fs/ocfs2/move_extents.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * move_extents.c
   5  *
   6  * Copyright (C) 2011 Oracle.  All rights reserved.
   7  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public
  10  * License version 2 as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License for more details.
  16  */
  17 #include <linux/fs.h>
  18 #include <linux/types.h>
  19 #include <linux/mount.h>
  20 #include <linux/swap.h>
  21
  22 #include <cluster/masklog.h>
  23
  24 #include "ocfs2.h"
  25 #include "ocfs2_ioctl.h"
  26
  27 #include "alloc.h"
  28 #include "localalloc.h"
  29 #include "aops.h"
  30 #include "dlmglue.h"
  31 #include "extent_map.h"
  32 #include "inode.h"
  33 #include "journal.h"
  34 #include "suballoc.h"
  35 #include "uptodate.h"
  36 #include "super.h"
  37 #include "dir.h"
  38 #include "buffer_head_io.h"
  39 #include "sysfile.h"
  40 #include "refcounttree.h"
  41 #include "move_extents.h"
  42
  43 struct ocfs2_move_extents_context {
  44         struct inode *inode;
  45         struct file *file;
  46         int auto_defrag;
  47         int partial;
  48         int credits;
  49         u32 new_phys_cpos;
  50         u32 clusters_moved;
  51         u64 refcount_loc;
  52         struct ocfs2_move_extents *range;
  53         struct ocfs2_extent_tree et;
  54         struct ocfs2_alloc_context *meta_ac;
  55         struct ocfs2_alloc_context *data_ac;
  56         struct ocfs2_cached_dealloc_ctxt dealloc;
  57 };
  58
  59 static int __ocfs2_move_extent(handle_t *handle,
  60                                struct ocfs2_move_extents_context *context,
  61                                u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
  62                                int ext_flags)
  63 {
  64         int ret = 0, index;
  65         struct inode *inode = context->inode;
  66         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  67         struct ocfs2_extent_rec *rec, replace_rec;
  68         struct ocfs2_path *path = NULL;
  69         struct ocfs2_extent_list *el;
  70         u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
  71         u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
  72
  73         ret = ocfs2_duplicate_clusters_by_page(handle, inode, cpos,
  74                                                p_cpos, new_p_cpos, len);
  75         if (ret) {
  76                 mlog_errno(ret);
  77                 goto out;
  78         }
  79
  80         memset(&replace_rec, 0, sizeof(replace_rec));
  81         replace_rec.e_cpos = cpu_to_le32(cpos);
  82         replace_rec.e_leaf_clusters = cpu_to_le16(len);
  83         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
  84                                                                    new_p_cpos));
  85
  86         path = ocfs2_new_path_from_et(&context->et);
  87         if (!path) {
  88                 ret = -ENOMEM;
  89                 mlog_errno(ret);
  90                 goto out;
  91         }
  92
  93         ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
  94         if (ret) {
  95                 mlog_errno(ret);
  96                 goto out;
  97         }
  98
  99         el = path_leaf_el(path);
 100
 101         index = ocfs2_search_extent_list(el, cpos);
 102         if (index == -1) {
 103                 ret = ocfs2_error(inode->i_sb,
 104                                   "Inode %llu has an extent at cpos %u which can no longer be found\n",
 105                                   (unsigned long long)ino, cpos);
 106                 goto out;
 107         }
 108
 109         rec = &el->l_recs[index];
 110
 111         BUG_ON(ext_flags != rec->e_flags);
 112         /*
 113          * after moving/defraging to new location, the extent is not going
 114          * to be refcounted anymore.
 115          */
 116         replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
 117
 118         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
 119                                       context->et.et_root_bh,
 120                                       OCFS2_JOURNAL_ACCESS_WRITE);
 121         if (ret) {
 122                 mlog_errno(ret);
 123                 goto out;
 124         }
 125
 126         ret = ocfs2_split_extent(handle, &context->et, path, index,
 127                                  &replace_rec, context->meta_ac,
 128                                  &context->dealloc);
 129         if (ret) {
 130                 mlog_errno(ret);
 131                 goto out;
 132         }
 133
 134         ocfs2_journal_dirty(handle, context->et.et_root_bh);
 135
 136         context->new_phys_cpos = new_p_cpos;
 137
 138         /*
 139          * need I to append truncate log for old clusters?
 140          */
 141         if (old_blkno) {
 142                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 143                         ret = ocfs2_decrease_refcount(inode, handle,
 144                                         ocfs2_blocks_to_clusters(osb->sb,
 145                                                                  old_blkno),
 146                                         len, context->meta_ac,
 147                                         &context->dealloc, 1);
 148                 else
 149                         ret = ocfs2_truncate_log_append(osb, handle,
 150                                                         old_blkno, len);
 151         }
 152
 153         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 154 out:
 155         ocfs2_free_path(path);
 156         return ret;
 157 }
 158
 159 /*
 160  * lock allocators, and reserving appropriate number of bits for
 161  * meta blocks and data clusters.
 162  *
 163  * in some cases, we don't need to reserve clusters, just let data_ac
 164  * be NULL.
 165  */
 166 static int ocfs2_lock_allocators_move_extents(struct inode *inode,
 167                                         struct ocfs2_extent_tree *et,
 168                                         u32 clusters_to_move,
 169                                         u32 extents_to_split,
 170                                         struct ocfs2_alloc_context **meta_ac,
 171                                         struct ocfs2_alloc_context **data_ac,
 172                                         int extra_blocks,
 173                                         int *credits)
 174 {
 175         int ret, num_free_extents;
 176         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 177         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 178
 179         num_free_extents = ocfs2_num_free_extents(et);
 180         if (num_free_extents < 0) {
 181                 ret = num_free_extents;
 182                 mlog_errno(ret);
 183                 goto out;
 184         }
 185
 186         if (!num_free_extents ||
 187             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 188                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 189
 190         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 191         if (ret) {
 192                 mlog_errno(ret);
 193                 goto out;
 194         }
 195
 196         if (data_ac) {
 197                 ret = ocfs2_reserve_clusters(osb, clusters_to_move, data_ac);
 198                 if (ret) {
 199                         mlog_errno(ret);
 200                         goto out;
 201                 }
 202         }
 203
 204         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 205
 206         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 207              extra_blocks, clusters_to_move, *credits);
 208 out:
 209         if (ret) {
 210                 if (*meta_ac) {
 211                         ocfs2_free_alloc_context(*meta_ac);
 212                         *meta_ac = NULL;
 213                 }
 214         }
 215
 216         return ret;
 217 }
 218
 219 /*
 220  * Using one journal handle to guarantee the data consistency in case
 221  * crash happens anywhere.
 222  *
 223  *  XXX: defrag can end up with finishing partial extent as requested,
 224  * due to not enough contiguous clusters can be found in allocator.
 225  */
 226 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 227                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 228 {
 229         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 230         handle_t *handle;
 231         struct inode *inode = context->inode;
 232         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 233         struct inode *tl_inode = osb->osb_tl_inode;
 234         struct ocfs2_refcount_tree *ref_tree = NULL;
 235         u32 new_phys_cpos, new_len;
 236         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 237         int need_free = 0;
 238
 239         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 240                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 241                 BUG_ON(!context->refcount_loc);
 242
 243                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 244                                                &ref_tree, NULL);
 245                 if (ret) {
 246                         mlog_errno(ret);
 247                         return ret;
 248                 }
 249
 250                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 251                                                         context->refcount_loc,
 252                                                         phys_blkno,
 253                                                         *len,
 254                                                         &credits,
 255                                                         &extra_blocks);
 256                 if (ret) {
 257                         mlog_errno(ret);
 258                         goto out;
 259                 }
 260         }
 261
 262         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, *len, 1,
 263                                                  &context->meta_ac,
 264                                                  &context->data_ac,
 265                                                  extra_blocks, &credits);
 266         if (ret) {
 267                 mlog_errno(ret);
 268                 goto out;
 269         }
 270
 271         /*
 272          * should be using allocation reservation strategy there?
 273          *
 274          * if (context->data_ac)
 275          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 276          */
 277
 278         inode_lock(tl_inode);
 279
 280         if (ocfs2_truncate_log_needs_flush(osb)) {
 281                 ret = __ocfs2_flush_truncate_log(osb);
 282                 if (ret < 0) {
 283                         mlog_errno(ret);
 284                         goto out_unlock_mutex;
 285                 }
 286         }
 287
 288         handle = ocfs2_start_trans(osb, credits);
 289         if (IS_ERR(handle)) {
 290                 ret = PTR_ERR(handle);
 291                 mlog_errno(ret);
 292                 goto out_unlock_mutex;
 293         }
 294
 295         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 296                                      &new_phys_cpos, &new_len);
 297         if (ret) {
 298                 mlog_errno(ret);
 299                 goto out_commit;
 300         }
 301
 302         /*
 303          * allowing partial extent moving is kind of 'pros and cons', it makes
 304          * whole defragmentation less likely to fail, on the contrary, the bad
 305          * thing is it may make the fs even more fragmented after moving, let
 306          * userspace make a good decision here.
 307          */
 308         if (new_len != *len) {
 309                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 310                 if (!partial) {
 311                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 312                         ret = -ENOSPC;
 313                         need_free = 1;
 314                         goto out_commit;
 315                 }
 316         }
 317
 318         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 319              phys_cpos, new_phys_cpos);
 320
 321         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 322                                   new_phys_cpos, ext_flags);
 323         if (ret)
 324                 mlog_errno(ret);
 325
 326         if (partial && (new_len != *len))
 327                 *len = new_len;
 328
 329         /*
 330          * Here we should write the new page out first if we are
 331          * in write-back mode.
 332          */
 333         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 334         if (ret)
 335                 mlog_errno(ret);
 336
 337 out_commit:
 338         if (need_free && context->data_ac) {
 339                 struct ocfs2_alloc_context *data_ac = context->data_ac;
 340
 341                 if (context->data_ac->ac_which == OCFS2_AC_USE_LOCAL)
 342                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
 343                                         new_phys_cpos, new_len);
 344                 else
 345                         ocfs2_free_clusters(handle,
 346                                         data_ac->ac_inode,
 347                                         data_ac->ac_bh,
 348                                         ocfs2_clusters_to_blocks(osb->sb, new_phys_cpos),
 349                                         new_len);
 350         }
 351
 352         ocfs2_commit_trans(osb, handle);
 353
 354 out_unlock_mutex:
 355         inode_unlock(tl_inode);
 356
 357         if (context->data_ac) {
 358                 ocfs2_free_alloc_context(context->data_ac);
 359                 context->data_ac = NULL;
 360         }
 361
 362         if (context->meta_ac) {
 363                 ocfs2_free_alloc_context(context->meta_ac);
 364                 context->meta_ac = NULL;
 365         }
 366
 367 out:
 368         if (ref_tree)
 369                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 370
 371         return ret;
 372 }
 373
 374 /*
 375  * find the victim alloc group, where #blkno fits.
 376  */
 377 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 378                                          u64 vict_blkno,
 379                                          int type, int slot,
 380                                          int *vict_bit,
 381                                          struct buffer_head **ret_bh)
 382 {
 383         int ret, i, bits_per_unit = 0;
 384         u64 blkno;
 385         char namebuf[40];
 386
 387         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 388         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 389         struct ocfs2_chain_list *cl;
 390         struct ocfs2_chain_rec *rec;
 391         struct ocfs2_dinode *ac_dinode;
 392         struct ocfs2_group_desc *bg;
 393
 394         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 395         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 396                                          strlen(namebuf), &blkno);
 397         if (ret) {
 398                 ret = -ENOENT;
 399                 goto out;
 400         }
 401
 402         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 403         if (ret) {
 404                 mlog_errno(ret);
 405                 goto out;
 406         }
 407
 408         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 409         cl = &(ac_dinode->id2.i_chain);
 410         rec = &(cl->cl_recs[0]);
 411
 412         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 413                 bits_per_unit = osb->s_clustersize_bits -
 414                                         inode->i_sb->s_blocksize_bits;
 415         /*
 416          * 'vict_blkno' was out of the valid range.
 417          */
 418         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 419             (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 420                                 bits_per_unit))) {
 421                 ret = -EINVAL;
 422                 goto out;
 423         }
 424
 425         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 426
 427                 rec = &(cl->cl_recs[i]);
 428                 if (!rec)
 429                         continue;
 430
 431                 bg = NULL;
 432
 433                 do {
 434                         if (!bg)
 435                                 blkno = le64_to_cpu(rec->c_blkno);
 436                         else
 437                                 blkno = le64_to_cpu(bg->bg_next_group);
 438
 439                         if (gd_bh) {
 440                                 brelse(gd_bh);
 441                                 gd_bh = NULL;
 442                         }
 443
 444                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 445                         if (ret) {
 446                                 mlog_errno(ret);
 447                                 goto out;
 448                         }
 449
 450                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 451
 452                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 453                                                 le16_to_cpu(bg->bg_bits))) {
 454
 455                                 *ret_bh = gd_bh;
 456                                 *vict_bit = (vict_blkno - blkno) >>
 457                                                         bits_per_unit;
 458                                 mlog(0, "find the victim group: #%llu, "
 459                                      "total_bits: %u, vict_bit: %u\n",
 460                                      blkno, le16_to_cpu(bg->bg_bits),
 461                                      *vict_bit);
 462                                 goto out;
 463                         }
 464
 465                 } while (le64_to_cpu(bg->bg_next_group));
 466         }
 467
 468         ret = -EINVAL;
 469 out:
 470         brelse(ac_bh);
 471
 472         /*
 473          * caller has to release the gd_bh properly.
 474          */
 475         return ret;
 476 }
 477
 478 /*
 479  * XXX: helper to validate and adjust moving goal.
 480  */
 481 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 482                                                struct ocfs2_move_extents *range)
 483 {
 484         int ret, goal_bit = 0;
 485
 486         struct buffer_head *gd_bh = NULL;
 487         struct ocfs2_group_desc *bg;
 488         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 489         int c_to_b = 1 << (osb->s_clustersize_bits -
 490                                         inode->i_sb->s_blocksize_bits);
 491
 492         /*
 493          * make goal become cluster aligned.
 494          */
 495         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 496                                                       range->me_goal);
 497         /*
 498          * validate goal sits within global_bitmap, and return the victim
 499          * group desc
 500          */
 501         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 502                                             GLOBAL_BITMAP_SYSTEM_INODE,
 503                                             OCFS2_INVALID_SLOT,
 504                                             &goal_bit, &gd_bh);
 505         if (ret)
 506                 goto out;
 507
 508         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 509
 510         /*
 511          * moving goal is not allowd to start with a group desc blok(#0 blk)
 512          * let's compromise to the latter cluster.
 513          */
 514         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 515                 range->me_goal += c_to_b;
 516
 517         /*
 518          * movement is not gonna cross two groups.
 519          */
 520         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 521                                                                 range->me_len) {
 522                 ret = -EINVAL;
 523                 goto out;
 524         }
 525         /*
 526          * more exact validations/adjustments will be performed later during
 527          * moving operation for each extent range.
 528          */
 529         mlog(0, "extents get ready to be moved to #%llu block\n",
 530              range->me_goal);
 531
 532 out:
 533         brelse(gd_bh);
 534
 535         return ret;
 536 }
 537
 538 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 539                                     int *goal_bit, u32 move_len, u32 max_hop,
 540                                     u32 *phys_cpos)
 541 {
 542         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 543         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 544         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 545                                                  le64_to_cpu(gd->bg_blkno));
 546
 547         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 548
 549                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 550                 if (used) {
 551                         /*
 552                          * we even tried searching the free chunk by jumping
 553                          * a 'max_hop' distance, but still failed.
 554                          */
 555                         if ((i - base_bit) > max_hop) {
 556                                 *phys_cpos = 0;
 557                                 break;
 558                         }
 559
 560                         if (last_free_bits)
 561                                 last_free_bits = 0;
 562
 563                         continue;
 564                 } else
 565                         last_free_bits++;
 566
 567                 if (last_free_bits == move_len) {
 568                         *goal_bit = i;
 569                         *phys_cpos = base_cpos + i;
 570                         break;
 571                 }
 572         }
 573
 574         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 575 }
 576
 577 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 578                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 579                              u32 len, int ext_flags)
 580 {
 581         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 582         handle_t *handle;
 583         struct inode *inode = context->inode;
 584         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 585         struct inode *tl_inode = osb->osb_tl_inode;
 586         struct inode *gb_inode = NULL;
 587         struct buffer_head *gb_bh = NULL;
 588         struct buffer_head *gd_bh = NULL;
 589         struct ocfs2_group_desc *gd;
 590         struct ocfs2_refcount_tree *ref_tree = NULL;
 591         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 592                                                     context->range->me_threshold);
 593         u64 phys_blkno, new_phys_blkno;
 594
 595         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 596
 597         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 598                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 599                 BUG_ON(!context->refcount_loc);
 600
 601                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 602                                                &ref_tree, NULL);
 603                 if (ret) {
 604                         mlog_errno(ret);
 605                         return ret;
 606                 }
 607
 608                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 609                                                         context->refcount_loc,
 610                                                         phys_blkno,
 611                                                         len,
 612                                                         &credits,
 613                                                         &extra_blocks);
 614                 if (ret) {
 615                         mlog_errno(ret);
 616                         goto out;
 617                 }
 618         }
 619
 620         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, len, 1,
 621                                                  &context->meta_ac,
 622                                                  NULL, extra_blocks, &credits);
 623         if (ret) {
 624                 mlog_errno(ret);
 625                 goto out;
 626         }
 627
 628         /*
 629          * need to count 2 extra credits for global_bitmap inode and
 630          * group descriptor.
 631          */
 632         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 633
 634         /*
 635          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 636          * logic, while we still need to lock the global_bitmap.
 637          */
 638         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 639                                                OCFS2_INVALID_SLOT);
 640         if (!gb_inode) {
 641                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 642                 ret = -EIO;
 643                 goto out;
 644         }
 645
 646         inode_lock(gb_inode);
 647
 648         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 649         if (ret) {
 650                 mlog_errno(ret);
 651                 goto out_unlock_gb_mutex;
 652         }
 653
 654         inode_lock(tl_inode);
 655
 656         handle = ocfs2_start_trans(osb, credits);
 657         if (IS_ERR(handle)) {
 658                 ret = PTR_ERR(handle);
 659                 mlog_errno(ret);
 660                 goto out_unlock_tl_inode;
 661         }
 662
 663         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 664         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 665                                             GLOBAL_BITMAP_SYSTEM_INODE,
 666                                             OCFS2_INVALID_SLOT,
 667                                             &goal_bit, &gd_bh);
 668         if (ret) {
 669                 mlog_errno(ret);
 670                 goto out_commit;
 671         }
 672
 673         /*
 674          * probe the victim cluster group to find a proper
 675          * region to fit wanted movement, it even will perfrom
 676          * a best-effort attempt by compromising to a threshold
 677          * around the goal.
 678          */
 679         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 680                                 new_phys_cpos);
 681         if (!*new_phys_cpos) {
 682                 ret = -ENOSPC;
 683                 goto out_commit;
 684         }
 685
 686         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 687                                   *new_phys_cpos, ext_flags);
 688         if (ret) {
 689                 mlog_errno(ret);
 690                 goto out_commit;
 691         }
 692
 693         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 694         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 695                                                le16_to_cpu(gd->bg_chain));
 696         if (ret) {
 697                 mlog_errno(ret);
 698                 goto out_commit;
 699         }
 700
 701         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 702                                          goal_bit, len);
 703         if (ret) {
 704                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 705                                                le16_to_cpu(gd->bg_chain));
 706                 mlog_errno(ret);
 707         }
 708
 709         /*
 710          * Here we should write the new page out first if we are
 711          * in write-back mode.
 712          */
 713         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 714         if (ret)
 715                 mlog_errno(ret);
 716
 717 out_commit:
 718         ocfs2_commit_trans(osb, handle);
 719         brelse(gd_bh);
 720
 721 out_unlock_tl_inode:
 722         inode_unlock(tl_inode);
 723
 724         ocfs2_inode_unlock(gb_inode, 1);
 725 out_unlock_gb_mutex:
 726         inode_unlock(gb_inode);
 727         brelse(gb_bh);
 728         iput(gb_inode);
 729
 730 out:
 731         if (context->meta_ac) {
 732                 ocfs2_free_alloc_context(context->meta_ac);
 733                 context->meta_ac = NULL;
 734         }
 735
 736         if (ref_tree)
 737                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 738
 739         return ret;
 740 }
 741
 742 /*
 743  * Helper to calculate the defraging length in one run according to threshold.
 744  */
 745 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 746                                          u32 threshold, int *skip)
 747 {
 748         if ((*alloc_size + *len_defraged) < threshold) {
 749                 /*
 750                  * proceed defragmentation until we meet the thresh
 751                  */
 752                 *len_defraged += *alloc_size;
 753         } else if (*len_defraged == 0) {
 754                 /*
 755                  * XXX: skip a large extent.
 756                  */
 757                 *skip = 1;
 758         } else {
 759                 /*
 760                  * split this extent to coalesce with former pieces as
 761                  * to reach the threshold.
 762                  *
 763                  * we're done here with one cycle of defragmentation
 764                  * in a size of 'thresh', resetting 'len_defraged'
 765                  * forces a new defragmentation.
 766                  */
 767                 *alloc_size = threshold - *len_defraged;
 768                 *len_defraged = 0;
 769         }
 770 }
 771
 772 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 773                                 struct ocfs2_move_extents_context *context)
 774 {
 775         int ret = 0, flags, do_defrag, skip = 0;
 776         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 777         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 778
 779         struct inode *inode = context->inode;
 780         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 781         struct ocfs2_move_extents *range = context->range;
 782         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 783
 784         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 785                 return 0;
 786
 787         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 788                 return 0;
 789
 790         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 791
 792         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 793         ocfs2_init_dealloc_ctxt(&context->dealloc);
 794
 795         /*
 796          * TO-DO XXX:
 797          *
 798          * - xattr extents.
 799          */
 800
 801         do_defrag = context->auto_defrag;
 802
 803         /*
 804          * extents moving happens in unit of clusters, for the sake
 805          * of simplicity, we may ignore two clusters where 'byte_start'
 806          * and 'byte_start + len' were within.
 807          */
 808         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 809         len_to_move = (range->me_start + range->me_len) >>
 810                                                 osb->s_clustersize_bits;
 811         if (len_to_move >= move_start)
 812                 len_to_move -= move_start;
 813         else
 814                 len_to_move = 0;
 815
 816         if (do_defrag) {
 817                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 818                 if (defrag_thresh <= 1)
 819                         goto done;
 820         } else
 821                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 822                                                          range->me_goal);
 823
 824         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 825              "thresh: %u\n",
 826              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 827              (unsigned long long)range->me_start,
 828              (unsigned long long)range->me_len,
 829              move_start, len_to_move, defrag_thresh);
 830
 831         cpos = move_start;
 832         while (len_to_move) {
 833                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 834                                          &flags);
 835                 if (ret) {
 836                         mlog_errno(ret);
 837                         goto out;
 838                 }
 839
 840                 if (alloc_size > len_to_move)
 841                         alloc_size = len_to_move;
 842
 843                 /*
 844                  * XXX: how to deal with a hole:
 845                  *
 846                  * - skip the hole of course
 847                  * - force a new defragmentation
 848                  */
 849                 if (!phys_cpos) {
 850                         if (do_defrag)
 851                                 len_defraged = 0;
 852
 853                         goto next;
 854                 }
 855
 856                 if (do_defrag) {
 857                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 858                                                      defrag_thresh, &skip);
 859                         /*
 860                          * skip large extents
 861                          */
 862                         if (skip) {
 863                                 skip = 0;
 864                                 goto next;
 865                         }
 866
 867                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 868                              "alloc_size: %u, len_defraged: %u\n",
 869                              cpos, phys_cpos, alloc_size, len_defraged);
 870
 871                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 872                                                   &alloc_size, flags);
 873                 } else {
 874                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 875                                                 &new_phys_cpos, alloc_size,
 876                                                 flags);
 877
 878                         new_phys_cpos += alloc_size;
 879                 }
 880
 881                 if (ret < 0) {
 882                         mlog_errno(ret);
 883                         goto out;
 884                 }
 885
 886                 context->clusters_moved += alloc_size;
 887 next:
 888                 cpos += alloc_size;
 889                 len_to_move -= alloc_size;
 890         }
 891
 892 done:
 893         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 894
 895 out:
 896         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 897                                                       context->clusters_moved);
 898         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 899                                                        context->new_phys_cpos);
 900
 901         ocfs2_schedule_truncate_log_flush(osb, 1);
 902         ocfs2_run_deallocs(osb, &context->dealloc);
 903
 904         return ret;
 905 }
 906
 907 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 908 {
 909         int status;
 910         handle_t *handle;
 911         struct inode *inode = context->inode;
 912         struct ocfs2_dinode *di;
 913         struct buffer_head *di_bh = NULL;
 914         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 915
 916         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 917                 return -EROFS;
 918
 919         inode_lock(inode);
 920
 921         /*
 922          * This prevents concurrent writes from other nodes
 923          */
 924         status = ocfs2_rw_lock(inode, 1);
 925         if (status) {
 926                 mlog_errno(status);
 927                 goto out;
 928         }
 929
 930         status = ocfs2_inode_lock(inode, &di_bh, 1);
 931         if (status) {
 932                 mlog_errno(status);
 933                 goto out_rw_unlock;
 934         }
 935
 936         /*
 937          * rememer ip_xattr_sem also needs to be held if necessary
 938          */
 939         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 940
 941         status = __ocfs2_move_extents_range(di_bh, context);
 942
 943         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 944         if (status) {
 945                 mlog_errno(status);
 946                 goto out_inode_unlock;
 947         }
 948
 949         /*
 950          * We update ctime for these changes
 951          */
 952         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 953         if (IS_ERR(handle)) {
 954                 status = PTR_ERR(handle);
 955                 mlog_errno(status);
 956                 goto out_inode_unlock;
 957         }
 958
 959         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 960                                          OCFS2_JOURNAL_ACCESS_WRITE);
 961         if (status) {
 962                 mlog_errno(status);
 963                 goto out_commit;
 964         }
 965
 966         di = (struct ocfs2_dinode *)di_bh->b_data;
 967         inode->i_ctime = current_time(inode);
 968         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 969         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 970         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 971
 972         ocfs2_journal_dirty(handle, di_bh);
 973
 974 out_commit:
 975         ocfs2_commit_trans(osb, handle);
 976
 977 out_inode_unlock:
 978         brelse(di_bh);
 979         ocfs2_inode_unlock(inode, 1);
 980 out_rw_unlock:
 981         ocfs2_rw_unlock(inode, 1);
 982 out:
 983         inode_unlock(inode);
 984
 985         return status;
 986 }
 987
 988 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 989 {
 990         int status;
 991
 992         struct inode *inode = file_inode(filp);
 993         struct ocfs2_move_extents range;
 994         struct ocfs2_move_extents_context *context;
 995
 996         if (!argp)
 997                 return -EINVAL;
 998
 999         status = mnt_want_write_file(filp);
1000         if (status)
1001                 return status;
1002
1003         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
1004                 status = -EPERM;
1005                 goto out_drop;
1006         }
1007
1008         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1009                 status = -EPERM;
1010                 goto out_drop;
1011         }
1012
1013         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1014         if (!context) {
1015                 status = -ENOMEM;
1016                 mlog_errno(status);
1017                 goto out_drop;
1018         }
1019
1020         context->inode = inode;
1021         context->file = filp;
1022
1023         if (copy_from_user(&range, argp, sizeof(range))) {
1024                 status = -EFAULT;
1025                 goto out_free;
1026         }
1027
1028         if (range.me_start > i_size_read(inode)) {
1029                 status = -EINVAL;
1030                 goto out_free;
1031         }
1032
1033         if (range.me_start + range.me_len > i_size_read(inode))
1034                         range.me_len = i_size_read(inode) - range.me_start;
1035
1036         context->range = &range;
1037
1038         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1039                 context->auto_defrag = 1;
1040                 /*
1041                  * ok, the default theshold for the defragmentation
1042                  * is 1M, since our maximum clustersize was 1M also.
1043                  * any thought?
1044                  */
1045                 if (!range.me_threshold)
1046                         range.me_threshold = 1024 * 1024;
1047
1048                 if (range.me_threshold > i_size_read(inode))
1049                         range.me_threshold = i_size_read(inode);
1050
1051                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1052                         context->partial = 1;
1053         } else {
1054                 /*
1055                  * first best-effort attempt to validate and adjust the goal
1056                  * (physical address in block), while it can't guarantee later
1057                  * operation can succeed all the time since global_bitmap may
1058                  * change a bit over time.
1059                  */
1060
1061                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1062                 if (status)
1063                         goto out_copy;
1064         }
1065
1066         status = ocfs2_move_extents(context);
1067         if (status)
1068                 mlog_errno(status);
1069 out_copy:
1070         /*
1071          * movement/defragmentation may end up being partially completed,
1072          * that's the reason why we need to return userspace the finished
1073          * length and new_offset even if failure happens somewhere.
1074          */
1075         if (copy_to_user(argp, &range, sizeof(range)))
1076                 status = -EFAULT;
1077
1078 out_free:
1079         kfree(context);
1080 out_drop:
1081         mnt_drop_write_file(filp);
1082
1083         return status;
1084 }