kernel/fs/ocfs2/move_extents.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * move_extents.c
   5  *
   6  * Copyright (C) 2011 Oracle.  All rights reserved.
   7  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public
  10  * License version 2 as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License for more details.
  16  */
  17 #include <linux/fs.h>
  18 #include <linux/types.h>
  19 #include <linux/mount.h>
  20 #include <linux/swap.h>
  21
  22 #include <cluster/masklog.h>
  23
  24 #include "ocfs2.h"
  25 #include "ocfs2_ioctl.h"
  26
  27 #include "alloc.h"
  28 #include "aops.h"
  29 #include "dlmglue.h"
  30 #include "extent_map.h"
  31 #include "inode.h"
  32 #include "journal.h"
  33 #include "suballoc.h"
  34 #include "uptodate.h"
  35 #include "super.h"
  36 #include "dir.h"
  37 #include "buffer_head_io.h"
  38 #include "sysfile.h"
  39 #include "refcounttree.h"
  40 #include "move_extents.h"
  41
  42 struct ocfs2_move_extents_context {
  43         struct inode *inode;
  44         struct file *file;
  45         int auto_defrag;
  46         int partial;
  47         int credits;
  48         u32 new_phys_cpos;
  49         u32 clusters_moved;
  50         u64 refcount_loc;
  51         struct ocfs2_move_extents *range;
  52         struct ocfs2_extent_tree et;
  53         struct ocfs2_alloc_context *meta_ac;
  54         struct ocfs2_alloc_context *data_ac;
  55         struct ocfs2_cached_dealloc_ctxt dealloc;
  56 };
  57
  58 static int __ocfs2_move_extent(handle_t *handle,
  59                                struct ocfs2_move_extents_context *context,
  60                                u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
  61                                int ext_flags)
  62 {
  63         int ret = 0, index;
  64         struct inode *inode = context->inode;
  65         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  66         struct ocfs2_extent_rec *rec, replace_rec;
  67         struct ocfs2_path *path = NULL;
  68         struct ocfs2_extent_list *el;
  69         u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
  70         u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
  71
  72         ret = ocfs2_duplicate_clusters_by_page(handle, inode, cpos,
  73                                                p_cpos, new_p_cpos, len);
  74         if (ret) {
  75                 mlog_errno(ret);
  76                 goto out;
  77         }
  78
  79         memset(&replace_rec, 0, sizeof(replace_rec));
  80         replace_rec.e_cpos = cpu_to_le32(cpos);
  81         replace_rec.e_leaf_clusters = cpu_to_le16(len);
  82         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
  83                                                                    new_p_cpos));
  84
  85         path = ocfs2_new_path_from_et(&context->et);
  86         if (!path) {
  87                 ret = -ENOMEM;
  88                 mlog_errno(ret);
  89                 goto out;
  90         }
  91
  92         ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
  93         if (ret) {
  94                 mlog_errno(ret);
  95                 goto out;
  96         }
  97
  98         el = path_leaf_el(path);
  99
 100         index = ocfs2_search_extent_list(el, cpos);
 101         if (index == -1) {
 102                 ocfs2_error(inode->i_sb,
 103                             "Inode %llu has an extent at cpos %u which can no "
 104                             "longer be found.\n",
 105                             (unsigned long long)ino, cpos);
 106                 ret = -EROFS;
 107                 goto out;
 108         }
 109
 110         rec = &el->l_recs[index];
 111
 112         BUG_ON(ext_flags != rec->e_flags);
 113         /*
 114          * after moving/defraging to new location, the extent is not going
 115          * to be refcounted anymore.
 116          */
 117         replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
 118
 119         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
 120                                       context->et.et_root_bh,
 121                                       OCFS2_JOURNAL_ACCESS_WRITE);
 122         if (ret) {
 123                 mlog_errno(ret);
 124                 goto out;
 125         }
 126
 127         ret = ocfs2_split_extent(handle, &context->et, path, index,
 128                                  &replace_rec, context->meta_ac,
 129                                  &context->dealloc);
 130         if (ret) {
 131                 mlog_errno(ret);
 132                 goto out;
 133         }
 134
 135         ocfs2_journal_dirty(handle, context->et.et_root_bh);
 136
 137         context->new_phys_cpos = new_p_cpos;
 138
 139         /*
 140          * need I to append truncate log for old clusters?
 141          */
 142         if (old_blkno) {
 143                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 144                         ret = ocfs2_decrease_refcount(inode, handle,
 145                                         ocfs2_blocks_to_clusters(osb->sb,
 146                                                                  old_blkno),
 147                                         len, context->meta_ac,
 148                                         &context->dealloc, 1);
 149                 else
 150                         ret = ocfs2_truncate_log_append(osb, handle,
 151                                                         old_blkno, len);
 152         }
 153
 154         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 155 out:
 156         ocfs2_free_path(path);
 157         return ret;
 158 }
 159
 160 /*
 161  * lock allocators, and reserving appropriate number of bits for
 162  * meta blocks and data clusters.
 163  *
 164  * in some cases, we don't need to reserve clusters, just let data_ac
 165  * be NULL.
 166  */
 167 static int ocfs2_lock_allocators_move_extents(struct inode *inode,
 168                                         struct ocfs2_extent_tree *et,
 169                                         u32 clusters_to_move,
 170                                         u32 extents_to_split,
 171                                         struct ocfs2_alloc_context **meta_ac,
 172                                         struct ocfs2_alloc_context **data_ac,
 173                                         int extra_blocks,
 174                                         int *credits)
 175 {
 176         int ret, num_free_extents;
 177         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 178         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 179
 180         num_free_extents = ocfs2_num_free_extents(osb, et);
 181         if (num_free_extents < 0) {
 182                 ret = num_free_extents;
 183                 mlog_errno(ret);
 184                 goto out;
 185         }
 186
 187         if (!num_free_extents ||
 188             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 189                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 190
 191         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 192         if (ret) {
 193                 mlog_errno(ret);
 194                 goto out;
 195         }
 196
 197         if (data_ac) {
 198                 ret = ocfs2_reserve_clusters(osb, clusters_to_move, data_ac);
 199                 if (ret) {
 200                         mlog_errno(ret);
 201                         goto out;
 202                 }
 203         }
 204
 205         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 206
 207         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 208              extra_blocks, clusters_to_move, *credits);
 209 out:
 210         if (ret) {
 211                 if (*meta_ac) {
 212                         ocfs2_free_alloc_context(*meta_ac);
 213                         *meta_ac = NULL;
 214                 }
 215         }
 216
 217         return ret;
 218 }
 219
 220 /*
 221  * Using one journal handle to guarantee the data consistency in case
 222  * crash happens anywhere.
 223  *
 224  *  XXX: defrag can end up with finishing partial extent as requested,
 225  * due to not enough contiguous clusters can be found in allocator.
 226  */
 227 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 228                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 229 {
 230         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 231         handle_t *handle;
 232         struct inode *inode = context->inode;
 233         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 234         struct inode *tl_inode = osb->osb_tl_inode;
 235         struct ocfs2_refcount_tree *ref_tree = NULL;
 236         u32 new_phys_cpos, new_len;
 237         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 238
 239         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 240
 241                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 242                          OCFS2_HAS_REFCOUNT_FL));
 243
 244                 BUG_ON(!context->refcount_loc);
 245
 246                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 247                                                &ref_tree, NULL);
 248                 if (ret) {
 249                         mlog_errno(ret);
 250                         return ret;
 251                 }
 252
 253                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 254                                                         context->refcount_loc,
 255                                                         phys_blkno,
 256                                                         *len,
 257                                                         &credits,
 258                                                         &extra_blocks);
 259                 if (ret) {
 260                         mlog_errno(ret);
 261                         goto out;
 262                 }
 263         }
 264
 265         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, *len, 1,
 266                                                  &context->meta_ac,
 267                                                  &context->data_ac,
 268                                                  extra_blocks, &credits);
 269         if (ret) {
 270                 mlog_errno(ret);
 271                 goto out;
 272         }
 273
 274         /*
 275          * should be using allocation reservation strategy there?
 276          *
 277          * if (context->data_ac)
 278          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 279          */
 280
 281         mutex_lock(&tl_inode->i_mutex);
 282
 283         if (ocfs2_truncate_log_needs_flush(osb)) {
 284                 ret = __ocfs2_flush_truncate_log(osb);
 285                 if (ret < 0) {
 286                         mlog_errno(ret);
 287                         goto out_unlock_mutex;
 288                 }
 289         }
 290
 291         handle = ocfs2_start_trans(osb, credits);
 292         if (IS_ERR(handle)) {
 293                 ret = PTR_ERR(handle);
 294                 mlog_errno(ret);
 295                 goto out_unlock_mutex;
 296         }
 297
 298         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 299                                      &new_phys_cpos, &new_len);
 300         if (ret) {
 301                 mlog_errno(ret);
 302                 goto out_commit;
 303         }
 304
 305         /*
 306          * allowing partial extent moving is kind of 'pros and cons', it makes
 307          * whole defragmentation less likely to fail, on the contrary, the bad
 308          * thing is it may make the fs even more fragmented after moving, let
 309          * userspace make a good decision here.
 310          */
 311         if (new_len != *len) {
 312                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 313                 if (!partial) {
 314                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 315                         ret = -ENOSPC;
 316                         goto out_commit;
 317                 }
 318         }
 319
 320         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 321              phys_cpos, new_phys_cpos);
 322
 323         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 324                                   new_phys_cpos, ext_flags);
 325         if (ret)
 326                 mlog_errno(ret);
 327
 328         if (partial && (new_len != *len))
 329                 *len = new_len;
 330
 331         /*
 332          * Here we should write the new page out first if we are
 333          * in write-back mode.
 334          */
 335         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 336         if (ret)
 337                 mlog_errno(ret);
 338
 339 out_commit:
 340         ocfs2_commit_trans(osb, handle);
 341
 342 out_unlock_mutex:
 343         mutex_unlock(&tl_inode->i_mutex);
 344
 345         if (context->data_ac) {
 346                 ocfs2_free_alloc_context(context->data_ac);
 347                 context->data_ac = NULL;
 348         }
 349
 350         if (context->meta_ac) {
 351                 ocfs2_free_alloc_context(context->meta_ac);
 352                 context->meta_ac = NULL;
 353         }
 354
 355 out:
 356         if (ref_tree)
 357                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 358
 359         return ret;
 360 }
 361
 362 /*
 363  * find the victim alloc group, where #blkno fits.
 364  */
 365 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 366                                          u64 vict_blkno,
 367                                          int type, int slot,
 368                                          int *vict_bit,
 369                                          struct buffer_head **ret_bh)
 370 {
 371         int ret, i, bits_per_unit = 0;
 372         u64 blkno;
 373         char namebuf[40];
 374
 375         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 376         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 377         struct ocfs2_chain_list *cl;
 378         struct ocfs2_chain_rec *rec;
 379         struct ocfs2_dinode *ac_dinode;
 380         struct ocfs2_group_desc *bg;
 381
 382         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 383         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 384                                          strlen(namebuf), &blkno);
 385         if (ret) {
 386                 ret = -ENOENT;
 387                 goto out;
 388         }
 389
 390         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 391         if (ret) {
 392                 mlog_errno(ret);
 393                 goto out;
 394         }
 395
 396         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 397         cl = &(ac_dinode->id2.i_chain);
 398         rec = &(cl->cl_recs[0]);
 399
 400         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 401                 bits_per_unit = osb->s_clustersize_bits -
 402                                         inode->i_sb->s_blocksize_bits;
 403         /*
 404          * 'vict_blkno' was out of the valid range.
 405          */
 406         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 407             (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 408                                 bits_per_unit))) {
 409                 ret = -EINVAL;
 410                 goto out;
 411         }
 412
 413         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 414
 415                 rec = &(cl->cl_recs[i]);
 416                 if (!rec)
 417                         continue;
 418
 419                 bg = NULL;
 420
 421                 do {
 422                         if (!bg)
 423                                 blkno = le64_to_cpu(rec->c_blkno);
 424                         else
 425                                 blkno = le64_to_cpu(bg->bg_next_group);
 426
 427                         if (gd_bh) {
 428                                 brelse(gd_bh);
 429                                 gd_bh = NULL;
 430                         }
 431
 432                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 433                         if (ret) {
 434                                 mlog_errno(ret);
 435                                 goto out;
 436                         }
 437
 438                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 439
 440                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 441                                                 le16_to_cpu(bg->bg_bits))) {
 442
 443                                 *ret_bh = gd_bh;
 444                                 *vict_bit = (vict_blkno - blkno) >>
 445                                                         bits_per_unit;
 446                                 mlog(0, "find the victim group: #%llu, "
 447                                      "total_bits: %u, vict_bit: %u\n",
 448                                      blkno, le16_to_cpu(bg->bg_bits),
 449                                      *vict_bit);
 450                                 goto out;
 451                         }
 452
 453                 } while (le64_to_cpu(bg->bg_next_group));
 454         }
 455
 456         ret = -EINVAL;
 457 out:
 458         brelse(ac_bh);
 459
 460         /*
 461          * caller has to release the gd_bh properly.
 462          */
 463         return ret;
 464 }
 465
 466 /*
 467  * XXX: helper to validate and adjust moving goal.
 468  */
 469 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 470                                                struct ocfs2_move_extents *range)
 471 {
 472         int ret, goal_bit = 0;
 473
 474         struct buffer_head *gd_bh = NULL;
 475         struct ocfs2_group_desc *bg;
 476         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 477         int c_to_b = 1 << (osb->s_clustersize_bits -
 478                                         inode->i_sb->s_blocksize_bits);
 479
 480         /*
 481          * make goal become cluster aligned.
 482          */
 483         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 484                                                       range->me_goal);
 485         /*
 486          * validate goal sits within global_bitmap, and return the victim
 487          * group desc
 488          */
 489         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 490                                             GLOBAL_BITMAP_SYSTEM_INODE,
 491                                             OCFS2_INVALID_SLOT,
 492                                             &goal_bit, &gd_bh);
 493         if (ret)
 494                 goto out;
 495
 496         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 497
 498         /*
 499          * moving goal is not allowd to start with a group desc blok(#0 blk)
 500          * let's compromise to the latter cluster.
 501          */
 502         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 503                 range->me_goal += c_to_b;
 504
 505         /*
 506          * movement is not gonna cross two groups.
 507          */
 508         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 509                                                                 range->me_len) {
 510                 ret = -EINVAL;
 511                 goto out;
 512         }
 513         /*
 514          * more exact validations/adjustments will be performed later during
 515          * moving operation for each extent range.
 516          */
 517         mlog(0, "extents get ready to be moved to #%llu block\n",
 518              range->me_goal);
 519
 520 out:
 521         brelse(gd_bh);
 522
 523         return ret;
 524 }
 525
 526 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 527                                     int *goal_bit, u32 move_len, u32 max_hop,
 528                                     u32 *phys_cpos)
 529 {
 530         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 531         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 532         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 533                                                  le64_to_cpu(gd->bg_blkno));
 534
 535         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 536
 537                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 538                 if (used) {
 539                         /*
 540                          * we even tried searching the free chunk by jumping
 541                          * a 'max_hop' distance, but still failed.
 542                          */
 543                         if ((i - base_bit) > max_hop) {
 544                                 *phys_cpos = 0;
 545                                 break;
 546                         }
 547
 548                         if (last_free_bits)
 549                                 last_free_bits = 0;
 550
 551                         continue;
 552                 } else
 553                         last_free_bits++;
 554
 555                 if (last_free_bits == move_len) {
 556                         *goal_bit = i;
 557                         *phys_cpos = base_cpos + i;
 558                         break;
 559                 }
 560         }
 561
 562         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 563 }
 564
 565 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 566                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 567                              u32 len, int ext_flags)
 568 {
 569         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 570         handle_t *handle;
 571         struct inode *inode = context->inode;
 572         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 573         struct inode *tl_inode = osb->osb_tl_inode;
 574         struct inode *gb_inode = NULL;
 575         struct buffer_head *gb_bh = NULL;
 576         struct buffer_head *gd_bh = NULL;
 577         struct ocfs2_group_desc *gd;
 578         struct ocfs2_refcount_tree *ref_tree = NULL;
 579         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 580                                                     context->range->me_threshold);
 581         u64 phys_blkno, new_phys_blkno;
 582
 583         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 584
 585         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 586
 587                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 588                          OCFS2_HAS_REFCOUNT_FL));
 589
 590                 BUG_ON(!context->refcount_loc);
 591
 592                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 593                                                &ref_tree, NULL);
 594                 if (ret) {
 595                         mlog_errno(ret);
 596                         return ret;
 597                 }
 598
 599                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 600                                                         context->refcount_loc,
 601                                                         phys_blkno,
 602                                                         len,
 603                                                         &credits,
 604                                                         &extra_blocks);
 605                 if (ret) {
 606                         mlog_errno(ret);
 607                         goto out;
 608                 }
 609         }
 610
 611         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, len, 1,
 612                                                  &context->meta_ac,
 613                                                  NULL, extra_blocks, &credits);
 614         if (ret) {
 615                 mlog_errno(ret);
 616                 goto out;
 617         }
 618
 619         /*
 620          * need to count 2 extra credits for global_bitmap inode and
 621          * group descriptor.
 622          */
 623         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 624
 625         /*
 626          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 627          * logic, while we still need to lock the global_bitmap.
 628          */
 629         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 630                                                OCFS2_INVALID_SLOT);
 631         if (!gb_inode) {
 632                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 633                 ret = -EIO;
 634                 goto out;
 635         }
 636
 637         mutex_lock(&gb_inode->i_mutex);
 638
 639         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 640         if (ret) {
 641                 mlog_errno(ret);
 642                 goto out_unlock_gb_mutex;
 643         }
 644
 645         mutex_lock(&tl_inode->i_mutex);
 646
 647         handle = ocfs2_start_trans(osb, credits);
 648         if (IS_ERR(handle)) {
 649                 ret = PTR_ERR(handle);
 650                 mlog_errno(ret);
 651                 goto out_unlock_tl_inode;
 652         }
 653
 654         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 655         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 656                                             GLOBAL_BITMAP_SYSTEM_INODE,
 657                                             OCFS2_INVALID_SLOT,
 658                                             &goal_bit, &gd_bh);
 659         if (ret) {
 660                 mlog_errno(ret);
 661                 goto out_commit;
 662         }
 663
 664         /*
 665          * probe the victim cluster group to find a proper
 666          * region to fit wanted movement, it even will perfrom
 667          * a best-effort attempt by compromising to a threshold
 668          * around the goal.
 669          */
 670         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 671                                 new_phys_cpos);
 672         if (!*new_phys_cpos) {
 673                 ret = -ENOSPC;
 674                 goto out_commit;
 675         }
 676
 677         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 678                                   *new_phys_cpos, ext_flags);
 679         if (ret) {
 680                 mlog_errno(ret);
 681                 goto out_commit;
 682         }
 683
 684         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 685         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 686                                                le16_to_cpu(gd->bg_chain));
 687         if (ret) {
 688                 mlog_errno(ret);
 689                 goto out_commit;
 690         }
 691
 692         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 693                                          goal_bit, len);
 694         if (ret) {
 695                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 696                                                le16_to_cpu(gd->bg_chain));
 697                 mlog_errno(ret);
 698         }
 699
 700         /*
 701          * Here we should write the new page out first if we are
 702          * in write-back mode.
 703          */
 704         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 705         if (ret)
 706                 mlog_errno(ret);
 707
 708 out_commit:
 709         ocfs2_commit_trans(osb, handle);
 710         brelse(gd_bh);
 711
 712 out_unlock_tl_inode:
 713         mutex_unlock(&tl_inode->i_mutex);
 714
 715         ocfs2_inode_unlock(gb_inode, 1);
 716 out_unlock_gb_mutex:
 717         mutex_unlock(&gb_inode->i_mutex);
 718         brelse(gb_bh);
 719         iput(gb_inode);
 720
 721 out:
 722         if (context->meta_ac) {
 723                 ocfs2_free_alloc_context(context->meta_ac);
 724                 context->meta_ac = NULL;
 725         }
 726
 727         if (ref_tree)
 728                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 729
 730         return ret;
 731 }
 732
 733 /*
 734  * Helper to calculate the defraging length in one run according to threshold.
 735  */
 736 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 737                                          u32 threshold, int *skip)
 738 {
 739         if ((*alloc_size + *len_defraged) < threshold) {
 740                 /*
 741                  * proceed defragmentation until we meet the thresh
 742                  */
 743                 *len_defraged += *alloc_size;
 744         } else if (*len_defraged == 0) {
 745                 /*
 746                  * XXX: skip a large extent.
 747                  */
 748                 *skip = 1;
 749         } else {
 750                 /*
 751                  * split this extent to coalesce with former pieces as
 752                  * to reach the threshold.
 753                  *
 754                  * we're done here with one cycle of defragmentation
 755                  * in a size of 'thresh', resetting 'len_defraged'
 756                  * forces a new defragmentation.
 757                  */
 758                 *alloc_size = threshold - *len_defraged;
 759                 *len_defraged = 0;
 760         }
 761 }
 762
 763 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 764                                 struct ocfs2_move_extents_context *context)
 765 {
 766         int ret = 0, flags, do_defrag, skip = 0;
 767         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 768         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 769
 770         struct inode *inode = context->inode;
 771         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 772         struct ocfs2_move_extents *range = context->range;
 773         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 774
 775         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 776                 return 0;
 777
 778         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 779                 return 0;
 780
 781         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 782
 783         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 784         ocfs2_init_dealloc_ctxt(&context->dealloc);
 785
 786         /*
 787          * TO-DO XXX:
 788          *
 789          * - xattr extents.
 790          */
 791
 792         do_defrag = context->auto_defrag;
 793
 794         /*
 795          * extents moving happens in unit of clusters, for the sake
 796          * of simplicity, we may ignore two clusters where 'byte_start'
 797          * and 'byte_start + len' were within.
 798          */
 799         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 800         len_to_move = (range->me_start + range->me_len) >>
 801                                                 osb->s_clustersize_bits;
 802         if (len_to_move >= move_start)
 803                 len_to_move -= move_start;
 804         else
 805                 len_to_move = 0;
 806
 807         if (do_defrag) {
 808                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 809                 if (defrag_thresh <= 1)
 810                         goto done;
 811         } else
 812                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 813                                                          range->me_goal);
 814
 815         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 816              "thresh: %u\n",
 817              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 818              (unsigned long long)range->me_start,
 819              (unsigned long long)range->me_len,
 820              move_start, len_to_move, defrag_thresh);
 821
 822         cpos = move_start;
 823         while (len_to_move) {
 824                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 825                                          &flags);
 826                 if (ret) {
 827                         mlog_errno(ret);
 828                         goto out;
 829                 }
 830
 831                 if (alloc_size > len_to_move)
 832                         alloc_size = len_to_move;
 833
 834                 /*
 835                  * XXX: how to deal with a hole:
 836                  *
 837                  * - skip the hole of course
 838                  * - force a new defragmentation
 839                  */
 840                 if (!phys_cpos) {
 841                         if (do_defrag)
 842                                 len_defraged = 0;
 843
 844                         goto next;
 845                 }
 846
 847                 if (do_defrag) {
 848                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 849                                                      defrag_thresh, &skip);
 850                         /*
 851                          * skip large extents
 852                          */
 853                         if (skip) {
 854                                 skip = 0;
 855                                 goto next;
 856                         }
 857
 858                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 859                              "alloc_size: %u, len_defraged: %u\n",
 860                              cpos, phys_cpos, alloc_size, len_defraged);
 861
 862                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 863                                                   &alloc_size, flags);
 864                 } else {
 865                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 866                                                 &new_phys_cpos, alloc_size,
 867                                                 flags);
 868
 869                         new_phys_cpos += alloc_size;
 870                 }
 871
 872                 if (ret < 0) {
 873                         mlog_errno(ret);
 874                         goto out;
 875                 }
 876
 877                 context->clusters_moved += alloc_size;
 878 next:
 879                 cpos += alloc_size;
 880                 len_to_move -= alloc_size;
 881         }
 882
 883 done:
 884         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 885
 886 out:
 887         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 888                                                       context->clusters_moved);
 889         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 890                                                        context->new_phys_cpos);
 891
 892         ocfs2_schedule_truncate_log_flush(osb, 1);
 893         ocfs2_run_deallocs(osb, &context->dealloc);
 894
 895         return ret;
 896 }
 897
 898 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 899 {
 900         int status;
 901         handle_t *handle;
 902         struct inode *inode = context->inode;
 903         struct ocfs2_dinode *di;
 904         struct buffer_head *di_bh = NULL;
 905         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 906
 907         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 908                 return -EROFS;
 909
 910         mutex_lock(&inode->i_mutex);
 911
 912         /*
 913          * This prevents concurrent writes from other nodes
 914          */
 915         status = ocfs2_rw_lock(inode, 1);
 916         if (status) {
 917                 mlog_errno(status);
 918                 goto out;
 919         }
 920
 921         status = ocfs2_inode_lock(inode, &di_bh, 1);
 922         if (status) {
 923                 mlog_errno(status);
 924                 goto out_rw_unlock;
 925         }
 926
 927         /*
 928          * rememer ip_xattr_sem also needs to be held if necessary
 929          */
 930         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 931
 932         status = __ocfs2_move_extents_range(di_bh, context);
 933
 934         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 935         if (status) {
 936                 mlog_errno(status);
 937                 goto out_inode_unlock;
 938         }
 939
 940         /*
 941          * We update ctime for these changes
 942          */
 943         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 944         if (IS_ERR(handle)) {
 945                 status = PTR_ERR(handle);
 946                 mlog_errno(status);
 947                 goto out_inode_unlock;
 948         }
 949
 950         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 951                                          OCFS2_JOURNAL_ACCESS_WRITE);
 952         if (status) {
 953                 mlog_errno(status);
 954                 goto out_commit;
 955         }
 956
 957         di = (struct ocfs2_dinode *)di_bh->b_data;
 958         inode->i_ctime = CURRENT_TIME;
 959         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 960         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 961         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 962
 963         ocfs2_journal_dirty(handle, di_bh);
 964
 965 out_commit:
 966         ocfs2_commit_trans(osb, handle);
 967
 968 out_inode_unlock:
 969         brelse(di_bh);
 970         ocfs2_inode_unlock(inode, 1);
 971 out_rw_unlock:
 972         ocfs2_rw_unlock(inode, 1);
 973 out:
 974         mutex_unlock(&inode->i_mutex);
 975
 976         return status;
 977 }
 978
 979 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 980 {
 981         int status;
 982
 983         struct inode *inode = file_inode(filp);
 984         struct ocfs2_move_extents range;
 985         struct ocfs2_move_extents_context *context;
 986
 987         if (!argp)
 988                 return -EINVAL;
 989
 990         status = mnt_want_write_file(filp);
 991         if (status)
 992                 return status;
 993
 994         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
 995                 status = -EPERM;
 996                 goto out_drop;
 997         }
 998
 999         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1000                 status = -EPERM;
1001                 goto out_drop;
1002         }
1003
1004         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1005         if (!context) {
1006                 status = -ENOMEM;
1007                 mlog_errno(status);
1008                 goto out_drop;
1009         }
1010
1011         context->inode = inode;
1012         context->file = filp;
1013
1014         if (copy_from_user(&range, argp, sizeof(range))) {
1015                 status = -EFAULT;
1016                 goto out_free;
1017         }
1018
1019         if (range.me_start > i_size_read(inode)) {
1020                 status = -EINVAL;
1021                 goto out_free;
1022         }
1023
1024         if (range.me_start + range.me_len > i_size_read(inode))
1025                         range.me_len = i_size_read(inode) - range.me_start;
1026
1027         context->range = &range;
1028
1029         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1030                 context->auto_defrag = 1;
1031                 /*
1032                  * ok, the default theshold for the defragmentation
1033                  * is 1M, since our maximum clustersize was 1M also.
1034                  * any thought?
1035                  */
1036                 if (!range.me_threshold)
1037                         range.me_threshold = 1024 * 1024;
1038
1039                 if (range.me_threshold > i_size_read(inode))
1040                         range.me_threshold = i_size_read(inode);
1041
1042                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1043                         context->partial = 1;
1044         } else {
1045                 /*
1046                  * first best-effort attempt to validate and adjust the goal
1047                  * (physical address in block), while it can't guarantee later
1048                  * operation can succeed all the time since global_bitmap may
1049                  * change a bit over time.
1050                  */
1051
1052                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1053                 if (status)
1054                         goto out_copy;
1055         }
1056
1057         status = ocfs2_move_extents(context);
1058         if (status)
1059                 mlog_errno(status);
1060 out_copy:
1061         /*
1062          * movement/defragmentation may end up being partially completed,
1063          * that's the reason why we need to return userspace the finished
1064          * length and new_offset even if failure happens somewhere.
1065          */
1066         if (copy_to_user(argp, &range, sizeof(range)))
1067                 status = -EFAULT;
1068
1069 out_free:
1070         kfree(context);
1071 out_drop:
1072         mnt_drop_write_file(filp);
1073
1074         return status;
1075 }