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, context->file, 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 || index >= le16_to_cpu(el->l_next_free_rec)) {
 102                 ocfs2_error(inode->i_sb,
 103                             "Inode %llu has an extent at cpos %u which can no "
 104                             "longer be found.\n",
 105                             (unsigned long long)ino, cpos);
 106                 ret = -EROFS;
 107                 goto out;
 108         }
 109
 110         rec = &el->l_recs[index];
 111
 112         BUG_ON(ext_flags != rec->e_flags);
 113         /*
 114          * after moving/defraging to new location, the extent is not going
 115          * to be refcounted anymore.
 116          */
 117         replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
 118
 119         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
 120                                       context->et.et_root_bh,
 121                                       OCFS2_JOURNAL_ACCESS_WRITE);
 122         if (ret) {
 123                 mlog_errno(ret);
 124                 goto out;
 125         }
 126
 127         ret = ocfs2_split_extent(handle, &context->et, path, index,
 128                                  &replace_rec, context->meta_ac,
 129                                  &context->dealloc);
 130         if (ret) {
 131                 mlog_errno(ret);
 132                 goto out;
 133         }
 134
 135         ocfs2_journal_dirty(handle, context->et.et_root_bh);
 136
 137         context->new_phys_cpos = new_p_cpos;
 138
 139         /*
 140          * need I to append truncate log for old clusters?
 141          */
 142         if (old_blkno) {
 143                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 144                         ret = ocfs2_decrease_refcount(inode, handle,
 145                                         ocfs2_blocks_to_clusters(osb->sb,
 146                                                                  old_blkno),
 147                                         len, context->meta_ac,
 148                                         &context->dealloc, 1);
 149                 else
 150                         ret = ocfs2_truncate_log_append(osb, handle,
 151                                                         old_blkno, len);
 152         }
 153
 154 out:
 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                                               clusters_to_move + 2);
 205
 206         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 207              extra_blocks, clusters_to_move, *credits);
 208 out:
 209         if (ret) {
 210                 if (*meta_ac) {
 211                         ocfs2_free_alloc_context(*meta_ac);
 212                         *meta_ac = NULL;
 213                 }
 214         }
 215
 216         return ret;
 217 }
 218
 219 /*
 220  * Using one journal handle to guarantee the data consistency in case
 221  * crash happens anywhere.
 222  *
 223  *  XXX: defrag can end up with finishing partial extent as requested,
 224  * due to not enough contiguous clusters can be found in allocator.
 225  */
 226 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 227                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 228 {
 229         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 230         handle_t *handle;
 231         struct inode *inode = context->inode;
 232         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 233         struct inode *tl_inode = osb->osb_tl_inode;
 234         struct ocfs2_refcount_tree *ref_tree = NULL;
 235         u32 new_phys_cpos, new_len;
 236         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 237
 238         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 239
 240                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 241                          OCFS2_HAS_REFCOUNT_FL));
 242
 243                 BUG_ON(!context->refcount_loc);
 244
 245                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 246                                                &ref_tree, NULL);
 247                 if (ret) {
 248                         mlog_errno(ret);
 249                         return ret;
 250                 }
 251
 252                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 253                                                         context->refcount_loc,
 254                                                         phys_blkno,
 255                                                         *len,
 256                                                         &credits,
 257                                                         &extra_blocks);
 258                 if (ret) {
 259                         mlog_errno(ret);
 260                         goto out;
 261                 }
 262         }
 263
 264         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, *len, 1,
 265                                                  &context->meta_ac,
 266                                                  &context->data_ac,
 267                                                  extra_blocks, &credits);
 268         if (ret) {
 269                 mlog_errno(ret);
 270                 goto out;
 271         }
 272
 273         /*
 274          * should be using allocation reservation strategy there?
 275          *
 276          * if (context->data_ac)
 277          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 278          */
 279
 280         mutex_lock(&tl_inode->i_mutex);
 281
 282         if (ocfs2_truncate_log_needs_flush(osb)) {
 283                 ret = __ocfs2_flush_truncate_log(osb);
 284                 if (ret < 0) {
 285                         mlog_errno(ret);
 286                         goto out_unlock_mutex;
 287                 }
 288         }
 289
 290         handle = ocfs2_start_trans(osb, credits);
 291         if (IS_ERR(handle)) {
 292                 ret = PTR_ERR(handle);
 293                 mlog_errno(ret);
 294                 goto out_unlock_mutex;
 295         }
 296
 297         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 298                                      &new_phys_cpos, &new_len);
 299         if (ret) {
 300                 mlog_errno(ret);
 301                 goto out_commit;
 302         }
 303
 304         /*
 305          * allowing partial extent moving is kind of 'pros and cons', it makes
 306          * whole defragmentation less likely to fail, on the contrary, the bad
 307          * thing is it may make the fs even more fragmented after moving, let
 308          * userspace make a good decision here.
 309          */
 310         if (new_len != *len) {
 311                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 312                 if (!partial) {
 313                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 314                         ret = -ENOSPC;
 315                         goto out_commit;
 316                 }
 317         }
 318
 319         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 320              phys_cpos, new_phys_cpos);
 321
 322         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 323                                   new_phys_cpos, ext_flags);
 324         if (ret)
 325                 mlog_errno(ret);
 326
 327         if (partial && (new_len != *len))
 328                 *len = new_len;
 329
 330         /*
 331          * Here we should write the new page out first if we are
 332          * in write-back mode.
 333          */
 334         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 335         if (ret)
 336                 mlog_errno(ret);
 337
 338 out_commit:
 339         ocfs2_commit_trans(osb, handle);
 340
 341 out_unlock_mutex:
 342         mutex_unlock(&tl_inode->i_mutex);
 343
 344         if (context->data_ac) {
 345                 ocfs2_free_alloc_context(context->data_ac);
 346                 context->data_ac = NULL;
 347         }
 348
 349         if (context->meta_ac) {
 350                 ocfs2_free_alloc_context(context->meta_ac);
 351                 context->meta_ac = NULL;
 352         }
 353
 354 out:
 355         if (ref_tree)
 356                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 357
 358         return ret;
 359 }
 360
 361 /*
 362  * find the victim alloc group, where #blkno fits.
 363  */
 364 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 365                                          u64 vict_blkno,
 366                                          int type, int slot,
 367                                          int *vict_bit,
 368                                          struct buffer_head **ret_bh)
 369 {
 370         int ret, i, bits_per_unit = 0;
 371         u64 blkno;
 372         char namebuf[40];
 373
 374         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 375         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 376         struct ocfs2_chain_list *cl;
 377         struct ocfs2_chain_rec *rec;
 378         struct ocfs2_dinode *ac_dinode;
 379         struct ocfs2_group_desc *bg;
 380
 381         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 382         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 383                                          strlen(namebuf), &blkno);
 384         if (ret) {
 385                 ret = -ENOENT;
 386                 goto out;
 387         }
 388
 389         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 390         if (ret) {
 391                 mlog_errno(ret);
 392                 goto out;
 393         }
 394
 395         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 396         cl = &(ac_dinode->id2.i_chain);
 397         rec = &(cl->cl_recs[0]);
 398
 399         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 400                 bits_per_unit = osb->s_clustersize_bits -
 401                                         inode->i_sb->s_blocksize_bits;
 402         /*
 403          * 'vict_blkno' was out of the valid range.
 404          */
 405         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 406             (vict_blkno >= (le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 407                                 bits_per_unit))) {
 408                 ret = -EINVAL;
 409                 goto out;
 410         }
 411
 412         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 413
 414                 rec = &(cl->cl_recs[i]);
 415                 if (!rec)
 416                         continue;
 417
 418                 bg = NULL;
 419
 420                 do {
 421                         if (!bg)
 422                                 blkno = le64_to_cpu(rec->c_blkno);
 423                         else
 424                                 blkno = le64_to_cpu(bg->bg_next_group);
 425
 426                         if (gd_bh) {
 427                                 brelse(gd_bh);
 428                                 gd_bh = NULL;
 429                         }
 430
 431                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 432                         if (ret) {
 433                                 mlog_errno(ret);
 434                                 goto out;
 435                         }
 436
 437                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 438
 439                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 440                                                 le16_to_cpu(bg->bg_bits))) {
 441
 442                                 *ret_bh = gd_bh;
 443                                 *vict_bit = (vict_blkno - blkno) >>
 444                                                         bits_per_unit;
 445                                 mlog(0, "find the victim group: #%llu, "
 446                                      "total_bits: %u, vict_bit: %u\n",
 447                                      blkno, le16_to_cpu(bg->bg_bits),
 448                                      *vict_bit);
 449                                 goto out;
 450                         }
 451
 452                 } while (le64_to_cpu(bg->bg_next_group));
 453         }
 454
 455         ret = -EINVAL;
 456 out:
 457         brelse(ac_bh);
 458
 459         /*
 460          * caller has to release the gd_bh properly.
 461          */
 462         return ret;
 463 }
 464
 465 /*
 466  * XXX: helper to validate and adjust moving goal.
 467  */
 468 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 469                                                struct ocfs2_move_extents *range)
 470 {
 471         int ret, goal_bit = 0;
 472
 473         struct buffer_head *gd_bh = NULL;
 474         struct ocfs2_group_desc *bg = NULL;
 475         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 476         int c_to_b = 1 << (osb->s_clustersize_bits -
 477                                         inode->i_sb->s_blocksize_bits);
 478
 479         /*
 480          * make goal become cluster aligned.
 481          */
 482         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 483                                                       range->me_goal);
 484         /*
 485          * moving goal is not allowd to start with a group desc blok(#0 blk)
 486          * let's compromise to the latter cluster.
 487          */
 488         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 489                 range->me_goal += c_to_b;
 490
 491         /*
 492          * validate goal sits within global_bitmap, and return the victim
 493          * group desc
 494          */
 495         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 496                                             GLOBAL_BITMAP_SYSTEM_INODE,
 497                                             OCFS2_INVALID_SLOT,
 498                                             &goal_bit, &gd_bh);
 499         if (ret)
 500                 goto out;
 501
 502         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 503
 504         /*
 505          * movement is not gonna cross two groups.
 506          */
 507         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 508                                                                 range->me_len) {
 509                 ret = -EINVAL;
 510                 goto out;
 511         }
 512         /*
 513          * more exact validations/adjustments will be performed later during
 514          * moving operation for each extent range.
 515          */
 516         mlog(0, "extents get ready to be moved to #%llu block\n",
 517              range->me_goal);
 518
 519 out:
 520         brelse(gd_bh);
 521
 522         return ret;
 523 }
 524
 525 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 526                                     int *goal_bit, u32 move_len, u32 max_hop,
 527                                     u32 *phys_cpos)
 528 {
 529         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 530         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 531         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 532                                                  le64_to_cpu(gd->bg_blkno));
 533
 534         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 535
 536                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 537                 if (used) {
 538                         /*
 539                          * we even tried searching the free chunk by jumping
 540                          * a 'max_hop' distance, but still failed.
 541                          */
 542                         if ((i - base_bit) > max_hop) {
 543                                 *phys_cpos = 0;
 544                                 break;
 545                         }
 546
 547                         if (last_free_bits)
 548                                 last_free_bits = 0;
 549
 550                         continue;
 551                 } else
 552                         last_free_bits++;
 553
 554                 if (last_free_bits == move_len) {
 555                         *goal_bit = i;
 556                         *phys_cpos = base_cpos + i;
 557                         break;
 558                 }
 559         }
 560
 561         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 562 }
 563
 564 static int ocfs2_alloc_dinode_update_counts(struct inode *inode,
 565                                        handle_t *handle,
 566                                        struct buffer_head *di_bh,
 567                                        u32 num_bits,
 568                                        u16 chain)
 569 {
 570         int ret;
 571         u32 tmp_used;
 572         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
 573         struct ocfs2_chain_list *cl =
 574                                 (struct ocfs2_chain_list *) &di->id2.i_chain;
 575
 576         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 577                                       OCFS2_JOURNAL_ACCESS_WRITE);
 578         if (ret < 0) {
 579                 mlog_errno(ret);
 580                 goto out;
 581         }
 582
 583         tmp_used = le32_to_cpu(di->id1.bitmap1.i_used);
 584         di->id1.bitmap1.i_used = cpu_to_le32(num_bits + tmp_used);
 585         le32_add_cpu(&cl->cl_recs[chain].c_free, -num_bits);
 586         ocfs2_journal_dirty(handle, di_bh);
 587
 588 out:
 589         return ret;
 590 }
 591
 592 static inline int ocfs2_block_group_set_bits(handle_t *handle,
 593                                              struct inode *alloc_inode,
 594                                              struct ocfs2_group_desc *bg,
 595                                              struct buffer_head *group_bh,
 596                                              unsigned int bit_off,
 597                                              unsigned int num_bits)
 598 {
 599         int status;
 600         void *bitmap = bg->bg_bitmap;
 601         int journal_type = OCFS2_JOURNAL_ACCESS_WRITE;
 602
 603         /* All callers get the descriptor via
 604          * ocfs2_read_group_descriptor().  Any corruption is a code bug. */
 605         BUG_ON(!OCFS2_IS_VALID_GROUP_DESC(bg));
 606         BUG_ON(le16_to_cpu(bg->bg_free_bits_count) < num_bits);
 607
 608         mlog(0, "block_group_set_bits: off = %u, num = %u\n", bit_off,
 609              num_bits);
 610
 611         if (ocfs2_is_cluster_bitmap(alloc_inode))
 612                 journal_type = OCFS2_JOURNAL_ACCESS_UNDO;
 613
 614         status = ocfs2_journal_access_gd(handle,
 615                                          INODE_CACHE(alloc_inode),
 616                                          group_bh,
 617                                          journal_type);
 618         if (status < 0) {
 619                 mlog_errno(status);
 620                 goto bail;
 621         }
 622
 623         le16_add_cpu(&bg->bg_free_bits_count, -num_bits);
 624         if (le16_to_cpu(bg->bg_free_bits_count) > le16_to_cpu(bg->bg_bits)) {
 625                 ocfs2_error(alloc_inode->i_sb, "Group descriptor # %llu has bit"
 626                             " count %u but claims %u are freed. num_bits %d",
 627                             (unsigned long long)le64_to_cpu(bg->bg_blkno),
 628                             le16_to_cpu(bg->bg_bits),
 629                             le16_to_cpu(bg->bg_free_bits_count), num_bits);
 630                 return -EROFS;
 631         }
 632         while (num_bits--)
 633                 ocfs2_set_bit(bit_off++, bitmap);
 634
 635         ocfs2_journal_dirty(handle, group_bh);
 636
 637 bail:
 638         return status;
 639 }
 640
 641 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 642                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 643                              u32 len, int ext_flags)
 644 {
 645         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 646         handle_t *handle;
 647         struct inode *inode = context->inode;
 648         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 649         struct inode *tl_inode = osb->osb_tl_inode;
 650         struct inode *gb_inode = NULL;
 651         struct buffer_head *gb_bh = NULL;
 652         struct buffer_head *gd_bh = NULL;
 653         struct ocfs2_group_desc *gd;
 654         struct ocfs2_refcount_tree *ref_tree = NULL;
 655         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 656                                                     context->range->me_threshold);
 657         u64 phys_blkno, new_phys_blkno;
 658
 659         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 660
 661         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 662
 663                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 664                          OCFS2_HAS_REFCOUNT_FL));
 665
 666                 BUG_ON(!context->refcount_loc);
 667
 668                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 669                                                &ref_tree, NULL);
 670                 if (ret) {
 671                         mlog_errno(ret);
 672                         return ret;
 673                 }
 674
 675                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 676                                                         context->refcount_loc,
 677                                                         phys_blkno,
 678                                                         len,
 679                                                         &credits,
 680                                                         &extra_blocks);
 681                 if (ret) {
 682                         mlog_errno(ret);
 683                         goto out;
 684                 }
 685         }
 686
 687         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, len, 1,
 688                                                  &context->meta_ac,
 689                                                  NULL, extra_blocks, &credits);
 690         if (ret) {
 691                 mlog_errno(ret);
 692                 goto out;
 693         }
 694
 695         /*
 696          * need to count 2 extra credits for global_bitmap inode and
 697          * group descriptor.
 698          */
 699         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 700
 701         /*
 702          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 703          * logic, while we still need to lock the global_bitmap.
 704          */
 705         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 706                                                OCFS2_INVALID_SLOT);
 707         if (!gb_inode) {
 708                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 709                 ret = -EIO;
 710                 goto out;
 711         }
 712
 713         mutex_lock(&gb_inode->i_mutex);
 714
 715         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 716         if (ret) {
 717                 mlog_errno(ret);
 718                 goto out_unlock_gb_mutex;
 719         }
 720
 721         mutex_lock(&tl_inode->i_mutex);
 722
 723         handle = ocfs2_start_trans(osb, credits);
 724         if (IS_ERR(handle)) {
 725                 ret = PTR_ERR(handle);
 726                 mlog_errno(ret);
 727                 goto out_unlock_tl_inode;
 728         }
 729
 730         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 731         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 732                                             GLOBAL_BITMAP_SYSTEM_INODE,
 733                                             OCFS2_INVALID_SLOT,
 734                                             &goal_bit, &gd_bh);
 735         if (ret) {
 736                 mlog_errno(ret);
 737                 goto out_commit;
 738         }
 739
 740         /*
 741          * probe the victim cluster group to find a proper
 742          * region to fit wanted movement, it even will perfrom
 743          * a best-effort attempt by compromising to a threshold
 744          * around the goal.
 745          */
 746         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 747                                 new_phys_cpos);
 748         if (!*new_phys_cpos) {
 749                 ret = -ENOSPC;
 750                 goto out_commit;
 751         }
 752
 753         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 754                                   *new_phys_cpos, ext_flags);
 755         if (ret) {
 756                 mlog_errno(ret);
 757                 goto out_commit;
 758         }
 759
 760         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 761         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 762                                                le16_to_cpu(gd->bg_chain));
 763         if (ret) {
 764                 mlog_errno(ret);
 765                 goto out_commit;
 766         }
 767
 768         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 769                                          goal_bit, len);
 770         if (ret)
 771                 mlog_errno(ret);
 772
 773         /*
 774          * Here we should write the new page out first if we are
 775          * in write-back mode.
 776          */
 777         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 778         if (ret)
 779                 mlog_errno(ret);
 780
 781 out_commit:
 782         ocfs2_commit_trans(osb, handle);
 783         brelse(gd_bh);
 784
 785 out_unlock_tl_inode:
 786         mutex_unlock(&tl_inode->i_mutex);
 787
 788         ocfs2_inode_unlock(gb_inode, 1);
 789 out_unlock_gb_mutex:
 790         mutex_unlock(&gb_inode->i_mutex);
 791         brelse(gb_bh);
 792         iput(gb_inode);
 793
 794 out:
 795         if (context->meta_ac) {
 796                 ocfs2_free_alloc_context(context->meta_ac);
 797                 context->meta_ac = NULL;
 798         }
 799
 800         if (ref_tree)
 801                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 802
 803         return ret;
 804 }
 805
 806 /*
 807  * Helper to calculate the defraging length in one run according to threshold.
 808  */
 809 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 810                                          u32 threshold, int *skip)
 811 {
 812         if ((*alloc_size + *len_defraged) < threshold) {
 813                 /*
 814                  * proceed defragmentation until we meet the thresh
 815                  */
 816                 *len_defraged += *alloc_size;
 817         } else if (*len_defraged == 0) {
 818                 /*
 819                  * XXX: skip a large extent.
 820                  */
 821                 *skip = 1;
 822         } else {
 823                 /*
 824                  * split this extent to coalesce with former pieces as
 825                  * to reach the threshold.
 826                  *
 827                  * we're done here with one cycle of defragmentation
 828                  * in a size of 'thresh', resetting 'len_defraged'
 829                  * forces a new defragmentation.
 830                  */
 831                 *alloc_size = threshold - *len_defraged;
 832                 *len_defraged = 0;
 833         }
 834 }
 835
 836 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 837                                 struct ocfs2_move_extents_context *context)
 838 {
 839         int ret = 0, flags, do_defrag, skip = 0;
 840         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 841         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 842
 843         struct inode *inode = context->inode;
 844         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 845         struct ocfs2_move_extents *range = context->range;
 846         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 847
 848         if ((inode->i_size == 0) || (range->me_len == 0))
 849                 return 0;
 850
 851         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 852                 return 0;
 853
 854         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 855
 856         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 857         ocfs2_init_dealloc_ctxt(&context->dealloc);
 858
 859         /*
 860          * TO-DO XXX:
 861          *
 862          * - xattr extents.
 863          */
 864
 865         do_defrag = context->auto_defrag;
 866
 867         /*
 868          * extents moving happens in unit of clusters, for the sake
 869          * of simplicity, we may ignore two clusters where 'byte_start'
 870          * and 'byte_start + len' were within.
 871          */
 872         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 873         len_to_move = (range->me_start + range->me_len) >>
 874                                                 osb->s_clustersize_bits;
 875         if (len_to_move >= move_start)
 876                 len_to_move -= move_start;
 877         else
 878                 len_to_move = 0;
 879
 880         if (do_defrag) {
 881                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 882                 if (defrag_thresh <= 1)
 883                         goto done;
 884         } else
 885                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 886                                                          range->me_goal);
 887
 888         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 889              "thresh: %u\n",
 890              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 891              (unsigned long long)range->me_start,
 892              (unsigned long long)range->me_len,
 893              move_start, len_to_move, defrag_thresh);
 894
 895         cpos = move_start;
 896         while (len_to_move) {
 897                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 898                                          &flags);
 899                 if (ret) {
 900                         mlog_errno(ret);
 901                         goto out;
 902                 }
 903
 904                 if (alloc_size > len_to_move)
 905                         alloc_size = len_to_move;
 906
 907                 /*
 908                  * XXX: how to deal with a hole:
 909                  *
 910                  * - skip the hole of course
 911                  * - force a new defragmentation
 912                  */
 913                 if (!phys_cpos) {
 914                         if (do_defrag)
 915                                 len_defraged = 0;
 916
 917                         goto next;
 918                 }
 919
 920                 if (do_defrag) {
 921                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 922                                                      defrag_thresh, &skip);
 923                         /*
 924                          * skip large extents
 925                          */
 926                         if (skip) {
 927                                 skip = 0;
 928                                 goto next;
 929                         }
 930
 931                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 932                              "alloc_size: %u, len_defraged: %u\n",
 933                              cpos, phys_cpos, alloc_size, len_defraged);
 934
 935                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 936                                                   &alloc_size, flags);
 937                 } else {
 938                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 939                                                 &new_phys_cpos, alloc_size,
 940                                                 flags);
 941
 942                         new_phys_cpos += alloc_size;
 943                 }
 944
 945                 if (ret < 0) {
 946                         mlog_errno(ret);
 947                         goto out;
 948                 }
 949
 950                 context->clusters_moved += alloc_size;
 951 next:
 952                 cpos += alloc_size;
 953                 len_to_move -= alloc_size;
 954         }
 955
 956 done:
 957         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 958
 959 out:
 960         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 961                                                       context->clusters_moved);
 962         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 963                                                        context->new_phys_cpos);
 964
 965         ocfs2_schedule_truncate_log_flush(osb, 1);
 966         ocfs2_run_deallocs(osb, &context->dealloc);
 967
 968         return ret;
 969 }
 970
 971 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 972 {
 973         int status;
 974         handle_t *handle;
 975         struct inode *inode = context->inode;
 976         struct ocfs2_dinode *di;
 977         struct buffer_head *di_bh = NULL;
 978         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 979
 980         if (!inode)
 981                 return -ENOENT;
 982
 983         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 984                 return -EROFS;
 985
 986         mutex_lock(&inode->i_mutex);
 987
 988         /*
 989          * This prevents concurrent writes from other nodes
 990          */
 991         status = ocfs2_rw_lock(inode, 1);
 992         if (status) {
 993                 mlog_errno(status);
 994                 goto out;
 995         }
 996
 997         status = ocfs2_inode_lock(inode, &di_bh, 1);
 998         if (status) {
 999                 mlog_errno(status);
1000                 goto out_rw_unlock;
1001         }
1002
1003         /*
1004          * rememer ip_xattr_sem also needs to be held if necessary
1005          */
1006         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1007
1008         status = __ocfs2_move_extents_range(di_bh, context);
1009
1010         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1011         if (status) {
1012                 mlog_errno(status);
1013                 goto out_inode_unlock;
1014         }
1015
1016         /*
1017          * We update ctime for these changes
1018          */
1019         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1020         if (IS_ERR(handle)) {
1021                 status = PTR_ERR(handle);
1022                 mlog_errno(status);
1023                 goto out_inode_unlock;
1024         }
1025
1026         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
1027                                          OCFS2_JOURNAL_ACCESS_WRITE);
1028         if (status) {
1029                 mlog_errno(status);
1030                 goto out_commit;
1031         }
1032
1033         di = (struct ocfs2_dinode *)di_bh->b_data;
1034         inode->i_ctime = CURRENT_TIME;
1035         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
1036         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
1037
1038         ocfs2_journal_dirty(handle, di_bh);
1039
1040 out_commit:
1041         ocfs2_commit_trans(osb, handle);
1042
1043 out_inode_unlock:
1044         brelse(di_bh);
1045         ocfs2_inode_unlock(inode, 1);
1046 out_rw_unlock:
1047         ocfs2_rw_unlock(inode, 1);
1048 out:
1049         mutex_unlock(&inode->i_mutex);
1050
1051         return status;
1052 }
1053
1054 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
1055 {
1056         int status;
1057
1058         struct inode *inode = filp->f_path.dentry->d_inode;
1059         struct ocfs2_move_extents range;
1060         struct ocfs2_move_extents_context *context = NULL;
1061
1062         status = mnt_want_write(filp->f_path.mnt);
1063         if (status)
1064                 return status;
1065
1066         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE))
1067                 goto out;
1068
1069         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1070                 status = -EPERM;
1071                 goto out;
1072         }
1073
1074         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1075         if (!context) {
1076                 status = -ENOMEM;
1077                 mlog_errno(status);
1078                 goto out;
1079         }
1080
1081         context->inode = inode;
1082         context->file = filp;
1083
1084         if (argp) {
1085                 if (copy_from_user(&range, (struct ocfs2_move_extents *)argp,
1086                                    sizeof(range))) {
1087                         status = -EFAULT;
1088                         goto out;
1089                 }
1090         } else {
1091                 status = -EINVAL;
1092                 goto out;
1093         }
1094
1095         if (range.me_start > i_size_read(inode))
1096                 goto out;
1097
1098         if (range.me_start + range.me_len > i_size_read(inode))
1099                         range.me_len = i_size_read(inode) - range.me_start;
1100
1101         context->range = &range;
1102
1103         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1104                 context->auto_defrag = 1;
1105                 /*
1106                  * ok, the default theshold for the defragmentation
1107                  * is 1M, since our maximum clustersize was 1M also.
1108                  * any thought?
1109                  */
1110                 if (!range.me_threshold)
1111                         range.me_threshold = 1024 * 1024;
1112
1113                 if (range.me_threshold > i_size_read(inode))
1114                         range.me_threshold = i_size_read(inode);
1115
1116                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1117                         context->partial = 1;
1118         } else {
1119                 /*
1120                  * first best-effort attempt to validate and adjust the goal
1121                  * (physical address in block), while it can't guarantee later
1122                  * operation can succeed all the time since global_bitmap may
1123                  * change a bit over time.
1124                  */
1125
1126                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1127                 if (status)
1128                         goto out;
1129         }
1130
1131         status = ocfs2_move_extents(context);
1132         if (status)
1133                 mlog_errno(status);
1134 out:
1135         /*
1136          * movement/defragmentation may end up being partially completed,
1137          * that's the reason why we need to return userspace the finished
1138          * length and new_offset even if failure happens somewhere.
1139          */
1140         if (argp) {
1141                 if (copy_to_user((struct ocfs2_move_extents *)argp, &range,
1142                                 sizeof(range)))
1143                         status = -EFAULT;
1144         }
1145
1146         kfree(context);
1147
1148         mnt_drop_write(filp->f_path.mnt);
1149
1150         return status;
1151 }