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                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 239                 BUG_ON(!context->refcount_loc);
 240
 241                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 242                                                &ref_tree, NULL);
 243                 if (ret) {
 244                         mlog_errno(ret);
 245                         return ret;
 246                 }
 247
 248                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 249                                                         context->refcount_loc,
 250                                                         phys_blkno,
 251                                                         *len,
 252                                                         &credits,
 253                                                         &extra_blocks);
 254                 if (ret) {
 255                         mlog_errno(ret);
 256                         goto out;
 257                 }
 258         }
 259
 260         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, *len, 1,
 261                                                  &context->meta_ac,
 262                                                  &context->data_ac,
 263                                                  extra_blocks, &credits);
 264         if (ret) {
 265                 mlog_errno(ret);
 266                 goto out;
 267         }
 268
 269         /*
 270          * should be using allocation reservation strategy there?
 271          *
 272          * if (context->data_ac)
 273          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 274          */
 275
 276         inode_lock(tl_inode);
 277
 278         if (ocfs2_truncate_log_needs_flush(osb)) {
 279                 ret = __ocfs2_flush_truncate_log(osb);
 280                 if (ret < 0) {
 281                         mlog_errno(ret);
 282                         goto out_unlock_mutex;
 283                 }
 284         }
 285
 286         handle = ocfs2_start_trans(osb, credits);
 287         if (IS_ERR(handle)) {
 288                 ret = PTR_ERR(handle);
 289                 mlog_errno(ret);
 290                 goto out_unlock_mutex;
 291         }
 292
 293         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 294                                      &new_phys_cpos, &new_len);
 295         if (ret) {
 296                 mlog_errno(ret);
 297                 goto out_commit;
 298         }
 299
 300         /*
 301          * allowing partial extent moving is kind of 'pros and cons', it makes
 302          * whole defragmentation less likely to fail, on the contrary, the bad
 303          * thing is it may make the fs even more fragmented after moving, let
 304          * userspace make a good decision here.
 305          */
 306         if (new_len != *len) {
 307                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 308                 if (!partial) {
 309                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 310                         ret = -ENOSPC;
 311                         goto out_commit;
 312                 }
 313         }
 314
 315         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 316              phys_cpos, new_phys_cpos);
 317
 318         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 319                                   new_phys_cpos, ext_flags);
 320         if (ret)
 321                 mlog_errno(ret);
 322
 323         if (partial && (new_len != *len))
 324                 *len = new_len;
 325
 326         /*
 327          * Here we should write the new page out first if we are
 328          * in write-back mode.
 329          */
 330         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 331         if (ret)
 332                 mlog_errno(ret);
 333
 334 out_commit:
 335         ocfs2_commit_trans(osb, handle);
 336
 337 out_unlock_mutex:
 338         inode_unlock(tl_inode);
 339
 340         if (context->data_ac) {
 341                 ocfs2_free_alloc_context(context->data_ac);
 342                 context->data_ac = NULL;
 343         }
 344
 345         if (context->meta_ac) {
 346                 ocfs2_free_alloc_context(context->meta_ac);
 347                 context->meta_ac = NULL;
 348         }
 349
 350 out:
 351         if (ref_tree)
 352                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 353
 354         return ret;
 355 }
 356
 357 /*
 358  * find the victim alloc group, where #blkno fits.
 359  */
 360 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 361                                          u64 vict_blkno,
 362                                          int type, int slot,
 363                                          int *vict_bit,
 364                                          struct buffer_head **ret_bh)
 365 {
 366         int ret, i, bits_per_unit = 0;
 367         u64 blkno;
 368         char namebuf[40];
 369
 370         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 371         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 372         struct ocfs2_chain_list *cl;
 373         struct ocfs2_chain_rec *rec;
 374         struct ocfs2_dinode *ac_dinode;
 375         struct ocfs2_group_desc *bg;
 376
 377         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 378         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 379                                          strlen(namebuf), &blkno);
 380         if (ret) {
 381                 ret = -ENOENT;
 382                 goto out;
 383         }
 384
 385         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 386         if (ret) {
 387                 mlog_errno(ret);
 388                 goto out;
 389         }
 390
 391         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 392         cl = &(ac_dinode->id2.i_chain);
 393         rec = &(cl->cl_recs[0]);
 394
 395         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 396                 bits_per_unit = osb->s_clustersize_bits -
 397                                         inode->i_sb->s_blocksize_bits;
 398         /*
 399          * 'vict_blkno' was out of the valid range.
 400          */
 401         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 402             (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 403                                 bits_per_unit))) {
 404                 ret = -EINVAL;
 405                 goto out;
 406         }
 407
 408         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 409
 410                 rec = &(cl->cl_recs[i]);
 411                 if (!rec)
 412                         continue;
 413
 414                 bg = NULL;
 415
 416                 do {
 417                         if (!bg)
 418                                 blkno = le64_to_cpu(rec->c_blkno);
 419                         else
 420                                 blkno = le64_to_cpu(bg->bg_next_group);
 421
 422                         if (gd_bh) {
 423                                 brelse(gd_bh);
 424                                 gd_bh = NULL;
 425                         }
 426
 427                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 428                         if (ret) {
 429                                 mlog_errno(ret);
 430                                 goto out;
 431                         }
 432
 433                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 434
 435                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 436                                                 le16_to_cpu(bg->bg_bits))) {
 437
 438                                 *ret_bh = gd_bh;
 439                                 *vict_bit = (vict_blkno - blkno) >>
 440                                                         bits_per_unit;
 441                                 mlog(0, "find the victim group: #%llu, "
 442                                      "total_bits: %u, vict_bit: %u\n",
 443                                      blkno, le16_to_cpu(bg->bg_bits),
 444                                      *vict_bit);
 445                                 goto out;
 446                         }
 447
 448                 } while (le64_to_cpu(bg->bg_next_group));
 449         }
 450
 451         ret = -EINVAL;
 452 out:
 453         brelse(ac_bh);
 454
 455         /*
 456          * caller has to release the gd_bh properly.
 457          */
 458         return ret;
 459 }
 460
 461 /*
 462  * XXX: helper to validate and adjust moving goal.
 463  */
 464 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 465                                                struct ocfs2_move_extents *range)
 466 {
 467         int ret, goal_bit = 0;
 468
 469         struct buffer_head *gd_bh = NULL;
 470         struct ocfs2_group_desc *bg;
 471         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 472         int c_to_b = 1 << (osb->s_clustersize_bits -
 473                                         inode->i_sb->s_blocksize_bits);
 474
 475         /*
 476          * make goal become cluster aligned.
 477          */
 478         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 479                                                       range->me_goal);
 480         /*
 481          * validate goal sits within global_bitmap, and return the victim
 482          * group desc
 483          */
 484         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 485                                             GLOBAL_BITMAP_SYSTEM_INODE,
 486                                             OCFS2_INVALID_SLOT,
 487                                             &goal_bit, &gd_bh);
 488         if (ret)
 489                 goto out;
 490
 491         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 492
 493         /*
 494          * moving goal is not allowd to start with a group desc blok(#0 blk)
 495          * let's compromise to the latter cluster.
 496          */
 497         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 498                 range->me_goal += c_to_b;
 499
 500         /*
 501          * movement is not gonna cross two groups.
 502          */
 503         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 504                                                                 range->me_len) {
 505                 ret = -EINVAL;
 506                 goto out;
 507         }
 508         /*
 509          * more exact validations/adjustments will be performed later during
 510          * moving operation for each extent range.
 511          */
 512         mlog(0, "extents get ready to be moved to #%llu block\n",
 513              range->me_goal);
 514
 515 out:
 516         brelse(gd_bh);
 517
 518         return ret;
 519 }
 520
 521 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 522                                     int *goal_bit, u32 move_len, u32 max_hop,
 523                                     u32 *phys_cpos)
 524 {
 525         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 526         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 527         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 528                                                  le64_to_cpu(gd->bg_blkno));
 529
 530         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 531
 532                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 533                 if (used) {
 534                         /*
 535                          * we even tried searching the free chunk by jumping
 536                          * a 'max_hop' distance, but still failed.
 537                          */
 538                         if ((i - base_bit) > max_hop) {
 539                                 *phys_cpos = 0;
 540                                 break;
 541                         }
 542
 543                         if (last_free_bits)
 544                                 last_free_bits = 0;
 545
 546                         continue;
 547                 } else
 548                         last_free_bits++;
 549
 550                 if (last_free_bits == move_len) {
 551                         *goal_bit = i;
 552                         *phys_cpos = base_cpos + i;
 553                         break;
 554                 }
 555         }
 556
 557         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 558 }
 559
 560 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 561                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 562                              u32 len, int ext_flags)
 563 {
 564         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 565         handle_t *handle;
 566         struct inode *inode = context->inode;
 567         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 568         struct inode *tl_inode = osb->osb_tl_inode;
 569         struct inode *gb_inode = NULL;
 570         struct buffer_head *gb_bh = NULL;
 571         struct buffer_head *gd_bh = NULL;
 572         struct ocfs2_group_desc *gd;
 573         struct ocfs2_refcount_tree *ref_tree = NULL;
 574         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 575                                                     context->range->me_threshold);
 576         u64 phys_blkno, new_phys_blkno;
 577
 578         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 579
 580         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 581                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 582                 BUG_ON(!context->refcount_loc);
 583
 584                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 585                                                &ref_tree, NULL);
 586                 if (ret) {
 587                         mlog_errno(ret);
 588                         return ret;
 589                 }
 590
 591                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 592                                                         context->refcount_loc,
 593                                                         phys_blkno,
 594                                                         len,
 595                                                         &credits,
 596                                                         &extra_blocks);
 597                 if (ret) {
 598                         mlog_errno(ret);
 599                         goto out;
 600                 }
 601         }
 602
 603         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, len, 1,
 604                                                  &context->meta_ac,
 605                                                  NULL, extra_blocks, &credits);
 606         if (ret) {
 607                 mlog_errno(ret);
 608                 goto out;
 609         }
 610
 611         /*
 612          * need to count 2 extra credits for global_bitmap inode and
 613          * group descriptor.
 614          */
 615         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 616
 617         /*
 618          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 619          * logic, while we still need to lock the global_bitmap.
 620          */
 621         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 622                                                OCFS2_INVALID_SLOT);
 623         if (!gb_inode) {
 624                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 625                 ret = -EIO;
 626                 goto out;
 627         }
 628
 629         inode_lock(gb_inode);
 630
 631         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 632         if (ret) {
 633                 mlog_errno(ret);
 634                 goto out_unlock_gb_mutex;
 635         }
 636
 637         inode_lock(tl_inode);
 638
 639         handle = ocfs2_start_trans(osb, credits);
 640         if (IS_ERR(handle)) {
 641                 ret = PTR_ERR(handle);
 642                 mlog_errno(ret);
 643                 goto out_unlock_tl_inode;
 644         }
 645
 646         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 647         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 648                                             GLOBAL_BITMAP_SYSTEM_INODE,
 649                                             OCFS2_INVALID_SLOT,
 650                                             &goal_bit, &gd_bh);
 651         if (ret) {
 652                 mlog_errno(ret);
 653                 goto out_commit;
 654         }
 655
 656         /*
 657          * probe the victim cluster group to find a proper
 658          * region to fit wanted movement, it even will perfrom
 659          * a best-effort attempt by compromising to a threshold
 660          * around the goal.
 661          */
 662         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 663                                 new_phys_cpos);
 664         if (!*new_phys_cpos) {
 665                 ret = -ENOSPC;
 666                 goto out_commit;
 667         }
 668
 669         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 670                                   *new_phys_cpos, ext_flags);
 671         if (ret) {
 672                 mlog_errno(ret);
 673                 goto out_commit;
 674         }
 675
 676         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 677         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 678                                                le16_to_cpu(gd->bg_chain));
 679         if (ret) {
 680                 mlog_errno(ret);
 681                 goto out_commit;
 682         }
 683
 684         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 685                                          goal_bit, len);
 686         if (ret) {
 687                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 688                                                le16_to_cpu(gd->bg_chain));
 689                 mlog_errno(ret);
 690         }
 691
 692         /*
 693          * Here we should write the new page out first if we are
 694          * in write-back mode.
 695          */
 696         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 697         if (ret)
 698                 mlog_errno(ret);
 699
 700 out_commit:
 701         ocfs2_commit_trans(osb, handle);
 702         brelse(gd_bh);
 703
 704 out_unlock_tl_inode:
 705         inode_unlock(tl_inode);
 706
 707         ocfs2_inode_unlock(gb_inode, 1);
 708 out_unlock_gb_mutex:
 709         inode_unlock(gb_inode);
 710         brelse(gb_bh);
 711         iput(gb_inode);
 712
 713 out:
 714         if (context->meta_ac) {
 715                 ocfs2_free_alloc_context(context->meta_ac);
 716                 context->meta_ac = NULL;
 717         }
 718
 719         if (ref_tree)
 720                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 721
 722         return ret;
 723 }
 724
 725 /*
 726  * Helper to calculate the defraging length in one run according to threshold.
 727  */
 728 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 729                                          u32 threshold, int *skip)
 730 {
 731         if ((*alloc_size + *len_defraged) < threshold) {
 732                 /*
 733                  * proceed defragmentation until we meet the thresh
 734                  */
 735                 *len_defraged += *alloc_size;
 736         } else if (*len_defraged == 0) {
 737                 /*
 738                  * XXX: skip a large extent.
 739                  */
 740                 *skip = 1;
 741         } else {
 742                 /*
 743                  * split this extent to coalesce with former pieces as
 744                  * to reach the threshold.
 745                  *
 746                  * we're done here with one cycle of defragmentation
 747                  * in a size of 'thresh', resetting 'len_defraged'
 748                  * forces a new defragmentation.
 749                  */
 750                 *alloc_size = threshold - *len_defraged;
 751                 *len_defraged = 0;
 752         }
 753 }
 754
 755 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 756                                 struct ocfs2_move_extents_context *context)
 757 {
 758         int ret = 0, flags, do_defrag, skip = 0;
 759         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 760         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 761
 762         struct inode *inode = context->inode;
 763         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 764         struct ocfs2_move_extents *range = context->range;
 765         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 766
 767         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 768                 return 0;
 769
 770         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 771                 return 0;
 772
 773         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 774
 775         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 776         ocfs2_init_dealloc_ctxt(&context->dealloc);
 777
 778         /*
 779          * TO-DO XXX:
 780          *
 781          * - xattr extents.
 782          */
 783
 784         do_defrag = context->auto_defrag;
 785
 786         /*
 787          * extents moving happens in unit of clusters, for the sake
 788          * of simplicity, we may ignore two clusters where 'byte_start'
 789          * and 'byte_start + len' were within.
 790          */
 791         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 792         len_to_move = (range->me_start + range->me_len) >>
 793                                                 osb->s_clustersize_bits;
 794         if (len_to_move >= move_start)
 795                 len_to_move -= move_start;
 796         else
 797                 len_to_move = 0;
 798
 799         if (do_defrag) {
 800                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 801                 if (defrag_thresh <= 1)
 802                         goto done;
 803         } else
 804                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 805                                                          range->me_goal);
 806
 807         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 808              "thresh: %u\n",
 809              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 810              (unsigned long long)range->me_start,
 811              (unsigned long long)range->me_len,
 812              move_start, len_to_move, defrag_thresh);
 813
 814         cpos = move_start;
 815         while (len_to_move) {
 816                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 817                                          &flags);
 818                 if (ret) {
 819                         mlog_errno(ret);
 820                         goto out;
 821                 }
 822
 823                 if (alloc_size > len_to_move)
 824                         alloc_size = len_to_move;
 825
 826                 /*
 827                  * XXX: how to deal with a hole:
 828                  *
 829                  * - skip the hole of course
 830                  * - force a new defragmentation
 831                  */
 832                 if (!phys_cpos) {
 833                         if (do_defrag)
 834                                 len_defraged = 0;
 835
 836                         goto next;
 837                 }
 838
 839                 if (do_defrag) {
 840                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 841                                                      defrag_thresh, &skip);
 842                         /*
 843                          * skip large extents
 844                          */
 845                         if (skip) {
 846                                 skip = 0;
 847                                 goto next;
 848                         }
 849
 850                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 851                              "alloc_size: %u, len_defraged: %u\n",
 852                              cpos, phys_cpos, alloc_size, len_defraged);
 853
 854                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 855                                                   &alloc_size, flags);
 856                 } else {
 857                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 858                                                 &new_phys_cpos, alloc_size,
 859                                                 flags);
 860
 861                         new_phys_cpos += alloc_size;
 862                 }
 863
 864                 if (ret < 0) {
 865                         mlog_errno(ret);
 866                         goto out;
 867                 }
 868
 869                 context->clusters_moved += alloc_size;
 870 next:
 871                 cpos += alloc_size;
 872                 len_to_move -= alloc_size;
 873         }
 874
 875 done:
 876         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 877
 878 out:
 879         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 880                                                       context->clusters_moved);
 881         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 882                                                        context->new_phys_cpos);
 883
 884         ocfs2_schedule_truncate_log_flush(osb, 1);
 885         ocfs2_run_deallocs(osb, &context->dealloc);
 886
 887         return ret;
 888 }
 889
 890 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 891 {
 892         int status;
 893         handle_t *handle;
 894         struct inode *inode = context->inode;
 895         struct ocfs2_dinode *di;
 896         struct buffer_head *di_bh = NULL;
 897         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 898
 899         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 900                 return -EROFS;
 901
 902         inode_lock(inode);
 903
 904         /*
 905          * This prevents concurrent writes from other nodes
 906          */
 907         status = ocfs2_rw_lock(inode, 1);
 908         if (status) {
 909                 mlog_errno(status);
 910                 goto out;
 911         }
 912
 913         status = ocfs2_inode_lock(inode, &di_bh, 1);
 914         if (status) {
 915                 mlog_errno(status);
 916                 goto out_rw_unlock;
 917         }
 918
 919         /*
 920          * rememer ip_xattr_sem also needs to be held if necessary
 921          */
 922         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 923
 924         status = __ocfs2_move_extents_range(di_bh, context);
 925
 926         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 927         if (status) {
 928                 mlog_errno(status);
 929                 goto out_inode_unlock;
 930         }
 931
 932         /*
 933          * We update ctime for these changes
 934          */
 935         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 936         if (IS_ERR(handle)) {
 937                 status = PTR_ERR(handle);
 938                 mlog_errno(status);
 939                 goto out_inode_unlock;
 940         }
 941
 942         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 943                                          OCFS2_JOURNAL_ACCESS_WRITE);
 944         if (status) {
 945                 mlog_errno(status);
 946                 goto out_commit;
 947         }
 948
 949         di = (struct ocfs2_dinode *)di_bh->b_data;
 950         inode->i_ctime = current_time(inode);
 951         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 952         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 953         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 954
 955         ocfs2_journal_dirty(handle, di_bh);
 956
 957 out_commit:
 958         ocfs2_commit_trans(osb, handle);
 959
 960 out_inode_unlock:
 961         brelse(di_bh);
 962         ocfs2_inode_unlock(inode, 1);
 963 out_rw_unlock:
 964         ocfs2_rw_unlock(inode, 1);
 965 out:
 966         inode_unlock(inode);
 967
 968         return status;
 969 }
 970
 971 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 972 {
 973         int status;
 974
 975         struct inode *inode = file_inode(filp);
 976         struct ocfs2_move_extents range;
 977         struct ocfs2_move_extents_context *context;
 978
 979         if (!argp)
 980                 return -EINVAL;
 981
 982         status = mnt_want_write_file(filp);
 983         if (status)
 984                 return status;
 985
 986         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
 987                 status = -EPERM;
 988                 goto out_drop;
 989         }
 990
 991         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
 992                 status = -EPERM;
 993                 goto out_drop;
 994         }
 995
 996         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
 997         if (!context) {
 998                 status = -ENOMEM;
 999                 mlog_errno(status);
1000                 goto out_drop;
1001         }
1002
1003         context->inode = inode;
1004         context->file = filp;
1005
1006         if (copy_from_user(&range, argp, sizeof(range))) {
1007                 status = -EFAULT;
1008                 goto out_free;
1009         }
1010
1011         if (range.me_start > i_size_read(inode)) {
1012                 status = -EINVAL;
1013                 goto out_free;
1014         }
1015
1016         if (range.me_start + range.me_len > i_size_read(inode))
1017                         range.me_len = i_size_read(inode) - range.me_start;
1018
1019         context->range = &range;
1020
1021         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1022                 context->auto_defrag = 1;
1023                 /*
1024                  * ok, the default theshold for the defragmentation
1025                  * is 1M, since our maximum clustersize was 1M also.
1026                  * any thought?
1027                  */
1028                 if (!range.me_threshold)
1029                         range.me_threshold = 1024 * 1024;
1030
1031                 if (range.me_threshold > i_size_read(inode))
1032                         range.me_threshold = i_size_read(inode);
1033
1034                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1035                         context->partial = 1;
1036         } else {
1037                 /*
1038                  * first best-effort attempt to validate and adjust the goal
1039                  * (physical address in block), while it can't guarantee later
1040                  * operation can succeed all the time since global_bitmap may
1041                  * change a bit over time.
1042                  */
1043
1044                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1045                 if (status)
1046                         goto out_copy;
1047         }
1048
1049         status = ocfs2_move_extents(context);
1050         if (status)
1051                 mlog_errno(status);
1052 out_copy:
1053         /*
1054          * movement/defragmentation may end up being partially completed,
1055          * that's the reason why we need to return userspace the finished
1056          * length and new_offset even if failure happens somewhere.
1057          */
1058         if (copy_to_user(argp, &range, sizeof(range)))
1059                 status = -EFAULT;
1060
1061 out_free:
1062         kfree(context);
1063 out_drop:
1064         mnt_drop_write_file(filp);
1065
1066         return status;
1067 }