write(inode, start + len);
7678
7679                 adjust_dio_outstanding_extents(inode, dio_data, len);
7680                 btrfs_free_reserved_data_space(inode, start, len);
7681                 WARN_ON(dio_data->reserve < len);
7682                 dio_data->reserve -= len;
7683                 current->journal_info = dio_data;
7684         }
7685
7686         /*
7687          * In the case of write we need to clear and unlock the entire range,
7688          * in the case of read we need to unlock only the end area that we
7689          * aren't using if there is any left over space.
7690          */
7691         if (lockstart < lockend) {
7692                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7693                                  lockend, unlock_bits, 1, 0,
7694                                  &cached_state, GFP_NOFS);
7695         } else {
7696                 free_extent_state(cached_state);
7697         }
7698
7699         free_extent_map(em);
7700
7701         return 0;
7702
7703 unlock_err:
7704         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7705                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7706 err:
7707         if (dio_data)
7708                 current->journal_info = dio_data;
7709         /*
7710          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7711          * write less data then expected, so that we don't underflow our inode's
7712          * outstanding extents counter.
7713          */
7714         if (create && dio_data)
7715                 adjust_dio_outstanding_extents(inode, dio_data, len);
7716
7717         return ret;
7718 }
7719
7720 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7721                                         int rw, int mirror_num)
7722 {
7723         struct btrfs_root *root = BTRFS_I(inode)->root;
7724         int ret;
7725
7726         BUG_ON(rw & REQ_WRITE);
7727
7728         bio_get(bio);
7729
7730         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7731                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7732         if (ret)
7733                 goto err;
7734
7735         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7736 err:
7737         bio_put(bio);
7738         return ret;
7739 }
7740
7741 static int btrfs_check_dio_repairable(struct inode *inode,
7742                                       struct bio *failed_bio,
7743                                       struct io_failure_record *failrec,
7744                                       int failed_mirror)
7745 {
7746         int num_copies;
7747
7748         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7749                                       failrec->logical, failrec->len);
7750         if (num_copies == 1) {
7751                 /*
7752                  * we only have a single copy of the data, so don't bother with
7753                  * all the retry and error correction code that follows. no
7754                  * matter what the error is, it is very likely to persist.
7755                  */
7756                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7757                          num_copies, failrec->this_mirror, failed_mirror);
7758                 return 0;
7759         }
7760
7761         failrec->failed_mirror = failed_mirror;
7762         failrec->this_mirror++;
7763         if (failrec->this_mirror == failed_mirror)
7764                 failrec->this_mirror++;
7765
7766         if (failrec->this_mirror > num_copies) {
7767                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7768                          num_copies, failrec->this_mirror, failed_mirror);
7769                 return 0;
7770         }
7771
7772         return 1;
7773 }
7774
7775 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7776                           struct page *page, u64 start, u64 end,
7777                           int failed_mirror, bio_end_io_t *repair_endio,
7778                           void *repair_arg)
7779 {
7780         struct io_failure_record *failrec;
7781         struct bio *bio;
7782         int isector;
7783         int read_mode;
7784         int ret;
7785
7786         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7787
7788         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7789         if (ret)
7790                 return ret;
7791
7792         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7793                                          failed_mirror);
7794         if (!ret) {
7795                 free_io_failure(inode, failrec);
7796                 return -EIO;
7797         }
7798
7799         if (failed_bio->bi_vcnt > 1)
7800                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7801         else
7802                 read_mode = READ_SYNC;
7803
7804         isector = start - btrfs_io_bio(failed_bio)->logical;
7805         isector >>= inode->i_sb->s_blocksize_bits;
7806         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7807                                       0, isector, repair_endio, repair_arg);
7808         if (!bio) {
7809                 free_io_failure(inode, failrec);
7810                 return -EIO;
7811         }
7812
7813         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7814                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7815                     read_mode, failrec->this_mirror, failrec->in_validation);
7816
7817         ret = submit_dio_repair_bio(inode, bio, read_mode,
7818                                     failrec->this_mirror);
7819         if (ret) {
7820                 free_io_failure(inode, failrec);
7821                 bio_put(bio);
7822         }
7823
7824         return ret;
7825 }
7826
7827 struct btrfs_retry_complete {
7828         struct completion done;
7829         struct inode *inode;
7830         u64 start;
7831         int uptodate;
7832 };
7833
7834 static void btrfs_retry_endio_nocsum(struct bio *bio)
7835 {
7836         struct btrfs_retry_complete *done = bio->bi_private;
7837         struct bio_vec *bvec;
7838         int i;
7839
7840         if (bio->bi_error)
7841                 goto end;
7842
7843         done->uptodate = 1;
7844         bio_for_each_segment_all(bvec, bio, i)
7845                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7846 end:
7847         complete(&done->done);
7848         bio_put(bio);
7849 }
7850
7851 static int __btrfs_correct_data_nocsum(struct inode *inode,
7852                                        struct btrfs_io_bio *io_bio)
7853 {
7854         struct bio_vec *bvec;
7855         struct btrfs_retry_complete done;
7856         u64 start;
7857         int i;
7858         int ret;
7859
7860         start = io_bio->logical;
7861         done.inode = inode;
7862
7863         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7864 try_again:
7865                 done.uptodate = 0;
7866                 done.start = start;
7867                 init_completion(&done.done);
7868
7869                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7870                                      start + bvec->bv_len - 1,
7871                                      io_bio->mirror_num,
7872                                      btrfs_retry_endio_nocsum, &done);
7873                 if (ret)
7874                         return ret;
7875
7876                 wait_for_completion(&done.done);
7877
7878                 if (!done.uptodate) {
7879                         /* We might have another mirror, so try again */
7880                         goto try_again;
7881                 }
7882
7883                 start += bvec->bv_len;
7884         }
7885
7886         return 0;
7887 }
7888
7889 static void btrfs_retry_endio(struct bio *bio)
7890 {
7891         struct btrfs_retry_complete *done = bio->bi_private;
7892         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7893         struct bio_vec *bvec;
7894         int uptodate;
7895         int ret;
7896         int i;
7897
7898         if (bio->bi_error)
7899                 goto end;
7900
7901         uptodate = 1;
7902         bio_for_each_segment_all(bvec, bio, i) {
7903                 ret = __readpage_endio_check(done->inode, io_bio, i,
7904                                              bvec->bv_page, 0,
7905                                              done->start, bvec->bv_len);
7906                 if (!ret)
7907                         clean_io_failure(done->inode, done->start,
7908                                          bvec->bv_page, 0);
7909                 else
7910                         uptodate = 0;
7911         }
7912
7913         done->uptodate = uptodate;
7914 end:
7915         complete(&done->done);
7916         bio_put(bio);
7917 }
7918
7919 static int __btrfs_subio_endio_read(struct inode *inode,
7920                                     struct btrfs_io_bio *io_bio, int err)
7921 {
7922         struct bio_vec *bvec;
7923         struct btrfs_retry_complete done;
7924         u64 start;
7925         u64 offset = 0;
7926         int i;
7927         int ret;
7928
7929         err = 0;
7930         start = io_bio->logical;
7931         done.inode = inode;
7932
7933         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7934                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7935                                              0, start, bvec->bv_len);
7936                 if (likely(!ret))
7937                         goto next;
7938 try_again:
7939                 done.uptodate = 0;
7940                 done.start = start;
7941                 init_completion(&done.done);
7942
7943                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7944                                      start + bvec->bv_len - 1,
7945                                      io_bio->mirror_num,
7946                                      btrfs_retry_endio, &done);
7947                 if (ret) {
7948                         err = ret;
7949                         goto next;
7950                 }
7951
7952                 wait_for_completion(&done.done);
7953
7954                 if (!done.uptodate) {
7955                         /* We might have another mirror, so try again */
7956                         goto try_again;
7957                 }
7958 next:
7959                 offset += bvec->bv_len;
7960                 start += bvec->bv_len;
7961         }
7962
7963         return err;
7964 }
7965
7966 static int btrfs_subio_endio_read(struct inode *inode,
7967                                   struct btrfs_io_bio *io_bio, int err)
7968 {
7969         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7970
7971         if (skip_csum) {
7972                 if (unlikely(err))
7973                         return __btrfs_correct_data_nocsum(inode, io_bio);
7974                 else
7975                         return 0;
7976         } else {
7977                 return __btrfs_subio_endio_read(inode, io_bio, err);
7978         }
7979 }
7980
7981 static void btrfs_endio_direct_read(struct bio *bio)
7982 {
7983         struct btrfs_dio_private *dip = bio->bi_private;
7984         struct inode *inode = dip->inode;
7985         struct bio *dio_bio;
7986         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7987         int err = bio->bi_error;
7988
7989         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
7990                 err = btrfs_subio_endio_read(inode, io_bio, err);
7991
7992         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
7993                       dip->logical_offset + dip->bytes - 1);
7994         dio_bio = dip->dio_bio;
7995
7996         kfree(dip);
7997
7998         dio_bio->bi_error = bio->bi_error;
7999         dio_end_io(dio_bio, bio->bi_error);
8000
8001         if (io_bio->end_io)
8002                 io_bio->end_io(io_bio, err);
8003         bio_put(bio);
8004 }
8005
8006 static void btrfs_endio_direct_write(struct bio *bio)
8007 {
8008         struct btrfs_dio_private *dip = bio->bi_private;
8009         struct inode *inode = dip->inode;
8010         struct btrfs_root *root = BTRFS_I(inode)->root;
8011         struct btrfs_ordered_extent *ordered = NULL;
8012         u64 ordered_offset = dip->logical_offset;
8013         u64 ordered_bytes = dip->bytes;
8014         struct bio *dio_bio;
8015         int ret;
8016
8017 again:
8018         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8019                                                    &ordered_offset,
8020                                                    ordered_bytes,
8021                                                    !bio->bi_error);
8022         if (!ret)
8023                 goto out_test;
8024
8025         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8026                         finish_ordered_fn, NULL, NULL);
8027         btrfs_queue_work(root->fs_info->endio_write_workers,
8028                          &ordered->work);
8029 out_test:
8030         /*
8031          * our bio might span multiple ordered extents.  If we haven't
8032          * completed the accounting for the whole dio, go back and try again
8033          */
8034         if (ordered_offset < dip->logical_offset + dip->bytes) {
8035                 ordered_bytes = dip->logical_offset + dip->bytes -
8036                         ordered_offset;
8037                 ordered = NULL;
8038                 goto again;
8039         }
8040         dio_bio = dip->dio_bio;
8041
8042         kfree(dip);
8043
8044         dio_bio->bi_error = bio->bi_error;
8045         dio_end_io(dio_bio, bio->bi_error);
8046         bio_put(bio);
8047 }
8048
8049 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8050                                     struct bio *bio, int mirror_num,
8051                                     unsigned long bio_flags, u64 offset)
8052 {
8053         int ret;
8054         struct btrfs_root *root = BTRFS_I(inode)->root;
8055         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8056         BUG_ON(ret); /* -ENOMEM */
8057         return 0;
8058 }
8059
8060 static void btrfs_end_dio_bio(struct bio *bio)
8061 {
8062         struct btrfs_dio_private *dip = bio->bi_private;
8063         int err = bio->bi_error;
8064
8065         if (err)
8066                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8067                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8068                            btrfs_ino(dip->inode), bio->bi_rw,
8069                            (unsigned long long)bio->bi_iter.bi_sector,
8070                            bio->bi_iter.bi_size, err);
8071
8072         if (dip->subio_endio)
8073                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8074
8075         if (err) {
8076                 dip->errors = 1;
8077
8078                 /*
8079                  * before atomic variable goto zero, we must make sure
8080                  * dip->errors is perceived to be set.
8081                  */
8082                 smp_mb__before_atomic();
8083         }
8084
8085         /* if there are more bios still pending for this dio, just exit */
8086         if (!atomic_dec_and_test(&dip->pending_bios))
8087                 goto out;
8088
8089         if (dip->errors) {
8090                 bio_io_error(dip->orig_bio);
8091         } else {
8092                 dip->dio_bio->bi_error = 0;
8093                 bio_endio(dip->orig_bio);
8094         }
8095 out:
8096         bio_put(bio);
8097 }
8098
8099 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8100                                        u64 first_sector, gfp_t gfp_flags)
8101 {
8102         struct bio *bio;
8103         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8104         if (bio)
8105                 bio_associate_current(bio);
8106         return bio;
8107 }
8108
8109 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8110                                                  struct inode *inode,
8111                                                  struct btrfs_dio_private *dip,
8112                                                  struct bio *bio,
8113                                                  u64 file_offset)
8114 {
8115         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8116         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8117         int ret;
8118
8119         /*
8120          * We load all the csum data we need when we submit
8121          * the first bio to reduce the csum tree search and
8122          * contention.
8123          */
8124         if (dip->logical_offset == file_offset) {
8125                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8126                                                 file_offset);
8127                 if (ret)
8128                         return ret;
8129         }
8130
8131         if (bio == dip->orig_bio)
8132                 return 0;
8133
8134         file_offset -= dip->logical_offset;
8135         file_offset >>= inode->i_sb->s_blocksize_bits;
8136         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8137
8138         return 0;
8139 }
8140
8141 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8142                                          int rw, u64 file_offset, int skip_sum,
8143                                          int async_submit)
8144 {
8145         struct btrfs_dio_private *dip = bio->bi_private;
8146         int write = rw & REQ_WRITE;
8147         struct btrfs_root *root = BTRFS_I(inode)->root;
8148         int ret;
8149
8150         if (async_submit)
8151                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8152
8153         bio_get(bio);
8154
8155         if (!write) {
8156                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8157                                 BTRFS_WQ_ENDIO_DATA);
8158                 if (ret)
8159                         goto err;
8160         }
8161
8162         if (skip_sum)
8163                 goto map;
8164
8165         if (write && async_submit) {
8166                 ret = btrfs_wq_submit_bio(root->fs_info,
8167                                    inode, rw, bio, 0, 0,
8168                                    file_offset,
8169                                    __btrfs_submit_bio_start_direct_io,
8170                                    __btrfs_submit_bio_done);
8171                 goto err;
8172         } else if (write) {
8173                 /*
8174                  * If we aren't doing async submit, calculate the csum of the
8175                  * bio now.
8176                  */
8177                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8178                 if (ret)
8179                         goto err;
8180         } else {
8181                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8182                                                      file_offset);
8183                 if (ret)
8184                         goto err;
8185         }
8186 map:
8187         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8188 err:
8189         bio_put(bio);
8190         return ret;
8191 }
8192
8193 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8194                                     int skip_sum)
8195 {
8196         struct inode *inode = dip->inode;
8197         struct btrfs_root *root = BTRFS_I(inode)->root;
8198         struct bio *bio;
8199         struct bio *orig_bio = dip->orig_bio;
8200         struct bio_vec *bvec = orig_bio->bi_io_vec;
8201         u64 start_sector = orig_bio->bi_iter.bi_sector;
8202         u64 file_offset = dip->logical_offset;
8203         u64 submit_len = 0;
8204         u64 map_length;
8205         int nr_pages = 0;
8206         int ret;
8207         int async_submit = 0;
8208
8209         map_length = orig_bio->bi_iter.bi_size;
8210         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8211                               &map_length, NULL, 0);
8212         if (ret)
8213                 return -EIO;
8214
8215         if (map_length >= orig_bio->bi_iter.bi_size) {
8216                 bio = orig_bio;
8217                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8218                 goto submit;
8219         }
8220
8221         /* async crcs make it difficult to collect full stripe writes. */
8222         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8223                 async_submit = 0;
8224         else
8225                 async_submit = 1;
8226
8227         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8228         if (!bio)
8229                 return -ENOMEM;
8230
8231         bio->bi_private = dip;
8232         bio->bi_end_io = btrfs_end_dio_bio;
8233         btrfs_io_bio(bio)->logical = file_offset;
8234         atomic_inc(&dip->pending_bios);
8235
8236         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8237                 if (map_length < submit_len + bvec->bv_len ||
8238                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
8239                                  bvec->bv_offset) < bvec->bv_len) {
8240                         /*
8241                          * inc the count before we submit the bio so
8242                          * we know the end IO handler won't happen before
8243                          * we inc the count. Otherwise, the dip might get freed
8244                          * before we're done setting it up
8245                          */
8246                         atomic_inc(&dip->pending_bios);
8247                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8248                                                      file_offset, skip_sum,
8249                                                      async_submit);
8250                         if (ret) {
8251                                 bio_put(bio);
8252                                 atomic_dec(&dip->pending_bios);
8253                                 goto out_err;
8254                         }
8255
8256                         start_sector += submit_len >> 9;
8257                         file_offset += submit_len;
8258
8259                         submit_len = 0;
8260                         nr_pages = 0;
8261
8262                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8263                                                   start_sector, GFP_NOFS);
8264                         if (!bio)
8265                                 goto out_err;
8266                         bio->bi_private = dip;
8267                         bio->bi_end_io = btrfs_end_dio_bio;
8268                         btrfs_io_bio(bio)->logical = file_offset;
8269
8270                         map_length = orig_bio->bi_iter.bi_size;
8271                         ret = btrfs_map_block(root->fs_info, rw,
8272                                               start_sector << 9,
8273                                               &map_length, NULL, 0);
8274                         if (ret) {
8275                                 bio_put(bio);
8276                                 goto out_err;
8277                         }
8278                 } else {
8279                         submit_len += bvec->bv_len;
8280                         nr_pages++;
8281                         bvec++;
8282                 }
8283         }
8284
8285 submit:
8286         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8287                                      async_submit);
8288         if (!ret)
8289                 return 0;
8290
8291         bio_put(bio);
8292 out_err:
8293         dip->errors = 1;
8294         /*
8295          * before atomic variable goto zero, we must
8296          * make sure dip->errors is perceived to be set.
8297          */
8298         smp_mb__before_atomic();
8299         if (atomic_dec_and_test(&dip->pending_bios))
8300                 bio_io_error(dip->orig_bio);
8301
8302         /* bio_end_io() will handle error, so we needn't return it */
8303         return 0;
8304 }
8305
8306 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8307                                 struct inode *inode, loff_t file_offset)
8308 {
8309         struct btrfs_dio_private *dip = NULL;
8310         struct bio *io_bio = NULL;
8311         struct btrfs_io_bio *btrfs_bio;
8312         int skip_sum;
8313         int write = rw & REQ_WRITE;
8314         int ret = 0;
8315
8316         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8317
8318         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8319         if (!io_bio) {
8320                 ret = -ENOMEM;
8321                 goto free_ordered;
8322         }
8323
8324         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8325         if (!dip) {
8326                 ret = -ENOMEM;
8327                 goto free_ordered;
8328         }
8329
8330         dip->private = dio_bio->bi_private;
8331         dip->inode = inode;
8332         dip->logical_offset = file_offset;
8333         dip->bytes = dio_bio->bi_iter.bi_size;
8334         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8335         io_bio->bi_private = dip;
8336         dip->orig_bio = io_bio;
8337         dip->dio_bio = dio_bio;
8338         atomic_set(&dip->pending_bios, 0);
8339         btrfs_bio = btrfs_io_bio(io_bio);
8340         btrfs_bio->logical = file_offset;
8341
8342         if (write) {
8343                 io_bio->bi_end_io = btrfs_endio_direct_write;
8344         } else {
8345                 io_bio->bi_end_io = btrfs_endio_direct_read;
8346                 dip->subio_endio = btrfs_subio_endio_read;
8347         }
8348
8349         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8350         if (!ret)
8351                 return;
8352
8353         if (btrfs_bio->end_io)
8354                 btrfs_bio->end_io(btrfs_bio, ret);
8355
8356 free_ordered:
8357         /*
8358          * If we arrived here it means either we failed to submit the dip
8359          * or we either failed to clone the dio_bio or failed to allocate the
8360          * dip. If we cloned the dio_bio and allocated the dip, we can just
8361          * call bio_endio against our io_bio so that we get proper resource
8362          * cleanup if we fail to submit the dip, otherwise, we must do the
8363          * same as btrfs_endio_direct_[write|read] because we can't call these
8364          * callbacks - they require an allocated dip and a clone of dio_bio.
8365          */
8366         if (io_bio && dip) {
8367                 io_bio->bi_error = -EIO;
8368                 bio_endio(io_bio);
8369                 /*
8370                  * The end io callbacks free our dip, do the final put on io_bio
8371                  * and all the cleanup and final put for dio_bio (through
8372                  * dio_end_io()).
8373                  */
8374                 dip = NULL;
8375                 io_bio = NULL;
8376         } else {
8377                 if (write) {
8378                         struct btrfs_ordered_extent *ordered;
8379
8380                         ordered = btrfs_lookup_ordered_extent(inode,
8381                                                               file_offset);
8382                         set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
8383                         /*
8384                          * Decrements our ref on the ordered extent and removes
8385                          * the ordered extent from the inode's ordered tree,
8386                          * doing all the proper resource cleanup such as for the
8387                          * reserved space and waking up any waiters for this
8388                          * ordered extent (through btrfs_remove_ordered_extent).
8389                          */
8390                         btrfs_finish_ordered_io(ordered);
8391                 } else {
8392                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8393                               file_offset + dio_bio->bi_iter.bi_size - 1);
8394                 }
8395                 dio_bio->bi_error = -EIO;
8396                 /*
8397                  * Releases and cleans up our dio_bio, no need to bio_put()
8398                  * nor bio_endio()/bio_io_error() against dio_bio.
8399                  */
8400                 dio_end_io(dio_bio, ret);
8401         }
8402         if (io_bio)
8403                 bio_put(io_bio);
8404         kfree(dip);
8405 }
8406
8407 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8408                         const struct iov_iter *iter, loff_t offset)
8409 {
8410         int seg;
8411         int i;
8412         unsigned blocksize_mask = root->sectorsize - 1;
8413         ssize_t retval = -EINVAL;
8414
8415         if (offset & blocksize_mask)
8416                 goto out;
8417
8418         if (iov_iter_alignment(iter) & blocksize_mask)
8419                 goto out;
8420
8421         /* If this is a write we don't need to check anymore */
8422         if (iov_iter_rw(iter) == WRITE)
8423                 return 0;
8424         /*
8425          * Check to make sure we don't have duplicate iov_base's in this
8426          * iovec, if so return EINVAL, otherwise we'll get csum errors
8427          * when reading back.
8428          */
8429         for (seg = 0; seg < iter->nr_segs; seg++) {
8430                 for (i = seg + 1; i < iter->nr_segs; i++) {
8431                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8432                                 goto out;
8433                 }
8434         }
8435         retval = 0;
8436 out:
8437         return retval;
8438 }
8439
8440 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
8441                                loff_t offset)
8442 {
8443         struct file *file = iocb->ki_filp;
8444         struct inode *inode = file->f_mapping->host;
8445         struct btrfs_root *root = BTRFS_I(inode)->root;
8446         struct btrfs_dio_data dio_data = { 0 };
8447         size_t count = 0;
8448         int flags = 0;
8449         bool wakeup = true;
8450         bool relock = false;
8451         ssize_t ret;
8452
8453         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8454                 return 0;
8455
8456         inode_dio_begin(inode);
8457         smp_mb__after_atomic();
8458
8459         /*
8460          * The generic stuff only does filemap_write_and_wait_range, which
8461          * isn't enough if we've written compressed pages to this area, so
8462          * we need to flush the dirty pages again to make absolutely sure
8463          * that any outstanding dirty pages are on disk.
8464          */
8465         count = iov_iter_count(iter);
8466         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8467                      &BTRFS_I(inode)->runtime_flags))
8468                 filemap_fdatawrite_range(inode->i_mapping, offset,
8469                                          offset + count - 1);
8470
8471         if (iov_iter_rw(iter) == WRITE) {
8472                 /*
8473                  * If the write DIO is beyond the EOF, we need update
8474                  * the isize, but it is protected by i_mutex. So we can
8475                  * not unlock the i_mutex at this case.
8476                  */
8477                 if (offset + count <= inode->i_size) {
8478                         mutex_unlock(&inode->i_mutex);
8479                         relock = true;
8480                 }
8481                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8482                 if (ret)
8483                         goto out;
8484                 dio_data.outstanding_extents = div64_u64(count +
8485                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8486                                                 BTRFS_MAX_EXTENT_SIZE);
8487
8488                 /*
8489                  * We need to know how many extents we reserved so that we can
8490                  * do the accounting properly if we go over the number we
8491                  * originally calculated.  Abuse current->journal_info for this.
8492                  */
8493                 dio_data.reserve = round_up(count, root->sectorsize);
8494                 current->journal_info = &dio_data;
8495         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8496                                      &BTRFS_I(inode)->runtime_flags)) {
8497                 inode_dio_end(inode);
8498                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8499                 wakeup = false;
8500         }
8501
8502         ret = __blockdev_direct_IO(iocb, inode,
8503                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8504                                    iter, offset, btrfs_get_blocks_direct, NULL,
8505                                    btrfs_submit_direct, flags);
8506         if (iov_iter_rw(iter) == WRITE) {
8507                 current->journal_info = NULL;
8508                 if (ret < 0 && ret != -EIOCBQUEUED) {
8509                         if (dio_data.reserve)
8510                                 btrfs_delalloc_release_space(inode, offset,
8511                                                              dio_data.reserve);
8512                 } else if (ret >= 0 && (size_t)ret < count)
8513                         btrfs_delalloc_release_space(inode, offset,
8514                                                      count - (size_t)ret);
8515         }
8516 out:
8517         if (wakeup)
8518                 inode_dio_end(inode);
8519         if (relock)
8520                 mutex_lock(&inode->i_mutex);
8521
8522         return ret;
8523 }
8524
8525 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8526
8527 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8528                 __u64 start, __u64 len)
8529 {
8530         int     ret;
8531
8532         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8533         if (ret)
8534                 return ret;
8535
8536         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8537 }
8538
8539 int btrfs_readpage(struct file *file, struct page *page)
8540 {
8541         struct extent_io_tree *tree;
8542         tree = &BTRFS_I(page->mapping->host)->io_tree;
8543         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8544 }
8545
8546 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8547 {
8548         struct extent_io_tree *tree;
8549         struct inode *inode = page->mapping->host;
8550         int ret;
8551
8552         if (current->flags & PF_MEMALLOC) {
8553                 redirty_page_for_writepage(wbc, page);
8554                 unlock_page(page);
8555                 return 0;
8556         }
8557
8558         /*
8559          * If we are under memory pressure we will call this directly from the
8560          * VM, we need to make sure we have the inode referenced for the ordered
8561          * extent.  If not just return like we didn't do anything.
8562          */
8563         if (!igrab(inode)) {
8564                 redirty_page_for_writepage(wbc, page);
8565                 return AOP_WRITEPAGE_ACTIVATE;
8566         }
8567         tree = &BTRFS_I(page->mapping->host)->io_tree;
8568         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8569         btrfs_add_delayed_iput(inode);
8570         return ret;
8571 }
8572
8573 static int btrfs_writepages(struct address_space *mapping,
8574                             struct writeback_control *wbc)
8575 {
8576         struct extent_io_tree *tree;
8577
8578         tree = &BTRFS_I(mapping->host)->io_tree;
8579         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8580 }
8581
8582 static int
8583 btrfs_readpages(struct file *file, struct address_space *mapping,
8584                 struct list_head *pages, unsigned nr_pages)
8585 {
8586         struct extent_io_tree *tree;
8587         tree = &BTRFS_I(mapping->host)->io_tree;
8588         return extent_readpages(tree, mapping, pages, nr_pages,
8589                                 btrfs_get_extent);
8590 }
8591 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8592 {
8593         struct extent_io_tree *tree;
8594         struct extent_map_tree *map;
8595         int ret;
8596
8597         tree = &BTRFS_I(page->mapping->host)->io_tree;
8598         map = &BTRFS_I(page->mapping->host)->extent_tree;
8599         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8600         if (ret == 1) {
8601                 ClearPagePrivate(page);
8602                 set_page_private(page, 0);
8603                 page_cache_release(page);
8604         }
8605         return ret;
8606 }
8607
8608 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8609 {
8610         if (PageWriteback(page) || PageDirty(page))
8611                 return 0;
8612         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8613 }
8614
8615 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8616                                  unsigned int length)
8617 {
8618         struct inode *inode = page->mapping->host;
8619         struct extent_io_tree *tree;
8620         struct btrfs_ordered_extent *ordered;
8621         struct extent_state *cached_state = NULL;
8622         u64 page_start = page_offset(page);
8623         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
8624         int inode_evicting = inode->i_state & I_FREEING;
8625
8626         /*
8627          * we have the page locked, so new writeback can't start,
8628          * and the dirty bit won't be cleared while we are here.
8629          *
8630          * Wait for IO on this page so that we can safely clear
8631          * the PagePrivate2 bit and do ordered accounting
8632          */
8633         wait_on_page_writeback(page);
8634
8635         tree = &BTRFS_I(inode)->io_tree;
8636         if (offset) {
8637                 btrfs_releasepage(page, GFP_NOFS);
8638                 return;
8639         }
8640
8641         if (!inode_evicting)
8642                 lock_extent_bits(tree, page_start, page_end, 0, &cached_state);
8643         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8644         if (ordered) {
8645                 /*
8646                  * IO on this page will never be started, so we need
8647                  * to account for any ordered extents now
8648                  */
8649                 if (!inode_evicting)
8650                         clear_extent_bit(tree, page_start, page_end,
8651                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8652                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8653                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8654                                          GFP_NOFS);
8655                 /*
8656                  * whoever cleared the private bit is responsible
8657                  * for the finish_ordered_io
8658                  */
8659                 if (TestClearPagePrivate2(page)) {
8660                         struct btrfs_ordered_inode_tree *tree;
8661                         u64 new_len;
8662
8663                         tree = &BTRFS_I(inode)->ordered_tree;
8664
8665                         spin_lock_irq(&tree->lock);
8666                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8667                         new_len = page_start - ordered->file_offset;
8668                         if (new_len < ordered->truncated_len)
8669                                 ordered->truncated_len = new_len;
8670                         spin_unlock_irq(&tree->lock);
8671
8672                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8673                                                            page_start,
8674                                                            PAGE_CACHE_SIZE, 1))
8675                                 btrfs_finish_ordered_io(ordered);
8676                 }
8677                 btrfs_put_ordered_extent(ordered);
8678                 if (!inode_evicting) {
8679                         cached_state = NULL;
8680                         lock_extent_bits(tree, page_start, page_end, 0,
8681                                          &cached_state);
8682                 }
8683         }
8684
8685         /*
8686          * Qgroup reserved space handler
8687          * Page here will be either
8688          * 1) Already written to disk
8689          *    In this case, its reserved space is released from data rsv map
8690          *    and will be freed by delayed_ref handler finally.
8691          *    So even we call qgroup_free_data(), it won't decrease reserved
8692          *    space.
8693          * 2) Not written to disk
8694          *    This means the reserved space should be freed here.
8695          */
8696         btrfs_qgroup_free_data(inode, page_start, PAGE_CACHE_SIZE);
8697         if (!inode_evicting) {
8698                 clear_extent_bit(tree, page_start, page_end,
8699                                  EXTENT_LOCKED | EXTENT_DIRTY |
8700                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8701                                  EXTENT_DEFRAG, 1, 1,
8702                                  &cached_state, GFP_NOFS);
8703
8704                 __btrfs_releasepage(page, GFP_NOFS);
8705         }
8706
8707         ClearPageChecked(page);
8708         if (PagePrivate(page)) {
8709                 ClearPagePrivate(page);
8710                 set_page_private(page, 0);
8711                 page_cache_release(page);
8712         }
8713 }
8714
8715 /*
8716  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8717  * called from a page fault handler when a page is first dirtied. Hence we must
8718  * be careful to check for EOF conditions here. We set the page up correctly
8719  * for a written page which means we get ENOSPC checking when writing into
8720  * holes and correct delalloc and unwritten extent mapping on filesystems that
8721  * support these features.
8722  *
8723  * We are not allowed to take the i_mutex here so we have to play games to
8724  * protect against truncate races as the page could now be beyond EOF.  Because
8725  * vmtruncate() writes the inode size before removing pages, once we have the
8726  * page lock we can determine safely if the page is beyond EOF. If it is not
8727  * beyond EOF, then the page is guaranteed safe against truncation until we
8728  * unlock the page.
8729  */
8730 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8731 {
8732         struct page *page = vmf->page;
8733         struct inode *inode = file_inode(vma->vm_file);
8734         struct btrfs_root *root = BTRFS_I(inode)->root;
8735         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8736         struct btrfs_ordered_extent *ordered;
8737         struct extent_state *cached_state = NULL;
8738         char *kaddr;
8739         unsigned long zero_start;
8740         loff_t size;
8741         int ret;
8742         int reserved = 0;
8743         u64 page_start;
8744         u64 page_end;
8745
8746         sb_start_pagefault(inode->i_sb);
8747         page_start = page_offset(page);
8748         page_end = page_start + PAGE_CACHE_SIZE - 1;
8749
8750         ret = btrfs_delalloc_reserve_space(inode, page_start,
8751                                            PAGE_CACHE_SIZE);
8752         if (!ret) {
8753                 ret = file_update_time(vma->vm_file);
8754                 reserved = 1;
8755         }
8756         if (ret) {
8757                 if (ret == -ENOMEM)
8758                         ret = VM_FAULT_OOM;
8759                 else /* -ENOSPC, -EIO, etc */
8760                         ret = VM_FAULT_SIGBUS;
8761                 if (reserved)
8762                         goto out;
8763                 goto out_noreserve;
8764         }
8765
8766         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8767 again:
8768         lock_page(page);
8769         size = i_size_read(inode);
8770
8771         if ((page->mapping != inode->i_mapping) ||
8772             (page_start >= size)) {
8773                 /* page got truncated out from underneath us */
8774                 goto out_unlock;
8775         }
8776         wait_on_page_writeback(page);
8777
8778         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state);
8779         set_page_extent_mapped(page);
8780
8781         /*
8782          * we can't set the delalloc bits if there are pending ordered
8783          * extents.  Drop our locks and wait for them to finish
8784          */
8785         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8786         if (ordered) {
8787                 unlock_extent_cached(io_tree, page_start, page_end,
8788                                      &cached_state, GFP_NOFS);
8789                 unlock_page(page);
8790                 btrfs_start_ordered_extent(inode, ordered, 1);
8791                 btrfs_put_ordered_extent(ordered);
8792                 goto again;
8793         }
8794
8795         /*
8796          * XXX - page_mkwrite gets called every time the page is dirtied, even
8797          * if it was already dirty, so for space accounting reasons we need to
8798          * clear any delalloc bits for the range we are fixing to save.  There
8799          * is probably a better way to do this, but for now keep consistent with
8800          * prepare_pages in the normal write path.
8801          */
8802         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
8803                           EXTENT_DIRTY | EXTENT_DELALLOC |
8804                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8805                           0, 0, &cached_state, GFP_NOFS);
8806
8807         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
8808                                         &cached_state);
8809         if (ret) {
8810                 unlock_extent_cached(io_tree, page_start, page_end,
8811                                      &cached_state, GFP_NOFS);
8812                 ret = VM_FAULT_SIGBUS;
8813                 goto out_unlock;
8814         }
8815         ret = 0;
8816
8817         /* page is wholly or partially inside EOF */
8818         if (page_start + PAGE_CACHE_SIZE > size)
8819                 zero_start = size & ~PAGE_CACHE_MASK;
8820         else
8821                 zero_start = PAGE_CACHE_SIZE;
8822
8823         if (zero_start != PAGE_CACHE_SIZE) {
8824                 kaddr = kmap(page);
8825                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
8826                 flush_dcache_page(page);
8827                 kunmap(page);
8828         }
8829         ClearPageChecked(page);
8830         set_page_dirty(page);
8831         SetPageUptodate(page);
8832
8833         BTRFS_I(inode)->last_trans = root->fs_info->generation;
8834         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8835         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8836
8837         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
8838
8839 out_unlock:
8840         if (!ret) {
8841                 sb_end_pagefault(inode->i_sb);
8842                 return VM_FAULT_LOCKED;
8843         }
8844         unlock_page(page);
8845 out:
8846         btrfs_delalloc_release_space(inode, page_start, PAGE_CACHE_SIZE);
8847 out_noreserve:
8848         sb_end_pagefault(inode->i_sb);
8849         return ret;
8850 }
8851
8852 static int btrfs_truncate(struct inode *inode)
8853 {
8854         struct btrfs_root *root = BTRFS_I(inode)->root;
8855         struct btrfs_block_rsv *rsv;
8856         int ret = 0;
8857         int err = 0;
8858         struct btrfs_trans_handle *trans;
8859         u64 mask = root->sectorsize - 1;
8860         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
8861
8862         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8863                                        (u64)-1);
8864         if (ret)
8865                 return ret;
8866
8867         /*
8868          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
8869          * 3 things going on here
8870          *
8871          * 1) We need to reserve space for our orphan item and the space to
8872          * delete our orphan item.  Lord knows we don't want to have a dangling
8873          * orphan item because we didn't reserve space to remove it.
8874          *
8875          * 2) We need to reserve space to update our inode.
8876          *
8877          * 3) We need to have something to cache all the space that is going to
8878          * be free'd up by the truncate operation, but also have some slack
8879          * space reserved in case it uses space during the truncate (thank you
8880          * very much snapshotting).
8881          *
8882          * And we need these to all be seperate.  The fact is we can use alot of
8883          * space doing the truncate, and we have no earthly idea how much space
8884          * we will use, so we need the truncate reservation to be seperate so it
8885          * doesn't end up using space reserved for updating the inode or
8886          * removing the orphan item.  We also need to be able to stop the
8887          * transaction and start a new one, which means we need to be able to
8888          * update the inode several times, and we have no idea of knowing how
8889          * many times that will be, so we can't just reserve 1 item for the
8890          * entirety of the opration, so that has to be done seperately as well.
8891          * Then there is the orphan item, which does indeed need to be held on
8892          * to for the whole operation, and we need nobody to touch this reserved
8893          * space except the orphan code.
8894          *
8895          * So that leaves us with
8896          *
8897          * 1) root->orphan_block_rsv - for the orphan deletion.
8898          * 2) rsv - for the truncate reservation, which we will steal from the
8899          * transaction reservation.
8900          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
8901          * updating the inode.
8902          */
8903         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
8904         if (!rsv)
8905                 return -ENOMEM;
8906         rsv->size = min_size;
8907         rsv->failfast = 1;
8908
8909         /*
8910          * 1 for the truncate slack space
8911          * 1 for updating the inode.
8912          */
8913         trans = btrfs_start_transaction(root, 2);
8914         if (IS_ERR(trans)) {
8915                 err = PTR_ERR(trans);
8916                 goto out;
8917         }
8918
8919         /* Migrate the slack space for the truncate to our reserve */
8920         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
8921                                       min_size);
8922         BUG_ON(ret);
8923
8924         /*
8925          * So if we truncate and then write and fsync we normally would just
8926          * write the extents that changed, which is a problem if we need to
8927          * first truncate that entire inode.  So set this flag so we write out
8928          * all of the extents in the inode to the sync log so we're completely
8929          * safe.
8930          */
8931         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8932         trans->block_rsv = rsv;
8933
8934         while (1) {
8935                 ret = btrfs_truncate_inode_items(trans, root, inode,
8936                                                  inode->i_size,
8937                                                  BTRFS_EXTENT_DATA_KEY);
8938                 if (ret != -ENOSPC && ret != -EAGAIN) {
8939                         err = ret;
8940                         break;
8941                 }
8942
8943                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8944                 ret = btrfs_update_inode(trans, root, inode);
8945                 if (ret) {
8946                         err = ret;
8947                         break;
8948                 }
8949
8950                 btrfs_end_transaction(trans, root);
8951                 btrfs_btree_balance_dirty(root);
8952
8953                 trans = btrfs_start_transaction(root, 2);
8954                 if (IS_ERR(trans)) {
8955                         ret = err = PTR_ERR(trans);
8956                         trans = NULL;
8957                         break;
8958                 }
8959
8960                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
8961                                               rsv, min_size);
8962                 BUG_ON(ret);    /* shouldn't happen */
8963                 trans->block_rsv = rsv;
8964         }
8965
8966         if (ret == 0 && inode->i_nlink > 0) {
8967                 trans->block_rsv = root->orphan_block_rsv;
8968                 ret = btrfs_orphan_del(trans, inode);
8969                 if (ret)
8970                         err = ret;
8971         }
8972
8973         if (trans) {
8974                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8975                 ret = btrfs_update_inode(trans, root, inode);
8976                 if (ret && !err)
8977                         err = ret;
8978
8979                 ret = btrfs_end_transaction(trans, root);
8980                 btrfs_btree_balance_dirty(root);
8981         }
8982
8983 out:
8984         btrfs_free_block_rsv(root, rsv);
8985
8986         if (ret && !err)
8987                 err = ret;
8988
8989         return err;
8990 }
8991
8992 /*
8993  * create a new subvolume directory/inode (helper for the ioctl).
8994  */
8995 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
8996                              struct btrfs_root *new_root,
8997                              struct btrfs_root *parent_root,
8998                              u64 new_dirid)
8999 {
9000         struct inode *inode;
9001         int err;
9002         u64 index = 0;
9003
9004         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9005                                 new_dirid, new_dirid,
9006                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9007                                 &index);
9008         if (IS_ERR(inode))
9009                 return PTR_ERR(inode);
9010         inode->i_op = &btrfs_dir_inode_operations;
9011         inode->i_fop = &btrfs_dir_file_operations;
9012
9013         set_nlink(inode, 1);
9014         btrfs_i_size_write(inode, 0);
9015         unlock_new_inode(inode);
9016
9017         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9018         if (err)
9019                 btrfs_err(new_root->fs_info,
9020                           "error inheriting subvolume %llu properties: %d",
9021                           new_root->root_key.objectid, err);
9022
9023         err = btrfs_update_inode(trans, new_root, inode);
9024
9025         iput(inode);
9026         return err;
9027 }
9028
9029 struct inode *btrfs_alloc_inode(struct super_block *sb)
9030 {
9031         struct btrfs_inode *ei;
9032         struct inode *inode;
9033
9034         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9035         if (!ei)
9036                 return NULL;
9037
9038         ei->root = NULL;
9039         ei->generation = 0;
9040         ei->last_trans = 0;
9041         ei->last_sub_trans = 0;
9042         ei->logged_trans = 0;
9043         ei->delalloc_bytes = 0;
9044         ei->defrag_bytes = 0;
9045         ei->disk_i_size = 0;
9046         ei->flags = 0;
9047         ei->csum_bytes = 0;
9048         ei->index_cnt = (u64)-1;
9049         ei->dir_index = 0;
9050         ei->last_unlink_trans = 0;
9051         ei->last_log_commit = 0;
9052
9053         spin_lock_init(&ei->lock);
9054         ei->outstanding_extents = 0;
9055         ei->reserved_extents = 0;
9056
9057         ei->runtime_flags = 0;
9058         ei->force_compress = BTRFS_COMPRESS_NONE;
9059
9060         ei->delayed_node = NULL;
9061
9062         ei->i_otime.tv_sec = 0;
9063         ei->i_otime.tv_nsec = 0;
9064
9065         inode = &ei->vfs_inode;
9066         extent_map_tree_init(&ei->extent_tree);
9067         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9068         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9069         ei->io_tree.track_uptodate = 1;
9070         ei->io_failure_tree.track_uptodate = 1;
9071         atomic_set(&ei->sync_writers, 0);
9072         mutex_init(&ei->log_mutex);
9073         mutex_init(&ei->delalloc_mutex);
9074         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9075         INIT_LIST_HEAD(&ei->delalloc_inodes);
9076         RB_CLEAR_NODE(&ei->rb_node);
9077
9078         return inode;
9079 }
9080
9081 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9082 void btrfs_test_destroy_inode(struct inode *inode)
9083 {
9084         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9085         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9086 }
9087 #endif
9088
9089 static void btrfs_i_callback(struct rcu_head *head)
9090 {
9091         struct inode *inode = container_of(head, struct inode, i_rcu);
9092         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9093 }
9094
9095 void btrfs_destroy_inode(struct inode *inode)
9096 {
9097         struct btrfs_ordered_extent *ordered;
9098         struct btrfs_root *root = BTRFS_I(inode)->root;
9099
9100         WARN_ON(!hlist_empty(&inode->i_dentry));
9101         WARN_ON(inode->i_data.nrpages);
9102         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9103         WARN_ON(BTRFS_I(inode)->reserved_extents);
9104         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9105         WARN_ON(BTRFS_I(inode)->csum_bytes);
9106         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9107
9108         /*
9109          * This can happen where we create an inode, but somebody else also
9110          * created the same inode and we need to destroy the one we already
9111          * created.
9112          */
9113         if (!root)
9114                 goto free;
9115
9116         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9117                      &BTRFS_I(inode)->runtime_flags)) {
9118                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9119                         btrfs_ino(inode));
9120                 atomic_dec(&root->orphan_inodes);
9121         }
9122
9123         while (1) {
9124                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9125                 if (!ordered)
9126                         break;
9127                 else {
9128                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9129                                 ordered->file_offset, ordered->len);
9130                         btrfs_remove_ordered_extent(inode, ordered);
9131                         btrfs_put_ordered_extent(ordered);
9132                         btrfs_put_ordered_extent(ordered);
9133                 }
9134         }
9135         btrfs_qgroup_check_reserved_leak(inode);
9136         inode_tree_del(inode);
9137         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9138 free:
9139         call_rcu(&inode->i_rcu, btrfs_i_callback);
9140 }
9141
9142 int btrfs_drop_inode(struct inode *inode)
9143 {
9144         struct btrfs_root *root = BTRFS_I(inode)->root;
9145
9146         if (root == NULL)
9147                 return 1;
9148
9149         /* the snap/subvol tree is on deleting */
9150         if (btrfs_root_refs(&root->root_item) == 0)
9151                 return 1;
9152         else
9153                 return generic_drop_inode(inode);
9154 }
9155
9156 static void init_once(void *foo)
9157 {
9158         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9159
9160         inode_init_once(&ei->vfs_inode);
9161 }
9162
9163 void btrfs_destroy_cachep(void)
9164 {
9165         /*
9166          * Make sure all delayed rcu free inodes are flushed before we
9167          * destroy cache.
9168          */
9169         rcu_barrier();
9170         if (btrfs_inode_cachep)
9171                 kmem_cache_destroy(btrfs_inode_cachep);
9172         if (btrfs_trans_handle_cachep)
9173                 kmem_cache_destroy(btrfs_trans_handle_cachep);
9174         if (btrfs_transaction_cachep)
9175                 kmem_cache_destroy(btrfs_transaction_cachep);
9176         if (btrfs_path_cachep)
9177                 kmem_cache_destroy(btrfs_path_cachep);
9178         if (btrfs_free_space_cachep)
9179                 kmem_cache_destroy(btrfs_free_space_cachep);
9180         if (btrfs_delalloc_work_cachep)
9181                 kmem_cache_destroy(btrfs_delalloc_work_cachep);
9182 }
9183
9184 int btrfs_init_cachep(void)
9185 {
9186         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9187                         sizeof(struct btrfs_inode), 0,
9188                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, init_once);
9189         if (!btrfs_inode_cachep)
9190                 goto fail;
9191
9192         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9193                         sizeof(struct btrfs_trans_handle), 0,
9194                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9195         if (!btrfs_trans_handle_cachep)
9196                 goto fail;
9197
9198         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9199                         sizeof(struct btrfs_transaction), 0,
9200                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9201         if (!btrfs_transaction_cachep)
9202                 goto fail;
9203
9204         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9205                         sizeof(struct btrfs_path), 0,
9206                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9207         if (!btrfs_path_cachep)
9208                 goto fail;
9209
9210         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9211                         sizeof(struct btrfs_free_space), 0,
9212                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9213         if (!btrfs_free_space_cachep)
9214                 goto fail;
9215
9216         btrfs_delalloc_work_cachep = kmem_cache_create("btrfs_delalloc_work",
9217                         sizeof(struct btrfs_delalloc_work), 0,
9218                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
9219                         NULL);
9220         if (!btrfs_delalloc_work_cachep)
9221                 goto fail;
9222
9223         return 0;
9224 fail:
9225         btrfs_destroy_cachep();
9226         return -ENOMEM;
9227 }
9228
9229 static int btrfs_getattr(struct vfsmount *mnt,
9230                          struct dentry *dentry, struct kstat *stat)
9231 {
9232         u64 delalloc_bytes;
9233         struct inode *inode = d_inode(dentry);
9234         u32 blocksize = inode->i_sb->s_blocksize;
9235
9236         generic_fillattr(inode, stat);
9237         stat->dev = BTRFS_I(inode)->root->anon_dev;
9238         stat->blksize = PAGE_CACHE_SIZE;
9239
9240         spin_lock(&BTRFS_I(inode)->lock);
9241         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9242         spin_unlock(&BTRFS_I(inode)->lock);
9243         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9244                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9245         return 0;
9246 }
9247
9248 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9249                            struct inode *new_dir, struct dentry *new_dentry)
9250 {
9251         struct btrfs_trans_handle *trans;
9252         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9253         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9254         struct inode *new_inode = d_inode(new_dentry);
9255         struct inode *old_inode = d_inode(old_dentry);
9256         struct timespec ctime = CURRENT_TIME;
9257         u64 index = 0;
9258         u64 root_objectid;
9259         int ret;
9260         u64 old_ino = btrfs_ino(old_inode);
9261
9262         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9263                 return -EPERM;
9264
9265         /* we only allow rename subvolume link between subvolumes */
9266         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9267                 return -EXDEV;
9268
9269         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9270             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9271                 return -ENOTEMPTY;
9272
9273         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9274             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9275                 return -ENOTEMPTY;
9276
9277
9278         /* check for collisions, even if the  name isn't there */
9279         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9280                              new_dentry->d_name.name,
9281                              new_dentry->d_name.len);
9282
9283         if (ret) {
9284                 if (ret == -EEXIST) {
9285                         /* we shouldn't get
9286                          * eexist without a new_inode */
9287                         if (WARN_ON(!new_inode)) {
9288                                 return ret;
9289                         }
9290                 } else {
9291                         /* maybe -EOVERFLOW */
9292                         return ret;
9293                 }
9294         }
9295         ret = 0;
9296
9297         /*
9298          * we're using rename to replace one file with another.  Start IO on it
9299          * now so  we don't add too much work to the end of the transaction
9300          */
9301         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9302                 filemap_flush(old_inode->i_mapping);
9303
9304         /* close the racy window with snapshot create/destroy ioctl */
9305         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9306                 down_read(&root->fs_info->subvol_sem);
9307         /*
9308          * We want to reserve the absolute worst case amount of items.  So if
9309          * both inodes are subvols and we need to unlink them then that would
9310          * require 4 item modifications, but if they are both normal inodes it
9311          * would require 5 item modifications, so we'll assume their normal
9312          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9313          * should cover the worst case number of items we'll modify.
9314          */
9315         trans = btrfs_start_transaction(root, 11);
9316         if (IS_ERR(trans)) {
9317                 ret = PTR_ERR(trans);
9318                 goto out_notrans;
9319         }
9320
9321         if (dest != root)
9322                 btrfs_record_root_in_trans(trans, dest);
9323
9324         ret = btrfs_set_inode_index(new_dir, &index);
9325         if (ret)
9326                 goto out_fail;
9327
9328         BTRFS_I(old_inode)->dir_index = 0ULL;
9329         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9330                 /* force full log commit if subvolume involved. */
9331                 btrfs_set_log_full_commit(root->fs_info, trans);
9332         } else {
9333                 ret = btrfs_insert_inode_ref(trans, dest,
9334                                              new_dentry->d_name.name,
9335                                              new_dentry->d_name.len,
9336                                              old_ino,
9337                                              btrfs_ino(new_dir), index);
9338                 if (ret)
9339                         goto out_fail;
9340                 /*
9341                  * this is an ugly little race, but the rename is required
9342                  * to make sure that if we crash, the inode is either at the
9343                  * old name or the new one.  pinning the log transaction lets
9344                  * us make sure we don't allow a log commit to come in after
9345                  * we unlink the name but before we add the new name back in.
9346                  */
9347                 btrfs_pin_log_trans(root);
9348         }
9349
9350         inode_inc_iversion(old_dir);
9351         inode_inc_iversion(new_dir);
9352         inode_inc_iversion(old_inode);
9353         old_dir->i_ctime = old_dir->i_mtime = ctime;
9354         new_dir->i_ctime = new_dir->i_mtime = ctime;
9355         old_inode->i_ctime = ctime;
9356
9357         if (old_dentry->d_parent != new_dentry->d_parent)
9358                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9359
9360         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9361                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9362                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9363                                         old_dentry->d_name.name,
9364                                         old_dentry->d_name.len);
9365         } else {
9366                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9367                                         d_inode(old_dentry),
9368                                         old_dentry->d_name.name,
9369                                         old_dentry->d_name.len);
9370                 if (!ret)
9371                         ret = btrfs_update_inode(trans, root, old_inode);
9372         }
9373         if (ret) {
9374                 btrfs_abort_transaction(trans, root, ret);
9375                 goto out_fail;
9376         }
9377
9378         if (new_inode) {
9379                 inode_inc_iversion(new_inode);
9380                 new_inode->i_ctime = CURRENT_TIME;
9381                 if (unlikely(btrfs_ino(new_inode) ==
9382                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9383                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9384                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9385                                                 root_objectid,
9386                                                 new_dentry->d_name.name,
9387                                                 new_dentry->d_name.len);
9388                         BUG_ON(new_inode->i_nlink == 0);
9389                 } else {
9390                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9391                                                  d_inode(new_dentry),
9392                                                  new_dentry->d_name.name,
9393                                                  new_dentry->d_name.len);
9394                 }
9395                 if (!ret && new_inode->i_nlink == 0)
9396                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9397                 if (ret) {
9398                         btrfs_abort_transaction(trans, root, ret);
9399                         goto out_fail;
9400                 }
9401         }
9402
9403         ret = btrfs_add_link(trans, new_dir, old_inode,
9404                              new_dentry->d_name.name,
9405                              new_dentry->d_name.len, 0, index);
9406         if (ret) {
9407                 btrfs_abort_transaction(trans, root, ret);
9408                 goto out_fail;
9409         }
9410
9411         if (old_inode->i_nlink == 1)
9412                 BTRFS_I(old_inode)->dir_index = index;
9413
9414         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
9415                 struct dentry *parent = new_dentry->d_parent;
9416                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9417                 btrfs_end_log_trans(root);
9418         }
9419 out_fail:
9420         btrfs_end_transaction(trans, root);
9421 out_notrans:
9422         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9423                 up_read(&root->fs_info->subvol_sem);
9424
9425         return ret;
9426 }
9427
9428 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9429                          struct inode *new_dir, struct dentry *new_dentry,
9430                          unsigned int flags)
9431 {
9432         if (flags & ~RENAME_NOREPLACE)
9433                 return -EINVAL;
9434
9435         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry);
9436 }
9437
9438 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9439 {
9440         struct btrfs_delalloc_work *delalloc_work;
9441         struct inode *inode;
9442
9443         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9444                                      work);
9445         inode = delalloc_work->inode;
9446         if (delalloc_work->wait) {
9447                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
9448         } else {
9449                 filemap_flush(inode->i_mapping);
9450                 if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9451                              &BTRFS_I(inode)->runtime_flags))
9452                         filemap_flush(inode->i_mapping);
9453         }
9454
9455         if (delalloc_work->delay_iput)
9456                 btrfs_add_delayed_iput(inode);
9457         else
9458                 iput(inode);
9459         complete(&delalloc_work->completion);
9460 }
9461
9462 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9463                                                     int wait, int delay_iput)
9464 {
9465         struct btrfs_delalloc_work *work;
9466
9467         work = kmem_cache_zalloc(btrfs_delalloc_work_cachep, GFP_NOFS);
9468         if (!work)
9469                 return NULL;
9470
9471         init_completion(&work->completion);
9472         INIT_LIST_HEAD(&work->list);
9473         work->inode = inode;
9474         work->wait = wait;
9475         work->delay_iput = delay_iput;
9476         WARN_ON_ONCE(!inode);
9477         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9478                         btrfs_run_delalloc_work, NULL, NULL);
9479
9480         return work;
9481 }
9482
9483 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9484 {
9485         wait_for_completion(&work->completion);
9486         kmem_cache_free(btrfs_delalloc_work_cachep, work);
9487 }
9488
9489 /*
9490  * some fairly slow code that needs optimization. This walks the list
9491  * of all the inodes with pending delalloc and forces them to disk.
9492  */
9493 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9494                                    int nr)
9495 {
9496         struct btrfs_inode *binode;
9497         struct inode *inode;
9498         struct btrfs_delalloc_work *work, *next;
9499         struct list_head works;
9500         struct list_head splice;
9501         int ret = 0;
9502
9503         INIT_LIST_HEAD(&works);
9504         INIT_LIST_HEAD(&splice);
9505
9506         mutex_lock(&root->delalloc_mutex);
9507         spin_lock(&root->delalloc_lock);
9508         list_splice_init(&root->delalloc_inodes, &splice);
9509         while (!list_empty(&splice)) {
9510                 binode = list_entry(splice.next, struct btrfs_inode,
9511                                     delalloc_inodes);
9512
9513                 list_move_tail(&binode->delalloc_inodes,
9514                                &root->delalloc_inodes);
9515                 inode = igrab(&binode->vfs_inode);
9516                 if (!inode) {
9517                         cond_resched_lock(&root->delalloc_lock);
9518                         continue;
9519                 }
9520                 spin_unlock(&root->delalloc_lock);
9521
9522                 work = btrfs_alloc_delalloc_work(inode, 0, delay_iput);
9523                 if (!work) {
9524                         if (delay_iput)
9525                                 btrfs_add_delayed_iput(inode);
9526                         else
9527                                 iput(inode);
9528                         ret = -ENOMEM;
9529                         goto out;
9530                 }
9531                 list_add_tail(&work->list, &works);
9532                 btrfs_queue_work(root->fs_info->flush_workers,
9533                                  &work->work);
9534                 ret++;
9535                 if (nr != -1 && ret >= nr)
9536                         goto out;
9537                 cond_resched();
9538                 spin_lock(&root->delalloc_lock);
9539         }
9540         spin_unlock(&root->delalloc_lock);
9541
9542 out:
9543         list_for_each_entry_safe(work, next, &works, list) {
9544                 list_del_init(&work->list);
9545                 btrfs_wait_and_free_delalloc_work(work);
9546         }
9547
9548         if (!list_empty_careful(&splice)) {
9549                 spin_lock(&root->delalloc_lock);
9550                 list_splice_tail(&splice, &root->delalloc_inodes);
9551                 spin_unlock(&root->delalloc_lock);
9552         }
9553         mutex_unlock(&root->delalloc_mutex);
9554         return ret;
9555 }
9556
9557 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
9558 {
9559         int ret;
9560
9561         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
9562                 return -EROFS;
9563
9564         ret = __start_delalloc_inodes(root, delay_iput, -1);
9565         if (ret > 0)
9566                 ret = 0;
9567         /*
9568          * the filemap_flush will queue IO into the worker threads, but
9569          * we have to make sure the IO is actually started and that
9570          * ordered extents get created before we return
9571          */
9572         atomic_inc(&root->fs_info->async_submit_draining);
9573         while (atomic_read(&root->fs_info->nr_async_submits) ||
9574               atomic_read(&root->fs_info->async_delalloc_pages)) {
9575                 wait_event(root->fs_info->async_submit_wait,
9576                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
9577                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
9578         }
9579         atomic_dec(&root->fs_info->async_submit_draining);
9580         return ret;
9581 }
9582
9583 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
9584                                int nr)
9585 {
9586         struct btrfs_root *root;
9587         struct list_head splice;
9588         int ret;
9589
9590         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9591                 return -EROFS;
9592
9593         INIT_LIST_HEAD(&splice);
9594
9595         mutex_lock(&fs_info->delalloc_root_mutex);
9596         spin_lock(&fs_info->delalloc_root_lock);
9597         list_splice_init(&fs_info->delalloc_roots, &splice);
9598         while (!list_empty(&splice) && nr) {
9599                 root = list_first_entry(&splice, struct btrfs_root,
9600                                         delalloc_root);
9601                 root = btrfs_grab_fs_root(root);
9602                 BUG_ON(!root);
9603                 list_move_tail(&root->delalloc_root,
9604                                &fs_info->delalloc_roots);
9605                 spin_unlock(&fs_info->delalloc_root_lock);
9606
9607                 ret = __start_delalloc_inodes(root, delay_iput, nr);
9608                 btrfs_put_fs_root(root);
9609                 if (ret < 0)
9610                         goto out;
9611
9612                 if (nr != -1) {
9613                         nr -= ret;
9614                         WARN_ON(nr < 0);
9615                 }
9616                 spin_lock(&fs_info->delalloc_root_lock);
9617         }
9618         spin_unlock(&fs_info->delalloc_root_lock);
9619
9620         ret = 0;
9621         atomic_inc(&fs_info->async_submit_draining);
9622         while (atomic_read(&fs_info->nr_async_submits) ||
9623               atomic_read(&fs_info->async_delalloc_pages)) {
9624                 wait_event(fs_info->async_submit_wait,
9625                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
9626                     atomic_read(&fs_info->async_delalloc_pages) == 0));
9627         }
9628         atomic_dec(&fs_info->async_submit_draining);
9629 out:
9630         if (!list_empty_careful(&splice)) {
9631                 spin_lock(&fs_info->delalloc_root_lock);
9632                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9633                 spin_unlock(&fs_info->delalloc_root_lock);
9634         }
9635         mutex_unlock(&fs_info->delalloc_root_mutex);
9636         return ret;
9637 }
9638
9639 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9640                          const char *symname)
9641 {
9642         struct btrfs_trans_handle *trans;
9643         struct btrfs_root *root = BTRFS_I(dir)->root;
9644         struct btrfs_path *path;
9645         struct btrfs_key key;
9646         struct inode *inode = NULL;
9647         int err;
9648         int drop_inode = 0;
9649         u64 objectid;
9650         u64 index = 0;
9651         int name_len;
9652         int datasize;
9653         unsigned long ptr;
9654         struct btrfs_file_extent_item *ei;
9655         struct extent_buffer *leaf;
9656
9657         name_len = strlen(symname);
9658         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
9659                 return -ENAMETOOLONG;
9660
9661         /*
9662          * 2 items for inode item and ref
9663          * 2 items for dir items
9664          * 1 item for updating parent inode item
9665          * 1 item for the inline extent item
9666          * 1 item for xattr if selinux is on
9667          */
9668         trans = btrfs_start_transaction(root, 7);
9669         if (IS_ERR(trans))
9670                 return PTR_ERR(trans);
9671
9672         err = btrfs_find_free_ino(root, &objectid);
9673         if (err)
9674                 goto out_unlock;
9675
9676         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9677                                 dentry->d_name.len, btrfs_ino(dir), objectid,
9678                                 S_IFLNK|S_IRWXUGO, &index);
9679         if (IS_ERR(inode)) {
9680                 err = PTR_ERR(inode);
9681                 goto out_unlock;
9682         }
9683
9684         /*
9685         * If the active LSM wants to access the inode during
9686         * d_instantiate it needs these. Smack checks to see
9687         * if the filesystem supports xattrs by looking at the
9688         * ops vector.
9689         */
9690         inode->i_fop = &btrfs_file_operations;
9691         inode->i_op = &btrfs_file_inode_operations;
9692         inode->i_mapping->a_ops = &btrfs_aops;
9693         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9694
9695         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9696         if (err)
9697                 goto out_unlock_inode;
9698
9699         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
9700         if (err)
9701                 goto out_unlock_inode;
9702
9703         path = btrfs_alloc_path();
9704         if (!path) {
9705                 err = -ENOMEM;
9706                 goto out_unlock_inode;
9707         }
9708         key.objectid = btrfs_ino(inode);
9709         key.offset = 0;
9710         key.type = BTRFS_EXTENT_DATA_KEY;
9711         datasize = btrfs_file_extent_calc_inline_size(name_len);
9712         err = btrfs_insert_empty_item(trans, root, path, &key,
9713                                       datasize);
9714         if (err) {
9715                 btrfs_free_path(path);
9716                 goto out_unlock_inode;
9717         }
9718         leaf = path->nodes[0];
9719         ei = btrfs_item_ptr(leaf, path->slots[0],
9720                             struct btrfs_file_extent_item);
9721         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9722         btrfs_set_file_extent_type(leaf, ei,
9723                                    BTRFS_FILE_EXTENT_INLINE);
9724         btrfs_set_file_extent_encryption(leaf, ei, 0);
9725         btrfs_set_file_extent_compression(leaf, ei, 0);
9726         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9727         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9728
9729         ptr = btrfs_file_extent_inline_start(ei);
9730         write_extent_buffer(leaf, symname, ptr, name_len);
9731         btrfs_mark_buffer_dirty(leaf);
9732         btrfs_free_path(path);
9733
9734         inode->i_op = &btrfs_symlink_inode_operations;
9735         inode->i_mapping->a_ops = &btrfs_symlink_aops;
9736         inode_set_bytes(inode, name_len);
9737         btrfs_i_size_write(inode, name_len);
9738         err = btrfs_update_inode(trans, root, inode);
9739         if (err) {
9740                 drop_inode = 1;
9741                 goto out_unlock_inode;
9742         }
9743
9744         unlock_new_inode(inode);
9745         d_instantiate(dentry, inode);
9746
9747 out_unlock:
9748         btrfs_end_transaction(trans, root);
9749         if (drop_inode) {
9750                 inode_dec_link_count(inode);
9751                 iput(inode);
9752         }
9753         btrfs_btree_balance_dirty(root);
9754         return err;
9755
9756 out_unlock_inode:
9757         drop_inode = 1;
9758         unlock_new_inode(inode);
9759         goto out_unlock;
9760 }
9761
9762 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9763                                        u64 start, u64 num_bytes, u64 min_size,
9764                                        loff_t actual_len, u64 *alloc_hint,
9765                                        struct btrfs_trans_handle *trans)
9766 {
9767         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9768         struct extent_map *em;
9769         struct btrfs_root *root = BTRFS_I(inode)->root;
9770         struct btrfs_key ins;
9771         u64 cur_offset = start;
9772         u64 i_size;
9773         u64 cur_bytes;
9774         u64 last_alloc = (u64)-1;
9775         int ret = 0;
9776         bool own_trans = true;
9777
9778         if (trans)
9779                 own_trans = false;
9780         while (num_bytes > 0) {
9781                 if (own_trans) {
9782                         trans = btrfs_start_transaction(root, 3);
9783                         if (IS_ERR(trans)) {
9784                                 ret = PTR_ERR(trans);
9785                                 break;
9786                         }
9787                 }
9788
9789                 cur_bytes = min(num_bytes, 256ULL * 1024 * 1024);
9790                 cur_bytes = max(cur_bytes, min_size);
9791                 /*
9792                  * If we are severely fragmented we could end up with really
9793                  * small allocations, so if the allocator is returning small
9794                  * chunks lets make its job easier by only searching for those
9795                  * sized chunks.
9796                  */
9797                 cur_bytes = min(cur_bytes, last_alloc);
9798                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
9799                                            *alloc_hint, &ins, 1, 0);
9800                 if (ret) {
9801                         if (own_trans)
9802                                 btrfs_end_transaction(trans, root);
9803                         break;
9804                 }
9805
9806                 last_alloc = ins.offset;
9807                 ret = insert_reserved_file_extent(trans, inode,
9808                                                   cur_offset, ins.objectid,
9809                                                   ins.offset, ins.offset,
9810                                                   ins.offset, 0, 0, 0,
9811                                                   BTRFS_FILE_EXTENT_PREALLOC);
9812                 if (ret) {
9813                         btrfs_free_reserved_extent(root, ins.objectid,
9814                                                    ins.offset, 0);
9815                         btrfs_abort_transaction(trans, root, ret);
9816                         if (own_trans)
9817                                 btrfs_end_transaction(trans, root);
9818                         break;
9819                 }
9820
9821                 btrfs_drop_extent_cache(inode, cur_offset,
9822                                         cur_offset + ins.offset -1, 0);
9823
9824                 em = alloc_extent_map();
9825                 if (!em) {
9826                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9827                                 &BTRFS_I(inode)->runtime_flags);
9828                         goto next;
9829                 }
9830
9831                 em->start = cur_offset;
9832                 em->orig_start = cur_offset;
9833                 em->len = ins.offset;
9834                 em->block_start = ins.objectid;
9835                 em->block_len = ins.offset;
9836                 em->orig_block_len = ins.offset;
9837                 em->ram_bytes = ins.offset;
9838                 em->bdev = root->fs_info->fs_devices->latest_bdev;
9839                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9840                 em->generation = trans->transid;
9841
9842                 while (1) {
9843                         write_lock(&em_tree->lock);
9844                         ret = add_extent_mapping(em_tree, em, 1);
9845                         write_unlock(&em_tree->lock);
9846                         if (ret != -EEXIST)
9847                                 break;
9848                         btrfs_drop_extent_cache(inode, cur_offset,
9849                                                 cur_offset + ins.offset - 1,
9850                                                 0);
9851                 }
9852                 free_extent_map(em);
9853 next:
9854                 num_bytes -= ins.offset;
9855                 cur_offset += ins.offset;
9856                 *alloc_hint = ins.objectid + ins.offset;
9857
9858                 inode_inc_iversion(inode);
9859                 inode->i_ctime = CURRENT_TIME;
9860                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9861                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9862                     (actual_len > inode->i_size) &&
9863                     (cur_offset > inode->i_size)) {
9864                         if (cur_offset > actual_len)
9865                                 i_size = actual_len;
9866                         else
9867                                 i_size = cur_offset;
9868                         i_size_write(inode, i_size);
9869                         btrfs_ordered_update_i_size(inode, i_size, NULL);
9870                 }
9871
9872                 ret = btrfs_update_inode(trans, root, inode);
9873
9874                 if (ret) {
9875                         btrfs_abort_transaction(trans, root, ret);
9876                         if (own_trans)
9877                                 btrfs_end_transaction(trans, root);
9878                         break;
9879                 }
9880
9881                 if (own_trans)
9882                         btrfs_end_transaction(trans, root);
9883         }
9884         return ret;
9885 }
9886
9887 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9888                               u64 start, u64 num_bytes, u64 min_size,
9889                               loff_t actual_len, u64 *alloc_hint)
9890 {
9891         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9892                                            min_size, actual_len, alloc_hint,
9893                                            NULL);
9894 }
9895
9896 int btrfs_prealloc_file_range_trans(struct inode *inode,
9897                                     struct btrfs_trans_handle *trans, int mode,
9898                                     u64 start, u64 num_bytes, u64 min_size,
9899                                     loff_t actual_len, u64 *alloc_hint)
9900 {
9901         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9902                                            min_size, actual_len, alloc_hint, trans);
9903 }
9904
9905 static int btrfs_set_page_dirty(struct page *page)
9906 {
9907         return __set_page_dirty_nobuffers(page);
9908 }
9909
9910 static int btrfs_permission(struct inode *inode, int mask)
9911 {
9912         struct btrfs_root *root = BTRFS_I(inode)->root;
9913         umode_t mode = inode->i_mode;
9914
9915         if (mask & MAY_WRITE &&
9916             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9917                 if (btrfs_root_readonly(root))
9918                         return -EROFS;
9919                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9920                         return -EACCES;
9921         }
9922         return generic_permission(inode, mask);
9923 }
9924
9925 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9926 {
9927         struct btrfs_trans_handle *trans;
9928         struct btrfs_root *root = BTRFS_I(dir)->root;
9929         struct inode *inode = NULL;
9930         u64 objectid;
9931         u64 index;
9932         int ret = 0;
9933
9934         /*
9935          * 5 units required for adding orphan entry
9936          */
9937         trans = btrfs_start_transaction(root, 5);
9938         if (IS_ERR(trans))
9939                 return PTR_ERR(trans);
9940
9941         ret = btrfs_find_free_ino(root, &objectid);
9942         if (ret)
9943                 goto out;
9944
9945         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9946                                 btrfs_ino(dir), objectid, mode, &index);
9947         if (IS_ERR(inode)) {
9948                 ret = PTR_ERR(inode);
9949                 inode = NULL;
9950                 goto out;
9951         }
9952
9953         inode->i_fop = &btrfs_file_operations;
9954         inode->i_op = &btrfs_file_inode_operations;
9955
9956         inode->i_mapping->a_ops = &btrfs_aops;
9957         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9958
9959         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9960         if (ret)
9961                 goto out_inode;
9962
9963         ret = btrfs_update_inode(trans, root, inode);
9964         if (ret)
9965                 goto out_inode;
9966         ret = btrfs_orphan_add(trans, inode);
9967         if (ret)
9968                 goto out_inode;
9969
9970         /*
9971          * We set number of links to 0 in btrfs_new_inode(), and here we set
9972          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9973          * through:
9974          *
9975          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9976          */
9977         set_nlink(inode, 1);
9978         unlock_new_inode(inode);
9979         d_tmpfile(dentry, inode);
9980         mark_inode_dirty(inode);
9981
9982 out:
9983         btrfs_end_transaction(trans, root);
9984         if (ret)
9985                 iput(inode);
9986         btrfs_balance_delayed_items(root);
9987         btrfs_btree_balance_dirty(root);
9988         return ret;
9989
9990 out_inode:
9991         unlock_new_inode(inode);
9992         goto out;
9993
9994 }
9995
9996 /* Inspired by filemap_check_errors() */
9997 int btrfs_inode_check_errors(struct inode *inode)
9998 {
9999         int ret = 0;
10000
10001         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10002             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10003                 ret = -ENOSPC;
10004         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10005             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10006                 ret = -EIO;
10007
10008         return ret;
10009 }
10010
10011 static const struct inode_operations btrfs_dir_inode_operations = {
10012         .getattr        = btrfs_getattr,
10013         .lookup         = btrfs_lookup,
10014         .create         = btrfs_create,
10015         .unlink         = btrfs_unlink,
10016         .link           = btrfs_link,
10017         .mkdir          = btrfs_mkdir,
10018         .rmdir          = btrfs_rmdir,
10019         .rename2        = btrfs_rename2,
10020         .symlink        = btrfs_symlink,
10021         .setattr        = btrfs_setattr,
10022         .mknod          = btrfs_mknod,
10023         .setxattr       = btrfs_setxattr,
10024         .getxattr       = btrfs_getxattr,
10025         .listxattr      = btrfs_listxattr,
10026         .removexattr    = btrfs_removexattr,
10027         .permission     = btrfs_permission,
10028         .get_acl        = btrfs_get_acl,
10029         .set_acl        = btrfs_set_acl,
10030         .update_time    = btrfs_update_time,
10031         .tmpfile        = btrfs_tmpfile,
10032 };
10033 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10034         .lookup         = btrfs_lookup,
10035         .permission     = btrfs_permission,
10036         .get_acl        = btrfs_get_acl,
10037         .set_acl        = btrfs_set_acl,
10038         .update_time    = btrfs_update_time,
10039 };
10040
10041 static const struct file_operations btrfs_dir_file_operations = {
10042         .llseek         = generic_file_llseek,
10043         .read           = generic_read_dir,
10044         .iterate        = btrfs_real_readdir,
10045         .unlocked_ioctl = btrfs_ioctl,
10046 #ifdef CONFIG_COMPAT
10047         .compat_ioctl   = btrfs_ioctl,
10048 #endif
10049         .release        = btrfs_release_file,
10050         .fsync          = btrfs_sync_file,
10051 };
10052
10053 static struct extent_io_ops btrfs_extent_io_ops = {
10054         .fill_delalloc = run_delalloc_range,
10055         .submit_bio_hook = btrfs_submit_bio_hook,
10056         .merge_bio_hook = btrfs_merge_bio_hook,
10057         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10058         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10059         .writepage_start_hook = btrfs_writepage_start_hook,
10060         .set_bit_hook = btrfs_set_bit_hook,
10061         .clear_bit_hook = btrfs_clear_bit_hook,
10062         .merge_extent_hook = btrfs_merge_extent_hook,
10063         .split_extent_hook = btrfs_split_extent_hook,
10064 };
10065
10066 /*
10067  * btrfs doesn't support the bmap operation because swapfiles
10068  * use bmap to make a mapping of extents in the file.  They assume
10069  * these extents won't change over the life of the file and they
10070  * use the bmap result to do IO directly to the drive.
10071  *
10072  * the btrfs bmap call would return logical addresses that aren't
10073  * suitable for IO and they also will change frequently as COW
10074  * operations happen.  So, swapfile + btrfs == corruption.
10075  *
10076  * For now we're avoiding this by dropping bmap.
10077  */
10078 static const struct address_space_operations btrfs_aops = {
10079         .readpage       = btrfs_readpage,
10080         .writepage      = btrfs_writepage,
10081         .writepages     = btrfs_writepages,
10082         .readpages      = btrfs_readpages,
10083         .direct_IO      = btrfs_direct_IO,
10084         .invalidatepage = btrfs_invalidatepage,
10085         .releasepage    = btrfs_releasepage,
10086         .set_page_dirty = btrfs_set_page_dirty,
10087         .error_remove_page = generic_error_remove_page,
10088 };
10089
10090 static const struct address_space_operations btrfs_symlink_aops = {
10091         .readpage       = btrfs_readpage,
10092         .writepage      = btrfs_writepage,
10093         .invalidatepage = btrfs_invalidatepage,
10094         .releasepage    = btrfs_releasepage,
10095 };
10096
10097 static const struct inode_operations btrfs_file_inode_operations = {
10098         .getattr        = btrfs_getattr,
10099         .setattr        = btrfs_setattr,
10100         .setxattr       = btrfs_setxattr,
10101         .getxattr       = btrfs_getxattr,
10102         .listxattr      = btrfs_listxattr,
10103         .removexattr    = btrfs_removexattr,
10104         .permission     = btrfs_permission,
10105         .fiemap         = btrfs_fiemap,
10106         .get_acl        = btrfs_get_acl,
10107         .set_acl        = btrfs_set_acl,
10108         .update_time    = btrfs_update_time,
10109 };
10110 static const struct inode_operations btrfs_special_inode_operations = {
10111         .getattr        = btrfs_getattr,
10112         .setattr        = btrfs_setattr,
10113         .permission     = btrfs_permission,
10114         .setxattr       = btrfs_setxattr,
10115         .getxattr       = btrfs_getxattr,
10116         .listxattr      = btrfs_listxattr,
10117         .removexattr    = btrfs_removexattr,
10118         .get_acl        = btrfs_get_acl,
10119         .set_acl        = btrfs_set_acl,
10120         .update_time    = btrfs_update_time,
10121 };
10122 static const struct inode_operations btrfs_symlink_inode_operations = {
10123         .readlink       = generic_readlink,
10124         .follow_link    = page_follow_link_light,
10125         .put_link       = page_put_link,
10126         .getattr        = btrfs_getattr,
10127         .setattr        = btrfs_setattr,
10128         .permission     = btrfs_permission,
10129         .setxattr       = btrfs_setxattr,
10130         .getxattr       = btrfs_getxattr,
10131         .listxattr      = btrfs_listxattr,
10132         .removexattr    = btrfs_removexattr,
10133         .update_time    = btrfs_update_time,
10134 };
10135
10136 const struct dentry_operations btrfs_dentry_operations = {
10137         .d_delete       = btrfs_dentry_delete,
10138         .d_release      = btrfs_dentry_release,
10139 };