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 allocator, and reserve appropriate number of bits for
 160  * meta blocks.
 161  */
 162 static int ocfs2_lock_meta_allocator_move_extents(struct inode *inode,
 163                                         struct ocfs2_extent_tree *et,
 164                                         u32 clusters_to_move,
 165                                         u32 extents_to_split,
 166                                         struct ocfs2_alloc_context **meta_ac,
 167                                         int extra_blocks,
 168                                         int *credits)
 169 {
 170         int ret, num_free_extents;
 171         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 172         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 173
 174         num_free_extents = ocfs2_num_free_extents(osb, et);
 175         if (num_free_extents < 0) {
 176                 ret = num_free_extents;
 177                 mlog_errno(ret);
 178                 goto out;
 179         }
 180
 181         if (!num_free_extents ||
 182             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 183                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 184
 185         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 186         if (ret) {
 187                 mlog_errno(ret);
 188                 goto out;
 189         }
 190
 191
 192         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 193
 194         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 195              extra_blocks, clusters_to_move, *credits);
 196 out:
 197         if (ret) {
 198                 if (*meta_ac) {
 199                         ocfs2_free_alloc_context(*meta_ac);
 200                         *meta_ac = NULL;
 201                 }
 202         }
 203
 204         return ret;
 205 }
 206
 207 /*
 208  * Using one journal handle to guarantee the data consistency in case
 209  * crash happens anywhere.
 210  *
 211  *  XXX: defrag can end up with finishing partial extent as requested,
 212  * due to not enough contiguous clusters can be found in allocator.
 213  */
 214 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 215                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 216 {
 217         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 218         handle_t *handle;
 219         struct inode *inode = context->inode;
 220         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 221         struct inode *tl_inode = osb->osb_tl_inode;
 222         struct ocfs2_refcount_tree *ref_tree = NULL;
 223         u32 new_phys_cpos, new_len;
 224         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 225
 226         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 227
 228                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 229                          OCFS2_HAS_REFCOUNT_FL));
 230
 231                 BUG_ON(!context->refcount_loc);
 232
 233                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 234                                                &ref_tree, NULL);
 235                 if (ret) {
 236                         mlog_errno(ret);
 237                         return ret;
 238                 }
 239
 240                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 241                                                         context->refcount_loc,
 242                                                         phys_blkno,
 243                                                         *len,
 244                                                         &credits,
 245                                                         &extra_blocks);
 246                 if (ret) {
 247                         mlog_errno(ret);
 248                         goto out;
 249                 }
 250         }
 251
 252         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 253                                                 *len, 1,
 254                                                 &context->meta_ac,
 255                                                 extra_blocks, &credits);
 256         if (ret) {
 257                 mlog_errno(ret);
 258                 goto out;
 259         }
 260
 261         /*
 262          * should be using allocation reservation strategy there?
 263          *
 264          * if (context->data_ac)
 265          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 266          */
 267
 268         inode_lock(tl_inode);
 269
 270         if (ocfs2_truncate_log_needs_flush(osb)) {
 271                 ret = __ocfs2_flush_truncate_log(osb);
 272                 if (ret < 0) {
 273                         mlog_errno(ret);
 274                         goto out_unlock_mutex;
 275                 }
 276         }
 277
 278         /*
 279          * Make sure ocfs2_reserve_cluster is called after
 280          * __ocfs2_flush_truncate_log, otherwise, dead lock may happen.
 281          *
 282          * If ocfs2_reserve_cluster is called
 283          * before __ocfs2_flush_truncate_log, dead lock on global bitmap
 284          * may happen.
 285          *
 286          */
 287         ret = ocfs2_reserve_clusters(osb, *len, &context->data_ac);
 288         if (ret) {
 289                 mlog_errno(ret);
 290                 goto out_unlock_mutex;
 291         }
 292
 293         handle = ocfs2_start_trans(osb, credits);
 294         if (IS_ERR(handle)) {
 295                 ret = PTR_ERR(handle);
 296                 mlog_errno(ret);
 297                 goto out_unlock_mutex;
 298         }
 299
 300         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 301                                      &new_phys_cpos, &new_len);
 302         if (ret) {
 303                 mlog_errno(ret);
 304                 goto out_commit;
 305         }
 306
 307         /*
 308          * allowing partial extent moving is kind of 'pros and cons', it makes
 309          * whole defragmentation less likely to fail, on the contrary, the bad
 310          * thing is it may make the fs even more fragmented after moving, let
 311          * userspace make a good decision here.
 312          */
 313         if (new_len != *len) {
 314                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 315                 if (!partial) {
 316                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 317                         ret = -ENOSPC;
 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         ocfs2_commit_trans(osb, handle);
 343
 344 out_unlock_mutex:
 345         inode_unlock(tl_inode);
 346
 347         if (context->data_ac) {
 348                 ocfs2_free_alloc_context(context->data_ac);
 349                 context->data_ac = NULL;
 350         }
 351
 352         if (context->meta_ac) {
 353                 ocfs2_free_alloc_context(context->meta_ac);
 354                 context->meta_ac = NULL;
 355         }
 356
 357 out:
 358         if (ref_tree)
 359                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 360
 361         return ret;
 362 }
 363
 364 /*
 365  * find the victim alloc group, where #blkno fits.
 366  */
 367 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 368                                          u64 vict_blkno,
 369                                          int type, int slot,
 370                                          int *vict_bit,
 371                                          struct buffer_head **ret_bh)
 372 {
 373         int ret, i, bits_per_unit = 0;
 374         u64 blkno;
 375         char namebuf[40];
 376
 377         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 378         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 379         struct ocfs2_chain_list *cl;
 380         struct ocfs2_chain_rec *rec;
 381         struct ocfs2_dinode *ac_dinode;
 382         struct ocfs2_group_desc *bg;
 383
 384         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 385         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 386                                          strlen(namebuf), &blkno);
 387         if (ret) {
 388                 ret = -ENOENT;
 389                 goto out;
 390         }
 391
 392         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 393         if (ret) {
 394                 mlog_errno(ret);
 395                 goto out;
 396         }
 397
 398         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 399         cl = &(ac_dinode->id2.i_chain);
 400         rec = &(cl->cl_recs[0]);
 401
 402         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 403                 bits_per_unit = osb->s_clustersize_bits -
 404                                         inode->i_sb->s_blocksize_bits;
 405         /*
 406          * 'vict_blkno' was out of the valid range.
 407          */
 408         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 409             (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 410                                 bits_per_unit))) {
 411                 ret = -EINVAL;
 412                 goto out;
 413         }
 414
 415         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 416
 417                 rec = &(cl->cl_recs[i]);
 418                 if (!rec)
 419                         continue;
 420
 421                 bg = NULL;
 422
 423                 do {
 424                         if (!bg)
 425                                 blkno = le64_to_cpu(rec->c_blkno);
 426                         else
 427                                 blkno = le64_to_cpu(bg->bg_next_group);
 428
 429                         if (gd_bh) {
 430                                 brelse(gd_bh);
 431                                 gd_bh = NULL;
 432                         }
 433
 434                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 435                         if (ret) {
 436                                 mlog_errno(ret);
 437                                 goto out;
 438                         }
 439
 440                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 441
 442                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 443                                                 le16_to_cpu(bg->bg_bits))) {
 444
 445                                 *ret_bh = gd_bh;
 446                                 *vict_bit = (vict_blkno - blkno) >>
 447                                                         bits_per_unit;
 448                                 mlog(0, "find the victim group: #%llu, "
 449                                      "total_bits: %u, vict_bit: %u\n",
 450                                      blkno, le16_to_cpu(bg->bg_bits),
 451                                      *vict_bit);
 452                                 goto out;
 453                         }
 454
 455                 } while (le64_to_cpu(bg->bg_next_group));
 456         }
 457
 458         ret = -EINVAL;
 459 out:
 460         brelse(ac_bh);
 461
 462         /*
 463          * caller has to release the gd_bh properly.
 464          */
 465         return ret;
 466 }
 467
 468 /*
 469  * XXX: helper to validate and adjust moving goal.
 470  */
 471 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 472                                                struct ocfs2_move_extents *range)
 473 {
 474         int ret, goal_bit = 0;
 475
 476         struct buffer_head *gd_bh = NULL;
 477         struct ocfs2_group_desc *bg;
 478         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 479         int c_to_b = 1 << (osb->s_clustersize_bits -
 480                                         inode->i_sb->s_blocksize_bits);
 481
 482         /*
 483          * make goal become cluster aligned.
 484          */
 485         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 486                                                       range->me_goal);
 487         /*
 488          * validate goal sits within global_bitmap, and return the victim
 489          * group desc
 490          */
 491         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 492                                             GLOBAL_BITMAP_SYSTEM_INODE,
 493                                             OCFS2_INVALID_SLOT,
 494                                             &goal_bit, &gd_bh);
 495         if (ret)
 496                 goto out;
 497
 498         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 499
 500         /*
 501          * moving goal is not allowd to start with a group desc blok(#0 blk)
 502          * let's compromise to the latter cluster.
 503          */
 504         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 505                 range->me_goal += c_to_b;
 506
 507         /*
 508          * movement is not gonna cross two groups.
 509          */
 510         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 511                                                                 range->me_len) {
 512                 ret = -EINVAL;
 513                 goto out;
 514         }
 515         /*
 516          * more exact validations/adjustments will be performed later during
 517          * moving operation for each extent range.
 518          */
 519         mlog(0, "extents get ready to be moved to #%llu block\n",
 520              range->me_goal);
 521
 522 out:
 523         brelse(gd_bh);
 524
 525         return ret;
 526 }
 527
 528 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 529                                     int *goal_bit, u32 move_len, u32 max_hop,
 530                                     u32 *phys_cpos)
 531 {
 532         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 533         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 534         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 535                                                  le64_to_cpu(gd->bg_blkno));
 536
 537         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 538
 539                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 540                 if (used) {
 541                         /*
 542                          * we even tried searching the free chunk by jumping
 543                          * a 'max_hop' distance, but still failed.
 544                          */
 545                         if ((i - base_bit) > max_hop) {
 546                                 *phys_cpos = 0;
 547                                 break;
 548                         }
 549
 550                         if (last_free_bits)
 551                                 last_free_bits = 0;
 552
 553                         continue;
 554                 } else
 555                         last_free_bits++;
 556
 557                 if (last_free_bits == move_len) {
 558                         *goal_bit = i;
 559                         *phys_cpos = base_cpos + i;
 560                         break;
 561                 }
 562         }
 563
 564         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 565 }
 566
 567 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 568                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 569                              u32 len, int ext_flags)
 570 {
 571         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 572         handle_t *handle;
 573         struct inode *inode = context->inode;
 574         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 575         struct inode *tl_inode = osb->osb_tl_inode;
 576         struct inode *gb_inode = NULL;
 577         struct buffer_head *gb_bh = NULL;
 578         struct buffer_head *gd_bh = NULL;
 579         struct ocfs2_group_desc *gd;
 580         struct ocfs2_refcount_tree *ref_tree = NULL;
 581         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 582                                                     context->range->me_threshold);
 583         u64 phys_blkno, new_phys_blkno;
 584
 585         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 586
 587         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 588
 589                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 590                          OCFS2_HAS_REFCOUNT_FL));
 591
 592                 BUG_ON(!context->refcount_loc);
 593
 594                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 595                                                &ref_tree, NULL);
 596                 if (ret) {
 597                         mlog_errno(ret);
 598                         return ret;
 599                 }
 600
 601                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 602                                                         context->refcount_loc,
 603                                                         phys_blkno,
 604                                                         len,
 605                                                         &credits,
 606                                                         &extra_blocks);
 607                 if (ret) {
 608                         mlog_errno(ret);
 609                         goto out;
 610                 }
 611         }
 612
 613         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 614                                                 len, 1,
 615                                                 &context->meta_ac,
 616                                                 extra_blocks, &credits);
 617         if (ret) {
 618                 mlog_errno(ret);
 619                 goto out;
 620         }
 621
 622         /*
 623          * need to count 2 extra credits for global_bitmap inode and
 624          * group descriptor.
 625          */
 626         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 627
 628         /*
 629          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 630          * logic, while we still need to lock the global_bitmap.
 631          */
 632         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 633                                                OCFS2_INVALID_SLOT);
 634         if (!gb_inode) {
 635                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 636                 ret = -EIO;
 637                 goto out;
 638         }
 639
 640         inode_lock(gb_inode);
 641
 642         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 643         if (ret) {
 644                 mlog_errno(ret);
 645                 goto out_unlock_gb_mutex;
 646         }
 647
 648         inode_lock(tl_inode);
 649
 650         handle = ocfs2_start_trans(osb, credits);
 651         if (IS_ERR(handle)) {
 652                 ret = PTR_ERR(handle);
 653                 mlog_errno(ret);
 654                 goto out_unlock_tl_inode;
 655         }
 656
 657         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 658         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 659                                             GLOBAL_BITMAP_SYSTEM_INODE,
 660                                             OCFS2_INVALID_SLOT,
 661                                             &goal_bit, &gd_bh);
 662         if (ret) {
 663                 mlog_errno(ret);
 664                 goto out_commit;
 665         }
 666
 667         /*
 668          * probe the victim cluster group to find a proper
 669          * region to fit wanted movement, it even will perfrom
 670          * a best-effort attempt by compromising to a threshold
 671          * around the goal.
 672          */
 673         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 674                                 new_phys_cpos);
 675         if (!*new_phys_cpos) {
 676                 ret = -ENOSPC;
 677                 goto out_commit;
 678         }
 679
 680         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 681                                   *new_phys_cpos, ext_flags);
 682         if (ret) {
 683                 mlog_errno(ret);
 684                 goto out_commit;
 685         }
 686
 687         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 688         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 689                                                le16_to_cpu(gd->bg_chain));
 690         if (ret) {
 691                 mlog_errno(ret);
 692                 goto out_commit;
 693         }
 694
 695         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 696                                          goal_bit, len);
 697         if (ret) {
 698                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 699                                                le16_to_cpu(gd->bg_chain));
 700                 mlog_errno(ret);
 701         }
 702
 703         /*
 704          * Here we should write the new page out first if we are
 705          * in write-back mode.
 706          */
 707         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 708         if (ret)
 709                 mlog_errno(ret);
 710
 711 out_commit:
 712         ocfs2_commit_trans(osb, handle);
 713         brelse(gd_bh);
 714
 715 out_unlock_tl_inode:
 716         inode_unlock(tl_inode);
 717
 718         ocfs2_inode_unlock(gb_inode, 1);
 719 out_unlock_gb_mutex:
 720         inode_unlock(gb_inode);
 721         brelse(gb_bh);
 722         iput(gb_inode);
 723
 724 out:
 725         if (context->meta_ac) {
 726                 ocfs2_free_alloc_context(context->meta_ac);
 727                 context->meta_ac = NULL;
 728         }
 729
 730         if (ref_tree)
 731                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 732
 733         return ret;
 734 }
 735
 736 /*
 737  * Helper to calculate the defraging length in one run according to threshold.
 738  */
 739 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 740                                          u32 threshold, int *skip)
 741 {
 742         if ((*alloc_size + *len_defraged) < threshold) {
 743                 /*
 744                  * proceed defragmentation until we meet the thresh
 745                  */
 746                 *len_defraged += *alloc_size;
 747         } else if (*len_defraged == 0) {
 748                 /*
 749                  * XXX: skip a large extent.
 750                  */
 751                 *skip = 1;
 752         } else {
 753                 /*
 754                  * split this extent to coalesce with former pieces as
 755                  * to reach the threshold.
 756                  *
 757                  * we're done here with one cycle of defragmentation
 758                  * in a size of 'thresh', resetting 'len_defraged'
 759                  * forces a new defragmentation.
 760                  */
 761                 *alloc_size = threshold - *len_defraged;
 762                 *len_defraged = 0;
 763         }
 764 }
 765
 766 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 767                                 struct ocfs2_move_extents_context *context)
 768 {
 769         int ret = 0, flags, do_defrag, skip = 0;
 770         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 771         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 772
 773         struct inode *inode = context->inode;
 774         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 775         struct ocfs2_move_extents *range = context->range;
 776         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 777
 778         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 779                 return 0;
 780
 781         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 782                 return 0;
 783
 784         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 785
 786         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 787         ocfs2_init_dealloc_ctxt(&context->dealloc);
 788
 789         /*
 790          * TO-DO XXX:
 791          *
 792          * - xattr extents.
 793          */
 794
 795         do_defrag = context->auto_defrag;
 796
 797         /*
 798          * extents moving happens in unit of clusters, for the sake
 799          * of simplicity, we may ignore two clusters where 'byte_start'
 800          * and 'byte_start + len' were within.
 801          */
 802         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 803         len_to_move = (range->me_start + range->me_len) >>
 804                                                 osb->s_clustersize_bits;
 805         if (len_to_move >= move_start)
 806                 len_to_move -= move_start;
 807         else
 808                 len_to_move = 0;
 809
 810         if (do_defrag) {
 811                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 812                 if (defrag_thresh <= 1)
 813                         goto done;
 814         } else
 815                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 816                                                          range->me_goal);
 817
 818         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 819              "thresh: %u\n",
 820              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 821              (unsigned long long)range->me_start,
 822              (unsigned long long)range->me_len,
 823              move_start, len_to_move, defrag_thresh);
 824
 825         cpos = move_start;
 826         while (len_to_move) {
 827                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 828                                          &flags);
 829                 if (ret) {
 830                         mlog_errno(ret);
 831                         goto out;
 832                 }
 833
 834                 if (alloc_size > len_to_move)
 835                         alloc_size = len_to_move;
 836
 837                 /*
 838                  * XXX: how to deal with a hole:
 839                  *
 840                  * - skip the hole of course
 841                  * - force a new defragmentation
 842                  */
 843                 if (!phys_cpos) {
 844                         if (do_defrag)
 845                                 len_defraged = 0;
 846
 847                         goto next;
 848                 }
 849
 850                 if (do_defrag) {
 851                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 852                                                      defrag_thresh, &skip);
 853                         /*
 854                          * skip large extents
 855                          */
 856                         if (skip) {
 857                                 skip = 0;
 858                                 goto next;
 859                         }
 860
 861                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 862                              "alloc_size: %u, len_defraged: %u\n",
 863                              cpos, phys_cpos, alloc_size, len_defraged);
 864
 865                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 866                                                   &alloc_size, flags);
 867                 } else {
 868                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 869                                                 &new_phys_cpos, alloc_size,
 870                                                 flags);
 871
 872                         new_phys_cpos += alloc_size;
 873                 }
 874
 875                 if (ret < 0) {
 876                         mlog_errno(ret);
 877                         goto out;
 878                 }
 879
 880                 context->clusters_moved += alloc_size;
 881 next:
 882                 cpos += alloc_size;
 883                 len_to_move -= alloc_size;
 884         }
 885
 886 done:
 887         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 888
 889 out:
 890         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 891                                                       context->clusters_moved);
 892         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 893                                                        context->new_phys_cpos);
 894
 895         ocfs2_schedule_truncate_log_flush(osb, 1);
 896         ocfs2_run_deallocs(osb, &context->dealloc);
 897
 898         return ret;
 899 }
 900
 901 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 902 {
 903         int status;
 904         handle_t *handle;
 905         struct inode *inode = context->inode;
 906         struct ocfs2_dinode *di;
 907         struct buffer_head *di_bh = NULL;
 908         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 909
 910         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 911                 return -EROFS;
 912
 913         inode_lock(inode);
 914
 915         /*
 916          * This prevents concurrent writes from other nodes
 917          */
 918         status = ocfs2_rw_lock(inode, 1);
 919         if (status) {
 920                 mlog_errno(status);
 921                 goto out;
 922         }
 923
 924         status = ocfs2_inode_lock(inode, &di_bh, 1);
 925         if (status) {
 926                 mlog_errno(status);
 927                 goto out_rw_unlock;
 928         }
 929
 930         /*
 931          * rememer ip_xattr_sem also needs to be held if necessary
 932          */
 933         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 934
 935         status = __ocfs2_move_extents_range(di_bh, context);
 936
 937         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 938         if (status) {
 939                 mlog_errno(status);
 940                 goto out_inode_unlock;
 941         }
 942
 943         /*
 944          * We update ctime for these changes
 945          */
 946         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 947         if (IS_ERR(handle)) {
 948                 status = PTR_ERR(handle);
 949                 mlog_errno(status);
 950                 goto out_inode_unlock;
 951         }
 952
 953         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 954                                          OCFS2_JOURNAL_ACCESS_WRITE);
 955         if (status) {
 956                 mlog_errno(status);
 957                 goto out_commit;
 958         }
 959
 960         di = (struct ocfs2_dinode *)di_bh->b_data;
 961         inode->i_ctime = current_time(inode);
 962         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 963         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 964         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 965
 966         ocfs2_journal_dirty(handle, di_bh);
 967
 968 out_commit:
 969         ocfs2_commit_trans(osb, handle);
 970
 971 out_inode_unlock:
 972         brelse(di_bh);
 973         ocfs2_inode_unlock(inode, 1);
 974 out_rw_unlock:
 975         ocfs2_rw_unlock(inode, 1);
 976 out:
 977         inode_unlock(inode);
 978
 979         return status;
 980 }
 981
 982 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 983 {
 984         int status;
 985
 986         struct inode *inode = file_inode(filp);
 987         struct ocfs2_move_extents range;
 988         struct ocfs2_move_extents_context *context;
 989
 990         if (!argp)
 991                 return -EINVAL;
 992
 993         status = mnt_want_write_file(filp);
 994         if (status)
 995                 return status;
 996
 997         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
 998                 status = -EPERM;
 999                 goto out_drop;
1000         }
1001
1002         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1003                 status = -EPERM;
1004                 goto out_drop;
1005         }
1006
1007         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1008         if (!context) {
1009                 status = -ENOMEM;
1010                 mlog_errno(status);
1011                 goto out_drop;
1012         }
1013
1014         context->inode = inode;
1015         context->file = filp;
1016
1017         if (copy_from_user(&range, argp, sizeof(range))) {
1018                 status = -EFAULT;
1019                 goto out_free;
1020         }
1021
1022         if (range.me_start > i_size_read(inode)) {
1023                 status = -EINVAL;
1024                 goto out_free;
1025         }
1026
1027         if (range.me_start + range.me_len > i_size_read(inode))
1028                         range.me_len = i_size_read(inode) - range.me_start;
1029
1030         context->range = &range;
1031
1032         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1033                 context->auto_defrag = 1;
1034                 /*
1035                  * ok, the default theshold for the defragmentation
1036                  * is 1M, since our maximum clustersize was 1M also.
1037                  * any thought?
1038                  */
1039                 if (!range.me_threshold)
1040                         range.me_threshold = 1024 * 1024;
1041
1042                 if (range.me_threshold > i_size_read(inode))
1043                         range.me_threshold = i_size_read(inode);
1044
1045                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1046                         context->partial = 1;
1047         } else {
1048                 /*
1049                  * first best-effort attempt to validate and adjust the goal
1050                  * (physical address in block), while it can't guarantee later
1051                  * operation can succeed all the time since global_bitmap may
1052                  * change a bit over time.
1053                  */
1054
1055                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1056                 if (status)
1057                         goto out_copy;
1058         }
1059
1060         status = ocfs2_move_extents(context);
1061         if (status)
1062                 mlog_errno(status);
1063 out_copy:
1064         /*
1065          * movement/defragmentation may end up being partially completed,
1066          * that's the reason why we need to return userspace the finished
1067          * length and new_offset even if failure happens somewhere.
1068          */
1069         if (copy_to_user(argp, &range, sizeof(range)))
1070                 status = -EFAULT;
1071
1072 out_free:
1073         kfree(context);
1074 out_drop:
1075         mnt_drop_write_file(filp);
1076
1077         return status;
1078 }