fs/ocfs2/refcounttree.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * refcounttree.c
   5  *
   6  * Copyright (C) 2009 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
  18 #include <linux/sort.h>
  19 #define MLOG_MASK_PREFIX ML_REFCOUNT
  20 #include <cluster/masklog.h>
  21 #include "ocfs2.h"
  22 #include "inode.h"
  23 #include "alloc.h"
  24 #include "suballoc.h"
  25 #include "journal.h"
  26 #include "uptodate.h"
  27 #include "super.h"
  28 #include "buffer_head_io.h"
  29 #include "blockcheck.h"
  30 #include "refcounttree.h"
  31 #include "sysfile.h"
  32 #include "dlmglue.h"
  33 #include "extent_map.h"
  34 #include "aops.h"
  35 #include "xattr.h"
  36 #include "namei.h"
  37
  38 #include <linux/bio.h>
  39 #include <linux/blkdev.h>
  40 #include <linux/gfp.h>
  41 #include <linux/slab.h>
  42 #include <linux/writeback.h>
  43 #include <linux/pagevec.h>
  44 #include <linux/swap.h>
  45 #include <linux/security.h>
  46 #include <linux/fsnotify.h>
  47 #include <linux/quotaops.h>
  48 #include <linux/namei.h>
  49 #include <linux/mount.h>
  50
  51 struct ocfs2_cow_context {
  52         struct inode *inode;
  53         u32 cow_start;
  54         u32 cow_len;
  55         struct ocfs2_extent_tree data_et;
  56         struct ocfs2_refcount_tree *ref_tree;
  57         struct buffer_head *ref_root_bh;
  58         struct ocfs2_alloc_context *meta_ac;
  59         struct ocfs2_alloc_context *data_ac;
  60         struct ocfs2_cached_dealloc_ctxt dealloc;
  61         void *cow_object;
  62         struct ocfs2_post_refcount *post_refcount;
  63         int extra_credits;
  64         int (*get_clusters)(struct ocfs2_cow_context *context,
  65                             u32 v_cluster, u32 *p_cluster,
  66                             u32 *num_clusters,
  67                             unsigned int *extent_flags);
  68         int (*cow_duplicate_clusters)(handle_t *handle,
  69                                       struct ocfs2_cow_context *context,
  70                                       u32 cpos, u32 old_cluster,
  71                                       u32 new_cluster, u32 new_len);
  72 };
  73
  74 static inline struct ocfs2_refcount_tree *
  75 cache_info_to_refcount(struct ocfs2_caching_info *ci)
  76 {
  77         return container_of(ci, struct ocfs2_refcount_tree, rf_ci);
  78 }
  79
  80 static int ocfs2_validate_refcount_block(struct super_block *sb,
  81                                          struct buffer_head *bh)
  82 {
  83         int rc;
  84         struct ocfs2_refcount_block *rb =
  85                 (struct ocfs2_refcount_block *)bh->b_data;
  86
  87         mlog(0, "Validating refcount block %llu\n",
  88              (unsigned long long)bh->b_blocknr);
  89
  90         BUG_ON(!buffer_uptodate(bh));
  91
  92         /*
  93          * If the ecc fails, we return the error but otherwise
  94          * leave the filesystem running.  We know any error is
  95          * local to this block.
  96          */
  97         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check);
  98         if (rc) {
  99                 mlog(ML_ERROR, "Checksum failed for refcount block %llu\n",
 100                      (unsigned long long)bh->b_blocknr);
 101                 return rc;
 102         }
 103
 104
 105         if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) {
 106                 ocfs2_error(sb,
 107                             "Refcount block #%llu has bad signature %.*s",
 108                             (unsigned long long)bh->b_blocknr, 7,
 109                             rb->rf_signature);
 110                 return -EINVAL;
 111         }
 112
 113         if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) {
 114                 ocfs2_error(sb,
 115                             "Refcount block #%llu has an invalid rf_blkno "
 116                             "of %llu",
 117                             (unsigned long long)bh->b_blocknr,
 118                             (unsigned long long)le64_to_cpu(rb->rf_blkno));
 119                 return -EINVAL;
 120         }
 121
 122         if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) {
 123                 ocfs2_error(sb,
 124                             "Refcount block #%llu has an invalid "
 125                             "rf_fs_generation of #%u",
 126                             (unsigned long long)bh->b_blocknr,
 127                             le32_to_cpu(rb->rf_fs_generation));
 128                 return -EINVAL;
 129         }
 130
 131         return 0;
 132 }
 133
 134 static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci,
 135                                      u64 rb_blkno,
 136                                      struct buffer_head **bh)
 137 {
 138         int rc;
 139         struct buffer_head *tmp = *bh;
 140
 141         rc = ocfs2_read_block(ci, rb_blkno, &tmp,
 142                               ocfs2_validate_refcount_block);
 143
 144         /* If ocfs2_read_block() got us a new bh, pass it up. */
 145         if (!rc && !*bh)
 146                 *bh = tmp;
 147
 148         return rc;
 149 }
 150
 151 static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci)
 152 {
 153         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 154
 155         return rf->rf_blkno;
 156 }
 157
 158 static struct super_block *
 159 ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci)
 160 {
 161         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 162
 163         return rf->rf_sb;
 164 }
 165
 166 static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci)
 167 {
 168         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 169
 170         spin_lock(&rf->rf_lock);
 171 }
 172
 173 static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci)
 174 {
 175         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 176
 177         spin_unlock(&rf->rf_lock);
 178 }
 179
 180 static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci)
 181 {
 182         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 183
 184         mutex_lock(&rf->rf_io_mutex);
 185 }
 186
 187 static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci)
 188 {
 189         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 190
 191         mutex_unlock(&rf->rf_io_mutex);
 192 }
 193
 194 static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = {
 195         .co_owner               = ocfs2_refcount_cache_owner,
 196         .co_get_super           = ocfs2_refcount_cache_get_super,
 197         .co_cache_lock          = ocfs2_refcount_cache_lock,
 198         .co_cache_unlock        = ocfs2_refcount_cache_unlock,
 199         .co_io_lock             = ocfs2_refcount_cache_io_lock,
 200         .co_io_unlock           = ocfs2_refcount_cache_io_unlock,
 201 };
 202
 203 static struct ocfs2_refcount_tree *
 204 ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno)
 205 {
 206         struct rb_node *n = osb->osb_rf_lock_tree.rb_node;
 207         struct ocfs2_refcount_tree *tree = NULL;
 208
 209         while (n) {
 210                 tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node);
 211
 212                 if (blkno < tree->rf_blkno)
 213                         n = n->rb_left;
 214                 else if (blkno > tree->rf_blkno)
 215                         n = n->rb_right;
 216                 else
 217                         return tree;
 218         }
 219
 220         return NULL;
 221 }
 222
 223 /* osb_lock is already locked. */
 224 static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb,
 225                                        struct ocfs2_refcount_tree *new)
 226 {
 227         u64 rf_blkno = new->rf_blkno;
 228         struct rb_node *parent = NULL;
 229         struct rb_node **p = &osb->osb_rf_lock_tree.rb_node;
 230         struct ocfs2_refcount_tree *tmp;
 231
 232         while (*p) {
 233                 parent = *p;
 234
 235                 tmp = rb_entry(parent, struct ocfs2_refcount_tree,
 236                                rf_node);
 237
 238                 if (rf_blkno < tmp->rf_blkno)
 239                         p = &(*p)->rb_left;
 240                 else if (rf_blkno > tmp->rf_blkno)
 241                         p = &(*p)->rb_right;
 242                 else {
 243                         /* This should never happen! */
 244                         mlog(ML_ERROR, "Duplicate refcount block %llu found!\n",
 245                              (unsigned long long)rf_blkno);
 246                         BUG();
 247                 }
 248         }
 249
 250         rb_link_node(&new->rf_node, parent, p);
 251         rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree);
 252 }
 253
 254 static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree)
 255 {
 256         ocfs2_metadata_cache_exit(&tree->rf_ci);
 257         ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres);
 258         ocfs2_lock_res_free(&tree->rf_lockres);
 259         kfree(tree);
 260 }
 261
 262 static inline void
 263 ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb,
 264                                         struct ocfs2_refcount_tree *tree)
 265 {
 266         rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree);
 267         if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree)
 268                 osb->osb_ref_tree_lru = NULL;
 269 }
 270
 271 static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb,
 272                                         struct ocfs2_refcount_tree *tree)
 273 {
 274         spin_lock(&osb->osb_lock);
 275         ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
 276         spin_unlock(&osb->osb_lock);
 277 }
 278
 279 static void ocfs2_kref_remove_refcount_tree(struct kref *kref)
 280 {
 281         struct ocfs2_refcount_tree *tree =
 282                 container_of(kref, struct ocfs2_refcount_tree, rf_getcnt);
 283
 284         ocfs2_free_refcount_tree(tree);
 285 }
 286
 287 static inline void
 288 ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree)
 289 {
 290         kref_get(&tree->rf_getcnt);
 291 }
 292
 293 static inline void
 294 ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree)
 295 {
 296         kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree);
 297 }
 298
 299 static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new,
 300                                                struct super_block *sb)
 301 {
 302         ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops);
 303         mutex_init(&new->rf_io_mutex);
 304         new->rf_sb = sb;
 305         spin_lock_init(&new->rf_lock);
 306 }
 307
 308 static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb,
 309                                         struct ocfs2_refcount_tree *new,
 310                                         u64 rf_blkno, u32 generation)
 311 {
 312         init_rwsem(&new->rf_sem);
 313         ocfs2_refcount_lock_res_init(&new->rf_lockres, osb,
 314                                      rf_blkno, generation);
 315 }
 316
 317 static struct ocfs2_refcount_tree*
 318 ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno)
 319 {
 320         struct ocfs2_refcount_tree *new;
 321
 322         new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS);
 323         if (!new)
 324                 return NULL;
 325
 326         new->rf_blkno = rf_blkno;
 327         kref_init(&new->rf_getcnt);
 328         ocfs2_init_refcount_tree_ci(new, osb->sb);
 329
 330         return new;
 331 }
 332
 333 static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno,
 334                                    struct ocfs2_refcount_tree **ret_tree)
 335 {
 336         int ret = 0;
 337         struct ocfs2_refcount_tree *tree, *new = NULL;
 338         struct buffer_head *ref_root_bh = NULL;
 339         struct ocfs2_refcount_block *ref_rb;
 340
 341         spin_lock(&osb->osb_lock);
 342         if (osb->osb_ref_tree_lru &&
 343             osb->osb_ref_tree_lru->rf_blkno == rf_blkno)
 344                 tree = osb->osb_ref_tree_lru;
 345         else
 346                 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
 347         if (tree)
 348                 goto out;
 349
 350         spin_unlock(&osb->osb_lock);
 351
 352         new = ocfs2_allocate_refcount_tree(osb, rf_blkno);
 353         if (!new) {
 354                 ret = -ENOMEM;
 355                 mlog_errno(ret);
 356                 return ret;
 357         }
 358         /*
 359          * We need the generation to create the refcount tree lock and since
 360          * it isn't changed during the tree modification, we are safe here to
 361          * read without protection.
 362          * We also have to purge the cache after we create the lock since the
 363          * refcount block may have the stale data. It can only be trusted when
 364          * we hold the refcount lock.
 365          */
 366         ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh);
 367         if (ret) {
 368                 mlog_errno(ret);
 369                 ocfs2_metadata_cache_exit(&new->rf_ci);
 370                 kfree(new);
 371                 return ret;
 372         }
 373
 374         ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 375         new->rf_generation = le32_to_cpu(ref_rb->rf_generation);
 376         ocfs2_init_refcount_tree_lock(osb, new, rf_blkno,
 377                                       new->rf_generation);
 378         ocfs2_metadata_cache_purge(&new->rf_ci);
 379
 380         spin_lock(&osb->osb_lock);
 381         tree = ocfs2_find_refcount_tree(osb, rf_blkno);
 382         if (tree)
 383                 goto out;
 384
 385         ocfs2_insert_refcount_tree(osb, new);
 386
 387         tree = new;
 388         new = NULL;
 389
 390 out:
 391         *ret_tree = tree;
 392
 393         osb->osb_ref_tree_lru = tree;
 394
 395         spin_unlock(&osb->osb_lock);
 396
 397         if (new)
 398                 ocfs2_free_refcount_tree(new);
 399
 400         brelse(ref_root_bh);
 401         return ret;
 402 }
 403
 404 static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno)
 405 {
 406         int ret;
 407         struct buffer_head *di_bh = NULL;
 408         struct ocfs2_dinode *di;
 409
 410         ret = ocfs2_read_inode_block(inode, &di_bh);
 411         if (ret) {
 412                 mlog_errno(ret);
 413                 goto out;
 414         }
 415
 416         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
 417
 418         di = (struct ocfs2_dinode *)di_bh->b_data;
 419         *ref_blkno = le64_to_cpu(di->i_refcount_loc);
 420         brelse(di_bh);
 421 out:
 422         return ret;
 423 }
 424
 425 static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
 426                                       struct ocfs2_refcount_tree *tree, int rw)
 427 {
 428         int ret;
 429
 430         ret = ocfs2_refcount_lock(tree, rw);
 431         if (ret) {
 432                 mlog_errno(ret);
 433                 goto out;
 434         }
 435
 436         if (rw)
 437                 down_write(&tree->rf_sem);
 438         else
 439                 down_read(&tree->rf_sem);
 440
 441 out:
 442         return ret;
 443 }
 444
 445 /*
 446  * Lock the refcount tree pointed by ref_blkno and return the tree.
 447  * In most case, we lock the tree and read the refcount block.
 448  * So read it here if the caller really needs it.
 449  *
 450  * If the tree has been re-created by other node, it will free the
 451  * old one and re-create it.
 452  */
 453 int ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
 454                              u64 ref_blkno, int rw,
 455                              struct ocfs2_refcount_tree **ret_tree,
 456                              struct buffer_head **ref_bh)
 457 {
 458         int ret, delete_tree = 0;
 459         struct ocfs2_refcount_tree *tree = NULL;
 460         struct buffer_head *ref_root_bh = NULL;
 461         struct ocfs2_refcount_block *rb;
 462
 463 again:
 464         ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree);
 465         if (ret) {
 466                 mlog_errno(ret);
 467                 return ret;
 468         }
 469
 470         ocfs2_refcount_tree_get(tree);
 471
 472         ret = __ocfs2_lock_refcount_tree(osb, tree, rw);
 473         if (ret) {
 474                 mlog_errno(ret);
 475                 ocfs2_refcount_tree_put(tree);
 476                 goto out;
 477         }
 478
 479         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
 480                                         &ref_root_bh);
 481         if (ret) {
 482                 mlog_errno(ret);
 483                 ocfs2_unlock_refcount_tree(osb, tree, rw);
 484                 ocfs2_refcount_tree_put(tree);
 485                 goto out;
 486         }
 487
 488         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 489         /*
 490          * If the refcount block has been freed and re-created, we may need
 491          * to recreate the refcount tree also.
 492          *
 493          * Here we just remove the tree from the rb-tree, and the last
 494          * kref holder will unlock and delete this refcount_tree.
 495          * Then we goto "again" and ocfs2_get_refcount_tree will create
 496          * the new refcount tree for us.
 497          */
 498         if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) {
 499                 if (!tree->rf_removed) {
 500                         ocfs2_erase_refcount_tree_from_list(osb, tree);
 501                         tree->rf_removed = 1;
 502                         delete_tree = 1;
 503                 }
 504
 505                 ocfs2_unlock_refcount_tree(osb, tree, rw);
 506                 /*
 507                  * We get an extra reference when we create the refcount
 508                  * tree, so another put will destroy it.
 509                  */
 510                 if (delete_tree)
 511                         ocfs2_refcount_tree_put(tree);
 512                 brelse(ref_root_bh);
 513                 ref_root_bh = NULL;
 514                 goto again;
 515         }
 516
 517         *ret_tree = tree;
 518         if (ref_bh) {
 519                 *ref_bh = ref_root_bh;
 520                 ref_root_bh = NULL;
 521         }
 522 out:
 523         brelse(ref_root_bh);
 524         return ret;
 525 }
 526
 527 void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
 528                                 struct ocfs2_refcount_tree *tree, int rw)
 529 {
 530         if (rw)
 531                 up_write(&tree->rf_sem);
 532         else
 533                 up_read(&tree->rf_sem);
 534
 535         ocfs2_refcount_unlock(tree, rw);
 536         ocfs2_refcount_tree_put(tree);
 537 }
 538
 539 void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
 540 {
 541         struct rb_node *node;
 542         struct ocfs2_refcount_tree *tree;
 543         struct rb_root *root = &osb->osb_rf_lock_tree;
 544
 545         while ((node = rb_last(root)) != NULL) {
 546                 tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
 547
 548                 mlog(0, "Purge tree %llu\n",
 549                      (unsigned long long) tree->rf_blkno);
 550
 551                 rb_erase(&tree->rf_node, root);
 552                 ocfs2_free_refcount_tree(tree);
 553         }
 554 }
 555
 556 /*
 557  * Create a refcount tree for an inode.
 558  * We take for granted that the inode is already locked.
 559  */
 560 static int ocfs2_create_refcount_tree(struct inode *inode,
 561                                       struct buffer_head *di_bh)
 562 {
 563         int ret;
 564         handle_t *handle = NULL;
 565         struct ocfs2_alloc_context *meta_ac = NULL;
 566         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 567         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 568         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 569         struct buffer_head *new_bh = NULL;
 570         struct ocfs2_refcount_block *rb;
 571         struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
 572         u16 suballoc_bit_start;
 573         u32 num_got;
 574         u64 suballoc_loc, first_blkno;
 575
 576         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
 577
 578         mlog(0, "create tree for inode %lu\n", inode->i_ino);
 579
 580         ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
 581         if (ret) {
 582                 mlog_errno(ret);
 583                 goto out;
 584         }
 585
 586         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
 587         if (IS_ERR(handle)) {
 588                 ret = PTR_ERR(handle);
 589                 mlog_errno(ret);
 590                 goto out;
 591         }
 592
 593         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 594                                       OCFS2_JOURNAL_ACCESS_WRITE);
 595         if (ret) {
 596                 mlog_errno(ret);
 597                 goto out_commit;
 598         }
 599
 600         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
 601                                    &suballoc_bit_start, &num_got,
 602                                    &first_blkno);
 603         if (ret) {
 604                 mlog_errno(ret);
 605                 goto out_commit;
 606         }
 607
 608         new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
 609         if (!new_tree) {
 610                 ret = -ENOMEM;
 611                 mlog_errno(ret);
 612                 goto out_commit;
 613         }
 614
 615         new_bh = sb_getblk(inode->i_sb, first_blkno);
 616         ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
 617
 618         ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
 619                                       OCFS2_JOURNAL_ACCESS_CREATE);
 620         if (ret) {
 621                 mlog_errno(ret);
 622                 goto out_commit;
 623         }
 624
 625         /* Initialize ocfs2_refcount_block. */
 626         rb = (struct ocfs2_refcount_block *)new_bh->b_data;
 627         memset(rb, 0, inode->i_sb->s_blocksize);
 628         strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
 629         rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
 630         rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
 631         rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
 632         rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
 633         rb->rf_blkno = cpu_to_le64(first_blkno);
 634         rb->rf_count = cpu_to_le32(1);
 635         rb->rf_records.rl_count =
 636                         cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
 637         spin_lock(&osb->osb_lock);
 638         rb->rf_generation = osb->s_next_generation++;
 639         spin_unlock(&osb->osb_lock);
 640
 641         ocfs2_journal_dirty(handle, new_bh);
 642
 643         spin_lock(&oi->ip_lock);
 644         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
 645         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 646         di->i_refcount_loc = cpu_to_le64(first_blkno);
 647         spin_unlock(&oi->ip_lock);
 648
 649         mlog(0, "created tree for inode %lu, refblock %llu\n",
 650              inode->i_ino, (unsigned long long)first_blkno);
 651
 652         ocfs2_journal_dirty(handle, di_bh);
 653
 654         /*
 655          * We have to init the tree lock here since it will use
 656          * the generation number to create it.
 657          */
 658         new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
 659         ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
 660                                       new_tree->rf_generation);
 661
 662         spin_lock(&osb->osb_lock);
 663         tree = ocfs2_find_refcount_tree(osb, first_blkno);
 664
 665         /*
 666          * We've just created a new refcount tree in this block.  If
 667          * we found a refcount tree on the ocfs2_super, it must be
 668          * one we just deleted.  We free the old tree before
 669          * inserting the new tree.
 670          */
 671         BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
 672         if (tree)
 673                 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
 674         ocfs2_insert_refcount_tree(osb, new_tree);
 675         spin_unlock(&osb->osb_lock);
 676         new_tree = NULL;
 677         if (tree)
 678                 ocfs2_refcount_tree_put(tree);
 679
 680 out_commit:
 681         ocfs2_commit_trans(osb, handle);
 682
 683 out:
 684         if (new_tree) {
 685                 ocfs2_metadata_cache_exit(&new_tree->rf_ci);
 686                 kfree(new_tree);
 687         }
 688
 689         brelse(new_bh);
 690         if (meta_ac)
 691                 ocfs2_free_alloc_context(meta_ac);
 692
 693         return ret;
 694 }
 695
 696 static int ocfs2_set_refcount_tree(struct inode *inode,
 697                                    struct buffer_head *di_bh,
 698                                    u64 refcount_loc)
 699 {
 700         int ret;
 701         handle_t *handle = NULL;
 702         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 703         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 704         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 705         struct buffer_head *ref_root_bh = NULL;
 706         struct ocfs2_refcount_block *rb;
 707         struct ocfs2_refcount_tree *ref_tree;
 708
 709         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
 710
 711         ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
 712                                        &ref_tree, &ref_root_bh);
 713         if (ret) {
 714                 mlog_errno(ret);
 715                 return ret;
 716         }
 717
 718         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
 719         if (IS_ERR(handle)) {
 720                 ret = PTR_ERR(handle);
 721                 mlog_errno(ret);
 722                 goto out;
 723         }
 724
 725         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 726                                       OCFS2_JOURNAL_ACCESS_WRITE);
 727         if (ret) {
 728                 mlog_errno(ret);
 729                 goto out_commit;
 730         }
 731
 732         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
 733                                       OCFS2_JOURNAL_ACCESS_WRITE);
 734         if (ret) {
 735                 mlog_errno(ret);
 736                 goto out_commit;
 737         }
 738
 739         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 740         le32_add_cpu(&rb->rf_count, 1);
 741
 742         ocfs2_journal_dirty(handle, ref_root_bh);
 743
 744         spin_lock(&oi->ip_lock);
 745         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
 746         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 747         di->i_refcount_loc = cpu_to_le64(refcount_loc);
 748         spin_unlock(&oi->ip_lock);
 749         ocfs2_journal_dirty(handle, di_bh);
 750
 751 out_commit:
 752         ocfs2_commit_trans(osb, handle);
 753 out:
 754         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 755         brelse(ref_root_bh);
 756
 757         return ret;
 758 }
 759
 760 int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
 761 {
 762         int ret, delete_tree = 0;
 763         handle_t *handle = NULL;
 764         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 765         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 766         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 767         struct ocfs2_refcount_block *rb;
 768         struct inode *alloc_inode = NULL;
 769         struct buffer_head *alloc_bh = NULL;
 770         struct buffer_head *blk_bh = NULL;
 771         struct ocfs2_refcount_tree *ref_tree;
 772         int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
 773         u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
 774         u16 bit = 0;
 775
 776         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
 777                 return 0;
 778
 779         BUG_ON(!ref_blkno);
 780         ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
 781         if (ret) {
 782                 mlog_errno(ret);
 783                 return ret;
 784         }
 785
 786         rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
 787
 788         /*
 789          * If we are the last user, we need to free the block.
 790          * So lock the allocator ahead.
 791          */
 792         if (le32_to_cpu(rb->rf_count) == 1) {
 793                 blk = le64_to_cpu(rb->rf_blkno);
 794                 bit = le16_to_cpu(rb->rf_suballoc_bit);
 795                 if (rb->rf_suballoc_loc)
 796                         bg_blkno = le64_to_cpu(rb->rf_suballoc_loc);
 797                 else
 798                         bg_blkno = ocfs2_which_suballoc_group(blk, bit);
 799
 800                 alloc_inode = ocfs2_get_system_file_inode(osb,
 801                                         EXTENT_ALLOC_SYSTEM_INODE,
 802                                         le16_to_cpu(rb->rf_suballoc_slot));
 803                 if (!alloc_inode) {
 804                         ret = -ENOMEM;
 805                         mlog_errno(ret);
 806                         goto out;
 807                 }
 808                 mutex_lock(&alloc_inode->i_mutex);
 809
 810                 ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
 811                 if (ret) {
 812                         mlog_errno(ret);
 813                         goto out_mutex;
 814                 }
 815
 816                 credits += OCFS2_SUBALLOC_FREE;
 817         }
 818
 819         handle = ocfs2_start_trans(osb, credits);
 820         if (IS_ERR(handle)) {
 821                 ret = PTR_ERR(handle);
 822                 mlog_errno(ret);
 823                 goto out_unlock;
 824         }
 825
 826         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 827                                       OCFS2_JOURNAL_ACCESS_WRITE);
 828         if (ret) {
 829                 mlog_errno(ret);
 830                 goto out_commit;
 831         }
 832
 833         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
 834                                       OCFS2_JOURNAL_ACCESS_WRITE);
 835         if (ret) {
 836                 mlog_errno(ret);
 837                 goto out_commit;
 838         }
 839
 840         spin_lock(&oi->ip_lock);
 841         oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
 842         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 843         di->i_refcount_loc = 0;
 844         spin_unlock(&oi->ip_lock);
 845         ocfs2_journal_dirty(handle, di_bh);
 846
 847         le32_add_cpu(&rb->rf_count , -1);
 848         ocfs2_journal_dirty(handle, blk_bh);
 849
 850         if (!rb->rf_count) {
 851                 delete_tree = 1;
 852                 ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
 853                 ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
 854                                                alloc_bh, bit, bg_blkno, 1);
 855                 if (ret)
 856                         mlog_errno(ret);
 857         }
 858
 859 out_commit:
 860         ocfs2_commit_trans(osb, handle);
 861 out_unlock:
 862         if (alloc_inode) {
 863                 ocfs2_inode_unlock(alloc_inode, 1);
 864                 brelse(alloc_bh);
 865         }
 866 out_mutex:
 867         if (alloc_inode) {
 868                 mutex_unlock(&alloc_inode->i_mutex);
 869                 iput(alloc_inode);
 870         }
 871 out:
 872         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 873         if (delete_tree)
 874                 ocfs2_refcount_tree_put(ref_tree);
 875         brelse(blk_bh);
 876
 877         return ret;
 878 }
 879
 880 static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
 881                                           struct buffer_head *ref_leaf_bh,
 882                                           u64 cpos, unsigned int len,
 883                                           struct ocfs2_refcount_rec *ret_rec,
 884                                           int *index)
 885 {
 886         int i = 0;
 887         struct ocfs2_refcount_block *rb =
 888                 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
 889         struct ocfs2_refcount_rec *rec = NULL;
 890
 891         for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
 892                 rec = &rb->rf_records.rl_recs[i];
 893
 894                 if (le64_to_cpu(rec->r_cpos) +
 895                     le32_to_cpu(rec->r_clusters) <= cpos)
 896                         continue;
 897                 else if (le64_to_cpu(rec->r_cpos) > cpos)
 898                         break;
 899
 900                 /* ok, cpos fail in this rec. Just return. */
 901                 if (ret_rec)
 902                         *ret_rec = *rec;
 903                 goto out;
 904         }
 905
 906         if (ret_rec) {
 907                 /* We meet with a hole here, so fake the rec. */
 908                 ret_rec->r_cpos = cpu_to_le64(cpos);
 909                 ret_rec->r_refcount = 0;
 910                 if (i < le16_to_cpu(rb->rf_records.rl_used) &&
 911                     le64_to_cpu(rec->r_cpos) < cpos + len)
 912                         ret_rec->r_clusters =
 913                                 cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
 914                 else
 915                         ret_rec->r_clusters = cpu_to_le32(len);
 916         }
 917
 918 out:
 919         *index = i;
 920 }
 921
 922 /*
 923  * Try to remove refcount tree. The mechanism is:
 924  * 1) Check whether i_clusters == 0, if no, exit.
 925  * 2) check whether we have i_xattr_loc in dinode. if yes, exit.
 926  * 3) Check whether we have inline xattr stored outside, if yes, exit.
 927  * 4) Remove the tree.
 928  */
 929 int ocfs2_try_remove_refcount_tree(struct inode *inode,
 930                                    struct buffer_head *di_bh)
 931 {
 932         int ret;
 933         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 934         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 935
 936         down_write(&oi->ip_xattr_sem);
 937         down_write(&oi->ip_alloc_sem);
 938
 939         if (oi->ip_clusters)
 940                 goto out;
 941
 942         if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc)
 943                 goto out;
 944
 945         if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL &&
 946             ocfs2_has_inline_xattr_value_outside(inode, di))
 947                 goto out;
 948
 949         ret = ocfs2_remove_refcount_tree(inode, di_bh);
 950         if (ret)
 951                 mlog_errno(ret);
 952 out:
 953         up_write(&oi->ip_alloc_sem);
 954         up_write(&oi->ip_xattr_sem);
 955         return 0;
 956 }
 957
 958 /*
 959  * Find the end range for a leaf refcount block indicated by
 960  * el->l_recs[index].e_blkno.
 961  */
 962 static int ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info *ci,
 963                                        struct buffer_head *ref_root_bh,
 964                                        struct ocfs2_extent_block *eb,
 965                                        struct ocfs2_extent_list *el,
 966                                        int index,  u32 *cpos_end)
 967 {
 968         int ret, i, subtree_root;
 969         u32 cpos;
 970         u64 blkno;
 971         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
 972         struct ocfs2_path *left_path = NULL, *right_path = NULL;
 973         struct ocfs2_extent_tree et;
 974         struct ocfs2_extent_list *tmp_el;
 975
 976         if (index < le16_to_cpu(el->l_next_free_rec) - 1) {
 977                 /*
 978                  * We have a extent rec after index, so just use the e_cpos
 979                  * of the next extent rec.
 980                  */
 981                 *cpos_end = le32_to_cpu(el->l_recs[index+1].e_cpos);
 982                 return 0;
 983         }
 984
 985         if (!eb || (eb && !eb->h_next_leaf_blk)) {
 986                 /*
 987                  * We are the last extent rec, so any high cpos should
 988                  * be stored in this leaf refcount block.
 989                  */
 990                 *cpos_end = UINT_MAX;
 991                 return 0;
 992         }
 993
 994         /*
 995          * If the extent block isn't the last one, we have to find
 996          * the subtree root between this extent block and the next
 997          * leaf extent block and get the corresponding e_cpos from
 998          * the subroot. Otherwise we may corrupt the b-tree.
 999          */
1000         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1001
1002         left_path = ocfs2_new_path_from_et(&et);
1003         if (!left_path) {
1004                 ret = -ENOMEM;
1005                 mlog_errno(ret);
1006                 goto out;
1007         }
1008
1009         cpos = le32_to_cpu(eb->h_list.l_recs[index].e_cpos);
1010         ret = ocfs2_find_path(ci, left_path, cpos);
1011         if (ret) {
1012                 mlog_errno(ret);
1013                 goto out;
1014         }
1015
1016         right_path = ocfs2_new_path_from_path(left_path);
1017         if (!right_path) {
1018                 ret = -ENOMEM;
1019                 mlog_errno(ret);
1020                 goto out;
1021         }
1022
1023         ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, &cpos);
1024         if (ret) {
1025                 mlog_errno(ret);
1026                 goto out;
1027         }
1028
1029         ret = ocfs2_find_path(ci, right_path, cpos);
1030         if (ret) {
1031                 mlog_errno(ret);
1032                 goto out;
1033         }
1034
1035         subtree_root = ocfs2_find_subtree_root(&et, left_path,
1036                                                right_path);
1037
1038         tmp_el = left_path->p_node[subtree_root].el;
1039         blkno = left_path->p_node[subtree_root+1].bh->b_blocknr;
1040         for (i = 0; i < le32_to_cpu(tmp_el->l_next_free_rec); i++) {
1041                 if (le64_to_cpu(tmp_el->l_recs[i].e_blkno) == blkno) {
1042                         *cpos_end = le32_to_cpu(tmp_el->l_recs[i+1].e_cpos);
1043                         break;
1044                 }
1045         }
1046
1047         BUG_ON(i == le32_to_cpu(tmp_el->l_next_free_rec));
1048
1049 out:
1050         ocfs2_free_path(left_path);
1051         ocfs2_free_path(right_path);
1052         return ret;
1053 }
1054
1055 /*
1056  * Given a cpos and len, try to find the refcount record which contains cpos.
1057  * 1. If cpos can be found in one refcount record, return the record.
1058  * 2. If cpos can't be found, return a fake record which start from cpos
1059  *    and end at a small value between cpos+len and start of the next record.
1060  *    This fake record has r_refcount = 0.
1061  */
1062 static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
1063                                   struct buffer_head *ref_root_bh,
1064                                   u64 cpos, unsigned int len,
1065                                   struct ocfs2_refcount_rec *ret_rec,
1066                                   int *index,
1067                                   struct buffer_head **ret_bh)
1068 {
1069         int ret = 0, i, found;
1070         u32 low_cpos, uninitialized_var(cpos_end);
1071         struct ocfs2_extent_list *el;
1072         struct ocfs2_extent_rec *rec = NULL;
1073         struct ocfs2_extent_block *eb = NULL;
1074         struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
1075         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1076         struct ocfs2_refcount_block *rb =
1077                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1078
1079         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
1080                 ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
1081                                               ret_rec, index);
1082                 *ret_bh = ref_root_bh;
1083                 get_bh(ref_root_bh);
1084                 return 0;
1085         }
1086
1087         el = &rb->rf_list;
1088         low_cpos = cpos & OCFS2_32BIT_POS_MASK;
1089
1090         if (el->l_tree_depth) {
1091                 ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
1092                 if (ret) {
1093                         mlog_errno(ret);
1094                         goto out;
1095                 }
1096
1097                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1098                 el = &eb->h_list;
1099
1100                 if (el->l_tree_depth) {
1101                         ocfs2_error(sb,
1102                         "refcount tree %llu has non zero tree "
1103                         "depth in leaf btree tree block %llu\n",
1104                         (unsigned long long)ocfs2_metadata_cache_owner(ci),
1105                         (unsigned long long)eb_bh->b_blocknr);
1106                         ret = -EROFS;
1107                         goto out;
1108                 }
1109         }
1110
1111         found = 0;
1112         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1113                 rec = &el->l_recs[i];
1114
1115                 if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
1116                         found = 1;
1117                         break;
1118                 }
1119         }
1120
1121         if (found) {
1122                 ret = ocfs2_get_refcount_cpos_end(ci, ref_root_bh,
1123                                                   eb, el, i, &cpos_end);
1124                 if (ret) {
1125                         mlog_errno(ret);
1126                         goto out;
1127                 }
1128
1129                 if (cpos_end < low_cpos + len)
1130                         len = cpos_end - low_cpos;
1131         }
1132
1133         ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
1134                                         &ref_leaf_bh);
1135         if (ret) {
1136                 mlog_errno(ret);
1137                 goto out;
1138         }
1139
1140         ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
1141                                       ret_rec, index);
1142         *ret_bh = ref_leaf_bh;
1143 out:
1144         brelse(eb_bh);
1145         return ret;
1146 }
1147
1148 enum ocfs2_ref_rec_contig {
1149         REF_CONTIG_NONE = 0,
1150         REF_CONTIG_LEFT,
1151         REF_CONTIG_RIGHT,
1152         REF_CONTIG_LEFTRIGHT,
1153 };
1154
1155 static enum ocfs2_ref_rec_contig
1156         ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
1157                                     int index)
1158 {
1159         if ((rb->rf_records.rl_recs[index].r_refcount ==
1160             rb->rf_records.rl_recs[index + 1].r_refcount) &&
1161             (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
1162             le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1163             le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1164                 return REF_CONTIG_RIGHT;
1165
1166         return REF_CONTIG_NONE;
1167 }
1168
1169 static enum ocfs2_ref_rec_contig
1170         ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1171                                   int index)
1172 {
1173         enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1174
1175         if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1176                 ret = ocfs2_refcount_rec_adjacent(rb, index);
1177
1178         if (index > 0) {
1179                 enum ocfs2_ref_rec_contig tmp;
1180
1181                 tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1182
1183                 if (tmp == REF_CONTIG_RIGHT) {
1184                         if (ret == REF_CONTIG_RIGHT)
1185                                 ret = REF_CONTIG_LEFTRIGHT;
1186                         else
1187                                 ret = REF_CONTIG_LEFT;
1188                 }
1189         }
1190
1191         return ret;
1192 }
1193
1194 static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1195                                            int index)
1196 {
1197         BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1198                rb->rf_records.rl_recs[index+1].r_refcount);
1199
1200         le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1201                      le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1202
1203         if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1204                 memmove(&rb->rf_records.rl_recs[index + 1],
1205                         &rb->rf_records.rl_recs[index + 2],
1206                         sizeof(struct ocfs2_refcount_rec) *
1207                         (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1208
1209         memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1210                0, sizeof(struct ocfs2_refcount_rec));
1211         le16_add_cpu(&rb->rf_records.rl_used, -1);
1212 }
1213
1214 /*
1215  * Merge the refcount rec if we are contiguous with the adjacent recs.
1216  */
1217 static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1218                                      int index)
1219 {
1220         enum ocfs2_ref_rec_contig contig =
1221                                 ocfs2_refcount_rec_contig(rb, index);
1222
1223         if (contig == REF_CONTIG_NONE)
1224                 return;
1225
1226         if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1227                 BUG_ON(index == 0);
1228                 index--;
1229         }
1230
1231         ocfs2_rotate_refcount_rec_left(rb, index);
1232
1233         if (contig == REF_CONTIG_LEFTRIGHT)
1234                 ocfs2_rotate_refcount_rec_left(rb, index);
1235 }
1236
1237 /*
1238  * Change the refcount indexed by "index" in ref_bh.
1239  * If refcount reaches 0, remove it.
1240  */
1241 static int ocfs2_change_refcount_rec(handle_t *handle,
1242                                      struct ocfs2_caching_info *ci,
1243                                      struct buffer_head *ref_leaf_bh,
1244                                      int index, int merge, int change)
1245 {
1246         int ret;
1247         struct ocfs2_refcount_block *rb =
1248                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1249         struct ocfs2_refcount_list *rl = &rb->rf_records;
1250         struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1251
1252         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1253                                       OCFS2_JOURNAL_ACCESS_WRITE);
1254         if (ret) {
1255                 mlog_errno(ret);
1256                 goto out;
1257         }
1258
1259         mlog(0, "change index %d, old count %u, change %d\n", index,
1260              le32_to_cpu(rec->r_refcount), change);
1261         le32_add_cpu(&rec->r_refcount, change);
1262
1263         if (!rec->r_refcount) {
1264                 if (index != le16_to_cpu(rl->rl_used) - 1) {
1265                         memmove(rec, rec + 1,
1266                                 (le16_to_cpu(rl->rl_used) - index - 1) *
1267                                 sizeof(struct ocfs2_refcount_rec));
1268                         memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
1269                                0, sizeof(struct ocfs2_refcount_rec));
1270                 }
1271
1272                 le16_add_cpu(&rl->rl_used, -1);
1273         } else if (merge)
1274                 ocfs2_refcount_rec_merge(rb, index);
1275
1276         ocfs2_journal_dirty(handle, ref_leaf_bh);
1277 out:
1278         return ret;
1279 }
1280
1281 static int ocfs2_expand_inline_ref_root(handle_t *handle,
1282                                         struct ocfs2_caching_info *ci,
1283                                         struct buffer_head *ref_root_bh,
1284                                         struct buffer_head **ref_leaf_bh,
1285                                         struct ocfs2_alloc_context *meta_ac)
1286 {
1287         int ret;
1288         u16 suballoc_bit_start;
1289         u32 num_got;
1290         u64 suballoc_loc, blkno;
1291         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1292         struct buffer_head *new_bh = NULL;
1293         struct ocfs2_refcount_block *new_rb;
1294         struct ocfs2_refcount_block *root_rb =
1295                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1296
1297         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1298                                       OCFS2_JOURNAL_ACCESS_WRITE);
1299         if (ret) {
1300                 mlog_errno(ret);
1301                 goto out;
1302         }
1303
1304         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
1305                                    &suballoc_bit_start, &num_got,
1306                                    &blkno);
1307         if (ret) {
1308                 mlog_errno(ret);
1309                 goto out;
1310         }
1311
1312         new_bh = sb_getblk(sb, blkno);
1313         if (new_bh == NULL) {
1314                 ret = -EIO;
1315                 mlog_errno(ret);
1316                 goto out;
1317         }
1318         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1319
1320         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1321                                       OCFS2_JOURNAL_ACCESS_CREATE);
1322         if (ret) {
1323                 mlog_errno(ret);
1324                 goto out;
1325         }
1326
1327         /*
1328          * Initialize ocfs2_refcount_block.
1329          * It should contain the same information as the old root.
1330          * so just memcpy it and change the corresponding field.
1331          */
1332         memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1333
1334         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1335         new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
1336         new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
1337         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1338         new_rb->rf_blkno = cpu_to_le64(blkno);
1339         new_rb->rf_cpos = cpu_to_le32(0);
1340         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1341         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1342         ocfs2_journal_dirty(handle, new_bh);
1343
1344         /* Now change the root. */
1345         memset(&root_rb->rf_list, 0, sb->s_blocksize -
1346                offsetof(struct ocfs2_refcount_block, rf_list));
1347         root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1348         root_rb->rf_clusters = cpu_to_le32(1);
1349         root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1350         root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1351         root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1352         root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1353
1354         ocfs2_journal_dirty(handle, ref_root_bh);
1355
1356         mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno,
1357              le16_to_cpu(new_rb->rf_records.rl_used));
1358
1359         *ref_leaf_bh = new_bh;
1360         new_bh = NULL;
1361 out:
1362         brelse(new_bh);
1363         return ret;
1364 }
1365
1366 static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1367                                            struct ocfs2_refcount_rec *next)
1368 {
1369         if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1370                 ocfs2_get_ref_rec_low_cpos(next))
1371                 return 1;
1372
1373         return 0;
1374 }
1375
1376 static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1377 {
1378         const struct ocfs2_refcount_rec *l = a, *r = b;
1379         u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1380         u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1381
1382         if (l_cpos > r_cpos)
1383                 return 1;
1384         if (l_cpos < r_cpos)
1385                 return -1;
1386         return 0;
1387 }
1388
1389 static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1390 {
1391         const struct ocfs2_refcount_rec *l = a, *r = b;
1392         u64 l_cpos = le64_to_cpu(l->r_cpos);
1393         u64 r_cpos = le64_to_cpu(r->r_cpos);
1394
1395         if (l_cpos > r_cpos)
1396                 return 1;
1397         if (l_cpos < r_cpos)
1398                 return -1;
1399         return 0;
1400 }
1401
1402 static void swap_refcount_rec(void *a, void *b, int size)
1403 {
1404         struct ocfs2_refcount_rec *l = a, *r = b, tmp;
1405
1406         tmp = *(struct ocfs2_refcount_rec *)l;
1407         *(struct ocfs2_refcount_rec *)l =
1408                         *(struct ocfs2_refcount_rec *)r;
1409         *(struct ocfs2_refcount_rec *)r = tmp;
1410 }
1411
1412 /*
1413  * The refcount cpos are ordered by their 64bit cpos,
1414  * But we will use the low 32 bit to be the e_cpos in the b-tree.
1415  * So we need to make sure that this pos isn't intersected with others.
1416  *
1417  * Note: The refcount block is already sorted by their low 32 bit cpos,
1418  *       So just try the middle pos first, and we will exit when we find
1419  *       the good position.
1420  */
1421 static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1422                                          u32 *split_pos, int *split_index)
1423 {
1424         int num_used = le16_to_cpu(rl->rl_used);
1425         int delta, middle = num_used / 2;
1426
1427         for (delta = 0; delta < middle; delta++) {
1428                 /* Let's check delta earlier than middle */
1429                 if (ocfs2_refcount_rec_no_intersect(
1430                                         &rl->rl_recs[middle - delta - 1],
1431                                         &rl->rl_recs[middle - delta])) {
1432                         *split_index = middle - delta;
1433                         break;
1434                 }
1435
1436                 /* For even counts, don't walk off the end */
1437                 if ((middle + delta + 1) == num_used)
1438                         continue;
1439
1440                 /* Now try delta past middle */
1441                 if (ocfs2_refcount_rec_no_intersect(
1442                                         &rl->rl_recs[middle + delta],
1443                                         &rl->rl_recs[middle + delta + 1])) {
1444                         *split_index = middle + delta + 1;
1445                         break;
1446                 }
1447         }
1448
1449         if (delta >= middle)
1450                 return -ENOSPC;
1451
1452         *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1453         return 0;
1454 }
1455
1456 static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1457                                             struct buffer_head *new_bh,
1458                                             u32 *split_cpos)
1459 {
1460         int split_index = 0, num_moved, ret;
1461         u32 cpos = 0;
1462         struct ocfs2_refcount_block *rb =
1463                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1464         struct ocfs2_refcount_list *rl = &rb->rf_records;
1465         struct ocfs2_refcount_block *new_rb =
1466                         (struct ocfs2_refcount_block *)new_bh->b_data;
1467         struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1468
1469         mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n",
1470              (unsigned long long)ref_leaf_bh->b_blocknr,
1471              le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used));
1472
1473         /*
1474          * XXX: Improvement later.
1475          * If we know all the high 32 bit cpos is the same, no need to sort.
1476          *
1477          * In order to make the whole process safe, we do:
1478          * 1. sort the entries by their low 32 bit cpos first so that we can
1479          *    find the split cpos easily.
1480          * 2. call ocfs2_insert_extent to insert the new refcount block.
1481          * 3. move the refcount rec to the new block.
1482          * 4. sort the entries by their 64 bit cpos.
1483          * 5. dirty the new_rb and rb.
1484          */
1485         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1486              sizeof(struct ocfs2_refcount_rec),
1487              cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1488
1489         ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1490         if (ret) {
1491                 mlog_errno(ret);
1492                 return ret;
1493         }
1494
1495         new_rb->rf_cpos = cpu_to_le32(cpos);
1496
1497         /* move refcount records starting from split_index to the new block. */
1498         num_moved = le16_to_cpu(rl->rl_used) - split_index;
1499         memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1500                num_moved * sizeof(struct ocfs2_refcount_rec));
1501
1502         /*ok, remove the entries we just moved over to the other block. */
1503         memset(&rl->rl_recs[split_index], 0,
1504                num_moved * sizeof(struct ocfs2_refcount_rec));
1505
1506         /* change old and new rl_used accordingly. */
1507         le16_add_cpu(&rl->rl_used, -num_moved);
1508         new_rl->rl_used = cpu_to_le16(num_moved);
1509
1510         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1511              sizeof(struct ocfs2_refcount_rec),
1512              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1513
1514         sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1515              sizeof(struct ocfs2_refcount_rec),
1516              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1517
1518         *split_cpos = cpos;
1519         return 0;
1520 }
1521
1522 static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1523                                          struct ocfs2_caching_info *ci,
1524                                          struct buffer_head *ref_root_bh,
1525                                          struct buffer_head *ref_leaf_bh,
1526                                          struct ocfs2_alloc_context *meta_ac)
1527 {
1528         int ret;
1529         u16 suballoc_bit_start;
1530         u32 num_got, new_cpos;
1531         u64 suballoc_loc, blkno;
1532         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1533         struct ocfs2_refcount_block *root_rb =
1534                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1535         struct buffer_head *new_bh = NULL;
1536         struct ocfs2_refcount_block *new_rb;
1537         struct ocfs2_extent_tree ref_et;
1538
1539         BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1540
1541         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1542                                       OCFS2_JOURNAL_ACCESS_WRITE);
1543         if (ret) {
1544                 mlog_errno(ret);
1545                 goto out;
1546         }
1547
1548         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1549                                       OCFS2_JOURNAL_ACCESS_WRITE);
1550         if (ret) {
1551                 mlog_errno(ret);
1552                 goto out;
1553         }
1554
1555         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
1556                                    &suballoc_bit_start, &num_got,
1557                                    &blkno);
1558         if (ret) {
1559                 mlog_errno(ret);
1560                 goto out;
1561         }
1562
1563         new_bh = sb_getblk(sb, blkno);
1564         if (new_bh == NULL) {
1565                 ret = -EIO;
1566                 mlog_errno(ret);
1567                 goto out;
1568         }
1569         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1570
1571         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1572                                       OCFS2_JOURNAL_ACCESS_CREATE);
1573         if (ret) {
1574                 mlog_errno(ret);
1575                 goto out;
1576         }
1577
1578         /* Initialize ocfs2_refcount_block. */
1579         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1580         memset(new_rb, 0, sb->s_blocksize);
1581         strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1582         new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
1583         new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
1584         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1585         new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1586         new_rb->rf_blkno = cpu_to_le64(blkno);
1587         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1588         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1589         new_rb->rf_records.rl_count =
1590                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1591         new_rb->rf_generation = root_rb->rf_generation;
1592
1593         ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1594         if (ret) {
1595                 mlog_errno(ret);
1596                 goto out;
1597         }
1598
1599         ocfs2_journal_dirty(handle, ref_leaf_bh);
1600         ocfs2_journal_dirty(handle, new_bh);
1601
1602         ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1603
1604         mlog(0, "insert new leaf block %llu at %u\n",
1605              (unsigned long long)new_bh->b_blocknr, new_cpos);
1606
1607         /* Insert the new leaf block with the specific offset cpos. */
1608         ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1609                                   1, 0, meta_ac);
1610         if (ret)
1611                 mlog_errno(ret);
1612
1613 out:
1614         brelse(new_bh);
1615         return ret;
1616 }
1617
1618 static int ocfs2_expand_refcount_tree(handle_t *handle,
1619                                       struct ocfs2_caching_info *ci,
1620                                       struct buffer_head *ref_root_bh,
1621                                       struct buffer_head *ref_leaf_bh,
1622                                       struct ocfs2_alloc_context *meta_ac)
1623 {
1624         int ret;
1625         struct buffer_head *expand_bh = NULL;
1626
1627         if (ref_root_bh == ref_leaf_bh) {
1628                 /*
1629                  * the old root bh hasn't been expanded to a b-tree,
1630                  * so expand it first.
1631                  */
1632                 ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1633                                                    &expand_bh, meta_ac);
1634                 if (ret) {
1635                         mlog_errno(ret);
1636                         goto out;
1637                 }
1638         } else {
1639                 expand_bh = ref_leaf_bh;
1640                 get_bh(expand_bh);
1641         }
1642
1643
1644         /* Now add a new refcount block into the tree.*/
1645         ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1646                                             expand_bh, meta_ac);
1647         if (ret)
1648                 mlog_errno(ret);
1649 out:
1650         brelse(expand_bh);
1651         return ret;
1652 }
1653
1654 /*
1655  * Adjust the extent rec in b-tree representing ref_leaf_bh.
1656  *
1657  * Only called when we have inserted a new refcount rec at index 0
1658  * which means ocfs2_extent_rec.e_cpos may need some change.
1659  */
1660 static int ocfs2_adjust_refcount_rec(handle_t *handle,
1661                                      struct ocfs2_caching_info *ci,
1662                                      struct buffer_head *ref_root_bh,
1663                                      struct buffer_head *ref_leaf_bh,
1664                                      struct ocfs2_refcount_rec *rec)
1665 {
1666         int ret = 0, i;
1667         u32 new_cpos, old_cpos;
1668         struct ocfs2_path *path = NULL;
1669         struct ocfs2_extent_tree et;
1670         struct ocfs2_refcount_block *rb =
1671                 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1672         struct ocfs2_extent_list *el;
1673
1674         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1675                 goto out;
1676
1677         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1678         old_cpos = le32_to_cpu(rb->rf_cpos);
1679         new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1680         if (old_cpos <= new_cpos)
1681                 goto out;
1682
1683         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1684
1685         path = ocfs2_new_path_from_et(&et);
1686         if (!path) {
1687                 ret = -ENOMEM;
1688                 mlog_errno(ret);
1689                 goto out;
1690         }
1691
1692         ret = ocfs2_find_path(ci, path, old_cpos);
1693         if (ret) {
1694                 mlog_errno(ret);
1695                 goto out;
1696         }
1697
1698         /*
1699          * 2 more credits, one for the leaf refcount block, one for
1700          * the extent block contains the extent rec.
1701          */
1702         ret = ocfs2_extend_trans(handle, 2);
1703         if (ret < 0) {
1704                 mlog_errno(ret);
1705                 goto out;
1706         }
1707
1708         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1709                                       OCFS2_JOURNAL_ACCESS_WRITE);
1710         if (ret < 0) {
1711                 mlog_errno(ret);
1712                 goto out;
1713         }
1714
1715         ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1716                                       OCFS2_JOURNAL_ACCESS_WRITE);
1717         if (ret < 0) {
1718                 mlog_errno(ret);
1719                 goto out;
1720         }
1721
1722         /* change the leaf extent block first. */
1723         el = path_leaf_el(path);
1724
1725         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1726                 if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1727                         break;
1728
1729         BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1730
1731         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1732
1733         /* change the r_cpos in the leaf block. */
1734         rb->rf_cpos = cpu_to_le32(new_cpos);
1735
1736         ocfs2_journal_dirty(handle, path_leaf_bh(path));
1737         ocfs2_journal_dirty(handle, ref_leaf_bh);
1738
1739 out:
1740         ocfs2_free_path(path);
1741         return ret;
1742 }
1743
1744 static int ocfs2_insert_refcount_rec(handle_t *handle,
1745                                      struct ocfs2_caching_info *ci,
1746                                      struct buffer_head *ref_root_bh,
1747                                      struct buffer_head *ref_leaf_bh,
1748                                      struct ocfs2_refcount_rec *rec,
1749                                      int index, int merge,
1750                                      struct ocfs2_alloc_context *meta_ac)
1751 {
1752         int ret;
1753         struct ocfs2_refcount_block *rb =
1754                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1755         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1756         struct buffer_head *new_bh = NULL;
1757
1758         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1759
1760         if (rf_list->rl_used == rf_list->rl_count) {
1761                 u64 cpos = le64_to_cpu(rec->r_cpos);
1762                 u32 len = le32_to_cpu(rec->r_clusters);
1763
1764                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1765                                                  ref_leaf_bh, meta_ac);
1766                 if (ret) {
1767                         mlog_errno(ret);
1768                         goto out;
1769                 }
1770
1771                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1772                                              cpos, len, NULL, &index,
1773                                              &new_bh);
1774                 if (ret) {
1775                         mlog_errno(ret);
1776                         goto out;
1777                 }
1778
1779                 ref_leaf_bh = new_bh;
1780                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1781                 rf_list = &rb->rf_records;
1782         }
1783
1784         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1785                                       OCFS2_JOURNAL_ACCESS_WRITE);
1786         if (ret) {
1787                 mlog_errno(ret);
1788                 goto out;
1789         }
1790
1791         if (index < le16_to_cpu(rf_list->rl_used))
1792                 memmove(&rf_list->rl_recs[index + 1],
1793                         &rf_list->rl_recs[index],
1794                         (le16_to_cpu(rf_list->rl_used) - index) *
1795                          sizeof(struct ocfs2_refcount_rec));
1796
1797         mlog(0, "insert refcount record start %llu, len %u, count %u "
1798              "to leaf block %llu at index %d\n",
1799              (unsigned long long)le64_to_cpu(rec->r_cpos),
1800              le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount),
1801              (unsigned long long)ref_leaf_bh->b_blocknr, index);
1802
1803         rf_list->rl_recs[index] = *rec;
1804
1805         le16_add_cpu(&rf_list->rl_used, 1);
1806
1807         if (merge)
1808                 ocfs2_refcount_rec_merge(rb, index);
1809
1810         ocfs2_journal_dirty(handle, ref_leaf_bh);
1811
1812         if (index == 0) {
1813                 ret = ocfs2_adjust_refcount_rec(handle, ci,
1814                                                 ref_root_bh,
1815                                                 ref_leaf_bh, rec);
1816                 if (ret)
1817                         mlog_errno(ret);
1818         }
1819 out:
1820         brelse(new_bh);
1821         return ret;
1822 }
1823
1824 /*
1825  * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1826  * This is much simple than our b-tree code.
1827  * split_rec is the new refcount rec we want to insert.
1828  * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1829  * increase refcount or decrease a refcount to non-zero).
1830  * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1831  * rec( in case we decrease a refcount to zero).
1832  */
1833 static int ocfs2_split_refcount_rec(handle_t *handle,
1834                                     struct ocfs2_caching_info *ci,
1835                                     struct buffer_head *ref_root_bh,
1836                                     struct buffer_head *ref_leaf_bh,
1837                                     struct ocfs2_refcount_rec *split_rec,
1838                                     int index, int merge,
1839                                     struct ocfs2_alloc_context *meta_ac,
1840                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
1841 {
1842         int ret, recs_need;
1843         u32 len;
1844         struct ocfs2_refcount_block *rb =
1845                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1846         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1847         struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1848         struct ocfs2_refcount_rec *tail_rec = NULL;
1849         struct buffer_head *new_bh = NULL;
1850
1851         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1852
1853         mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n",
1854              le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters),
1855              le64_to_cpu(split_rec->r_cpos),
1856              le32_to_cpu(split_rec->r_clusters));
1857
1858         /*
1859          * If we just need to split the header or tail clusters,
1860          * no more recs are needed, just split is OK.
1861          * Otherwise we at least need one new recs.
1862          */
1863         if (!split_rec->r_refcount &&
1864             (split_rec->r_cpos == orig_rec->r_cpos ||
1865              le64_to_cpu(split_rec->r_cpos) +
1866              le32_to_cpu(split_rec->r_clusters) ==
1867              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1868                 recs_need = 0;
1869         else
1870                 recs_need = 1;
1871
1872         /*
1873          * We need one more rec if we split in the middle and the new rec have
1874          * some refcount in it.
1875          */
1876         if (split_rec->r_refcount &&
1877             (split_rec->r_cpos != orig_rec->r_cpos &&
1878              le64_to_cpu(split_rec->r_cpos) +
1879              le32_to_cpu(split_rec->r_clusters) !=
1880              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1881                 recs_need++;
1882
1883         /* If the leaf block don't have enough record, expand it. */
1884         if (le16_to_cpu(rf_list->rl_used) + recs_need >
1885                                          le16_to_cpu(rf_list->rl_count)) {
1886                 struct ocfs2_refcount_rec tmp_rec;
1887                 u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1888                 len = le32_to_cpu(orig_rec->r_clusters);
1889                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1890                                                  ref_leaf_bh, meta_ac);
1891                 if (ret) {
1892                         mlog_errno(ret);
1893                         goto out;
1894                 }
1895
1896                 /*
1897                  * We have to re-get it since now cpos may be moved to
1898                  * another leaf block.
1899                  */
1900                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1901                                              cpos, len, &tmp_rec, &index,
1902                                              &new_bh);
1903                 if (ret) {
1904                         mlog_errno(ret);
1905                         goto out;
1906                 }
1907
1908                 ref_leaf_bh = new_bh;
1909                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1910                 rf_list = &rb->rf_records;
1911                 orig_rec = &rf_list->rl_recs[index];
1912         }
1913
1914         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1915                                       OCFS2_JOURNAL_ACCESS_WRITE);
1916         if (ret) {
1917                 mlog_errno(ret);
1918                 goto out;
1919         }
1920
1921         /*
1922          * We have calculated out how many new records we need and store
1923          * in recs_need, so spare enough space first by moving the records
1924          * after "index" to the end.
1925          */
1926         if (index != le16_to_cpu(rf_list->rl_used) - 1)
1927                 memmove(&rf_list->rl_recs[index + 1 + recs_need],
1928                         &rf_list->rl_recs[index + 1],
1929                         (le16_to_cpu(rf_list->rl_used) - index - 1) *
1930                          sizeof(struct ocfs2_refcount_rec));
1931
1932         len = (le64_to_cpu(orig_rec->r_cpos) +
1933               le32_to_cpu(orig_rec->r_clusters)) -
1934               (le64_to_cpu(split_rec->r_cpos) +
1935               le32_to_cpu(split_rec->r_clusters));
1936
1937         /*
1938          * If we have "len", the we will split in the tail and move it
1939          * to the end of the space we have just spared.
1940          */
1941         if (len) {
1942                 tail_rec = &rf_list->rl_recs[index + recs_need];
1943
1944                 memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1945                 le64_add_cpu(&tail_rec->r_cpos,
1946                              le32_to_cpu(tail_rec->r_clusters) - len);
1947                 tail_rec->r_clusters = cpu_to_le32(len);
1948         }
1949
1950         /*
1951          * If the split pos isn't the same as the original one, we need to
1952          * split in the head.
1953          *
1954          * Note: We have the chance that split_rec.r_refcount = 0,
1955          * recs_need = 0 and len > 0, which means we just cut the head from
1956          * the orig_rec and in that case we have done some modification in
1957          * orig_rec above, so the check for r_cpos is faked.
1958          */
1959         if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1960                 len = le64_to_cpu(split_rec->r_cpos) -
1961                       le64_to_cpu(orig_rec->r_cpos);
1962                 orig_rec->r_clusters = cpu_to_le32(len);
1963                 index++;
1964         }
1965
1966         le16_add_cpu(&rf_list->rl_used, recs_need);
1967
1968         if (split_rec->r_refcount) {
1969                 rf_list->rl_recs[index] = *split_rec;
1970                 mlog(0, "insert refcount record start %llu, len %u, count %u "
1971                      "to leaf block %llu at index %d\n",
1972                      (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1973                      le32_to_cpu(split_rec->r_clusters),
1974                      le32_to_cpu(split_rec->r_refcount),
1975                      (unsigned long long)ref_leaf_bh->b_blocknr, index);
1976
1977                 if (merge)
1978                         ocfs2_refcount_rec_merge(rb, index);
1979         }
1980
1981         ocfs2_journal_dirty(handle, ref_leaf_bh);
1982
1983 out:
1984         brelse(new_bh);
1985         return ret;
1986 }
1987
1988 static int __ocfs2_increase_refcount(handle_t *handle,
1989                                      struct ocfs2_caching_info *ci,
1990                                      struct buffer_head *ref_root_bh,
1991                                      u64 cpos, u32 len, int merge,
1992                                      struct ocfs2_alloc_context *meta_ac,
1993                                      struct ocfs2_cached_dealloc_ctxt *dealloc)
1994 {
1995         int ret = 0, index;
1996         struct buffer_head *ref_leaf_bh = NULL;
1997         struct ocfs2_refcount_rec rec;
1998         unsigned int set_len = 0;
1999
2000         mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n",
2001              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2002              (unsigned long long)cpos, len);
2003
2004         while (len) {
2005                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2006                                              cpos, len, &rec, &index,
2007                                              &ref_leaf_bh);
2008                 if (ret) {
2009                         mlog_errno(ret);
2010                         goto out;
2011                 }
2012
2013                 set_len = le32_to_cpu(rec.r_clusters);
2014
2015                 /*
2016                  * Here we may meet with 3 situations:
2017                  *
2018                  * 1. If we find an already existing record, and the length
2019                  *    is the same, cool, we just need to increase the r_refcount
2020                  *    and it is OK.
2021                  * 2. If we find a hole, just insert it with r_refcount = 1.
2022                  * 3. If we are in the middle of one extent record, split
2023                  *    it.
2024                  */
2025                 if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
2026                     set_len <= len) {
2027                         mlog(0, "increase refcount rec, start %llu, len %u, "
2028                              "count %u\n", (unsigned long long)cpos, set_len,
2029                              le32_to_cpu(rec.r_refcount));
2030                         ret = ocfs2_change_refcount_rec(handle, ci,
2031                                                         ref_leaf_bh, index,
2032                                                         merge, 1);
2033                         if (ret) {
2034                                 mlog_errno(ret);
2035                                 goto out;
2036                         }
2037                 } else if (!rec.r_refcount) {
2038                         rec.r_refcount = cpu_to_le32(1);
2039
2040                         mlog(0, "insert refcount rec, start %llu, len %u\n",
2041                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2042                              set_len);
2043                         ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
2044                                                         ref_leaf_bh,
2045                                                         &rec, index,
2046                                                         merge, meta_ac);
2047                         if (ret) {
2048                                 mlog_errno(ret);
2049                                 goto out;
2050                         }
2051                 } else  {
2052                         set_len = min((u64)(cpos + len),
2053                                       le64_to_cpu(rec.r_cpos) + set_len) - cpos;
2054                         rec.r_cpos = cpu_to_le64(cpos);
2055                         rec.r_clusters = cpu_to_le32(set_len);
2056                         le32_add_cpu(&rec.r_refcount, 1);
2057
2058                         mlog(0, "split refcount rec, start %llu, "
2059                              "len %u, count %u\n",
2060                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2061                              set_len, le32_to_cpu(rec.r_refcount));
2062                         ret = ocfs2_split_refcount_rec(handle, ci,
2063                                                        ref_root_bh, ref_leaf_bh,
2064                                                        &rec, index, merge,
2065                                                        meta_ac, dealloc);
2066                         if (ret) {
2067                                 mlog_errno(ret);
2068                                 goto out;
2069                         }
2070                 }
2071
2072                 cpos += set_len;
2073                 len -= set_len;
2074                 brelse(ref_leaf_bh);
2075                 ref_leaf_bh = NULL;
2076         }
2077
2078 out:
2079         brelse(ref_leaf_bh);
2080         return ret;
2081 }
2082
2083 static int ocfs2_remove_refcount_extent(handle_t *handle,
2084                                 struct ocfs2_caching_info *ci,
2085                                 struct buffer_head *ref_root_bh,
2086                                 struct buffer_head *ref_leaf_bh,
2087                                 struct ocfs2_alloc_context *meta_ac,
2088                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2089 {
2090         int ret;
2091         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2092         struct ocfs2_refcount_block *rb =
2093                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2094         struct ocfs2_extent_tree et;
2095
2096         BUG_ON(rb->rf_records.rl_used);
2097
2098         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2099         ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
2100                                   1, meta_ac, dealloc);
2101         if (ret) {
2102                 mlog_errno(ret);
2103                 goto out;
2104         }
2105
2106         ocfs2_remove_from_cache(ci, ref_leaf_bh);
2107
2108         /*
2109          * add the freed block to the dealloc so that it will be freed
2110          * when we run dealloc.
2111          */
2112         ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
2113                                         le16_to_cpu(rb->rf_suballoc_slot),
2114                                         le64_to_cpu(rb->rf_suballoc_loc),
2115                                         le64_to_cpu(rb->rf_blkno),
2116                                         le16_to_cpu(rb->rf_suballoc_bit));
2117         if (ret) {
2118                 mlog_errno(ret);
2119                 goto out;
2120         }
2121
2122         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
2123                                       OCFS2_JOURNAL_ACCESS_WRITE);
2124         if (ret) {
2125                 mlog_errno(ret);
2126                 goto out;
2127         }
2128
2129         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2130
2131         le32_add_cpu(&rb->rf_clusters, -1);
2132
2133         /*
2134          * check whether we need to restore the root refcount block if
2135          * there is no leaf extent block at atll.
2136          */
2137         if (!rb->rf_list.l_next_free_rec) {
2138                 BUG_ON(rb->rf_clusters);
2139
2140                 mlog(0, "reset refcount tree root %llu to be a record block.\n",
2141                      (unsigned long long)ref_root_bh->b_blocknr);
2142
2143                 rb->rf_flags = 0;
2144                 rb->rf_parent = 0;
2145                 rb->rf_cpos = 0;
2146                 memset(&rb->rf_records, 0, sb->s_blocksize -
2147                        offsetof(struct ocfs2_refcount_block, rf_records));
2148                 rb->rf_records.rl_count =
2149                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
2150         }
2151
2152         ocfs2_journal_dirty(handle, ref_root_bh);
2153
2154 out:
2155         return ret;
2156 }
2157
2158 int ocfs2_increase_refcount(handle_t *handle,
2159                             struct ocfs2_caching_info *ci,
2160                             struct buffer_head *ref_root_bh,
2161                             u64 cpos, u32 len,
2162                             struct ocfs2_alloc_context *meta_ac,
2163                             struct ocfs2_cached_dealloc_ctxt *dealloc)
2164 {
2165         return __ocfs2_increase_refcount(handle, ci, ref_root_bh,
2166                                          cpos, len, 1,
2167                                          meta_ac, dealloc);
2168 }
2169
2170 static int ocfs2_decrease_refcount_rec(handle_t *handle,
2171                                 struct ocfs2_caching_info *ci,
2172                                 struct buffer_head *ref_root_bh,
2173                                 struct buffer_head *ref_leaf_bh,
2174                                 int index, u64 cpos, unsigned int len,
2175                                 struct ocfs2_alloc_context *meta_ac,
2176                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2177 {
2178         int ret;
2179         struct ocfs2_refcount_block *rb =
2180                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2181         struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
2182
2183         BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
2184         BUG_ON(cpos + len >
2185                le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
2186
2187         if (cpos == le64_to_cpu(rec->r_cpos) &&
2188             len == le32_to_cpu(rec->r_clusters))
2189                 ret = ocfs2_change_refcount_rec(handle, ci,
2190                                                 ref_leaf_bh, index, 1, -1);
2191         else {
2192                 struct ocfs2_refcount_rec split = *rec;
2193                 split.r_cpos = cpu_to_le64(cpos);
2194                 split.r_clusters = cpu_to_le32(len);
2195
2196                 le32_add_cpu(&split.r_refcount, -1);
2197
2198                 mlog(0, "split refcount rec, start %llu, "
2199                      "len %u, count %u, original start %llu, len %u\n",
2200                      (unsigned long long)le64_to_cpu(split.r_cpos),
2201                      len, le32_to_cpu(split.r_refcount),
2202                      (unsigned long long)le64_to_cpu(rec->r_cpos),
2203                      le32_to_cpu(rec->r_clusters));
2204                 ret = ocfs2_split_refcount_rec(handle, ci,
2205                                                ref_root_bh, ref_leaf_bh,
2206                                                &split, index, 1,
2207                                                meta_ac, dealloc);
2208         }
2209
2210         if (ret) {
2211                 mlog_errno(ret);
2212                 goto out;
2213         }
2214
2215         /* Remove the leaf refcount block if it contains no refcount record. */
2216         if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
2217                 ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
2218                                                    ref_leaf_bh, meta_ac,
2219                                                    dealloc);
2220                 if (ret)
2221                         mlog_errno(ret);
2222         }
2223
2224 out:
2225         return ret;
2226 }
2227
2228 static int __ocfs2_decrease_refcount(handle_t *handle,
2229                                      struct ocfs2_caching_info *ci,
2230                                      struct buffer_head *ref_root_bh,
2231                                      u64 cpos, u32 len,
2232                                      struct ocfs2_alloc_context *meta_ac,
2233                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2234                                      int delete)
2235 {
2236         int ret = 0, index = 0;
2237         struct ocfs2_refcount_rec rec;
2238         unsigned int r_count = 0, r_len;
2239         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2240         struct buffer_head *ref_leaf_bh = NULL;
2241
2242         mlog(0, "Tree owner %llu, decrease refcount start %llu, "
2243              "len %u, delete %u\n",
2244              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2245              (unsigned long long)cpos, len, delete);
2246
2247         while (len) {
2248                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2249                                              cpos, len, &rec, &index,
2250                                              &ref_leaf_bh);
2251                 if (ret) {
2252                         mlog_errno(ret);
2253                         goto out;
2254                 }
2255
2256                 r_count = le32_to_cpu(rec.r_refcount);
2257                 BUG_ON(r_count == 0);
2258                 if (!delete)
2259                         BUG_ON(r_count > 1);
2260
2261                 r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
2262                               le32_to_cpu(rec.r_clusters)) - cpos;
2263
2264                 ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
2265                                                   ref_leaf_bh, index,
2266                                                   cpos, r_len,
2267                                                   meta_ac, dealloc);
2268                 if (ret) {
2269                         mlog_errno(ret);
2270                         goto out;
2271                 }
2272
2273                 if (le32_to_cpu(rec.r_refcount) == 1 && delete) {
2274                         ret = ocfs2_cache_cluster_dealloc(dealloc,
2275                                           ocfs2_clusters_to_blocks(sb, cpos),
2276                                                           r_len);
2277                         if (ret) {
2278                                 mlog_errno(ret);
2279                                 goto out;
2280                         }
2281                 }
2282
2283                 cpos += r_len;
2284                 len -= r_len;
2285                 brelse(ref_leaf_bh);
2286                 ref_leaf_bh = NULL;
2287         }
2288
2289 out:
2290         brelse(ref_leaf_bh);
2291         return ret;
2292 }
2293
2294 /* Caller must hold refcount tree lock. */
2295 int ocfs2_decrease_refcount(struct inode *inode,
2296                             handle_t *handle, u32 cpos, u32 len,
2297                             struct ocfs2_alloc_context *meta_ac,
2298                             struct ocfs2_cached_dealloc_ctxt *dealloc,
2299                             int delete)
2300 {
2301         int ret;
2302         u64 ref_blkno;
2303         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2304         struct buffer_head *ref_root_bh = NULL;
2305         struct ocfs2_refcount_tree *tree;
2306
2307         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2308
2309         ret = ocfs2_get_refcount_block(inode, &ref_blkno);
2310         if (ret) {
2311                 mlog_errno(ret);
2312                 goto out;
2313         }
2314
2315         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
2316         if (ret) {
2317                 mlog_errno(ret);
2318                 goto out;
2319         }
2320
2321         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
2322                                         &ref_root_bh);
2323         if (ret) {
2324                 mlog_errno(ret);
2325                 goto out;
2326         }
2327
2328         ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
2329                                         cpos, len, meta_ac, dealloc, delete);
2330         if (ret)
2331                 mlog_errno(ret);
2332 out:
2333         brelse(ref_root_bh);
2334         return ret;
2335 }
2336
2337 /*
2338  * Mark the already-existing extent at cpos as refcounted for len clusters.
2339  * This adds the refcount extent flag.
2340  *
2341  * If the existing extent is larger than the request, initiate a
2342  * split. An attempt will be made at merging with adjacent extents.
2343  *
2344  * The caller is responsible for passing down meta_ac if we'll need it.
2345  */
2346 static int ocfs2_mark_extent_refcounted(struct inode *inode,
2347                                 struct ocfs2_extent_tree *et,
2348                                 handle_t *handle, u32 cpos,
2349                                 u32 len, u32 phys,
2350                                 struct ocfs2_alloc_context *meta_ac,
2351                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2352 {
2353         int ret;
2354
2355         mlog(0, "Inode %lu refcount tree cpos %u, len %u, phys cluster %u\n",
2356              inode->i_ino, cpos, len, phys);
2357
2358         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2359                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2360                             "tree, but the feature bit is not set in the "
2361                             "super block.", inode->i_ino);
2362                 ret = -EROFS;
2363                 goto out;
2364         }
2365
2366         ret = ocfs2_change_extent_flag(handle, et, cpos,
2367                                        len, phys, meta_ac, dealloc,
2368                                        OCFS2_EXT_REFCOUNTED, 0);
2369         if (ret)
2370                 mlog_errno(ret);
2371
2372 out:
2373         return ret;
2374 }
2375
2376 /*
2377  * Given some contiguous physical clusters, calculate what we need
2378  * for modifying their refcount.
2379  */
2380 static int ocfs2_calc_refcount_meta_credits(struct super_block *sb,
2381                                             struct ocfs2_caching_info *ci,
2382                                             struct buffer_head *ref_root_bh,
2383                                             u64 start_cpos,
2384                                             u32 clusters,
2385                                             int *meta_add,
2386                                             int *credits)
2387 {
2388         int ret = 0, index, ref_blocks = 0, recs_add = 0;
2389         u64 cpos = start_cpos;
2390         struct ocfs2_refcount_block *rb;
2391         struct ocfs2_refcount_rec rec;
2392         struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL;
2393         u32 len;
2394
2395         mlog(0, "start_cpos %llu, clusters %u\n",
2396              (unsigned long long)start_cpos, clusters);
2397         while (clusters) {
2398                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2399                                              cpos, clusters, &rec,
2400                                              &index, &ref_leaf_bh);
2401                 if (ret) {
2402                         mlog_errno(ret);
2403                         goto out;
2404                 }
2405
2406                 if (ref_leaf_bh != prev_bh) {
2407                         /*
2408                          * Now we encounter a new leaf block, so calculate
2409                          * whether we need to extend the old leaf.
2410                          */
2411                         if (prev_bh) {
2412                                 rb = (struct ocfs2_refcount_block *)
2413                                                         prev_bh->b_data;
2414
2415                                 if (le64_to_cpu(rb->rf_records.rl_used) +
2416                                     recs_add >
2417                                     le16_to_cpu(rb->rf_records.rl_count))
2418                                         ref_blocks++;
2419                         }
2420
2421                         recs_add = 0;
2422                         *credits += 1;
2423                         brelse(prev_bh);
2424                         prev_bh = ref_leaf_bh;
2425                         get_bh(prev_bh);
2426                 }
2427
2428                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2429
2430                 mlog(0, "recs_add %d,cpos %llu, clusters %u, rec->r_cpos %llu,"
2431                      "rec->r_clusters %u, rec->r_refcount %u, index %d\n",
2432                      recs_add, (unsigned long long)cpos, clusters,
2433                      (unsigned long long)le64_to_cpu(rec.r_cpos),
2434                      le32_to_cpu(rec.r_clusters),
2435                      le32_to_cpu(rec.r_refcount), index);
2436
2437                 len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) +
2438                           le32_to_cpu(rec.r_clusters)) - cpos;
2439                 /*
2440                  * If the refcount rec already exist, cool. We just need
2441                  * to check whether there is a split. Otherwise we just need
2442                  * to increase the refcount.
2443                  * If we will insert one, increases recs_add.
2444                  *
2445                  * We record all the records which will be inserted to the
2446                  * same refcount block, so that we can tell exactly whether
2447                  * we need a new refcount block or not.
2448                  */
2449                 if (rec.r_refcount) {
2450                         /* Check whether we need a split at the beginning. */
2451                         if (cpos == start_cpos &&
2452                             cpos != le64_to_cpu(rec.r_cpos))
2453                                 recs_add++;
2454
2455                         /* Check whether we need a split in the end. */
2456                         if (cpos + clusters < le64_to_cpu(rec.r_cpos) +
2457                             le32_to_cpu(rec.r_clusters))
2458                                 recs_add++;
2459                 } else
2460                         recs_add++;
2461
2462                 brelse(ref_leaf_bh);
2463                 ref_leaf_bh = NULL;
2464                 clusters -= len;
2465                 cpos += len;
2466         }
2467
2468         if (prev_bh) {
2469                 rb = (struct ocfs2_refcount_block *)prev_bh->b_data;
2470
2471                 if (le64_to_cpu(rb->rf_records.rl_used) + recs_add >
2472                     le16_to_cpu(rb->rf_records.rl_count))
2473                         ref_blocks++;
2474
2475                 *credits += 1;
2476         }
2477
2478         if (!ref_blocks)
2479                 goto out;
2480
2481         mlog(0, "we need ref_blocks %d\n", ref_blocks);
2482         *meta_add += ref_blocks;
2483         *credits += ref_blocks;
2484
2485         /*
2486          * So we may need ref_blocks to insert into the tree.
2487          * That also means we need to change the b-tree and add that number
2488          * of records since we never merge them.
2489          * We need one more block for expansion since the new created leaf
2490          * block is also full and needs split.
2491          */
2492         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2493         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) {
2494                 struct ocfs2_extent_tree et;
2495
2496                 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2497                 *meta_add += ocfs2_extend_meta_needed(et.et_root_el);
2498                 *credits += ocfs2_calc_extend_credits(sb,
2499                                                       et.et_root_el,
2500                                                       ref_blocks);
2501         } else {
2502                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
2503                 *meta_add += 1;
2504         }
2505
2506 out:
2507         brelse(ref_leaf_bh);
2508         brelse(prev_bh);
2509         return ret;
2510 }
2511
2512 /*
2513  * For refcount tree, we will decrease some contiguous clusters
2514  * refcount count, so just go through it to see how many blocks
2515  * we gonna touch and whether we need to create new blocks.
2516  *
2517  * Normally the refcount blocks store these refcount should be
2518  * contiguous also, so that we can get the number easily.
2519  * We will at most add split 2 refcount records and 2 more
2520  * refcount blocks, so just check it in a rough way.
2521  *
2522  * Caller must hold refcount tree lock.
2523  */
2524 int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
2525                                           u64 refcount_loc,
2526                                           u64 phys_blkno,
2527                                           u32 clusters,
2528                                           int *credits,
2529                                           int *ref_blocks)
2530 {
2531         int ret;
2532         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2533         struct buffer_head *ref_root_bh = NULL;
2534         struct ocfs2_refcount_tree *tree;
2535         u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno);
2536
2537         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2538                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2539                             "tree, but the feature bit is not set in the "
2540                             "super block.", inode->i_ino);
2541                 ret = -EROFS;
2542                 goto out;
2543         }
2544
2545         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2546
2547         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb),
2548                                       refcount_loc, &tree);
2549         if (ret) {
2550                 mlog_errno(ret);
2551                 goto out;
2552         }
2553
2554         ret = ocfs2_read_refcount_block(&tree->rf_ci, refcount_loc,
2555                                         &ref_root_bh);
2556         if (ret) {
2557                 mlog_errno(ret);
2558                 goto out;
2559         }
2560
2561         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
2562                                                &tree->rf_ci,
2563                                                ref_root_bh,
2564                                                start_cpos, clusters,
2565                                                ref_blocks, credits);
2566         if (ret) {
2567                 mlog_errno(ret);
2568                 goto out;
2569         }
2570
2571         mlog(0, "reserve new metadata %d blocks, credits = %d\n",
2572              *ref_blocks, *credits);
2573
2574 out:
2575         brelse(ref_root_bh);
2576         return ret;
2577 }
2578
2579 #define MAX_CONTIG_BYTES        1048576
2580
2581 static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb)
2582 {
2583         return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES);
2584 }
2585
2586 static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb)
2587 {
2588         return ~(ocfs2_cow_contig_clusters(sb) - 1);
2589 }
2590
2591 /*
2592  * Given an extent that starts at 'start' and an I/O that starts at 'cpos',
2593  * find an offset (start + (n * contig_clusters)) that is closest to cpos
2594  * while still being less than or equal to it.
2595  *
2596  * The goal is to break the extent at a multiple of contig_clusters.
2597  */
2598 static inline unsigned int ocfs2_cow_align_start(struct super_block *sb,
2599                                                  unsigned int start,
2600                                                  unsigned int cpos)
2601 {
2602         BUG_ON(start > cpos);
2603
2604         return start + ((cpos - start) & ocfs2_cow_contig_mask(sb));
2605 }
2606
2607 /*
2608  * Given a cluster count of len, pad it out so that it is a multiple
2609  * of contig_clusters.
2610  */
2611 static inline unsigned int ocfs2_cow_align_length(struct super_block *sb,
2612                                                   unsigned int len)
2613 {
2614         unsigned int padded =
2615                 (len + (ocfs2_cow_contig_clusters(sb) - 1)) &
2616                 ocfs2_cow_contig_mask(sb);
2617
2618         /* Did we wrap? */
2619         if (padded < len)
2620                 padded = UINT_MAX;
2621
2622         return padded;
2623 }
2624
2625 /*
2626  * Calculate out the start and number of virtual clusters we need to to CoW.
2627  *
2628  * cpos is vitual start cluster position we want to do CoW in a
2629  * file and write_len is the cluster length.
2630  * max_cpos is the place where we want to stop CoW intentionally.
2631  *
2632  * Normal we will start CoW from the beginning of extent record cotaining cpos.
2633  * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we
2634  * get good I/O from the resulting extent tree.
2635  */
2636 static int ocfs2_refcount_cal_cow_clusters(struct inode *inode,
2637                                            struct ocfs2_extent_list *el,
2638                                            u32 cpos,
2639                                            u32 write_len,
2640                                            u32 max_cpos,
2641                                            u32 *cow_start,
2642                                            u32 *cow_len)
2643 {
2644         int ret = 0;
2645         int tree_height = le16_to_cpu(el->l_tree_depth), i;
2646         struct buffer_head *eb_bh = NULL;
2647         struct ocfs2_extent_block *eb = NULL;
2648         struct ocfs2_extent_rec *rec;
2649         unsigned int want_clusters, rec_end = 0;
2650         int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb);
2651         int leaf_clusters;
2652
2653         BUG_ON(cpos + write_len > max_cpos);
2654
2655         if (tree_height > 0) {
2656                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh);
2657                 if (ret) {
2658                         mlog_errno(ret);
2659                         goto out;
2660                 }
2661
2662                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2663                 el = &eb->h_list;
2664
2665                 if (el->l_tree_depth) {
2666                         ocfs2_error(inode->i_sb,
2667                                     "Inode %lu has non zero tree depth in "
2668                                     "leaf block %llu\n", inode->i_ino,
2669                                     (unsigned long long)eb_bh->b_blocknr);
2670                         ret = -EROFS;
2671                         goto out;
2672                 }
2673         }
2674
2675         *cow_len = 0;
2676         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
2677                 rec = &el->l_recs[i];
2678
2679                 if (ocfs2_is_empty_extent(rec)) {
2680                         mlog_bug_on_msg(i != 0, "Inode %lu has empty record in "
2681                                         "index %d\n", inode->i_ino, i);
2682                         continue;
2683                 }
2684
2685                 if (le32_to_cpu(rec->e_cpos) +
2686                     le16_to_cpu(rec->e_leaf_clusters) <= cpos)
2687                         continue;
2688
2689                 if (*cow_len == 0) {
2690                         /*
2691                          * We should find a refcounted record in the
2692                          * first pass.
2693                          */
2694                         BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED));
2695                         *cow_start = le32_to_cpu(rec->e_cpos);
2696                 }
2697
2698                 /*
2699                  * If we encounter a hole, a non-refcounted record or
2700                  * pass the max_cpos, stop the search.
2701                  */
2702                 if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) ||
2703                     (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) ||
2704                     (max_cpos <= le32_to_cpu(rec->e_cpos)))
2705                         break;
2706
2707                 leaf_clusters = le16_to_cpu(rec->e_leaf_clusters);
2708                 rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters;
2709                 if (rec_end > max_cpos) {
2710                         rec_end = max_cpos;
2711                         leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos);
2712                 }
2713
2714                 /*
2715                  * How many clusters do we actually need from
2716                  * this extent?  First we see how many we actually
2717                  * need to complete the write.  If that's smaller
2718                  * than contig_clusters, we try for contig_clusters.
2719                  */
2720                 if (!*cow_len)
2721                         want_clusters = write_len;
2722                 else
2723                         want_clusters = (cpos + write_len) -
2724                                 (*cow_start + *cow_len);
2725                 if (want_clusters < contig_clusters)
2726                         want_clusters = contig_clusters;
2727
2728                 /*
2729                  * If the write does not cover the whole extent, we
2730                  * need to calculate how we're going to split the extent.
2731                  * We try to do it on contig_clusters boundaries.
2732                  *
2733                  * Any extent smaller than contig_clusters will be
2734                  * CoWed in its entirety.
2735                  */
2736                 if (leaf_clusters <= contig_clusters)
2737                         *cow_len += leaf_clusters;
2738                 else if (*cow_len || (*cow_start == cpos)) {
2739                         /*
2740                          * This extent needs to be CoW'd from its
2741                          * beginning, so all we have to do is compute
2742                          * how many clusters to grab.  We align
2743                          * want_clusters to the edge of contig_clusters
2744                          * to get better I/O.
2745                          */
2746                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2747                                                                want_clusters);
2748
2749                         if (leaf_clusters < want_clusters)
2750                                 *cow_len += leaf_clusters;
2751                         else
2752                                 *cow_len += want_clusters;
2753                 } else if ((*cow_start + contig_clusters) >=
2754                            (cpos + write_len)) {
2755                         /*
2756                          * Breaking off contig_clusters at the front
2757                          * of the extent will cover our write.  That's
2758                          * easy.
2759                          */
2760                         *cow_len = contig_clusters;
2761                 } else if ((rec_end - cpos) <= contig_clusters) {
2762                         /*
2763                          * Breaking off contig_clusters at the tail of
2764                          * this extent will cover cpos.
2765                          */
2766                         *cow_start = rec_end - contig_clusters;
2767                         *cow_len = contig_clusters;
2768                 } else if ((rec_end - cpos) <= want_clusters) {
2769                         /*
2770                          * While we can't fit the entire write in this
2771                          * extent, we know that the write goes from cpos
2772                          * to the end of the extent.  Break that off.
2773                          * We try to break it at some multiple of
2774                          * contig_clusters from the front of the extent.
2775                          * Failing that (ie, cpos is within
2776                          * contig_clusters of the front), we'll CoW the
2777                          * entire extent.
2778                          */
2779                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2780                                                            *cow_start, cpos);
2781                         *cow_len = rec_end - *cow_start;
2782                 } else {
2783                         /*
2784                          * Ok, the entire write lives in the middle of
2785                          * this extent.  Let's try to slice the extent up
2786                          * nicely.  Optimally, our CoW region starts at
2787                          * m*contig_clusters from the beginning of the
2788                          * extent and goes for n*contig_clusters,
2789                          * covering the entire write.
2790                          */
2791                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2792                                                            *cow_start, cpos);
2793
2794                         want_clusters = (cpos + write_len) - *cow_start;
2795                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2796                                                                want_clusters);
2797                         if (*cow_start + want_clusters <= rec_end)
2798                                 *cow_len = want_clusters;
2799                         else
2800                                 *cow_len = rec_end - *cow_start;
2801                 }
2802
2803                 /* Have we covered our entire write yet? */
2804                 if ((*cow_start + *cow_len) >= (cpos + write_len))
2805                         break;
2806
2807                 /*
2808                  * If we reach the end of the extent block and don't get enough
2809                  * clusters, continue with the next extent block if possible.
2810                  */
2811                 if (i + 1 == le16_to_cpu(el->l_next_free_rec) &&
2812                     eb && eb->h_next_leaf_blk) {
2813                         brelse(eb_bh);
2814                         eb_bh = NULL;
2815
2816                         ret = ocfs2_read_extent_block(INODE_CACHE(inode),
2817                                                le64_to_cpu(eb->h_next_leaf_blk),
2818                                                &eb_bh);
2819                         if (ret) {
2820                                 mlog_errno(ret);
2821                                 goto out;
2822                         }
2823
2824                         eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2825                         el = &eb->h_list;
2826                         i = -1;
2827                 }
2828         }
2829
2830 out:
2831         brelse(eb_bh);
2832         return ret;
2833 }
2834
2835 /*
2836  * Prepare meta_ac, data_ac and calculate credits when we want to add some
2837  * num_clusters in data_tree "et" and change the refcount for the old
2838  * clusters(starting form p_cluster) in the refcount tree.
2839  *
2840  * Note:
2841  * 1. since we may split the old tree, so we at most will need num_clusters + 2
2842  *    more new leaf records.
2843  * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so
2844  *    just give data_ac = NULL.
2845  */
2846 static int ocfs2_lock_refcount_allocators(struct super_block *sb,
2847                                         u32 p_cluster, u32 num_clusters,
2848                                         struct ocfs2_extent_tree *et,
2849                                         struct ocfs2_caching_info *ref_ci,
2850                                         struct buffer_head *ref_root_bh,
2851                                         struct ocfs2_alloc_context **meta_ac,
2852                                         struct ocfs2_alloc_context **data_ac,
2853                                         int *credits)
2854 {
2855         int ret = 0, meta_add = 0;
2856         int num_free_extents = ocfs2_num_free_extents(OCFS2_SB(sb), et);
2857
2858         if (num_free_extents < 0) {
2859                 ret = num_free_extents;
2860                 mlog_errno(ret);
2861                 goto out;
2862         }
2863
2864         if (num_free_extents < num_clusters + 2)
2865                 meta_add =
2866                         ocfs2_extend_meta_needed(et->et_root_el);
2867
2868         *credits += ocfs2_calc_extend_credits(sb, et->et_root_el,
2869                                               num_clusters + 2);
2870
2871         ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh,
2872                                                p_cluster, num_clusters,
2873                                                &meta_add, credits);
2874         if (ret) {
2875                 mlog_errno(ret);
2876                 goto out;
2877         }
2878
2879         mlog(0, "reserve new metadata %d, clusters %u, credits = %d\n",
2880              meta_add, num_clusters, *credits);
2881         ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add,
2882                                                 meta_ac);
2883         if (ret) {
2884                 mlog_errno(ret);
2885                 goto out;
2886         }
2887
2888         if (data_ac) {
2889                 ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters,
2890                                              data_ac);
2891                 if (ret)
2892                         mlog_errno(ret);
2893         }
2894
2895 out:
2896         if (ret) {
2897                 if (*meta_ac) {
2898                         ocfs2_free_alloc_context(*meta_ac);
2899                         *meta_ac = NULL;
2900                 }
2901         }
2902
2903         return ret;
2904 }
2905
2906 static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh)
2907 {
2908         BUG_ON(buffer_dirty(bh));
2909
2910         clear_buffer_mapped(bh);
2911
2912         return 0;
2913 }
2914
2915 static int ocfs2_duplicate_clusters_by_page(handle_t *handle,
2916                                             struct ocfs2_cow_context *context,
2917                                             u32 cpos, u32 old_cluster,
2918                                             u32 new_cluster, u32 new_len)
2919 {
2920         int ret = 0, partial;
2921         struct ocfs2_caching_info *ci = context->data_et.et_ci;
2922         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2923         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
2924         struct page *page;
2925         pgoff_t page_index;
2926         unsigned int from, to;
2927         loff_t offset, end, map_end;
2928         struct address_space *mapping = context->inode->i_mapping;
2929
2930         mlog(0, "old_cluster %u, new %u, len %u at offset %u\n", old_cluster,
2931              new_cluster, new_len, cpos);
2932
2933         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
2934         end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
2935
2936         while (offset < end) {
2937                 page_index = offset >> PAGE_CACHE_SHIFT;
2938                 map_end = ((loff_t)page_index + 1) << PAGE_CACHE_SHIFT;
2939                 if (map_end > end)
2940                         map_end = end;
2941
2942                 /* from, to is the offset within the page. */
2943                 from = offset & (PAGE_CACHE_SIZE - 1);
2944                 to = PAGE_CACHE_SIZE;
2945                 if (map_end & (PAGE_CACHE_SIZE - 1))
2946                         to = map_end & (PAGE_CACHE_SIZE - 1);
2947
2948                 page = grab_cache_page(mapping, page_index);
2949
2950                 /*
2951                  * In case PAGE_CACHE_SIZE <= CLUSTER_SIZE, This page
2952                  * can't be dirtied before we CoW it out.
2953                  */
2954                 if (PAGE_CACHE_SIZE <= OCFS2_SB(sb)->s_clustersize)
2955                         BUG_ON(PageDirty(page));
2956
2957                 if (!PageUptodate(page)) {
2958                         ret = block_read_full_page(page, ocfs2_get_block);
2959                         if (ret) {
2960                                 mlog_errno(ret);
2961                                 goto unlock;
2962                         }
2963                         lock_page(page);
2964                 }
2965
2966                 if (page_has_buffers(page)) {
2967                         ret = walk_page_buffers(handle, page_buffers(page),
2968                                                 from, to, &partial,
2969                                                 ocfs2_clear_cow_buffer);
2970                         if (ret) {
2971                                 mlog_errno(ret);
2972                                 goto unlock;
2973                         }
2974                 }
2975
2976                 ocfs2_map_and_dirty_page(context->inode,
2977                                          handle, from, to,
2978                                          page, 0, &new_block);
2979                 mark_page_accessed(page);
2980 unlock:
2981                 unlock_page(page);
2982                 page_cache_release(page);
2983                 page = NULL;
2984                 offset = map_end;
2985                 if (ret)
2986                         break;
2987         }
2988
2989         return ret;
2990 }
2991
2992 static int ocfs2_duplicate_clusters_by_jbd(handle_t *handle,
2993                                            struct ocfs2_cow_context *context,
2994                                            u32 cpos, u32 old_cluster,
2995                                            u32 new_cluster, u32 new_len)
2996 {
2997         int ret = 0;
2998         struct super_block *sb = context->inode->i_sb;
2999         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3000         int i, blocks = ocfs2_clusters_to_blocks(sb, new_len);
3001         u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster);
3002         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
3003         struct ocfs2_super *osb = OCFS2_SB(sb);
3004         struct buffer_head *old_bh = NULL;
3005         struct buffer_head *new_bh = NULL;
3006
3007         mlog(0, "old_cluster %u, new %u, len %u\n", old_cluster,
3008              new_cluster, new_len);
3009
3010         for (i = 0; i < blocks; i++, old_block++, new_block++) {
3011                 new_bh = sb_getblk(osb->sb, new_block);
3012                 if (new_bh == NULL) {
3013                         ret = -EIO;
3014                         mlog_errno(ret);
3015                         break;
3016                 }
3017
3018                 ocfs2_set_new_buffer_uptodate(ci, new_bh);
3019
3020                 ret = ocfs2_read_block(ci, old_block, &old_bh, NULL);
3021                 if (ret) {
3022                         mlog_errno(ret);
3023                         break;
3024                 }
3025
3026                 ret = ocfs2_journal_access(handle, ci, new_bh,
3027                                            OCFS2_JOURNAL_ACCESS_CREATE);
3028                 if (ret) {
3029                         mlog_errno(ret);
3030                         break;
3031                 }
3032
3033                 memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize);
3034                 ocfs2_journal_dirty(handle, new_bh);
3035
3036                 brelse(new_bh);
3037                 brelse(old_bh);
3038                 new_bh = NULL;
3039                 old_bh = NULL;
3040         }
3041
3042         brelse(new_bh);
3043         brelse(old_bh);
3044         return ret;
3045 }
3046
3047 static int ocfs2_clear_ext_refcount(handle_t *handle,
3048                                     struct ocfs2_extent_tree *et,
3049                                     u32 cpos, u32 p_cluster, u32 len,
3050                                     unsigned int ext_flags,
3051                                     struct ocfs2_alloc_context *meta_ac,
3052                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
3053 {
3054         int ret, index;
3055         struct ocfs2_extent_rec replace_rec;
3056         struct ocfs2_path *path = NULL;
3057         struct ocfs2_extent_list *el;
3058         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
3059         u64 ino = ocfs2_metadata_cache_owner(et->et_ci);
3060
3061         mlog(0, "inode %llu cpos %u, len %u, p_cluster %u, ext_flags %u\n",
3062              (unsigned long long)ino, cpos, len, p_cluster, ext_flags);
3063
3064         memset(&replace_rec, 0, sizeof(replace_rec));
3065         replace_rec.e_cpos = cpu_to_le32(cpos);
3066         replace_rec.e_leaf_clusters = cpu_to_le16(len);
3067         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb,
3068                                                                    p_cluster));
3069         replace_rec.e_flags = ext_flags;
3070         replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED;
3071
3072         path = ocfs2_new_path_from_et(et);
3073         if (!path) {
3074                 ret = -ENOMEM;
3075                 mlog_errno(ret);
3076                 goto out;
3077         }
3078
3079         ret = ocfs2_find_path(et->et_ci, path, cpos);
3080         if (ret) {
3081                 mlog_errno(ret);
3082                 goto out;
3083         }
3084
3085         el = path_leaf_el(path);
3086
3087         index = ocfs2_search_extent_list(el, cpos);
3088         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
3089                 ocfs2_error(sb,
3090                             "Inode %llu has an extent at cpos %u which can no "
3091                             "longer be found.\n",
3092                             (unsigned long long)ino, cpos);
3093                 ret = -EROFS;
3094                 goto out;
3095         }
3096
3097         ret = ocfs2_split_extent(handle, et, path, index,
3098                                  &replace_rec, meta_ac, dealloc);
3099         if (ret)
3100                 mlog_errno(ret);
3101
3102 out:
3103         ocfs2_free_path(path);
3104         return ret;
3105 }
3106
3107 static int ocfs2_replace_clusters(handle_t *handle,
3108                                   struct ocfs2_cow_context *context,
3109                                   u32 cpos, u32 old,
3110                                   u32 new, u32 len,
3111                                   unsigned int ext_flags)
3112 {
3113         int ret;
3114         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3115         u64 ino = ocfs2_metadata_cache_owner(ci);
3116
3117         mlog(0, "inode %llu, cpos %u, old %u, new %u, len %u, ext_flags %u\n",
3118              (unsigned long long)ino, cpos, old, new, len, ext_flags);
3119
3120         /*If the old clusters is unwritten, no need to duplicate. */
3121         if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) {
3122                 ret = context->cow_duplicate_clusters(handle, context, cpos,
3123                                                       old, new, len);
3124                 if (ret) {
3125                         mlog_errno(ret);
3126                         goto out;
3127                 }
3128         }
3129
3130         ret = ocfs2_clear_ext_refcount(handle, &context->data_et,
3131                                        cpos, new, len, ext_flags,
3132                                        context->meta_ac, &context->dealloc);
3133         if (ret)
3134                 mlog_errno(ret);
3135 out:
3136         return ret;
3137 }
3138
3139 static int ocfs2_cow_sync_writeback(struct super_block *sb,
3140                                     struct ocfs2_cow_context *context,
3141                                     u32 cpos, u32 num_clusters)
3142 {
3143         int ret = 0;
3144         loff_t offset, end, map_end;
3145         pgoff_t page_index;
3146         struct page *page;
3147
3148         if (ocfs2_should_order_data(context->inode))
3149                 return 0;
3150
3151         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
3152         end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits);
3153
3154         ret = filemap_fdatawrite_range(context->inode->i_mapping,
3155                                        offset, end - 1);
3156         if (ret < 0) {
3157                 mlog_errno(ret);
3158                 return ret;
3159         }
3160
3161         while (offset < end) {
3162                 page_index = offset >> PAGE_CACHE_SHIFT;
3163                 map_end = ((loff_t)page_index + 1) << PAGE_CACHE_SHIFT;
3164                 if (map_end > end)
3165                         map_end = end;
3166
3167                 page = grab_cache_page(context->inode->i_mapping, page_index);
3168                 BUG_ON(!page);
3169
3170                 wait_on_page_writeback(page);
3171                 if (PageError(page)) {
3172                         ret = -EIO;
3173                         mlog_errno(ret);
3174                 } else
3175                         mark_page_accessed(page);
3176
3177                 unlock_page(page);
3178                 page_cache_release(page);
3179                 page = NULL;
3180                 offset = map_end;
3181                 if (ret)
3182                         break;
3183         }
3184
3185         return ret;
3186 }
3187
3188 static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context,
3189                                  u32 v_cluster, u32 *p_cluster,
3190                                  u32 *num_clusters,
3191                                  unsigned int *extent_flags)
3192 {
3193         return ocfs2_get_clusters(context->inode, v_cluster, p_cluster,
3194                                   num_clusters, extent_flags);
3195 }
3196
3197 static int ocfs2_make_clusters_writable(struct super_block *sb,
3198                                         struct ocfs2_cow_context *context,
3199                                         u32 cpos, u32 p_cluster,
3200                                         u32 num_clusters, unsigned int e_flags)
3201 {
3202         int ret, delete, index, credits =  0;
3203         u32 new_bit, new_len;
3204         unsigned int set_len;
3205         struct ocfs2_super *osb = OCFS2_SB(sb);
3206         handle_t *handle;
3207         struct buffer_head *ref_leaf_bh = NULL;
3208         struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci;
3209         struct ocfs2_refcount_rec rec;
3210
3211         mlog(0, "cpos %u, p_cluster %u, num_clusters %u, e_flags %u\n",
3212              cpos, p_cluster, num_clusters, e_flags);
3213
3214         ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters,
3215                                              &context->data_et,
3216                                              ref_ci,
3217                                              context->ref_root_bh,
3218                                              &context->meta_ac,
3219                                              &context->data_ac, &credits);
3220         if (ret) {
3221                 mlog_errno(ret);
3222                 return ret;
3223         }
3224
3225         if (context->post_refcount)
3226                 credits += context->post_refcount->credits;
3227
3228         credits += context->extra_credits;
3229         handle = ocfs2_start_trans(osb, credits);
3230         if (IS_ERR(handle)) {
3231                 ret = PTR_ERR(handle);
3232                 mlog_errno(ret);
3233                 goto out;
3234         }
3235
3236         while (num_clusters) {
3237                 ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh,
3238                                              p_cluster, num_clusters,
3239                                              &rec, &index, &ref_leaf_bh);
3240                 if (ret) {
3241                         mlog_errno(ret);
3242                         goto out_commit;
3243                 }
3244
3245                 BUG_ON(!rec.r_refcount);
3246                 set_len = min((u64)p_cluster + num_clusters,
3247                               le64_to_cpu(rec.r_cpos) +
3248                               le32_to_cpu(rec.r_clusters)) - p_cluster;
3249
3250                 /*
3251                  * There are many different situation here.
3252                  * 1. If refcount == 1, remove the flag and don't COW.
3253                  * 2. If refcount > 1, allocate clusters.
3254                  *    Here we may not allocate r_len once at a time, so continue
3255                  *    until we reach num_clusters.
3256                  */
3257                 if (le32_to_cpu(rec.r_refcount) == 1) {
3258                         delete = 0;
3259                         ret = ocfs2_clear_ext_refcount(handle,
3260                                                        &context->data_et,
3261                                                        cpos, p_cluster,
3262                                                        set_len, e_flags,
3263                                                        context->meta_ac,
3264                                                        &context->dealloc);
3265                         if (ret) {
3266                                 mlog_errno(ret);
3267                                 goto out_commit;
3268                         }
3269                 } else {
3270                         delete = 1;
3271
3272                         ret = __ocfs2_claim_clusters(handle,
3273                                                      context->data_ac,
3274                                                      1, set_len,
3275                                                      &new_bit, &new_len);
3276                         if (ret) {
3277                                 mlog_errno(ret);
3278                                 goto out_commit;
3279                         }
3280
3281                         ret = ocfs2_replace_clusters(handle, context,
3282                                                      cpos, p_cluster, new_bit,
3283                                                      new_len, e_flags);
3284                         if (ret) {
3285                                 mlog_errno(ret);
3286                                 goto out_commit;
3287                         }
3288                         set_len = new_len;
3289                 }
3290
3291                 ret = __ocfs2_decrease_refcount(handle, ref_ci,
3292                                                 context->ref_root_bh,
3293                                                 p_cluster, set_len,
3294                                                 context->meta_ac,
3295                                                 &context->dealloc, delete);
3296                 if (ret) {
3297                         mlog_errno(ret);
3298                         goto out_commit;
3299                 }
3300
3301                 cpos += set_len;
3302                 p_cluster += set_len;
3303                 num_clusters -= set_len;
3304                 brelse(ref_leaf_bh);
3305                 ref_leaf_bh = NULL;
3306         }
3307
3308         /* handle any post_cow action. */
3309         if (context->post_refcount && context->post_refcount->func) {
3310                 ret = context->post_refcount->func(context->inode, handle,
3311                                                 context->post_refcount->para);
3312                 if (ret) {
3313                         mlog_errno(ret);
3314                         goto out_commit;
3315                 }
3316         }
3317
3318         /*
3319          * Here we should write the new page out first if we are
3320          * in write-back mode.
3321          */
3322         if (context->get_clusters == ocfs2_di_get_clusters) {
3323                 ret = ocfs2_cow_sync_writeback(sb, context, cpos, num_clusters);
3324                 if (ret)
3325                         mlog_errno(ret);
3326         }
3327
3328 out_commit:
3329         ocfs2_commit_trans(osb, handle);
3330
3331 out:
3332         if (context->data_ac) {
3333                 ocfs2_free_alloc_context(context->data_ac);
3334                 context->data_ac = NULL;
3335         }
3336         if (context->meta_ac) {
3337                 ocfs2_free_alloc_context(context->meta_ac);
3338                 context->meta_ac = NULL;
3339         }
3340         brelse(ref_leaf_bh);
3341
3342         return ret;
3343 }
3344
3345 static int ocfs2_replace_cow(struct ocfs2_cow_context *context)
3346 {
3347         int ret = 0;
3348         struct inode *inode = context->inode;
3349         u32 cow_start = context->cow_start, cow_len = context->cow_len;
3350         u32 p_cluster, num_clusters;
3351         unsigned int ext_flags;
3352         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3353
3354         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
3355                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
3356                             "tree, but the feature bit is not set in the "
3357                             "super block.", inode->i_ino);
3358                 return -EROFS;
3359         }
3360
3361         ocfs2_init_dealloc_ctxt(&context->dealloc);
3362
3363         while (cow_len) {
3364                 ret = context->get_clusters(context, cow_start, &p_cluster,
3365                                             &num_clusters, &ext_flags);
3366                 if (ret) {
3367                         mlog_errno(ret);
3368                         break;
3369                 }
3370
3371                 BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED));
3372
3373                 if (cow_len < num_clusters)
3374                         num_clusters = cow_len;
3375
3376                 ret = ocfs2_make_clusters_writable(inode->i_sb, context,
3377                                                    cow_start, p_cluster,
3378                                                    num_clusters, ext_flags);
3379                 if (ret) {
3380                         mlog_errno(ret);
3381                         break;
3382                 }
3383
3384                 cow_len -= num_clusters;
3385                 cow_start += num_clusters;
3386         }
3387
3388         if (ocfs2_dealloc_has_cluster(&context->dealloc)) {
3389                 ocfs2_schedule_truncate_log_flush(osb, 1);
3390                 ocfs2_run_deallocs(osb, &context->dealloc);
3391         }
3392
3393         return ret;
3394 }
3395
3396 /*
3397  * Starting at cpos, try to CoW write_len clusters.  Don't CoW
3398  * past max_cpos.  This will stop when it runs into a hole or an
3399  * unrefcounted extent.
3400  */
3401 static int ocfs2_refcount_cow_hunk(struct inode *inode,
3402                                    struct buffer_head *di_bh,
3403                                    u32 cpos, u32 write_len, u32 max_cpos)
3404 {
3405         int ret;
3406         u32 cow_start = 0, cow_len = 0;
3407         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3408         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3409         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3410         struct buffer_head *ref_root_bh = NULL;
3411         struct ocfs2_refcount_tree *ref_tree;
3412         struct ocfs2_cow_context *context = NULL;
3413
3414         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3415
3416         ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list,
3417                                               cpos, write_len, max_cpos,
3418                                               &cow_start, &cow_len);
3419         if (ret) {
3420                 mlog_errno(ret);
3421                 goto out;
3422         }
3423
3424         mlog(0, "CoW inode %lu, cpos %u, write_len %u, cow_start %u, "
3425              "cow_len %u\n", inode->i_ino,
3426              cpos, write_len, cow_start, cow_len);
3427
3428         BUG_ON(cow_len == 0);
3429
3430         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3431         if (!context) {
3432                 ret = -ENOMEM;
3433                 mlog_errno(ret);
3434                 goto out;
3435         }
3436
3437         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
3438                                        1, &ref_tree, &ref_root_bh);
3439         if (ret) {
3440                 mlog_errno(ret);
3441                 goto out;
3442         }
3443
3444         context->inode = inode;
3445         context->cow_start = cow_start;
3446         context->cow_len = cow_len;
3447         context->ref_tree = ref_tree;
3448         context->ref_root_bh = ref_root_bh;
3449         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
3450         context->get_clusters = ocfs2_di_get_clusters;
3451
3452         ocfs2_init_dinode_extent_tree(&context->data_et,
3453                                       INODE_CACHE(inode), di_bh);
3454
3455         ret = ocfs2_replace_cow(context);
3456         if (ret)
3457                 mlog_errno(ret);
3458
3459         /*
3460          * truncate the extent map here since no matter whether we meet with
3461          * any error during the action, we shouldn't trust cached extent map
3462          * any more.
3463          */
3464         ocfs2_extent_map_trunc(inode, cow_start);
3465
3466         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3467         brelse(ref_root_bh);
3468 out:
3469         kfree(context);
3470         return ret;
3471 }
3472
3473 /*
3474  * CoW any and all clusters between cpos and cpos+write_len.
3475  * Don't CoW past max_cpos.  If this returns successfully, all
3476  * clusters between cpos and cpos+write_len are safe to modify.
3477  */
3478 int ocfs2_refcount_cow(struct inode *inode,
3479                        struct buffer_head *di_bh,
3480                        u32 cpos, u32 write_len, u32 max_cpos)
3481 {
3482         int ret = 0;
3483         u32 p_cluster, num_clusters;
3484         unsigned int ext_flags;
3485
3486         while (write_len) {
3487                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3488                                          &num_clusters, &ext_flags);
3489                 if (ret) {
3490                         mlog_errno(ret);
3491                         break;
3492                 }
3493
3494                 if (write_len < num_clusters)
3495                         num_clusters = write_len;
3496
3497                 if (ext_flags & OCFS2_EXT_REFCOUNTED) {
3498                         ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos,
3499                                                       num_clusters, max_cpos);
3500                         if (ret) {
3501                                 mlog_errno(ret);
3502                                 break;
3503                         }
3504                 }
3505
3506                 write_len -= num_clusters;
3507                 cpos += num_clusters;
3508         }
3509
3510         return ret;
3511 }
3512
3513 static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context,
3514                                           u32 v_cluster, u32 *p_cluster,
3515                                           u32 *num_clusters,
3516                                           unsigned int *extent_flags)
3517 {
3518         struct inode *inode = context->inode;
3519         struct ocfs2_xattr_value_root *xv = context->cow_object;
3520
3521         return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster,
3522                                         num_clusters, &xv->xr_list,
3523                                         extent_flags);
3524 }
3525
3526 /*
3527  * Given a xattr value root, calculate the most meta/credits we need for
3528  * refcount tree change if we truncate it to 0.
3529  */
3530 int ocfs2_refcounted_xattr_delete_need(struct inode *inode,
3531                                        struct ocfs2_caching_info *ref_ci,
3532                                        struct buffer_head *ref_root_bh,
3533                                        struct ocfs2_xattr_value_root *xv,
3534                                        int *meta_add, int *credits)
3535 {
3536         int ret = 0, index, ref_blocks = 0;
3537         u32 p_cluster, num_clusters;
3538         u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters);
3539         struct ocfs2_refcount_block *rb;
3540         struct ocfs2_refcount_rec rec;
3541         struct buffer_head *ref_leaf_bh = NULL;
3542
3543         while (cpos < clusters) {
3544                 ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster,
3545                                                &num_clusters, &xv->xr_list,
3546                                                NULL);
3547                 if (ret) {
3548                         mlog_errno(ret);
3549                         goto out;
3550                 }
3551
3552                 cpos += num_clusters;
3553
3554                 while (num_clusters) {
3555                         ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh,
3556                                                      p_cluster, num_clusters,
3557                                                      &rec, &index,
3558                                                      &ref_leaf_bh);
3559                         if (ret) {
3560                                 mlog_errno(ret);
3561                                 goto out;
3562                         }
3563
3564                         BUG_ON(!rec.r_refcount);
3565
3566                         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
3567
3568                         /*
3569                          * We really don't know whether the other clusters is in
3570                          * this refcount block or not, so just take the worst
3571                          * case that all the clusters are in this block and each
3572                          * one will split a refcount rec, so totally we need
3573                          * clusters * 2 new refcount rec.
3574                          */
3575                         if (le64_to_cpu(rb->rf_records.rl_used) + clusters * 2 >
3576                             le16_to_cpu(rb->rf_records.rl_count))
3577                                 ref_blocks++;
3578
3579                         *credits += 1;
3580                         brelse(ref_leaf_bh);
3581                         ref_leaf_bh = NULL;
3582
3583                         if (num_clusters <= le32_to_cpu(rec.r_clusters))
3584                                 break;
3585                         else
3586                                 num_clusters -= le32_to_cpu(rec.r_clusters);
3587                         p_cluster += num_clusters;
3588                 }
3589         }
3590
3591         *meta_add += ref_blocks;
3592         if (!ref_blocks)
3593                 goto out;
3594
3595         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
3596         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)
3597                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
3598         else {
3599                 struct ocfs2_extent_tree et;
3600
3601                 ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh);
3602                 *credits += ocfs2_calc_extend_credits(inode->i_sb,
3603                                                       et.et_root_el,
3604                                                       ref_blocks);
3605         }
3606
3607 out:
3608         brelse(ref_leaf_bh);
3609         return ret;
3610 }
3611
3612 /*
3613  * Do CoW for xattr.
3614  */
3615 int ocfs2_refcount_cow_xattr(struct inode *inode,
3616                              struct ocfs2_dinode *di,
3617                              struct ocfs2_xattr_value_buf *vb,
3618                              struct ocfs2_refcount_tree *ref_tree,
3619                              struct buffer_head *ref_root_bh,
3620                              u32 cpos, u32 write_len,
3621                              struct ocfs2_post_refcount *post)
3622 {
3623         int ret;
3624         struct ocfs2_xattr_value_root *xv = vb->vb_xv;
3625         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3626         struct ocfs2_cow_context *context = NULL;
3627         u32 cow_start, cow_len;
3628
3629         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3630
3631         ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list,
3632                                               cpos, write_len, UINT_MAX,
3633                                               &cow_start, &cow_len);
3634         if (ret) {
3635                 mlog_errno(ret);
3636                 goto out;
3637         }
3638
3639         BUG_ON(cow_len == 0);
3640
3641         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3642         if (!context) {
3643                 ret = -ENOMEM;
3644                 mlog_errno(ret);
3645                 goto out;
3646         }
3647
3648         context->inode = inode;
3649         context->cow_start = cow_start;
3650         context->cow_len = cow_len;
3651         context->ref_tree = ref_tree;
3652         context->ref_root_bh = ref_root_bh;;
3653         context->cow_object = xv;
3654
3655         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd;
3656         /* We need the extra credits for duplicate_clusters by jbd. */
3657         context->extra_credits =
3658                 ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len;
3659         context->get_clusters = ocfs2_xattr_value_get_clusters;
3660         context->post_refcount = post;
3661
3662         ocfs2_init_xattr_value_extent_tree(&context->data_et,
3663                                            INODE_CACHE(inode), vb);
3664
3665         ret = ocfs2_replace_cow(context);
3666         if (ret)
3667                 mlog_errno(ret);
3668
3669 out:
3670         kfree(context);
3671         return ret;
3672 }
3673
3674 /*
3675  * Insert a new extent into refcount tree and mark a extent rec
3676  * as refcounted in the dinode tree.
3677  */
3678 int ocfs2_add_refcount_flag(struct inode *inode,
3679                             struct ocfs2_extent_tree *data_et,
3680                             struct ocfs2_caching_info *ref_ci,
3681                             struct buffer_head *ref_root_bh,
3682                             u32 cpos, u32 p_cluster, u32 num_clusters,
3683                             struct ocfs2_cached_dealloc_ctxt *dealloc,
3684                             struct ocfs2_post_refcount *post)
3685 {
3686         int ret;
3687         handle_t *handle;
3688         int credits = 1, ref_blocks = 0;
3689         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3690         struct ocfs2_alloc_context *meta_ac = NULL;
3691
3692         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
3693                                                ref_ci, ref_root_bh,
3694                                                p_cluster, num_clusters,
3695                                                &ref_blocks, &credits);
3696         if (ret) {
3697                 mlog_errno(ret);
3698                 goto out;
3699         }
3700
3701         mlog(0, "reserve new metadata %d, credits = %d\n",
3702              ref_blocks, credits);
3703
3704         if (ref_blocks) {
3705                 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
3706                                                         ref_blocks, &meta_ac);
3707                 if (ret) {
3708                         mlog_errno(ret);
3709                         goto out;
3710                 }
3711         }
3712
3713         if (post)
3714                 credits += post->credits;
3715
3716         handle = ocfs2_start_trans(osb, credits);
3717         if (IS_ERR(handle)) {
3718                 ret = PTR_ERR(handle);
3719                 mlog_errno(ret);
3720                 goto out;
3721         }
3722
3723         ret = ocfs2_mark_extent_refcounted(inode, data_et, handle,
3724                                            cpos, num_clusters, p_cluster,
3725                                            meta_ac, dealloc);
3726         if (ret) {
3727                 mlog_errno(ret);
3728                 goto out_commit;
3729         }
3730
3731         ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3732                                         p_cluster, num_clusters, 0,
3733                                         meta_ac, dealloc);
3734         if (ret) {
3735                 mlog_errno(ret);
3736                 goto out_commit;
3737         }
3738
3739         if (post && post->func) {
3740                 ret = post->func(inode, handle, post->para);
3741                 if (ret)
3742                         mlog_errno(ret);
3743         }
3744
3745 out_commit:
3746         ocfs2_commit_trans(osb, handle);
3747 out:
3748         if (meta_ac)
3749                 ocfs2_free_alloc_context(meta_ac);
3750         return ret;
3751 }
3752
3753 static int ocfs2_change_ctime(struct inode *inode,
3754                               struct buffer_head *di_bh)
3755 {
3756         int ret;
3757         handle_t *handle;
3758         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3759
3760         handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
3761                                    OCFS2_INODE_UPDATE_CREDITS);
3762         if (IS_ERR(handle)) {
3763                 ret = PTR_ERR(handle);
3764                 mlog_errno(ret);
3765                 goto out;
3766         }
3767
3768         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
3769                                       OCFS2_JOURNAL_ACCESS_WRITE);
3770         if (ret) {
3771                 mlog_errno(ret);
3772                 goto out_commit;
3773         }
3774
3775         inode->i_ctime = CURRENT_TIME;
3776         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
3777         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
3778
3779         ocfs2_journal_dirty(handle, di_bh);
3780
3781 out_commit:
3782         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
3783 out:
3784         return ret;
3785 }
3786
3787 static int ocfs2_attach_refcount_tree(struct inode *inode,
3788                                       struct buffer_head *di_bh)
3789 {
3790         int ret, data_changed = 0;
3791         struct buffer_head *ref_root_bh = NULL;
3792         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3793         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3794         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3795         struct ocfs2_refcount_tree *ref_tree;
3796         unsigned int ext_flags;
3797         loff_t size;
3798         u32 cpos, num_clusters, clusters, p_cluster;
3799         struct ocfs2_cached_dealloc_ctxt dealloc;
3800         struct ocfs2_extent_tree di_et;
3801
3802         ocfs2_init_dealloc_ctxt(&dealloc);
3803
3804         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) {
3805                 ret = ocfs2_create_refcount_tree(inode, di_bh);
3806                 if (ret) {
3807                         mlog_errno(ret);
3808                         goto out;
3809                 }
3810         }
3811
3812         BUG_ON(!di->i_refcount_loc);
3813         ret = ocfs2_lock_refcount_tree(osb,
3814                                        le64_to_cpu(di->i_refcount_loc), 1,
3815                                        &ref_tree, &ref_root_bh);
3816         if (ret) {
3817                 mlog_errno(ret);
3818                 goto out;
3819         }
3820
3821         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
3822                 goto attach_xattr;
3823
3824         ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh);
3825
3826         size = i_size_read(inode);
3827         clusters = ocfs2_clusters_for_bytes(inode->i_sb, size);
3828
3829         cpos = 0;
3830         while (cpos < clusters) {
3831                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3832                                          &num_clusters, &ext_flags);
3833
3834                 if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) {
3835                         ret = ocfs2_add_refcount_flag(inode, &di_et,
3836                                                       &ref_tree->rf_ci,
3837                                                       ref_root_bh, cpos,
3838                                                       p_cluster, num_clusters,
3839                                                       &dealloc, NULL);
3840                         if (ret) {
3841                                 mlog_errno(ret);
3842                                 goto unlock;
3843                         }
3844
3845                         data_changed = 1;
3846                 }
3847                 cpos += num_clusters;
3848         }
3849
3850 attach_xattr:
3851         if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
3852                 ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh,
3853                                                        &ref_tree->rf_ci,
3854                                                        ref_root_bh,
3855                                                        &dealloc);
3856                 if (ret) {
3857                         mlog_errno(ret);
3858                         goto unlock;
3859                 }
3860         }
3861
3862         if (data_changed) {
3863                 ret = ocfs2_change_ctime(inode, di_bh);
3864                 if (ret)
3865                         mlog_errno(ret);
3866         }
3867
3868 unlock:
3869         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3870         brelse(ref_root_bh);
3871
3872         if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) {
3873                 ocfs2_schedule_truncate_log_flush(osb, 1);
3874                 ocfs2_run_deallocs(osb, &dealloc);
3875         }
3876 out:
3877         /*
3878          * Empty the extent map so that we may get the right extent
3879          * record from the disk.
3880          */
3881         ocfs2_extent_map_trunc(inode, 0);
3882
3883         return ret;
3884 }
3885
3886 static int ocfs2_add_refcounted_extent(struct inode *inode,
3887                                    struct ocfs2_extent_tree *et,
3888                                    struct ocfs2_caching_info *ref_ci,
3889                                    struct buffer_head *ref_root_bh,
3890                                    u32 cpos, u32 p_cluster, u32 num_clusters,
3891                                    unsigned int ext_flags,
3892                                    struct ocfs2_cached_dealloc_ctxt *dealloc)
3893 {
3894         int ret;
3895         handle_t *handle;
3896         int credits = 0;
3897         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3898         struct ocfs2_alloc_context *meta_ac = NULL;
3899
3900         ret = ocfs2_lock_refcount_allocators(inode->i_sb,
3901                                              p_cluster, num_clusters,
3902                                              et, ref_ci,
3903                                              ref_root_bh, &meta_ac,
3904                                              NULL, &credits);
3905         if (ret) {
3906                 mlog_errno(ret);
3907                 goto out;
3908         }
3909
3910         handle = ocfs2_start_trans(osb, credits);
3911         if (IS_ERR(handle)) {
3912                 ret = PTR_ERR(handle);
3913                 mlog_errno(ret);
3914                 goto out;
3915         }
3916
3917         ret = ocfs2_insert_extent(handle, et, cpos,
3918                         ocfs2_clusters_to_blocks(inode->i_sb, p_cluster),
3919                         num_clusters, ext_flags, meta_ac);
3920         if (ret) {
3921                 mlog_errno(ret);
3922                 goto out_commit;
3923         }
3924
3925         ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3926                                       p_cluster, num_clusters,
3927                                       meta_ac, dealloc);
3928         if (ret)
3929                 mlog_errno(ret);
3930
3931 out_commit:
3932         ocfs2_commit_trans(osb, handle);
3933 out:
3934         if (meta_ac)
3935                 ocfs2_free_alloc_context(meta_ac);
3936         return ret;
3937 }
3938
3939 static int ocfs2_duplicate_inline_data(struct inode *s_inode,
3940                                        struct buffer_head *s_bh,
3941                                        struct inode *t_inode,
3942                                        struct buffer_head *t_bh)
3943 {
3944         int ret;
3945         handle_t *handle;
3946         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
3947         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
3948         struct ocfs2_dinode *t_di = (struct ocfs2_dinode *)t_bh->b_data;
3949
3950         BUG_ON(!(OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
3951
3952         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
3953         if (IS_ERR(handle)) {
3954                 ret = PTR_ERR(handle);
3955                 mlog_errno(ret);
3956                 goto out;
3957         }
3958
3959         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
3960                                       OCFS2_JOURNAL_ACCESS_WRITE);
3961         if (ret) {
3962                 mlog_errno(ret);
3963                 goto out_commit;
3964         }
3965
3966         t_di->id2.i_data.id_count = s_di->id2.i_data.id_count;
3967         memcpy(t_di->id2.i_data.id_data, s_di->id2.i_data.id_data,
3968                le16_to_cpu(s_di->id2.i_data.id_count));
3969         spin_lock(&OCFS2_I(t_inode)->ip_lock);
3970         OCFS2_I(t_inode)->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
3971         t_di->i_dyn_features = cpu_to_le16(OCFS2_I(t_inode)->ip_dyn_features);
3972         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
3973
3974         ocfs2_journal_dirty(handle, t_bh);
3975
3976 out_commit:
3977         ocfs2_commit_trans(osb, handle);
3978 out:
3979         return ret;
3980 }
3981
3982 static int ocfs2_duplicate_extent_list(struct inode *s_inode,
3983                                 struct inode *t_inode,
3984                                 struct buffer_head *t_bh,
3985                                 struct ocfs2_caching_info *ref_ci,
3986                                 struct buffer_head *ref_root_bh,
3987                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
3988 {
3989         int ret = 0;
3990         u32 p_cluster, num_clusters, clusters, cpos;
3991         loff_t size;
3992         unsigned int ext_flags;
3993         struct ocfs2_extent_tree et;
3994
3995         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh);
3996
3997         size = i_size_read(s_inode);
3998         clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size);
3999
4000         cpos = 0;
4001         while (cpos < clusters) {
4002                 ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster,
4003                                          &num_clusters, &ext_flags);
4004
4005                 if (p_cluster) {
4006                         ret = ocfs2_add_refcounted_extent(t_inode, &et,
4007                                                           ref_ci, ref_root_bh,
4008                                                           cpos, p_cluster,
4009                                                           num_clusters,
4010                                                           ext_flags,
4011                                                           dealloc);
4012                         if (ret) {
4013                                 mlog_errno(ret);
4014                                 goto out;
4015                         }
4016                 }
4017
4018                 cpos += num_clusters;
4019         }
4020
4021 out:
4022         return ret;
4023 }
4024
4025 /*
4026  * change the new file's attributes to the src.
4027  *
4028  * reflink creates a snapshot of a file, that means the attributes
4029  * must be identical except for three exceptions - nlink, ino, and ctime.
4030  */
4031 static int ocfs2_complete_reflink(struct inode *s_inode,
4032                                   struct buffer_head *s_bh,
4033                                   struct inode *t_inode,
4034                                   struct buffer_head *t_bh,
4035                                   bool preserve)
4036 {
4037         int ret;
4038         handle_t *handle;
4039         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
4040         struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data;
4041         loff_t size = i_size_read(s_inode);
4042
4043         handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb),
4044                                    OCFS2_INODE_UPDATE_CREDITS);
4045         if (IS_ERR(handle)) {
4046                 ret = PTR_ERR(handle);
4047                 mlog_errno(ret);
4048                 return ret;
4049         }
4050
4051         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
4052                                       OCFS2_JOURNAL_ACCESS_WRITE);
4053         if (ret) {
4054                 mlog_errno(ret);
4055                 goto out_commit;
4056         }
4057
4058         spin_lock(&OCFS2_I(t_inode)->ip_lock);
4059         OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters;
4060         OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr;
4061         OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features;
4062         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
4063         i_size_write(t_inode, size);
4064         t_inode->i_blocks = s_inode->i_blocks;
4065
4066         di->i_xattr_inline_size = s_di->i_xattr_inline_size;
4067         di->i_clusters = s_di->i_clusters;
4068         di->i_size = s_di->i_size;
4069         di->i_dyn_features = s_di->i_dyn_features;
4070         di->i_attr = s_di->i_attr;
4071
4072         if (preserve) {
4073                 di->i_uid = s_di->i_uid;
4074                 di->i_gid = s_di->i_gid;
4075                 di->i_mode = s_di->i_mode;
4076
4077                 /*
4078                  * update time.
4079                  * we want mtime to appear identical to the source and
4080                  * update ctime.
4081                  */
4082                 t_inode->i_ctime = CURRENT_TIME;
4083
4084                 di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec);
4085                 di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec);
4086
4087                 t_inode->i_mtime = s_inode->i_mtime;
4088                 di->i_mtime = s_di->i_mtime;
4089                 di->i_mtime_nsec = s_di->i_mtime_nsec;
4090         }
4091
4092         ocfs2_journal_dirty(handle, t_bh);
4093
4094 out_commit:
4095         ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle);
4096         return ret;
4097 }
4098
4099 static int ocfs2_create_reflink_node(struct inode *s_inode,
4100                                      struct buffer_head *s_bh,
4101                                      struct inode *t_inode,
4102                                      struct buffer_head *t_bh,
4103                                      bool preserve)
4104 {
4105         int ret;
4106         struct buffer_head *ref_root_bh = NULL;
4107         struct ocfs2_cached_dealloc_ctxt dealloc;
4108         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
4109         struct ocfs2_refcount_block *rb;
4110         struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data;
4111         struct ocfs2_refcount_tree *ref_tree;
4112
4113         ocfs2_init_dealloc_ctxt(&dealloc);
4114
4115         ret = ocfs2_set_refcount_tree(t_inode, t_bh,
4116                                       le64_to_cpu(di->i_refcount_loc));
4117         if (ret) {
4118                 mlog_errno(ret);
4119                 goto out;
4120         }
4121
4122         if (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4123                 ret = ocfs2_duplicate_inline_data(s_inode, s_bh,
4124                                                   t_inode, t_bh);
4125                 if (ret)
4126                         mlog_errno(ret);
4127                 goto out;
4128         }
4129
4130         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
4131                                        1, &ref_tree, &ref_root_bh);
4132         if (ret) {
4133                 mlog_errno(ret);
4134                 goto out;
4135         }
4136         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
4137
4138         ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh,
4139                                           &ref_tree->rf_ci, ref_root_bh,
4140                                           &dealloc);
4141         if (ret) {
4142                 mlog_errno(ret);
4143                 goto out_unlock_refcount;
4144         }
4145
4146 out_unlock_refcount:
4147         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4148         brelse(ref_root_bh);
4149 out:
4150         if (ocfs2_dealloc_has_cluster(&dealloc)) {
4151                 ocfs2_schedule_truncate_log_flush(osb, 1);
4152                 ocfs2_run_deallocs(osb, &dealloc);
4153         }
4154
4155         return ret;
4156 }
4157
4158 static int __ocfs2_reflink(struct dentry *old_dentry,
4159                            struct buffer_head *old_bh,
4160                            struct inode *new_inode,
4161                            bool preserve)
4162 {
4163         int ret;
4164         struct inode *inode = old_dentry->d_inode;
4165         struct buffer_head *new_bh = NULL;
4166
4167         ret = filemap_fdatawrite(inode->i_mapping);
4168         if (ret) {
4169                 mlog_errno(ret);
4170                 goto out;
4171         }
4172
4173         ret = ocfs2_attach_refcount_tree(inode, old_bh);
4174         if (ret) {
4175                 mlog_errno(ret);
4176                 goto out;
4177         }
4178
4179         mutex_lock(&new_inode->i_mutex);
4180         ret = ocfs2_inode_lock(new_inode, &new_bh, 1);
4181         if (ret) {
4182                 mlog_errno(ret);
4183                 goto out_unlock;
4184         }
4185
4186         ret = ocfs2_create_reflink_node(inode, old_bh,
4187                                         new_inode, new_bh, preserve);
4188         if (ret) {
4189                 mlog_errno(ret);
4190                 goto inode_unlock;
4191         }
4192
4193         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
4194                 ret = ocfs2_reflink_xattrs(inode, old_bh,
4195                                            new_inode, new_bh,
4196                                            preserve);
4197                 if (ret) {
4198                         mlog_errno(ret);
4199                         goto inode_unlock;
4200                 }
4201         }
4202
4203         ret = ocfs2_complete_reflink(inode, old_bh,
4204                                      new_inode, new_bh, preserve);
4205         if (ret)
4206                 mlog_errno(ret);
4207
4208 inode_unlock:
4209         ocfs2_inode_unlock(new_inode, 1);
4210         brelse(new_bh);
4211 out_unlock:
4212         mutex_unlock(&new_inode->i_mutex);
4213 out:
4214         if (!ret) {
4215                 ret = filemap_fdatawait(inode->i_mapping);
4216                 if (ret)
4217                         mlog_errno(ret);
4218         }
4219         return ret;
4220 }
4221
4222 static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir,
4223                          struct dentry *new_dentry, bool preserve)
4224 {
4225         int error;
4226         struct inode *inode = old_dentry->d_inode;
4227         struct buffer_head *old_bh = NULL;
4228         struct inode *new_orphan_inode = NULL;
4229
4230         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4231                 return -EOPNOTSUPP;
4232
4233         error = ocfs2_create_inode_in_orphan(dir, inode->i_mode,
4234                                              &new_orphan_inode);
4235         if (error) {
4236                 mlog_errno(error);
4237                 goto out;
4238         }
4239
4240         error = ocfs2_inode_lock(inode, &old_bh, 1);
4241         if (error) {
4242                 mlog_errno(error);
4243                 goto out;
4244         }
4245
4246         down_write(&OCFS2_I(inode)->ip_xattr_sem);
4247         down_write(&OCFS2_I(inode)->ip_alloc_sem);
4248         error = __ocfs2_reflink(old_dentry, old_bh,
4249                                 new_orphan_inode, preserve);
4250         up_write(&OCFS2_I(inode)->ip_alloc_sem);
4251         up_write(&OCFS2_I(inode)->ip_xattr_sem);
4252
4253         ocfs2_inode_unlock(inode, 1);
4254         brelse(old_bh);
4255
4256         if (error) {
4257                 mlog_errno(error);
4258                 goto out;
4259         }
4260
4261         /* If the security isn't preserved, we need to re-initialize them. */
4262         if (!preserve) {
4263                 error = ocfs2_init_security_and_acl(dir, new_orphan_inode);
4264                 if (error)
4265                         mlog_errno(error);
4266         }
4267 out:
4268         if (!error) {
4269                 error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode,
4270                                                        new_dentry);
4271                 if (error)
4272                         mlog_errno(error);
4273         }
4274
4275         if (new_orphan_inode) {
4276                 /*
4277                  * We need to open_unlock the inode no matter whether we
4278                  * succeed or not, so that other nodes can delete it later.
4279                  */
4280                 ocfs2_open_unlock(new_orphan_inode);
4281                 if (error)
4282                         iput(new_orphan_inode);
4283         }
4284
4285         return error;
4286 }
4287
4288 /*
4289  * Below here are the bits used by OCFS2_IOC_REFLINK() to fake
4290  * sys_reflink().  This will go away when vfs_reflink() exists in
4291  * fs/namei.c.
4292  */
4293
4294 /* copied from may_create in VFS. */
4295 static inline int ocfs2_may_create(struct inode *dir, struct dentry *child)
4296 {
4297         if (child->d_inode)
4298                 return -EEXIST;
4299         if (IS_DEADDIR(dir))
4300                 return -ENOENT;
4301         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
4302 }
4303
4304 /* copied from user_path_parent. */
4305 static int ocfs2_user_path_parent(const char __user *path,
4306                                   struct nameidata *nd, char **name)
4307 {
4308         char *s = getname(path);
4309         int error;
4310
4311         if (IS_ERR(s))
4312                 return PTR_ERR(s);
4313
4314         error = path_lookup(s, LOOKUP_PARENT, nd);
4315         if (error)
4316                 putname(s);
4317         else
4318                 *name = s;
4319
4320         return error;
4321 }
4322
4323 /**
4324  * ocfs2_vfs_reflink - Create a reference-counted link
4325  *
4326  * @old_dentry:        source dentry + inode
4327  * @dir:       directory to create the target
4328  * @new_dentry:        target dentry
4329  * @preserve:  if true, preserve all file attributes
4330  */
4331 static int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir,
4332                              struct dentry *new_dentry, bool preserve)
4333 {
4334         struct inode *inode = old_dentry->d_inode;
4335         int error;
4336
4337         if (!inode)
4338                 return -ENOENT;
4339
4340         error = ocfs2_may_create(dir, new_dentry);
4341         if (error)
4342                 return error;
4343
4344         if (dir->i_sb != inode->i_sb)
4345                 return -EXDEV;
4346
4347         /*
4348          * A reflink to an append-only or immutable file cannot be created.
4349          */
4350         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4351                 return -EPERM;
4352
4353         /* Only regular files can be reflinked. */
4354         if (!S_ISREG(inode->i_mode))
4355                 return -EPERM;
4356
4357         /*
4358          * If the caller wants to preserve ownership, they require the
4359          * rights to do so.
4360          */
4361         if (preserve) {
4362                 if ((current_fsuid() != inode->i_uid) && !capable(CAP_CHOWN))
4363                         return -EPERM;
4364                 if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN))
4365                         return -EPERM;
4366         }
4367
4368         /*
4369          * If the caller is modifying any aspect of the attributes, they
4370          * are not creating a snapshot.  They need read permission on the
4371          * file.
4372          */
4373         if (!preserve) {
4374                 error = inode_permission(inode, MAY_READ);
4375                 if (error)
4376                         return error;
4377         }
4378
4379         mutex_lock(&inode->i_mutex);
4380         dquot_initialize(dir);
4381         error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve);
4382         mutex_unlock(&inode->i_mutex);
4383         if (!error)
4384                 fsnotify_create(dir, new_dentry);
4385         return error;
4386 }
4387 /*
4388  * Most codes are copied from sys_linkat.
4389  */
4390 int ocfs2_reflink_ioctl(struct inode *inode,
4391                         const char __user *oldname,
4392                         const char __user *newname,
4393                         bool preserve)
4394 {
4395         struct dentry *new_dentry;
4396         struct nameidata nd;
4397         struct path old_path;
4398         int error;
4399         char *to = NULL;
4400
4401         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4402                 return -EOPNOTSUPP;
4403
4404         error = user_path_at(AT_FDCWD, oldname, 0, &old_path);
4405         if (error) {
4406                 mlog_errno(error);
4407                 return error;
4408         }
4409
4410         error = ocfs2_user_path_parent(newname, &nd, &to);
4411         if (error) {
4412                 mlog_errno(error);
4413                 goto out;
4414         }
4415
4416         error = -EXDEV;
4417         if (old_path.mnt != nd.path.mnt)
4418                 goto out_release;
4419         new_dentry = lookup_create(&nd, 0);
4420         error = PTR_ERR(new_dentry);
4421         if (IS_ERR(new_dentry)) {
4422                 mlog_errno(error);
4423                 goto out_unlock;
4424         }
4425
4426         error = mnt_want_write(nd.path.mnt);
4427         if (error) {
4428                 mlog_errno(error);
4429                 goto out_dput;
4430         }
4431
4432         error = ocfs2_vfs_reflink(old_path.dentry,
4433                                   nd.path.dentry->d_inode,
4434                                   new_dentry, preserve);
4435         mnt_drop_write(nd.path.mnt);
4436 out_dput:
4437         dput(new_dentry);
4438 out_unlock:
4439         mutex_unlock(&nd.path.dentry->d_inode->i_mutex);
4440 out_release:
4441         path_put(&nd.path);
4442         putname(to);
4443 out:
4444         path_put(&old_path);
4445
4446         return error;
4447 }