fs/ocfs2/move_extents.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * move_extents.c
   5  *
   6  * Copyright (C) 2011 Oracle.  All rights reserved.
   7  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public
  10  * License version 2 as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License for more details.
  16  */
  17 #include <linux/fs.h>
  18 #include <linux/types.h>
  19 #include <linux/mount.h>
  20 #include <linux/swap.h>
  21
  22 #include <cluster/masklog.h>
  23
  24 #include "ocfs2.h"
  25 #include "ocfs2_ioctl.h"
  26
  27 #include "alloc.h"
  28 #include "aops.h"
  29 #include "dlmglue.h"
  30 #include "extent_map.h"
  31 #include "inode.h"
  32 #include "journal.h"
  33 #include "suballoc.h"
  34 #include "uptodate.h"
  35 #include "super.h"
  36 #include "dir.h"
  37 #include "buffer_head_io.h"
  38 #include "sysfile.h"
  39 #include "refcounttree.h"
  40 #include "move_extents.h"
  41
  42 struct ocfs2_move_extents_context {
  43         struct inode *inode;
  44         struct file *file;
  45         int auto_defrag;
  46         int partial;
  47         int credits;
  48         u32 new_phys_cpos;
  49         u32 clusters_moved;
  50         u64 refcount_loc;
  51         struct ocfs2_move_extents *range;
  52         struct ocfs2_extent_tree et;
  53         struct ocfs2_alloc_context *meta_ac;
  54         struct ocfs2_alloc_context *data_ac;
  55         struct ocfs2_cached_dealloc_ctxt dealloc;
  56 };
  57
  58 static int __ocfs2_move_extent(handle_t *handle,
  59                                struct ocfs2_move_extents_context *context,
  60                                u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
  61                                int ext_flags)
  62 {
  63         int ret = 0, index;
  64         struct inode *inode = context->inode;
  65         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  66         struct ocfs2_extent_rec *rec, replace_rec;
  67         struct ocfs2_path *path = NULL;
  68         struct ocfs2_extent_list *el;
  69         u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
  70         u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
  71
  72         ret = ocfs2_duplicate_clusters_by_page(handle, inode, cpos,
  73                                                p_cpos, new_p_cpos, len);
  74         if (ret) {
  75                 mlog_errno(ret);
  76                 goto out;
  77         }
  78
  79         memset(&replace_rec, 0, sizeof(replace_rec));
  80         replace_rec.e_cpos = cpu_to_le32(cpos);
  81         replace_rec.e_leaf_clusters = cpu_to_le16(len);
  82         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
  83                                                                    new_p_cpos));
  84
  85         path = ocfs2_new_path_from_et(&context->et);
  86         if (!path) {
  87                 ret = -ENOMEM;
  88                 mlog_errno(ret);
  89                 goto out;
  90         }
  91
  92         ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
  93         if (ret) {
  94                 mlog_errno(ret);
  95                 goto out;
  96         }
  97
  98         el = path_leaf_el(path);
  99
 100         index = ocfs2_search_extent_list(el, cpos);
 101         if (index == -1 || 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         ocfs2_free_path(path);
 156         return ret;
 157 }
 158
 159 /*
 160  * lock allocators, and reserving appropriate number of bits for
 161  * meta blocks and data clusters.
 162  *
 163  * in some cases, we don't need to reserve clusters, just let data_ac
 164  * be NULL.
 165  */
 166 static int ocfs2_lock_allocators_move_extents(struct inode *inode,
 167                                         struct ocfs2_extent_tree *et,
 168                                         u32 clusters_to_move,
 169                                         u32 extents_to_split,
 170                                         struct ocfs2_alloc_context **meta_ac,
 171                                         struct ocfs2_alloc_context **data_ac,
 172                                         int extra_blocks,
 173                                         int *credits)
 174 {
 175         int ret, num_free_extents;
 176         unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 177         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 178
 179         num_free_extents = ocfs2_num_free_extents(osb, et);
 180         if (num_free_extents < 0) {
 181                 ret = num_free_extents;
 182                 mlog_errno(ret);
 183                 goto out;
 184         }
 185
 186         if (!num_free_extents ||
 187             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 188                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 189
 190         ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 191         if (ret) {
 192                 mlog_errno(ret);
 193                 goto out;
 194         }
 195
 196         if (data_ac) {
 197                 ret = ocfs2_reserve_clusters(osb, clusters_to_move, data_ac);
 198                 if (ret) {
 199                         mlog_errno(ret);
 200                         goto out;
 201                 }
 202         }
 203
 204         *credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 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;
 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          * validate goal sits within global_bitmap, and return the victim
 486          * group desc
 487          */
 488         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 489                                             GLOBAL_BITMAP_SYSTEM_INODE,
 490                                             OCFS2_INVALID_SLOT,
 491                                             &goal_bit, &gd_bh);
 492         if (ret)
 493                 goto out;
 494
 495         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 496
 497         /*
 498          * moving goal is not allowd to start with a group desc blok(#0 blk)
 499          * let's compromise to the latter cluster.
 500          */
 501         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 502                 range->me_goal += c_to_b;
 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_move_extent(struct ocfs2_move_extents_context *context,
 565                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 566                              u32 len, int ext_flags)
 567 {
 568         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 569         handle_t *handle;
 570         struct inode *inode = context->inode;
 571         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 572         struct inode *tl_inode = osb->osb_tl_inode;
 573         struct inode *gb_inode = NULL;
 574         struct buffer_head *gb_bh = NULL;
 575         struct buffer_head *gd_bh = NULL;
 576         struct ocfs2_group_desc *gd;
 577         struct ocfs2_refcount_tree *ref_tree = NULL;
 578         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 579                                                     context->range->me_threshold);
 580         u64 phys_blkno, new_phys_blkno;
 581
 582         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 583
 584         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 585
 586                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 587                          OCFS2_HAS_REFCOUNT_FL));
 588
 589                 BUG_ON(!context->refcount_loc);
 590
 591                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 592                                                &ref_tree, NULL);
 593                 if (ret) {
 594                         mlog_errno(ret);
 595                         return ret;
 596                 }
 597
 598                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 599                                                         context->refcount_loc,
 600                                                         phys_blkno,
 601                                                         len,
 602                                                         &credits,
 603                                                         &extra_blocks);
 604                 if (ret) {
 605                         mlog_errno(ret);
 606                         goto out;
 607                 }
 608         }
 609
 610         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, len, 1,
 611                                                  &context->meta_ac,
 612                                                  NULL, extra_blocks, &credits);
 613         if (ret) {
 614                 mlog_errno(ret);
 615                 goto out;
 616         }
 617
 618         /*
 619          * need to count 2 extra credits for global_bitmap inode and
 620          * group descriptor.
 621          */
 622         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 623
 624         /*
 625          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 626          * logic, while we still need to lock the global_bitmap.
 627          */
 628         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 629                                                OCFS2_INVALID_SLOT);
 630         if (!gb_inode) {
 631                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 632                 ret = -EIO;
 633                 goto out;
 634         }
 635
 636         mutex_lock(&gb_inode->i_mutex);
 637
 638         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 639         if (ret) {
 640                 mlog_errno(ret);
 641                 goto out_unlock_gb_mutex;
 642         }
 643
 644         mutex_lock(&tl_inode->i_mutex);
 645
 646         handle = ocfs2_start_trans(osb, credits);
 647         if (IS_ERR(handle)) {
 648                 ret = PTR_ERR(handle);
 649                 mlog_errno(ret);
 650                 goto out_unlock_tl_inode;
 651         }
 652
 653         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 654         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 655                                             GLOBAL_BITMAP_SYSTEM_INODE,
 656                                             OCFS2_INVALID_SLOT,
 657                                             &goal_bit, &gd_bh);
 658         if (ret) {
 659                 mlog_errno(ret);
 660                 goto out_commit;
 661         }
 662
 663         /*
 664          * probe the victim cluster group to find a proper
 665          * region to fit wanted movement, it even will perfrom
 666          * a best-effort attempt by compromising to a threshold
 667          * around the goal.
 668          */
 669         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 670                                 new_phys_cpos);
 671         if (!*new_phys_cpos) {
 672                 ret = -ENOSPC;
 673                 goto out_commit;
 674         }
 675
 676         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 677                                   *new_phys_cpos, ext_flags);
 678         if (ret) {
 679                 mlog_errno(ret);
 680                 goto out_commit;
 681         }
 682
 683         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 684         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 685                                                le16_to_cpu(gd->bg_chain));
 686         if (ret) {
 687                 mlog_errno(ret);
 688                 goto out_commit;
 689         }
 690
 691         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 692                                          goal_bit, len);
 693         if (ret)
 694                 mlog_errno(ret);
 695
 696         /*
 697          * Here we should write the new page out first if we are
 698          * in write-back mode.
 699          */
 700         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 701         if (ret)
 702                 mlog_errno(ret);
 703
 704 out_commit:
 705         ocfs2_commit_trans(osb, handle);
 706         brelse(gd_bh);
 707
 708 out_unlock_tl_inode:
 709         mutex_unlock(&tl_inode->i_mutex);
 710
 711         ocfs2_inode_unlock(gb_inode, 1);
 712 out_unlock_gb_mutex:
 713         mutex_unlock(&gb_inode->i_mutex);
 714         brelse(gb_bh);
 715         iput(gb_inode);
 716
 717 out:
 718         if (context->meta_ac) {
 719                 ocfs2_free_alloc_context(context->meta_ac);
 720                 context->meta_ac = NULL;
 721         }
 722
 723         if (ref_tree)
 724                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 725
 726         return ret;
 727 }
 728
 729 /*
 730  * Helper to calculate the defraging length in one run according to threshold.
 731  */
 732 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 733                                          u32 threshold, int *skip)
 734 {
 735         if ((*alloc_size + *len_defraged) < threshold) {
 736                 /*
 737                  * proceed defragmentation until we meet the thresh
 738                  */
 739                 *len_defraged += *alloc_size;
 740         } else if (*len_defraged == 0) {
 741                 /*
 742                  * XXX: skip a large extent.
 743                  */
 744                 *skip = 1;
 745         } else {
 746                 /*
 747                  * split this extent to coalesce with former pieces as
 748                  * to reach the threshold.
 749                  *
 750                  * we're done here with one cycle of defragmentation
 751                  * in a size of 'thresh', resetting 'len_defraged'
 752                  * forces a new defragmentation.
 753                  */
 754                 *alloc_size = threshold - *len_defraged;
 755                 *len_defraged = 0;
 756         }
 757 }
 758
 759 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 760                                 struct ocfs2_move_extents_context *context)
 761 {
 762         int ret = 0, flags, do_defrag, skip = 0;
 763         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 764         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 765
 766         struct inode *inode = context->inode;
 767         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 768         struct ocfs2_move_extents *range = context->range;
 769         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 770
 771         if ((i_size_read(inode) == 0) || (range->me_len == 0))
 772                 return 0;
 773
 774         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 775                 return 0;
 776
 777         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 778
 779         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 780         ocfs2_init_dealloc_ctxt(&context->dealloc);
 781
 782         /*
 783          * TO-DO XXX:
 784          *
 785          * - xattr extents.
 786          */
 787
 788         do_defrag = context->auto_defrag;
 789
 790         /*
 791          * extents moving happens in unit of clusters, for the sake
 792          * of simplicity, we may ignore two clusters where 'byte_start'
 793          * and 'byte_start + len' were within.
 794          */
 795         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 796         len_to_move = (range->me_start + range->me_len) >>
 797                                                 osb->s_clustersize_bits;
 798         if (len_to_move >= move_start)
 799                 len_to_move -= move_start;
 800         else
 801                 len_to_move = 0;
 802
 803         if (do_defrag) {
 804                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 805                 if (defrag_thresh <= 1)
 806                         goto done;
 807         } else
 808                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 809                                                          range->me_goal);
 810
 811         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 812              "thresh: %u\n",
 813              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 814              (unsigned long long)range->me_start,
 815              (unsigned long long)range->me_len,
 816              move_start, len_to_move, defrag_thresh);
 817
 818         cpos = move_start;
 819         while (len_to_move) {
 820                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 821                                          &flags);
 822                 if (ret) {
 823                         mlog_errno(ret);
 824                         goto out;
 825                 }
 826
 827                 if (alloc_size > len_to_move)
 828                         alloc_size = len_to_move;
 829
 830                 /*
 831                  * XXX: how to deal with a hole:
 832                  *
 833                  * - skip the hole of course
 834                  * - force a new defragmentation
 835                  */
 836                 if (!phys_cpos) {
 837                         if (do_defrag)
 838                                 len_defraged = 0;
 839
 840                         goto next;
 841                 }
 842
 843                 if (do_defrag) {
 844                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 845                                                      defrag_thresh, &skip);
 846                         /*
 847                          * skip large extents
 848                          */
 849                         if (skip) {
 850                                 skip = 0;
 851                                 goto next;
 852                         }
 853
 854                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 855                              "alloc_size: %u, len_defraged: %u\n",
 856                              cpos, phys_cpos, alloc_size, len_defraged);
 857
 858                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 859                                                   &alloc_size, flags);
 860                 } else {
 861                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 862                                                 &new_phys_cpos, alloc_size,
 863                                                 flags);
 864
 865                         new_phys_cpos += alloc_size;
 866                 }
 867
 868                 if (ret < 0) {
 869                         mlog_errno(ret);
 870                         goto out;
 871                 }
 872
 873                 context->clusters_moved += alloc_size;
 874 next:
 875                 cpos += alloc_size;
 876                 len_to_move -= alloc_size;
 877         }
 878
 879 done:
 880         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 881
 882 out:
 883         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 884                                                       context->clusters_moved);
 885         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 886                                                        context->new_phys_cpos);
 887
 888         ocfs2_schedule_truncate_log_flush(osb, 1);
 889         ocfs2_run_deallocs(osb, &context->dealloc);
 890
 891         return ret;
 892 }
 893
 894 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 895 {
 896         int status;
 897         handle_t *handle;
 898         struct inode *inode = context->inode;
 899         struct ocfs2_dinode *di;
 900         struct buffer_head *di_bh = NULL;
 901         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 902
 903         if (!inode)
 904                 return -ENOENT;
 905
 906         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 907                 return -EROFS;
 908
 909         mutex_lock(&inode->i_mutex);
 910
 911         /*
 912          * This prevents concurrent writes from other nodes
 913          */
 914         status = ocfs2_rw_lock(inode, 1);
 915         if (status) {
 916                 mlog_errno(status);
 917                 goto out;
 918         }
 919
 920         status = ocfs2_inode_lock(inode, &di_bh, 1);
 921         if (status) {
 922                 mlog_errno(status);
 923                 goto out_rw_unlock;
 924         }
 925
 926         /*
 927          * rememer ip_xattr_sem also needs to be held if necessary
 928          */
 929         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 930
 931         status = __ocfs2_move_extents_range(di_bh, context);
 932
 933         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 934         if (status) {
 935                 mlog_errno(status);
 936                 goto out_inode_unlock;
 937         }
 938
 939         /*
 940          * We update ctime for these changes
 941          */
 942         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 943         if (IS_ERR(handle)) {
 944                 status = PTR_ERR(handle);
 945                 mlog_errno(status);
 946                 goto out_inode_unlock;
 947         }
 948
 949         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 950                                          OCFS2_JOURNAL_ACCESS_WRITE);
 951         if (status) {
 952                 mlog_errno(status);
 953                 goto out_commit;
 954         }
 955
 956         di = (struct ocfs2_dinode *)di_bh->b_data;
 957         inode->i_ctime = CURRENT_TIME;
 958         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
 959         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 960
 961         ocfs2_journal_dirty(handle, di_bh);
 962
 963 out_commit:
 964         ocfs2_commit_trans(osb, handle);
 965
 966 out_inode_unlock:
 967         brelse(di_bh);
 968         ocfs2_inode_unlock(inode, 1);
 969 out_rw_unlock:
 970         ocfs2_rw_unlock(inode, 1);
 971 out:
 972         mutex_unlock(&inode->i_mutex);
 973
 974         return status;
 975 }
 976
 977 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 978 {
 979         int status;
 980
 981         struct inode *inode = file_inode(filp);
 982         struct ocfs2_move_extents range;
 983         struct ocfs2_move_extents_context *context;
 984
 985         if (!argp)
 986                 return -EINVAL;
 987
 988         status = mnt_want_write_file(filp);
 989         if (status)
 990                 return status;
 991
 992         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
 993                 status = -EPERM;
 994                 goto out_drop;
 995         }
 996
 997         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
 998                 status = -EPERM;
 999                 goto out_drop;
1000         }
1001
1002         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1003         if (!context) {
1004                 status = -ENOMEM;
1005                 mlog_errno(status);
1006                 goto out_drop;
1007         }
1008
1009         context->inode = inode;
1010         context->file = filp;
1011
1012         if (copy_from_user(&range, argp, sizeof(range))) {
1013                 status = -EFAULT;
1014                 goto out_free;
1015         }
1016
1017         if (range.me_start > i_size_read(inode)) {
1018                 status = -EINVAL;
1019                 goto out_free;
1020         }
1021
1022         if (range.me_start + range.me_len > i_size_read(inode))
1023                         range.me_len = i_size_read(inode) - range.me_start;
1024
1025         context->range = &range;
1026
1027         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1028                 context->auto_defrag = 1;
1029                 /*
1030                  * ok, the default theshold for the defragmentation
1031                  * is 1M, since our maximum clustersize was 1M also.
1032                  * any thought?
1033                  */
1034                 if (!range.me_threshold)
1035                         range.me_threshold = 1024 * 1024;
1036
1037                 if (range.me_threshold > i_size_read(inode))
1038                         range.me_threshold = i_size_read(inode);
1039
1040                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1041                         context->partial = 1;
1042         } else {
1043                 /*
1044                  * first best-effort attempt to validate and adjust the goal
1045                  * (physical address in block), while it can't guarantee later
1046                  * operation can succeed all the time since global_bitmap may
1047                  * change a bit over time.
1048                  */
1049
1050                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1051                 if (status)
1052                         goto out_copy;
1053         }
1054
1055         status = ocfs2_move_extents(context);
1056         if (status)
1057                 mlog_errno(status);
1058 out_copy:
1059         /*
1060          * movement/defragmentation may end up being partially completed,
1061          * that's the reason why we need to return userspace the finished
1062          * length and new_offset even if failure happens somewhere.
1063          */
1064         if (copy_to_user(argp, &range, sizeof(range)))
1065                 status = -EFAULT;
1066
1067 out_free:
1068         kfree(context);
1069 out_drop:
1070         mnt_drop_write_file(filp);
1071
1072         return status;
1073 }