Pileus Git - ~andy/linux/blob - 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 partial;
  48         int credits;
  49         u32 new_phys_cpos;
  50         u32 clusters_moved;
  51         u64 refcount_loc;
  52         struct ocfs2_move_extents *range;
  53         struct ocfs2_extent_tree et;
  54         struct ocfs2_alloc_context *meta_ac;
  55         struct ocfs2_alloc_context *data_ac;
  56         struct ocfs2_cached_dealloc_ctxt dealloc;
  57 };
  58
  59 static int __ocfs2_move_extent(handle_t *handle,
  60                                struct ocfs2_move_extents_context *context,
  61                                u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
  62                                int ext_flags)
  63 {
  64         int ret = 0, index;
  65         struct inode *inode = context->inode;
  66         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  67         struct ocfs2_extent_rec *rec, replace_rec;
  68         struct ocfs2_path *path = NULL;
  69         struct ocfs2_extent_list *el;
  70         u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
  71         u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
  72
  73         ret = ocfs2_duplicate_clusters_by_page(handle, context->file, cpos,
  74                                                p_cpos, new_p_cpos, len);
  75         if (ret) {
  76                 mlog_errno(ret);
  77                 goto out;
  78         }
  79
  80         memset(&replace_rec, 0, sizeof(replace_rec));
  81         replace_rec.e_cpos = cpu_to_le32(cpos);
  82         replace_rec.e_leaf_clusters = cpu_to_le16(len);
  83         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
  84                                                                    new_p_cpos));
  85
  86         path = ocfs2_new_path_from_et(&context->et);
  87         if (!path) {
  88                 ret = -ENOMEM;
  89                 mlog_errno(ret);
  90                 goto out;
  91         }
  92
  93         ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
  94         if (ret) {
  95                 mlog_errno(ret);
  96                 goto out;
  97         }
  98
  99         el = path_leaf_el(path);
 100
 101         index = ocfs2_search_extent_list(el, cpos);
 102         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
 103                 ocfs2_error(inode->i_sb,
 104                             "Inode %llu has an extent at cpos %u which can no "
 105                             "longer be found.\n",
 106                             (unsigned long long)ino, cpos);
 107                 ret = -EROFS;
 108                 goto out;
 109         }
 110
 111         rec = &el->l_recs[index];
 112
 113         BUG_ON(ext_flags != rec->e_flags);
 114         /*
 115          * after moving/defraging to new location, the extent is not going
 116          * to be refcounted anymore.
 117          */
 118         replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
 119
 120         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
 121                                       context->et.et_root_bh,
 122                                       OCFS2_JOURNAL_ACCESS_WRITE);
 123         if (ret) {
 124                 mlog_errno(ret);
 125                 goto out;
 126         }
 127
 128         ret = ocfs2_split_extent(handle, &context->et, path, index,
 129                                  &replace_rec, context->meta_ac,
 130                                  &context->dealloc);
 131         if (ret) {
 132                 mlog_errno(ret);
 133                 goto out;
 134         }
 135
 136         ocfs2_journal_dirty(handle, context->et.et_root_bh);
 137
 138         context->new_phys_cpos = new_p_cpos;
 139
 140         /*
 141          * need I to append truncate log for old clusters?
 142          */
 143         if (old_blkno) {
 144                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 145                         ret = ocfs2_decrease_refcount(inode, handle,
 146                                         ocfs2_blocks_to_clusters(osb->sb,
 147                                                                  old_blkno),
 148                                         len, context->meta_ac,
 149                                         &context->dealloc, 1);
 150                 else
 151                         ret = ocfs2_truncate_log_append(osb, handle,
 152                                                         old_blkno, len);
 153         }
 154
 155 out:
 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                                               clusters_to_move + 2);
 206
 207         mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 208              extra_blocks, clusters_to_move, *credits);
 209 out:
 210         if (ret) {
 211                 if (*meta_ac) {
 212                         ocfs2_free_alloc_context(*meta_ac);
 213                         *meta_ac = NULL;
 214                 }
 215         }
 216
 217         return ret;
 218 }
 219
 220 /*
 221  * Using one journal handle to guarantee the data consistency in case
 222  * crash happens anywhere.
 223  *
 224  *  XXX: defrag can end up with finishing partial extent as requested,
 225  * due to not enough contiguous clusters can be found in allocator.
 226  */
 227 static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 228                                u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 229 {
 230         int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 231         handle_t *handle;
 232         struct inode *inode = context->inode;
 233         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 234         struct inode *tl_inode = osb->osb_tl_inode;
 235         struct ocfs2_refcount_tree *ref_tree = NULL;
 236         u32 new_phys_cpos, new_len;
 237         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 238
 239         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 240
 241                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 242                          OCFS2_HAS_REFCOUNT_FL));
 243
 244                 BUG_ON(!context->refcount_loc);
 245
 246                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 247                                                &ref_tree, NULL);
 248                 if (ret) {
 249                         mlog_errno(ret);
 250                         return ret;
 251                 }
 252
 253                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 254                                                         context->refcount_loc,
 255                                                         phys_blkno,
 256                                                         *len,
 257                                                         &credits,
 258                                                         &extra_blocks);
 259                 if (ret) {
 260                         mlog_errno(ret);
 261                         goto out;
 262                 }
 263         }
 264
 265         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, *len, 1,
 266                                                  &context->meta_ac,
 267                                                  &context->data_ac,
 268                                                  extra_blocks, &credits);
 269         if (ret) {
 270                 mlog_errno(ret);
 271                 goto out;
 272         }
 273
 274         /*
 275          * should be using allocation reservation strategy there?
 276          *
 277          * if (context->data_ac)
 278          *      context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 279          */
 280
 281         mutex_lock(&tl_inode->i_mutex);
 282
 283         if (ocfs2_truncate_log_needs_flush(osb)) {
 284                 ret = __ocfs2_flush_truncate_log(osb);
 285                 if (ret < 0) {
 286                         mlog_errno(ret);
 287                         goto out_unlock_mutex;
 288                 }
 289         }
 290
 291         handle = ocfs2_start_trans(osb, credits);
 292         if (IS_ERR(handle)) {
 293                 ret = PTR_ERR(handle);
 294                 mlog_errno(ret);
 295                 goto out_unlock_mutex;
 296         }
 297
 298         ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 299                                      &new_phys_cpos, &new_len);
 300         if (ret) {
 301                 mlog_errno(ret);
 302                 goto out_commit;
 303         }
 304
 305         /*
 306          * allowing partial extent moving is kind of 'pros and cons', it makes
 307          * whole defragmentation less likely to fail, on the contrary, the bad
 308          * thing is it may make the fs even more fragmented after moving, let
 309          * userspace make a good decision here.
 310          */
 311         if (new_len != *len) {
 312                 mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 313                 if (!partial) {
 314                         context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 315                         ret = -ENOSPC;
 316                         goto out_commit;
 317                 }
 318         }
 319
 320         mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 321              phys_cpos, new_phys_cpos);
 322
 323         ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 324                                   new_phys_cpos, ext_flags);
 325         if (ret)
 326                 mlog_errno(ret);
 327
 328         if (partial && (new_len != *len))
 329                 *len = new_len;
 330
 331         /*
 332          * Here we should write the new page out first if we are
 333          * in write-back mode.
 334          */
 335         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 336         if (ret)
 337                 mlog_errno(ret);
 338
 339 out_commit:
 340         ocfs2_commit_trans(osb, handle);
 341
 342 out_unlock_mutex:
 343         mutex_unlock(&tl_inode->i_mutex);
 344
 345         if (context->data_ac) {
 346                 ocfs2_free_alloc_context(context->data_ac);
 347                 context->data_ac = NULL;
 348         }
 349
 350         if (context->meta_ac) {
 351                 ocfs2_free_alloc_context(context->meta_ac);
 352                 context->meta_ac = NULL;
 353         }
 354
 355 out:
 356         if (ref_tree)
 357                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 358
 359         return ret;
 360 }
 361
 362 /*
 363  * find the victim alloc group, where #blkno fits.
 364  */
 365 static int ocfs2_find_victim_alloc_group(struct inode *inode,
 366                                          u64 vict_blkno,
 367                                          int type, int slot,
 368                                          int *vict_bit,
 369                                          struct buffer_head **ret_bh)
 370 {
 371         int ret, i, blocks_per_unit = 1;
 372         u64 blkno;
 373         char namebuf[40];
 374
 375         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 376         struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 377         struct ocfs2_chain_list *cl;
 378         struct ocfs2_chain_rec *rec;
 379         struct ocfs2_dinode *ac_dinode;
 380         struct ocfs2_group_desc *bg;
 381
 382         ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 383         ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf,
 384                                          strlen(namebuf), &blkno);
 385         if (ret) {
 386                 ret = -ENOENT;
 387                 goto out;
 388         }
 389
 390         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 391         if (ret) {
 392                 mlog_errno(ret);
 393                 goto out;
 394         }
 395
 396         ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 397         cl = &(ac_dinode->id2.i_chain);
 398         rec = &(cl->cl_recs[0]);
 399
 400         if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 401                 blocks_per_unit <<= (osb->s_clustersize_bits -
 402                                                 inode->i_sb->s_blocksize_bits);
 403         /*
 404          * 'vict_blkno' was out of the valid range.
 405          */
 406         if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 407             (vict_blkno >= (le32_to_cpu(ac_dinode->id1.bitmap1.i_total) *
 408                                 blocks_per_unit))) {
 409                 ret = -EINVAL;
 410                 goto out;
 411         }
 412
 413         for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 414
 415                 rec = &(cl->cl_recs[i]);
 416                 if (!rec)
 417                         continue;
 418
 419                 bg = NULL;
 420
 421                 do {
 422                         if (!bg)
 423                                 blkno = le64_to_cpu(rec->c_blkno);
 424                         else
 425                                 blkno = le64_to_cpu(bg->bg_next_group);
 426
 427                         if (gd_bh) {
 428                                 brelse(gd_bh);
 429                                 gd_bh = NULL;
 430                         }
 431
 432                         ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 433                         if (ret) {
 434                                 mlog_errno(ret);
 435                                 goto out;
 436                         }
 437
 438                         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 439
 440                         if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 441                                                 le16_to_cpu(bg->bg_bits))) {
 442
 443                                 *ret_bh = gd_bh;
 444                                 *vict_bit = (vict_blkno - blkno) /
 445                                                         blocks_per_unit;
 446                                 mlog(0, "find the victim group: #%llu, "
 447                                      "total_bits: %u, vict_bit: %u\n",
 448                                      blkno, le16_to_cpu(bg->bg_bits),
 449                                      *vict_bit);
 450                                 goto out;
 451                         }
 452
 453                 } while (le64_to_cpu(bg->bg_next_group));
 454         }
 455
 456         ret = -EINVAL;
 457 out:
 458         brelse(ac_bh);
 459
 460         /*
 461          * caller has to release the gd_bh properly.
 462          */
 463         return ret;
 464 }
 465
 466 /*
 467  * XXX: helper to validate and adjust moving goal.
 468  */
 469 static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 470                                                struct ocfs2_move_extents *range)
 471 {
 472         int ret, goal_bit = 0;
 473
 474         struct buffer_head *gd_bh = NULL;
 475         struct ocfs2_group_desc *bg;
 476         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 477         int c_to_b = 1 << (osb->s_clustersize_bits -
 478                                         inode->i_sb->s_blocksize_bits);
 479
 480         /*
 481          * validate goal sits within global_bitmap, and return the victim
 482          * group desc
 483          */
 484         ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 485                                             GLOBAL_BITMAP_SYSTEM_INODE,
 486                                             OCFS2_INVALID_SLOT,
 487                                             &goal_bit, &gd_bh);
 488         if (ret)
 489                 goto out;
 490
 491         bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 492
 493         /*
 494          * make goal become cluster aligned.
 495          */
 496         if (range->me_goal % c_to_b)
 497                 range->me_goal = range->me_goal / c_to_b * c_to_b;
 498
 499         /*
 500          * moving goal is not allowd to start with a group desc blok(#0 blk)
 501          * let's compromise to the latter cluster.
 502          */
 503         if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 504                 range->me_goal += c_to_b;
 505
 506         /*
 507          * movement is not gonna cross two groups.
 508          */
 509         if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 510                                                                 range->me_len) {
 511                 ret = -EINVAL;
 512                 goto out;
 513         }
 514         /*
 515          * more exact validations/adjustments will be performed later during
 516          * moving operation for each extent range.
 517          */
 518         mlog(0, "extents get ready to be moved to #%llu block\n",
 519              range->me_goal);
 520
 521 out:
 522         brelse(gd_bh);
 523
 524         return ret;
 525 }
 526
 527 static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 528                                     int *goal_bit, u32 move_len, u32 max_hop,
 529                                     u32 *phys_cpos)
 530 {
 531         int i, used, last_free_bits = 0, base_bit = *goal_bit;
 532         struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 533         u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 534                                                  le64_to_cpu(gd->bg_blkno));
 535
 536         for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 537
 538                 used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 539                 if (used) {
 540                         /*
 541                          * we even tried searching the free chunk by jumping
 542                          * a 'max_hop' distance, but still failed.
 543                          */
 544                         if ((i - base_bit) > max_hop) {
 545                                 *phys_cpos = 0;
 546                                 break;
 547                         }
 548
 549                         if (last_free_bits)
 550                                 last_free_bits = 0;
 551
 552                         continue;
 553                 } else
 554                         last_free_bits++;
 555
 556                 if (last_free_bits == move_len) {
 557                         *goal_bit = i;
 558                         *phys_cpos = base_cpos + i;
 559                         break;
 560                 }
 561         }
 562
 563         mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 564 }
 565
 566 static int ocfs2_alloc_dinode_update_counts(struct inode *inode,
 567                                        handle_t *handle,
 568                                        struct buffer_head *di_bh,
 569                                        u32 num_bits,
 570                                        u16 chain)
 571 {
 572         int ret;
 573         u32 tmp_used;
 574         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
 575         struct ocfs2_chain_list *cl =
 576                                 (struct ocfs2_chain_list *) &di->id2.i_chain;
 577
 578         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 579                                       OCFS2_JOURNAL_ACCESS_WRITE);
 580         if (ret < 0) {
 581                 mlog_errno(ret);
 582                 goto out;
 583         }
 584
 585         tmp_used = le32_to_cpu(di->id1.bitmap1.i_used);
 586         di->id1.bitmap1.i_used = cpu_to_le32(num_bits + tmp_used);
 587         le32_add_cpu(&cl->cl_recs[chain].c_free, -num_bits);
 588         ocfs2_journal_dirty(handle, di_bh);
 589
 590 out:
 591         return ret;
 592 }
 593
 594 static inline int ocfs2_block_group_set_bits(handle_t *handle,
 595                                              struct inode *alloc_inode,
 596                                              struct ocfs2_group_desc *bg,
 597                                              struct buffer_head *group_bh,
 598                                              unsigned int bit_off,
 599                                              unsigned int num_bits)
 600 {
 601         int status;
 602         void *bitmap = bg->bg_bitmap;
 603         int journal_type = OCFS2_JOURNAL_ACCESS_WRITE;
 604
 605         /* All callers get the descriptor via
 606          * ocfs2_read_group_descriptor().  Any corruption is a code bug. */
 607         BUG_ON(!OCFS2_IS_VALID_GROUP_DESC(bg));
 608         BUG_ON(le16_to_cpu(bg->bg_free_bits_count) < num_bits);
 609
 610         mlog(0, "block_group_set_bits: off = %u, num = %u\n", bit_off,
 611              num_bits);
 612
 613         if (ocfs2_is_cluster_bitmap(alloc_inode))
 614                 journal_type = OCFS2_JOURNAL_ACCESS_UNDO;
 615
 616         status = ocfs2_journal_access_gd(handle,
 617                                          INODE_CACHE(alloc_inode),
 618                                          group_bh,
 619                                          journal_type);
 620         if (status < 0) {
 621                 mlog_errno(status);
 622                 goto bail;
 623         }
 624
 625         le16_add_cpu(&bg->bg_free_bits_count, -num_bits);
 626         if (le16_to_cpu(bg->bg_free_bits_count) > le16_to_cpu(bg->bg_bits)) {
 627                 ocfs2_error(alloc_inode->i_sb, "Group descriptor # %llu has bit"
 628                             " count %u but claims %u are freed. num_bits %d",
 629                             (unsigned long long)le64_to_cpu(bg->bg_blkno),
 630                             le16_to_cpu(bg->bg_bits),
 631                             le16_to_cpu(bg->bg_free_bits_count), num_bits);
 632                 return -EROFS;
 633         }
 634         while (num_bits--)
 635                 ocfs2_set_bit(bit_off++, bitmap);
 636
 637         ocfs2_journal_dirty(handle, group_bh);
 638
 639 bail:
 640         return status;
 641 }
 642
 643 static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 644                              u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 645                              u32 len, int ext_flags)
 646 {
 647         int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 648         handle_t *handle;
 649         struct inode *inode = context->inode;
 650         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 651         struct inode *tl_inode = osb->osb_tl_inode;
 652         struct inode *gb_inode = NULL;
 653         struct buffer_head *gb_bh = NULL;
 654         struct buffer_head *gd_bh = NULL;
 655         struct ocfs2_group_desc *gd;
 656         struct ocfs2_refcount_tree *ref_tree = NULL;
 657         u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 658                                                     context->range->me_threshold);
 659         u64 phys_blkno, new_phys_blkno;
 660
 661         phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 662
 663         if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 664
 665                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
 666                          OCFS2_HAS_REFCOUNT_FL));
 667
 668                 BUG_ON(!context->refcount_loc);
 669
 670                 ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 671                                                &ref_tree, NULL);
 672                 if (ret) {
 673                         mlog_errno(ret);
 674                         return ret;
 675                 }
 676
 677                 ret = ocfs2_prepare_refcount_change_for_del(inode,
 678                                                         context->refcount_loc,
 679                                                         phys_blkno,
 680                                                         len,
 681                                                         &credits,
 682                                                         &extra_blocks);
 683                 if (ret) {
 684                         mlog_errno(ret);
 685                         goto out;
 686                 }
 687         }
 688
 689         ret = ocfs2_lock_allocators_move_extents(inode, &context->et, len, 1,
 690                                                  &context->meta_ac,
 691                                                  NULL, extra_blocks, &credits);
 692         if (ret) {
 693                 mlog_errno(ret);
 694                 goto out;
 695         }
 696
 697         /*
 698          * need to count 2 extra credits for global_bitmap inode and
 699          * group descriptor.
 700          */
 701         credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 702
 703         /*
 704          * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 705          * logic, while we still need to lock the global_bitmap.
 706          */
 707         gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 708                                                OCFS2_INVALID_SLOT);
 709         if (!gb_inode) {
 710                 mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 711                 ret = -EIO;
 712                 goto out;
 713         }
 714
 715         mutex_lock(&gb_inode->i_mutex);
 716
 717         ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 718         if (ret) {
 719                 mlog_errno(ret);
 720                 goto out_unlock_gb_mutex;
 721         }
 722
 723         mutex_lock(&tl_inode->i_mutex);
 724
 725         handle = ocfs2_start_trans(osb, credits);
 726         if (IS_ERR(handle)) {
 727                 ret = PTR_ERR(handle);
 728                 mlog_errno(ret);
 729                 goto out_unlock_tl_inode;
 730         }
 731
 732         new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 733         ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 734                                             GLOBAL_BITMAP_SYSTEM_INODE,
 735                                             OCFS2_INVALID_SLOT,
 736                                             &goal_bit, &gd_bh);
 737         if (ret) {
 738                 mlog_errno(ret);
 739                 goto out_commit;
 740         }
 741
 742         /*
 743          * probe the victim cluster group to find a proper
 744          * region to fit wanted movement, it even will perfrom
 745          * a best-effort attempt by compromising to a threshold
 746          * around the goal.
 747          */
 748         ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 749                                 new_phys_cpos);
 750         if (!new_phys_cpos) {
 751                 ret = -ENOSPC;
 752                 goto out_commit;
 753         }
 754
 755         ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 756                                   *new_phys_cpos, ext_flags);
 757         if (ret) {
 758                 mlog_errno(ret);
 759                 goto out_commit;
 760         }
 761
 762         gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 763         ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 764                                                le16_to_cpu(gd->bg_chain));
 765         if (ret) {
 766                 mlog_errno(ret);
 767                 goto out_commit;
 768         }
 769
 770         ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 771                                          goal_bit, len);
 772         if (ret)
 773                 mlog_errno(ret);
 774
 775         /*
 776          * Here we should write the new page out first if we are
 777          * in write-back mode.
 778          */
 779         ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 780         if (ret)
 781                 mlog_errno(ret);
 782
 783 out_commit:
 784         ocfs2_commit_trans(osb, handle);
 785         brelse(gd_bh);
 786
 787 out_unlock_tl_inode:
 788         mutex_unlock(&tl_inode->i_mutex);
 789
 790         ocfs2_inode_unlock(gb_inode, 1);
 791 out_unlock_gb_mutex:
 792         mutex_unlock(&gb_inode->i_mutex);
 793         brelse(gb_bh);
 794         iput(gb_inode);
 795
 796 out:
 797         if (context->meta_ac) {
 798                 ocfs2_free_alloc_context(context->meta_ac);
 799                 context->meta_ac = NULL;
 800         }
 801
 802         if (ref_tree)
 803                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 804
 805         return ret;
 806 }
 807
 808 /*
 809  * Helper to calculate the defraging length in one run according to threshold.
 810  */
 811 static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 812                                          u32 threshold, int *skip)
 813 {
 814         if ((*alloc_size + *len_defraged) < threshold) {
 815                 /*
 816                  * proceed defragmentation until we meet the thresh
 817                  */
 818                 *len_defraged += *alloc_size;
 819         } else if (*len_defraged == 0) {
 820                 /*
 821                  * XXX: skip a large extent.
 822                  */
 823                 *skip = 1;
 824         } else {
 825                 /*
 826                  * split this extent to coalesce with former pieces as
 827                  * to reach the threshold.
 828                  *
 829                  * we're done here with one cycle of defragmentation
 830                  * in a size of 'thresh', resetting 'len_defraged'
 831                  * forces a new defragmentation.
 832                  */
 833                 *alloc_size = threshold - *len_defraged;
 834                 *len_defraged = 0;
 835         }
 836 }
 837
 838 static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 839                                 struct ocfs2_move_extents_context *context)
 840 {
 841         int ret = 0, flags, do_defrag, skip = 0;
 842         u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 843         u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 844
 845         struct inode *inode = context->inode;
 846         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 847         struct ocfs2_move_extents *range = context->range;
 848         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 849
 850         if ((inode->i_size == 0) || (range->me_len == 0))
 851                 return 0;
 852
 853         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 854                 return 0;
 855
 856         context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 857
 858         ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 859         ocfs2_init_dealloc_ctxt(&context->dealloc);
 860
 861         /*
 862          * TO-DO XXX:
 863          *
 864          * - xattr extents.
 865          */
 866
 867         do_defrag = context->auto_defrag;
 868
 869         /*
 870          * extents moving happens in unit of clusters, for the sake
 871          * of simplicity, we may ignore two clusters where 'byte_start'
 872          * and 'byte_start + len' were within.
 873          */
 874         move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 875         len_to_move = (range->me_start + range->me_len) >>
 876                                                 osb->s_clustersize_bits;
 877         if (len_to_move >= move_start)
 878                 len_to_move -= move_start;
 879         else
 880                 len_to_move = 0;
 881
 882         if (do_defrag) {
 883                 defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 884                 if (defrag_thresh <= 1)
 885                         goto done;
 886         } else
 887                 new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 888                                                          range->me_goal);
 889
 890         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 891              "thresh: %u\n",
 892              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 893              (unsigned long long)range->me_start,
 894              (unsigned long long)range->me_len,
 895              move_start, len_to_move, defrag_thresh);
 896
 897         cpos = move_start;
 898         while (len_to_move) {
 899                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 900                                          &flags);
 901                 if (ret) {
 902                         mlog_errno(ret);
 903                         goto out;
 904                 }
 905
 906                 if (alloc_size > len_to_move)
 907                         alloc_size = len_to_move;
 908
 909                 /*
 910                  * XXX: how to deal with a hole:
 911                  *
 912                  * - skip the hole of course
 913                  * - force a new defragmentation
 914                  */
 915                 if (!phys_cpos) {
 916                         if (do_defrag)
 917                                 len_defraged = 0;
 918
 919                         goto next;
 920                 }
 921
 922                 if (do_defrag) {
 923                         ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 924                                                      defrag_thresh, &skip);
 925                         /*
 926                          * skip large extents
 927                          */
 928                         if (skip) {
 929                                 skip = 0;
 930                                 goto next;
 931                         }
 932
 933                         mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 934                              "alloc_size: %u, len_defraged: %u\n",
 935                              cpos, phys_cpos, alloc_size, len_defraged);
 936
 937                         ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 938                                                   &alloc_size, flags);
 939                 } else {
 940                         ret = ocfs2_move_extent(context, cpos, phys_cpos,
 941                                                 &new_phys_cpos, alloc_size,
 942                                                 flags);
 943
 944                         new_phys_cpos += alloc_size;
 945                 }
 946
 947                 if (ret < 0) {
 948                         mlog_errno(ret);
 949                         goto out;
 950                 }
 951
 952                 context->clusters_moved += alloc_size;
 953 next:
 954                 cpos += alloc_size;
 955                 len_to_move -= alloc_size;
 956         }
 957
 958 done:
 959         range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 960
 961 out:
 962         range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 963                                                       context->clusters_moved);
 964         range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 965                                                        context->new_phys_cpos);
 966
 967         ocfs2_schedule_truncate_log_flush(osb, 1);
 968         ocfs2_run_deallocs(osb, &context->dealloc);
 969
 970         return ret;
 971 }
 972
 973 static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 974 {
 975         int status;
 976         handle_t *handle;
 977         struct inode *inode = context->inode;
 978         struct ocfs2_dinode *di;
 979         struct buffer_head *di_bh = NULL;
 980         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 981
 982         if (!inode)
 983                 return -ENOENT;
 984
 985         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 986                 return -EROFS;
 987
 988         mutex_lock(&inode->i_mutex);
 989
 990         /*
 991          * This prevents concurrent writes from other nodes
 992          */
 993         status = ocfs2_rw_lock(inode, 1);
 994         if (status) {
 995                 mlog_errno(status);
 996                 goto out;
 997         }
 998
 999         status = ocfs2_inode_lock(inode, &di_bh, 1);
1000         if (status) {
1001                 mlog_errno(status);
1002                 goto out_rw_unlock;
1003         }
1004
1005         /*
1006          * rememer ip_xattr_sem also needs to be held if necessary
1007          */
1008         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1009
1010         status = __ocfs2_move_extents_range(di_bh, context);
1011
1012         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1013         if (status) {
1014                 mlog_errno(status);
1015                 goto out_inode_unlock;
1016         }
1017
1018         /*
1019          * We update ctime for these changes
1020          */
1021         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1022         if (IS_ERR(handle)) {
1023                 status = PTR_ERR(handle);
1024                 mlog_errno(status);
1025                 goto out_inode_unlock;
1026         }
1027
1028         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
1029                                          OCFS2_JOURNAL_ACCESS_WRITE);
1030         if (status) {
1031                 mlog_errno(status);
1032                 goto out_commit;
1033         }
1034
1035         di = (struct ocfs2_dinode *)di_bh->b_data;
1036         inode->i_ctime = CURRENT_TIME;
1037         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
1038         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
1039
1040         ocfs2_journal_dirty(handle, di_bh);
1041
1042 out_commit:
1043         ocfs2_commit_trans(osb, handle);
1044
1045 out_inode_unlock:
1046         brelse(di_bh);
1047         ocfs2_inode_unlock(inode, 1);
1048 out_rw_unlock:
1049         ocfs2_rw_unlock(inode, 1);
1050 out:
1051         mutex_unlock(&inode->i_mutex);
1052
1053         return status;
1054 }
1055
1056 int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
1057 {
1058         int status;
1059
1060         struct inode *inode = filp->f_path.dentry->d_inode;
1061         struct ocfs2_move_extents range;
1062         struct ocfs2_move_extents_context *context = NULL;
1063
1064         status = mnt_want_write(filp->f_path.mnt);
1065         if (status)
1066                 return status;
1067
1068         if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE))
1069                 goto out;
1070
1071         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1072                 status = -EPERM;
1073                 goto out;
1074         }
1075
1076         context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1077         if (!context) {
1078                 status = -ENOMEM;
1079                 mlog_errno(status);
1080                 goto out;
1081         }
1082
1083         context->inode = inode;
1084         context->file = filp;
1085
1086         if (argp) {
1087                 if (copy_from_user(&range, (struct ocfs2_move_extents *)argp,
1088                                    sizeof(range))) {
1089                         status = -EFAULT;
1090                         goto out;
1091                 }
1092         } else {
1093                 status = -EINVAL;
1094                 goto out;
1095         }
1096
1097         if (range.me_start > i_size_read(inode))
1098                 goto out;
1099
1100         if (range.me_start + range.me_len > i_size_read(inode))
1101                         range.me_len = i_size_read(inode) - range.me_start;
1102
1103         context->range = &range;
1104
1105         if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1106                 context->auto_defrag = 1;
1107                 /*
1108                  * ok, the default theshold for the defragmentation
1109                  * is 1M, since our maximum clustersize was 1M also.
1110                  * any thought?
1111                  */
1112                 if (!range.me_threshold)
1113                         range.me_threshold = 1024 * 1024;
1114
1115                 if (range.me_threshold > i_size_read(inode))
1116                         range.me_threshold = i_size_read(inode);
1117
1118                 if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1119                         context->partial = 1;
1120         } else {
1121                 /*
1122                  * first best-effort attempt to validate and adjust the goal
1123                  * (physical address in block), while it can't guarantee later
1124                  * operation can succeed all the time since global_bitmap may
1125                  * change a bit over time.
1126                  */
1127
1128                 status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1129                 if (status)
1130                         goto out;
1131         }
1132
1133         status = ocfs2_move_extents(context);
1134         if (status)
1135                 mlog_errno(status);
1136 out:
1137         /*
1138          * movement/defragmentation may end up being partially completed,
1139          * that's the reason why we need to return userspace the finished
1140          * length and new_offset even if failure happens somewhere.
1141          */
1142         if (argp) {
1143                 if (copy_to_user((struct ocfs2_move_extents *)argp, &range,
1144                                 sizeof(range)))
1145                         status = -EFAULT;
1146         }
1147
1148         kfree(context);
1149
1150         mnt_drop_write(filp->f_path.mnt);
1151
1152         return status;
1153 }