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 allocator, and reserve appropriate number of bits for
 161  * meta blocks.
 162  */
 163 static int ocfs2_lock_meta_allocator_move_extents(struct inode *inode,
 164                                         struct ocfs2_extent_tree *et,
 165                                         u32 clusters_to_move,
 166                                         u32 extents_to_split,
 167                                         struct ocfs2_alloc_context **meta_ac,
 168                                         int extra_blocks,
 169                                         int *credits)
 170 {
 171         int ret, num_free_extents;
 172         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 173         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 174
 175         num_free_extents = ocfs2_num_free_extents(et);
 176         if (num_free_extents < 0) {
 177                 ret = num_free_extents;
 178                 mlog_errno(ret);
 179                 goto out;
 180         }
 181
 182         if (!num_free_extents ||
 183             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 184                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 185
 186         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 187         if (ret) {
 188                 mlog_errno(ret);
 189                 goto out;
 190         }
 191
 192
 193         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 194
 195         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 196              extra_blocks, clusters_to_move, *credits);
 197 out:
 198         if (ret) {
 199                 if (*meta_ac) {
 200                         ocfs2_free_alloc_context(*meta_ac);
 201                         *meta_ac = NULL;
 202                 }
 203         }
 204
 205         return ret;
 206 }
 207
 208 /*
 209  * Using one journal handle to guarantee the data consistency in case
 210  * crash happens anywhere.
 211  *
 212  *  XXX: defrag can end up with finishing partial extent as requested,
 213  * due to not enough contiguous clusters can be found in allocator.
 214  */
 215 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 216                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 217 {
 218         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 219         handle_t *handle;
 220         struct inode *inode = context->inode;
 221         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 222         struct inode *tl_inode = osb->osb_tl_inode;
 223         struct ocfs2_refcount_tree *ref_tree = NULL;
 224         u32 new_phys_cpos, new_len;
 225         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 226         int need_free = 0;
 227
 228         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 229                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 230                 BUG_ON(!context->refcount_loc);
 231
 232                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 233                                                &ref_tree, NULL);
 234                 if (ret) {
 235                         mlog_errno(ret);
 236                         return ret;
 237                 }
 238
 239                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 240                                                         context->refcount_loc,
 241                                                         phys_blkno,
 242                                                         *len,
 243                                                         &credits,
 244                                                         &extra_blocks);
 245                 if (ret) {
 246                         mlog_errno(ret);
 247                         goto out;
 248                 }
 249         }
 250
 251         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 252                                                 *len, 1,
 253                                                 &context->meta_ac,
 254                                                 extra_blocks, &credits);
 255         if (ret) {
 256                 mlog_errno(ret);
 257                 goto out;
 258         }
 259
 260         /*
 261          * should be using allocation reservation strategy there?
 262          *
 263          * if (context->data_ac)
 264          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 265          */
 266
 267         inode_lock(tl_inode);
 268
 269         if (ocfs2_truncate_log_needs_flush(osb)) {
 270                 ret = __ocfs2_flush_truncate_log(osb);
 271                 if (ret < 0) {
 272                         mlog_errno(ret);
 273                         goto out_unlock_mutex;
 274                 }
 275         }
 276
 277         /*
 278          * Make sure ocfs2_reserve_cluster is called after
 279          * __ocfs2_flush_truncate_log, otherwise, dead lock may happen.
 280          *
 281          * If ocfs2_reserve_cluster is called
 282          * before __ocfs2_flush_truncate_log, dead lock on global bitmap
 283          * may happen.
 284          *
 285          */
 286         ret = ocfs2_reserve_clusters(osb, *len, &context->data_ac);
 287         if (ret) {
 288                 mlog_errno(ret);
 289                 goto out_unlock_mutex;
 290         }
 291
 292         handle = ocfs2_start_trans(osb, credits);
 293         if (IS_ERR(handle)) {
 294                 ret = PTR_ERR(handle);
 295                 mlog_errno(ret);
 296                 goto out_unlock_mutex;
 297         }
 298
 299         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 300                                      &new_phys_cpos, &new_len);
 301         if (ret) {
 302                 mlog_errno(ret);
 303                 goto out_commit;
 304         }
 305
 306         /*
 307          * allowing partial extent moving is kind of 'pros and cons', it makes
 308          * whole defragmentation less likely to fail, on the contrary, the bad
 309          * thing is it may make the fs even more fragmented after moving, let
 310          * userspace make a good decision here.
 311          */
 312         if (new_len != *len) {
 313                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 314                 if (!partial) {
 315                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 316                         ret = -ENOSPC;
 317                         need_free = 1;
 318                         goto out_commit;
 319                 }
 320         }
 321
 322         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 323              phys_cpos, new_phys_cpos);
 324
 325         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 326                                   new_phys_cpos, ext_flags);
 327         if (ret)
 328                 mlog_errno(ret);
 329
 330         if (partial && (new_len != *len))
 331                 *len = new_len;
 332
 333         /*
 334          * Here we should write the new page out first if we are
 335          * in write-back mode.
 336          */
 337         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 338         if (ret)
 339                 mlog_errno(ret);
 340
 341 out_commit:
 342         if (need_free && context->data_ac) {
 343                 struct ocfs2_alloc_context *data_ac = context->data_ac;
 344
 345                 if (context->data_ac->ac_which == OCFS2_AC_USE_LOCAL)
 346                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
 347                                         new_phys_cpos, new_len);
 348                 else
 349                         ocfs2_free_clusters(handle,
 350                                         data_ac->ac_inode,
 351                                         data_ac->ac_bh,
 352                                         ocfs2_clusters_to_blocks(osb->sb, new_phys_cpos),
 353                                         new_len);
 354         }
 355
 356         ocfs2_commit_trans(osb, handle);
 357
 358 out_unlock_mutex:
 359         inode_unlock(tl_inode);
 360
 361         if (context->data_ac) {
 362                 ocfs2_free_alloc_context(context->data_ac);
 363                 context->data_ac = NULL;
 364         }
 365
 366         if (context->meta_ac) {
 367                 ocfs2_free_alloc_context(context->meta_ac);
 368                 context->meta_ac = NULL;
 369         }
 370
 371 out:
 372         if (ref_tree)
 373                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 374
 375         return ret;
 376 }
 377
 378 /*
 379  * find the victim alloc group, where #blkno fits.
 380  */
 381 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 382                                          u64 vict_blkno,
 383                                          int type, int slot,
 384                                          int *vict_bit,
 385                                          struct buffer_head **ret_bh)
 386 {
 387         int ret, i, bits_per_unit = 0;
 388         u64 blkno;
 389         char namebuf[40];
 390
 391         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 392         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 393         struct ocfs2_chain_list *cl;
 394         struct ocfs2_chain_rec *rec;
 395         struct ocfs2_dinode *ac_dinode;
 396         struct ocfs2_group_desc *bg;
 397
 398         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 399         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 400                                          strlen(namebuf), &blkno);
 401         if (ret) {
 402                 ret = -ENOENT;
 403                 goto out;
 404         }
 405
 406         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 407         if (ret) {
 408                 mlog_errno(ret);
 409                 goto out;
 410         }
 411
 412         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 413         cl = &(ac_dinode->id2.i_chain);
 414         rec = &(cl->cl_recs[0]);
 415
 416         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 417                 bits_per_unit = osb->s_clustersize_bits -
 418                                         inode->i_sb->s_blocksize_bits;
 419         /*
 420          * 'vict_blkno' was out of the valid range.
 421          */
 422         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 423             (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 424                                 bits_per_unit))) {
 425                 ret = -EINVAL;
 426                 goto out;
 427         }
 428
 429         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 430
 431                 rec = &(cl->cl_recs[i]);
 432                 if (!rec)
 433                         continue;
 434
 435                 bg = NULL;
 436
 437                 do {
 438                         if (!bg)
 439                                 blkno = le64_to_cpu(rec->c_blkno);
 440                         else
 441                                 blkno = le64_to_cpu(bg->bg_next_group);
 442
 443                         if (gd_bh) {
 444                                 brelse(gd_bh);
 445                                 gd_bh = NULL;
 446                         }
 447
 448                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 449                         if (ret) {
 450                                 mlog_errno(ret);
 451                                 goto out;
 452                         }
 453
 454                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 455
 456                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 457                                                 le16_to_cpu(bg->bg_bits))) {
 458
 459                                 *ret_bh = gd_bh;
 460                                 *vict_bit = (vict_blkno - blkno) >>
 461                                                         bits_per_unit;
 462                                 mlog(0, "find the victim group: #%llu, "
 463                                      "total_bits: %u, vict_bit: %u\n",
 464                                      blkno, le16_to_cpu(bg->bg_bits),
 465                                      *vict_bit);
 466                                 goto out;
 467                         }
 468
 469                 } while (le64_to_cpu(bg->bg_next_group));
 470         }
 471
 472         ret = -EINVAL;
 473 out:
 474         brelse(ac_bh);
 475
 476         /*
 477          * caller has to release the gd_bh properly.
 478          */
 479         return ret;
 480 }
 481
 482 /*
 483  * XXX: helper to validate and adjust moving goal.
 484  */
 485 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 486                                                struct ocfs2_move_extents *range)
 487 {
 488         int ret, goal_bit = 0;
 489
 490         struct buffer_head *gd_bh = NULL;
 491         struct ocfs2_group_desc *bg;
 492         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 493         int c_to_b = 1 << (osb->s_clustersize_bits -
 494                                         inode->i_sb->s_blocksize_bits);
 495
 496         /*
 497          * make goal become cluster aligned.
 498          */
 499         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 500                                                       range->me_goal);
 501         /*
 502          * validate goal sits within global_bitmap, and return the victim
 503          * group desc
 504          */
 505         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 506                                             GLOBAL_BITMAP_SYSTEM_INODE,
 507                                             OCFS2_INVALID_SLOT,
 508                                             &goal_bit, &gd_bh);
 509         if (ret)
 510                 goto out;
 511
 512         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 513
 514         /*
 515          * moving goal is not allowd to start with a group desc blok(#0 blk)
 516          * let's compromise to the latter cluster.
 517          */
 518         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 519                 range->me_goal += c_to_b;
 520
 521         /*
 522          * movement is not gonna cross two groups.
 523          */
 524         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 525                                                                 range->me_len) {
 526                 ret = -EINVAL;
 527                 goto out;
 528         }
 529         /*
 530          * more exact validations/adjustments will be performed later during
 531          * moving operation for each extent range.
 532          */
 533         mlog(0, "extents get ready to be moved to #%llu block\n",
 534              range->me_goal);
 535
 536 out:
 537         brelse(gd_bh);
 538
 539         return ret;
 540 }
 541
 542 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 543                                     int *goal_bit, u32 move_len, u32 max_hop,
 544                                     u32 *phys_cpos)
 545 {
 546         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 547         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 548         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 549                                                  le64_to_cpu(gd->bg_blkno));
 550
 551         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 552
 553                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 554                 if (used) {
 555                         /*
 556                          * we even tried searching the free chunk by jumping
 557                          * a 'max_hop' distance, but still failed.
 558                          */
 559                         if ((i - base_bit) > max_hop) {
 560                                 *phys_cpos = 0;
 561                                 break;
 562                         }
 563
 564                         if (last_free_bits)
 565                                 last_free_bits = 0;
 566
 567                         continue;
 568                 } else
 569                         last_free_bits++;
 570
 571                 if (last_free_bits == move_len) {
 572                         *goal_bit = i;
 573                         *phys_cpos = base_cpos + i;
 574                         break;
 575                 }
 576         }
 577
 578         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 579 }
 580
 581 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 582                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 583                              u32 len, int ext_flags)
 584 {
 585         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 586         handle_t *handle;
 587         struct inode *inode = context->inode;
 588         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 589         struct inode *tl_inode = osb->osb_tl_inode;
 590         struct inode *gb_inode = NULL;
 591         struct buffer_head *gb_bh = NULL;
 592         struct buffer_head *gd_bh = NULL;
 593         struct ocfs2_group_desc *gd;
 594         struct ocfs2_refcount_tree *ref_tree = NULL;
 595         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 596                                                     context->range->me_threshold);
 597         u64 phys_blkno, new_phys_blkno;
 598
 599         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 600
 601         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 602                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 603                 BUG_ON(!context->refcount_loc);
 604
 605                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 606                                                &ref_tree, NULL);
 607                 if (ret) {
 608                         mlog_errno(ret);
 609                         return ret;
 610                 }
 611
 612                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 613                                                         context->refcount_loc,
 614                                                         phys_blkno,
 615                                                         len,
 616                                                         &credits,
 617                                                         &extra_blocks);
 618                 if (ret) {
 619                         mlog_errno(ret);
 620                         goto out;
 621                 }
 622         }
 623
 624         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 625                                                 len, 1,
 626                                                 &context->meta_ac,
 627                                                 extra_blocks, &credits);
 628         if (ret) {
 629                 mlog_errno(ret);
 630                 goto out;
 631         }
 632
 633         /*
 634          * need to count 2 extra credits for global_bitmap inode and
 635          * group descriptor.
 636          */
 637         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 638
 639         /*
 640          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 641          * logic, while we still need to lock the global_bitmap.
 642          */
 643         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 644                                                OCFS2_INVALID_SLOT);
 645         if (!gb_inode) {
 646                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 647                 ret = -EIO;
 648                 goto out;
 649         }
 650
 651         inode_lock(gb_inode);
 652
 653         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 654         if (ret) {
 655                 mlog_errno(ret);
 656                 goto out_unlock_gb_mutex;
 657         }
 658
 659         inode_lock(tl_inode);
 660
 661         handle = ocfs2_start_trans(osb, credits);
 662         if (IS_ERR(handle)) {
 663                 ret = PTR_ERR(handle);
 664                 mlog_errno(ret);
 665                 goto out_unlock_tl_inode;
 666         }
 667
 668         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 669         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 670                                             GLOBAL_BITMAP_SYSTEM_INODE,
 671                                             OCFS2_INVALID_SLOT,
 672                                             &goal_bit, &gd_bh);
 673         if (ret) {
 674                 mlog_errno(ret);
 675                 goto out_commit;
 676         }
 677
 678         /*
 679          * probe the victim cluster group to find a proper
 680          * region to fit wanted movement, it even will perfrom
 681          * a best-effort attempt by compromising to a threshold
 682          * around the goal.
 683          */
 684         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 685                                 new_phys_cpos);
 686         if (!*new_phys_cpos) {
 687                 ret = -ENOSPC;
 688                 goto out_commit;
 689         }
 690
 691         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 692                                   *new_phys_cpos, ext_flags);
 693         if (ret) {
 694                 mlog_errno(ret);
 695                 goto out_commit;
 696         }
 697
 698         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 699         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 700                                                le16_to_cpu(gd->bg_chain));
 701         if (ret) {
 702                 mlog_errno(ret);
 703                 goto out_commit;
 704         }
 705
 706         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 707                                          goal_bit, len);
 708         if (ret) {
 709                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 710                                                le16_to_cpu(gd->bg_chain));
 711                 mlog_errno(ret);
 712         }
 713
 714         /*
 715          * Here we should write the new page out first if we are
 716          * in write-back mode.
 717          */
 718         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 719         if (ret)
 720                 mlog_errno(ret);
 721
 722 out_commit:
 723         ocfs2_commit_trans(osb, handle);
 724         brelse(gd_bh);
 725
 726 out_unlock_tl_inode:
 727         inode_unlock(tl_inode);
 728
 729         ocfs2_inode_unlock(gb_inode, 1);
 730 out_unlock_gb_mutex:
 731         inode_unlock(gb_inode);
 732         brelse(gb_bh);
 733         iput(gb_inode);
 734
 735 out:
 736         if (context->meta_ac) {
 737                 ocfs2_free_alloc_context(context->meta_ac);
 738                 context->meta_ac = NULL;
 739         }
 740
 741         if (ref_tree)
 742                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 743
 744         return ret;
 745 }
 746
 747 /*
 748  * Helper to calculate the defraging length in one run according to threshold.
 749  */
 750 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 751                                          u32 threshold, int *skip)
 752 {
 753         if ((*alloc_size + *len_defraged) < threshold) {
 754                 /*
 755                  * proceed defragmentation until we meet the thresh
 756                  */
 757                 *len_defraged += *alloc_size;
 758         } else if (*len_defraged == 0) {
 759                 /*
 760                  * XXX: skip a large extent.
 761                  */
 762                 *skip = 1;
 763         } else {
 764                 /*
 765                  * split this extent to coalesce with former pieces as
 766                  * to reach the threshold.
 767                  *
 768                  * we're done here with one cycle of defragmentation
 769                  * in a size of 'thresh', resetting 'len_defraged'
 770                  * forces a new defragmentation.
 771                  */
 772                 *alloc_size = threshold - *len_defraged;
 773                 *len_defraged = 0;
 774         }
 775 }
 776
 777 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 778                                 struct ocfs2_move_extents_context *context)
 779 {
 780         int ret = 0, flags, do_defrag, skip = 0;
 781         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 782         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 783
 784         struct inode *inode = context->inode;
 785         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 786         struct ocfs2_move_extents *range = context->range;
 787         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 788
 789         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 790                 return 0;
 791
 792         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 793                 return 0;
 794
 795         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 796
 797         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 798         ocfs2_init_dealloc_ctxt(&context->dealloc);
 799
 800         /*
 801          * TO-DO XXX:
 802          *
 803          * - xattr extents.
 804          */
 805
 806         do_defrag = context->auto_defrag;
 807
 808         /*
 809          * extents moving happens in unit of clusters, for the sake
 810          * of simplicity, we may ignore two clusters where 'byte_start'
 811          * and 'byte_start + len' were within.
 812          */
 813         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 814         len_to_move = (range->me_start + range->me_len) >>
 815                                                 osb->s_clustersize_bits;
 816         if (len_to_move >= move_start)
 817                 len_to_move -= move_start;
 818         else
 819                 len_to_move = 0;
 820
 821         if (do_defrag) {
 822                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 823                 if (defrag_thresh <= 1)
 824                         goto done;
 825         } else
 826                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 827                                                          range->me_goal);
 828
 829         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 830              "thresh: %u\n",
 831              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 832              (unsigned long long)range->me_start,
 833              (unsigned long long)range->me_len,
 834              move_start, len_to_move, defrag_thresh);
 835
 836         cpos = move_start;
 837         while (len_to_move) {
 838                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 839                                          &flags);
 840                 if (ret) {
 841                         mlog_errno(ret);
 842                         goto out;
 843                 }
 844
 845                 if (alloc_size > len_to_move)
 846                         alloc_size = len_to_move;
 847
 848                 /*
 849                  * XXX: how to deal with a hole:
 850                  *
 851                  * - skip the hole of course
 852                  * - force a new defragmentation
 853                  */
 854                 if (!phys_cpos) {
 855                         if (do_defrag)
 856                                 len_defraged = 0;
 857
 858                         goto next;
 859                 }
 860
 861                 if (do_defrag) {
 862                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 863                                                      defrag_thresh, &skip);
 864                         /*
 865                          * skip large extents
 866                          */
 867                         if (skip) {
 868                                 skip = 0;
 869                                 goto next;
 870                         }
 871
 872                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 873                              "alloc_size: %u, len_defraged: %u\n",
 874                              cpos, phys_cpos, alloc_size, len_defraged);
 875
 876                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 877                                                   &alloc_size, flags);
 878                 } else {
 879                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 880                                                 &new_phys_cpos, alloc_size,
 881                                                 flags);
 882
 883                         new_phys_cpos += alloc_size;
 884                 }
 885
 886                 if (ret < 0) {
 887                         mlog_errno(ret);
 888                         goto out;
 889                 }
 890
 891                 context->clusters_moved += alloc_size;
 892 next:
 893                 cpos += alloc_size;
 894                 len_to_move -= alloc_size;
 895         }
 896
 897 done:
 898         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 899
 900 out:
 901         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 902                                                       context->clusters_moved);
 903         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 904                                                        context->new_phys_cpos);
 905
 906         ocfs2_schedule_truncate_log_flush(osb, 1);
 907         ocfs2_run_deallocs(osb, &context->dealloc);
 908
 909         return ret;
 910 }
 911
 912 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 913 {
 914         int status;
 915         handle_t *handle;
 916         struct inode *inode = context->inode;
 917         struct ocfs2_dinode *di;
 918         struct buffer_head *di_bh = NULL;
 919         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 920
 921         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 922                 return -EROFS;
 923
 924         inode_lock(inode);
 925
 926         /*
 927          * This prevents concurrent writes from other nodes
 928          */
 929         status = ocfs2_rw_lock(inode, 1);
 930         if (status) {
 931                 mlog_errno(status);
 932                 goto out;
 933         }
 934
 935         status = ocfs2_inode_lock(inode, &di_bh, 1);
 936         if (status) {
 937                 mlog_errno(status);
 938                 goto out_rw_unlock;
 939         }
 940
 941         /*
 942          * rememer ip_xattr_sem also needs to be held if necessary
 943          */
 944         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 945
 946         status = __ocfs2_move_extents_range(di_bh, context);
 947
 948         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 949         if (status) {
 950                 mlog_errno(status);
 951                 goto out_inode_unlock;
 952         }
 953
 954         /*
 955          * We update ctime for these changes
 956          */
 957         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 958         if (IS_ERR(handle)) {
 959                 status = PTR_ERR(handle);
 960                 mlog_errno(status);
 961                 goto out_inode_unlock;
 962         }
 963
 964         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 965                                          OCFS2_JOURNAL_ACCESS_WRITE);
 966         if (status) {
 967                 mlog_errno(status);
 968                 goto out_commit;
 969         }
 970
 971         di = (struct ocfs2_dinode *)di_bh->b_data;
 972         inode->i_ctime = current_time(inode);
 973         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 974         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 975         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 976
 977         ocfs2_journal_dirty(handle, di_bh);
 978
 979 out_commit:
 980         ocfs2_commit_trans(osb, handle);
 981
 982 out_inode_unlock:
 983         brelse(di_bh);
 984         ocfs2_inode_unlock(inode, 1);
 985 out_rw_unlock:
 986         ocfs2_rw_unlock(inode, 1);
 987 out:
 988         inode_unlock(inode);
 989
 990         return status;
 991 }
 992
 993 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 994 {
 995         int status;
 996
 997         struct inode *inode = file_inode(filp);
 998         struct ocfs2_move_extents range;
 999         struct ocfs2_move_extents_context *context;
1000
1001         if (!argp)
1002                 return -EINVAL;
1003
1004         status = mnt_want_write_file(filp);
1005         if (status)
1006                 return status;
1007
1008         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
1009                 status = -EPERM;
1010                 goto out_drop;
1011         }
1012
1013         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1014                 status = -EPERM;
1015                 goto out_drop;
1016         }
1017
1018         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1019         if (!context) {
1020                 status = -ENOMEM;
1021                 mlog_errno(status);
1022                 goto out_drop;
1023         }
1024
1025         context->inode = inode;
1026         context->file = filp;
1027
1028         if (copy_from_user(&range, argp, sizeof(range))) {
1029                 status = -EFAULT;
1030                 goto out_free;
1031         }
1032
1033         if (range.me_start > i_size_read(inode)) {
1034                 status = -EINVAL;
1035                 goto out_free;
1036         }
1037
1038         if (range.me_start + range.me_len > i_size_read(inode))
1039                         range.me_len = i_size_read(inode) - range.me_start;
1040
1041         context->range = &range;
1042
1043         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1044                 context->auto_defrag = 1;
1045                 /*
1046                  * ok, the default theshold for the defragmentation
1047                  * is 1M, since our maximum clustersize was 1M also.
1048                  * any thought?
1049                  */
1050                 if (!range.me_threshold)
1051                         range.me_threshold = 1024 * 1024;
1052
1053                 if (range.me_threshold > i_size_read(inode))
1054                         range.me_threshold = i_size_read(inode);
1055
1056                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1057                         context->partial = 1;
1058         } else {
1059                 /*
1060                  * first best-effort attempt to validate and adjust the goal
1061                  * (physical address in block), while it can't guarantee later
1062                  * operation can succeed all the time since global_bitmap may
1063                  * change a bit over time.
1064                  */
1065
1066                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1067                 if (status)
1068                         goto out_copy;
1069         }
1070
1071         status = ocfs2_move_extents(context);
1072         if (status)
1073                 mlog_errno(status);
1074 out_copy:
1075         /*
1076          * movement/defragmentation may end up being partially completed,
1077          * that's the reason why we need to return userspace the finished
1078          * length and new_offset even if failure happens somewhere.
1079          */
1080         if (copy_to_user(argp, &range, sizeof(range)))
1081                 status = -EFAULT;
1082
1083 out_free:
1084         kfree(context);
1085 out_drop:
1086         mnt_drop_write_file(filp);
1087
1088         return status;
1089 }