fs/ocfs2/move_extents.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* -*- mode: c; c-basic-offset: 8; -*-
   3  * vim: noexpandtab sw=8 ts=8 sts=0:
   4  *
   5  * move_extents.c
   6  *
   7  * Copyright (C) 2011 Oracle.  All rights reserved.
   8  */
   9 #include <linux/fs.h>
  10 #include <linux/types.h>
  11 #include <linux/mount.h>
  12 #include <linux/swap.h>
  13
  14 #include <cluster/masklog.h>
  15
  16 #include "ocfs2.h"
  17 #include "ocfs2_ioctl.h"
  18
  19 #include "alloc.h"
  20 #include "localalloc.h"
  21 #include "aops.h"
  22 #include "dlmglue.h"
  23 #include "extent_map.h"
  24 #include "inode.h"
  25 #include "journal.h"
  26 #include "suballoc.h"
  27 #include "uptodate.h"
  28 #include "super.h"
  29 #include "dir.h"
  30 #include "buffer_head_io.h"
  31 #include "sysfile.h"
  32 #include "refcounttree.h"
  33 #include "move_extents.h"
  34
  35 struct ocfs2_move_extents_context {
  36         struct inode *inode;
  37         struct file *file;
  38         int auto_defrag;
  39         int partial;
  40         int credits;
  41         u32 new_phys_cpos;
  42         u32 clusters_moved;
  43         u64 refcount_loc;
  44         struct ocfs2_move_extents *range;
  45         struct ocfs2_extent_tree et;
  46         struct ocfs2_alloc_context *meta_ac;
  47         struct ocfs2_alloc_context *data_ac;
  48         struct ocfs2_cached_dealloc_ctxt dealloc;
  49 };
  50
  51 static int __ocfs2_move_extent(handle_t *handle,
  52                                struct ocfs2_move_extents_context *context,
  53                                u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
  54                                int ext_flags)
  55 {
  56         int ret = 0, index;
  57         struct inode *inode = context->inode;
  58         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  59         struct ocfs2_extent_rec *rec, replace_rec;
  60         struct ocfs2_path *path = NULL;
  61         struct ocfs2_extent_list *el;
  62         u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
  63         u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
  64
  65         ret = ocfs2_duplicate_clusters_by_page(handle, inode, cpos,
  66                                                p_cpos, new_p_cpos, len);
  67         if (ret) {
  68                 mlog_errno(ret);
  69                 goto out;
  70         }
  71
  72         memset(&replace_rec, 0, sizeof(replace_rec));
  73         replace_rec.e_cpos = cpu_to_le32(cpos);
  74         replace_rec.e_leaf_clusters = cpu_to_le16(len);
  75         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
  76                                                                    new_p_cpos));
  77
  78         path = ocfs2_new_path_from_et(&context->et);
  79         if (!path) {
  80                 ret = -ENOMEM;
  81                 mlog_errno(ret);
  82                 goto out;
  83         }
  84
  85         ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
  86         if (ret) {
  87                 mlog_errno(ret);
  88                 goto out;
  89         }
  90
  91         el = path_leaf_el(path);
  92
  93         index = ocfs2_search_extent_list(el, cpos);
  94         if (index == -1) {
  95                 ret = ocfs2_error(inode->i_sb,
  96                                   "Inode %llu has an extent at cpos %u which can no longer be found\n",
  97                                   (unsigned long long)ino, cpos);
  98                 goto out;
  99         }
 100
 101         rec = &el->l_recs[index];
 102
 103         BUG_ON(ext_flags != rec->e_flags);
 104         /*
 105          * after moving/defraging to new location, the extent is not going
 106          * to be refcounted anymore.
 107          */
 108         replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
 109
 110         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
 111                                       context->et.et_root_bh,
 112                                       OCFS2_JOURNAL_ACCESS_WRITE);
 113         if (ret) {
 114                 mlog_errno(ret);
 115                 goto out;
 116         }
 117
 118         ret = ocfs2_split_extent(handle, &context->et, path, index,
 119                                  &replace_rec, context->meta_ac,
 120                                  &context->dealloc);
 121         if (ret) {
 122                 mlog_errno(ret);
 123                 goto out;
 124         }
 125
 126         ocfs2_journal_dirty(handle, context->et.et_root_bh);
 127
 128         context->new_phys_cpos = new_p_cpos;
 129
 130         /*
 131          * need I to append truncate log for old clusters?
 132          */
 133         if (old_blkno) {
 134                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 135                         ret = ocfs2_decrease_refcount(inode, handle,
 136                                         ocfs2_blocks_to_clusters(osb->sb,
 137                                                                  old_blkno),
 138                                         len, context->meta_ac,
 139                                         &context->dealloc, 1);
 140                 else
 141                         ret = ocfs2_truncate_log_append(osb, handle,
 142                                                         old_blkno, len);
 143         }
 144
 145         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 146 out:
 147         ocfs2_free_path(path);
 148         return ret;
 149 }
 150
 151 /*
 152  * lock allocator, and reserve appropriate number of bits for
 153  * meta blocks.
 154  */
 155 static int ocfs2_lock_meta_allocator_move_extents(struct inode *inode,
 156                                         struct ocfs2_extent_tree *et,
 157                                         u32 clusters_to_move,
 158                                         u32 extents_to_split,
 159                                         struct ocfs2_alloc_context **meta_ac,
 160                                         int extra_blocks,
 161                                         int *credits)
 162 {
 163         int ret, num_free_extents;
 164         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 165         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 166
 167         num_free_extents = ocfs2_num_free_extents(et);
 168         if (num_free_extents < 0) {
 169                 ret = num_free_extents;
 170                 mlog_errno(ret);
 171                 goto out;
 172         }
 173
 174         if (!num_free_extents ||
 175             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 176                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 177
 178         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 179         if (ret) {
 180                 mlog_errno(ret);
 181                 goto out;
 182         }
 183
 184
 185         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 186
 187         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 188              extra_blocks, clusters_to_move, *credits);
 189 out:
 190         if (ret) {
 191                 if (*meta_ac) {
 192                         ocfs2_free_alloc_context(*meta_ac);
 193                         *meta_ac = NULL;
 194                 }
 195         }
 196
 197         return ret;
 198 }
 199
 200 /*
 201  * Using one journal handle to guarantee the data consistency in case
 202  * crash happens anywhere.
 203  *
 204  *  XXX: defrag can end up with finishing partial extent as requested,
 205  * due to not enough contiguous clusters can be found in allocator.
 206  */
 207 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 208                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 209 {
 210         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 211         handle_t *handle;
 212         struct inode *inode = context->inode;
 213         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 214         struct inode *tl_inode = osb->osb_tl_inode;
 215         struct ocfs2_refcount_tree *ref_tree = NULL;
 216         u32 new_phys_cpos, new_len;
 217         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 218         int need_free = 0;
 219
 220         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 221                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 222                 BUG_ON(!context->refcount_loc);
 223
 224                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 225                                                &ref_tree, NULL);
 226                 if (ret) {
 227                         mlog_errno(ret);
 228                         return ret;
 229                 }
 230
 231                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 232                                                         context->refcount_loc,
 233                                                         phys_blkno,
 234                                                         *len,
 235                                                         &credits,
 236                                                         &extra_blocks);
 237                 if (ret) {
 238                         mlog_errno(ret);
 239                         goto out;
 240                 }
 241         }
 242
 243         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 244                                                 *len, 1,
 245                                                 &context->meta_ac,
 246                                                 extra_blocks, &credits);
 247         if (ret) {
 248                 mlog_errno(ret);
 249                 goto out;
 250         }
 251
 252         /*
 253          * should be using allocation reservation strategy there?
 254          *
 255          * if (context->data_ac)
 256          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 257          */
 258
 259         inode_lock(tl_inode);
 260
 261         if (ocfs2_truncate_log_needs_flush(osb)) {
 262                 ret = __ocfs2_flush_truncate_log(osb);
 263                 if (ret < 0) {
 264                         mlog_errno(ret);
 265                         goto out_unlock_mutex;
 266                 }
 267         }
 268
 269         /*
 270          * Make sure ocfs2_reserve_cluster is called after
 271          * __ocfs2_flush_truncate_log, otherwise, dead lock may happen.
 272          *
 273          * If ocfs2_reserve_cluster is called
 274          * before __ocfs2_flush_truncate_log, dead lock on global bitmap
 275          * may happen.
 276          *
 277          */
 278         ret = ocfs2_reserve_clusters(osb, *len, &context->data_ac);
 279         if (ret) {
 280                 mlog_errno(ret);
 281                 goto out_unlock_mutex;
 282         }
 283
 284         handle = ocfs2_start_trans(osb, credits);
 285         if (IS_ERR(handle)) {
 286                 ret = PTR_ERR(handle);
 287                 mlog_errno(ret);
 288                 goto out_unlock_mutex;
 289         }
 290
 291         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 292                                      &new_phys_cpos, &new_len);
 293         if (ret) {
 294                 mlog_errno(ret);
 295                 goto out_commit;
 296         }
 297
 298         /*
 299          * allowing partial extent moving is kind of 'pros and cons', it makes
 300          * whole defragmentation less likely to fail, on the contrary, the bad
 301          * thing is it may make the fs even more fragmented after moving, let
 302          * userspace make a good decision here.
 303          */
 304         if (new_len != *len) {
 305                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 306                 if (!partial) {
 307                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 308                         ret = -ENOSPC;
 309                         need_free = 1;
 310                         goto out_commit;
 311                 }
 312         }
 313
 314         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 315              phys_cpos, new_phys_cpos);
 316
 317         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 318                                   new_phys_cpos, ext_flags);
 319         if (ret)
 320                 mlog_errno(ret);
 321
 322         if (partial && (new_len != *len))
 323                 *len = new_len;
 324
 325         /*
 326          * Here we should write the new page out first if we are
 327          * in write-back mode.
 328          */
 329         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 330         if (ret)
 331                 mlog_errno(ret);
 332
 333 out_commit:
 334         if (need_free && context->data_ac) {
 335                 struct ocfs2_alloc_context *data_ac = context->data_ac;
 336
 337                 if (context->data_ac->ac_which == OCFS2_AC_USE_LOCAL)
 338                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
 339                                         new_phys_cpos, new_len);
 340                 else
 341                         ocfs2_free_clusters(handle,
 342                                         data_ac->ac_inode,
 343                                         data_ac->ac_bh,
 344                                         ocfs2_clusters_to_blocks(osb->sb, new_phys_cpos),
 345                                         new_len);
 346         }
 347
 348         ocfs2_commit_trans(osb, handle);
 349
 350 out_unlock_mutex:
 351         inode_unlock(tl_inode);
 352
 353         if (context->data_ac) {
 354                 ocfs2_free_alloc_context(context->data_ac);
 355                 context->data_ac = NULL;
 356         }
 357
 358         if (context->meta_ac) {
 359                 ocfs2_free_alloc_context(context->meta_ac);
 360                 context->meta_ac = NULL;
 361         }
 362
 363 out:
 364         if (ref_tree)
 365                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 366
 367         return ret;
 368 }
 369
 370 /*
 371  * find the victim alloc group, where #blkno fits.
 372  */
 373 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 374                                          u64 vict_blkno,
 375                                          int type, int slot,
 376                                          int *vict_bit,
 377                                          struct buffer_head **ret_bh)
 378 {
 379         int ret, i, bits_per_unit = 0;
 380         u64 blkno;
 381         char namebuf[40];
 382
 383         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 384         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 385         struct ocfs2_chain_list *cl;
 386         struct ocfs2_chain_rec *rec;
 387         struct ocfs2_dinode *ac_dinode;
 388         struct ocfs2_group_desc *bg;
 389
 390         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 391         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 392                                          strlen(namebuf), &blkno);
 393         if (ret) {
 394                 ret = -ENOENT;
 395                 goto out;
 396         }
 397
 398         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 399         if (ret) {
 400                 mlog_errno(ret);
 401                 goto out;
 402         }
 403
 404         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 405         cl = &(ac_dinode->id2.i_chain);
 406         rec = &(cl->cl_recs[0]);
 407
 408         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 409                 bits_per_unit = osb->s_clustersize_bits -
 410                                         inode->i_sb->s_blocksize_bits;
 411         /*
 412          * 'vict_blkno' was out of the valid range.
 413          */
 414         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 415             (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 416                                 bits_per_unit))) {
 417                 ret = -EINVAL;
 418                 goto out;
 419         }
 420
 421         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 422
 423                 rec = &(cl->cl_recs[i]);
 424                 if (!rec)
 425                         continue;
 426
 427                 bg = NULL;
 428
 429                 do {
 430                         if (!bg)
 431                                 blkno = le64_to_cpu(rec->c_blkno);
 432                         else
 433                                 blkno = le64_to_cpu(bg->bg_next_group);
 434
 435                         if (gd_bh) {
 436                                 brelse(gd_bh);
 437                                 gd_bh = NULL;
 438                         }
 439
 440                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 441                         if (ret) {
 442                                 mlog_errno(ret);
 443                                 goto out;
 444                         }
 445
 446                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 447
 448                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 449                                                 le16_to_cpu(bg->bg_bits))) {
 450
 451                                 *ret_bh = gd_bh;
 452                                 *vict_bit = (vict_blkno - blkno) >>
 453                                                         bits_per_unit;
 454                                 mlog(0, "find the victim group: #%llu, "
 455                                      "total_bits: %u, vict_bit: %u\n",
 456                                      blkno, le16_to_cpu(bg->bg_bits),
 457                                      *vict_bit);
 458                                 goto out;
 459                         }
 460
 461                 } while (le64_to_cpu(bg->bg_next_group));
 462         }
 463
 464         ret = -EINVAL;
 465 out:
 466         brelse(ac_bh);
 467
 468         /*
 469          * caller has to release the gd_bh properly.
 470          */
 471         return ret;
 472 }
 473
 474 /*
 475  * XXX: helper to validate and adjust moving goal.
 476  */
 477 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 478                                                struct ocfs2_move_extents *range)
 479 {
 480         int ret, goal_bit = 0;
 481
 482         struct buffer_head *gd_bh = NULL;
 483         struct ocfs2_group_desc *bg;
 484         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 485         int c_to_b = 1 << (osb->s_clustersize_bits -
 486                                         inode->i_sb->s_blocksize_bits);
 487
 488         /*
 489          * make goal become cluster aligned.
 490          */
 491         range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 492                                                       range->me_goal);
 493         /*
 494          * validate goal sits within global_bitmap, and return the victim
 495          * group desc
 496          */
 497         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 498                                             GLOBAL_BITMAP_SYSTEM_INODE,
 499                                             OCFS2_INVALID_SLOT,
 500                                             &goal_bit, &gd_bh);
 501         if (ret)
 502                 goto out;
 503
 504         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 505
 506         /*
 507          * moving goal is not allowd to start with a group desc blok(#0 blk)
 508          * let's compromise to the latter cluster.
 509          */
 510         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 511                 range->me_goal += c_to_b;
 512
 513         /*
 514          * movement is not gonna cross two groups.
 515          */
 516         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 517                                                                 range->me_len) {
 518                 ret = -EINVAL;
 519                 goto out;
 520         }
 521         /*
 522          * more exact validations/adjustments will be performed later during
 523          * moving operation for each extent range.
 524          */
 525         mlog(0, "extents get ready to be moved to #%llu block\n",
 526              range->me_goal);
 527
 528 out:
 529         brelse(gd_bh);
 530
 531         return ret;
 532 }
 533
 534 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 535                                     int *goal_bit, u32 move_len, u32 max_hop,
 536                                     u32 *phys_cpos)
 537 {
 538         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 539         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 540         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 541                                                  le64_to_cpu(gd->bg_blkno));
 542
 543         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 544
 545                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 546                 if (used) {
 547                         /*
 548                          * we even tried searching the free chunk by jumping
 549                          * a 'max_hop' distance, but still failed.
 550                          */
 551                         if ((i - base_bit) > max_hop) {
 552                                 *phys_cpos = 0;
 553                                 break;
 554                         }
 555
 556                         if (last_free_bits)
 557                                 last_free_bits = 0;
 558
 559                         continue;
 560                 } else
 561                         last_free_bits++;
 562
 563                 if (last_free_bits == move_len) {
 564                         *goal_bit = i;
 565                         *phys_cpos = base_cpos + i;
 566                         break;
 567                 }
 568         }
 569
 570         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 571 }
 572
 573 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 574                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 575                              u32 len, int ext_flags)
 576 {
 577         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 578         handle_t *handle;
 579         struct inode *inode = context->inode;
 580         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 581         struct inode *tl_inode = osb->osb_tl_inode;
 582         struct inode *gb_inode = NULL;
 583         struct buffer_head *gb_bh = NULL;
 584         struct buffer_head *gd_bh = NULL;
 585         struct ocfs2_group_desc *gd;
 586         struct ocfs2_refcount_tree *ref_tree = NULL;
 587         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 588                                                     context->range->me_threshold);
 589         u64 phys_blkno, new_phys_blkno;
 590
 591         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 592
 593         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 594                 BUG_ON(!ocfs2_is_refcount_inode(inode));
 595                 BUG_ON(!context->refcount_loc);
 596
 597                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 598                                                &ref_tree, NULL);
 599                 if (ret) {
 600                         mlog_errno(ret);
 601                         return ret;
 602                 }
 603
 604                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 605                                                         context->refcount_loc,
 606                                                         phys_blkno,
 607                                                         len,
 608                                                         &credits,
 609                                                         &extra_blocks);
 610                 if (ret) {
 611                         mlog_errno(ret);
 612                         goto out;
 613                 }
 614         }
 615
 616         ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 617                                                 len, 1,
 618                                                 &context->meta_ac,
 619                                                 extra_blocks, &credits);
 620         if (ret) {
 621                 mlog_errno(ret);
 622                 goto out;
 623         }
 624
 625         /*
 626          * need to count 2 extra credits for global_bitmap inode and
 627          * group descriptor.
 628          */
 629         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 630
 631         /*
 632          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 633          * logic, while we still need to lock the global_bitmap.
 634          */
 635         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 636                                                OCFS2_INVALID_SLOT);
 637         if (!gb_inode) {
 638                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 639                 ret = -EIO;
 640                 goto out;
 641         }
 642
 643         inode_lock(gb_inode);
 644
 645         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 646         if (ret) {
 647                 mlog_errno(ret);
 648                 goto out_unlock_gb_mutex;
 649         }
 650
 651         inode_lock(tl_inode);
 652
 653         handle = ocfs2_start_trans(osb, credits);
 654         if (IS_ERR(handle)) {
 655                 ret = PTR_ERR(handle);
 656                 mlog_errno(ret);
 657                 goto out_unlock_tl_inode;
 658         }
 659
 660         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 661         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 662                                             GLOBAL_BITMAP_SYSTEM_INODE,
 663                                             OCFS2_INVALID_SLOT,
 664                                             &goal_bit, &gd_bh);
 665         if (ret) {
 666                 mlog_errno(ret);
 667                 goto out_commit;
 668         }
 669
 670         /*
 671          * probe the victim cluster group to find a proper
 672          * region to fit wanted movement, it even will perfrom
 673          * a best-effort attempt by compromising to a threshold
 674          * around the goal.
 675          */
 676         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 677                                 new_phys_cpos);
 678         if (!*new_phys_cpos) {
 679                 ret = -ENOSPC;
 680                 goto out_commit;
 681         }
 682
 683         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 684                                   *new_phys_cpos, ext_flags);
 685         if (ret) {
 686                 mlog_errno(ret);
 687                 goto out_commit;
 688         }
 689
 690         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 691         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 692                                                le16_to_cpu(gd->bg_chain));
 693         if (ret) {
 694                 mlog_errno(ret);
 695                 goto out_commit;
 696         }
 697
 698         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 699                                          goal_bit, len);
 700         if (ret) {
 701                 ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 702                                                le16_to_cpu(gd->bg_chain));
 703                 mlog_errno(ret);
 704         }
 705
 706         /*
 707          * Here we should write the new page out first if we are
 708          * in write-back mode.
 709          */
 710         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 711         if (ret)
 712                 mlog_errno(ret);
 713
 714 out_commit:
 715         ocfs2_commit_trans(osb, handle);
 716         brelse(gd_bh);
 717
 718 out_unlock_tl_inode:
 719         inode_unlock(tl_inode);
 720
 721         ocfs2_inode_unlock(gb_inode, 1);
 722 out_unlock_gb_mutex:
 723         inode_unlock(gb_inode);
 724         brelse(gb_bh);
 725         iput(gb_inode);
 726
 727 out:
 728         if (context->meta_ac) {
 729                 ocfs2_free_alloc_context(context->meta_ac);
 730                 context->meta_ac = NULL;
 731         }
 732
 733         if (ref_tree)
 734                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 735
 736         return ret;
 737 }
 738
 739 /*
 740  * Helper to calculate the defraging length in one run according to threshold.
 741  */
 742 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 743                                          u32 threshold, int *skip)
 744 {
 745         if ((*alloc_size + *len_defraged) < threshold) {
 746                 /*
 747                  * proceed defragmentation until we meet the thresh
 748                  */
 749                 *len_defraged += *alloc_size;
 750         } else if (*len_defraged == 0) {
 751                 /*
 752                  * XXX: skip a large extent.
 753                  */
 754                 *skip = 1;
 755         } else {
 756                 /*
 757                  * split this extent to coalesce with former pieces as
 758                  * to reach the threshold.
 759                  *
 760                  * we're done here with one cycle of defragmentation
 761                  * in a size of 'thresh', resetting 'len_defraged'
 762                  * forces a new defragmentation.
 763                  */
 764                 *alloc_size = threshold - *len_defraged;
 765                 *len_defraged = 0;
 766         }
 767 }
 768
 769 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 770                                 struct ocfs2_move_extents_context *context)
 771 {
 772         int ret = 0, flags, do_defrag, skip = 0;
 773         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 774         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 775
 776         struct inode *inode = context->inode;
 777         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 778         struct ocfs2_move_extents *range = context->range;
 779         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 780
 781         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 782                 return 0;
 783
 784         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 785                 return 0;
 786
 787         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 788
 789         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 790         ocfs2_init_dealloc_ctxt(&context->dealloc);
 791
 792         /*
 793          * TO-DO XXX:
 794          *
 795          * - xattr extents.
 796          */
 797
 798         do_defrag = context->auto_defrag;
 799
 800         /*
 801          * extents moving happens in unit of clusters, for the sake
 802          * of simplicity, we may ignore two clusters where 'byte_start'
 803          * and 'byte_start + len' were within.
 804          */
 805         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 806         len_to_move = (range->me_start + range->me_len) >>
 807                                                 osb->s_clustersize_bits;
 808         if (len_to_move >= move_start)
 809                 len_to_move -= move_start;
 810         else
 811                 len_to_move = 0;
 812
 813         if (do_defrag) {
 814                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 815                 if (defrag_thresh <= 1)
 816                         goto done;
 817         } else
 818                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 819                                                          range->me_goal);
 820
 821         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 822              "thresh: %u\n",
 823              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 824              (unsigned long long)range->me_start,
 825              (unsigned long long)range->me_len,
 826              move_start, len_to_move, defrag_thresh);
 827
 828         cpos = move_start;
 829         while (len_to_move) {
 830                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 831                                          &flags);
 832                 if (ret) {
 833                         mlog_errno(ret);
 834                         goto out;
 835                 }
 836
 837                 if (alloc_size > len_to_move)
 838                         alloc_size = len_to_move;
 839
 840                 /*
 841                  * XXX: how to deal with a hole:
 842                  *
 843                  * - skip the hole of course
 844                  * - force a new defragmentation
 845                  */
 846                 if (!phys_cpos) {
 847                         if (do_defrag)
 848                                 len_defraged = 0;
 849
 850                         goto next;
 851                 }
 852
 853                 if (do_defrag) {
 854                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 855                                                      defrag_thresh, &skip);
 856                         /*
 857                          * skip large extents
 858                          */
 859                         if (skip) {
 860                                 skip = 0;
 861                                 goto next;
 862                         }
 863
 864                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 865                              "alloc_size: %u, len_defraged: %u\n",
 866                              cpos, phys_cpos, alloc_size, len_defraged);
 867
 868                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 869                                                   &alloc_size, flags);
 870                 } else {
 871                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 872                                                 &new_phys_cpos, alloc_size,
 873                                                 flags);
 874
 875                         new_phys_cpos += alloc_size;
 876                 }
 877
 878                 if (ret < 0) {
 879                         mlog_errno(ret);
 880                         goto out;
 881                 }
 882
 883                 context->clusters_moved += alloc_size;
 884 next:
 885                 cpos += alloc_size;
 886                 len_to_move -= alloc_size;
 887         }
 888
 889 done:
 890         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 891
 892 out:
 893         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 894                                                       context->clusters_moved);
 895         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 896                                                        context->new_phys_cpos);
 897
 898         ocfs2_schedule_truncate_log_flush(osb, 1);
 899         ocfs2_run_deallocs(osb, &context->dealloc);
 900
 901         return ret;
 902 }
 903
 904 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 905 {
 906         int status;
 907         handle_t *handle;
 908         struct inode *inode = context->inode;
 909         struct ocfs2_dinode *di;
 910         struct buffer_head *di_bh = NULL;
 911         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 912
 913         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 914                 return -EROFS;
 915
 916         inode_lock(inode);
 917
 918         /*
 919          * This prevents concurrent writes from other nodes
 920          */
 921         status = ocfs2_rw_lock(inode, 1);
 922         if (status) {
 923                 mlog_errno(status);
 924                 goto out;
 925         }
 926
 927         status = ocfs2_inode_lock(inode, &di_bh, 1);
 928         if (status) {
 929                 mlog_errno(status);
 930                 goto out_rw_unlock;
 931         }
 932
 933         /*
 934          * rememer ip_xattr_sem also needs to be held if necessary
 935          */
 936         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 937
 938         status = __ocfs2_move_extents_range(di_bh, context);
 939
 940         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 941         if (status) {
 942                 mlog_errno(status);
 943                 goto out_inode_unlock;
 944         }
 945
 946         /*
 947          * We update ctime for these changes
 948          */
 949         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 950         if (IS_ERR(handle)) {
 951                 status = PTR_ERR(handle);
 952                 mlog_errno(status);
 953                 goto out_inode_unlock;
 954         }
 955
 956         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 957                                          OCFS2_JOURNAL_ACCESS_WRITE);
 958         if (status) {
 959                 mlog_errno(status);
 960                 goto out_commit;
 961         }
 962
 963         di = (struct ocfs2_dinode *)di_bh->b_data;
 964         inode->i_ctime = current_time(inode);
 965         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 966         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 967         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 968
 969         ocfs2_journal_dirty(handle, di_bh);
 970
 971 out_commit:
 972         ocfs2_commit_trans(osb, handle);
 973
 974 out_inode_unlock:
 975         brelse(di_bh);
 976         ocfs2_inode_unlock(inode, 1);
 977 out_rw_unlock:
 978         ocfs2_rw_unlock(inode, 1);
 979 out:
 980         inode_unlock(inode);
 981
 982         return status;
 983 }
 984
 985 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 986 {
 987         int status;
 988
 989         struct inode *inode = file_inode(filp);
 990         struct ocfs2_move_extents range;
 991         struct ocfs2_move_extents_context *context;
 992
 993         if (!argp)
 994                 return -EINVAL;
 995
 996         status = mnt_want_write_file(filp);
 997         if (status)
 998                 return status;
 999
1000         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
1001                 status = -EPERM;
1002                 goto out_drop;
1003         }
1004
1005         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1006                 status = -EPERM;
1007                 goto out_drop;
1008         }
1009
1010         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1011         if (!context) {
1012                 status = -ENOMEM;
1013                 mlog_errno(status);
1014                 goto out_drop;
1015         }
1016
1017         context->inode = inode;
1018         context->file = filp;
1019
1020         if (copy_from_user(&range, argp, sizeof(range))) {
1021                 status = -EFAULT;
1022                 goto out_free;
1023         }
1024
1025         if (range.me_start > i_size_read(inode)) {
1026                 status = -EINVAL;
1027                 goto out_free;
1028         }
1029
1030         if (range.me_start + range.me_len > i_size_read(inode))
1031                         range.me_len = i_size_read(inode) - range.me_start;
1032
1033         context->range = &range;
1034
1035         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1036                 context->auto_defrag = 1;
1037                 /*
1038                  * ok, the default theshold for the defragmentation
1039                  * is 1M, since our maximum clustersize was 1M also.
1040                  * any thought?
1041                  */
1042                 if (!range.me_threshold)
1043                         range.me_threshold = 1024 * 1024;
1044
1045                 if (range.me_threshold > i_size_read(inode))
1046                         range.me_threshold = i_size_read(inode);
1047
1048                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1049                         context->partial = 1;
1050         } else {
1051                 /*
1052                  * first best-effort attempt to validate and adjust the goal
1053                  * (physical address in block), while it can't guarantee later
1054                  * operation can succeed all the time since global_bitmap may
1055                  * change a bit over time.
1056                  */
1057
1058                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1059                 if (status)
1060                         goto out_copy;
1061         }
1062
1063         status = ocfs2_move_extents(context);
1064         if (status)
1065                 mlog_errno(status);
1066 out_copy:
1067         /*
1068          * movement/defragmentation may end up being partially completed,
1069          * that's the reason why we need to return userspace the finished
1070          * length and new_offset even if failure happens somewhere.
1071          */
1072         if (copy_to_user(argp, &range, sizeof(range)))
1073                 status = -EFAULT;
1074
1075 out_free:
1076         kfree(context);
1077 out_drop:
1078         mnt_drop_write_file(filp);
1079
1080         return status;
1081 }