kernel/fs/ceph/addr.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/backing-dev.h>
   4 #include <linux/fs.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/writeback.h>    /* generic_writepages */
   8 #include <linux/slab.h>
   9 #include <linux/pagevec.h>
  10 #include <linux/task_io_accounting_ops.h>
  11
  12 #include "super.h"
  13 #include "mds_client.h"
  14 #include "cache.h"
  15 #include <linux/ceph/osd_client.h>
  16
  17 /*
  18  * Ceph address space ops.
  19  *
  20  * There are a few funny things going on here.
  21  *
  22  * The page->private field is used to reference a struct
  23  * ceph_snap_context for _every_ dirty page.  This indicates which
  24  * snapshot the page was logically dirtied in, and thus which snap
  25  * context needs to be associated with the osd write during writeback.
  26  *
  27  * Similarly, struct ceph_inode_info maintains a set of counters to
  28  * count dirty pages on the inode.  In the absence of snapshots,
  29  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
  30  *
  31  * When a snapshot is taken (that is, when the client receives
  32  * notification that a snapshot was taken), each inode with caps and
  33  * with dirty pages (dirty pages implies there is a cap) gets a new
  34  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
  35  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
  36  * moved to capsnap->dirty. (Unless a sync write is currently in
  37  * progress.  In that case, the capsnap is said to be "pending", new
  38  * writes cannot start, and the capsnap isn't "finalized" until the
  39  * write completes (or fails) and a final size/mtime for the inode for
  40  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
  41  *
  42  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
  43  * we look for the first capsnap in i_cap_snaps and write out pages in
  44  * that snap context _only_.  Then we move on to the next capsnap,
  45  * eventually reaching the "live" or "head" context (i.e., pages that
  46  * are not yet snapped) and are writing the most recently dirtied
  47  * pages.
  48  *
  49  * Invalidate and so forth must take care to ensure the dirty page
  50  * accounting is preserved.
  51  */
  52
  53 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
  54 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
  55         (CONGESTION_ON_THRESH(congestion_kb) -                          \
  56          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
  57
  58 static inline struct ceph_snap_context *page_snap_context(struct page *page)
  59 {
  60         if (PagePrivate(page))
  61                 return (void *)page->private;
  62         return NULL;
  63 }
  64
  65 /*
  66  * Dirty a page.  Optimistically adjust accounting, on the assumption
  67  * that we won't race with invalidate.  If we do, readjust.
  68  */
  69 static int ceph_set_page_dirty(struct page *page)
  70 {
  71         struct address_space *mapping = page->mapping;
  72         struct inode *inode;
  73         struct ceph_inode_info *ci;
  74         struct ceph_snap_context *snapc;
  75         int ret;
  76
  77         if (unlikely(!mapping))
  78                 return !TestSetPageDirty(page);
  79
  80         if (PageDirty(page)) {
  81                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
  82                      mapping->host, page, page->index);
  83                 BUG_ON(!PagePrivate(page));
  84                 return 0;
  85         }
  86
  87         inode = mapping->host;
  88         ci = ceph_inode(inode);
  89
  90         /*
  91          * Note that we're grabbing a snapc ref here without holding
  92          * any locks!
  93          */
  94         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
  95
  96         /* dirty the head */
  97         spin_lock(&ci->i_ceph_lock);
  98         if (ci->i_head_snapc == NULL)
  99                 ci->i_head_snapc = ceph_get_snap_context(snapc);
 100         ++ci->i_wrbuffer_ref_head;
 101         if (ci->i_wrbuffer_ref == 0)
 102                 ihold(inode);
 103         ++ci->i_wrbuffer_ref;
 104         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
 105              "snapc %p seq %lld (%d snaps)\n",
 106              mapping->host, page, page->index,
 107              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
 108              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
 109              snapc, snapc->seq, snapc->num_snaps);
 110         spin_unlock(&ci->i_ceph_lock);
 111
 112         /*
 113          * Reference snap context in page->private.  Also set
 114          * PagePrivate so that we get invalidatepage callback.
 115          */
 116         BUG_ON(PagePrivate(page));
 117         page->private = (unsigned long)snapc;
 118         SetPagePrivate(page);
 119
 120         ret = __set_page_dirty_nobuffers(page);
 121         WARN_ON(!PageLocked(page));
 122         WARN_ON(!page->mapping);
 123
 124         return ret;
 125 }
 126
 127 /*
 128  * If we are truncating the full page (i.e. offset == 0), adjust the
 129  * dirty page counters appropriately.  Only called if there is private
 130  * data on the page.
 131  */
 132 static void ceph_invalidatepage(struct page *page, unsigned int offset,
 133                                 unsigned int length)
 134 {
 135         struct inode *inode;
 136         struct ceph_inode_info *ci;
 137         struct ceph_snap_context *snapc = page_snap_context(page);
 138
 139         inode = page->mapping->host;
 140         ci = ceph_inode(inode);
 141
 142         if (offset != 0 || length != PAGE_CACHE_SIZE) {
 143                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
 144                      inode, page, page->index, offset, length);
 145                 return;
 146         }
 147
 148         ceph_invalidate_fscache_page(inode, page);
 149
 150         if (!PagePrivate(page))
 151                 return;
 152
 153         /*
 154          * We can get non-dirty pages here due to races between
 155          * set_page_dirty and truncate_complete_page; just spit out a
 156          * warning, in case we end up with accounting problems later.
 157          */
 158         if (!PageDirty(page))
 159                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
 160
 161         ClearPageChecked(page);
 162
 163         dout("%p invalidatepage %p idx %lu full dirty page\n",
 164              inode, page, page->index);
 165
 166         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 167         ceph_put_snap_context(snapc);
 168         page->private = 0;
 169         ClearPagePrivate(page);
 170 }
 171
 172 static int ceph_releasepage(struct page *page, gfp_t g)
 173 {
 174         struct inode *inode = page->mapping ? page->mapping->host : NULL;
 175         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
 176         WARN_ON(PageDirty(page));
 177
 178         /* Can we release the page from the cache? */
 179         if (!ceph_release_fscache_page(page, g))
 180                 return 0;
 181
 182         return !PagePrivate(page);
 183 }
 184
 185 /*
 186  * read a single page, without unlocking it.
 187  */
 188 static int readpage_nounlock(struct file *filp, struct page *page)
 189 {
 190         struct inode *inode = file_inode(filp);
 191         struct ceph_inode_info *ci = ceph_inode(inode);
 192         struct ceph_osd_client *osdc =
 193                 &ceph_inode_to_client(inode)->client->osdc;
 194         int err = 0;
 195         u64 off = page_offset(page);
 196         u64 len = PAGE_CACHE_SIZE;
 197
 198         if (off >= i_size_read(inode)) {
 199                 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
 200                 SetPageUptodate(page);
 201                 return 0;
 202         }
 203
 204         if (ci->i_inline_version != CEPH_INLINE_NONE) {
 205                 /*
 206                  * Uptodate inline data should have been added
 207                  * into page cache while getting Fcr caps.
 208                  */
 209                 if (off == 0)
 210                         return -EINVAL;
 211                 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
 212                 SetPageUptodate(page);
 213                 return 0;
 214         }
 215
 216         err = ceph_readpage_from_fscache(inode, page);
 217         if (err == 0)
 218                 goto out;
 219
 220         dout("readpage inode %p file %p page %p index %lu\n",
 221              inode, filp, page, page->index);
 222         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
 223                                   off, &len,
 224                                   ci->i_truncate_seq, ci->i_truncate_size,
 225                                   &page, 1, 0);
 226         if (err == -ENOENT)
 227                 err = 0;
 228         if (err < 0) {
 229                 SetPageError(page);
 230                 ceph_fscache_readpage_cancel(inode, page);
 231                 goto out;
 232         }
 233         if (err < PAGE_CACHE_SIZE)
 234                 /* zero fill remainder of page */
 235                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
 236         else
 237                 flush_dcache_page(page);
 238
 239         SetPageUptodate(page);
 240         ceph_readpage_to_fscache(inode, page);
 241
 242 out:
 243         return err < 0 ? err : 0;
 244 }
 245
 246 static int ceph_readpage(struct file *filp, struct page *page)
 247 {
 248         int r = readpage_nounlock(filp, page);
 249         unlock_page(page);
 250         return r;
 251 }
 252
 253 /*
 254  * Finish an async read(ahead) op.
 255  */
 256 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
 257 {
 258         struct inode *inode = req->r_inode;
 259         struct ceph_osd_data *osd_data;
 260         int rc = req->r_result;
 261         int bytes = le32_to_cpu(msg->hdr.data_len);
 262         int num_pages;
 263         int i;
 264
 265         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
 266
 267         /* unlock all pages, zeroing any data we didn't read */
 268         osd_data = osd_req_op_extent_osd_data(req, 0);
 269         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 270         num_pages = calc_pages_for((u64)osd_data->alignment,
 271                                         (u64)osd_data->length);
 272         for (i = 0; i < num_pages; i++) {
 273                 struct page *page = osd_data->pages[i];
 274
 275                 if (rc < 0)
 276                         goto unlock;
 277                 if (bytes < (int)PAGE_CACHE_SIZE) {
 278                         /* zero (remainder of) page */
 279                         int s = bytes < 0 ? 0 : bytes;
 280                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
 281                 }
 282                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
 283                      page->index);
 284                 flush_dcache_page(page);
 285                 SetPageUptodate(page);
 286                 ceph_readpage_to_fscache(inode, page);
 287 unlock:
 288                 unlock_page(page);
 289                 page_cache_release(page);
 290                 bytes -= PAGE_CACHE_SIZE;
 291         }
 292         kfree(osd_data->pages);
 293 }
 294
 295 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
 296 {
 297         int i;
 298
 299         for (i = 0; i < num_pages; i++)
 300                 unlock_page(pages[i]);
 301 }
 302
 303 /*
 304  * start an async read(ahead) operation.  return nr_pages we submitted
 305  * a read for on success, or negative error code.
 306  */
 307 static int start_read(struct inode *inode, struct list_head *page_list, int max)
 308 {
 309         struct ceph_osd_client *osdc =
 310                 &ceph_inode_to_client(inode)->client->osdc;
 311         struct ceph_inode_info *ci = ceph_inode(inode);
 312         struct page *page = list_entry(page_list->prev, struct page, lru);
 313         struct ceph_vino vino;
 314         struct ceph_osd_request *req;
 315         u64 off;
 316         u64 len;
 317         int i;
 318         struct page **pages;
 319         pgoff_t next_index;
 320         int nr_pages = 0;
 321         int ret;
 322
 323         off = (u64) page_offset(page);
 324
 325         /* count pages */
 326         next_index = page->index;
 327         list_for_each_entry_reverse(page, page_list, lru) {
 328                 if (page->index != next_index)
 329                         break;
 330                 nr_pages++;
 331                 next_index++;
 332                 if (max && nr_pages == max)
 333                         break;
 334         }
 335         len = nr_pages << PAGE_CACHE_SHIFT;
 336         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
 337              off, len);
 338         vino = ceph_vino(inode);
 339         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
 340                                     0, 1, CEPH_OSD_OP_READ,
 341                                     CEPH_OSD_FLAG_READ, NULL,
 342                                     ci->i_truncate_seq, ci->i_truncate_size,
 343                                     false);
 344         if (IS_ERR(req))
 345                 return PTR_ERR(req);
 346
 347         /* build page vector */
 348         nr_pages = calc_pages_for(0, len);
 349         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
 350         ret = -ENOMEM;
 351         if (!pages)
 352                 goto out;
 353         for (i = 0; i < nr_pages; ++i) {
 354                 page = list_entry(page_list->prev, struct page, lru);
 355                 BUG_ON(PageLocked(page));
 356                 list_del(&page->lru);
 357
 358                 dout("start_read %p adding %p idx %lu\n", inode, page,
 359                      page->index);
 360                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
 361                                           GFP_NOFS)) {
 362                         ceph_fscache_uncache_page(inode, page);
 363                         page_cache_release(page);
 364                         dout("start_read %p add_to_page_cache failed %p\n",
 365                              inode, page);
 366                         nr_pages = i;
 367                         goto out_pages;
 368                 }
 369                 pages[i] = page;
 370         }
 371         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
 372         req->r_callback = finish_read;
 373         req->r_inode = inode;
 374
 375         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
 376
 377         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
 378         ret = ceph_osdc_start_request(osdc, req, false);
 379         if (ret < 0)
 380                 goto out_pages;
 381         ceph_osdc_put_request(req);
 382         return nr_pages;
 383
 384 out_pages:
 385         ceph_unlock_page_vector(pages, nr_pages);
 386         ceph_release_page_vector(pages, nr_pages);
 387 out:
 388         ceph_osdc_put_request(req);
 389         return ret;
 390 }
 391
 392
 393 /*
 394  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
 395  * the caller (VM) cleans them up.
 396  */
 397 static int ceph_readpages(struct file *file, struct address_space *mapping,
 398                           struct list_head *page_list, unsigned nr_pages)
 399 {
 400         struct inode *inode = file_inode(file);
 401         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 402         int rc = 0;
 403         int max = 0;
 404
 405         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
 406                 return -EINVAL;
 407
 408         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
 409                                          &nr_pages);
 410
 411         if (rc == 0)
 412                 goto out;
 413
 414         if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
 415                 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
 416                         >> PAGE_SHIFT;
 417
 418         dout("readpages %p file %p nr_pages %d max %d\n", inode,
 419                 file, nr_pages,
 420              max);
 421         while (!list_empty(page_list)) {
 422                 rc = start_read(inode, page_list, max);
 423                 if (rc < 0)
 424                         goto out;
 425                 BUG_ON(rc == 0);
 426         }
 427 out:
 428         ceph_fscache_readpages_cancel(inode, page_list);
 429
 430         dout("readpages %p file %p ret %d\n", inode, file, rc);
 431         return rc;
 432 }
 433
 434 /*
 435  * Get ref for the oldest snapc for an inode with dirty data... that is, the
 436  * only snap context we are allowed to write back.
 437  */
 438 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
 439                                                     u64 *snap_size)
 440 {
 441         struct ceph_inode_info *ci = ceph_inode(inode);
 442         struct ceph_snap_context *snapc = NULL;
 443         struct ceph_cap_snap *capsnap = NULL;
 444
 445         spin_lock(&ci->i_ceph_lock);
 446         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
 447                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
 448                      capsnap->context, capsnap->dirty_pages);
 449                 if (capsnap->dirty_pages) {
 450                         snapc = ceph_get_snap_context(capsnap->context);
 451                         if (snap_size)
 452                                 *snap_size = capsnap->size;
 453                         break;
 454                 }
 455         }
 456         if (!snapc && ci->i_wrbuffer_ref_head) {
 457                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 458                 dout(" head snapc %p has %d dirty pages\n",
 459                      snapc, ci->i_wrbuffer_ref_head);
 460         }
 461         spin_unlock(&ci->i_ceph_lock);
 462         return snapc;
 463 }
 464
 465 /*
 466  * Write a single page, but leave the page locked.
 467  *
 468  * If we get a write error, set the page error bit, but still adjust the
 469  * dirty page accounting (i.e., page is no longer dirty).
 470  */
 471 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
 472 {
 473         struct inode *inode;
 474         struct ceph_inode_info *ci;
 475         struct ceph_fs_client *fsc;
 476         struct ceph_osd_client *osdc;
 477         struct ceph_snap_context *snapc, *oldest;
 478         loff_t page_off = page_offset(page);
 479         long writeback_stat;
 480         u64 truncate_size, snap_size = 0;
 481         u32 truncate_seq;
 482         int err = 0, len = PAGE_CACHE_SIZE;
 483
 484         dout("writepage %p idx %lu\n", page, page->index);
 485
 486         if (!page->mapping || !page->mapping->host) {
 487                 dout("writepage %p - no mapping\n", page);
 488                 return -EFAULT;
 489         }
 490         inode = page->mapping->host;
 491         ci = ceph_inode(inode);
 492         fsc = ceph_inode_to_client(inode);
 493         osdc = &fsc->client->osdc;
 494
 495         /* verify this is a writeable snap context */
 496         snapc = page_snap_context(page);
 497         if (snapc == NULL) {
 498                 dout("writepage %p page %p not dirty?\n", inode, page);
 499                 goto out;
 500         }
 501         oldest = get_oldest_context(inode, &snap_size);
 502         if (snapc->seq > oldest->seq) {
 503                 dout("writepage %p page %p snapc %p not writeable - noop\n",
 504                      inode, page, snapc);
 505                 /* we should only noop if called by kswapd */
 506                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
 507                 ceph_put_snap_context(oldest);
 508                 goto out;
 509         }
 510         ceph_put_snap_context(oldest);
 511
 512         spin_lock(&ci->i_ceph_lock);
 513         truncate_seq = ci->i_truncate_seq;
 514         truncate_size = ci->i_truncate_size;
 515         if (!snap_size)
 516                 snap_size = i_size_read(inode);
 517         spin_unlock(&ci->i_ceph_lock);
 518
 519         /* is this a partial page at end of file? */
 520         if (page_off >= snap_size) {
 521                 dout("%p page eof %llu\n", page, snap_size);
 522                 goto out;
 523         }
 524         if (snap_size < page_off + len)
 525                 len = snap_size - page_off;
 526
 527         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
 528              inode, page, page->index, page_off, len, snapc);
 529
 530         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
 531         if (writeback_stat >
 532             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
 533                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
 534
 535         ceph_readpage_to_fscache(inode, page);
 536
 537         set_page_writeback(page);
 538         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
 539                                    &ci->i_layout, snapc,
 540                                    page_off, len,
 541                                    truncate_seq, truncate_size,
 542                                    &inode->i_mtime, &page, 1);
 543         if (err < 0) {
 544                 dout("writepage setting page/mapping error %d %p\n", err, page);
 545                 SetPageError(page);
 546                 mapping_set_error(&inode->i_data, err);
 547                 if (wbc)
 548                         wbc->pages_skipped++;
 549         } else {
 550                 dout("writepage cleaned page %p\n", page);
 551                 err = 0;  /* vfs expects us to return 0 */
 552         }
 553         page->private = 0;
 554         ClearPagePrivate(page);
 555         end_page_writeback(page);
 556         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
 557         ceph_put_snap_context(snapc);  /* page's reference */
 558 out:
 559         return err;
 560 }
 561
 562 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
 563 {
 564         int err;
 565         struct inode *inode = page->mapping->host;
 566         BUG_ON(!inode);
 567         ihold(inode);
 568         err = writepage_nounlock(page, wbc);
 569         unlock_page(page);
 570         iput(inode);
 571         return err;
 572 }
 573
 574
 575 /*
 576  * lame release_pages helper.  release_pages() isn't exported to
 577  * modules.
 578  */
 579 static void ceph_release_pages(struct page **pages, int num)
 580 {
 581         struct pagevec pvec;
 582         int i;
 583
 584         pagevec_init(&pvec, 0);
 585         for (i = 0; i < num; i++) {
 586                 if (pagevec_add(&pvec, pages[i]) == 0)
 587                         pagevec_release(&pvec);
 588         }
 589         pagevec_release(&pvec);
 590 }
 591
 592 /*
 593  * async writeback completion handler.
 594  *
 595  * If we get an error, set the mapping error bit, but not the individual
 596  * page error bits.
 597  */
 598 static void writepages_finish(struct ceph_osd_request *req,
 599                               struct ceph_msg *msg)
 600 {
 601         struct inode *inode = req->r_inode;
 602         struct ceph_inode_info *ci = ceph_inode(inode);
 603         struct ceph_osd_data *osd_data;
 604         unsigned wrote;
 605         struct page *page;
 606         int num_pages;
 607         int i;
 608         struct ceph_snap_context *snapc = req->r_snapc;
 609         struct address_space *mapping = inode->i_mapping;
 610         int rc = req->r_result;
 611         u64 bytes = req->r_ops[0].extent.length;
 612         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 613         long writeback_stat;
 614         unsigned issued = ceph_caps_issued(ci);
 615
 616         osd_data = osd_req_op_extent_osd_data(req, 0);
 617         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
 618         num_pages = calc_pages_for((u64)osd_data->alignment,
 619                                         (u64)osd_data->length);
 620         if (rc >= 0) {
 621                 /*
 622                  * Assume we wrote the pages we originally sent.  The
 623                  * osd might reply with fewer pages if our writeback
 624                  * raced with a truncation and was adjusted at the osd,
 625                  * so don't believe the reply.
 626                  */
 627                 wrote = num_pages;
 628         } else {
 629                 wrote = 0;
 630                 mapping_set_error(mapping, rc);
 631         }
 632         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
 633              inode, rc, bytes, wrote);
 634
 635         /* clean all pages */
 636         for (i = 0; i < num_pages; i++) {
 637                 page = osd_data->pages[i];
 638                 BUG_ON(!page);
 639                 WARN_ON(!PageUptodate(page));
 640
 641                 writeback_stat =
 642                         atomic_long_dec_return(&fsc->writeback_count);
 643                 if (writeback_stat <
 644                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
 645                         clear_bdi_congested(&fsc->backing_dev_info,
 646                                             BLK_RW_ASYNC);
 647
 648                 ceph_put_snap_context(page_snap_context(page));
 649                 page->private = 0;
 650                 ClearPagePrivate(page);
 651                 dout("unlocking %d %p\n", i, page);
 652                 end_page_writeback(page);
 653
 654                 /*
 655                  * We lost the cache cap, need to truncate the page before
 656                  * it is unlocked, otherwise we'd truncate it later in the
 657                  * page truncation thread, possibly losing some data that
 658                  * raced its way in
 659                  */
 660                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
 661                         generic_error_remove_page(inode->i_mapping, page);
 662
 663                 unlock_page(page);
 664         }
 665         dout("%p wrote+cleaned %d pages\n", inode, wrote);
 666         ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
 667
 668         ceph_release_pages(osd_data->pages, num_pages);
 669         if (osd_data->pages_from_pool)
 670                 mempool_free(osd_data->pages,
 671                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
 672         else
 673                 kfree(osd_data->pages);
 674         ceph_osdc_put_request(req);
 675 }
 676
 677 /*
 678  * initiate async writeback
 679  */
 680 static int ceph_writepages_start(struct address_space *mapping,
 681                                  struct writeback_control *wbc)
 682 {
 683         struct inode *inode = mapping->host;
 684         struct ceph_inode_info *ci = ceph_inode(inode);
 685         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 686         struct ceph_vino vino = ceph_vino(inode);
 687         pgoff_t index, start, end;
 688         int range_whole = 0;
 689         int should_loop = 1;
 690         pgoff_t max_pages = 0, max_pages_ever = 0;
 691         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
 692         struct pagevec pvec;
 693         int done = 0;
 694         int rc = 0;
 695         unsigned wsize = 1 << inode->i_blkbits;
 696         struct ceph_osd_request *req = NULL;
 697         int do_sync = 0;
 698         u64 truncate_size, snap_size;
 699         u32 truncate_seq;
 700
 701         /*
 702          * Include a 'sync' in the OSD request if this is a data
 703          * integrity write (e.g., O_SYNC write or fsync()), or if our
 704          * cap is being revoked.
 705          */
 706         if ((wbc->sync_mode == WB_SYNC_ALL) ||
 707                 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
 708                 do_sync = 1;
 709         dout("writepages_start %p dosync=%d (mode=%s)\n",
 710              inode, do_sync,
 711              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
 712              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
 713
 714         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
 715                 pr_warn("writepage_start %p on forced umount\n", inode);
 716                 return -EIO; /* we're in a forced umount, don't write! */
 717         }
 718         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
 719                 wsize = fsc->mount_options->wsize;
 720         if (wsize < PAGE_CACHE_SIZE)
 721                 wsize = PAGE_CACHE_SIZE;
 722         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
 723
 724         pagevec_init(&pvec, 0);
 725
 726         /* where to start/end? */
 727         if (wbc->range_cyclic) {
 728                 start = mapping->writeback_index; /* Start from prev offset */
 729                 end = -1;
 730                 dout(" cyclic, start at %lu\n", start);
 731         } else {
 732                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
 733                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
 734                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 735                         range_whole = 1;
 736                 should_loop = 0;
 737                 dout(" not cyclic, %lu to %lu\n", start, end);
 738         }
 739         index = start;
 740
 741 retry:
 742         /* find oldest snap context with dirty data */
 743         ceph_put_snap_context(snapc);
 744         snap_size = 0;
 745         snapc = get_oldest_context(inode, &snap_size);
 746         if (!snapc) {
 747                 /* hmm, why does writepages get called when there
 748                    is no dirty data? */
 749                 dout(" no snap context with dirty data?\n");
 750                 goto out;
 751         }
 752         if (snap_size == 0)
 753                 snap_size = i_size_read(inode);
 754         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
 755              snapc, snapc->seq, snapc->num_snaps);
 756
 757         spin_lock(&ci->i_ceph_lock);
 758         truncate_seq = ci->i_truncate_seq;
 759         truncate_size = ci->i_truncate_size;
 760         if (!snap_size)
 761                 snap_size = i_size_read(inode);
 762         spin_unlock(&ci->i_ceph_lock);
 763
 764         if (last_snapc && snapc != last_snapc) {
 765                 /* if we switched to a newer snapc, restart our scan at the
 766                  * start of the original file range. */
 767                 dout("  snapc differs from last pass, restarting at %lu\n",
 768                      index);
 769                 index = start;
 770         }
 771         last_snapc = snapc;
 772
 773         while (!done && index <= end) {
 774                 unsigned i;
 775                 int first;
 776                 pgoff_t next;
 777                 int pvec_pages, locked_pages;
 778                 struct page **pages = NULL;
 779                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
 780                 struct page *page;
 781                 int want;
 782                 u64 offset, len;
 783                 long writeback_stat;
 784
 785                 next = 0;
 786                 locked_pages = 0;
 787                 max_pages = max_pages_ever;
 788
 789 get_more_pages:
 790                 first = -1;
 791                 want = min(end - index,
 792                            min((pgoff_t)PAGEVEC_SIZE,
 793                                max_pages - (pgoff_t)locked_pages) - 1)
 794                         + 1;
 795                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 796                                                 PAGECACHE_TAG_DIRTY,
 797                                                 want);
 798                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
 799                 if (!pvec_pages && !locked_pages)
 800                         break;
 801                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
 802                         page = pvec.pages[i];
 803                         dout("? %p idx %lu\n", page, page->index);
 804                         if (locked_pages == 0)
 805                                 lock_page(page);  /* first page */
 806                         else if (!trylock_page(page))
 807                                 break;
 808
 809                         /* only dirty pages, or our accounting breaks */
 810                         if (unlikely(!PageDirty(page)) ||
 811                             unlikely(page->mapping != mapping)) {
 812                                 dout("!dirty or !mapping %p\n", page);
 813                                 unlock_page(page);
 814                                 break;
 815                         }
 816                         if (!wbc->range_cyclic && page->index > end) {
 817                                 dout("end of range %p\n", page);
 818                                 done = 1;
 819                                 unlock_page(page);
 820                                 break;
 821                         }
 822                         if (next && (page->index != next)) {
 823                                 dout("not consecutive %p\n", page);
 824                                 unlock_page(page);
 825                                 break;
 826                         }
 827                         if (wbc->sync_mode != WB_SYNC_NONE) {
 828                                 dout("waiting on writeback %p\n", page);
 829                                 wait_on_page_writeback(page);
 830                         }
 831                         if (page_offset(page) >= snap_size) {
 832                                 dout("%p page eof %llu\n", page, snap_size);
 833                                 done = 1;
 834                                 unlock_page(page);
 835                                 break;
 836                         }
 837                         if (PageWriteback(page)) {
 838                                 dout("%p under writeback\n", page);
 839                                 unlock_page(page);
 840                                 break;
 841                         }
 842
 843                         /* only if matching snap context */
 844                         pgsnapc = page_snap_context(page);
 845                         if (pgsnapc->seq > snapc->seq) {
 846                                 dout("page snapc %p %lld > oldest %p %lld\n",
 847                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
 848                                 unlock_page(page);
 849                                 if (!locked_pages)
 850                                         continue; /* keep looking for snap */
 851                                 break;
 852                         }
 853
 854                         if (!clear_page_dirty_for_io(page)) {
 855                                 dout("%p !clear_page_dirty_for_io\n", page);
 856                                 unlock_page(page);
 857                                 break;
 858                         }
 859
 860                         /*
 861                          * We have something to write.  If this is
 862                          * the first locked page this time through,
 863                          * allocate an osd request and a page array
 864                          * that it will use.
 865                          */
 866                         if (locked_pages == 0) {
 867                                 BUG_ON(pages);
 868                                 /* prepare async write request */
 869                                 offset = (u64)page_offset(page);
 870                                 len = wsize;
 871                                 req = ceph_osdc_new_request(&fsc->client->osdc,
 872                                                         &ci->i_layout, vino,
 873                                                         offset, &len, 0,
 874                                                         do_sync ? 2 : 1,
 875                                                         CEPH_OSD_OP_WRITE,
 876                                                         CEPH_OSD_FLAG_WRITE |
 877                                                         CEPH_OSD_FLAG_ONDISK,
 878                                                         snapc, truncate_seq,
 879                                                         truncate_size, true);
 880                                 if (IS_ERR(req)) {
 881                                         rc = PTR_ERR(req);
 882                                         unlock_page(page);
 883                                         break;
 884                                 }
 885
 886                                 if (do_sync)
 887                                         osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
 888
 889                                 req->r_callback = writepages_finish;
 890                                 req->r_inode = inode;
 891
 892                                 max_pages = calc_pages_for(0, (u64)len);
 893                                 pages = kmalloc(max_pages * sizeof (*pages),
 894                                                 GFP_NOFS);
 895                                 if (!pages) {
 896                                         pool = fsc->wb_pagevec_pool;
 897                                         pages = mempool_alloc(pool, GFP_NOFS);
 898                                         BUG_ON(!pages);
 899                                 }
 900                         }
 901
 902                         /* note position of first page in pvec */
 903                         if (first < 0)
 904                                 first = i;
 905                         dout("%p will write page %p idx %lu\n",
 906                              inode, page, page->index);
 907
 908                         writeback_stat =
 909                                atomic_long_inc_return(&fsc->writeback_count);
 910                         if (writeback_stat > CONGESTION_ON_THRESH(
 911                                     fsc->mount_options->congestion_kb)) {
 912                                 set_bdi_congested(&fsc->backing_dev_info,
 913                                                   BLK_RW_ASYNC);
 914                         }
 915
 916                         set_page_writeback(page);
 917                         pages[locked_pages] = page;
 918                         locked_pages++;
 919                         next = page->index + 1;
 920                 }
 921
 922                 /* did we get anything? */
 923                 if (!locked_pages)
 924                         goto release_pvec_pages;
 925                 if (i) {
 926                         int j;
 927                         BUG_ON(!locked_pages || first < 0);
 928
 929                         if (pvec_pages && i == pvec_pages &&
 930                             locked_pages < max_pages) {
 931                                 dout("reached end pvec, trying for more\n");
 932                                 pagevec_reinit(&pvec);
 933                                 goto get_more_pages;
 934                         }
 935
 936                         /* shift unused pages over in the pvec...  we
 937                          * will need to release them below. */
 938                         for (j = i; j < pvec_pages; j++) {
 939                                 dout(" pvec leftover page %p\n",
 940                                      pvec.pages[j]);
 941                                 pvec.pages[j-i+first] = pvec.pages[j];
 942                         }
 943                         pvec.nr -= i-first;
 944                 }
 945
 946                 /* Format the osd request message and submit the write */
 947
 948                 offset = page_offset(pages[0]);
 949                 len = min(snap_size - offset,
 950                           (u64)locked_pages << PAGE_CACHE_SHIFT);
 951                 dout("writepages got %d pages at %llu~%llu\n",
 952                      locked_pages, offset, len);
 953
 954                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
 955                                                         !!pool, false);
 956
 957                 pages = NULL;   /* request message now owns the pages array */
 958                 pool = NULL;
 959
 960                 /* Update the write op length in case we changed it */
 961
 962                 osd_req_op_extent_update(req, 0, len);
 963
 964                 vino = ceph_vino(inode);
 965                 ceph_osdc_build_request(req, offset, snapc, vino.snap,
 966                                         &inode->i_mtime);
 967
 968                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
 969                 BUG_ON(rc);
 970                 req = NULL;
 971
 972                 /* continue? */
 973                 index = next;
 974                 wbc->nr_to_write -= locked_pages;
 975                 if (wbc->nr_to_write <= 0)
 976                         done = 1;
 977
 978 release_pvec_pages:
 979                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
 980                      pvec.nr ? pvec.pages[0] : NULL);
 981                 pagevec_release(&pvec);
 982
 983                 if (locked_pages && !done)
 984                         goto retry;
 985         }
 986
 987         if (should_loop && !done) {
 988                 /* more to do; loop back to beginning of file */
 989                 dout("writepages looping back to beginning of file\n");
 990                 should_loop = 0;
 991                 index = 0;
 992                 goto retry;
 993         }
 994
 995         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 996                 mapping->writeback_index = index;
 997
 998 out:
 999         if (req)
1000                 ceph_osdc_put_request(req);
1001         ceph_put_snap_context(snapc);
1002         dout("writepages done, rc = %d\n", rc);
1003         return rc;
1004 }
1005
1006
1007
1008 /*
1009  * See if a given @snapc is either writeable, or already written.
1010  */
1011 static int context_is_writeable_or_written(struct inode *inode,
1012                                            struct ceph_snap_context *snapc)
1013 {
1014         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
1015         int ret = !oldest || snapc->seq <= oldest->seq;
1016
1017         ceph_put_snap_context(oldest);
1018         return ret;
1019 }
1020
1021 /*
1022  * We are only allowed to write into/dirty the page if the page is
1023  * clean, or already dirty within the same snap context.
1024  *
1025  * called with page locked.
1026  * return success with page locked,
1027  * or any failure (incl -EAGAIN) with page unlocked.
1028  */
1029 static int ceph_update_writeable_page(struct file *file,
1030                             loff_t pos, unsigned len,
1031                             struct page *page)
1032 {
1033         struct inode *inode = file_inode(file);
1034         struct ceph_inode_info *ci = ceph_inode(inode);
1035         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1036         loff_t page_off = pos & PAGE_CACHE_MASK;
1037         int pos_in_page = pos & ~PAGE_CACHE_MASK;
1038         int end_in_page = pos_in_page + len;
1039         loff_t i_size;
1040         int r;
1041         struct ceph_snap_context *snapc, *oldest;
1042
1043 retry_locked:
1044         /* writepages currently holds page lock, but if we change that later, */
1045         wait_on_page_writeback(page);
1046
1047         /* check snap context */
1048         BUG_ON(!ci->i_snap_realm);
1049         down_read(&mdsc->snap_rwsem);
1050         BUG_ON(!ci->i_snap_realm->cached_context);
1051         snapc = page_snap_context(page);
1052         if (snapc && snapc != ci->i_head_snapc) {
1053                 /*
1054                  * this page is already dirty in another (older) snap
1055                  * context!  is it writeable now?
1056                  */
1057                 oldest = get_oldest_context(inode, NULL);
1058                 up_read(&mdsc->snap_rwsem);
1059
1060                 if (snapc->seq > oldest->seq) {
1061                         ceph_put_snap_context(oldest);
1062                         dout(" page %p snapc %p not current or oldest\n",
1063                              page, snapc);
1064                         /*
1065                          * queue for writeback, and wait for snapc to
1066                          * be writeable or written
1067                          */
1068                         snapc = ceph_get_snap_context(snapc);
1069                         unlock_page(page);
1070                         ceph_queue_writeback(inode);
1071                         r = wait_event_interruptible(ci->i_cap_wq,
1072                                context_is_writeable_or_written(inode, snapc));
1073                         ceph_put_snap_context(snapc);
1074                         if (r == -ERESTARTSYS)
1075                                 return r;
1076                         return -EAGAIN;
1077                 }
1078                 ceph_put_snap_context(oldest);
1079
1080                 /* yay, writeable, do it now (without dropping page lock) */
1081                 dout(" page %p snapc %p not current, but oldest\n",
1082                      page, snapc);
1083                 if (!clear_page_dirty_for_io(page))
1084                         goto retry_locked;
1085                 r = writepage_nounlock(page, NULL);
1086                 if (r < 0)
1087                         goto fail_nosnap;
1088                 goto retry_locked;
1089         }
1090
1091         if (PageUptodate(page)) {
1092                 dout(" page %p already uptodate\n", page);
1093                 return 0;
1094         }
1095
1096         /* full page? */
1097         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1098                 return 0;
1099
1100         /* past end of file? */
1101         i_size = inode->i_size;   /* caller holds i_mutex */
1102
1103         if (page_off >= i_size ||
1104             (pos_in_page == 0 && (pos+len) >= i_size &&
1105              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1106                 dout(" zeroing %p 0 - %d and %d - %d\n",
1107                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1108                 zero_user_segments(page,
1109                                    0, pos_in_page,
1110                                    end_in_page, PAGE_CACHE_SIZE);
1111                 return 0;
1112         }
1113
1114         /* we need to read it. */
1115         up_read(&mdsc->snap_rwsem);
1116         r = readpage_nounlock(file, page);
1117         if (r < 0)
1118                 goto fail_nosnap;
1119         goto retry_locked;
1120 fail_nosnap:
1121         unlock_page(page);
1122         return r;
1123 }
1124
1125 /*
1126  * We are only allowed to write into/dirty the page if the page is
1127  * clean, or already dirty within the same snap context.
1128  */
1129 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1130                             loff_t pos, unsigned len, unsigned flags,
1131                             struct page **pagep, void **fsdata)
1132 {
1133         struct inode *inode = file_inode(file);
1134         struct page *page;
1135         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1136         int r;
1137
1138         do {
1139                 /* get a page */
1140                 page = grab_cache_page_write_begin(mapping, index, 0);
1141                 if (!page)
1142                         return -ENOMEM;
1143                 *pagep = page;
1144
1145                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1146                      inode, page, (int)pos, (int)len);
1147
1148                 r = ceph_update_writeable_page(file, pos, len, page);
1149                 if (r < 0)
1150                         page_cache_release(page);
1151                 else
1152                         *pagep = page;
1153         } while (r == -EAGAIN);
1154
1155         return r;
1156 }
1157
1158 /*
1159  * we don't do anything in here that simple_write_end doesn't do
1160  * except adjust dirty page accounting and drop read lock on
1161  * mdsc->snap_rwsem.
1162  */
1163 static int ceph_write_end(struct file *file, struct address_space *mapping,
1164                           loff_t pos, unsigned len, unsigned copied,
1165                           struct page *page, void *fsdata)
1166 {
1167         struct inode *inode = file_inode(file);
1168         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1169         struct ceph_mds_client *mdsc = fsc->mdsc;
1170         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1171         int check_cap = 0;
1172
1173         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1174              inode, page, (int)pos, (int)copied, (int)len);
1175
1176         /* zero the stale part of the page if we did a short copy */
1177         if (copied < len)
1178                 zero_user_segment(page, from+copied, len);
1179
1180         /* did file size increase? */
1181         /* (no need for i_size_read(); we caller holds i_mutex */
1182         if (pos+copied > inode->i_size)
1183                 check_cap = ceph_inode_set_size(inode, pos+copied);
1184
1185         if (!PageUptodate(page))
1186                 SetPageUptodate(page);
1187
1188         set_page_dirty(page);
1189
1190         unlock_page(page);
1191         up_read(&mdsc->snap_rwsem);
1192         page_cache_release(page);
1193
1194         if (check_cap)
1195                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1196
1197         return copied;
1198 }
1199
1200 /*
1201  * we set .direct_IO to indicate direct io is supported, but since we
1202  * intercept O_DIRECT reads and writes early, this function should
1203  * never get called.
1204  */
1205 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter,
1206                               loff_t pos)
1207 {
1208         WARN_ON(1);
1209         return -EINVAL;
1210 }
1211
1212 const struct address_space_operations ceph_aops = {
1213         .readpage = ceph_readpage,
1214         .readpages = ceph_readpages,
1215         .writepage = ceph_writepage,
1216         .writepages = ceph_writepages_start,
1217         .write_begin = ceph_write_begin,
1218         .write_end = ceph_write_end,
1219         .set_page_dirty = ceph_set_page_dirty,
1220         .invalidatepage = ceph_invalidatepage,
1221         .releasepage = ceph_releasepage,
1222         .direct_IO = ceph_direct_io,
1223 };
1224
1225
1226 /*
1227  * vm ops
1228  */
1229 static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1230 {
1231         struct inode *inode = file_inode(vma->vm_file);
1232         struct ceph_inode_info *ci = ceph_inode(inode);
1233         struct ceph_file_info *fi = vma->vm_file->private_data;
1234         struct page *pinned_page = NULL;
1235         loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT;
1236         int want, got, ret;
1237
1238         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1239              inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1240         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1241                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1242         else
1243                 want = CEPH_CAP_FILE_CACHE;
1244         while (1) {
1245                 got = 0;
1246                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want,
1247                                     -1, &got, &pinned_page);
1248                 if (ret == 0)
1249                         break;
1250                 if (ret != -ERESTARTSYS) {
1251                         WARN_ON(1);
1252                         return VM_FAULT_SIGBUS;
1253                 }
1254         }
1255         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1256              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1257
1258         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1259             ci->i_inline_version == CEPH_INLINE_NONE)
1260                 ret = filemap_fault(vma, vmf);
1261         else
1262                 ret = -EAGAIN;
1263
1264         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1265              inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1266         if (pinned_page)
1267                 page_cache_release(pinned_page);
1268         ceph_put_cap_refs(ci, got);
1269
1270         if (ret != -EAGAIN)
1271                 return ret;
1272
1273         /* read inline data */
1274         if (off >= PAGE_CACHE_SIZE) {
1275                 /* does not support inline data > PAGE_SIZE */
1276                 ret = VM_FAULT_SIGBUS;
1277         } else {
1278                 int ret1;
1279                 struct address_space *mapping = inode->i_mapping;
1280                 struct page *page = find_or_create_page(mapping, 0,
1281                                                 mapping_gfp_mask(mapping) &
1282                                                 ~__GFP_FS);
1283                 if (!page) {
1284                         ret = VM_FAULT_OOM;
1285                         goto out;
1286                 }
1287                 ret1 = __ceph_do_getattr(inode, page,
1288                                          CEPH_STAT_CAP_INLINE_DATA, true);
1289                 if (ret1 < 0 || off >= i_size_read(inode)) {
1290                         unlock_page(page);
1291                         page_cache_release(page);
1292                         ret = VM_FAULT_SIGBUS;
1293                         goto out;
1294                 }
1295                 if (ret1 < PAGE_CACHE_SIZE)
1296                         zero_user_segment(page, ret1, PAGE_CACHE_SIZE);
1297                 else
1298                         flush_dcache_page(page);
1299                 SetPageUptodate(page);
1300                 vmf->page = page;
1301                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1302         }
1303 out:
1304         dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1305              inode, off, (size_t)PAGE_CACHE_SIZE, ret);
1306         return ret;
1307 }
1308
1309 /*
1310  * Reuse write_begin here for simplicity.
1311  */
1312 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1313 {
1314         struct inode *inode = file_inode(vma->vm_file);
1315         struct ceph_inode_info *ci = ceph_inode(inode);
1316         struct ceph_file_info *fi = vma->vm_file->private_data;
1317         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1318         struct page *page = vmf->page;
1319         loff_t off = page_offset(page);
1320         loff_t size = i_size_read(inode);
1321         size_t len;
1322         int want, got, ret;
1323
1324         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1325                 struct page *locked_page = NULL;
1326                 if (off == 0) {
1327                         lock_page(page);
1328                         locked_page = page;
1329                 }
1330                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1331                 if (locked_page)
1332                         unlock_page(locked_page);
1333                 if (ret < 0)
1334                         return VM_FAULT_SIGBUS;
1335         }
1336
1337         if (off + PAGE_CACHE_SIZE <= size)
1338                 len = PAGE_CACHE_SIZE;
1339         else
1340                 len = size & ~PAGE_CACHE_MASK;
1341
1342         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1343              inode, ceph_vinop(inode), off, len, size);
1344         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1345                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1346         else
1347                 want = CEPH_CAP_FILE_BUFFER;
1348         while (1) {
1349                 got = 0;
1350                 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1351                                     &got, NULL);
1352                 if (ret == 0)
1353                         break;
1354                 if (ret != -ERESTARTSYS) {
1355                         WARN_ON(1);
1356                         return VM_FAULT_SIGBUS;
1357                 }
1358         }
1359         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1360              inode, off, len, ceph_cap_string(got));
1361
1362         /* Update time before taking page lock */
1363         file_update_time(vma->vm_file);
1364
1365         lock_page(page);
1366
1367         ret = VM_FAULT_NOPAGE;
1368         if ((off > size) ||
1369             (page->mapping != inode->i_mapping))
1370                 goto out;
1371
1372         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1373         if (ret == 0) {
1374                 /* success.  we'll keep the page locked. */
1375                 set_page_dirty(page);
1376                 up_read(&mdsc->snap_rwsem);
1377                 ret = VM_FAULT_LOCKED;
1378         } else {
1379                 if (ret == -ENOMEM)
1380                         ret = VM_FAULT_OOM;
1381                 else
1382                         ret = VM_FAULT_SIGBUS;
1383         }
1384 out:
1385         if (ret != VM_FAULT_LOCKED)
1386                 unlock_page(page);
1387         if (ret == VM_FAULT_LOCKED ||
1388             ci->i_inline_version != CEPH_INLINE_NONE) {
1389                 int dirty;
1390                 spin_lock(&ci->i_ceph_lock);
1391                 ci->i_inline_version = CEPH_INLINE_NONE;
1392                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1393                 spin_unlock(&ci->i_ceph_lock);
1394                 if (dirty)
1395                         __mark_inode_dirty(inode, dirty);
1396         }
1397
1398         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1399              inode, off, len, ceph_cap_string(got), ret);
1400         ceph_put_cap_refs(ci, got);
1401
1402         return ret;
1403 }
1404
1405 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1406                            char *data, size_t len)
1407 {
1408         struct address_space *mapping = inode->i_mapping;
1409         struct page *page;
1410
1411         if (locked_page) {
1412                 page = locked_page;
1413         } else {
1414                 if (i_size_read(inode) == 0)
1415                         return;
1416                 page = find_or_create_page(mapping, 0,
1417                                            mapping_gfp_mask(mapping) & ~__GFP_FS);
1418                 if (!page)
1419                         return;
1420                 if (PageUptodate(page)) {
1421                         unlock_page(page);
1422                         page_cache_release(page);
1423                         return;
1424                 }
1425         }
1426
1427         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1428              inode, ceph_vinop(inode), len, locked_page);
1429
1430         if (len > 0) {
1431                 void *kaddr = kmap_atomic(page);
1432                 memcpy(kaddr, data, len);
1433                 kunmap_atomic(kaddr);
1434         }
1435
1436         if (page != locked_page) {
1437                 if (len < PAGE_CACHE_SIZE)
1438                         zero_user_segment(page, len, PAGE_CACHE_SIZE);
1439                 else
1440                         flush_dcache_page(page);
1441
1442                 SetPageUptodate(page);
1443                 unlock_page(page);
1444                 page_cache_release(page);
1445         }
1446 }
1447
1448 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1449 {
1450         struct inode *inode = file_inode(filp);
1451         struct ceph_inode_info *ci = ceph_inode(inode);
1452         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1453         struct ceph_osd_request *req;
1454         struct page *page = NULL;
1455         u64 len, inline_version;
1456         int err = 0;
1457         bool from_pagecache = false;
1458
1459         spin_lock(&ci->i_ceph_lock);
1460         inline_version = ci->i_inline_version;
1461         spin_unlock(&ci->i_ceph_lock);
1462
1463         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1464              inode, ceph_vinop(inode), inline_version);
1465
1466         if (inline_version == 1 || /* initial version, no data */
1467             inline_version == CEPH_INLINE_NONE)
1468                 goto out;
1469
1470         if (locked_page) {
1471                 page = locked_page;
1472                 WARN_ON(!PageUptodate(page));
1473         } else if (ceph_caps_issued(ci) &
1474                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1475                 page = find_get_page(inode->i_mapping, 0);
1476                 if (page) {
1477                         if (PageUptodate(page)) {
1478                                 from_pagecache = true;
1479                                 lock_page(page);
1480                         } else {
1481                                 page_cache_release(page);
1482                                 page = NULL;
1483                         }
1484                 }
1485         }
1486
1487         if (page) {
1488                 len = i_size_read(inode);
1489                 if (len > PAGE_CACHE_SIZE)
1490                         len = PAGE_CACHE_SIZE;
1491         } else {
1492                 page = __page_cache_alloc(GFP_NOFS);
1493                 if (!page) {
1494                         err = -ENOMEM;
1495                         goto out;
1496                 }
1497                 err = __ceph_do_getattr(inode, page,
1498                                         CEPH_STAT_CAP_INLINE_DATA, true);
1499                 if (err < 0) {
1500                         /* no inline data */
1501                         if (err == -ENODATA)
1502                                 err = 0;
1503                         goto out;
1504                 }
1505                 len = err;
1506         }
1507
1508         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1509                                     ceph_vino(inode), 0, &len, 0, 1,
1510                                     CEPH_OSD_OP_CREATE,
1511                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1512                                     ci->i_snap_realm->cached_context,
1513                                     0, 0, false);
1514         if (IS_ERR(req)) {
1515                 err = PTR_ERR(req);
1516                 goto out;
1517         }
1518
1519         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1520         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1521         if (!err)
1522                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1523         ceph_osdc_put_request(req);
1524         if (err < 0)
1525                 goto out;
1526
1527         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1528                                     ceph_vino(inode), 0, &len, 1, 3,
1529                                     CEPH_OSD_OP_WRITE,
1530                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1531                                     ci->i_snap_realm->cached_context,
1532                                     ci->i_truncate_seq, ci->i_truncate_size,
1533                                     false);
1534         if (IS_ERR(req)) {
1535                 err = PTR_ERR(req);
1536                 goto out;
1537         }
1538
1539         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1540
1541         {
1542                 __le64 xattr_buf = cpu_to_le64(inline_version);
1543                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1544                                             "inline_version", &xattr_buf,
1545                                             sizeof(xattr_buf),
1546                                             CEPH_OSD_CMPXATTR_OP_GT,
1547                                             CEPH_OSD_CMPXATTR_MODE_U64);
1548                 if (err)
1549                         goto out_put;
1550         }
1551
1552         {
1553                 char xattr_buf[32];
1554                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1555                                          "%llu", inline_version);
1556                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1557                                             "inline_version",
1558                                             xattr_buf, xattr_len, 0, 0);
1559                 if (err)
1560                         goto out_put;
1561         }
1562
1563         ceph_osdc_build_request(req, 0, NULL, CEPH_NOSNAP, &inode->i_mtime);
1564         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1565         if (!err)
1566                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1567 out_put:
1568         ceph_osdc_put_request(req);
1569         if (err == -ECANCELED)
1570                 err = 0;
1571 out:
1572         if (page && page != locked_page) {
1573                 if (from_pagecache) {
1574                         unlock_page(page);
1575                         page_cache_release(page);
1576                 } else
1577                         __free_pages(page, 0);
1578         }
1579
1580         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1581              inode, ceph_vinop(inode), inline_version, err);
1582         return err;
1583 }
1584
1585 static struct vm_operations_struct ceph_vmops = {
1586         .fault          = ceph_filemap_fault,
1587         .page_mkwrite   = ceph_page_mkwrite,
1588 };
1589
1590 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1591 {
1592         struct address_space *mapping = file->f_mapping;
1593
1594         if (!mapping->a_ops->readpage)
1595                 return -ENOEXEC;
1596         file_accessed(file);
1597         vma->vm_ops = &ceph_vmops;
1598         return 0;
1599 }