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