fs/ocfs2/move_extents.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * move_extents.c
   5  *
   6  * Copyright (C) 2011 Oracle.  All rights reserved.
   7  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public
  10  * License version 2 as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License for more details.
  16  */
  17 #include <linux/fs.h>
  18 #include <linux/types.h>
  19 #include <linux/mount.h>
  20 #include <linux/swap.h>
  21
  22 #include <cluster/masklog.h>
  23
  24 #include "ocfs2.h"
  25 #include "ocfs2_ioctl.h"
  26
  27 #include "alloc.h"
  28 #include "aops.h"
  29 #include "dlmglue.h"
  30 #include "extent_map.h"
  31 #include "inode.h"
  32 #include "journal.h"
  33 #include "suballoc.h"
  34 #include "uptodate.h"
  35 #include "super.h"
  36 #include "dir.h"
  37 #include "buffer_head_io.h"
  38 #include "sysfile.h"
  39 #include "refcounttree.h"
  40 #include "move_extents.h"
  41
  42 struct ocfs2_move_extents_context {
  43         struct inode *inode;
  44         struct file *file;
  45         int auto_defrag;
  46         int partial;
  47         int credits;
  48         u32 new_phys_cpos;
  49         u32 clusters_moved;
  50         u64 refcount_loc;
  51         struct ocfs2_move_extents *range;
  52         struct ocfs2_extent_tree et;
  53         struct ocfs2_alloc_context *meta_ac;
  54         struct ocfs2_alloc_context *data_ac;
  55         struct ocfs2_cached_dealloc_ctxt dealloc;
  56 };
  57
  58 static int __ocfs2_move_extent(handle_t *handle,
  59                                struct ocfs2_move_extents_context *context,
  60                                u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
  61                                int ext_flags)
  62 {
  63         int ret = 0, index;
  64         struct inode *inode = context->inode;
  65         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  66         struct ocfs2_extent_rec *rec, replace_rec;
  67         struct ocfs2_path *path = NULL;
  68         struct ocfs2_extent_list *el;
  69         u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
  70         u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
  71
  72         ret = ocfs2_duplicate_clusters_by_page(handle, inode, cpos,
  73                                                p_cpos, new_p_cpos, len);
  74         if (ret) {
  75                 mlog_errno(ret);
  76                 goto out;
  77         }
  78
  79         memset(&replace_rec, 0, sizeof(replace_rec));
  80         replace_rec.e_cpos = cpu_to_le32(cpos);
  81         replace_rec.e_leaf_clusters = cpu_to_le16(len);
  82         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
  83                                                                    new_p_cpos));
  84
  85         path = ocfs2_new_path_from_et(&context->et);
  86         if (!path) {
  87                 ret = -ENOMEM;
  88                 mlog_errno(ret);
  89                 goto out;
  90         }
  91
  92         ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
  93         if (ret) {
  94                 mlog_errno(ret);
  95                 goto out;
  96         }
  97
  98         el = path_leaf_el(path);
  99
 100         index = ocfs2_search_extent_list(el, cpos);
 101         if (index == -1) {
 102                 ret = ocfs2_error(inode->i_sb,
 103                                   "Inode %llu has an extent at cpos %u which can no longer be found\n",
 104                                   (unsigned long long)ino, cpos);
 105                 goto out;
 106         }
 107
 108         rec = &el->l_recs[index];
 109
 110         BUG_ON(ext_flags != rec->e_flags);
 111         /*
 112          * after moving/defraging to new location, the extent is not going
 113          * to be refcounted anymore.
 114          */
 115         replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
 116
 117         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
 118                                       context->et.et_root_bh,
 119                                       OCFS2_JOURNAL_ACCESS_WRITE);
 120         if (ret) {
 121                 mlog_errno(ret);
 122                 goto out;
 123         }
 124
 125         ret = ocfs2_split_extent(handle, &context->et, path, index,
 126                                  &replace_rec, context->meta_ac,
 127                                  &context->dealloc);
 128         if (ret) {
 129                 mlog_errno(ret);
 130                 goto out;
 131         }
 132
 133         ocfs2_journal_dirty(handle, context->et.et_root_bh);
 134
 135         context->new_phys_cpos = new_p_cpos;
 136
 137         /*
 138          * need I to append truncate log for old clusters?
 139          */
 140         if (old_blkno) {
 141                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 142                         ret = ocfs2_decrease_refcount(inode, handle,
 143                                         ocfs2_blocks_to_clusters(osb->sb,
 144                                                                  old_blkno),
 145                                         len, context->meta_ac,
 146                                         &context->dealloc, 1);
 147                 else
 148                         ret = ocfs2_truncate_log_append(osb, handle,
 149                                                         old_blkno, len);
 150         }
 151
 152         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 153 out:
 154         ocfs2_free_path(path);
 155         return ret;
 156 }
 157
 158 /*
 159  * lock allocators, and reserving appropriate number of bits for
 160  * meta blocks and data clusters.
 161  *
 162  * in some cases, we don't need to reserve clusters, just let data_ac
 163  * be NULL.
 164  */
 165 static int ocfs2_lock_allocators_move_extents(struct inode *inode,
 166                                         struct ocfs2_extent_tree *et,
 167                                         u32 clusters_to_move,
 168                                         u32 extents_to_split,
 169                                         struct ocfs2_alloc_context **meta_ac,
 170                                         struct ocfs2_alloc_context **data_ac,
 171                                         int extra_blocks,
 172                                         int *credits)
 173 {
 174         int ret, num_free_extents;
 175         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 176         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 177
 178         num_free_extents = ocfs2_num_free_extents(osb, et);
 179         if (num_free_extents < 0) {
 180                 ret = num_free_extents;
 181                 mlog_errno(ret);
 182                 goto out;
 183         }
 184
 185         if (!num_free_extents ||
 186             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 187                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 188
 189         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 190         if (ret) {
 191                 mlog_errno(ret);
 192                 goto out;
 193         }
 194
 195         if (data_ac) {
 196                 ret = ocfs2_reserve_clusters(osb, clusters_to_move, data_ac);
 197                 if (ret) {
 198                         mlog_errno(ret);
 199                         goto out;
 200                 }
 201         }
 202
 203         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 204
 205         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 206              extra_blocks, clusters_to_move, *credits);
 207 out:
 208         if (ret) {
 209                 if (*meta_ac) {
 210                         ocfs2_free_alloc_context(*meta_ac);
 211                         *meta_ac = NULL;
 212                 }
 213         }
 214
 215         return ret;
 216 }
 217
 218 /*
 219  * Using one journal handle to guarantee the data consistency in case
 220  * crash happens anywhere.
 221  *
 222  *  XXX: defrag can end up with finishing partial extent as requested,
 223  * due to not enough contiguous clusters can be found in allocator.
 224  */
 225 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 226                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 227 {
 228         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 229         handle_t *handle;
 230         struct inode *inode = context->inode;
 231         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 232         struct inode *tl_inode = osb->osb_tl_inode;
 233         struct ocfs2_refcount_tree *ref_tree = NULL;
 234         u32 new_phys_cpos, new_len;
 235         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 236
 237         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 238
 239                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 240                          OCFS2_HAS_REFCOUNT_FL));
 241
 242                 BUG_ON(!context->refcount_loc);
 243
 244                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 245                                                &ref_tree, NULL);
 246                 if (ret) {
 247                         mlog_errno(ret);
 248                         return ret;
 249                 }
 250
 251                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 252                                                         context->refcount_loc,
 253                                                         phys_blkno,
 254                                                         *len,
 255                                                         &credits,
 256                                                         &extra_blocks);
 257                 if (ret) {
 258                         mlog_errno(ret);
 259                         goto out;
 260                 }
 261         }
 262
 263         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, *len, 1,
 264                                                  &context->meta_ac,
 265                                                  &context->data_ac,
 266                                                  extra_blocks, &credits);
 267         if (ret) {
 268                 mlog_errno(ret);
 269                 goto out;
 270         }
 271
 272         /*
 273          * should be using allocation reservation strategy there?
 274          *
 275          * if (context->data_ac)
 276          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 277          */
 278
 279         mutex_lock(&tl_inode->i_mutex);
 280
 281         if (ocfs2_truncate_log_needs_flush(osb)) {
 282                 ret = __ocfs2_flush_truncate_log(osb);
 283                 if (ret < 0) {
 284                         mlog_errno(ret);
 285                         goto out_unlock_mutex;
 286                 }
 287         }
 288
 289         handle = ocfs2_start_trans(osb, credits);
 290         if (IS_ERR(handle)) {
 291                 ret = PTR_ERR(handle);
 292                 mlog_errno(ret);
 293                 goto out_unlock_mutex;
 294         }
 295
 296         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 297                                      &new_phys_cpos, &new_len);
 298         if (ret) {
 299                 mlog_errno(ret);
 300                 goto out_commit;
 301         }
 302
 303         /*
 304          * allowing partial extent moving is kind of 'pros and cons', it makes
 305          * whole defragmentation less likely to fail, on the contrary, the bad
 306          * thing is it may make the fs even more fragmented after moving, let
 307          * userspace make a good decision here.
 308          */
 309         if (new_len != *len) {
 310                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 311                 if (!partial) {
 312                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 313                         ret = -ENOSPC;
 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         ocfs2_commit_trans(osb, handle);
 339
 340 out_unlock_mutex:
 341         mutex_unlock(&tl_inode->i_mutex);
 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))) {
 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                         *goal_bit = i;
 555                         *phys_cpos = base_cpos + i;
 556                         break;
 557                 }
 558         }
 559
 560         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 561 }
 562
 563 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 564                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 565                              u32 len, int ext_flags)
 566 {
 567         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 568         handle_t *handle;
 569         struct inode *inode = context->inode;
 570         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 571         struct inode *tl_inode = osb->osb_tl_inode;
 572         struct inode *gb_inode = NULL;
 573         struct buffer_head *gb_bh = NULL;
 574         struct buffer_head *gd_bh = NULL;
 575         struct ocfs2_group_desc *gd;
 576         struct ocfs2_refcount_tree *ref_tree = NULL;
 577         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 578                                                     context->range->me_threshold);
 579         u64 phys_blkno, new_phys_blkno;
 580
 581         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 582
 583         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 584
 585                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 586                          OCFS2_HAS_REFCOUNT_FL));
 587
 588                 BUG_ON(!context->refcount_loc);
 589
 590                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 591                                                &ref_tree, NULL);
 592                 if (ret) {
 593                         mlog_errno(ret);
 594                         return ret;
 595                 }
 596
 597                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 598                                                         context->refcount_loc,
 599                                                         phys_blkno,
 600                                                         len,
 601                                                         &credits,
 602                                                         &extra_blocks);
 603                 if (ret) {
 604                         mlog_errno(ret);
 605                         goto out;
 606                 }
 607         }
 608
 609         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, len, 1,
 610                                                  &context->meta_ac,
 611                                                  NULL, extra_blocks, &credits);
 612         if (ret) {
 613                 mlog_errno(ret);
 614                 goto out;
 615         }
 616
 617         /*
 618          * need to count 2 extra credits for global_bitmap inode and
 619          * group descriptor.
 620          */
 621         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 622
 623         /*
 624          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 625          * logic, while we still need to lock the global_bitmap.
 626          */
 627         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 628                                                OCFS2_INVALID_SLOT);
 629         if (!gb_inode) {
 630                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 631                 ret = -EIO;
 632                 goto out;
 633         }
 634
 635         mutex_lock(&gb_inode->i_mutex);
 636
 637         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 638         if (ret) {
 639                 mlog_errno(ret);
 640                 goto out_unlock_gb_mutex;
 641         }
 642
 643         mutex_lock(&tl_inode->i_mutex);
 644
 645         handle = ocfs2_start_trans(osb, credits);
 646         if (IS_ERR(handle)) {
 647                 ret = PTR_ERR(handle);
 648                 mlog_errno(ret);
 649                 goto out_unlock_tl_inode;
 650         }
 651
 652         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 653         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 654                                             GLOBAL_BITMAP_SYSTEM_INODE,
 655                                             OCFS2_INVALID_SLOT,
 656                                             &goal_bit, &gd_bh);
 657         if (ret) {
 658                 mlog_errno(ret);
 659                 goto out_commit;
 660         }
 661
 662         /*
 663          * probe the victim cluster group to find a proper
 664          * region to fit wanted movement, it even will perfrom
 665          * a best-effort attempt by compromising to a threshold
 666          * around the goal.
 667          */
 668         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 669                                 new_phys_cpos);
 670         if (!*new_phys_cpos) {
 671                 ret = -ENOSPC;
 672                 goto out_commit;
 673         }
 674
 675         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 676                                   *new_phys_cpos, ext_flags);
 677         if (ret) {
 678                 mlog_errno(ret);
 679                 goto out_commit;
 680         }
 681
 682         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 683         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 684                                                le16_to_cpu(gd->bg_chain));
 685         if (ret) {
 686                 mlog_errno(ret);
 687                 goto out_commit;
 688         }
 689
 690         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 691                                          goal_bit, len);
 692         if (ret) {
 693                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 694                                                le16_to_cpu(gd->bg_chain));
 695                 mlog_errno(ret);
 696         }
 697
 698         /*
 699          * Here we should write the new page out first if we are
 700          * in write-back mode.
 701          */
 702         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 703         if (ret)
 704                 mlog_errno(ret);
 705
 706 out_commit:
 707         ocfs2_commit_trans(osb, handle);
 708         brelse(gd_bh);
 709
 710 out_unlock_tl_inode:
 711         mutex_unlock(&tl_inode->i_mutex);
 712
 713         ocfs2_inode_unlock(gb_inode, 1);
 714 out_unlock_gb_mutex:
 715         mutex_unlock(&gb_inode->i_mutex);
 716         brelse(gb_bh);
 717         iput(gb_inode);
 718
 719 out:
 720         if (context->meta_ac) {
 721                 ocfs2_free_alloc_context(context->meta_ac);
 722                 context->meta_ac = NULL;
 723         }
 724
 725         if (ref_tree)
 726                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 727
 728         return ret;
 729 }
 730
 731 /*
 732  * Helper to calculate the defraging length in one run according to threshold.
 733  */
 734 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 735                                          u32 threshold, int *skip)
 736 {
 737         if ((*alloc_size + *len_defraged) < threshold) {
 738                 /*
 739                  * proceed defragmentation until we meet the thresh
 740                  */
 741                 *len_defraged += *alloc_size;
 742         } else if (*len_defraged == 0) {
 743                 /*
 744                  * XXX: skip a large extent.
 745                  */
 746                 *skip = 1;
 747         } else {
 748                 /*
 749                  * split this extent to coalesce with former pieces as
 750                  * to reach the threshold.
 751                  *
 752                  * we're done here with one cycle of defragmentation
 753                  * in a size of 'thresh', resetting 'len_defraged'
 754                  * forces a new defragmentation.
 755                  */
 756                 *alloc_size = threshold - *len_defraged;
 757                 *len_defraged = 0;
 758         }
 759 }
 760
 761 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 762                                 struct ocfs2_move_extents_context *context)
 763 {
 764         int ret = 0, flags, do_defrag, skip = 0;
 765         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 766         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 767
 768         struct inode *inode = context->inode;
 769         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 770         struct ocfs2_move_extents *range = context->range;
 771         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 772
 773         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 774                 return 0;
 775
 776         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 777                 return 0;
 778
 779         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 780
 781         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 782         ocfs2_init_dealloc_ctxt(&context->dealloc);
 783
 784         /*
 785          * TO-DO XXX:
 786          *
 787          * - xattr extents.
 788          */
 789
 790         do_defrag = context->auto_defrag;
 791
 792         /*
 793          * extents moving happens in unit of clusters, for the sake
 794          * of simplicity, we may ignore two clusters where 'byte_start'
 795          * and 'byte_start + len' were within.
 796          */
 797         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 798         len_to_move = (range->me_start + range->me_len) >>
 799                                                 osb->s_clustersize_bits;
 800         if (len_to_move >= move_start)
 801                 len_to_move -= move_start;
 802         else
 803                 len_to_move = 0;
 804
 805         if (do_defrag) {
 806                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 807                 if (defrag_thresh <= 1)
 808                         goto done;
 809         } else
 810                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 811                                                          range->me_goal);
 812
 813         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 814              "thresh: %u\n",
 815              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 816              (unsigned long long)range->me_start,
 817              (unsigned long long)range->me_len,
 818              move_start, len_to_move, defrag_thresh);
 819
 820         cpos = move_start;
 821         while (len_to_move) {
 822                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 823                                          &flags);
 824                 if (ret) {
 825                         mlog_errno(ret);
 826                         goto out;
 827                 }
 828
 829                 if (alloc_size > len_to_move)
 830                         alloc_size = len_to_move;
 831
 832                 /*
 833                  * XXX: how to deal with a hole:
 834                  *
 835                  * - skip the hole of course
 836                  * - force a new defragmentation
 837                  */
 838                 if (!phys_cpos) {
 839                         if (do_defrag)
 840                                 len_defraged = 0;
 841
 842                         goto next;
 843                 }
 844
 845                 if (do_defrag) {
 846                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 847                                                      defrag_thresh, &skip);
 848                         /*
 849                          * skip large extents
 850                          */
 851                         if (skip) {
 852                                 skip = 0;
 853                                 goto next;
 854                         }
 855
 856                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 857                              "alloc_size: %u, len_defraged: %u\n",
 858                              cpos, phys_cpos, alloc_size, len_defraged);
 859
 860                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 861                                                   &alloc_size, flags);
 862                 } else {
 863                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 864                                                 &new_phys_cpos, alloc_size,
 865                                                 flags);
 866
 867                         new_phys_cpos += alloc_size;
 868                 }
 869
 870                 if (ret < 0) {
 871                         mlog_errno(ret);
 872                         goto out;
 873                 }
 874
 875                 context->clusters_moved += alloc_size;
 876 next:
 877                 cpos += alloc_size;
 878                 len_to_move -= alloc_size;
 879         }
 880
 881 done:
 882         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 883
 884 out:
 885         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 886                                                       context->clusters_moved);
 887         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 888                                                        context->new_phys_cpos);
 889
 890         ocfs2_schedule_truncate_log_flush(osb, 1);
 891         ocfs2_run_deallocs(osb, &context->dealloc);
 892
 893         return ret;
 894 }
 895
 896 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 897 {
 898         int status;
 899         handle_t *handle;
 900         struct inode *inode = context->inode;
 901         struct ocfs2_dinode *di;
 902         struct buffer_head *di_bh = NULL;
 903         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 904
 905         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 906                 return -EROFS;
 907
 908         mutex_lock(&inode->i_mutex);
 909
 910         /*
 911          * This prevents concurrent writes from other nodes
 912          */
 913         status = ocfs2_rw_lock(inode, 1);
 914         if (status) {
 915                 mlog_errno(status);
 916                 goto out;
 917         }
 918
 919         status = ocfs2_inode_lock(inode, &di_bh, 1);
 920         if (status) {
 921                 mlog_errno(status);
 922                 goto out_rw_unlock;
 923         }
 924
 925         /*
 926          * rememer ip_xattr_sem also needs to be held if necessary
 927          */
 928         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 929
 930         status = __ocfs2_move_extents_range(di_bh, context);
 931
 932         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 933         if (status) {
 934                 mlog_errno(status);
 935                 goto out_inode_unlock;
 936         }
 937
 938         /*
 939          * We update ctime for these changes
 940          */
 941         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 942         if (IS_ERR(handle)) {
 943                 status = PTR_ERR(handle);
 944                 mlog_errno(status);
 945                 goto out_inode_unlock;
 946         }
 947
 948         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 949                                          OCFS2_JOURNAL_ACCESS_WRITE);
 950         if (status) {
 951                 mlog_errno(status);
 952                 goto out_commit;
 953         }
 954
 955         di = (struct ocfs2_dinode *)di_bh->b_data;
 956         inode->i_ctime = CURRENT_TIME;
 957         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 958         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 959         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 960
 961         ocfs2_journal_dirty(handle, di_bh);
 962
 963 out_commit:
 964         ocfs2_commit_trans(osb, handle);
 965
 966 out_inode_unlock:
 967         brelse(di_bh);
 968         ocfs2_inode_unlock(inode, 1);
 969 out_rw_unlock:
 970         ocfs2_rw_unlock(inode, 1);
 971 out:
 972         mutex_unlock(&inode->i_mutex);
 973
 974         return status;
 975 }
 976
 977 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 978 {
 979         int status;
 980
 981         struct inode *inode = file_inode(filp);
 982         struct ocfs2_move_extents range;
 983         struct ocfs2_move_extents_context *context;
 984
 985         if (!argp)
 986                 return -EINVAL;
 987
 988         status = mnt_want_write_file(filp);
 989         if (status)
 990                 return status;
 991
 992         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
 993                 status = -EPERM;
 994                 goto out_drop;
 995         }
 996
 997         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
 998                 status = -EPERM;
 999                 goto out_drop;
1000         }
1001
1002         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1003         if (!context) {
1004                 status = -ENOMEM;
1005                 mlog_errno(status);
1006                 goto out_drop;
1007         }
1008
1009         context->inode = inode;
1010         context->file = filp;
1011
1012         if (copy_from_user(&range, argp, sizeof(range))) {
1013                 status = -EFAULT;
1014                 goto out_free;
1015         }
1016
1017         if (range.me_start > i_size_read(inode)) {
1018                 status = -EINVAL;
1019                 goto out_free;
1020         }
1021
1022         if (range.me_start + range.me_len > i_size_read(inode))
1023                         range.me_len = i_size_read(inode) - range.me_start;
1024
1025         context->range = &range;
1026
1027         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1028                 context->auto_defrag = 1;
1029                 /*
1030                  * ok, the default theshold for the defragmentation
1031                  * is 1M, since our maximum clustersize was 1M also.
1032                  * any thought?
1033                  */
1034                 if (!range.me_threshold)
1035                         range.me_threshold = 1024 * 1024;
1036
1037                 if (range.me_threshold > i_size_read(inode))
1038                         range.me_threshold = i_size_read(inode);
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 }