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