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 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 int ocfs2_lock_refcount_tree_by_inode(struct inode *inode, int rw,
 528                                       struct ocfs2_refcount_tree **ret_tree,
 529                                       struct buffer_head **ref_bh)
 530 {
 531         int ret;
 532         u64 ref_blkno;
 533
 534         ret = ocfs2_get_refcount_block(inode, &ref_blkno);
 535         if (ret) {
 536                 mlog_errno(ret);
 537                 return ret;
 538         }
 539
 540         return ocfs2_lock_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno,
 541                                         rw, ret_tree, ref_bh);
 542 }
 543
 544 void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
 545                                 struct ocfs2_refcount_tree *tree, int rw)
 546 {
 547         if (rw)
 548                 up_write(&tree->rf_sem);
 549         else
 550                 up_read(&tree->rf_sem);
 551
 552         ocfs2_refcount_unlock(tree, rw);
 553         ocfs2_refcount_tree_put(tree);
 554 }
 555
 556 void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
 557 {
 558         struct rb_node *node;
 559         struct ocfs2_refcount_tree *tree;
 560         struct rb_root *root = &osb->osb_rf_lock_tree;
 561
 562         while ((node = rb_last(root)) != NULL) {
 563                 tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
 564
 565                 mlog(0, "Purge tree %llu\n",
 566                      (unsigned long long) tree->rf_blkno);
 567
 568                 rb_erase(&tree->rf_node, root);
 569                 ocfs2_free_refcount_tree(tree);
 570         }
 571 }
 572
 573 /*
 574  * Create a refcount tree for an inode.
 575  * We take for granted that the inode is already locked.
 576  */
 577 static int ocfs2_create_refcount_tree(struct inode *inode,
 578                                       struct buffer_head *di_bh)
 579 {
 580         int ret;
 581         handle_t *handle = NULL;
 582         struct ocfs2_alloc_context *meta_ac = NULL;
 583         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 584         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 585         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 586         struct buffer_head *new_bh = NULL;
 587         struct ocfs2_refcount_block *rb;
 588         struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
 589         u16 suballoc_bit_start;
 590         u32 num_got;
 591         u64 first_blkno;
 592
 593         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
 594
 595         mlog(0, "create tree for inode %lu\n", inode->i_ino);
 596
 597         ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
 598         if (ret) {
 599                 mlog_errno(ret);
 600                 goto out;
 601         }
 602
 603         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
 604         if (IS_ERR(handle)) {
 605                 ret = PTR_ERR(handle);
 606                 mlog_errno(ret);
 607                 goto out;
 608         }
 609
 610         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 611                                       OCFS2_JOURNAL_ACCESS_WRITE);
 612         if (ret) {
 613                 mlog_errno(ret);
 614                 goto out_commit;
 615         }
 616
 617         ret = ocfs2_claim_metadata(osb, handle, meta_ac, 1,
 618                                    &suballoc_bit_start, &num_got,
 619                                    &first_blkno);
 620         if (ret) {
 621                 mlog_errno(ret);
 622                 goto out_commit;
 623         }
 624
 625         new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
 626         if (!new_tree) {
 627                 ret = -ENOMEM;
 628                 mlog_errno(ret);
 629                 goto out_commit;
 630         }
 631
 632         new_bh = sb_getblk(inode->i_sb, first_blkno);
 633         ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
 634
 635         ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
 636                                       OCFS2_JOURNAL_ACCESS_CREATE);
 637         if (ret) {
 638                 mlog_errno(ret);
 639                 goto out_commit;
 640         }
 641
 642         /* Initialize ocfs2_refcount_block. */
 643         rb = (struct ocfs2_refcount_block *)new_bh->b_data;
 644         memset(rb, 0, inode->i_sb->s_blocksize);
 645         strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
 646         rb->rf_suballoc_slot = cpu_to_le16(osb->slot_num);
 647         rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
 648         rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
 649         rb->rf_blkno = cpu_to_le64(first_blkno);
 650         rb->rf_count = cpu_to_le32(1);
 651         rb->rf_records.rl_count =
 652                         cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
 653         spin_lock(&osb->osb_lock);
 654         rb->rf_generation = osb->s_next_generation++;
 655         spin_unlock(&osb->osb_lock);
 656
 657         ocfs2_journal_dirty(handle, new_bh);
 658
 659         spin_lock(&oi->ip_lock);
 660         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
 661         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 662         di->i_refcount_loc = cpu_to_le64(first_blkno);
 663         spin_unlock(&oi->ip_lock);
 664
 665         mlog(0, "created tree for inode %lu, refblock %llu\n",
 666              inode->i_ino, (unsigned long long)first_blkno);
 667
 668         ocfs2_journal_dirty(handle, di_bh);
 669
 670         /*
 671          * We have to init the tree lock here since it will use
 672          * the generation number to create it.
 673          */
 674         new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
 675         ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
 676                                       new_tree->rf_generation);
 677
 678         spin_lock(&osb->osb_lock);
 679         tree = ocfs2_find_refcount_tree(osb, first_blkno);
 680
 681         /*
 682          * We've just created a new refcount tree in this block.  If
 683          * we found a refcount tree on the ocfs2_super, it must be
 684          * one we just deleted.  We free the old tree before
 685          * inserting the new tree.
 686          */
 687         BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
 688         if (tree)
 689                 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
 690         ocfs2_insert_refcount_tree(osb, new_tree);
 691         spin_unlock(&osb->osb_lock);
 692         new_tree = NULL;
 693         if (tree)
 694                 ocfs2_refcount_tree_put(tree);
 695
 696 out_commit:
 697         ocfs2_commit_trans(osb, handle);
 698
 699 out:
 700         if (new_tree) {
 701                 ocfs2_metadata_cache_exit(&new_tree->rf_ci);
 702                 kfree(new_tree);
 703         }
 704
 705         brelse(new_bh);
 706         if (meta_ac)
 707                 ocfs2_free_alloc_context(meta_ac);
 708
 709         return ret;
 710 }
 711
 712 static int ocfs2_set_refcount_tree(struct inode *inode,
 713                                    struct buffer_head *di_bh,
 714                                    u64 refcount_loc)
 715 {
 716         int ret;
 717         handle_t *handle = NULL;
 718         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 719         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 720         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 721         struct buffer_head *ref_root_bh = NULL;
 722         struct ocfs2_refcount_block *rb;
 723         struct ocfs2_refcount_tree *ref_tree;
 724
 725         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
 726
 727         ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
 728                                        &ref_tree, &ref_root_bh);
 729         if (ret) {
 730                 mlog_errno(ret);
 731                 return ret;
 732         }
 733
 734         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
 735         if (IS_ERR(handle)) {
 736                 ret = PTR_ERR(handle);
 737                 mlog_errno(ret);
 738                 goto out;
 739         }
 740
 741         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 742                                       OCFS2_JOURNAL_ACCESS_WRITE);
 743         if (ret) {
 744                 mlog_errno(ret);
 745                 goto out_commit;
 746         }
 747
 748         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
 749                                       OCFS2_JOURNAL_ACCESS_WRITE);
 750         if (ret) {
 751                 mlog_errno(ret);
 752                 goto out_commit;
 753         }
 754
 755         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 756         le32_add_cpu(&rb->rf_count, 1);
 757
 758         ocfs2_journal_dirty(handle, ref_root_bh);
 759
 760         spin_lock(&oi->ip_lock);
 761         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
 762         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 763         di->i_refcount_loc = cpu_to_le64(refcount_loc);
 764         spin_unlock(&oi->ip_lock);
 765         ocfs2_journal_dirty(handle, di_bh);
 766
 767 out_commit:
 768         ocfs2_commit_trans(osb, handle);
 769 out:
 770         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 771         brelse(ref_root_bh);
 772
 773         return ret;
 774 }
 775
 776 int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
 777 {
 778         int ret, delete_tree = 0;
 779         handle_t *handle = NULL;
 780         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 781         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 782         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 783         struct ocfs2_refcount_block *rb;
 784         struct inode *alloc_inode = NULL;
 785         struct buffer_head *alloc_bh = NULL;
 786         struct buffer_head *blk_bh = NULL;
 787         struct ocfs2_refcount_tree *ref_tree;
 788         int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
 789         u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
 790         u16 bit = 0;
 791
 792         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
 793                 return 0;
 794
 795         BUG_ON(!ref_blkno);
 796         ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
 797         if (ret) {
 798                 mlog_errno(ret);
 799                 return ret;
 800         }
 801
 802         rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
 803
 804         /*
 805          * If we are the last user, we need to free the block.
 806          * So lock the allocator ahead.
 807          */
 808         if (le32_to_cpu(rb->rf_count) == 1) {
 809                 blk = le64_to_cpu(rb->rf_blkno);
 810                 bit = le16_to_cpu(rb->rf_suballoc_bit);
 811                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
 812
 813                 alloc_inode = ocfs2_get_system_file_inode(osb,
 814                                         EXTENT_ALLOC_SYSTEM_INODE,
 815                                         le16_to_cpu(rb->rf_suballoc_slot));
 816                 if (!alloc_inode) {
 817                         ret = -ENOMEM;
 818                         mlog_errno(ret);
 819                         goto out;
 820                 }
 821                 mutex_lock(&alloc_inode->i_mutex);
 822
 823                 ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
 824                 if (ret) {
 825                         mlog_errno(ret);
 826                         goto out_mutex;
 827                 }
 828
 829                 credits += OCFS2_SUBALLOC_FREE;
 830         }
 831
 832         handle = ocfs2_start_trans(osb, credits);
 833         if (IS_ERR(handle)) {
 834                 ret = PTR_ERR(handle);
 835                 mlog_errno(ret);
 836                 goto out_unlock;
 837         }
 838
 839         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 840                                       OCFS2_JOURNAL_ACCESS_WRITE);
 841         if (ret) {
 842                 mlog_errno(ret);
 843                 goto out_commit;
 844         }
 845
 846         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
 847                                       OCFS2_JOURNAL_ACCESS_WRITE);
 848         if (ret) {
 849                 mlog_errno(ret);
 850                 goto out_commit;
 851         }
 852
 853         spin_lock(&oi->ip_lock);
 854         oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
 855         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 856         di->i_refcount_loc = 0;
 857         spin_unlock(&oi->ip_lock);
 858         ocfs2_journal_dirty(handle, di_bh);
 859
 860         le32_add_cpu(&rb->rf_count , -1);
 861         ocfs2_journal_dirty(handle, blk_bh);
 862
 863         if (!rb->rf_count) {
 864                 delete_tree = 1;
 865                 ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
 866                 ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
 867                                                alloc_bh, bit, bg_blkno, 1);
 868                 if (ret)
 869                         mlog_errno(ret);
 870         }
 871
 872 out_commit:
 873         ocfs2_commit_trans(osb, handle);
 874 out_unlock:
 875         if (alloc_inode) {
 876                 ocfs2_inode_unlock(alloc_inode, 1);
 877                 brelse(alloc_bh);
 878         }
 879 out_mutex:
 880         if (alloc_inode) {
 881                 mutex_unlock(&alloc_inode->i_mutex);
 882                 iput(alloc_inode);
 883         }
 884 out:
 885         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 886         if (delete_tree)
 887                 ocfs2_refcount_tree_put(ref_tree);
 888         brelse(blk_bh);
 889
 890         return ret;
 891 }
 892
 893 static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
 894                                           struct buffer_head *ref_leaf_bh,
 895                                           u64 cpos, unsigned int len,
 896                                           struct ocfs2_refcount_rec *ret_rec,
 897                                           int *index)
 898 {
 899         int i = 0;
 900         struct ocfs2_refcount_block *rb =
 901                 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
 902         struct ocfs2_refcount_rec *rec = NULL;
 903
 904         for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
 905                 rec = &rb->rf_records.rl_recs[i];
 906
 907                 if (le64_to_cpu(rec->r_cpos) +
 908                     le32_to_cpu(rec->r_clusters) <= cpos)
 909                         continue;
 910                 else if (le64_to_cpu(rec->r_cpos) > cpos)
 911                         break;
 912
 913                 /* ok, cpos fail in this rec. Just return. */
 914                 if (ret_rec)
 915                         *ret_rec = *rec;
 916                 goto out;
 917         }
 918
 919         if (ret_rec) {
 920                 /* We meet with a hole here, so fake the rec. */
 921                 ret_rec->r_cpos = cpu_to_le64(cpos);
 922                 ret_rec->r_refcount = 0;
 923                 if (i < le16_to_cpu(rb->rf_records.rl_used) &&
 924                     le64_to_cpu(rec->r_cpos) < cpos + len)
 925                         ret_rec->r_clusters =
 926                                 cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
 927                 else
 928                         ret_rec->r_clusters = cpu_to_le32(len);
 929         }
 930
 931 out:
 932         *index = i;
 933 }
 934
 935 /*
 936  * Try to remove refcount tree. The mechanism is:
 937  * 1) Check whether i_clusters == 0, if no, exit.
 938  * 2) check whether we have i_xattr_loc in dinode. if yes, exit.
 939  * 3) Check whether we have inline xattr stored outside, if yes, exit.
 940  * 4) Remove the tree.
 941  */
 942 int ocfs2_try_remove_refcount_tree(struct inode *inode,
 943                                    struct buffer_head *di_bh)
 944 {
 945         int ret;
 946         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 947         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 948
 949         down_write(&oi->ip_xattr_sem);
 950         down_write(&oi->ip_alloc_sem);
 951
 952         if (oi->ip_clusters)
 953                 goto out;
 954
 955         if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc)
 956                 goto out;
 957
 958         if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL &&
 959             ocfs2_has_inline_xattr_value_outside(inode, di))
 960                 goto out;
 961
 962         ret = ocfs2_remove_refcount_tree(inode, di_bh);
 963         if (ret)
 964                 mlog_errno(ret);
 965 out:
 966         up_write(&oi->ip_alloc_sem);
 967         up_write(&oi->ip_xattr_sem);
 968         return 0;
 969 }
 970
 971 /*
 972  * Find the end range for a leaf refcount block indicated by
 973  * el->l_recs[index].e_blkno.
 974  */
 975 static int ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info *ci,
 976                                        struct buffer_head *ref_root_bh,
 977                                        struct ocfs2_extent_block *eb,
 978                                        struct ocfs2_extent_list *el,
 979                                        int index,  u32 *cpos_end)
 980 {
 981         int ret, i, subtree_root;
 982         u32 cpos;
 983         u64 blkno;
 984         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
 985         struct ocfs2_path *left_path = NULL, *right_path = NULL;
 986         struct ocfs2_extent_tree et;
 987         struct ocfs2_extent_list *tmp_el;
 988
 989         if (index < le16_to_cpu(el->l_next_free_rec) - 1) {
 990                 /*
 991                  * We have a extent rec after index, so just use the e_cpos
 992                  * of the next extent rec.
 993                  */
 994                 *cpos_end = le32_to_cpu(el->l_recs[index+1].e_cpos);
 995                 return 0;
 996         }
 997
 998         if (!eb || (eb && !eb->h_next_leaf_blk)) {
 999                 /*
1000                  * We are the last extent rec, so any high cpos should
1001                  * be stored in this leaf refcount block.
1002                  */
1003                 *cpos_end = UINT_MAX;
1004                 return 0;
1005         }
1006
1007         /*
1008          * If the extent block isn't the last one, we have to find
1009          * the subtree root between this extent block and the next
1010          * leaf extent block and get the corresponding e_cpos from
1011          * the subroot. Otherwise we may corrupt the b-tree.
1012          */
1013         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1014
1015         left_path = ocfs2_new_path_from_et(&et);
1016         if (!left_path) {
1017                 ret = -ENOMEM;
1018                 mlog_errno(ret);
1019                 goto out;
1020         }
1021
1022         cpos = le32_to_cpu(eb->h_list.l_recs[index].e_cpos);
1023         ret = ocfs2_find_path(ci, left_path, cpos);
1024         if (ret) {
1025                 mlog_errno(ret);
1026                 goto out;
1027         }
1028
1029         right_path = ocfs2_new_path_from_path(left_path);
1030         if (!right_path) {
1031                 ret = -ENOMEM;
1032                 mlog_errno(ret);
1033                 goto out;
1034         }
1035
1036         ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, &cpos);
1037         if (ret) {
1038                 mlog_errno(ret);
1039                 goto out;
1040         }
1041
1042         ret = ocfs2_find_path(ci, right_path, cpos);
1043         if (ret) {
1044                 mlog_errno(ret);
1045                 goto out;
1046         }
1047
1048         subtree_root = ocfs2_find_subtree_root(&et, left_path,
1049                                                right_path);
1050
1051         tmp_el = left_path->p_node[subtree_root].el;
1052         blkno = left_path->p_node[subtree_root+1].bh->b_blocknr;
1053         for (i = 0; i < le32_to_cpu(tmp_el->l_next_free_rec); i++) {
1054                 if (le64_to_cpu(tmp_el->l_recs[i].e_blkno) == blkno) {
1055                         *cpos_end = le32_to_cpu(tmp_el->l_recs[i+1].e_cpos);
1056                         break;
1057                 }
1058         }
1059
1060         BUG_ON(i == le32_to_cpu(tmp_el->l_next_free_rec));
1061
1062 out:
1063         ocfs2_free_path(left_path);
1064         ocfs2_free_path(right_path);
1065         return ret;
1066 }
1067
1068 /*
1069  * Given a cpos and len, try to find the refcount record which contains cpos.
1070  * 1. If cpos can be found in one refcount record, return the record.
1071  * 2. If cpos can't be found, return a fake record which start from cpos
1072  *    and end at a small value between cpos+len and start of the next record.
1073  *    This fake record has r_refcount = 0.
1074  */
1075 static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
1076                                   struct buffer_head *ref_root_bh,
1077                                   u64 cpos, unsigned int len,
1078                                   struct ocfs2_refcount_rec *ret_rec,
1079                                   int *index,
1080                                   struct buffer_head **ret_bh)
1081 {
1082         int ret = 0, i, found;
1083         u32 low_cpos, uninitialized_var(cpos_end);
1084         struct ocfs2_extent_list *el;
1085         struct ocfs2_extent_rec *rec = NULL;
1086         struct ocfs2_extent_block *eb = NULL;
1087         struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
1088         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1089         struct ocfs2_refcount_block *rb =
1090                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1091
1092         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
1093                 ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
1094                                               ret_rec, index);
1095                 *ret_bh = ref_root_bh;
1096                 get_bh(ref_root_bh);
1097                 return 0;
1098         }
1099
1100         el = &rb->rf_list;
1101         low_cpos = cpos & OCFS2_32BIT_POS_MASK;
1102
1103         if (el->l_tree_depth) {
1104                 ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
1105                 if (ret) {
1106                         mlog_errno(ret);
1107                         goto out;
1108                 }
1109
1110                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1111                 el = &eb->h_list;
1112
1113                 if (el->l_tree_depth) {
1114                         ocfs2_error(sb,
1115                         "refcount tree %llu has non zero tree "
1116                         "depth in leaf btree tree block %llu\n",
1117                         (unsigned long long)ocfs2_metadata_cache_owner(ci),
1118                         (unsigned long long)eb_bh->b_blocknr);
1119                         ret = -EROFS;
1120                         goto out;
1121                 }
1122         }
1123
1124         found = 0;
1125         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1126                 rec = &el->l_recs[i];
1127
1128                 if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
1129                         found = 1;
1130                         break;
1131                 }
1132         }
1133
1134         if (found) {
1135                 ret = ocfs2_get_refcount_cpos_end(ci, ref_root_bh,
1136                                                   eb, el, i, &cpos_end);
1137                 if (ret) {
1138                         mlog_errno(ret);
1139                         goto out;
1140                 }
1141
1142                 if (cpos_end < low_cpos + len)
1143                         len = cpos_end - low_cpos;
1144         }
1145
1146         ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
1147                                         &ref_leaf_bh);
1148         if (ret) {
1149                 mlog_errno(ret);
1150                 goto out;
1151         }
1152
1153         ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
1154                                       ret_rec, index);
1155         *ret_bh = ref_leaf_bh;
1156 out:
1157         brelse(eb_bh);
1158         return ret;
1159 }
1160
1161 enum ocfs2_ref_rec_contig {
1162         REF_CONTIG_NONE = 0,
1163         REF_CONTIG_LEFT,
1164         REF_CONTIG_RIGHT,
1165         REF_CONTIG_LEFTRIGHT,
1166 };
1167
1168 static enum ocfs2_ref_rec_contig
1169         ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
1170                                     int index)
1171 {
1172         if ((rb->rf_records.rl_recs[index].r_refcount ==
1173             rb->rf_records.rl_recs[index + 1].r_refcount) &&
1174             (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
1175             le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1176             le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1177                 return REF_CONTIG_RIGHT;
1178
1179         return REF_CONTIG_NONE;
1180 }
1181
1182 static enum ocfs2_ref_rec_contig
1183         ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1184                                   int index)
1185 {
1186         enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1187
1188         if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1189                 ret = ocfs2_refcount_rec_adjacent(rb, index);
1190
1191         if (index > 0) {
1192                 enum ocfs2_ref_rec_contig tmp;
1193
1194                 tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1195
1196                 if (tmp == REF_CONTIG_RIGHT) {
1197                         if (ret == REF_CONTIG_RIGHT)
1198                                 ret = REF_CONTIG_LEFTRIGHT;
1199                         else
1200                                 ret = REF_CONTIG_LEFT;
1201                 }
1202         }
1203
1204         return ret;
1205 }
1206
1207 static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1208                                            int index)
1209 {
1210         BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1211                rb->rf_records.rl_recs[index+1].r_refcount);
1212
1213         le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1214                      le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1215
1216         if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1217                 memmove(&rb->rf_records.rl_recs[index + 1],
1218                         &rb->rf_records.rl_recs[index + 2],
1219                         sizeof(struct ocfs2_refcount_rec) *
1220                         (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1221
1222         memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1223                0, sizeof(struct ocfs2_refcount_rec));
1224         le16_add_cpu(&rb->rf_records.rl_used, -1);
1225 }
1226
1227 /*
1228  * Merge the refcount rec if we are contiguous with the adjacent recs.
1229  */
1230 static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1231                                      int index)
1232 {
1233         enum ocfs2_ref_rec_contig contig =
1234                                 ocfs2_refcount_rec_contig(rb, index);
1235
1236         if (contig == REF_CONTIG_NONE)
1237                 return;
1238
1239         if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1240                 BUG_ON(index == 0);
1241                 index--;
1242         }
1243
1244         ocfs2_rotate_refcount_rec_left(rb, index);
1245
1246         if (contig == REF_CONTIG_LEFTRIGHT)
1247                 ocfs2_rotate_refcount_rec_left(rb, index);
1248 }
1249
1250 /*
1251  * Change the refcount indexed by "index" in ref_bh.
1252  * If refcount reaches 0, remove it.
1253  */
1254 static int ocfs2_change_refcount_rec(handle_t *handle,
1255                                      struct ocfs2_caching_info *ci,
1256                                      struct buffer_head *ref_leaf_bh,
1257                                      int index, int merge, int change)
1258 {
1259         int ret;
1260         struct ocfs2_refcount_block *rb =
1261                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1262         struct ocfs2_refcount_list *rl = &rb->rf_records;
1263         struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1264
1265         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1266                                       OCFS2_JOURNAL_ACCESS_WRITE);
1267         if (ret) {
1268                 mlog_errno(ret);
1269                 goto out;
1270         }
1271
1272         mlog(0, "change index %d, old count %u, change %d\n", index,
1273              le32_to_cpu(rec->r_refcount), change);
1274         le32_add_cpu(&rec->r_refcount, change);
1275
1276         if (!rec->r_refcount) {
1277                 if (index != le16_to_cpu(rl->rl_used) - 1) {
1278                         memmove(rec, rec + 1,
1279                                 (le16_to_cpu(rl->rl_used) - index - 1) *
1280                                 sizeof(struct ocfs2_refcount_rec));
1281                         memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
1282                                0, sizeof(struct ocfs2_refcount_rec));
1283                 }
1284
1285                 le16_add_cpu(&rl->rl_used, -1);
1286         } else if (merge)
1287                 ocfs2_refcount_rec_merge(rb, index);
1288
1289         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1290         if (ret)
1291                 mlog_errno(ret);
1292 out:
1293         return ret;
1294 }
1295
1296 static int ocfs2_expand_inline_ref_root(handle_t *handle,
1297                                         struct ocfs2_caching_info *ci,
1298                                         struct buffer_head *ref_root_bh,
1299                                         struct buffer_head **ref_leaf_bh,
1300                                         struct ocfs2_alloc_context *meta_ac)
1301 {
1302         int ret;
1303         u16 suballoc_bit_start;
1304         u32 num_got;
1305         u64 blkno;
1306         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1307         struct buffer_head *new_bh = NULL;
1308         struct ocfs2_refcount_block *new_rb;
1309         struct ocfs2_refcount_block *root_rb =
1310                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1311
1312         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1313                                       OCFS2_JOURNAL_ACCESS_WRITE);
1314         if (ret) {
1315                 mlog_errno(ret);
1316                 goto out;
1317         }
1318
1319         ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1320                                    &suballoc_bit_start, &num_got,
1321                                    &blkno);
1322         if (ret) {
1323                 mlog_errno(ret);
1324                 goto out;
1325         }
1326
1327         new_bh = sb_getblk(sb, blkno);
1328         if (new_bh == NULL) {
1329                 ret = -EIO;
1330                 mlog_errno(ret);
1331                 goto out;
1332         }
1333         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1334
1335         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1336                                       OCFS2_JOURNAL_ACCESS_CREATE);
1337         if (ret) {
1338                 mlog_errno(ret);
1339                 goto out;
1340         }
1341
1342         /*
1343          * Initialize ocfs2_refcount_block.
1344          * It should contain the same information as the old root.
1345          * so just memcpy it and change the corresponding field.
1346          */
1347         memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1348
1349         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1350         new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num);
1351         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1352         new_rb->rf_blkno = cpu_to_le64(blkno);
1353         new_rb->rf_cpos = cpu_to_le32(0);
1354         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1355         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1356         ocfs2_journal_dirty(handle, new_bh);
1357
1358         /* Now change the root. */
1359         memset(&root_rb->rf_list, 0, sb->s_blocksize -
1360                offsetof(struct ocfs2_refcount_block, rf_list));
1361         root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1362         root_rb->rf_clusters = cpu_to_le32(1);
1363         root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1364         root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1365         root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1366         root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1367
1368         ocfs2_journal_dirty(handle, ref_root_bh);
1369
1370         mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno,
1371              le16_to_cpu(new_rb->rf_records.rl_used));
1372
1373         *ref_leaf_bh = new_bh;
1374         new_bh = NULL;
1375 out:
1376         brelse(new_bh);
1377         return ret;
1378 }
1379
1380 static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1381                                            struct ocfs2_refcount_rec *next)
1382 {
1383         if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1384                 ocfs2_get_ref_rec_low_cpos(next))
1385                 return 1;
1386
1387         return 0;
1388 }
1389
1390 static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1391 {
1392         const struct ocfs2_refcount_rec *l = a, *r = b;
1393         u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1394         u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1395
1396         if (l_cpos > r_cpos)
1397                 return 1;
1398         if (l_cpos < r_cpos)
1399                 return -1;
1400         return 0;
1401 }
1402
1403 static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1404 {
1405         const struct ocfs2_refcount_rec *l = a, *r = b;
1406         u64 l_cpos = le64_to_cpu(l->r_cpos);
1407         u64 r_cpos = le64_to_cpu(r->r_cpos);
1408
1409         if (l_cpos > r_cpos)
1410                 return 1;
1411         if (l_cpos < r_cpos)
1412                 return -1;
1413         return 0;
1414 }
1415
1416 static void swap_refcount_rec(void *a, void *b, int size)
1417 {
1418         struct ocfs2_refcount_rec *l = a, *r = b, tmp;
1419
1420         tmp = *(struct ocfs2_refcount_rec *)l;
1421         *(struct ocfs2_refcount_rec *)l =
1422                         *(struct ocfs2_refcount_rec *)r;
1423         *(struct ocfs2_refcount_rec *)r = tmp;
1424 }
1425
1426 /*
1427  * The refcount cpos are ordered by their 64bit cpos,
1428  * But we will use the low 32 bit to be the e_cpos in the b-tree.
1429  * So we need to make sure that this pos isn't intersected with others.
1430  *
1431  * Note: The refcount block is already sorted by their low 32 bit cpos,
1432  *       So just try the middle pos first, and we will exit when we find
1433  *       the good position.
1434  */
1435 static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1436                                          u32 *split_pos, int *split_index)
1437 {
1438         int num_used = le16_to_cpu(rl->rl_used);
1439         int delta, middle = num_used / 2;
1440
1441         for (delta = 0; delta < middle; delta++) {
1442                 /* Let's check delta earlier than middle */
1443                 if (ocfs2_refcount_rec_no_intersect(
1444                                         &rl->rl_recs[middle - delta - 1],
1445                                         &rl->rl_recs[middle - delta])) {
1446                         *split_index = middle - delta;
1447                         break;
1448                 }
1449
1450                 /* For even counts, don't walk off the end */
1451                 if ((middle + delta + 1) == num_used)
1452                         continue;
1453
1454                 /* Now try delta past middle */
1455                 if (ocfs2_refcount_rec_no_intersect(
1456                                         &rl->rl_recs[middle + delta],
1457                                         &rl->rl_recs[middle + delta + 1])) {
1458                         *split_index = middle + delta + 1;
1459                         break;
1460                 }
1461         }
1462
1463         if (delta >= middle)
1464                 return -ENOSPC;
1465
1466         *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1467         return 0;
1468 }
1469
1470 static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1471                                             struct buffer_head *new_bh,
1472                                             u32 *split_cpos)
1473 {
1474         int split_index = 0, num_moved, ret;
1475         u32 cpos = 0;
1476         struct ocfs2_refcount_block *rb =
1477                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1478         struct ocfs2_refcount_list *rl = &rb->rf_records;
1479         struct ocfs2_refcount_block *new_rb =
1480                         (struct ocfs2_refcount_block *)new_bh->b_data;
1481         struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1482
1483         mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n",
1484              (unsigned long long)ref_leaf_bh->b_blocknr,
1485              le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used));
1486
1487         /*
1488          * XXX: Improvement later.
1489          * If we know all the high 32 bit cpos is the same, no need to sort.
1490          *
1491          * In order to make the whole process safe, we do:
1492          * 1. sort the entries by their low 32 bit cpos first so that we can
1493          *    find the split cpos easily.
1494          * 2. call ocfs2_insert_extent to insert the new refcount block.
1495          * 3. move the refcount rec to the new block.
1496          * 4. sort the entries by their 64 bit cpos.
1497          * 5. dirty the new_rb and rb.
1498          */
1499         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1500              sizeof(struct ocfs2_refcount_rec),
1501              cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1502
1503         ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1504         if (ret) {
1505                 mlog_errno(ret);
1506                 return ret;
1507         }
1508
1509         new_rb->rf_cpos = cpu_to_le32(cpos);
1510
1511         /* move refcount records starting from split_index to the new block. */
1512         num_moved = le16_to_cpu(rl->rl_used) - split_index;
1513         memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1514                num_moved * sizeof(struct ocfs2_refcount_rec));
1515
1516         /*ok, remove the entries we just moved over to the other block. */
1517         memset(&rl->rl_recs[split_index], 0,
1518                num_moved * sizeof(struct ocfs2_refcount_rec));
1519
1520         /* change old and new rl_used accordingly. */
1521         le16_add_cpu(&rl->rl_used, -num_moved);
1522         new_rl->rl_used = cpu_to_le32(num_moved);
1523
1524         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1525              sizeof(struct ocfs2_refcount_rec),
1526              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1527
1528         sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1529              sizeof(struct ocfs2_refcount_rec),
1530              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1531
1532         *split_cpos = cpos;
1533         return 0;
1534 }
1535
1536 static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1537                                          struct ocfs2_caching_info *ci,
1538                                          struct buffer_head *ref_root_bh,
1539                                          struct buffer_head *ref_leaf_bh,
1540                                          struct ocfs2_alloc_context *meta_ac)
1541 {
1542         int ret;
1543         u16 suballoc_bit_start;
1544         u32 num_got, new_cpos;
1545         u64 blkno;
1546         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1547         struct ocfs2_refcount_block *root_rb =
1548                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1549         struct buffer_head *new_bh = NULL;
1550         struct ocfs2_refcount_block *new_rb;
1551         struct ocfs2_extent_tree ref_et;
1552
1553         BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1554
1555         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1556                                       OCFS2_JOURNAL_ACCESS_WRITE);
1557         if (ret) {
1558                 mlog_errno(ret);
1559                 goto out;
1560         }
1561
1562         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1563                                       OCFS2_JOURNAL_ACCESS_WRITE);
1564         if (ret) {
1565                 mlog_errno(ret);
1566                 goto out;
1567         }
1568
1569         ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1570                                    &suballoc_bit_start, &num_got,
1571                                    &blkno);
1572         if (ret) {
1573                 mlog_errno(ret);
1574                 goto out;
1575         }
1576
1577         new_bh = sb_getblk(sb, blkno);
1578         if (new_bh == NULL) {
1579                 ret = -EIO;
1580                 mlog_errno(ret);
1581                 goto out;
1582         }
1583         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1584
1585         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1586                                       OCFS2_JOURNAL_ACCESS_CREATE);
1587         if (ret) {
1588                 mlog_errno(ret);
1589                 goto out;
1590         }
1591
1592         /* Initialize ocfs2_refcount_block. */
1593         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1594         memset(new_rb, 0, sb->s_blocksize);
1595         strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1596         new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num);
1597         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1598         new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1599         new_rb->rf_blkno = cpu_to_le64(blkno);
1600         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1601         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1602         new_rb->rf_records.rl_count =
1603                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1604         new_rb->rf_generation = root_rb->rf_generation;
1605
1606         ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1607         if (ret) {
1608                 mlog_errno(ret);
1609                 goto out;
1610         }
1611
1612         ocfs2_journal_dirty(handle, ref_leaf_bh);
1613         ocfs2_journal_dirty(handle, new_bh);
1614
1615         ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1616
1617         mlog(0, "insert new leaf block %llu at %u\n",
1618              (unsigned long long)new_bh->b_blocknr, new_cpos);
1619
1620         /* Insert the new leaf block with the specific offset cpos. */
1621         ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1622                                   1, 0, meta_ac);
1623         if (ret)
1624                 mlog_errno(ret);
1625
1626 out:
1627         brelse(new_bh);
1628         return ret;
1629 }
1630
1631 static int ocfs2_expand_refcount_tree(handle_t *handle,
1632                                       struct ocfs2_caching_info *ci,
1633                                       struct buffer_head *ref_root_bh,
1634                                       struct buffer_head *ref_leaf_bh,
1635                                       struct ocfs2_alloc_context *meta_ac)
1636 {
1637         int ret;
1638         struct buffer_head *expand_bh = NULL;
1639
1640         if (ref_root_bh == ref_leaf_bh) {
1641                 /*
1642                  * the old root bh hasn't been expanded to a b-tree,
1643                  * so expand it first.
1644                  */
1645                 ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1646                                                    &expand_bh, meta_ac);
1647                 if (ret) {
1648                         mlog_errno(ret);
1649                         goto out;
1650                 }
1651         } else {
1652                 expand_bh = ref_leaf_bh;
1653                 get_bh(expand_bh);
1654         }
1655
1656
1657         /* Now add a new refcount block into the tree.*/
1658         ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1659                                             expand_bh, meta_ac);
1660         if (ret)
1661                 mlog_errno(ret);
1662 out:
1663         brelse(expand_bh);
1664         return ret;
1665 }
1666
1667 /*
1668  * Adjust the extent rec in b-tree representing ref_leaf_bh.
1669  *
1670  * Only called when we have inserted a new refcount rec at index 0
1671  * which means ocfs2_extent_rec.e_cpos may need some change.
1672  */
1673 static int ocfs2_adjust_refcount_rec(handle_t *handle,
1674                                      struct ocfs2_caching_info *ci,
1675                                      struct buffer_head *ref_root_bh,
1676                                      struct buffer_head *ref_leaf_bh,
1677                                      struct ocfs2_refcount_rec *rec)
1678 {
1679         int ret = 0, i;
1680         u32 new_cpos, old_cpos;
1681         struct ocfs2_path *path = NULL;
1682         struct ocfs2_extent_tree et;
1683         struct ocfs2_refcount_block *rb =
1684                 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1685         struct ocfs2_extent_list *el;
1686
1687         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1688                 goto out;
1689
1690         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1691         old_cpos = le32_to_cpu(rb->rf_cpos);
1692         new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1693         if (old_cpos <= new_cpos)
1694                 goto out;
1695
1696         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1697
1698         path = ocfs2_new_path_from_et(&et);
1699         if (!path) {
1700                 ret = -ENOMEM;
1701                 mlog_errno(ret);
1702                 goto out;
1703         }
1704
1705         ret = ocfs2_find_path(ci, path, old_cpos);
1706         if (ret) {
1707                 mlog_errno(ret);
1708                 goto out;
1709         }
1710
1711         /*
1712          * 2 more credits, one for the leaf refcount block, one for
1713          * the extent block contains the extent rec.
1714          */
1715         ret = ocfs2_extend_trans(handle, handle->h_buffer_credits + 2);
1716         if (ret < 0) {
1717                 mlog_errno(ret);
1718                 goto out;
1719         }
1720
1721         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1722                                       OCFS2_JOURNAL_ACCESS_WRITE);
1723         if (ret < 0) {
1724                 mlog_errno(ret);
1725                 goto out;
1726         }
1727
1728         ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1729                                       OCFS2_JOURNAL_ACCESS_WRITE);
1730         if (ret < 0) {
1731                 mlog_errno(ret);
1732                 goto out;
1733         }
1734
1735         /* change the leaf extent block first. */
1736         el = path_leaf_el(path);
1737
1738         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1739                 if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1740                         break;
1741
1742         BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1743
1744         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1745
1746         /* change the r_cpos in the leaf block. */
1747         rb->rf_cpos = cpu_to_le32(new_cpos);
1748
1749         ocfs2_journal_dirty(handle, path_leaf_bh(path));
1750         ocfs2_journal_dirty(handle, ref_leaf_bh);
1751
1752 out:
1753         ocfs2_free_path(path);
1754         return ret;
1755 }
1756
1757 static int ocfs2_insert_refcount_rec(handle_t *handle,
1758                                      struct ocfs2_caching_info *ci,
1759                                      struct buffer_head *ref_root_bh,
1760                                      struct buffer_head *ref_leaf_bh,
1761                                      struct ocfs2_refcount_rec *rec,
1762                                      int index, int merge,
1763                                      struct ocfs2_alloc_context *meta_ac)
1764 {
1765         int ret;
1766         struct ocfs2_refcount_block *rb =
1767                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1768         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1769         struct buffer_head *new_bh = NULL;
1770
1771         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1772
1773         if (rf_list->rl_used == rf_list->rl_count) {
1774                 u64 cpos = le64_to_cpu(rec->r_cpos);
1775                 u32 len = le32_to_cpu(rec->r_clusters);
1776
1777                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1778                                                  ref_leaf_bh, meta_ac);
1779                 if (ret) {
1780                         mlog_errno(ret);
1781                         goto out;
1782                 }
1783
1784                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1785                                              cpos, len, NULL, &index,
1786                                              &new_bh);
1787                 if (ret) {
1788                         mlog_errno(ret);
1789                         goto out;
1790                 }
1791
1792                 ref_leaf_bh = new_bh;
1793                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1794                 rf_list = &rb->rf_records;
1795         }
1796
1797         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1798                                       OCFS2_JOURNAL_ACCESS_WRITE);
1799         if (ret) {
1800                 mlog_errno(ret);
1801                 goto out;
1802         }
1803
1804         if (index < le16_to_cpu(rf_list->rl_used))
1805                 memmove(&rf_list->rl_recs[index + 1],
1806                         &rf_list->rl_recs[index],
1807                         (le16_to_cpu(rf_list->rl_used) - index) *
1808                          sizeof(struct ocfs2_refcount_rec));
1809
1810         mlog(0, "insert refcount record start %llu, len %u, count %u "
1811              "to leaf block %llu at index %d\n",
1812              (unsigned long long)le64_to_cpu(rec->r_cpos),
1813              le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount),
1814              (unsigned long long)ref_leaf_bh->b_blocknr, index);
1815
1816         rf_list->rl_recs[index] = *rec;
1817
1818         le16_add_cpu(&rf_list->rl_used, 1);
1819
1820         if (merge)
1821                 ocfs2_refcount_rec_merge(rb, index);
1822
1823         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1824         if (ret) {
1825                 mlog_errno(ret);
1826                 goto out;
1827         }
1828
1829         if (index == 0) {
1830                 ret = ocfs2_adjust_refcount_rec(handle, ci,
1831                                                 ref_root_bh,
1832                                                 ref_leaf_bh, rec);
1833                 if (ret)
1834                         mlog_errno(ret);
1835         }
1836 out:
1837         brelse(new_bh);
1838         return ret;
1839 }
1840
1841 /*
1842  * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1843  * This is much simple than our b-tree code.
1844  * split_rec is the new refcount rec we want to insert.
1845  * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1846  * increase refcount or decrease a refcount to non-zero).
1847  * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1848  * rec( in case we decrease a refcount to zero).
1849  */
1850 static int ocfs2_split_refcount_rec(handle_t *handle,
1851                                     struct ocfs2_caching_info *ci,
1852                                     struct buffer_head *ref_root_bh,
1853                                     struct buffer_head *ref_leaf_bh,
1854                                     struct ocfs2_refcount_rec *split_rec,
1855                                     int index, int merge,
1856                                     struct ocfs2_alloc_context *meta_ac,
1857                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
1858 {
1859         int ret, recs_need;
1860         u32 len;
1861         struct ocfs2_refcount_block *rb =
1862                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1863         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1864         struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1865         struct ocfs2_refcount_rec *tail_rec = NULL;
1866         struct buffer_head *new_bh = NULL;
1867
1868         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1869
1870         mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n",
1871              le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters),
1872              le64_to_cpu(split_rec->r_cpos),
1873              le32_to_cpu(split_rec->r_clusters));
1874
1875         /*
1876          * If we just need to split the header or tail clusters,
1877          * no more recs are needed, just split is OK.
1878          * Otherwise we at least need one new recs.
1879          */
1880         if (!split_rec->r_refcount &&
1881             (split_rec->r_cpos == orig_rec->r_cpos ||
1882              le64_to_cpu(split_rec->r_cpos) +
1883              le32_to_cpu(split_rec->r_clusters) ==
1884              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1885                 recs_need = 0;
1886         else
1887                 recs_need = 1;
1888
1889         /*
1890          * We need one more rec if we split in the middle and the new rec have
1891          * some refcount in it.
1892          */
1893         if (split_rec->r_refcount &&
1894             (split_rec->r_cpos != orig_rec->r_cpos &&
1895              le64_to_cpu(split_rec->r_cpos) +
1896              le32_to_cpu(split_rec->r_clusters) !=
1897              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1898                 recs_need++;
1899
1900         /* If the leaf block don't have enough record, expand it. */
1901         if (le16_to_cpu(rf_list->rl_used) + recs_need > rf_list->rl_count) {
1902                 struct ocfs2_refcount_rec tmp_rec;
1903                 u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1904                 len = le32_to_cpu(orig_rec->r_clusters);
1905                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1906                                                  ref_leaf_bh, meta_ac);
1907                 if (ret) {
1908                         mlog_errno(ret);
1909                         goto out;
1910                 }
1911
1912                 /*
1913                  * We have to re-get it since now cpos may be moved to
1914                  * another leaf block.
1915                  */
1916                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1917                                              cpos, len, &tmp_rec, &index,
1918                                              &new_bh);
1919                 if (ret) {
1920                         mlog_errno(ret);
1921                         goto out;
1922                 }
1923
1924                 ref_leaf_bh = new_bh;
1925                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1926                 rf_list = &rb->rf_records;
1927                 orig_rec = &rf_list->rl_recs[index];
1928         }
1929
1930         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1931                                       OCFS2_JOURNAL_ACCESS_WRITE);
1932         if (ret) {
1933                 mlog_errno(ret);
1934                 goto out;
1935         }
1936
1937         /*
1938          * We have calculated out how many new records we need and store
1939          * in recs_need, so spare enough space first by moving the records
1940          * after "index" to the end.
1941          */
1942         if (index != le16_to_cpu(rf_list->rl_used) - 1)
1943                 memmove(&rf_list->rl_recs[index + 1 + recs_need],
1944                         &rf_list->rl_recs[index + 1],
1945                         (le16_to_cpu(rf_list->rl_used) - index - 1) *
1946                          sizeof(struct ocfs2_refcount_rec));
1947
1948         len = (le64_to_cpu(orig_rec->r_cpos) +
1949               le32_to_cpu(orig_rec->r_clusters)) -
1950               (le64_to_cpu(split_rec->r_cpos) +
1951               le32_to_cpu(split_rec->r_clusters));
1952
1953         /*
1954          * If we have "len", the we will split in the tail and move it
1955          * to the end of the space we have just spared.
1956          */
1957         if (len) {
1958                 tail_rec = &rf_list->rl_recs[index + recs_need];
1959
1960                 memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1961                 le64_add_cpu(&tail_rec->r_cpos,
1962                              le32_to_cpu(tail_rec->r_clusters) - len);
1963                 tail_rec->r_clusters = le32_to_cpu(len);
1964         }
1965
1966         /*
1967          * If the split pos isn't the same as the original one, we need to
1968          * split in the head.
1969          *
1970          * Note: We have the chance that split_rec.r_refcount = 0,
1971          * recs_need = 0 and len > 0, which means we just cut the head from
1972          * the orig_rec and in that case we have done some modification in
1973          * orig_rec above, so the check for r_cpos is faked.
1974          */
1975         if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1976                 len = le64_to_cpu(split_rec->r_cpos) -
1977                       le64_to_cpu(orig_rec->r_cpos);
1978                 orig_rec->r_clusters = cpu_to_le32(len);
1979                 index++;
1980         }
1981
1982         le16_add_cpu(&rf_list->rl_used, recs_need);
1983
1984         if (split_rec->r_refcount) {
1985                 rf_list->rl_recs[index] = *split_rec;
1986                 mlog(0, "insert refcount record start %llu, len %u, count %u "
1987                      "to leaf block %llu at index %d\n",
1988                      (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1989                      le32_to_cpu(split_rec->r_clusters),
1990                      le32_to_cpu(split_rec->r_refcount),
1991                      (unsigned long long)ref_leaf_bh->b_blocknr, index);
1992
1993                 if (merge)
1994                         ocfs2_refcount_rec_merge(rb, index);
1995         }
1996
1997         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1998         if (ret)
1999                 mlog_errno(ret);
2000
2001 out:
2002         brelse(new_bh);
2003         return ret;
2004 }
2005
2006 static int __ocfs2_increase_refcount(handle_t *handle,
2007                                      struct ocfs2_caching_info *ci,
2008                                      struct buffer_head *ref_root_bh,
2009                                      u64 cpos, u32 len, int merge,
2010                                      struct ocfs2_alloc_context *meta_ac,
2011                                      struct ocfs2_cached_dealloc_ctxt *dealloc)
2012 {
2013         int ret = 0, index;
2014         struct buffer_head *ref_leaf_bh = NULL;
2015         struct ocfs2_refcount_rec rec;
2016         unsigned int set_len = 0;
2017
2018         mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n",
2019              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2020              (unsigned long long)cpos, len);
2021
2022         while (len) {
2023                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2024                                              cpos, len, &rec, &index,
2025                                              &ref_leaf_bh);
2026                 if (ret) {
2027                         mlog_errno(ret);
2028                         goto out;
2029                 }
2030
2031                 set_len = le32_to_cpu(rec.r_clusters);
2032
2033                 /*
2034                  * Here we may meet with 3 situations:
2035                  *
2036                  * 1. If we find an already existing record, and the length
2037                  *    is the same, cool, we just need to increase the r_refcount
2038                  *    and it is OK.
2039                  * 2. If we find a hole, just insert it with r_refcount = 1.
2040                  * 3. If we are in the middle of one extent record, split
2041                  *    it.
2042                  */
2043                 if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
2044                     set_len <= len) {
2045                         mlog(0, "increase refcount rec, start %llu, len %u, "
2046                              "count %u\n", (unsigned long long)cpos, set_len,
2047                              le32_to_cpu(rec.r_refcount));
2048                         ret = ocfs2_change_refcount_rec(handle, ci,
2049                                                         ref_leaf_bh, index,
2050                                                         merge, 1);
2051                         if (ret) {
2052                                 mlog_errno(ret);
2053                                 goto out;
2054                         }
2055                 } else if (!rec.r_refcount) {
2056                         rec.r_refcount = cpu_to_le32(1);
2057
2058                         mlog(0, "insert refcount rec, start %llu, len %u\n",
2059                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2060                              set_len);
2061                         ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
2062                                                         ref_leaf_bh,
2063                                                         &rec, index,
2064                                                         merge, meta_ac);
2065                         if (ret) {
2066                                 mlog_errno(ret);
2067                                 goto out;
2068                         }
2069                 } else  {
2070                         set_len = min((u64)(cpos + len),
2071                                       le64_to_cpu(rec.r_cpos) + set_len) - cpos;
2072                         rec.r_cpos = cpu_to_le64(cpos);
2073                         rec.r_clusters = cpu_to_le32(set_len);
2074                         le32_add_cpu(&rec.r_refcount, 1);
2075
2076                         mlog(0, "split refcount rec, start %llu, "
2077                              "len %u, count %u\n",
2078                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2079                              set_len, le32_to_cpu(rec.r_refcount));
2080                         ret = ocfs2_split_refcount_rec(handle, ci,
2081                                                        ref_root_bh, ref_leaf_bh,
2082                                                        &rec, index, merge,
2083                                                        meta_ac, dealloc);
2084                         if (ret) {
2085                                 mlog_errno(ret);
2086                                 goto out;
2087                         }
2088                 }
2089
2090                 cpos += set_len;
2091                 len -= set_len;
2092                 brelse(ref_leaf_bh);
2093                 ref_leaf_bh = NULL;
2094         }
2095
2096 out:
2097         brelse(ref_leaf_bh);
2098         return ret;
2099 }
2100
2101 static int ocfs2_remove_refcount_extent(handle_t *handle,
2102                                 struct ocfs2_caching_info *ci,
2103                                 struct buffer_head *ref_root_bh,
2104                                 struct buffer_head *ref_leaf_bh,
2105                                 struct ocfs2_alloc_context *meta_ac,
2106                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2107 {
2108         int ret;
2109         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2110         struct ocfs2_refcount_block *rb =
2111                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2112         struct ocfs2_extent_tree et;
2113
2114         BUG_ON(rb->rf_records.rl_used);
2115
2116         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2117         ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
2118                                   1, meta_ac, dealloc);
2119         if (ret) {
2120                 mlog_errno(ret);
2121                 goto out;
2122         }
2123
2124         ocfs2_remove_from_cache(ci, ref_leaf_bh);
2125
2126         /*
2127          * add the freed block to the dealloc so that it will be freed
2128          * when we run dealloc.
2129          */
2130         ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
2131                                         le16_to_cpu(rb->rf_suballoc_slot),
2132                                         le64_to_cpu(rb->rf_blkno),
2133                                         le16_to_cpu(rb->rf_suballoc_bit));
2134         if (ret) {
2135                 mlog_errno(ret);
2136                 goto out;
2137         }
2138
2139         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
2140                                       OCFS2_JOURNAL_ACCESS_WRITE);
2141         if (ret) {
2142                 mlog_errno(ret);
2143                 goto out;
2144         }
2145
2146         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2147
2148         le32_add_cpu(&rb->rf_clusters, -1);
2149
2150         /*
2151          * check whether we need to restore the root refcount block if
2152          * there is no leaf extent block at atll.
2153          */
2154         if (!rb->rf_list.l_next_free_rec) {
2155                 BUG_ON(rb->rf_clusters);
2156
2157                 mlog(0, "reset refcount tree root %llu to be a record block.\n",
2158                      (unsigned long long)ref_root_bh->b_blocknr);
2159
2160                 rb->rf_flags = 0;
2161                 rb->rf_parent = 0;
2162                 rb->rf_cpos = 0;
2163                 memset(&rb->rf_records, 0, sb->s_blocksize -
2164                        offsetof(struct ocfs2_refcount_block, rf_records));
2165                 rb->rf_records.rl_count =
2166                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
2167         }
2168
2169         ocfs2_journal_dirty(handle, ref_root_bh);
2170
2171 out:
2172         return ret;
2173 }
2174
2175 int ocfs2_increase_refcount(handle_t *handle,
2176                             struct ocfs2_caching_info *ci,
2177                             struct buffer_head *ref_root_bh,
2178                             u64 cpos, u32 len,
2179                             struct ocfs2_alloc_context *meta_ac,
2180                             struct ocfs2_cached_dealloc_ctxt *dealloc)
2181 {
2182         return __ocfs2_increase_refcount(handle, ci, ref_root_bh,
2183                                          cpos, len, 1,
2184                                          meta_ac, dealloc);
2185 }
2186
2187 static int ocfs2_decrease_refcount_rec(handle_t *handle,
2188                                 struct ocfs2_caching_info *ci,
2189                                 struct buffer_head *ref_root_bh,
2190                                 struct buffer_head *ref_leaf_bh,
2191                                 int index, u64 cpos, unsigned int len,
2192                                 struct ocfs2_alloc_context *meta_ac,
2193                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2194 {
2195         int ret;
2196         struct ocfs2_refcount_block *rb =
2197                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2198         struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
2199
2200         BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
2201         BUG_ON(cpos + len >
2202                le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
2203
2204         if (cpos == le64_to_cpu(rec->r_cpos) &&
2205             len == le32_to_cpu(rec->r_clusters))
2206                 ret = ocfs2_change_refcount_rec(handle, ci,
2207                                                 ref_leaf_bh, index, 1, -1);
2208         else {
2209                 struct ocfs2_refcount_rec split = *rec;
2210                 split.r_cpos = cpu_to_le64(cpos);
2211                 split.r_clusters = cpu_to_le32(len);
2212
2213                 le32_add_cpu(&split.r_refcount, -1);
2214
2215                 mlog(0, "split refcount rec, start %llu, "
2216                      "len %u, count %u, original start %llu, len %u\n",
2217                      (unsigned long long)le64_to_cpu(split.r_cpos),
2218                      len, le32_to_cpu(split.r_refcount),
2219                      (unsigned long long)le64_to_cpu(rec->r_cpos),
2220                      le32_to_cpu(rec->r_clusters));
2221                 ret = ocfs2_split_refcount_rec(handle, ci,
2222                                                ref_root_bh, ref_leaf_bh,
2223                                                &split, index, 1,
2224                                                meta_ac, dealloc);
2225         }
2226
2227         if (ret) {
2228                 mlog_errno(ret);
2229                 goto out;
2230         }
2231
2232         /* Remove the leaf refcount block if it contains no refcount record. */
2233         if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
2234                 ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
2235                                                    ref_leaf_bh, meta_ac,
2236                                                    dealloc);
2237                 if (ret)
2238                         mlog_errno(ret);
2239         }
2240
2241 out:
2242         return ret;
2243 }
2244
2245 static int __ocfs2_decrease_refcount(handle_t *handle,
2246                                      struct ocfs2_caching_info *ci,
2247                                      struct buffer_head *ref_root_bh,
2248                                      u64 cpos, u32 len,
2249                                      struct ocfs2_alloc_context *meta_ac,
2250                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2251                                      int delete)
2252 {
2253         int ret = 0, index = 0;
2254         struct ocfs2_refcount_rec rec;
2255         unsigned int r_count = 0, r_len;
2256         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2257         struct buffer_head *ref_leaf_bh = NULL;
2258
2259         mlog(0, "Tree owner %llu, decrease refcount start %llu, "
2260              "len %u, delete %u\n",
2261              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2262              (unsigned long long)cpos, len, delete);
2263
2264         while (len) {
2265                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2266                                              cpos, len, &rec, &index,
2267                                              &ref_leaf_bh);
2268                 if (ret) {
2269                         mlog_errno(ret);
2270                         goto out;
2271                 }
2272
2273                 r_count = le32_to_cpu(rec.r_refcount);
2274                 BUG_ON(r_count == 0);
2275                 if (!delete)
2276                         BUG_ON(r_count > 1);
2277
2278                 r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
2279                               le32_to_cpu(rec.r_clusters)) - cpos;
2280
2281                 ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
2282                                                   ref_leaf_bh, index,
2283                                                   cpos, r_len,
2284                                                   meta_ac, dealloc);
2285                 if (ret) {
2286                         mlog_errno(ret);
2287                         goto out;
2288                 }
2289
2290                 if (le32_to_cpu(rec.r_refcount) == 1 && delete) {
2291                         ret = ocfs2_cache_cluster_dealloc(dealloc,
2292                                           ocfs2_clusters_to_blocks(sb, cpos),
2293                                                           r_len);
2294                         if (ret) {
2295                                 mlog_errno(ret);
2296                                 goto out;
2297                         }
2298                 }
2299
2300                 cpos += r_len;
2301                 len -= r_len;
2302                 brelse(ref_leaf_bh);
2303                 ref_leaf_bh = NULL;
2304         }
2305
2306 out:
2307         brelse(ref_leaf_bh);
2308         return ret;
2309 }
2310
2311 /* Caller must hold refcount tree lock. */
2312 int ocfs2_decrease_refcount(struct inode *inode,
2313                             handle_t *handle, u32 cpos, u32 len,
2314                             struct ocfs2_alloc_context *meta_ac,
2315                             struct ocfs2_cached_dealloc_ctxt *dealloc,
2316                             int delete)
2317 {
2318         int ret;
2319         u64 ref_blkno;
2320         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2321         struct buffer_head *ref_root_bh = NULL;
2322         struct ocfs2_refcount_tree *tree;
2323
2324         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2325
2326         ret = ocfs2_get_refcount_block(inode, &ref_blkno);
2327         if (ret) {
2328                 mlog_errno(ret);
2329                 goto out;
2330         }
2331
2332         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
2333         if (ret) {
2334                 mlog_errno(ret);
2335                 goto out;
2336         }
2337
2338         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
2339                                         &ref_root_bh);
2340         if (ret) {
2341                 mlog_errno(ret);
2342                 goto out;
2343         }
2344
2345         ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
2346                                         cpos, len, meta_ac, dealloc, delete);
2347         if (ret)
2348                 mlog_errno(ret);
2349 out:
2350         brelse(ref_root_bh);
2351         return ret;
2352 }
2353
2354 /*
2355  * Mark the already-existing extent at cpos as refcounted for len clusters.
2356  * This adds the refcount extent flag.
2357  *
2358  * If the existing extent is larger than the request, initiate a
2359  * split. An attempt will be made at merging with adjacent extents.
2360  *
2361  * The caller is responsible for passing down meta_ac if we'll need it.
2362  */
2363 static int ocfs2_mark_extent_refcounted(struct inode *inode,
2364                                 struct ocfs2_extent_tree *et,
2365                                 handle_t *handle, u32 cpos,
2366                                 u32 len, u32 phys,
2367                                 struct ocfs2_alloc_context *meta_ac,
2368                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2369 {
2370         int ret;
2371
2372         mlog(0, "Inode %lu refcount tree cpos %u, len %u, phys cluster %u\n",
2373              inode->i_ino, cpos, len, phys);
2374
2375         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2376                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2377                             "tree, but the feature bit is not set in the "
2378                             "super block.", inode->i_ino);
2379                 ret = -EROFS;
2380                 goto out;
2381         }
2382
2383         ret = ocfs2_change_extent_flag(handle, et, cpos,
2384                                        len, phys, meta_ac, dealloc,
2385                                        OCFS2_EXT_REFCOUNTED, 0);
2386         if (ret)
2387                 mlog_errno(ret);
2388
2389 out:
2390         return ret;
2391 }
2392
2393 /*
2394  * Given some contiguous physical clusters, calculate what we need
2395  * for modifying their refcount.
2396  */
2397 static int ocfs2_calc_refcount_meta_credits(struct super_block *sb,
2398                                             struct ocfs2_caching_info *ci,
2399                                             struct buffer_head *ref_root_bh,
2400                                             u64 start_cpos,
2401                                             u32 clusters,
2402                                             int *meta_add,
2403                                             int *credits)
2404 {
2405         int ret = 0, index, ref_blocks = 0, recs_add = 0;
2406         u64 cpos = start_cpos;
2407         struct ocfs2_refcount_block *rb;
2408         struct ocfs2_refcount_rec rec;
2409         struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL;
2410         u32 len;
2411
2412         mlog(0, "start_cpos %llu, clusters %u\n",
2413              (unsigned long long)start_cpos, clusters);
2414         while (clusters) {
2415                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2416                                              cpos, clusters, &rec,
2417                                              &index, &ref_leaf_bh);
2418                 if (ret) {
2419                         mlog_errno(ret);
2420                         goto out;
2421                 }
2422
2423                 if (ref_leaf_bh != prev_bh) {
2424                         /*
2425                          * Now we encounter a new leaf block, so calculate
2426                          * whether we need to extend the old leaf.
2427                          */
2428                         if (prev_bh) {
2429                                 rb = (struct ocfs2_refcount_block *)
2430                                                         prev_bh->b_data;
2431
2432                                 if (le64_to_cpu(rb->rf_records.rl_used) +
2433                                     recs_add >
2434                                     le16_to_cpu(rb->rf_records.rl_count))
2435                                         ref_blocks++;
2436                         }
2437
2438                         recs_add = 0;
2439                         *credits += 1;
2440                         brelse(prev_bh);
2441                         prev_bh = ref_leaf_bh;
2442                         get_bh(prev_bh);
2443                 }
2444
2445                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2446
2447                 mlog(0, "recs_add %d,cpos %llu, clusters %u, rec->r_cpos %llu,"
2448                      "rec->r_clusters %u, rec->r_refcount %u, index %d\n",
2449                      recs_add, (unsigned long long)cpos, clusters,
2450                      (unsigned long long)le64_to_cpu(rec.r_cpos),
2451                      le32_to_cpu(rec.r_clusters),
2452                      le32_to_cpu(rec.r_refcount), index);
2453
2454                 len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) +
2455                           le32_to_cpu(rec.r_clusters)) - cpos;
2456                 /*
2457                  * We record all the records which will be inserted to the
2458                  * same refcount block, so that we can tell exactly whether
2459                  * we need a new refcount block or not.
2460                  *
2461                  * If we will insert a new one, this is easy and only happens
2462                  * during adding refcounted flag to the extent, so we don't
2463                  * have a chance of spliting. We just need one record.
2464                  *
2465                  * If the refcount rec already exists, that would be a little
2466                  * complicated. we may have to:
2467                  * 1) split at the beginning if the start pos isn't aligned.
2468                  *    we need 1 more record in this case.
2469                  * 2) split int the end if the end pos isn't aligned.
2470                  *    we need 1 more record in this case.
2471                  * 3) split in the middle because of file system fragmentation.
2472                  *    we need 2 more records in this case(we can't detect this
2473                  *    beforehand, so always think of the worst case).
2474                  */
2475                 if (rec.r_refcount) {
2476                         recs_add += 2;
2477                         /* Check whether we need a split at the beginning. */
2478                         if (cpos == start_cpos &&
2479                             cpos != le64_to_cpu(rec.r_cpos))
2480                                 recs_add++;
2481
2482                         /* Check whether we need a split in the end. */
2483                         if (cpos + clusters < le64_to_cpu(rec.r_cpos) +
2484                             le32_to_cpu(rec.r_clusters))
2485                                 recs_add++;
2486                 } else
2487                         recs_add++;
2488
2489                 brelse(ref_leaf_bh);
2490                 ref_leaf_bh = NULL;
2491                 clusters -= len;
2492                 cpos += len;
2493         }
2494
2495         if (prev_bh) {
2496                 rb = (struct ocfs2_refcount_block *)prev_bh->b_data;
2497
2498                 if (le64_to_cpu(rb->rf_records.rl_used) + recs_add >
2499                     le16_to_cpu(rb->rf_records.rl_count))
2500                         ref_blocks++;
2501
2502                 *credits += 1;
2503         }
2504
2505         if (!ref_blocks)
2506                 goto out;
2507
2508         mlog(0, "we need ref_blocks %d\n", ref_blocks);
2509         *meta_add += ref_blocks;
2510         *credits += ref_blocks;
2511
2512         /*
2513          * So we may need ref_blocks to insert into the tree.
2514          * That also means we need to change the b-tree and add that number
2515          * of records since we never merge them.
2516          * We need one more block for expansion since the new created leaf
2517          * block is also full and needs split.
2518          */
2519         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2520         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) {
2521                 struct ocfs2_extent_tree et;
2522
2523                 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2524                 *meta_add += ocfs2_extend_meta_needed(et.et_root_el);
2525                 *credits += ocfs2_calc_extend_credits(sb,
2526                                                       et.et_root_el,
2527                                                       ref_blocks);
2528         } else {
2529                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
2530                 *meta_add += 1;
2531         }
2532
2533 out:
2534         brelse(ref_leaf_bh);
2535         brelse(prev_bh);
2536         return ret;
2537 }
2538
2539 /*
2540  * For refcount tree, we will decrease some contiguous clusters
2541  * refcount count, so just go through it to see how many blocks
2542  * we gonna touch and whether we need to create new blocks.
2543  *
2544  * Normally the refcount blocks store these refcount should be
2545  * continguous also, so that we can get the number easily.
2546  * As for meta_ac, we will at most add split 2 refcount record and
2547  * 2 more refcount block, so just check it in a rough way.
2548  *
2549  * Caller must hold refcount tree lock.
2550  */
2551 int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
2552                                           struct buffer_head *di_bh,
2553                                           u64 phys_blkno,
2554                                           u32 clusters,
2555                                           int *credits,
2556                                           struct ocfs2_alloc_context **meta_ac)
2557 {
2558         int ret, ref_blocks = 0;
2559         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2560         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2561         struct buffer_head *ref_root_bh = NULL;
2562         struct ocfs2_refcount_tree *tree;
2563         u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno);
2564
2565         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2566                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2567                             "tree, but the feature bit is not set in the "
2568                             "super block.", inode->i_ino);
2569                 ret = -EROFS;
2570                 goto out;
2571         }
2572
2573         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2574
2575         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb),
2576                                       le64_to_cpu(di->i_refcount_loc), &tree);
2577         if (ret) {
2578                 mlog_errno(ret);
2579                 goto out;
2580         }
2581
2582         ret = ocfs2_read_refcount_block(&tree->rf_ci,
2583                                         le64_to_cpu(di->i_refcount_loc),
2584                                         &ref_root_bh);
2585         if (ret) {
2586                 mlog_errno(ret);
2587                 goto out;
2588         }
2589
2590         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
2591                                                &tree->rf_ci,
2592                                                ref_root_bh,
2593                                                start_cpos, clusters,
2594                                                &ref_blocks, credits);
2595         if (ret) {
2596                 mlog_errno(ret);
2597                 goto out;
2598         }
2599
2600         mlog(0, "reserve new metadata %d, credits = %d\n",
2601              ref_blocks, *credits);
2602
2603         if (ref_blocks) {
2604                 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
2605                                                         ref_blocks, meta_ac);
2606                 if (ret)
2607                         mlog_errno(ret);
2608         }
2609
2610 out:
2611         brelse(ref_root_bh);
2612         return ret;
2613 }
2614
2615 #define MAX_CONTIG_BYTES        1048576
2616
2617 static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb)
2618 {
2619         return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES);
2620 }
2621
2622 static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb)
2623 {
2624         return ~(ocfs2_cow_contig_clusters(sb) - 1);
2625 }
2626
2627 /*
2628  * Given an extent that starts at 'start' and an I/O that starts at 'cpos',
2629  * find an offset (start + (n * contig_clusters)) that is closest to cpos
2630  * while still being less than or equal to it.
2631  *
2632  * The goal is to break the extent at a multiple of contig_clusters.
2633  */
2634 static inline unsigned int ocfs2_cow_align_start(struct super_block *sb,
2635                                                  unsigned int start,
2636                                                  unsigned int cpos)
2637 {
2638         BUG_ON(start > cpos);
2639
2640         return start + ((cpos - start) & ocfs2_cow_contig_mask(sb));
2641 }
2642
2643 /*
2644  * Given a cluster count of len, pad it out so that it is a multiple
2645  * of contig_clusters.
2646  */
2647 static inline unsigned int ocfs2_cow_align_length(struct super_block *sb,
2648                                                   unsigned int len)
2649 {
2650         unsigned int padded =
2651                 (len + (ocfs2_cow_contig_clusters(sb) - 1)) &
2652                 ocfs2_cow_contig_mask(sb);
2653
2654         /* Did we wrap? */
2655         if (padded < len)
2656                 padded = UINT_MAX;
2657
2658         return padded;
2659 }
2660
2661 /*
2662  * Calculate out the start and number of virtual clusters we need to to CoW.
2663  *
2664  * cpos is vitual start cluster position we want to do CoW in a
2665  * file and write_len is the cluster length.
2666  * max_cpos is the place where we want to stop CoW intentionally.
2667  *
2668  * Normal we will start CoW from the beginning of extent record cotaining cpos.
2669  * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we
2670  * get good I/O from the resulting extent tree.
2671  */
2672 static int ocfs2_refcount_cal_cow_clusters(struct inode *inode,
2673                                            struct ocfs2_extent_list *el,
2674                                            u32 cpos,
2675                                            u32 write_len,
2676                                            u32 max_cpos,
2677                                            u32 *cow_start,
2678                                            u32 *cow_len)
2679 {
2680         int ret = 0;
2681         int tree_height = le16_to_cpu(el->l_tree_depth), i;
2682         struct buffer_head *eb_bh = NULL;
2683         struct ocfs2_extent_block *eb = NULL;
2684         struct ocfs2_extent_rec *rec;
2685         unsigned int want_clusters, rec_end = 0;
2686         int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb);
2687         int leaf_clusters;
2688
2689         BUG_ON(cpos + write_len > max_cpos);
2690
2691         if (tree_height > 0) {
2692                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh);
2693                 if (ret) {
2694                         mlog_errno(ret);
2695                         goto out;
2696                 }
2697
2698                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2699                 el = &eb->h_list;
2700
2701                 if (el->l_tree_depth) {
2702                         ocfs2_error(inode->i_sb,
2703                                     "Inode %lu has non zero tree depth in "
2704                                     "leaf block %llu\n", inode->i_ino,
2705                                     (unsigned long long)eb_bh->b_blocknr);
2706                         ret = -EROFS;
2707                         goto out;
2708                 }
2709         }
2710
2711         *cow_len = 0;
2712         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
2713                 rec = &el->l_recs[i];
2714
2715                 if (ocfs2_is_empty_extent(rec)) {
2716                         mlog_bug_on_msg(i != 0, "Inode %lu has empty record in "
2717                                         "index %d\n", inode->i_ino, i);
2718                         continue;
2719                 }
2720
2721                 if (le32_to_cpu(rec->e_cpos) +
2722                     le16_to_cpu(rec->e_leaf_clusters) <= cpos)
2723                         continue;
2724
2725                 if (*cow_len == 0) {
2726                         /*
2727                          * We should find a refcounted record in the
2728                          * first pass.
2729                          */
2730                         BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED));
2731                         *cow_start = le32_to_cpu(rec->e_cpos);
2732                 }
2733
2734                 /*
2735                  * If we encounter a hole, a non-refcounted record or
2736                  * pass the max_cpos, stop the search.
2737                  */
2738                 if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) ||
2739                     (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) ||
2740                     (max_cpos <= le32_to_cpu(rec->e_cpos)))
2741                         break;
2742
2743                 leaf_clusters = le16_to_cpu(rec->e_leaf_clusters);
2744                 rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters;
2745                 if (rec_end > max_cpos) {
2746                         rec_end = max_cpos;
2747                         leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos);
2748                 }
2749
2750                 /*
2751                  * How many clusters do we actually need from
2752                  * this extent?  First we see how many we actually
2753                  * need to complete the write.  If that's smaller
2754                  * than contig_clusters, we try for contig_clusters.
2755                  */
2756                 if (!*cow_len)
2757                         want_clusters = write_len;
2758                 else
2759                         want_clusters = (cpos + write_len) -
2760                                 (*cow_start + *cow_len);
2761                 if (want_clusters < contig_clusters)
2762                         want_clusters = contig_clusters;
2763
2764                 /*
2765                  * If the write does not cover the whole extent, we
2766                  * need to calculate how we're going to split the extent.
2767                  * We try to do it on contig_clusters boundaries.
2768                  *
2769                  * Any extent smaller than contig_clusters will be
2770                  * CoWed in its entirety.
2771                  */
2772                 if (leaf_clusters <= contig_clusters)
2773                         *cow_len += leaf_clusters;
2774                 else if (*cow_len || (*cow_start == cpos)) {
2775                         /*
2776                          * This extent needs to be CoW'd from its
2777                          * beginning, so all we have to do is compute
2778                          * how many clusters to grab.  We align
2779                          * want_clusters to the edge of contig_clusters
2780                          * to get better I/O.
2781                          */
2782                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2783                                                                want_clusters);
2784
2785                         if (leaf_clusters < want_clusters)
2786                                 *cow_len += leaf_clusters;
2787                         else
2788                                 *cow_len += want_clusters;
2789                 } else if ((*cow_start + contig_clusters) >=
2790                            (cpos + write_len)) {
2791                         /*
2792                          * Breaking off contig_clusters at the front
2793                          * of the extent will cover our write.  That's
2794                          * easy.
2795                          */
2796                         *cow_len = contig_clusters;
2797                 } else if ((rec_end - cpos) <= contig_clusters) {
2798                         /*
2799                          * Breaking off contig_clusters at the tail of
2800                          * this extent will cover cpos.
2801                          */
2802                         *cow_start = rec_end - contig_clusters;
2803                         *cow_len = contig_clusters;
2804                 } else if ((rec_end - cpos) <= want_clusters) {
2805                         /*
2806                          * While we can't fit the entire write in this
2807                          * extent, we know that the write goes from cpos
2808                          * to the end of the extent.  Break that off.
2809                          * We try to break it at some multiple of
2810                          * contig_clusters from the front of the extent.
2811                          * Failing that (ie, cpos is within
2812                          * contig_clusters of the front), we'll CoW the
2813                          * entire extent.
2814                          */
2815                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2816                                                            *cow_start, cpos);
2817                         *cow_len = rec_end - *cow_start;
2818                 } else {
2819                         /*
2820                          * Ok, the entire write lives in the middle of
2821                          * this extent.  Let's try to slice the extent up
2822                          * nicely.  Optimally, our CoW region starts at
2823                          * m*contig_clusters from the beginning of the
2824                          * extent and goes for n*contig_clusters,
2825                          * covering the entire write.
2826                          */
2827                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2828                                                            *cow_start, cpos);
2829
2830                         want_clusters = (cpos + write_len) - *cow_start;
2831                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2832                                                                want_clusters);
2833                         if (*cow_start + want_clusters <= rec_end)
2834                                 *cow_len = want_clusters;
2835                         else
2836                                 *cow_len = rec_end - *cow_start;
2837                 }
2838
2839                 /* Have we covered our entire write yet? */
2840                 if ((*cow_start + *cow_len) >= (cpos + write_len))
2841                         break;
2842
2843                 /*
2844                  * If we reach the end of the extent block and don't get enough
2845                  * clusters, continue with the next extent block if possible.
2846                  */
2847                 if (i + 1 == le16_to_cpu(el->l_next_free_rec) &&
2848                     eb && eb->h_next_leaf_blk) {
2849                         brelse(eb_bh);
2850                         eb_bh = NULL;
2851
2852                         ret = ocfs2_read_extent_block(INODE_CACHE(inode),
2853                                                le64_to_cpu(eb->h_next_leaf_blk),
2854                                                &eb_bh);
2855                         if (ret) {
2856                                 mlog_errno(ret);
2857                                 goto out;
2858                         }
2859
2860                         eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2861                         el = &eb->h_list;
2862                         i = -1;
2863                 }
2864         }
2865
2866 out:
2867         brelse(eb_bh);
2868         return ret;
2869 }
2870
2871 /*
2872  * Prepare meta_ac, data_ac and calculate credits when we want to add some
2873  * num_clusters in data_tree "et" and change the refcount for the old
2874  * clusters(starting form p_cluster) in the refcount tree.
2875  *
2876  * Note:
2877  * 1. since we may split the old tree, so we at most will need num_clusters + 2
2878  *    more new leaf records.
2879  * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so
2880  *    just give data_ac = NULL.
2881  */
2882 static int ocfs2_lock_refcount_allocators(struct super_block *sb,
2883                                         u32 p_cluster, u32 num_clusters,
2884                                         struct ocfs2_extent_tree *et,
2885                                         struct ocfs2_caching_info *ref_ci,
2886                                         struct buffer_head *ref_root_bh,
2887                                         struct ocfs2_alloc_context **meta_ac,
2888                                         struct ocfs2_alloc_context **data_ac,
2889                                         int *credits)
2890 {
2891         int ret = 0, meta_add = 0;
2892         int num_free_extents = ocfs2_num_free_extents(OCFS2_SB(sb), et);
2893
2894         if (num_free_extents < 0) {
2895                 ret = num_free_extents;
2896                 mlog_errno(ret);
2897                 goto out;
2898         }
2899
2900         if (num_free_extents < num_clusters + 2)
2901                 meta_add =
2902                         ocfs2_extend_meta_needed(et->et_root_el);
2903
2904         *credits += ocfs2_calc_extend_credits(sb, et->et_root_el,
2905                                               num_clusters + 2);
2906
2907         ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh,
2908                                                p_cluster, num_clusters,
2909                                                &meta_add, credits);
2910         if (ret) {
2911                 mlog_errno(ret);
2912                 goto out;
2913         }
2914
2915         mlog(0, "reserve new metadata %d, clusters %u, credits = %d\n",
2916              meta_add, num_clusters, *credits);
2917         ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add,
2918                                                 meta_ac);
2919         if (ret) {
2920                 mlog_errno(ret);
2921                 goto out;
2922         }
2923
2924         if (data_ac) {
2925                 ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters,
2926                                              data_ac);
2927                 if (ret)
2928                         mlog_errno(ret);
2929         }
2930
2931 out:
2932         if (ret) {
2933                 if (*meta_ac) {
2934                         ocfs2_free_alloc_context(*meta_ac);
2935                         *meta_ac = NULL;
2936                 }
2937         }
2938
2939         return ret;
2940 }
2941
2942 static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh)
2943 {
2944         BUG_ON(buffer_dirty(bh));
2945
2946         clear_buffer_mapped(bh);
2947
2948         return 0;
2949 }
2950
2951 static int ocfs2_duplicate_clusters_by_page(handle_t *handle,
2952                                             struct ocfs2_cow_context *context,
2953                                             u32 cpos, u32 old_cluster,
2954                                             u32 new_cluster, u32 new_len)
2955 {
2956         int ret = 0, partial;
2957         struct ocfs2_caching_info *ci = context->data_et.et_ci;
2958         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2959         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
2960         struct page *page;
2961         pgoff_t page_index;
2962         unsigned int from, to;
2963         loff_t offset, end, map_end;
2964         struct address_space *mapping = context->inode->i_mapping;
2965
2966         mlog(0, "old_cluster %u, new %u, len %u at offset %u\n", old_cluster,
2967              new_cluster, new_len, cpos);
2968
2969         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
2970         end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
2971
2972         while (offset < end) {
2973                 page_index = offset >> PAGE_CACHE_SHIFT;
2974                 map_end = (page_index + 1) << PAGE_CACHE_SHIFT;
2975                 if (map_end > end)
2976                         map_end = end;
2977
2978                 /* from, to is the offset within the page. */
2979                 from = offset & (PAGE_CACHE_SIZE - 1);
2980                 to = PAGE_CACHE_SIZE;
2981                 if (map_end & (PAGE_CACHE_SIZE - 1))
2982                         to = map_end & (PAGE_CACHE_SIZE - 1);
2983
2984                 page = grab_cache_page(mapping, page_index);
2985
2986                 /* This page can't be dirtied before we CoW it out. */
2987                 BUG_ON(PageDirty(page));
2988
2989                 if (!PageUptodate(page)) {
2990                         ret = block_read_full_page(page, ocfs2_get_block);
2991                         if (ret) {
2992                                 mlog_errno(ret);
2993                                 goto unlock;
2994                         }
2995                         lock_page(page);
2996                 }
2997
2998                 if (page_has_buffers(page)) {
2999                         ret = walk_page_buffers(handle, page_buffers(page),
3000                                                 from, to, &partial,
3001                                                 ocfs2_clear_cow_buffer);
3002                         if (ret) {
3003                                 mlog_errno(ret);
3004                                 goto unlock;
3005                         }
3006                 }
3007
3008                 ocfs2_map_and_dirty_page(context->inode,
3009                                          handle, from, to,
3010                                          page, 0, &new_block);
3011                 mark_page_accessed(page);
3012 unlock:
3013                 unlock_page(page);
3014                 page_cache_release(page);
3015                 page = NULL;
3016                 offset = map_end;
3017                 if (ret)
3018                         break;
3019         }
3020
3021         return ret;
3022 }
3023
3024 static int ocfs2_duplicate_clusters_by_jbd(handle_t *handle,
3025                                            struct ocfs2_cow_context *context,
3026                                            u32 cpos, u32 old_cluster,
3027                                            u32 new_cluster, u32 new_len)
3028 {
3029         int ret = 0;
3030         struct super_block *sb = context->inode->i_sb;
3031         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3032         int i, blocks = ocfs2_clusters_to_blocks(sb, new_len);
3033         u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster);
3034         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
3035         struct ocfs2_super *osb = OCFS2_SB(sb);
3036         struct buffer_head *old_bh = NULL;
3037         struct buffer_head *new_bh = NULL;
3038
3039         mlog(0, "old_cluster %u, new %u, len %u\n", old_cluster,
3040              new_cluster, new_len);
3041
3042         for (i = 0; i < blocks; i++, old_block++, new_block++) {
3043                 new_bh = sb_getblk(osb->sb, new_block);
3044                 if (new_bh == NULL) {
3045                         ret = -EIO;
3046                         mlog_errno(ret);
3047                         break;
3048                 }
3049
3050                 ocfs2_set_new_buffer_uptodate(ci, new_bh);
3051
3052                 ret = ocfs2_read_block(ci, old_block, &old_bh, NULL);
3053                 if (ret) {
3054                         mlog_errno(ret);
3055                         break;
3056                 }
3057
3058                 ret = ocfs2_journal_access(handle, ci, new_bh,
3059                                            OCFS2_JOURNAL_ACCESS_CREATE);
3060                 if (ret) {
3061                         mlog_errno(ret);
3062                         break;
3063                 }
3064
3065                 memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize);
3066                 ret = ocfs2_journal_dirty(handle, new_bh);
3067                 if (ret) {
3068                         mlog_errno(ret);
3069                         break;
3070                 }
3071
3072                 brelse(new_bh);
3073                 brelse(old_bh);
3074                 new_bh = NULL;
3075                 old_bh = NULL;
3076         }
3077
3078         brelse(new_bh);
3079         brelse(old_bh);
3080         return ret;
3081 }
3082
3083 static int ocfs2_clear_ext_refcount(handle_t *handle,
3084                                     struct ocfs2_extent_tree *et,
3085                                     u32 cpos, u32 p_cluster, u32 len,
3086                                     unsigned int ext_flags,
3087                                     struct ocfs2_alloc_context *meta_ac,
3088                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
3089 {
3090         int ret, index;
3091         struct ocfs2_extent_rec replace_rec;
3092         struct ocfs2_path *path = NULL;
3093         struct ocfs2_extent_list *el;
3094         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
3095         u64 ino = ocfs2_metadata_cache_owner(et->et_ci);
3096
3097         mlog(0, "inode %llu cpos %u, len %u, p_cluster %u, ext_flags %u\n",
3098              (unsigned long long)ino, cpos, len, p_cluster, ext_flags);
3099
3100         memset(&replace_rec, 0, sizeof(replace_rec));
3101         replace_rec.e_cpos = cpu_to_le32(cpos);
3102         replace_rec.e_leaf_clusters = cpu_to_le16(len);
3103         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb,
3104                                                                    p_cluster));
3105         replace_rec.e_flags = ext_flags;
3106         replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED;
3107
3108         path = ocfs2_new_path_from_et(et);
3109         if (!path) {
3110                 ret = -ENOMEM;
3111                 mlog_errno(ret);
3112                 goto out;
3113         }
3114
3115         ret = ocfs2_find_path(et->et_ci, path, cpos);
3116         if (ret) {
3117                 mlog_errno(ret);
3118                 goto out;
3119         }
3120
3121         el = path_leaf_el(path);
3122
3123         index = ocfs2_search_extent_list(el, cpos);
3124         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
3125                 ocfs2_error(sb,
3126                             "Inode %llu has an extent at cpos %u which can no "
3127                             "longer be found.\n",
3128                             (unsigned long long)ino, cpos);
3129                 ret = -EROFS;
3130                 goto out;
3131         }
3132
3133         ret = ocfs2_split_extent(handle, et, path, index,
3134                                  &replace_rec, meta_ac, dealloc);
3135         if (ret)
3136                 mlog_errno(ret);
3137
3138 out:
3139         ocfs2_free_path(path);
3140         return ret;
3141 }
3142
3143 static int ocfs2_replace_clusters(handle_t *handle,
3144                                   struct ocfs2_cow_context *context,
3145                                   u32 cpos, u32 old,
3146                                   u32 new, u32 len,
3147                                   unsigned int ext_flags)
3148 {
3149         int ret;
3150         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3151         u64 ino = ocfs2_metadata_cache_owner(ci);
3152
3153         mlog(0, "inode %llu, cpos %u, old %u, new %u, len %u, ext_flags %u\n",
3154              (unsigned long long)ino, cpos, old, new, len, ext_flags);
3155
3156         /*If the old clusters is unwritten, no need to duplicate. */
3157         if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) {
3158                 ret = context->cow_duplicate_clusters(handle, context, cpos,
3159                                                       old, new, len);
3160                 if (ret) {
3161                         mlog_errno(ret);
3162                         goto out;
3163                 }
3164         }
3165
3166         ret = ocfs2_clear_ext_refcount(handle, &context->data_et,
3167                                        cpos, new, len, ext_flags,
3168                                        context->meta_ac, &context->dealloc);
3169         if (ret)
3170                 mlog_errno(ret);
3171 out:
3172         return ret;
3173 }
3174
3175 static int ocfs2_cow_sync_writeback(struct super_block *sb,
3176                                     struct ocfs2_cow_context *context,
3177                                     u32 cpos, u32 num_clusters)
3178 {
3179         int ret = 0;
3180         loff_t offset, end, map_end;
3181         pgoff_t page_index;
3182         struct page *page;
3183
3184         if (ocfs2_should_order_data(context->inode))
3185                 return 0;
3186
3187         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
3188         end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits);
3189
3190         ret = filemap_fdatawrite_range(context->inode->i_mapping,
3191                                        offset, end - 1);
3192         if (ret < 0) {
3193                 mlog_errno(ret);
3194                 return ret;
3195         }
3196
3197         while (offset < end) {
3198                 page_index = offset >> PAGE_CACHE_SHIFT;
3199                 map_end = (page_index + 1) << PAGE_CACHE_SHIFT;
3200                 if (map_end > end)
3201                         map_end = end;
3202
3203                 page = grab_cache_page(context->inode->i_mapping, page_index);
3204                 BUG_ON(!page);
3205
3206                 wait_on_page_writeback(page);
3207                 if (PageError(page)) {
3208                         ret = -EIO;
3209                         mlog_errno(ret);
3210                 } else
3211                         mark_page_accessed(page);
3212
3213                 unlock_page(page);
3214                 page_cache_release(page);
3215                 page = NULL;
3216                 offset = map_end;
3217                 if (ret)
3218                         break;
3219         }
3220
3221         return ret;
3222 }
3223
3224 static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context,
3225                                  u32 v_cluster, u32 *p_cluster,
3226                                  u32 *num_clusters,
3227                                  unsigned int *extent_flags)
3228 {
3229         return ocfs2_get_clusters(context->inode, v_cluster, p_cluster,
3230                                   num_clusters, extent_flags);
3231 }
3232
3233 static int ocfs2_make_clusters_writable(struct super_block *sb,
3234                                         struct ocfs2_cow_context *context,
3235                                         u32 cpos, u32 p_cluster,
3236                                         u32 num_clusters, unsigned int e_flags)
3237 {
3238         int ret, delete, index, credits =  0;
3239         u32 new_bit, new_len, orig_num_clusters;
3240         unsigned int set_len;
3241         struct ocfs2_super *osb = OCFS2_SB(sb);
3242         handle_t *handle;
3243         struct buffer_head *ref_leaf_bh = NULL;
3244         struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci;
3245         struct ocfs2_refcount_rec rec;
3246
3247         mlog(0, "cpos %u, p_cluster %u, num_clusters %u, e_flags %u\n",
3248              cpos, p_cluster, num_clusters, e_flags);
3249
3250         ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters,
3251                                              &context->data_et,
3252                                              ref_ci,
3253                                              context->ref_root_bh,
3254                                              &context->meta_ac,
3255                                              &context->data_ac, &credits);
3256         if (ret) {
3257                 mlog_errno(ret);
3258                 return ret;
3259         }
3260
3261         if (context->post_refcount)
3262                 credits += context->post_refcount->credits;
3263
3264         credits += context->extra_credits;
3265         handle = ocfs2_start_trans(osb, credits);
3266         if (IS_ERR(handle)) {
3267                 ret = PTR_ERR(handle);
3268                 mlog_errno(ret);
3269                 goto out;
3270         }
3271
3272         orig_num_clusters = num_clusters;
3273
3274         while (num_clusters) {
3275                 ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh,
3276                                              p_cluster, num_clusters,
3277                                              &rec, &index, &ref_leaf_bh);
3278                 if (ret) {
3279                         mlog_errno(ret);
3280                         goto out_commit;
3281                 }
3282
3283                 BUG_ON(!rec.r_refcount);
3284                 set_len = min((u64)p_cluster + num_clusters,
3285                               le64_to_cpu(rec.r_cpos) +
3286                               le32_to_cpu(rec.r_clusters)) - p_cluster;
3287
3288                 /*
3289                  * There are many different situation here.
3290                  * 1. If refcount == 1, remove the flag and don't COW.
3291                  * 2. If refcount > 1, allocate clusters.
3292                  *    Here we may not allocate r_len once at a time, so continue
3293                  *    until we reach num_clusters.
3294                  */
3295                 if (le32_to_cpu(rec.r_refcount) == 1) {
3296                         delete = 0;
3297                         ret = ocfs2_clear_ext_refcount(handle,
3298                                                        &context->data_et,
3299                                                        cpos, p_cluster,
3300                                                        set_len, e_flags,
3301                                                        context->meta_ac,
3302                                                        &context->dealloc);
3303                         if (ret) {
3304                                 mlog_errno(ret);
3305                                 goto out_commit;
3306                         }
3307                 } else {
3308                         delete = 1;
3309
3310                         ret = __ocfs2_claim_clusters(osb, handle,
3311                                                      context->data_ac,
3312                                                      1, set_len,
3313                                                      &new_bit, &new_len);
3314                         if (ret) {
3315                                 mlog_errno(ret);
3316                                 goto out_commit;
3317                         }
3318
3319                         ret = ocfs2_replace_clusters(handle, context,
3320                                                      cpos, p_cluster, new_bit,
3321                                                      new_len, e_flags);
3322                         if (ret) {
3323                                 mlog_errno(ret);
3324                                 goto out_commit;
3325                         }
3326                         set_len = new_len;
3327                 }
3328
3329                 ret = __ocfs2_decrease_refcount(handle, ref_ci,
3330                                                 context->ref_root_bh,
3331                                                 p_cluster, set_len,
3332                                                 context->meta_ac,
3333                                                 &context->dealloc, delete);
3334                 if (ret) {
3335                         mlog_errno(ret);
3336                         goto out_commit;
3337                 }
3338
3339                 cpos += set_len;
3340                 p_cluster += set_len;
3341                 num_clusters -= set_len;
3342                 brelse(ref_leaf_bh);
3343                 ref_leaf_bh = NULL;
3344         }
3345
3346         /* handle any post_cow action. */
3347         if (context->post_refcount && context->post_refcount->func) {
3348                 ret = context->post_refcount->func(context->inode, handle,
3349                                                 context->post_refcount->para);
3350                 if (ret) {
3351                         mlog_errno(ret);
3352                         goto out_commit;
3353                 }
3354         }
3355
3356         /*
3357          * Here we should write the new page out first if we are
3358          * in write-back mode.
3359          */
3360         if (context->get_clusters == ocfs2_di_get_clusters) {
3361                 ret = ocfs2_cow_sync_writeback(sb, context, cpos,
3362                                                orig_num_clusters);
3363                 if (ret)
3364                         mlog_errno(ret);
3365         }
3366
3367 out_commit:
3368         ocfs2_commit_trans(osb, handle);
3369
3370 out:
3371         if (context->data_ac) {
3372                 ocfs2_free_alloc_context(context->data_ac);
3373                 context->data_ac = NULL;
3374         }
3375         if (context->meta_ac) {
3376                 ocfs2_free_alloc_context(context->meta_ac);
3377                 context->meta_ac = NULL;
3378         }
3379         brelse(ref_leaf_bh);
3380
3381         return ret;
3382 }
3383
3384 static int ocfs2_replace_cow(struct ocfs2_cow_context *context)
3385 {
3386         int ret = 0;
3387         struct inode *inode = context->inode;
3388         u32 cow_start = context->cow_start, cow_len = context->cow_len;
3389         u32 p_cluster, num_clusters;
3390         unsigned int ext_flags;
3391         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3392
3393         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
3394                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
3395                             "tree, but the feature bit is not set in the "
3396                             "super block.", inode->i_ino);
3397                 return -EROFS;
3398         }
3399
3400         ocfs2_init_dealloc_ctxt(&context->dealloc);
3401
3402         while (cow_len) {
3403                 ret = context->get_clusters(context, cow_start, &p_cluster,
3404                                             &num_clusters, &ext_flags);
3405                 if (ret) {
3406                         mlog_errno(ret);
3407                         break;
3408                 }
3409
3410                 BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED));
3411
3412                 if (cow_len < num_clusters)
3413                         num_clusters = cow_len;
3414
3415                 ret = ocfs2_make_clusters_writable(inode->i_sb, context,
3416                                                    cow_start, p_cluster,
3417                                                    num_clusters, ext_flags);
3418                 if (ret) {
3419                         mlog_errno(ret);
3420                         break;
3421                 }
3422
3423                 cow_len -= num_clusters;
3424                 cow_start += num_clusters;
3425         }
3426
3427         if (ocfs2_dealloc_has_cluster(&context->dealloc)) {
3428                 ocfs2_schedule_truncate_log_flush(osb, 1);
3429                 ocfs2_run_deallocs(osb, &context->dealloc);
3430         }
3431
3432         return ret;
3433 }
3434
3435 /*
3436  * Starting at cpos, try to CoW write_len clusters.  Don't CoW
3437  * past max_cpos.  This will stop when it runs into a hole or an
3438  * unrefcounted extent.
3439  */
3440 static int ocfs2_refcount_cow_hunk(struct inode *inode,
3441                                    struct buffer_head *di_bh,
3442                                    u32 cpos, u32 write_len, u32 max_cpos)
3443 {
3444         int ret;
3445         u32 cow_start = 0, cow_len = 0;
3446         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3447         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3448         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3449         struct buffer_head *ref_root_bh = NULL;
3450         struct ocfs2_refcount_tree *ref_tree;
3451         struct ocfs2_cow_context *context = NULL;
3452
3453         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3454
3455         ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list,
3456                                               cpos, write_len, max_cpos,
3457                                               &cow_start, &cow_len);
3458         if (ret) {
3459                 mlog_errno(ret);
3460                 goto out;
3461         }
3462
3463         mlog(0, "CoW inode %lu, cpos %u, write_len %u, cow_start %u, "
3464              "cow_len %u\n", inode->i_ino,
3465              cpos, write_len, cow_start, cow_len);
3466
3467         BUG_ON(cow_len == 0);
3468
3469         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3470         if (!context) {
3471                 ret = -ENOMEM;
3472                 mlog_errno(ret);
3473                 goto out;
3474         }
3475
3476         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
3477                                        1, &ref_tree, &ref_root_bh);
3478         if (ret) {
3479                 mlog_errno(ret);
3480                 goto out;
3481         }
3482
3483         context->inode = inode;
3484         context->cow_start = cow_start;
3485         context->cow_len = cow_len;
3486         context->ref_tree = ref_tree;
3487         context->ref_root_bh = ref_root_bh;
3488         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
3489         context->get_clusters = ocfs2_di_get_clusters;
3490
3491         ocfs2_init_dinode_extent_tree(&context->data_et,
3492                                       INODE_CACHE(inode), di_bh);
3493
3494         ret = ocfs2_replace_cow(context);
3495         if (ret)
3496                 mlog_errno(ret);
3497
3498         /*
3499          * truncate the extent map here since no matter whether we meet with
3500          * any error during the action, we shouldn't trust cached extent map
3501          * any more.
3502          */
3503         ocfs2_extent_map_trunc(inode, cow_start);
3504
3505         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3506         brelse(ref_root_bh);
3507 out:
3508         kfree(context);
3509         return ret;
3510 }
3511
3512 /*
3513  * CoW any and all clusters between cpos and cpos+write_len.
3514  * Don't CoW past max_cpos.  If this returns successfully, all
3515  * clusters between cpos and cpos+write_len are safe to modify.
3516  */
3517 int ocfs2_refcount_cow(struct inode *inode,
3518                        struct buffer_head *di_bh,
3519                        u32 cpos, u32 write_len, u32 max_cpos)
3520 {
3521         int ret = 0;
3522         u32 p_cluster, num_clusters;
3523         unsigned int ext_flags;
3524
3525         while (write_len) {
3526                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3527                                          &num_clusters, &ext_flags);
3528                 if (ret) {
3529                         mlog_errno(ret);
3530                         break;
3531                 }
3532
3533                 if (write_len < num_clusters)
3534                         num_clusters = write_len;
3535
3536                 if (ext_flags & OCFS2_EXT_REFCOUNTED) {
3537                         ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos,
3538                                                       num_clusters, max_cpos);
3539                         if (ret) {
3540                                 mlog_errno(ret);
3541                                 break;
3542                         }
3543                 }
3544
3545                 write_len -= num_clusters;
3546                 cpos += num_clusters;
3547         }
3548
3549         return ret;
3550 }
3551
3552 static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context,
3553                                           u32 v_cluster, u32 *p_cluster,
3554                                           u32 *num_clusters,
3555                                           unsigned int *extent_flags)
3556 {
3557         struct inode *inode = context->inode;
3558         struct ocfs2_xattr_value_root *xv = context->cow_object;
3559
3560         return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster,
3561                                         num_clusters, &xv->xr_list,
3562                                         extent_flags);
3563 }
3564
3565 /*
3566  * Given a xattr value root, calculate the most meta/credits we need for
3567  * refcount tree change if we truncate it to 0.
3568  */
3569 int ocfs2_refcounted_xattr_delete_need(struct inode *inode,
3570                                        struct ocfs2_caching_info *ref_ci,
3571                                        struct buffer_head *ref_root_bh,
3572                                        struct ocfs2_xattr_value_root *xv,
3573                                        int *meta_add, int *credits)
3574 {
3575         int ret = 0, index, ref_blocks = 0;
3576         u32 p_cluster, num_clusters;
3577         u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters);
3578         struct ocfs2_refcount_block *rb;
3579         struct ocfs2_refcount_rec rec;
3580         struct buffer_head *ref_leaf_bh = NULL;
3581
3582         while (cpos < clusters) {
3583                 ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster,
3584                                                &num_clusters, &xv->xr_list,
3585                                                NULL);
3586                 if (ret) {
3587                         mlog_errno(ret);
3588                         goto out;
3589                 }
3590
3591                 cpos += num_clusters;
3592
3593                 while (num_clusters) {
3594                         ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh,
3595                                                      p_cluster, num_clusters,
3596                                                      &rec, &index,
3597                                                      &ref_leaf_bh);
3598                         if (ret) {
3599                                 mlog_errno(ret);
3600                                 goto out;
3601                         }
3602
3603                         BUG_ON(!rec.r_refcount);
3604
3605                         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
3606
3607                         /*
3608                          * We really don't know whether the other clusters is in
3609                          * this refcount block or not, so just take the worst
3610                          * case that all the clusters are in this block and each
3611                          * one will split a refcount rec, so totally we need
3612                          * clusters * 2 new refcount rec.
3613                          */
3614                         if (le64_to_cpu(rb->rf_records.rl_used) + clusters * 2 >
3615                             le16_to_cpu(rb->rf_records.rl_count))
3616                                 ref_blocks++;
3617
3618                         *credits += 1;
3619                         brelse(ref_leaf_bh);
3620                         ref_leaf_bh = NULL;
3621
3622                         if (num_clusters <= le32_to_cpu(rec.r_clusters))
3623                                 break;
3624                         else
3625                                 num_clusters -= le32_to_cpu(rec.r_clusters);
3626                         p_cluster += num_clusters;
3627                 }
3628         }
3629
3630         *meta_add += ref_blocks;
3631         if (!ref_blocks)
3632                 goto out;
3633
3634         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
3635         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)
3636                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
3637         else {
3638                 struct ocfs2_extent_tree et;
3639
3640                 ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh);
3641                 *credits += ocfs2_calc_extend_credits(inode->i_sb,
3642                                                       et.et_root_el,
3643                                                       ref_blocks);
3644         }
3645
3646 out:
3647         brelse(ref_leaf_bh);
3648         return ret;
3649 }
3650
3651 /*
3652  * Do CoW for xattr.
3653  */
3654 int ocfs2_refcount_cow_xattr(struct inode *inode,
3655                              struct ocfs2_dinode *di,
3656                              struct ocfs2_xattr_value_buf *vb,
3657                              struct ocfs2_refcount_tree *ref_tree,
3658                              struct buffer_head *ref_root_bh,
3659                              u32 cpos, u32 write_len,
3660                              struct ocfs2_post_refcount *post)
3661 {
3662         int ret;
3663         struct ocfs2_xattr_value_root *xv = vb->vb_xv;
3664         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3665         struct ocfs2_cow_context *context = NULL;
3666         u32 cow_start, cow_len;
3667
3668         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3669
3670         ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list,
3671                                               cpos, write_len, UINT_MAX,
3672                                               &cow_start, &cow_len);
3673         if (ret) {
3674                 mlog_errno(ret);
3675                 goto out;
3676         }
3677
3678         BUG_ON(cow_len == 0);
3679
3680         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3681         if (!context) {
3682                 ret = -ENOMEM;
3683                 mlog_errno(ret);
3684                 goto out;
3685         }
3686
3687         context->inode = inode;
3688         context->cow_start = cow_start;
3689         context->cow_len = cow_len;
3690         context->ref_tree = ref_tree;
3691         context->ref_root_bh = ref_root_bh;;
3692         context->cow_object = xv;
3693
3694         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd;
3695         /* We need the extra credits for duplicate_clusters by jbd. */
3696         context->extra_credits =
3697                 ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len;
3698         context->get_clusters = ocfs2_xattr_value_get_clusters;
3699         context->post_refcount = post;
3700
3701         ocfs2_init_xattr_value_extent_tree(&context->data_et,
3702                                            INODE_CACHE(inode), vb);
3703
3704         ret = ocfs2_replace_cow(context);
3705         if (ret)
3706                 mlog_errno(ret);
3707
3708 out:
3709         kfree(context);
3710         return ret;
3711 }
3712
3713 /*
3714  * Insert a new extent into refcount tree and mark a extent rec
3715  * as refcounted in the dinode tree.
3716  */
3717 int ocfs2_add_refcount_flag(struct inode *inode,
3718                             struct ocfs2_extent_tree *data_et,
3719                             struct ocfs2_caching_info *ref_ci,
3720                             struct buffer_head *ref_root_bh,
3721                             u32 cpos, u32 p_cluster, u32 num_clusters,
3722                             struct ocfs2_cached_dealloc_ctxt *dealloc,
3723                             struct ocfs2_post_refcount *post)
3724 {
3725         int ret;
3726         handle_t *handle;
3727         int credits = 1, ref_blocks = 0;
3728         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3729         struct ocfs2_alloc_context *meta_ac = NULL;
3730
3731         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
3732                                                ref_ci, ref_root_bh,
3733                                                p_cluster, num_clusters,
3734                                                &ref_blocks, &credits);
3735         if (ret) {
3736                 mlog_errno(ret);
3737                 goto out;
3738         }
3739
3740         mlog(0, "reserve new metadata %d, credits = %d\n",
3741              ref_blocks, credits);
3742
3743         if (ref_blocks) {
3744                 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
3745                                                         ref_blocks, &meta_ac);
3746                 if (ret) {
3747                         mlog_errno(ret);
3748                         goto out;
3749                 }
3750         }
3751
3752         if (post)
3753                 credits += post->credits;
3754
3755         handle = ocfs2_start_trans(osb, credits);
3756         if (IS_ERR(handle)) {
3757                 ret = PTR_ERR(handle);
3758                 mlog_errno(ret);
3759                 goto out;
3760         }
3761
3762         ret = ocfs2_mark_extent_refcounted(inode, data_et, handle,
3763                                            cpos, num_clusters, p_cluster,
3764                                            meta_ac, dealloc);
3765         if (ret) {
3766                 mlog_errno(ret);
3767                 goto out_commit;
3768         }
3769
3770         ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3771                                         p_cluster, num_clusters, 0,
3772                                         meta_ac, dealloc);
3773         if (ret) {
3774                 mlog_errno(ret);
3775                 goto out_commit;
3776         }
3777
3778         if (post && post->func) {
3779                 ret = post->func(inode, handle, post->para);
3780                 if (ret)
3781                         mlog_errno(ret);
3782         }
3783
3784 out_commit:
3785         ocfs2_commit_trans(osb, handle);
3786 out:
3787         if (meta_ac)
3788                 ocfs2_free_alloc_context(meta_ac);
3789         return ret;
3790 }
3791
3792 static int ocfs2_change_ctime(struct inode *inode,
3793                               struct buffer_head *di_bh)
3794 {
3795         int ret;
3796         handle_t *handle;
3797         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3798
3799         handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
3800                                    OCFS2_INODE_UPDATE_CREDITS);
3801         if (IS_ERR(handle)) {
3802                 ret = PTR_ERR(handle);
3803                 mlog_errno(ret);
3804                 goto out;
3805         }
3806
3807         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
3808                                       OCFS2_JOURNAL_ACCESS_WRITE);
3809         if (ret) {
3810                 mlog_errno(ret);
3811                 goto out_commit;
3812         }
3813
3814         inode->i_ctime = CURRENT_TIME;
3815         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
3816         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
3817
3818         ocfs2_journal_dirty(handle, di_bh);
3819
3820 out_commit:
3821         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
3822 out:
3823         return ret;
3824 }
3825
3826 static int ocfs2_attach_refcount_tree(struct inode *inode,
3827                                       struct buffer_head *di_bh)
3828 {
3829         int ret, data_changed = 0;
3830         struct buffer_head *ref_root_bh = NULL;
3831         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3832         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3833         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3834         struct ocfs2_refcount_tree *ref_tree;
3835         unsigned int ext_flags;
3836         loff_t size;
3837         u32 cpos, num_clusters, clusters, p_cluster;
3838         struct ocfs2_cached_dealloc_ctxt dealloc;
3839         struct ocfs2_extent_tree di_et;
3840
3841         ocfs2_init_dealloc_ctxt(&dealloc);
3842
3843         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) {
3844                 ret = ocfs2_create_refcount_tree(inode, di_bh);
3845                 if (ret) {
3846                         mlog_errno(ret);
3847                         goto out;
3848                 }
3849         }
3850
3851         BUG_ON(!di->i_refcount_loc);
3852         ret = ocfs2_lock_refcount_tree(osb,
3853                                        le64_to_cpu(di->i_refcount_loc), 1,
3854                                        &ref_tree, &ref_root_bh);
3855         if (ret) {
3856                 mlog_errno(ret);
3857                 goto out;
3858         }
3859
3860         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
3861                 goto attach_xattr;
3862
3863         ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh);
3864
3865         size = i_size_read(inode);
3866         clusters = ocfs2_clusters_for_bytes(inode->i_sb, size);
3867
3868         cpos = 0;
3869         while (cpos < clusters) {
3870                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3871                                          &num_clusters, &ext_flags);
3872
3873                 if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) {
3874                         ret = ocfs2_add_refcount_flag(inode, &di_et,
3875                                                       &ref_tree->rf_ci,
3876                                                       ref_root_bh, cpos,
3877                                                       p_cluster, num_clusters,
3878                                                       &dealloc, NULL);
3879                         if (ret) {
3880                                 mlog_errno(ret);
3881                                 goto unlock;
3882                         }
3883
3884                         data_changed = 1;
3885                 }
3886                 cpos += num_clusters;
3887         }
3888
3889 attach_xattr:
3890         if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
3891                 ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh,
3892                                                        &ref_tree->rf_ci,
3893                                                        ref_root_bh,
3894                                                        &dealloc);
3895                 if (ret) {
3896                         mlog_errno(ret);
3897                         goto unlock;
3898                 }
3899         }
3900
3901         if (data_changed) {
3902                 ret = ocfs2_change_ctime(inode, di_bh);
3903                 if (ret)
3904                         mlog_errno(ret);
3905         }
3906
3907 unlock:
3908         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3909         brelse(ref_root_bh);
3910
3911         if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) {
3912                 ocfs2_schedule_truncate_log_flush(osb, 1);
3913                 ocfs2_run_deallocs(osb, &dealloc);
3914         }
3915 out:
3916         /*
3917          * Empty the extent map so that we may get the right extent
3918          * record from the disk.
3919          */
3920         ocfs2_extent_map_trunc(inode, 0);
3921
3922         return ret;
3923 }
3924
3925 static int ocfs2_add_refcounted_extent(struct inode *inode,
3926                                    struct ocfs2_extent_tree *et,
3927                                    struct ocfs2_caching_info *ref_ci,
3928                                    struct buffer_head *ref_root_bh,
3929                                    u32 cpos, u32 p_cluster, u32 num_clusters,
3930                                    unsigned int ext_flags,
3931                                    struct ocfs2_cached_dealloc_ctxt *dealloc)
3932 {
3933         int ret;
3934         handle_t *handle;
3935         int credits = 0;
3936         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3937         struct ocfs2_alloc_context *meta_ac = NULL;
3938
3939         ret = ocfs2_lock_refcount_allocators(inode->i_sb,
3940                                              p_cluster, num_clusters,
3941                                              et, ref_ci,
3942                                              ref_root_bh, &meta_ac,
3943                                              NULL, &credits);
3944         if (ret) {
3945                 mlog_errno(ret);
3946                 goto out;
3947         }
3948
3949         handle = ocfs2_start_trans(osb, credits);
3950         if (IS_ERR(handle)) {
3951                 ret = PTR_ERR(handle);
3952                 mlog_errno(ret);
3953                 goto out;
3954         }
3955
3956         ret = ocfs2_insert_extent(handle, et, cpos,
3957                         cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
3958                                                              p_cluster)),
3959                         num_clusters, ext_flags, meta_ac);
3960         if (ret) {
3961                 mlog_errno(ret);
3962                 goto out_commit;
3963         }
3964
3965         ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3966                                       p_cluster, num_clusters,
3967                                       meta_ac, dealloc);
3968         if (ret)
3969                 mlog_errno(ret);
3970
3971 out_commit:
3972         ocfs2_commit_trans(osb, handle);
3973 out:
3974         if (meta_ac)
3975                 ocfs2_free_alloc_context(meta_ac);
3976         return ret;
3977 }
3978
3979 static int ocfs2_duplicate_inline_data(struct inode *s_inode,
3980                                        struct buffer_head *s_bh,
3981                                        struct inode *t_inode,
3982                                        struct buffer_head *t_bh)
3983 {
3984         int ret;
3985         handle_t *handle;
3986         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
3987         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
3988         struct ocfs2_dinode *t_di = (struct ocfs2_dinode *)t_bh->b_data;
3989
3990         BUG_ON(!(OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
3991
3992         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
3993         if (IS_ERR(handle)) {
3994                 ret = PTR_ERR(handle);
3995                 mlog_errno(ret);
3996                 goto out;
3997         }
3998
3999         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
4000                                       OCFS2_JOURNAL_ACCESS_WRITE);
4001         if (ret) {
4002                 mlog_errno(ret);
4003                 goto out_commit;
4004         }
4005
4006         t_di->id2.i_data.id_count = s_di->id2.i_data.id_count;
4007         memcpy(t_di->id2.i_data.id_data, s_di->id2.i_data.id_data,
4008                le16_to_cpu(s_di->id2.i_data.id_count));
4009         spin_lock(&OCFS2_I(t_inode)->ip_lock);
4010         OCFS2_I(t_inode)->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
4011         t_di->i_dyn_features = cpu_to_le16(OCFS2_I(t_inode)->ip_dyn_features);
4012         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
4013
4014         ocfs2_journal_dirty(handle, t_bh);
4015
4016 out_commit:
4017         ocfs2_commit_trans(osb, handle);
4018 out:
4019         return ret;
4020 }
4021
4022 static int ocfs2_duplicate_extent_list(struct inode *s_inode,
4023                                 struct inode *t_inode,
4024                                 struct buffer_head *t_bh,
4025                                 struct ocfs2_caching_info *ref_ci,
4026                                 struct buffer_head *ref_root_bh,
4027                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
4028 {
4029         int ret = 0;
4030         u32 p_cluster, num_clusters, clusters, cpos;
4031         loff_t size;
4032         unsigned int ext_flags;
4033         struct ocfs2_extent_tree et;
4034
4035         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh);
4036
4037         size = i_size_read(s_inode);
4038         clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size);
4039
4040         cpos = 0;
4041         while (cpos < clusters) {
4042                 ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster,
4043                                          &num_clusters, &ext_flags);
4044
4045                 if (p_cluster) {
4046                         ret = ocfs2_add_refcounted_extent(t_inode, &et,
4047                                                           ref_ci, ref_root_bh,
4048                                                           cpos, p_cluster,
4049                                                           num_clusters,
4050                                                           ext_flags,
4051                                                           dealloc);
4052                         if (ret) {
4053                                 mlog_errno(ret);
4054                                 goto out;
4055                         }
4056                 }
4057
4058                 cpos += num_clusters;
4059         }
4060
4061 out:
4062         return ret;
4063 }
4064
4065 /*
4066  * change the new file's attributes to the src.
4067  *
4068  * reflink creates a snapshot of a file, that means the attributes
4069  * must be identical except for three exceptions - nlink, ino, and ctime.
4070  */
4071 static int ocfs2_complete_reflink(struct inode *s_inode,
4072                                   struct buffer_head *s_bh,
4073                                   struct inode *t_inode,
4074                                   struct buffer_head *t_bh,
4075                                   bool preserve)
4076 {
4077         int ret;
4078         handle_t *handle;
4079         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
4080         struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data;
4081         loff_t size = i_size_read(s_inode);
4082
4083         handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb),
4084                                    OCFS2_INODE_UPDATE_CREDITS);
4085         if (IS_ERR(handle)) {
4086                 ret = PTR_ERR(handle);
4087                 mlog_errno(ret);
4088                 return ret;
4089         }
4090
4091         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
4092                                       OCFS2_JOURNAL_ACCESS_WRITE);
4093         if (ret) {
4094                 mlog_errno(ret);
4095                 goto out_commit;
4096         }
4097
4098         spin_lock(&OCFS2_I(t_inode)->ip_lock);
4099         OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters;
4100         OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr;
4101         OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features;
4102         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
4103         i_size_write(t_inode, size);
4104
4105         di->i_xattr_inline_size = s_di->i_xattr_inline_size;
4106         di->i_clusters = s_di->i_clusters;
4107         di->i_size = s_di->i_size;
4108         di->i_dyn_features = s_di->i_dyn_features;
4109         di->i_attr = s_di->i_attr;
4110
4111         if (preserve) {
4112                 t_inode->i_uid = s_inode->i_uid;
4113                 t_inode->i_gid = s_inode->i_gid;
4114                 t_inode->i_mode = s_inode->i_mode;
4115                 di->i_uid = s_di->i_uid;
4116                 di->i_gid = s_di->i_gid;
4117                 di->i_mode = s_di->i_mode;
4118
4119                 /*
4120                  * update time.
4121                  * we want mtime to appear identical to the source and
4122                  * update ctime.
4123                  */
4124                 t_inode->i_ctime = CURRENT_TIME;
4125
4126                 di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec);
4127                 di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec);
4128
4129                 t_inode->i_mtime = s_inode->i_mtime;
4130                 di->i_mtime = s_di->i_mtime;
4131                 di->i_mtime_nsec = s_di->i_mtime_nsec;
4132         }
4133
4134         ocfs2_journal_dirty(handle, t_bh);
4135
4136 out_commit:
4137         ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle);
4138         return ret;
4139 }
4140
4141 static int ocfs2_create_reflink_node(struct inode *s_inode,
4142                                      struct buffer_head *s_bh,
4143                                      struct inode *t_inode,
4144                                      struct buffer_head *t_bh,
4145                                      bool preserve)
4146 {
4147         int ret;
4148         struct buffer_head *ref_root_bh = NULL;
4149         struct ocfs2_cached_dealloc_ctxt dealloc;
4150         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
4151         struct ocfs2_refcount_block *rb;
4152         struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data;
4153         struct ocfs2_refcount_tree *ref_tree;
4154
4155         ocfs2_init_dealloc_ctxt(&dealloc);
4156
4157         ret = ocfs2_set_refcount_tree(t_inode, t_bh,
4158                                       le64_to_cpu(di->i_refcount_loc));
4159         if (ret) {
4160                 mlog_errno(ret);
4161                 goto out;
4162         }
4163
4164         if (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4165                 ret = ocfs2_duplicate_inline_data(s_inode, s_bh,
4166                                                   t_inode, t_bh);
4167                 if (ret)
4168                         mlog_errno(ret);
4169                 goto out;
4170         }
4171
4172         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
4173                                        1, &ref_tree, &ref_root_bh);
4174         if (ret) {
4175                 mlog_errno(ret);
4176                 goto out;
4177         }
4178         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
4179
4180         ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh,
4181                                           &ref_tree->rf_ci, ref_root_bh,
4182                                           &dealloc);
4183         if (ret) {
4184                 mlog_errno(ret);
4185                 goto out_unlock_refcount;
4186         }
4187
4188 out_unlock_refcount:
4189         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4190         brelse(ref_root_bh);
4191 out:
4192         if (ocfs2_dealloc_has_cluster(&dealloc)) {
4193                 ocfs2_schedule_truncate_log_flush(osb, 1);
4194                 ocfs2_run_deallocs(osb, &dealloc);
4195         }
4196
4197         return ret;
4198 }
4199
4200 static int __ocfs2_reflink(struct dentry *old_dentry,
4201                            struct buffer_head *old_bh,
4202                            struct inode *new_inode,
4203                            bool preserve)
4204 {
4205         int ret;
4206         struct inode *inode = old_dentry->d_inode;
4207         struct buffer_head *new_bh = NULL;
4208
4209         ret = filemap_fdatawrite(inode->i_mapping);
4210         if (ret) {
4211                 mlog_errno(ret);
4212                 goto out;
4213         }
4214
4215         ret = ocfs2_attach_refcount_tree(inode, old_bh);
4216         if (ret) {
4217                 mlog_errno(ret);
4218                 goto out;
4219         }
4220
4221         mutex_lock(&new_inode->i_mutex);
4222         ret = ocfs2_inode_lock(new_inode, &new_bh, 1);
4223         if (ret) {
4224                 mlog_errno(ret);
4225                 goto out_unlock;
4226         }
4227
4228         ret = ocfs2_create_reflink_node(inode, old_bh,
4229                                         new_inode, new_bh, preserve);
4230         if (ret) {
4231                 mlog_errno(ret);
4232                 goto inode_unlock;
4233         }
4234
4235         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
4236                 ret = ocfs2_reflink_xattrs(inode, old_bh,
4237                                            new_inode, new_bh,
4238                                            preserve);
4239                 if (ret) {
4240                         mlog_errno(ret);
4241                         goto inode_unlock;
4242                 }
4243         }
4244
4245         ret = ocfs2_complete_reflink(inode, old_bh,
4246                                      new_inode, new_bh, preserve);
4247         if (ret)
4248                 mlog_errno(ret);
4249
4250 inode_unlock:
4251         ocfs2_inode_unlock(new_inode, 1);
4252         brelse(new_bh);
4253 out_unlock:
4254         mutex_unlock(&new_inode->i_mutex);
4255 out:
4256         if (!ret) {
4257                 ret = filemap_fdatawait(inode->i_mapping);
4258                 if (ret)
4259                         mlog_errno(ret);
4260         }
4261         return ret;
4262 }
4263
4264 static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir,
4265                          struct dentry *new_dentry, bool preserve)
4266 {
4267         int error;
4268         struct inode *inode = old_dentry->d_inode;
4269         struct buffer_head *old_bh = NULL;
4270         struct inode *new_orphan_inode = NULL;
4271
4272         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4273                 return -EOPNOTSUPP;
4274
4275         error = ocfs2_create_inode_in_orphan(dir, inode->i_mode,
4276                                              &new_orphan_inode);
4277         if (error) {
4278                 mlog_errno(error);
4279                 goto out;
4280         }
4281
4282         error = ocfs2_inode_lock(inode, &old_bh, 1);
4283         if (error) {
4284                 mlog_errno(error);
4285                 goto out;
4286         }
4287
4288         down_write(&OCFS2_I(inode)->ip_xattr_sem);
4289         down_write(&OCFS2_I(inode)->ip_alloc_sem);
4290         error = __ocfs2_reflink(old_dentry, old_bh,
4291                                 new_orphan_inode, preserve);
4292         up_write(&OCFS2_I(inode)->ip_alloc_sem);
4293         up_write(&OCFS2_I(inode)->ip_xattr_sem);
4294
4295         ocfs2_inode_unlock(inode, 1);
4296         brelse(old_bh);
4297
4298         if (error) {
4299                 mlog_errno(error);
4300                 goto out;
4301         }
4302
4303         /* If the security isn't preserved, we need to re-initialize them. */
4304         if (!preserve) {
4305                 error = ocfs2_init_security_and_acl(dir, new_orphan_inode);
4306                 if (error)
4307                         mlog_errno(error);
4308         }
4309 out:
4310         if (!error) {
4311                 error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode,
4312                                                        new_dentry);
4313                 if (error)
4314                         mlog_errno(error);
4315         }
4316
4317         if (new_orphan_inode) {
4318                 /*
4319                  * We need to open_unlock the inode no matter whether we
4320                  * succeed or not, so that other nodes can delete it later.
4321                  */
4322                 ocfs2_open_unlock(new_orphan_inode);
4323                 if (error)
4324                         iput(new_orphan_inode);
4325         }
4326
4327         return error;
4328 }
4329
4330 /*
4331  * Below here are the bits used by OCFS2_IOC_REFLINK() to fake
4332  * sys_reflink().  This will go away when vfs_reflink() exists in
4333  * fs/namei.c.
4334  */
4335
4336 /* copied from may_create in VFS. */
4337 static inline int ocfs2_may_create(struct inode *dir, struct dentry *child)
4338 {
4339         if (child->d_inode)
4340                 return -EEXIST;
4341         if (IS_DEADDIR(dir))
4342                 return -ENOENT;
4343         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
4344 }
4345
4346 /* copied from user_path_parent. */
4347 static int ocfs2_user_path_parent(const char __user *path,
4348                                   struct nameidata *nd, char **name)
4349 {
4350         char *s = getname(path);
4351         int error;
4352
4353         if (IS_ERR(s))
4354                 return PTR_ERR(s);
4355
4356         error = path_lookup(s, LOOKUP_PARENT, nd);
4357         if (error)
4358                 putname(s);
4359         else
4360                 *name = s;
4361
4362         return error;
4363 }
4364
4365 /**
4366  * ocfs2_vfs_reflink - Create a reference-counted link
4367  *
4368  * @old_dentry:        source dentry + inode
4369  * @dir:       directory to create the target
4370  * @new_dentry:        target dentry
4371  * @preserve:  if true, preserve all file attributes
4372  */
4373 int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir,
4374                       struct dentry *new_dentry, bool preserve)
4375 {
4376         struct inode *inode = old_dentry->d_inode;
4377         int error;
4378
4379         if (!inode)
4380                 return -ENOENT;
4381
4382         error = ocfs2_may_create(dir, new_dentry);
4383         if (error)
4384                 return error;
4385
4386         if (dir->i_sb != inode->i_sb)
4387                 return -EXDEV;
4388
4389         /*
4390          * A reflink to an append-only or immutable file cannot be created.
4391          */
4392         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4393                 return -EPERM;
4394
4395         /* Only regular files can be reflinked. */
4396         if (!S_ISREG(inode->i_mode))
4397                 return -EPERM;
4398
4399         /*
4400          * If the caller wants to preserve ownership, they require the
4401          * rights to do so.
4402          */
4403         if (preserve) {
4404                 if ((current_fsuid() != inode->i_uid) && !capable(CAP_CHOWN))
4405                         return -EPERM;
4406                 if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN))
4407                         return -EPERM;
4408         }
4409
4410         /*
4411          * If the caller is modifying any aspect of the attributes, they
4412          * are not creating a snapshot.  They need read permission on the
4413          * file.
4414          */
4415         if (!preserve) {
4416                 error = inode_permission(inode, MAY_READ);
4417                 if (error)
4418                         return error;
4419         }
4420
4421         mutex_lock(&inode->i_mutex);
4422         vfs_dq_init(dir);
4423         error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve);
4424         mutex_unlock(&inode->i_mutex);
4425         if (!error)
4426                 fsnotify_create(dir, new_dentry);
4427         return error;
4428 }
4429 /*
4430  * Most codes are copied from sys_linkat.
4431  */
4432 int ocfs2_reflink_ioctl(struct inode *inode,
4433                         const char __user *oldname,
4434                         const char __user *newname,
4435                         bool preserve)
4436 {
4437         struct dentry *new_dentry;
4438         struct nameidata nd;
4439         struct path old_path;
4440         int error;
4441         char *to = NULL;
4442
4443         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4444                 return -EOPNOTSUPP;
4445
4446         error = user_path_at(AT_FDCWD, oldname, 0, &old_path);
4447         if (error) {
4448                 mlog_errno(error);
4449                 return error;
4450         }
4451
4452         error = ocfs2_user_path_parent(newname, &nd, &to);
4453         if (error) {
4454                 mlog_errno(error);
4455                 goto out;
4456         }
4457
4458         error = -EXDEV;
4459         if (old_path.mnt != nd.path.mnt)
4460                 goto out_release;
4461         new_dentry = lookup_create(&nd, 0);
4462         error = PTR_ERR(new_dentry);
4463         if (IS_ERR(new_dentry)) {
4464                 mlog_errno(error);
4465                 goto out_unlock;
4466         }
4467
4468         error = mnt_want_write(nd.path.mnt);
4469         if (error) {
4470                 mlog_errno(error);
4471                 goto out_dput;
4472         }
4473
4474         error = ocfs2_vfs_reflink(old_path.dentry,
4475                                   nd.path.dentry->d_inode,
4476                                   new_dentry, preserve);
4477         mnt_drop_write(nd.path.mnt);
4478 out_dput:
4479         dput(new_dentry);
4480 out_unlock:
4481         mutex_unlock(&nd.path.dentry->d_inode->i_mutex);
4482 out_release:
4483         path_put(&nd.path);
4484         putname(to);
4485 out:
4486         path_put(&old_path);
4487
4488         return error;
4489 }