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