fs/ocfs2/move_extents.c

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