fs/ocfs2/move_extents.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * move_extents.c
   4  *
   5  * Copyright (C) 2011 Oracle.  All rights reserved.
   6  */
   7 #include <linux/fs.h>
   8 #include <linux/types.h>
   9 #include <linux/mount.h>
  10 #include <linux/swap.h>
  11
  12 #include <cluster/masklog.h>
  13
  14 #include "ocfs2.h"
  15 #include "ocfs2_ioctl.h"
  16
  17 #include "alloc.h"
  18 #include "localalloc.h"
  19 #include "aops.h"
  20 #include "dlmglue.h"
  21 #include "extent_map.h"
  22 #include "inode.h"
  23 #include "journal.h"
  24 #include "suballoc.h"
  25 #include "uptodate.h"
  26 #include "super.h"
  27 #include "dir.h"
  28 #include "buffer_head_io.h"
  29 #include "sysfile.h"
  30 #include "refcounttree.h"
  31 #include "move_extents.h"
  32
  33 struct ocfs2_move_extents_context {
  34         struct inode *inode;
  35         struct file *file;
  36         int auto_defrag;
  37         int partial;
  38         int credits;
  39         u32 new_phys_cpos;
  40         u32 clusters_moved;
  41         u64 refcount_loc;
  42         struct ocfs2_move_extents *range;
  43         struct ocfs2_extent_tree et;
  44         struct ocfs2_alloc_context *meta_ac;
  45         struct ocfs2_alloc_context *data_ac;
  46         struct ocfs2_cached_dealloc_ctxt dealloc;
  47 };
  48
  49 static int __ocfs2_move_extent(handle_t *handle,
  50                                struct ocfs2_move_extents_context *context,
  51                                u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
  52                                int ext_flags)
  53 {
  54         int ret = 0, index;
  55         struct inode *inode = context->inode;
  56         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  57         struct ocfs2_extent_rec *rec, replace_rec;
  58         struct ocfs2_path *path = NULL;
  59         struct ocfs2_extent_list *el;
  60         u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
  61         u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
  62
  63         ret = ocfs2_duplicate_clusters_by_page(handle, inode, cpos,
  64                                                p_cpos, new_p_cpos, len);
  65         if (ret) {
  66                 mlog_errno(ret);
  67                 goto out;
  68         }
  69
  70         memset(&replace_rec, 0, sizeof(replace_rec));
  71         replace_rec.e_cpos = cpu_to_le32(cpos);
  72         replace_rec.e_leaf_clusters = cpu_to_le16(len);
  73         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
  74                                                                    new_p_cpos));
  75
  76         path = ocfs2_new_path_from_et(&context->et);
  77         if (!path) {
  78                 ret = -ENOMEM;
  79                 mlog_errno(ret);
  80                 goto out;
  81         }
  82
  83         ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
  84         if (ret) {
  85                 mlog_errno(ret);
  86                 goto out;
  87         }
  88
  89         el = path_leaf_el(path);
  90
  91         index = ocfs2_search_extent_list(el, cpos);
  92         if (index == -1) {
  93                 ret = ocfs2_error(inode->i_sb,
  94                                   "Inode %llu has an extent at cpos %u which can no longer be found\n",
  95                                   (unsigned long long)ino, cpos);
  96                 goto out;
  97         }
  98
  99         rec = &el->l_recs[index];
 100
 101         BUG_ON(ext_flags != rec->e_flags);
 102         /*
 103          * after moving/defraging to new location, the extent is not going
 104          * to be refcounted anymore.
 105          */
 106         replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
 107
 108         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
 109                                       context->et.et_root_bh,
 110                                       OCFS2_JOURNAL_ACCESS_WRITE);
 111         if (ret) {
 112                 mlog_errno(ret);
 113                 goto out;
 114         }
 115
 116         ret = ocfs2_split_extent(handle, &context->et, path, index,
 117                                  &replace_rec, context->meta_ac,
 118                                  &context->dealloc);
 119         if (ret) {
 120                 mlog_errno(ret);
 121                 goto out;
 122         }
 123
 124         ocfs2_journal_dirty(handle, context->et.et_root_bh);
 125
 126         context->new_phys_cpos = new_p_cpos;
 127
 128         /*
 129          * need I to append truncate log for old clusters?
 130          */
 131         if (old_blkno) {
 132                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 133                         ret = ocfs2_decrease_refcount(inode, handle,
 134                                         ocfs2_blocks_to_clusters(osb->sb,
 135                                                                  old_blkno),
 136                                         len, context->meta_ac,
 137                                         &context->dealloc, 1);
 138                 else
 139                         ret = ocfs2_truncate_log_append(osb, handle,
 140                                                         old_blkno, len);
 141         }
 142
 143         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 144 out:
 145         ocfs2_free_path(path);
 146         return ret;
 147 }
 148
 149 /*
 150  * lock allocator, and reserve appropriate number of bits for
 151  * meta blocks.
 152  */
 153 static int ocfs2_lock_meta_allocator_move_extents(struct inode *inode,
 154                                         struct ocfs2_extent_tree *et,
 155                                         u32 clusters_to_move,
 156                                         u32 extents_to_split,
 157                                         struct ocfs2_alloc_context **meta_ac,
 158                                         int extra_blocks,
 159                                         int *credits)
 160 {
 161         int ret, num_free_extents;
 162         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 163         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 164
 165         num_free_extents = ocfs2_num_free_extents(et);
 166         if (num_free_extents < 0) {
 167                 ret = num_free_extents;
 168                 mlog_errno(ret);
 169                 goto out;
 170         }
 171
 172         if (!num_free_extents ||
 173             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 174                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 175
 176         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 177         if (ret) {
 178                 mlog_errno(ret);
 179                 goto out;
 180         }
 181
 182
 183         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 184
 185         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 186              extra_blocks, clusters_to_move, *credits);
 187 out:
 188         if (ret) {
 189                 if (*meta_ac) {
 190                         ocfs2_free_alloc_context(*meta_ac);
 191                         *meta_ac = NULL;
 192                 }
 193         }
 194
 195         return ret;
 196 }
 197
 198 /*
 199  * Using one journal handle to guarantee the data consistency in case
 200  * crash happens anywhere.
 201  *
 202  *  XXX: defrag can end up with finishing partial extent as requested,
 203  * due to not enough contiguous clusters can be found in allocator.
 204  */
 205 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 206                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 207 {
 208         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 209         handle_t *handle;
 210         struct inode *inode = context->inode;
 211         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 212         struct inode *tl_inode = osb->osb_tl_inode;
 213         struct ocfs2_refcount_tree *ref_tree = NULL;
 214         u32 new_phys_cpos, new_len;
 215         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 216         int need_free = 0;
 217
 218         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 219                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 220                 BUG_ON(!context->refcount_loc);
 221
 222                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 223                                                &ref_tree, NULL);
 224                 if (ret) {
 225                         mlog_errno(ret);
 226                         return ret;
 227                 }
 228
 229                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 230                                                         context->refcount_loc,
 231                                                         phys_blkno,
 232                                                         *len,
 233                                                         &credits,
 234                                                         &extra_blocks);
 235                 if (ret) {
 236                         mlog_errno(ret);
 237                         goto out;
 238                 }
 239         }
 240
 241         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 242                                                 *len, 1,
 243                                                 &context->meta_ac,
 244                                                 extra_blocks, &credits);
 245         if (ret) {
 246                 mlog_errno(ret);
 247                 goto out;
 248         }
 249
 250         /*
 251          * should be using allocation reservation strategy there?
 252          *
 253          * if (context->data_ac)
 254          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 255          */
 256
 257         inode_lock(tl_inode);
 258
 259         if (ocfs2_truncate_log_needs_flush(osb)) {
 260                 ret = __ocfs2_flush_truncate_log(osb);
 261                 if (ret < 0) {
 262                         mlog_errno(ret);
 263                         goto out_unlock_mutex;
 264                 }
 265         }
 266
 267         /*
 268          * Make sure ocfs2_reserve_cluster is called after
 269          * __ocfs2_flush_truncate_log, otherwise, dead lock may happen.
 270          *
 271          * If ocfs2_reserve_cluster is called
 272          * before __ocfs2_flush_truncate_log, dead lock on global bitmap
 273          * may happen.
 274          *
 275          */
 276         ret = ocfs2_reserve_clusters(osb, *len, &context->data_ac);
 277         if (ret) {
 278                 mlog_errno(ret);
 279                 goto out_unlock_mutex;
 280         }
 281
 282         handle = ocfs2_start_trans(osb, credits);
 283         if (IS_ERR(handle)) {
 284                 ret = PTR_ERR(handle);
 285                 mlog_errno(ret);
 286                 goto out_unlock_mutex;
 287         }
 288
 289         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 290                                      &new_phys_cpos, &new_len);
 291         if (ret) {
 292                 mlog_errno(ret);
 293                 goto out_commit;
 294         }
 295
 296         /*
 297          * allowing partial extent moving is kind of 'pros and cons', it makes
 298          * whole defragmentation less likely to fail, on the contrary, the bad
 299          * thing is it may make the fs even more fragmented after moving, let
 300          * userspace make a good decision here.
 301          */
 302         if (new_len != *len) {
 303                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 304                 if (!partial) {
 305                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 306                         ret = -ENOSPC;
 307                         need_free = 1;
 308                         goto out_commit;
 309                 }
 310         }
 311
 312         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 313              phys_cpos, new_phys_cpos);
 314
 315         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 316                                   new_phys_cpos, ext_flags);
 317         if (ret)
 318                 mlog_errno(ret);
 319
 320         if (partial && (new_len != *len))
 321                 *len = new_len;
 322
 323         /*
 324          * Here we should write the new page out first if we are
 325          * in write-back mode.
 326          */
 327         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 328         if (ret)
 329                 mlog_errno(ret);
 330
 331 out_commit:
 332         if (need_free && context->data_ac) {
 333                 struct ocfs2_alloc_context *data_ac = context->data_ac;
 334
 335                 if (context->data_ac->ac_which == OCFS2_AC_USE_LOCAL)
 336                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
 337                                         new_phys_cpos, new_len);
 338                 else
 339                         ocfs2_free_clusters(handle,
 340                                         data_ac->ac_inode,
 341                                         data_ac->ac_bh,
 342                                         ocfs2_clusters_to_blocks(osb->sb, new_phys_cpos),
 343                                         new_len);
 344         }
 345
 346         ocfs2_commit_trans(osb, handle);
 347
 348 out_unlock_mutex:
 349         inode_unlock(tl_inode);
 350
 351         if (context->data_ac) {
 352                 ocfs2_free_alloc_context(context->data_ac);
 353                 context->data_ac = NULL;
 354         }
 355
 356         if (context->meta_ac) {
 357                 ocfs2_free_alloc_context(context->meta_ac);
 358                 context->meta_ac = NULL;
 359         }
 360
 361 out:
 362         if (ref_tree)
 363                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 364
 365         return ret;
 366 }
 367
 368 /*
 369  * find the victim alloc group, where #blkno fits.
 370  */
 371 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 372                                          u64 vict_blkno,
 373                                          int type, int slot,
 374                                          int *vict_bit,
 375                                          struct buffer_head **ret_bh)
 376 {
 377         int ret, i, bits_per_unit = 0;
 378         u64 blkno;
 379         char namebuf[40];
 380
 381         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 382         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 383         struct ocfs2_chain_list *cl;
 384         struct ocfs2_chain_rec *rec;
 385         struct ocfs2_dinode *ac_dinode;
 386         struct ocfs2_group_desc *bg;
 387
 388         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 389         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 390                                          strlen(namebuf), &blkno);
 391         if (ret) {
 392                 ret = -ENOENT;
 393                 goto out;
 394         }
 395
 396         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 397         if (ret) {
 398                 mlog_errno(ret);
 399                 goto out;
 400         }
 401
 402         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 403         cl = &(ac_dinode->id2.i_chain);
 404         rec = &(cl->cl_recs[0]);
 405
 406         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 407                 bits_per_unit = osb->s_clustersize_bits -
 408                                         inode->i_sb->s_blocksize_bits;
 409         /*
 410          * 'vict_blkno' was out of the valid range.
 411          */
 412         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 413             (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 414                                 bits_per_unit))) {
 415                 ret = -EINVAL;
 416                 goto out;
 417         }
 418
 419         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 420
 421                 rec = &(cl->cl_recs[i]);
 422                 if (!rec)
 423                         continue;
 424
 425                 bg = NULL;
 426
 427                 do {
 428                         if (!bg)
 429                                 blkno = le64_to_cpu(rec->c_blkno);
 430                         else
 431                                 blkno = le64_to_cpu(bg->bg_next_group);
 432
 433                         if (gd_bh) {
 434                                 brelse(gd_bh);
 435                                 gd_bh = NULL;
 436                         }
 437
 438                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 439                         if (ret) {
 440                                 mlog_errno(ret);
 441                                 goto out;
 442                         }
 443
 444                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 445
 446                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 447                                                 le16_to_cpu(bg->bg_bits))) {
 448
 449                                 *ret_bh = gd_bh;
 450                                 *vict_bit = (vict_blkno - blkno) >>
 451                                                         bits_per_unit;
 452                                 mlog(0, "find the victim group: #%llu, "
 453                                      "total_bits: %u, vict_bit: %u\n",
 454                                      blkno, le16_to_cpu(bg->bg_bits),
 455                                      *vict_bit);
 456                                 goto out;
 457                         }
 458
 459                 } while (le64_to_cpu(bg->bg_next_group));
 460         }
 461
 462         ret = -EINVAL;
 463 out:
 464         brelse(ac_bh);
 465
 466         /*
 467          * caller has to release the gd_bh properly.
 468          */
 469         return ret;
 470 }
 471
 472 /*
 473  * XXX: helper to validate and adjust moving goal.
 474  */
 475 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 476                                                struct ocfs2_move_extents *range)
 477 {
 478         int ret, goal_bit = 0;
 479
 480         struct buffer_head *gd_bh = NULL;
 481         struct ocfs2_group_desc *bg;
 482         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 483         int c_to_b = 1 << (osb->s_clustersize_bits -
 484                                         inode->i_sb->s_blocksize_bits);
 485
 486         /*
 487          * make goal become cluster aligned.
 488          */
 489         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 490                                                       range->me_goal);
 491         /*
 492          * validate goal sits within global_bitmap, and return the victim
 493          * group desc
 494          */
 495         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 496                                             GLOBAL_BITMAP_SYSTEM_INODE,
 497                                             OCFS2_INVALID_SLOT,
 498                                             &goal_bit, &gd_bh);
 499         if (ret)
 500                 goto out;
 501
 502         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 503
 504         /*
 505          * moving goal is not allowd to start with a group desc blok(#0 blk)
 506          * let's compromise to the latter cluster.
 507          */
 508         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 509                 range->me_goal += c_to_b;
 510
 511         /*
 512          * movement is not gonna cross two groups.
 513          */
 514         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 515                                                                 range->me_len) {
 516                 ret = -EINVAL;
 517                 goto out;
 518         }
 519         /*
 520          * more exact validations/adjustments will be performed later during
 521          * moving operation for each extent range.
 522          */
 523         mlog(0, "extents get ready to be moved to #%llu block\n",
 524              range->me_goal);
 525
 526 out:
 527         brelse(gd_bh);
 528
 529         return ret;
 530 }
 531
 532 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 533                                     int *goal_bit, u32 move_len, u32 max_hop,
 534                                     u32 *phys_cpos)
 535 {
 536         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 537         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 538         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 539                                                  le64_to_cpu(gd->bg_blkno));
 540
 541         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 542
 543                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 544                 if (used) {
 545                         /*
 546                          * we even tried searching the free chunk by jumping
 547                          * a 'max_hop' distance, but still failed.
 548                          */
 549                         if ((i - base_bit) > max_hop) {
 550                                 *phys_cpos = 0;
 551                                 break;
 552                         }
 553
 554                         if (last_free_bits)
 555                                 last_free_bits = 0;
 556
 557                         continue;
 558                 } else
 559                         last_free_bits++;
 560
 561                 if (last_free_bits == move_len) {
 562                         *goal_bit = i;
 563                         *phys_cpos = base_cpos + i;
 564                         break;
 565                 }
 566         }
 567
 568         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 569 }
 570
 571 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 572                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 573                              u32 len, int ext_flags)
 574 {
 575         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 576         handle_t *handle;
 577         struct inode *inode = context->inode;
 578         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 579         struct inode *tl_inode = osb->osb_tl_inode;
 580         struct inode *gb_inode = NULL;
 581         struct buffer_head *gb_bh = NULL;
 582         struct buffer_head *gd_bh = NULL;
 583         struct ocfs2_group_desc *gd;
 584         struct ocfs2_refcount_tree *ref_tree = NULL;
 585         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 586                                                     context->range->me_threshold);
 587         u64 phys_blkno, new_phys_blkno;
 588
 589         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 590
 591         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 592                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 593                 BUG_ON(!context->refcount_loc);
 594
 595                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 596                                                &ref_tree, NULL);
 597                 if (ret) {
 598                         mlog_errno(ret);
 599                         return ret;
 600                 }
 601
 602                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 603                                                         context->refcount_loc,
 604                                                         phys_blkno,
 605                                                         len,
 606                                                         &credits,
 607                                                         &extra_blocks);
 608                 if (ret) {
 609                         mlog_errno(ret);
 610                         goto out;
 611                 }
 612         }
 613
 614         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 615                                                 len, 1,
 616                                                 &context->meta_ac,
 617                                                 extra_blocks, &credits);
 618         if (ret) {
 619                 mlog_errno(ret);
 620                 goto out;
 621         }
 622
 623         /*
 624          * need to count 2 extra credits for global_bitmap inode and
 625          * group descriptor.
 626          */
 627         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 628
 629         /*
 630          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 631          * logic, while we still need to lock the global_bitmap.
 632          */
 633         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 634                                                OCFS2_INVALID_SLOT);
 635         if (!gb_inode) {
 636                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 637                 ret = -EIO;
 638                 goto out;
 639         }
 640
 641         inode_lock(gb_inode);
 642
 643         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 644         if (ret) {
 645                 mlog_errno(ret);
 646                 goto out_unlock_gb_mutex;
 647         }
 648
 649         inode_lock(tl_inode);
 650
 651         handle = ocfs2_start_trans(osb, credits);
 652         if (IS_ERR(handle)) {
 653                 ret = PTR_ERR(handle);
 654                 mlog_errno(ret);
 655                 goto out_unlock_tl_inode;
 656         }
 657
 658         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 659         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 660                                             GLOBAL_BITMAP_SYSTEM_INODE,
 661                                             OCFS2_INVALID_SLOT,
 662                                             &goal_bit, &gd_bh);
 663         if (ret) {
 664                 mlog_errno(ret);
 665                 goto out_commit;
 666         }
 667
 668         /*
 669          * probe the victim cluster group to find a proper
 670          * region to fit wanted movement, it even will perfrom
 671          * a best-effort attempt by compromising to a threshold
 672          * around the goal.
 673          */
 674         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 675                                 new_phys_cpos);
 676         if (!*new_phys_cpos) {
 677                 ret = -ENOSPC;
 678                 goto out_commit;
 679         }
 680
 681         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 682                                   *new_phys_cpos, ext_flags);
 683         if (ret) {
 684                 mlog_errno(ret);
 685                 goto out_commit;
 686         }
 687
 688         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 689         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 690                                                le16_to_cpu(gd->bg_chain));
 691         if (ret) {
 692                 mlog_errno(ret);
 693                 goto out_commit;
 694         }
 695
 696         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 697                                          goal_bit, len);
 698         if (ret) {
 699                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 700                                                le16_to_cpu(gd->bg_chain));
 701                 mlog_errno(ret);
 702         }
 703
 704         /*
 705          * Here we should write the new page out first if we are
 706          * in write-back mode.
 707          */
 708         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 709         if (ret)
 710                 mlog_errno(ret);
 711
 712 out_commit:
 713         ocfs2_commit_trans(osb, handle);
 714         brelse(gd_bh);
 715
 716 out_unlock_tl_inode:
 717         inode_unlock(tl_inode);
 718
 719         ocfs2_inode_unlock(gb_inode, 1);
 720 out_unlock_gb_mutex:
 721         inode_unlock(gb_inode);
 722         brelse(gb_bh);
 723         iput(gb_inode);
 724
 725 out:
 726         if (context->meta_ac) {
 727                 ocfs2_free_alloc_context(context->meta_ac);
 728                 context->meta_ac = NULL;
 729         }
 730
 731         if (ref_tree)
 732                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 733
 734         return ret;
 735 }
 736
 737 /*
 738  * Helper to calculate the defraging length in one run according to threshold.
 739  */
 740 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 741                                          u32 threshold, int *skip)
 742 {
 743         if ((*alloc_size + *len_defraged) < threshold) {
 744                 /*
 745                  * proceed defragmentation until we meet the thresh
 746                  */
 747                 *len_defraged += *alloc_size;
 748         } else if (*len_defraged == 0) {
 749                 /*
 750                  * XXX: skip a large extent.
 751                  */
 752                 *skip = 1;
 753         } else {
 754                 /*
 755                  * split this extent to coalesce with former pieces as
 756                  * to reach the threshold.
 757                  *
 758                  * we're done here with one cycle of defragmentation
 759                  * in a size of 'thresh', resetting 'len_defraged'
 760                  * forces a new defragmentation.
 761                  */
 762                 *alloc_size = threshold - *len_defraged;
 763                 *len_defraged = 0;
 764         }
 765 }
 766
 767 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 768                                 struct ocfs2_move_extents_context *context)
 769 {
 770         int ret = 0, flags, do_defrag, skip = 0;
 771         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 772         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 773
 774         struct inode *inode = context->inode;
 775         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 776         struct ocfs2_move_extents *range = context->range;
 777         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 778
 779         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 780                 return 0;
 781
 782         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 783                 return 0;
 784
 785         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 786
 787         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 788         ocfs2_init_dealloc_ctxt(&context->dealloc);
 789
 790         /*
 791          * TO-DO XXX:
 792          *
 793          * - xattr extents.
 794          */
 795
 796         do_defrag = context->auto_defrag;
 797
 798         /*
 799          * extents moving happens in unit of clusters, for the sake
 800          * of simplicity, we may ignore two clusters where 'byte_start'
 801          * and 'byte_start + len' were within.
 802          */
 803         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 804         len_to_move = (range->me_start + range->me_len) >>
 805                                                 osb->s_clustersize_bits;
 806         if (len_to_move >= move_start)
 807                 len_to_move -= move_start;
 808         else
 809                 len_to_move = 0;
 810
 811         if (do_defrag) {
 812                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 813                 if (defrag_thresh <= 1)
 814                         goto done;
 815         } else
 816                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 817                                                          range->me_goal);
 818
 819         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 820              "thresh: %u\n",
 821              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 822              (unsigned long long)range->me_start,
 823              (unsigned long long)range->me_len,
 824              move_start, len_to_move, defrag_thresh);
 825
 826         cpos = move_start;
 827         while (len_to_move) {
 828                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 829                                          &flags);
 830                 if (ret) {
 831                         mlog_errno(ret);
 832                         goto out;
 833                 }
 834
 835                 if (alloc_size > len_to_move)
 836                         alloc_size = len_to_move;
 837
 838                 /*
 839                  * XXX: how to deal with a hole:
 840                  *
 841                  * - skip the hole of course
 842                  * - force a new defragmentation
 843                  */
 844                 if (!phys_cpos) {
 845                         if (do_defrag)
 846                                 len_defraged = 0;
 847
 848                         goto next;
 849                 }
 850
 851                 if (do_defrag) {
 852                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 853                                                      defrag_thresh, &skip);
 854                         /*
 855                          * skip large extents
 856                          */
 857                         if (skip) {
 858                                 skip = 0;
 859                                 goto next;
 860                         }
 861
 862                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 863                              "alloc_size: %u, len_defraged: %u\n",
 864                              cpos, phys_cpos, alloc_size, len_defraged);
 865
 866                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 867                                                   &alloc_size, flags);
 868                 } else {
 869                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 870                                                 &new_phys_cpos, alloc_size,
 871                                                 flags);
 872
 873                         new_phys_cpos += alloc_size;
 874                 }
 875
 876                 if (ret < 0) {
 877                         mlog_errno(ret);
 878                         goto out;
 879                 }
 880
 881                 context->clusters_moved += alloc_size;
 882 next:
 883                 cpos += alloc_size;
 884                 len_to_move -= alloc_size;
 885         }
 886
 887 done:
 888         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 889
 890 out:
 891         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 892                                                       context->clusters_moved);
 893         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 894                                                        context->new_phys_cpos);
 895
 896         ocfs2_schedule_truncate_log_flush(osb, 1);
 897         ocfs2_run_deallocs(osb, &context->dealloc);
 898
 899         return ret;
 900 }
 901
 902 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 903 {
 904         int status;
 905         handle_t *handle;
 906         struct inode *inode = context->inode;
 907         struct ocfs2_dinode *di;
 908         struct buffer_head *di_bh = NULL;
 909         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 910
 911         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 912                 return -EROFS;
 913
 914         inode_lock(inode);
 915
 916         /*
 917          * This prevents concurrent writes from other nodes
 918          */
 919         status = ocfs2_rw_lock(inode, 1);
 920         if (status) {
 921                 mlog_errno(status);
 922                 goto out;
 923         }
 924
 925         status = ocfs2_inode_lock(inode, &di_bh, 1);
 926         if (status) {
 927                 mlog_errno(status);
 928                 goto out_rw_unlock;
 929         }
 930
 931         /*
 932          * rememer ip_xattr_sem also needs to be held if necessary
 933          */
 934         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 935
 936         status = __ocfs2_move_extents_range(di_bh, context);
 937
 938         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 939         if (status) {
 940                 mlog_errno(status);
 941                 goto out_inode_unlock;
 942         }
 943
 944         /*
 945          * We update ctime for these changes
 946          */
 947         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 948         if (IS_ERR(handle)) {
 949                 status = PTR_ERR(handle);
 950                 mlog_errno(status);
 951                 goto out_inode_unlock;
 952         }
 953
 954         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 955                                          OCFS2_JOURNAL_ACCESS_WRITE);
 956         if (status) {
 957                 mlog_errno(status);
 958                 goto out_commit;
 959         }
 960
 961         di = (struct ocfs2_dinode *)di_bh->b_data;
 962         inode->i_ctime = current_time(inode);
 963         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 964         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 965         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 966
 967         ocfs2_journal_dirty(handle, di_bh);
 968
 969 out_commit:
 970         ocfs2_commit_trans(osb, handle);
 971
 972 out_inode_unlock:
 973         brelse(di_bh);
 974         ocfs2_inode_unlock(inode, 1);
 975 out_rw_unlock:
 976         ocfs2_rw_unlock(inode, 1);
 977 out:
 978         inode_unlock(inode);
 979
 980         return status;
 981 }
 982
 983 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 984 {
 985         int status;
 986
 987         struct inode *inode = file_inode(filp);
 988         struct ocfs2_move_extents range;
 989         struct ocfs2_move_extents_context *context;
 990
 991         if (!argp)
 992                 return -EINVAL;
 993
 994         status = mnt_want_write_file(filp);
 995         if (status)
 996                 return status;
 997
 998         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
 999                 status = -EPERM;
1000                 goto out_drop;
1001         }
1002
1003         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1004                 status = -EPERM;
1005                 goto out_drop;
1006         }
1007
1008         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1009         if (!context) {
1010                 status = -ENOMEM;
1011                 mlog_errno(status);
1012                 goto out_drop;
1013         }
1014
1015         context->inode = inode;
1016         context->file = filp;
1017
1018         if (copy_from_user(&range, argp, sizeof(range))) {
1019                 status = -EFAULT;
1020                 goto out_free;
1021         }
1022
1023         if (range.me_start > i_size_read(inode)) {
1024                 status = -EINVAL;
1025                 goto out_free;
1026         }
1027
1028         if (range.me_start + range.me_len > i_size_read(inode))
1029                         range.me_len = i_size_read(inode) - range.me_start;
1030
1031         context->range = &range;
1032
1033         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1034                 context->auto_defrag = 1;
1035                 /*
1036                  * ok, the default theshold for the defragmentation
1037                  * is 1M, since our maximum clustersize was 1M also.
1038                  * any thought?
1039                  */
1040                 if (!range.me_threshold)
1041                         range.me_threshold = 1024 * 1024;
1042
1043                 if (range.me_threshold > i_size_read(inode))
1044                         range.me_threshold = i_size_read(inode);
1045
1046                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1047                         context->partial = 1;
1048         } else {
1049                 /*
1050                  * first best-effort attempt to validate and adjust the goal
1051                  * (physical address in block), while it can't guarantee later
1052                  * operation can succeed all the time since global_bitmap may
1053                  * change a bit over time.
1054                  */
1055
1056                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1057                 if (status)
1058                         goto out_copy;
1059         }
1060
1061         status = ocfs2_move_extents(context);
1062         if (status)
1063                 mlog_errno(status);
1064 out_copy:
1065         /*
1066          * movement/defragmentation may end up being partially completed,
1067          * that's the reason why we need to return userspace the finished
1068          * length and new_offset even if failure happens somewhere.
1069          */
1070         if (copy_to_user(argp, &range, sizeof(range)))
1071                 status = -EFAULT;
1072
1073 out_free:
1074         kfree(context);
1075 out_drop:
1076         mnt_drop_write_file(filp);
1077
1078         return status;
1079 }