fs/ocfs2/refcounttree.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * refcounttree.c
   5  *
   6  * Copyright (C) 2009 Oracle.  All rights reserved.
   7  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public
  10  * License version 2 as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License for more details.
  16  */
  17
  18 #include <linux/sort.h>
  19 #define MLOG_MASK_PREFIX ML_REFCOUNT
  20 #include <cluster/masklog.h>
  21 #include "ocfs2.h"
  22 #include "inode.h"
  23 #include "alloc.h"
  24 #include "suballoc.h"
  25 #include "journal.h"
  26 #include "uptodate.h"
  27 #include "super.h"
  28 #include "buffer_head_io.h"
  29 #include "blockcheck.h"
  30 #include "refcounttree.h"
  31 #include "sysfile.h"
  32 #include "dlmglue.h"
  33 #include "extent_map.h"
  34 #include "aops.h"
  35 #include "xattr.h"
  36 #include "namei.h"
  37
  38 #include <linux/bio.h>
  39 #include <linux/blkdev.h>
  40 #include <linux/gfp.h>
  41 #include <linux/slab.h>
  42 #include <linux/writeback.h>
  43 #include <linux/pagevec.h>
  44 #include <linux/swap.h>
  45 #include <linux/security.h>
  46 #include <linux/fsnotify.h>
  47 #include <linux/quotaops.h>
  48 #include <linux/namei.h>
  49 #include <linux/mount.h>
  50
  51 struct ocfs2_cow_context {
  52         struct inode *inode;
  53         u32 cow_start;
  54         u32 cow_len;
  55         struct ocfs2_extent_tree data_et;
  56         struct ocfs2_refcount_tree *ref_tree;
  57         struct buffer_head *ref_root_bh;
  58         struct ocfs2_alloc_context *meta_ac;
  59         struct ocfs2_alloc_context *data_ac;
  60         struct ocfs2_cached_dealloc_ctxt dealloc;
  61         void *cow_object;
  62         struct ocfs2_post_refcount *post_refcount;
  63         int extra_credits;
  64         int (*get_clusters)(struct ocfs2_cow_context *context,
  65                             u32 v_cluster, u32 *p_cluster,
  66                             u32 *num_clusters,
  67                             unsigned int *extent_flags);
  68         int (*cow_duplicate_clusters)(handle_t *handle,
  69                                       struct ocfs2_cow_context *context,
  70                                       u32 cpos, u32 old_cluster,
  71                                       u32 new_cluster, u32 new_len);
  72 };
  73
  74 static inline struct ocfs2_refcount_tree *
  75 cache_info_to_refcount(struct ocfs2_caching_info *ci)
  76 {
  77         return container_of(ci, struct ocfs2_refcount_tree, rf_ci);
  78 }
  79
  80 static int ocfs2_validate_refcount_block(struct super_block *sb,
  81                                          struct buffer_head *bh)
  82 {
  83         int rc;
  84         struct ocfs2_refcount_block *rb =
  85                 (struct ocfs2_refcount_block *)bh->b_data;
  86
  87         mlog(0, "Validating refcount block %llu\n",
  88              (unsigned long long)bh->b_blocknr);
  89
  90         BUG_ON(!buffer_uptodate(bh));
  91
  92         /*
  93          * If the ecc fails, we return the error but otherwise
  94          * leave the filesystem running.  We know any error is
  95          * local to this block.
  96          */
  97         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check);
  98         if (rc) {
  99                 mlog(ML_ERROR, "Checksum failed for refcount block %llu\n",
 100                      (unsigned long long)bh->b_blocknr);
 101                 return rc;
 102         }
 103
 104
 105         if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) {
 106                 ocfs2_error(sb,
 107                             "Refcount block #%llu has bad signature %.*s",
 108                             (unsigned long long)bh->b_blocknr, 7,
 109                             rb->rf_signature);
 110                 return -EINVAL;
 111         }
 112
 113         if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) {
 114                 ocfs2_error(sb,
 115                             "Refcount block #%llu has an invalid rf_blkno "
 116                             "of %llu",
 117                             (unsigned long long)bh->b_blocknr,
 118                             (unsigned long long)le64_to_cpu(rb->rf_blkno));
 119                 return -EINVAL;
 120         }
 121
 122         if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) {
 123                 ocfs2_error(sb,
 124                             "Refcount block #%llu has an invalid "
 125                             "rf_fs_generation of #%u",
 126                             (unsigned long long)bh->b_blocknr,
 127                             le32_to_cpu(rb->rf_fs_generation));
 128                 return -EINVAL;
 129         }
 130
 131         return 0;
 132 }
 133
 134 static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci,
 135                                      u64 rb_blkno,
 136                                      struct buffer_head **bh)
 137 {
 138         int rc;
 139         struct buffer_head *tmp = *bh;
 140
 141         rc = ocfs2_read_block(ci, rb_blkno, &tmp,
 142                               ocfs2_validate_refcount_block);
 143
 144         /* If ocfs2_read_block() got us a new bh, pass it up. */
 145         if (!rc && !*bh)
 146                 *bh = tmp;
 147
 148         return rc;
 149 }
 150
 151 static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci)
 152 {
 153         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 154
 155         return rf->rf_blkno;
 156 }
 157
 158 static struct super_block *
 159 ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci)
 160 {
 161         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 162
 163         return rf->rf_sb;
 164 }
 165
 166 static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci)
 167 {
 168         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 169
 170         spin_lock(&rf->rf_lock);
 171 }
 172
 173 static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci)
 174 {
 175         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 176
 177         spin_unlock(&rf->rf_lock);
 178 }
 179
 180 static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci)
 181 {
 182         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 183
 184         mutex_lock(&rf->rf_io_mutex);
 185 }
 186
 187 static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci)
 188 {
 189         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 190
 191         mutex_unlock(&rf->rf_io_mutex);
 192 }
 193
 194 static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = {
 195         .co_owner               = ocfs2_refcount_cache_owner,
 196         .co_get_super           = ocfs2_refcount_cache_get_super,
 197         .co_cache_lock          = ocfs2_refcount_cache_lock,
 198         .co_cache_unlock        = ocfs2_refcount_cache_unlock,
 199         .co_io_lock             = ocfs2_refcount_cache_io_lock,
 200         .co_io_unlock           = ocfs2_refcount_cache_io_unlock,
 201 };
 202
 203 static struct ocfs2_refcount_tree *
 204 ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno)
 205 {
 206         struct rb_node *n = osb->osb_rf_lock_tree.rb_node;
 207         struct ocfs2_refcount_tree *tree = NULL;
 208
 209         while (n) {
 210                 tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node);
 211
 212                 if (blkno < tree->rf_blkno)
 213                         n = n->rb_left;
 214                 else if (blkno > tree->rf_blkno)
 215                         n = n->rb_right;
 216                 else
 217                         return tree;
 218         }
 219
 220         return NULL;
 221 }
 222
 223 /* osb_lock is already locked. */
 224 static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb,
 225                                        struct ocfs2_refcount_tree *new)
 226 {
 227         u64 rf_blkno = new->rf_blkno;
 228         struct rb_node *parent = NULL;
 229         struct rb_node **p = &osb->osb_rf_lock_tree.rb_node;
 230         struct ocfs2_refcount_tree *tmp;
 231
 232         while (*p) {
 233                 parent = *p;
 234
 235                 tmp = rb_entry(parent, struct ocfs2_refcount_tree,
 236                                rf_node);
 237
 238                 if (rf_blkno < tmp->rf_blkno)
 239                         p = &(*p)->rb_left;
 240                 else if (rf_blkno > tmp->rf_blkno)
 241                         p = &(*p)->rb_right;
 242                 else {
 243                         /* This should never happen! */
 244                         mlog(ML_ERROR, "Duplicate refcount block %llu found!\n",
 245                              (unsigned long long)rf_blkno);
 246                         BUG();
 247                 }
 248         }
 249
 250         rb_link_node(&new->rf_node, parent, p);
 251         rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree);
 252 }
 253
 254 static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree)
 255 {
 256         ocfs2_metadata_cache_exit(&tree->rf_ci);
 257         ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres);
 258         ocfs2_lock_res_free(&tree->rf_lockres);
 259         kfree(tree);
 260 }
 261
 262 static inline void
 263 ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb,
 264                                         struct ocfs2_refcount_tree *tree)
 265 {
 266         rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree);
 267         if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree)
 268                 osb->osb_ref_tree_lru = NULL;
 269 }
 270
 271 static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb,
 272                                         struct ocfs2_refcount_tree *tree)
 273 {
 274         spin_lock(&osb->osb_lock);
 275         ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
 276         spin_unlock(&osb->osb_lock);
 277 }
 278
 279 static void ocfs2_kref_remove_refcount_tree(struct kref *kref)
 280 {
 281         struct ocfs2_refcount_tree *tree =
 282                 container_of(kref, struct ocfs2_refcount_tree, rf_getcnt);
 283
 284         ocfs2_free_refcount_tree(tree);
 285 }
 286
 287 static inline void
 288 ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree)
 289 {
 290         kref_get(&tree->rf_getcnt);
 291 }
 292
 293 static inline void
 294 ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree)
 295 {
 296         kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree);
 297 }
 298
 299 static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new,
 300                                                struct super_block *sb)
 301 {
 302         ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops);
 303         mutex_init(&new->rf_io_mutex);
 304         new->rf_sb = sb;
 305         spin_lock_init(&new->rf_lock);
 306 }
 307
 308 static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb,
 309                                         struct ocfs2_refcount_tree *new,
 310                                         u64 rf_blkno, u32 generation)
 311 {
 312         init_rwsem(&new->rf_sem);
 313         ocfs2_refcount_lock_res_init(&new->rf_lockres, osb,
 314                                      rf_blkno, generation);
 315 }
 316
 317 static struct ocfs2_refcount_tree*
 318 ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno)
 319 {
 320         struct ocfs2_refcount_tree *new;
 321
 322         new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS);
 323         if (!new)
 324                 return NULL;
 325
 326         new->rf_blkno = rf_blkno;
 327         kref_init(&new->rf_getcnt);
 328         ocfs2_init_refcount_tree_ci(new, osb->sb);
 329
 330         return new;
 331 }
 332
 333 static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno,
 334                                    struct ocfs2_refcount_tree **ret_tree)
 335 {
 336         int ret = 0;
 337         struct ocfs2_refcount_tree *tree, *new = NULL;
 338         struct buffer_head *ref_root_bh = NULL;
 339         struct ocfs2_refcount_block *ref_rb;
 340
 341         spin_lock(&osb->osb_lock);
 342         if (osb->osb_ref_tree_lru &&
 343             osb->osb_ref_tree_lru->rf_blkno == rf_blkno)
 344                 tree = osb->osb_ref_tree_lru;
 345         else
 346                 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
 347         if (tree)
 348                 goto out;
 349
 350         spin_unlock(&osb->osb_lock);
 351
 352         new = ocfs2_allocate_refcount_tree(osb, rf_blkno);
 353         if (!new) {
 354                 ret = -ENOMEM;
 355                 mlog_errno(ret);
 356                 return ret;
 357         }
 358         /*
 359          * We need the generation to create the refcount tree lock and since
 360          * it isn't changed during the tree modification, we are safe here to
 361          * read without protection.
 362          * We also have to purge the cache after we create the lock since the
 363          * refcount block may have the stale data. It can only be trusted when
 364          * we hold the refcount lock.
 365          */
 366         ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh);
 367         if (ret) {
 368                 mlog_errno(ret);
 369                 ocfs2_metadata_cache_exit(&new->rf_ci);
 370                 kfree(new);
 371                 return ret;
 372         }
 373
 374         ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 375         new->rf_generation = le32_to_cpu(ref_rb->rf_generation);
 376         ocfs2_init_refcount_tree_lock(osb, new, rf_blkno,
 377                                       new->rf_generation);
 378         ocfs2_metadata_cache_purge(&new->rf_ci);
 379
 380         spin_lock(&osb->osb_lock);
 381         tree = ocfs2_find_refcount_tree(osb, rf_blkno);
 382         if (tree)
 383                 goto out;
 384
 385         ocfs2_insert_refcount_tree(osb, new);
 386
 387         tree = new;
 388         new = NULL;
 389
 390 out:
 391         *ret_tree = tree;
 392
 393         osb->osb_ref_tree_lru = tree;
 394
 395         spin_unlock(&osb->osb_lock);
 396
 397         if (new)
 398                 ocfs2_free_refcount_tree(new);
 399
 400         brelse(ref_root_bh);
 401         return ret;
 402 }
 403
 404 static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno)
 405 {
 406         int ret;
 407         struct buffer_head *di_bh = NULL;
 408         struct ocfs2_dinode *di;
 409
 410         ret = ocfs2_read_inode_block(inode, &di_bh);
 411         if (ret) {
 412                 mlog_errno(ret);
 413                 goto out;
 414         }
 415
 416         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
 417
 418         di = (struct ocfs2_dinode *)di_bh->b_data;
 419         *ref_blkno = le64_to_cpu(di->i_refcount_loc);
 420         brelse(di_bh);
 421 out:
 422         return ret;
 423 }
 424
 425 static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
 426                                       struct ocfs2_refcount_tree *tree, int rw)
 427 {
 428         int ret;
 429
 430         ret = ocfs2_refcount_lock(tree, rw);
 431         if (ret) {
 432                 mlog_errno(ret);
 433                 goto out;
 434         }
 435
 436         if (rw)
 437                 down_write(&tree->rf_sem);
 438         else
 439                 down_read(&tree->rf_sem);
 440
 441 out:
 442         return ret;
 443 }
 444
 445 /*
 446  * Lock the refcount tree pointed by ref_blkno and return the tree.
 447  * In most case, we lock the tree and read the refcount block.
 448  * So read it here if the caller really needs it.
 449  *
 450  * If the tree has been re-created by other node, it will free the
 451  * old one and re-create it.
 452  */
 453 int ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
 454                              u64 ref_blkno, int rw,
 455                              struct ocfs2_refcount_tree **ret_tree,
 456                              struct buffer_head **ref_bh)
 457 {
 458         int ret, delete_tree = 0;
 459         struct ocfs2_refcount_tree *tree = NULL;
 460         struct buffer_head *ref_root_bh = NULL;
 461         struct ocfs2_refcount_block *rb;
 462
 463 again:
 464         ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree);
 465         if (ret) {
 466                 mlog_errno(ret);
 467                 return ret;
 468         }
 469
 470         ocfs2_refcount_tree_get(tree);
 471
 472         ret = __ocfs2_lock_refcount_tree(osb, tree, rw);
 473         if (ret) {
 474                 mlog_errno(ret);
 475                 ocfs2_refcount_tree_put(tree);
 476                 goto out;
 477         }
 478
 479         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
 480                                         &ref_root_bh);
 481         if (ret) {
 482                 mlog_errno(ret);
 483                 ocfs2_unlock_refcount_tree(osb, tree, rw);
 484                 ocfs2_refcount_tree_put(tree);
 485                 goto out;
 486         }
 487
 488         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 489         /*
 490          * If the refcount block has been freed and re-created, we may need
 491          * to recreate the refcount tree also.
 492          *
 493          * Here we just remove the tree from the rb-tree, and the last
 494          * kref holder will unlock and delete this refcount_tree.
 495          * Then we goto "again" and ocfs2_get_refcount_tree will create
 496          * the new refcount tree for us.
 497          */
 498         if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) {
 499                 if (!tree->rf_removed) {
 500                         ocfs2_erase_refcount_tree_from_list(osb, tree);
 501                         tree->rf_removed = 1;
 502                         delete_tree = 1;
 503                 }
 504
 505                 ocfs2_unlock_refcount_tree(osb, tree, rw);
 506                 /*
 507                  * We get an extra reference when we create the refcount
 508                  * tree, so another put will destroy it.
 509                  */
 510                 if (delete_tree)
 511                         ocfs2_refcount_tree_put(tree);
 512                 brelse(ref_root_bh);
 513                 ref_root_bh = NULL;
 514                 goto again;
 515         }
 516
 517         *ret_tree = tree;
 518         if (ref_bh) {
 519                 *ref_bh = ref_root_bh;
 520                 ref_root_bh = NULL;
 521         }
 522 out:
 523         brelse(ref_root_bh);
 524         return ret;
 525 }
 526
 527 void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
 528                                 struct ocfs2_refcount_tree *tree, int rw)
 529 {
 530         if (rw)
 531                 up_write(&tree->rf_sem);
 532         else
 533                 up_read(&tree->rf_sem);
 534
 535         ocfs2_refcount_unlock(tree, rw);
 536         ocfs2_refcount_tree_put(tree);
 537 }
 538
 539 void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
 540 {
 541         struct rb_node *node;
 542         struct ocfs2_refcount_tree *tree;
 543         struct rb_root *root = &osb->osb_rf_lock_tree;
 544
 545         while ((node = rb_last(root)) != NULL) {
 546                 tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
 547
 548                 mlog(0, "Purge tree %llu\n",
 549                      (unsigned long long) tree->rf_blkno);
 550
 551                 rb_erase(&tree->rf_node, root);
 552                 ocfs2_free_refcount_tree(tree);
 553         }
 554 }
 555
 556 /*
 557  * Create a refcount tree for an inode.
 558  * We take for granted that the inode is already locked.
 559  */
 560 static int ocfs2_create_refcount_tree(struct inode *inode,
 561                                       struct buffer_head *di_bh)
 562 {
 563         int ret;
 564         handle_t *handle = NULL;
 565         struct ocfs2_alloc_context *meta_ac = NULL;
 566         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 567         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 568         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 569         struct buffer_head *new_bh = NULL;
 570         struct ocfs2_refcount_block *rb;
 571         struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
 572         u16 suballoc_bit_start;
 573         u32 num_got;
 574         u64 first_blkno;
 575
 576         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
 577
 578         mlog(0, "create tree for inode %lu\n", inode->i_ino);
 579
 580         ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
 581         if (ret) {
 582                 mlog_errno(ret);
 583                 goto out;
 584         }
 585
 586         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
 587         if (IS_ERR(handle)) {
 588                 ret = PTR_ERR(handle);
 589                 mlog_errno(ret);
 590                 goto out;
 591         }
 592
 593         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 594                                       OCFS2_JOURNAL_ACCESS_WRITE);
 595         if (ret) {
 596                 mlog_errno(ret);
 597                 goto out_commit;
 598         }
 599
 600         ret = ocfs2_claim_metadata(osb, handle, meta_ac, 1,
 601                                    &suballoc_bit_start, &num_got,
 602                                    &first_blkno);
 603         if (ret) {
 604                 mlog_errno(ret);
 605                 goto out_commit;
 606         }
 607
 608         new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
 609         if (!new_tree) {
 610                 ret = -ENOMEM;
 611                 mlog_errno(ret);
 612                 goto out_commit;
 613         }
 614
 615         new_bh = sb_getblk(inode->i_sb, first_blkno);
 616         ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
 617
 618         ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
 619                                       OCFS2_JOURNAL_ACCESS_CREATE);
 620         if (ret) {
 621                 mlog_errno(ret);
 622                 goto out_commit;
 623         }
 624
 625         /* Initialize ocfs2_refcount_block. */
 626         rb = (struct ocfs2_refcount_block *)new_bh->b_data;
 627         memset(rb, 0, inode->i_sb->s_blocksize);
 628         strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
 629         rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
 630         rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
 631         rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
 632         rb->rf_blkno = cpu_to_le64(first_blkno);
 633         rb->rf_count = cpu_to_le32(1);
 634         rb->rf_records.rl_count =
 635                         cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
 636         spin_lock(&osb->osb_lock);
 637         rb->rf_generation = osb->s_next_generation++;
 638         spin_unlock(&osb->osb_lock);
 639
 640         ocfs2_journal_dirty(handle, new_bh);
 641
 642         spin_lock(&oi->ip_lock);
 643         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
 644         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 645         di->i_refcount_loc = cpu_to_le64(first_blkno);
 646         spin_unlock(&oi->ip_lock);
 647
 648         mlog(0, "created tree for inode %lu, refblock %llu\n",
 649              inode->i_ino, (unsigned long long)first_blkno);
 650
 651         ocfs2_journal_dirty(handle, di_bh);
 652
 653         /*
 654          * We have to init the tree lock here since it will use
 655          * the generation number to create it.
 656          */
 657         new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
 658         ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
 659                                       new_tree->rf_generation);
 660
 661         spin_lock(&osb->osb_lock);
 662         tree = ocfs2_find_refcount_tree(osb, first_blkno);
 663
 664         /*
 665          * We've just created a new refcount tree in this block.  If
 666          * we found a refcount tree on the ocfs2_super, it must be
 667          * one we just deleted.  We free the old tree before
 668          * inserting the new tree.
 669          */
 670         BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
 671         if (tree)
 672                 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
 673         ocfs2_insert_refcount_tree(osb, new_tree);
 674         spin_unlock(&osb->osb_lock);
 675         new_tree = NULL;
 676         if (tree)
 677                 ocfs2_refcount_tree_put(tree);
 678
 679 out_commit:
 680         ocfs2_commit_trans(osb, handle);
 681
 682 out:
 683         if (new_tree) {
 684                 ocfs2_metadata_cache_exit(&new_tree->rf_ci);
 685                 kfree(new_tree);
 686         }
 687
 688         brelse(new_bh);
 689         if (meta_ac)
 690                 ocfs2_free_alloc_context(meta_ac);
 691
 692         return ret;
 693 }
 694
 695 static int ocfs2_set_refcount_tree(struct inode *inode,
 696                                    struct buffer_head *di_bh,
 697                                    u64 refcount_loc)
 698 {
 699         int ret;
 700         handle_t *handle = NULL;
 701         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 702         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 703         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 704         struct buffer_head *ref_root_bh = NULL;
 705         struct ocfs2_refcount_block *rb;
 706         struct ocfs2_refcount_tree *ref_tree;
 707
 708         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
 709
 710         ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
 711                                        &ref_tree, &ref_root_bh);
 712         if (ret) {
 713                 mlog_errno(ret);
 714                 return ret;
 715         }
 716
 717         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
 718         if (IS_ERR(handle)) {
 719                 ret = PTR_ERR(handle);
 720                 mlog_errno(ret);
 721                 goto out;
 722         }
 723
 724         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 725                                       OCFS2_JOURNAL_ACCESS_WRITE);
 726         if (ret) {
 727                 mlog_errno(ret);
 728                 goto out_commit;
 729         }
 730
 731         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
 732                                       OCFS2_JOURNAL_ACCESS_WRITE);
 733         if (ret) {
 734                 mlog_errno(ret);
 735                 goto out_commit;
 736         }
 737
 738         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 739         le32_add_cpu(&rb->rf_count, 1);
 740
 741         ocfs2_journal_dirty(handle, ref_root_bh);
 742
 743         spin_lock(&oi->ip_lock);
 744         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
 745         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 746         di->i_refcount_loc = cpu_to_le64(refcount_loc);
 747         spin_unlock(&oi->ip_lock);
 748         ocfs2_journal_dirty(handle, di_bh);
 749
 750 out_commit:
 751         ocfs2_commit_trans(osb, handle);
 752 out:
 753         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 754         brelse(ref_root_bh);
 755
 756         return ret;
 757 }
 758
 759 int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
 760 {
 761         int ret, delete_tree = 0;
 762         handle_t *handle = NULL;
 763         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 764         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 765         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 766         struct ocfs2_refcount_block *rb;
 767         struct inode *alloc_inode = NULL;
 768         struct buffer_head *alloc_bh = NULL;
 769         struct buffer_head *blk_bh = NULL;
 770         struct ocfs2_refcount_tree *ref_tree;
 771         int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
 772         u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
 773         u16 bit = 0;
 774
 775         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
 776                 return 0;
 777
 778         BUG_ON(!ref_blkno);
 779         ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
 780         if (ret) {
 781                 mlog_errno(ret);
 782                 return ret;
 783         }
 784
 785         rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
 786
 787         /*
 788          * If we are the last user, we need to free the block.
 789          * So lock the allocator ahead.
 790          */
 791         if (le32_to_cpu(rb->rf_count) == 1) {
 792                 blk = le64_to_cpu(rb->rf_blkno);
 793                 bit = le16_to_cpu(rb->rf_suballoc_bit);
 794                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
 795
 796                 alloc_inode = ocfs2_get_system_file_inode(osb,
 797                                         EXTENT_ALLOC_SYSTEM_INODE,
 798                                         le16_to_cpu(rb->rf_suballoc_slot));
 799                 if (!alloc_inode) {
 800                         ret = -ENOMEM;
 801                         mlog_errno(ret);
 802                         goto out;
 803                 }
 804                 mutex_lock(&alloc_inode->i_mutex);
 805
 806                 ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
 807                 if (ret) {
 808                         mlog_errno(ret);
 809                         goto out_mutex;
 810                 }
 811
 812                 credits += OCFS2_SUBALLOC_FREE;
 813         }
 814
 815         handle = ocfs2_start_trans(osb, credits);
 816         if (IS_ERR(handle)) {
 817                 ret = PTR_ERR(handle);
 818                 mlog_errno(ret);
 819                 goto out_unlock;
 820         }
 821
 822         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 823                                       OCFS2_JOURNAL_ACCESS_WRITE);
 824         if (ret) {
 825                 mlog_errno(ret);
 826                 goto out_commit;
 827         }
 828
 829         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
 830                                       OCFS2_JOURNAL_ACCESS_WRITE);
 831         if (ret) {
 832                 mlog_errno(ret);
 833                 goto out_commit;
 834         }
 835
 836         spin_lock(&oi->ip_lock);
 837         oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
 838         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 839         di->i_refcount_loc = 0;
 840         spin_unlock(&oi->ip_lock);
 841         ocfs2_journal_dirty(handle, di_bh);
 842
 843         le32_add_cpu(&rb->rf_count , -1);
 844         ocfs2_journal_dirty(handle, blk_bh);
 845
 846         if (!rb->rf_count) {
 847                 delete_tree = 1;
 848                 ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
 849                 ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
 850                                                alloc_bh, bit, bg_blkno, 1);
 851                 if (ret)
 852                         mlog_errno(ret);
 853         }
 854
 855 out_commit:
 856         ocfs2_commit_trans(osb, handle);
 857 out_unlock:
 858         if (alloc_inode) {
 859                 ocfs2_inode_unlock(alloc_inode, 1);
 860                 brelse(alloc_bh);
 861         }
 862 out_mutex:
 863         if (alloc_inode) {
 864                 mutex_unlock(&alloc_inode->i_mutex);
 865                 iput(alloc_inode);
 866         }
 867 out:
 868         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 869         if (delete_tree)
 870                 ocfs2_refcount_tree_put(ref_tree);
 871         brelse(blk_bh);
 872
 873         return ret;
 874 }
 875
 876 static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
 877                                           struct buffer_head *ref_leaf_bh,
 878                                           u64 cpos, unsigned int len,
 879                                           struct ocfs2_refcount_rec *ret_rec,
 880                                           int *index)
 881 {
 882         int i = 0;
 883         struct ocfs2_refcount_block *rb =
 884                 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
 885         struct ocfs2_refcount_rec *rec = NULL;
 886
 887         for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
 888                 rec = &rb->rf_records.rl_recs[i];
 889
 890                 if (le64_to_cpu(rec->r_cpos) +
 891                     le32_to_cpu(rec->r_clusters) <= cpos)
 892                         continue;
 893                 else if (le64_to_cpu(rec->r_cpos) > cpos)
 894                         break;
 895
 896                 /* ok, cpos fail in this rec. Just return. */
 897                 if (ret_rec)
 898                         *ret_rec = *rec;
 899                 goto out;
 900         }
 901
 902         if (ret_rec) {
 903                 /* We meet with a hole here, so fake the rec. */
 904                 ret_rec->r_cpos = cpu_to_le64(cpos);
 905                 ret_rec->r_refcount = 0;
 906                 if (i < le16_to_cpu(rb->rf_records.rl_used) &&
 907                     le64_to_cpu(rec->r_cpos) < cpos + len)
 908                         ret_rec->r_clusters =
 909                                 cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
 910                 else
 911                         ret_rec->r_clusters = cpu_to_le32(len);
 912         }
 913
 914 out:
 915         *index = i;
 916 }
 917
 918 /*
 919  * Try to remove refcount tree. The mechanism is:
 920  * 1) Check whether i_clusters == 0, if no, exit.
 921  * 2) check whether we have i_xattr_loc in dinode. if yes, exit.
 922  * 3) Check whether we have inline xattr stored outside, if yes, exit.
 923  * 4) Remove the tree.
 924  */
 925 int ocfs2_try_remove_refcount_tree(struct inode *inode,
 926                                    struct buffer_head *di_bh)
 927 {
 928         int ret;
 929         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 930         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 931
 932         down_write(&oi->ip_xattr_sem);
 933         down_write(&oi->ip_alloc_sem);
 934
 935         if (oi->ip_clusters)
 936                 goto out;
 937
 938         if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc)
 939                 goto out;
 940
 941         if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL &&
 942             ocfs2_has_inline_xattr_value_outside(inode, di))
 943                 goto out;
 944
 945         ret = ocfs2_remove_refcount_tree(inode, di_bh);
 946         if (ret)
 947                 mlog_errno(ret);
 948 out:
 949         up_write(&oi->ip_alloc_sem);
 950         up_write(&oi->ip_xattr_sem);
 951         return 0;
 952 }
 953
 954 /*
 955  * Find the end range for a leaf refcount block indicated by
 956  * el->l_recs[index].e_blkno.
 957  */
 958 static int ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info *ci,
 959                                        struct buffer_head *ref_root_bh,
 960                                        struct ocfs2_extent_block *eb,
 961                                        struct ocfs2_extent_list *el,
 962                                        int index,  u32 *cpos_end)
 963 {
 964         int ret, i, subtree_root;
 965         u32 cpos;
 966         u64 blkno;
 967         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
 968         struct ocfs2_path *left_path = NULL, *right_path = NULL;
 969         struct ocfs2_extent_tree et;
 970         struct ocfs2_extent_list *tmp_el;
 971
 972         if (index < le16_to_cpu(el->l_next_free_rec) - 1) {
 973                 /*
 974                  * We have a extent rec after index, so just use the e_cpos
 975                  * of the next extent rec.
 976                  */
 977                 *cpos_end = le32_to_cpu(el->l_recs[index+1].e_cpos);
 978                 return 0;
 979         }
 980
 981         if (!eb || (eb && !eb->h_next_leaf_blk)) {
 982                 /*
 983                  * We are the last extent rec, so any high cpos should
 984                  * be stored in this leaf refcount block.
 985                  */
 986                 *cpos_end = UINT_MAX;
 987                 return 0;
 988         }
 989
 990         /*
 991          * If the extent block isn't the last one, we have to find
 992          * the subtree root between this extent block and the next
 993          * leaf extent block and get the corresponding e_cpos from
 994          * the subroot. Otherwise we may corrupt the b-tree.
 995          */
 996         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
 997
 998         left_path = ocfs2_new_path_from_et(&et);
 999         if (!left_path) {
1000                 ret = -ENOMEM;
1001                 mlog_errno(ret);
1002                 goto out;
1003         }
1004
1005         cpos = le32_to_cpu(eb->h_list.l_recs[index].e_cpos);
1006         ret = ocfs2_find_path(ci, left_path, cpos);
1007         if (ret) {
1008                 mlog_errno(ret);
1009                 goto out;
1010         }
1011
1012         right_path = ocfs2_new_path_from_path(left_path);
1013         if (!right_path) {
1014                 ret = -ENOMEM;
1015                 mlog_errno(ret);
1016                 goto out;
1017         }
1018
1019         ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, &cpos);
1020         if (ret) {
1021                 mlog_errno(ret);
1022                 goto out;
1023         }
1024
1025         ret = ocfs2_find_path(ci, right_path, cpos);
1026         if (ret) {
1027                 mlog_errno(ret);
1028                 goto out;
1029         }
1030
1031         subtree_root = ocfs2_find_subtree_root(&et, left_path,
1032                                                right_path);
1033
1034         tmp_el = left_path->p_node[subtree_root].el;
1035         blkno = left_path->p_node[subtree_root+1].bh->b_blocknr;
1036         for (i = 0; i < le32_to_cpu(tmp_el->l_next_free_rec); i++) {
1037                 if (le64_to_cpu(tmp_el->l_recs[i].e_blkno) == blkno) {
1038                         *cpos_end = le32_to_cpu(tmp_el->l_recs[i+1].e_cpos);
1039                         break;
1040                 }
1041         }
1042
1043         BUG_ON(i == le32_to_cpu(tmp_el->l_next_free_rec));
1044
1045 out:
1046         ocfs2_free_path(left_path);
1047         ocfs2_free_path(right_path);
1048         return ret;
1049 }
1050
1051 /*
1052  * Given a cpos and len, try to find the refcount record which contains cpos.
1053  * 1. If cpos can be found in one refcount record, return the record.
1054  * 2. If cpos can't be found, return a fake record which start from cpos
1055  *    and end at a small value between cpos+len and start of the next record.
1056  *    This fake record has r_refcount = 0.
1057  */
1058 static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
1059                                   struct buffer_head *ref_root_bh,
1060                                   u64 cpos, unsigned int len,
1061                                   struct ocfs2_refcount_rec *ret_rec,
1062                                   int *index,
1063                                   struct buffer_head **ret_bh)
1064 {
1065         int ret = 0, i, found;
1066         u32 low_cpos, uninitialized_var(cpos_end);
1067         struct ocfs2_extent_list *el;
1068         struct ocfs2_extent_rec *rec = NULL;
1069         struct ocfs2_extent_block *eb = NULL;
1070         struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
1071         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1072         struct ocfs2_refcount_block *rb =
1073                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1074
1075         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
1076                 ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
1077                                               ret_rec, index);
1078                 *ret_bh = ref_root_bh;
1079                 get_bh(ref_root_bh);
1080                 return 0;
1081         }
1082
1083         el = &rb->rf_list;
1084         low_cpos = cpos & OCFS2_32BIT_POS_MASK;
1085
1086         if (el->l_tree_depth) {
1087                 ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
1088                 if (ret) {
1089                         mlog_errno(ret);
1090                         goto out;
1091                 }
1092
1093                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1094                 el = &eb->h_list;
1095
1096                 if (el->l_tree_depth) {
1097                         ocfs2_error(sb,
1098                         "refcount tree %llu has non zero tree "
1099                         "depth in leaf btree tree block %llu\n",
1100                         (unsigned long long)ocfs2_metadata_cache_owner(ci),
1101                         (unsigned long long)eb_bh->b_blocknr);
1102                         ret = -EROFS;
1103                         goto out;
1104                 }
1105         }
1106
1107         found = 0;
1108         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1109                 rec = &el->l_recs[i];
1110
1111                 if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
1112                         found = 1;
1113                         break;
1114                 }
1115         }
1116
1117         if (found) {
1118                 ret = ocfs2_get_refcount_cpos_end(ci, ref_root_bh,
1119                                                   eb, el, i, &cpos_end);
1120                 if (ret) {
1121                         mlog_errno(ret);
1122                         goto out;
1123                 }
1124
1125                 if (cpos_end < low_cpos + len)
1126                         len = cpos_end - low_cpos;
1127         }
1128
1129         ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
1130                                         &ref_leaf_bh);
1131         if (ret) {
1132                 mlog_errno(ret);
1133                 goto out;
1134         }
1135
1136         ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
1137                                       ret_rec, index);
1138         *ret_bh = ref_leaf_bh;
1139 out:
1140         brelse(eb_bh);
1141         return ret;
1142 }
1143
1144 enum ocfs2_ref_rec_contig {
1145         REF_CONTIG_NONE = 0,
1146         REF_CONTIG_LEFT,
1147         REF_CONTIG_RIGHT,
1148         REF_CONTIG_LEFTRIGHT,
1149 };
1150
1151 static enum ocfs2_ref_rec_contig
1152         ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
1153                                     int index)
1154 {
1155         if ((rb->rf_records.rl_recs[index].r_refcount ==
1156             rb->rf_records.rl_recs[index + 1].r_refcount) &&
1157             (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
1158             le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1159             le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1160                 return REF_CONTIG_RIGHT;
1161
1162         return REF_CONTIG_NONE;
1163 }
1164
1165 static enum ocfs2_ref_rec_contig
1166         ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1167                                   int index)
1168 {
1169         enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1170
1171         if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1172                 ret = ocfs2_refcount_rec_adjacent(rb, index);
1173
1174         if (index > 0) {
1175                 enum ocfs2_ref_rec_contig tmp;
1176
1177                 tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1178
1179                 if (tmp == REF_CONTIG_RIGHT) {
1180                         if (ret == REF_CONTIG_RIGHT)
1181                                 ret = REF_CONTIG_LEFTRIGHT;
1182                         else
1183                                 ret = REF_CONTIG_LEFT;
1184                 }
1185         }
1186
1187         return ret;
1188 }
1189
1190 static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1191                                            int index)
1192 {
1193         BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1194                rb->rf_records.rl_recs[index+1].r_refcount);
1195
1196         le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1197                      le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1198
1199         if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1200                 memmove(&rb->rf_records.rl_recs[index + 1],
1201                         &rb->rf_records.rl_recs[index + 2],
1202                         sizeof(struct ocfs2_refcount_rec) *
1203                         (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1204
1205         memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1206                0, sizeof(struct ocfs2_refcount_rec));
1207         le16_add_cpu(&rb->rf_records.rl_used, -1);
1208 }
1209
1210 /*
1211  * Merge the refcount rec if we are contiguous with the adjacent recs.
1212  */
1213 static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1214                                      int index)
1215 {
1216         enum ocfs2_ref_rec_contig contig =
1217                                 ocfs2_refcount_rec_contig(rb, index);
1218
1219         if (contig == REF_CONTIG_NONE)
1220                 return;
1221
1222         if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1223                 BUG_ON(index == 0);
1224                 index--;
1225         }
1226
1227         ocfs2_rotate_refcount_rec_left(rb, index);
1228
1229         if (contig == REF_CONTIG_LEFTRIGHT)
1230                 ocfs2_rotate_refcount_rec_left(rb, index);
1231 }
1232
1233 /*
1234  * Change the refcount indexed by "index" in ref_bh.
1235  * If refcount reaches 0, remove it.
1236  */
1237 static int ocfs2_change_refcount_rec(handle_t *handle,
1238                                      struct ocfs2_caching_info *ci,
1239                                      struct buffer_head *ref_leaf_bh,
1240                                      int index, int merge, int change)
1241 {
1242         int ret;
1243         struct ocfs2_refcount_block *rb =
1244                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1245         struct ocfs2_refcount_list *rl = &rb->rf_records;
1246         struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1247
1248         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1249                                       OCFS2_JOURNAL_ACCESS_WRITE);
1250         if (ret) {
1251                 mlog_errno(ret);
1252                 goto out;
1253         }
1254
1255         mlog(0, "change index %d, old count %u, change %d\n", index,
1256              le32_to_cpu(rec->r_refcount), change);
1257         le32_add_cpu(&rec->r_refcount, change);
1258
1259         if (!rec->r_refcount) {
1260                 if (index != le16_to_cpu(rl->rl_used) - 1) {
1261                         memmove(rec, rec + 1,
1262                                 (le16_to_cpu(rl->rl_used) - index - 1) *
1263                                 sizeof(struct ocfs2_refcount_rec));
1264                         memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
1265                                0, sizeof(struct ocfs2_refcount_rec));
1266                 }
1267
1268                 le16_add_cpu(&rl->rl_used, -1);
1269         } else if (merge)
1270                 ocfs2_refcount_rec_merge(rb, index);
1271
1272         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1273         if (ret)
1274                 mlog_errno(ret);
1275 out:
1276         return ret;
1277 }
1278
1279 static int ocfs2_expand_inline_ref_root(handle_t *handle,
1280                                         struct ocfs2_caching_info *ci,
1281                                         struct buffer_head *ref_root_bh,
1282                                         struct buffer_head **ref_leaf_bh,
1283                                         struct ocfs2_alloc_context *meta_ac)
1284 {
1285         int ret;
1286         u16 suballoc_bit_start;
1287         u32 num_got;
1288         u64 blkno;
1289         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1290         struct buffer_head *new_bh = NULL;
1291         struct ocfs2_refcount_block *new_rb;
1292         struct ocfs2_refcount_block *root_rb =
1293                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1294
1295         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1296                                       OCFS2_JOURNAL_ACCESS_WRITE);
1297         if (ret) {
1298                 mlog_errno(ret);
1299                 goto out;
1300         }
1301
1302         ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1303                                    &suballoc_bit_start, &num_got,
1304                                    &blkno);
1305         if (ret) {
1306                 mlog_errno(ret);
1307                 goto out;
1308         }
1309
1310         new_bh = sb_getblk(sb, blkno);
1311         if (new_bh == NULL) {
1312                 ret = -EIO;
1313                 mlog_errno(ret);
1314                 goto out;
1315         }
1316         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1317
1318         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1319                                       OCFS2_JOURNAL_ACCESS_CREATE);
1320         if (ret) {
1321                 mlog_errno(ret);
1322                 goto out;
1323         }
1324
1325         /*
1326          * Initialize ocfs2_refcount_block.
1327          * It should contain the same information as the old root.
1328          * so just memcpy it and change the corresponding field.
1329          */
1330         memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1331
1332         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1333         new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
1334         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1335         new_rb->rf_blkno = cpu_to_le64(blkno);
1336         new_rb->rf_cpos = cpu_to_le32(0);
1337         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1338         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1339         ocfs2_journal_dirty(handle, new_bh);
1340
1341         /* Now change the root. */
1342         memset(&root_rb->rf_list, 0, sb->s_blocksize -
1343                offsetof(struct ocfs2_refcount_block, rf_list));
1344         root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1345         root_rb->rf_clusters = cpu_to_le32(1);
1346         root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1347         root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1348         root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1349         root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1350
1351         ocfs2_journal_dirty(handle, ref_root_bh);
1352
1353         mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno,
1354              le16_to_cpu(new_rb->rf_records.rl_used));
1355
1356         *ref_leaf_bh = new_bh;
1357         new_bh = NULL;
1358 out:
1359         brelse(new_bh);
1360         return ret;
1361 }
1362
1363 static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1364                                            struct ocfs2_refcount_rec *next)
1365 {
1366         if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1367                 ocfs2_get_ref_rec_low_cpos(next))
1368                 return 1;
1369
1370         return 0;
1371 }
1372
1373 static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1374 {
1375         const struct ocfs2_refcount_rec *l = a, *r = b;
1376         u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1377         u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1378
1379         if (l_cpos > r_cpos)
1380                 return 1;
1381         if (l_cpos < r_cpos)
1382                 return -1;
1383         return 0;
1384 }
1385
1386 static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1387 {
1388         const struct ocfs2_refcount_rec *l = a, *r = b;
1389         u64 l_cpos = le64_to_cpu(l->r_cpos);
1390         u64 r_cpos = le64_to_cpu(r->r_cpos);
1391
1392         if (l_cpos > r_cpos)
1393                 return 1;
1394         if (l_cpos < r_cpos)
1395                 return -1;
1396         return 0;
1397 }
1398
1399 static void swap_refcount_rec(void *a, void *b, int size)
1400 {
1401         struct ocfs2_refcount_rec *l = a, *r = b, tmp;
1402
1403         tmp = *(struct ocfs2_refcount_rec *)l;
1404         *(struct ocfs2_refcount_rec *)l =
1405                         *(struct ocfs2_refcount_rec *)r;
1406         *(struct ocfs2_refcount_rec *)r = tmp;
1407 }
1408
1409 /*
1410  * The refcount cpos are ordered by their 64bit cpos,
1411  * But we will use the low 32 bit to be the e_cpos in the b-tree.
1412  * So we need to make sure that this pos isn't intersected with others.
1413  *
1414  * Note: The refcount block is already sorted by their low 32 bit cpos,
1415  *       So just try the middle pos first, and we will exit when we find
1416  *       the good position.
1417  */
1418 static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1419                                          u32 *split_pos, int *split_index)
1420 {
1421         int num_used = le16_to_cpu(rl->rl_used);
1422         int delta, middle = num_used / 2;
1423
1424         for (delta = 0; delta < middle; delta++) {
1425                 /* Let's check delta earlier than middle */
1426                 if (ocfs2_refcount_rec_no_intersect(
1427                                         &rl->rl_recs[middle - delta - 1],
1428                                         &rl->rl_recs[middle - delta])) {
1429                         *split_index = middle - delta;
1430                         break;
1431                 }
1432
1433                 /* For even counts, don't walk off the end */
1434                 if ((middle + delta + 1) == num_used)
1435                         continue;
1436
1437                 /* Now try delta past middle */
1438                 if (ocfs2_refcount_rec_no_intersect(
1439                                         &rl->rl_recs[middle + delta],
1440                                         &rl->rl_recs[middle + delta + 1])) {
1441                         *split_index = middle + delta + 1;
1442                         break;
1443                 }
1444         }
1445
1446         if (delta >= middle)
1447                 return -ENOSPC;
1448
1449         *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1450         return 0;
1451 }
1452
1453 static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1454                                             struct buffer_head *new_bh,
1455                                             u32 *split_cpos)
1456 {
1457         int split_index = 0, num_moved, ret;
1458         u32 cpos = 0;
1459         struct ocfs2_refcount_block *rb =
1460                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1461         struct ocfs2_refcount_list *rl = &rb->rf_records;
1462         struct ocfs2_refcount_block *new_rb =
1463                         (struct ocfs2_refcount_block *)new_bh->b_data;
1464         struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1465
1466         mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n",
1467              (unsigned long long)ref_leaf_bh->b_blocknr,
1468              le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used));
1469
1470         /*
1471          * XXX: Improvement later.
1472          * If we know all the high 32 bit cpos is the same, no need to sort.
1473          *
1474          * In order to make the whole process safe, we do:
1475          * 1. sort the entries by their low 32 bit cpos first so that we can
1476          *    find the split cpos easily.
1477          * 2. call ocfs2_insert_extent to insert the new refcount block.
1478          * 3. move the refcount rec to the new block.
1479          * 4. sort the entries by their 64 bit cpos.
1480          * 5. dirty the new_rb and rb.
1481          */
1482         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1483              sizeof(struct ocfs2_refcount_rec),
1484              cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1485
1486         ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1487         if (ret) {
1488                 mlog_errno(ret);
1489                 return ret;
1490         }
1491
1492         new_rb->rf_cpos = cpu_to_le32(cpos);
1493
1494         /* move refcount records starting from split_index to the new block. */
1495         num_moved = le16_to_cpu(rl->rl_used) - split_index;
1496         memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1497                num_moved * sizeof(struct ocfs2_refcount_rec));
1498
1499         /*ok, remove the entries we just moved over to the other block. */
1500         memset(&rl->rl_recs[split_index], 0,
1501                num_moved * sizeof(struct ocfs2_refcount_rec));
1502
1503         /* change old and new rl_used accordingly. */
1504         le16_add_cpu(&rl->rl_used, -num_moved);
1505         new_rl->rl_used = cpu_to_le16(num_moved);
1506
1507         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1508              sizeof(struct ocfs2_refcount_rec),
1509              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1510
1511         sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1512              sizeof(struct ocfs2_refcount_rec),
1513              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1514
1515         *split_cpos = cpos;
1516         return 0;
1517 }
1518
1519 static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1520                                          struct ocfs2_caching_info *ci,
1521                                          struct buffer_head *ref_root_bh,
1522                                          struct buffer_head *ref_leaf_bh,
1523                                          struct ocfs2_alloc_context *meta_ac)
1524 {
1525         int ret;
1526         u16 suballoc_bit_start;
1527         u32 num_got, new_cpos;
1528         u64 blkno;
1529         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1530         struct ocfs2_refcount_block *root_rb =
1531                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1532         struct buffer_head *new_bh = NULL;
1533         struct ocfs2_refcount_block *new_rb;
1534         struct ocfs2_extent_tree ref_et;
1535
1536         BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1537
1538         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1539                                       OCFS2_JOURNAL_ACCESS_WRITE);
1540         if (ret) {
1541                 mlog_errno(ret);
1542                 goto out;
1543         }
1544
1545         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1546                                       OCFS2_JOURNAL_ACCESS_WRITE);
1547         if (ret) {
1548                 mlog_errno(ret);
1549                 goto out;
1550         }
1551
1552         ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1,
1553                                    &suballoc_bit_start, &num_got,
1554                                    &blkno);
1555         if (ret) {
1556                 mlog_errno(ret);
1557                 goto out;
1558         }
1559
1560         new_bh = sb_getblk(sb, blkno);
1561         if (new_bh == NULL) {
1562                 ret = -EIO;
1563                 mlog_errno(ret);
1564                 goto out;
1565         }
1566         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1567
1568         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1569                                       OCFS2_JOURNAL_ACCESS_CREATE);
1570         if (ret) {
1571                 mlog_errno(ret);
1572                 goto out;
1573         }
1574
1575         /* Initialize ocfs2_refcount_block. */
1576         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1577         memset(new_rb, 0, sb->s_blocksize);
1578         strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1579         new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
1580         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1581         new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1582         new_rb->rf_blkno = cpu_to_le64(blkno);
1583         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1584         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1585         new_rb->rf_records.rl_count =
1586                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1587         new_rb->rf_generation = root_rb->rf_generation;
1588
1589         ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1590         if (ret) {
1591                 mlog_errno(ret);
1592                 goto out;
1593         }
1594
1595         ocfs2_journal_dirty(handle, ref_leaf_bh);
1596         ocfs2_journal_dirty(handle, new_bh);
1597
1598         ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1599
1600         mlog(0, "insert new leaf block %llu at %u\n",
1601              (unsigned long long)new_bh->b_blocknr, new_cpos);
1602
1603         /* Insert the new leaf block with the specific offset cpos. */
1604         ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1605                                   1, 0, meta_ac);
1606         if (ret)
1607                 mlog_errno(ret);
1608
1609 out:
1610         brelse(new_bh);
1611         return ret;
1612 }
1613
1614 static int ocfs2_expand_refcount_tree(handle_t *handle,
1615                                       struct ocfs2_caching_info *ci,
1616                                       struct buffer_head *ref_root_bh,
1617                                       struct buffer_head *ref_leaf_bh,
1618                                       struct ocfs2_alloc_context *meta_ac)
1619 {
1620         int ret;
1621         struct buffer_head *expand_bh = NULL;
1622
1623         if (ref_root_bh == ref_leaf_bh) {
1624                 /*
1625                  * the old root bh hasn't been expanded to a b-tree,
1626                  * so expand it first.
1627                  */
1628                 ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1629                                                    &expand_bh, meta_ac);
1630                 if (ret) {
1631                         mlog_errno(ret);
1632                         goto out;
1633                 }
1634         } else {
1635                 expand_bh = ref_leaf_bh;
1636                 get_bh(expand_bh);
1637         }
1638
1639
1640         /* Now add a new refcount block into the tree.*/
1641         ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1642                                             expand_bh, meta_ac);
1643         if (ret)
1644                 mlog_errno(ret);
1645 out:
1646         brelse(expand_bh);
1647         return ret;
1648 }
1649
1650 /*
1651  * Adjust the extent rec in b-tree representing ref_leaf_bh.
1652  *
1653  * Only called when we have inserted a new refcount rec at index 0
1654  * which means ocfs2_extent_rec.e_cpos may need some change.
1655  */
1656 static int ocfs2_adjust_refcount_rec(handle_t *handle,
1657                                      struct ocfs2_caching_info *ci,
1658                                      struct buffer_head *ref_root_bh,
1659                                      struct buffer_head *ref_leaf_bh,
1660                                      struct ocfs2_refcount_rec *rec)
1661 {
1662         int ret = 0, i;
1663         u32 new_cpos, old_cpos;
1664         struct ocfs2_path *path = NULL;
1665         struct ocfs2_extent_tree et;
1666         struct ocfs2_refcount_block *rb =
1667                 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1668         struct ocfs2_extent_list *el;
1669
1670         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1671                 goto out;
1672
1673         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1674         old_cpos = le32_to_cpu(rb->rf_cpos);
1675         new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1676         if (old_cpos <= new_cpos)
1677                 goto out;
1678
1679         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1680
1681         path = ocfs2_new_path_from_et(&et);
1682         if (!path) {
1683                 ret = -ENOMEM;
1684                 mlog_errno(ret);
1685                 goto out;
1686         }
1687
1688         ret = ocfs2_find_path(ci, path, old_cpos);
1689         if (ret) {
1690                 mlog_errno(ret);
1691                 goto out;
1692         }
1693
1694         /*
1695          * 2 more credits, one for the leaf refcount block, one for
1696          * the extent block contains the extent rec.
1697          */
1698         ret = ocfs2_extend_trans(handle, handle->h_buffer_credits + 2);
1699         if (ret < 0) {
1700                 mlog_errno(ret);
1701                 goto out;
1702         }
1703
1704         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1705                                       OCFS2_JOURNAL_ACCESS_WRITE);
1706         if (ret < 0) {
1707                 mlog_errno(ret);
1708                 goto out;
1709         }
1710
1711         ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1712                                       OCFS2_JOURNAL_ACCESS_WRITE);
1713         if (ret < 0) {
1714                 mlog_errno(ret);
1715                 goto out;
1716         }
1717
1718         /* change the leaf extent block first. */
1719         el = path_leaf_el(path);
1720
1721         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1722                 if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1723                         break;
1724
1725         BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1726
1727         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1728
1729         /* change the r_cpos in the leaf block. */
1730         rb->rf_cpos = cpu_to_le32(new_cpos);
1731
1732         ocfs2_journal_dirty(handle, path_leaf_bh(path));
1733         ocfs2_journal_dirty(handle, ref_leaf_bh);
1734
1735 out:
1736         ocfs2_free_path(path);
1737         return ret;
1738 }
1739
1740 static int ocfs2_insert_refcount_rec(handle_t *handle,
1741                                      struct ocfs2_caching_info *ci,
1742                                      struct buffer_head *ref_root_bh,
1743                                      struct buffer_head *ref_leaf_bh,
1744                                      struct ocfs2_refcount_rec *rec,
1745                                      int index, int merge,
1746                                      struct ocfs2_alloc_context *meta_ac)
1747 {
1748         int ret;
1749         struct ocfs2_refcount_block *rb =
1750                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1751         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1752         struct buffer_head *new_bh = NULL;
1753
1754         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1755
1756         if (rf_list->rl_used == rf_list->rl_count) {
1757                 u64 cpos = le64_to_cpu(rec->r_cpos);
1758                 u32 len = le32_to_cpu(rec->r_clusters);
1759
1760                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1761                                                  ref_leaf_bh, meta_ac);
1762                 if (ret) {
1763                         mlog_errno(ret);
1764                         goto out;
1765                 }
1766
1767                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1768                                              cpos, len, NULL, &index,
1769                                              &new_bh);
1770                 if (ret) {
1771                         mlog_errno(ret);
1772                         goto out;
1773                 }
1774
1775                 ref_leaf_bh = new_bh;
1776                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1777                 rf_list = &rb->rf_records;
1778         }
1779
1780         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1781                                       OCFS2_JOURNAL_ACCESS_WRITE);
1782         if (ret) {
1783                 mlog_errno(ret);
1784                 goto out;
1785         }
1786
1787         if (index < le16_to_cpu(rf_list->rl_used))
1788                 memmove(&rf_list->rl_recs[index + 1],
1789                         &rf_list->rl_recs[index],
1790                         (le16_to_cpu(rf_list->rl_used) - index) *
1791                          sizeof(struct ocfs2_refcount_rec));
1792
1793         mlog(0, "insert refcount record start %llu, len %u, count %u "
1794              "to leaf block %llu at index %d\n",
1795              (unsigned long long)le64_to_cpu(rec->r_cpos),
1796              le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount),
1797              (unsigned long long)ref_leaf_bh->b_blocknr, index);
1798
1799         rf_list->rl_recs[index] = *rec;
1800
1801         le16_add_cpu(&rf_list->rl_used, 1);
1802
1803         if (merge)
1804                 ocfs2_refcount_rec_merge(rb, index);
1805
1806         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1807         if (ret) {
1808                 mlog_errno(ret);
1809                 goto out;
1810         }
1811
1812         if (index == 0) {
1813                 ret = ocfs2_adjust_refcount_rec(handle, ci,
1814                                                 ref_root_bh,
1815                                                 ref_leaf_bh, rec);
1816                 if (ret)
1817                         mlog_errno(ret);
1818         }
1819 out:
1820         brelse(new_bh);
1821         return ret;
1822 }
1823
1824 /*
1825  * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1826  * This is much simple than our b-tree code.
1827  * split_rec is the new refcount rec we want to insert.
1828  * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1829  * increase refcount or decrease a refcount to non-zero).
1830  * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1831  * rec( in case we decrease a refcount to zero).
1832  */
1833 static int ocfs2_split_refcount_rec(handle_t *handle,
1834                                     struct ocfs2_caching_info *ci,
1835                                     struct buffer_head *ref_root_bh,
1836                                     struct buffer_head *ref_leaf_bh,
1837                                     struct ocfs2_refcount_rec *split_rec,
1838                                     int index, int merge,
1839                                     struct ocfs2_alloc_context *meta_ac,
1840                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
1841 {
1842         int ret, recs_need;
1843         u32 len;
1844         struct ocfs2_refcount_block *rb =
1845                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1846         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1847         struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1848         struct ocfs2_refcount_rec *tail_rec = NULL;
1849         struct buffer_head *new_bh = NULL;
1850
1851         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1852
1853         mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n",
1854              le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters),
1855              le64_to_cpu(split_rec->r_cpos),
1856              le32_to_cpu(split_rec->r_clusters));
1857
1858         /*
1859          * If we just need to split the header or tail clusters,
1860          * no more recs are needed, just split is OK.
1861          * Otherwise we at least need one new recs.
1862          */
1863         if (!split_rec->r_refcount &&
1864             (split_rec->r_cpos == orig_rec->r_cpos ||
1865              le64_to_cpu(split_rec->r_cpos) +
1866              le32_to_cpu(split_rec->r_clusters) ==
1867              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1868                 recs_need = 0;
1869         else
1870                 recs_need = 1;
1871
1872         /*
1873          * We need one more rec if we split in the middle and the new rec have
1874          * some refcount in it.
1875          */
1876         if (split_rec->r_refcount &&
1877             (split_rec->r_cpos != orig_rec->r_cpos &&
1878              le64_to_cpu(split_rec->r_cpos) +
1879              le32_to_cpu(split_rec->r_clusters) !=
1880              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1881                 recs_need++;
1882
1883         /* If the leaf block don't have enough record, expand it. */
1884         if (le16_to_cpu(rf_list->rl_used) + recs_need >
1885                                          le16_to_cpu(rf_list->rl_count)) {
1886                 struct ocfs2_refcount_rec tmp_rec;
1887                 u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1888                 len = le32_to_cpu(orig_rec->r_clusters);
1889                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1890                                                  ref_leaf_bh, meta_ac);
1891                 if (ret) {
1892                         mlog_errno(ret);
1893                         goto out;
1894                 }
1895
1896                 /*
1897                  * We have to re-get it since now cpos may be moved to
1898                  * another leaf block.
1899                  */
1900                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1901                                              cpos, len, &tmp_rec, &index,
1902                                              &new_bh);
1903                 if (ret) {
1904                         mlog_errno(ret);
1905                         goto out;
1906                 }
1907
1908                 ref_leaf_bh = new_bh;
1909                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1910                 rf_list = &rb->rf_records;
1911                 orig_rec = &rf_list->rl_recs[index];
1912         }
1913
1914         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1915                                       OCFS2_JOURNAL_ACCESS_WRITE);
1916         if (ret) {
1917                 mlog_errno(ret);
1918                 goto out;
1919         }
1920
1921         /*
1922          * We have calculated out how many new records we need and store
1923          * in recs_need, so spare enough space first by moving the records
1924          * after "index" to the end.
1925          */
1926         if (index != le16_to_cpu(rf_list->rl_used) - 1)
1927                 memmove(&rf_list->rl_recs[index + 1 + recs_need],
1928                         &rf_list->rl_recs[index + 1],
1929                         (le16_to_cpu(rf_list->rl_used) - index - 1) *
1930                          sizeof(struct ocfs2_refcount_rec));
1931
1932         len = (le64_to_cpu(orig_rec->r_cpos) +
1933               le32_to_cpu(orig_rec->r_clusters)) -
1934               (le64_to_cpu(split_rec->r_cpos) +
1935               le32_to_cpu(split_rec->r_clusters));
1936
1937         /*
1938          * If we have "len", the we will split in the tail and move it
1939          * to the end of the space we have just spared.
1940          */
1941         if (len) {
1942                 tail_rec = &rf_list->rl_recs[index + recs_need];
1943
1944                 memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1945                 le64_add_cpu(&tail_rec->r_cpos,
1946                              le32_to_cpu(tail_rec->r_clusters) - len);
1947                 tail_rec->r_clusters = cpu_to_le32(len);
1948         }
1949
1950         /*
1951          * If the split pos isn't the same as the original one, we need to
1952          * split in the head.
1953          *
1954          * Note: We have the chance that split_rec.r_refcount = 0,
1955          * recs_need = 0 and len > 0, which means we just cut the head from
1956          * the orig_rec and in that case we have done some modification in
1957          * orig_rec above, so the check for r_cpos is faked.
1958          */
1959         if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1960                 len = le64_to_cpu(split_rec->r_cpos) -
1961                       le64_to_cpu(orig_rec->r_cpos);
1962                 orig_rec->r_clusters = cpu_to_le32(len);
1963                 index++;
1964         }
1965
1966         le16_add_cpu(&rf_list->rl_used, recs_need);
1967
1968         if (split_rec->r_refcount) {
1969                 rf_list->rl_recs[index] = *split_rec;
1970                 mlog(0, "insert refcount record start %llu, len %u, count %u "
1971                      "to leaf block %llu at index %d\n",
1972                      (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1973                      le32_to_cpu(split_rec->r_clusters),
1974                      le32_to_cpu(split_rec->r_refcount),
1975                      (unsigned long long)ref_leaf_bh->b_blocknr, index);
1976
1977                 if (merge)
1978                         ocfs2_refcount_rec_merge(rb, index);
1979         }
1980
1981         ret = ocfs2_journal_dirty(handle, ref_leaf_bh);
1982         if (ret)
1983                 mlog_errno(ret);
1984
1985 out:
1986         brelse(new_bh);
1987         return ret;
1988 }
1989
1990 static int __ocfs2_increase_refcount(handle_t *handle,
1991                                      struct ocfs2_caching_info *ci,
1992                                      struct buffer_head *ref_root_bh,
1993                                      u64 cpos, u32 len, int merge,
1994                                      struct ocfs2_alloc_context *meta_ac,
1995                                      struct ocfs2_cached_dealloc_ctxt *dealloc)
1996 {
1997         int ret = 0, index;
1998         struct buffer_head *ref_leaf_bh = NULL;
1999         struct ocfs2_refcount_rec rec;
2000         unsigned int set_len = 0;
2001
2002         mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n",
2003              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2004              (unsigned long long)cpos, len);
2005
2006         while (len) {
2007                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2008                                              cpos, len, &rec, &index,
2009                                              &ref_leaf_bh);
2010                 if (ret) {
2011                         mlog_errno(ret);
2012                         goto out;
2013                 }
2014
2015                 set_len = le32_to_cpu(rec.r_clusters);
2016
2017                 /*
2018                  * Here we may meet with 3 situations:
2019                  *
2020                  * 1. If we find an already existing record, and the length
2021                  *    is the same, cool, we just need to increase the r_refcount
2022                  *    and it is OK.
2023                  * 2. If we find a hole, just insert it with r_refcount = 1.
2024                  * 3. If we are in the middle of one extent record, split
2025                  *    it.
2026                  */
2027                 if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
2028                     set_len <= len) {
2029                         mlog(0, "increase refcount rec, start %llu, len %u, "
2030                              "count %u\n", (unsigned long long)cpos, set_len,
2031                              le32_to_cpu(rec.r_refcount));
2032                         ret = ocfs2_change_refcount_rec(handle, ci,
2033                                                         ref_leaf_bh, index,
2034                                                         merge, 1);
2035                         if (ret) {
2036                                 mlog_errno(ret);
2037                                 goto out;
2038                         }
2039                 } else if (!rec.r_refcount) {
2040                         rec.r_refcount = cpu_to_le32(1);
2041
2042                         mlog(0, "insert refcount rec, start %llu, len %u\n",
2043                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2044                              set_len);
2045                         ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
2046                                                         ref_leaf_bh,
2047                                                         &rec, index,
2048                                                         merge, meta_ac);
2049                         if (ret) {
2050                                 mlog_errno(ret);
2051                                 goto out;
2052                         }
2053                 } else  {
2054                         set_len = min((u64)(cpos + len),
2055                                       le64_to_cpu(rec.r_cpos) + set_len) - cpos;
2056                         rec.r_cpos = cpu_to_le64(cpos);
2057                         rec.r_clusters = cpu_to_le32(set_len);
2058                         le32_add_cpu(&rec.r_refcount, 1);
2059
2060                         mlog(0, "split refcount rec, start %llu, "
2061                              "len %u, count %u\n",
2062                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2063                              set_len, le32_to_cpu(rec.r_refcount));
2064                         ret = ocfs2_split_refcount_rec(handle, ci,
2065                                                        ref_root_bh, ref_leaf_bh,
2066                                                        &rec, index, merge,
2067                                                        meta_ac, dealloc);
2068                         if (ret) {
2069                                 mlog_errno(ret);
2070                                 goto out;
2071                         }
2072                 }
2073
2074                 cpos += set_len;
2075                 len -= set_len;
2076                 brelse(ref_leaf_bh);
2077                 ref_leaf_bh = NULL;
2078         }
2079
2080 out:
2081         brelse(ref_leaf_bh);
2082         return ret;
2083 }
2084
2085 static int ocfs2_remove_refcount_extent(handle_t *handle,
2086                                 struct ocfs2_caching_info *ci,
2087                                 struct buffer_head *ref_root_bh,
2088                                 struct buffer_head *ref_leaf_bh,
2089                                 struct ocfs2_alloc_context *meta_ac,
2090                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2091 {
2092         int ret;
2093         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2094         struct ocfs2_refcount_block *rb =
2095                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2096         struct ocfs2_extent_tree et;
2097
2098         BUG_ON(rb->rf_records.rl_used);
2099
2100         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2101         ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
2102                                   1, meta_ac, dealloc);
2103         if (ret) {
2104                 mlog_errno(ret);
2105                 goto out;
2106         }
2107
2108         ocfs2_remove_from_cache(ci, ref_leaf_bh);
2109
2110         /*
2111          * add the freed block to the dealloc so that it will be freed
2112          * when we run dealloc.
2113          */
2114         ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
2115                                         le16_to_cpu(rb->rf_suballoc_slot),
2116                                         le64_to_cpu(rb->rf_blkno),
2117                                         le16_to_cpu(rb->rf_suballoc_bit));
2118         if (ret) {
2119                 mlog_errno(ret);
2120                 goto out;
2121         }
2122
2123         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
2124                                       OCFS2_JOURNAL_ACCESS_WRITE);
2125         if (ret) {
2126                 mlog_errno(ret);
2127                 goto out;
2128         }
2129
2130         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2131
2132         le32_add_cpu(&rb->rf_clusters, -1);
2133
2134         /*
2135          * check whether we need to restore the root refcount block if
2136          * there is no leaf extent block at atll.
2137          */
2138         if (!rb->rf_list.l_next_free_rec) {
2139                 BUG_ON(rb->rf_clusters);
2140
2141                 mlog(0, "reset refcount tree root %llu to be a record block.\n",
2142                      (unsigned long long)ref_root_bh->b_blocknr);
2143
2144                 rb->rf_flags = 0;
2145                 rb->rf_parent = 0;
2146                 rb->rf_cpos = 0;
2147                 memset(&rb->rf_records, 0, sb->s_blocksize -
2148                        offsetof(struct ocfs2_refcount_block, rf_records));
2149                 rb->rf_records.rl_count =
2150                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
2151         }
2152
2153         ocfs2_journal_dirty(handle, ref_root_bh);
2154
2155 out:
2156         return ret;
2157 }
2158
2159 int ocfs2_increase_refcount(handle_t *handle,
2160                             struct ocfs2_caching_info *ci,
2161                             struct buffer_head *ref_root_bh,
2162                             u64 cpos, u32 len,
2163                             struct ocfs2_alloc_context *meta_ac,
2164                             struct ocfs2_cached_dealloc_ctxt *dealloc)
2165 {
2166         return __ocfs2_increase_refcount(handle, ci, ref_root_bh,
2167                                          cpos, len, 1,
2168                                          meta_ac, dealloc);
2169 }
2170
2171 static int ocfs2_decrease_refcount_rec(handle_t *handle,
2172                                 struct ocfs2_caching_info *ci,
2173                                 struct buffer_head *ref_root_bh,
2174                                 struct buffer_head *ref_leaf_bh,
2175                                 int index, u64 cpos, unsigned int len,
2176                                 struct ocfs2_alloc_context *meta_ac,
2177                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2178 {
2179         int ret;
2180         struct ocfs2_refcount_block *rb =
2181                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2182         struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
2183
2184         BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
2185         BUG_ON(cpos + len >
2186                le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
2187
2188         if (cpos == le64_to_cpu(rec->r_cpos) &&
2189             len == le32_to_cpu(rec->r_clusters))
2190                 ret = ocfs2_change_refcount_rec(handle, ci,
2191                                                 ref_leaf_bh, index, 1, -1);
2192         else {
2193                 struct ocfs2_refcount_rec split = *rec;
2194                 split.r_cpos = cpu_to_le64(cpos);
2195                 split.r_clusters = cpu_to_le32(len);
2196
2197                 le32_add_cpu(&split.r_refcount, -1);
2198
2199                 mlog(0, "split refcount rec, start %llu, "
2200                      "len %u, count %u, original start %llu, len %u\n",
2201                      (unsigned long long)le64_to_cpu(split.r_cpos),
2202                      len, le32_to_cpu(split.r_refcount),
2203                      (unsigned long long)le64_to_cpu(rec->r_cpos),
2204                      le32_to_cpu(rec->r_clusters));
2205                 ret = ocfs2_split_refcount_rec(handle, ci,
2206                                                ref_root_bh, ref_leaf_bh,
2207                                                &split, index, 1,
2208                                                meta_ac, dealloc);
2209         }
2210
2211         if (ret) {
2212                 mlog_errno(ret);
2213                 goto out;
2214         }
2215
2216         /* Remove the leaf refcount block if it contains no refcount record. */
2217         if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
2218                 ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
2219                                                    ref_leaf_bh, meta_ac,
2220                                                    dealloc);
2221                 if (ret)
2222                         mlog_errno(ret);
2223         }
2224
2225 out:
2226         return ret;
2227 }
2228
2229 static int __ocfs2_decrease_refcount(handle_t *handle,
2230                                      struct ocfs2_caching_info *ci,
2231                                      struct buffer_head *ref_root_bh,
2232                                      u64 cpos, u32 len,
2233                                      struct ocfs2_alloc_context *meta_ac,
2234                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2235                                      int delete)
2236 {
2237         int ret = 0, index = 0;
2238         struct ocfs2_refcount_rec rec;
2239         unsigned int r_count = 0, r_len;
2240         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2241         struct buffer_head *ref_leaf_bh = NULL;
2242
2243         mlog(0, "Tree owner %llu, decrease refcount start %llu, "
2244              "len %u, delete %u\n",
2245              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2246              (unsigned long long)cpos, len, delete);
2247
2248         while (len) {
2249                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2250                                              cpos, len, &rec, &index,
2251                                              &ref_leaf_bh);
2252                 if (ret) {
2253                         mlog_errno(ret);
2254                         goto out;
2255                 }
2256
2257                 r_count = le32_to_cpu(rec.r_refcount);
2258                 BUG_ON(r_count == 0);
2259                 if (!delete)
2260                         BUG_ON(r_count > 1);
2261
2262                 r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
2263                               le32_to_cpu(rec.r_clusters)) - cpos;
2264
2265                 ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
2266                                                   ref_leaf_bh, index,
2267                                                   cpos, r_len,
2268                                                   meta_ac, dealloc);
2269                 if (ret) {
2270                         mlog_errno(ret);
2271                         goto out;
2272                 }
2273
2274                 if (le32_to_cpu(rec.r_refcount) == 1 && delete) {
2275                         ret = ocfs2_cache_cluster_dealloc(dealloc,
2276                                           ocfs2_clusters_to_blocks(sb, cpos),
2277                                                           r_len);
2278                         if (ret) {
2279                                 mlog_errno(ret);
2280                                 goto out;
2281                         }
2282                 }
2283
2284                 cpos += r_len;
2285                 len -= r_len;
2286                 brelse(ref_leaf_bh);
2287                 ref_leaf_bh = NULL;
2288         }
2289
2290 out:
2291         brelse(ref_leaf_bh);
2292         return ret;
2293 }
2294
2295 /* Caller must hold refcount tree lock. */
2296 int ocfs2_decrease_refcount(struct inode *inode,
2297                             handle_t *handle, u32 cpos, u32 len,
2298                             struct ocfs2_alloc_context *meta_ac,
2299                             struct ocfs2_cached_dealloc_ctxt *dealloc,
2300                             int delete)
2301 {
2302         int ret;
2303         u64 ref_blkno;
2304         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2305         struct buffer_head *ref_root_bh = NULL;
2306         struct ocfs2_refcount_tree *tree;
2307
2308         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2309
2310         ret = ocfs2_get_refcount_block(inode, &ref_blkno);
2311         if (ret) {
2312                 mlog_errno(ret);
2313                 goto out;
2314         }
2315
2316         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
2317         if (ret) {
2318                 mlog_errno(ret);
2319                 goto out;
2320         }
2321
2322         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
2323                                         &ref_root_bh);
2324         if (ret) {
2325                 mlog_errno(ret);
2326                 goto out;
2327         }
2328
2329         ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
2330                                         cpos, len, meta_ac, dealloc, delete);
2331         if (ret)
2332                 mlog_errno(ret);
2333 out:
2334         brelse(ref_root_bh);
2335         return ret;
2336 }
2337
2338 /*
2339  * Mark the already-existing extent at cpos as refcounted for len clusters.
2340  * This adds the refcount extent flag.
2341  *
2342  * If the existing extent is larger than the request, initiate a
2343  * split. An attempt will be made at merging with adjacent extents.
2344  *
2345  * The caller is responsible for passing down meta_ac if we'll need it.
2346  */
2347 static int ocfs2_mark_extent_refcounted(struct inode *inode,
2348                                 struct ocfs2_extent_tree *et,
2349                                 handle_t *handle, u32 cpos,
2350                                 u32 len, u32 phys,
2351                                 struct ocfs2_alloc_context *meta_ac,
2352                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2353 {
2354         int ret;
2355
2356         mlog(0, "Inode %lu refcount tree cpos %u, len %u, phys cluster %u\n",
2357              inode->i_ino, cpos, len, phys);
2358
2359         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2360                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2361                             "tree, but the feature bit is not set in the "
2362                             "super block.", inode->i_ino);
2363                 ret = -EROFS;
2364                 goto out;
2365         }
2366
2367         ret = ocfs2_change_extent_flag(handle, et, cpos,
2368                                        len, phys, meta_ac, dealloc,
2369                                        OCFS2_EXT_REFCOUNTED, 0);
2370         if (ret)
2371                 mlog_errno(ret);
2372
2373 out:
2374         return ret;
2375 }
2376
2377 /*
2378  * Given some contiguous physical clusters, calculate what we need
2379  * for modifying their refcount.
2380  */
2381 static int ocfs2_calc_refcount_meta_credits(struct super_block *sb,
2382                                             struct ocfs2_caching_info *ci,
2383                                             struct buffer_head *ref_root_bh,
2384                                             u64 start_cpos,
2385                                             u32 clusters,
2386                                             int *meta_add,
2387                                             int *credits)
2388 {
2389         int ret = 0, index, ref_blocks = 0, recs_add = 0;
2390         u64 cpos = start_cpos;
2391         struct ocfs2_refcount_block *rb;
2392         struct ocfs2_refcount_rec rec;
2393         struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL;
2394         u32 len;
2395
2396         mlog(0, "start_cpos %llu, clusters %u\n",
2397              (unsigned long long)start_cpos, clusters);
2398         while (clusters) {
2399                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2400                                              cpos, clusters, &rec,
2401                                              &index, &ref_leaf_bh);
2402                 if (ret) {
2403                         mlog_errno(ret);
2404                         goto out;
2405                 }
2406
2407                 if (ref_leaf_bh != prev_bh) {
2408                         /*
2409                          * Now we encounter a new leaf block, so calculate
2410                          * whether we need to extend the old leaf.
2411                          */
2412                         if (prev_bh) {
2413                                 rb = (struct ocfs2_refcount_block *)
2414                                                         prev_bh->b_data;
2415
2416                                 if (le64_to_cpu(rb->rf_records.rl_used) +
2417                                     recs_add >
2418                                     le16_to_cpu(rb->rf_records.rl_count))
2419                                         ref_blocks++;
2420                         }
2421
2422                         recs_add = 0;
2423                         *credits += 1;
2424                         brelse(prev_bh);
2425                         prev_bh = ref_leaf_bh;
2426                         get_bh(prev_bh);
2427                 }
2428
2429                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2430
2431                 mlog(0, "recs_add %d,cpos %llu, clusters %u, rec->r_cpos %llu,"
2432                      "rec->r_clusters %u, rec->r_refcount %u, index %d\n",
2433                      recs_add, (unsigned long long)cpos, clusters,
2434                      (unsigned long long)le64_to_cpu(rec.r_cpos),
2435                      le32_to_cpu(rec.r_clusters),
2436                      le32_to_cpu(rec.r_refcount), index);
2437
2438                 len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) +
2439                           le32_to_cpu(rec.r_clusters)) - cpos;
2440                 /*
2441                  * If the refcount rec already exist, cool. We just need
2442                  * to check whether there is a split. Otherwise we just need
2443                  * to increase the refcount.
2444                  * If we will insert one, increases recs_add.
2445                  *
2446                  * We record all the records which will be inserted to the
2447                  * same refcount block, so that we can tell exactly whether
2448                  * we need a new refcount block or not.
2449                  */
2450                 if (rec.r_refcount) {
2451                         /* Check whether we need a split at the beginning. */
2452                         if (cpos == start_cpos &&
2453                             cpos != le64_to_cpu(rec.r_cpos))
2454                                 recs_add++;
2455
2456                         /* Check whether we need a split in the end. */
2457                         if (cpos + clusters < le64_to_cpu(rec.r_cpos) +
2458                             le32_to_cpu(rec.r_clusters))
2459                                 recs_add++;
2460                 } else
2461                         recs_add++;
2462
2463                 brelse(ref_leaf_bh);
2464                 ref_leaf_bh = NULL;
2465                 clusters -= len;
2466                 cpos += len;
2467         }
2468
2469         if (prev_bh) {
2470                 rb = (struct ocfs2_refcount_block *)prev_bh->b_data;
2471
2472                 if (le64_to_cpu(rb->rf_records.rl_used) + recs_add >
2473                     le16_to_cpu(rb->rf_records.rl_count))
2474                         ref_blocks++;
2475
2476                 *credits += 1;
2477         }
2478
2479         if (!ref_blocks)
2480                 goto out;
2481
2482         mlog(0, "we need ref_blocks %d\n", ref_blocks);
2483         *meta_add += ref_blocks;
2484         *credits += ref_blocks;
2485
2486         /*
2487          * So we may need ref_blocks to insert into the tree.
2488          * That also means we need to change the b-tree and add that number
2489          * of records since we never merge them.
2490          * We need one more block for expansion since the new created leaf
2491          * block is also full and needs split.
2492          */
2493         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2494         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) {
2495                 struct ocfs2_extent_tree et;
2496
2497                 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2498                 *meta_add += ocfs2_extend_meta_needed(et.et_root_el);
2499                 *credits += ocfs2_calc_extend_credits(sb,
2500                                                       et.et_root_el,
2501                                                       ref_blocks);
2502         } else {
2503                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
2504                 *meta_add += 1;
2505         }
2506
2507 out:
2508         brelse(ref_leaf_bh);
2509         brelse(prev_bh);
2510         return ret;
2511 }
2512
2513 /*
2514  * For refcount tree, we will decrease some contiguous clusters
2515  * refcount count, so just go through it to see how many blocks
2516  * we gonna touch and whether we need to create new blocks.
2517  *
2518  * Normally the refcount blocks store these refcount should be
2519  * contiguous also, so that we can get the number easily.
2520  * As for meta_ac, we will at most add split 2 refcount record and
2521  * 2 more refcount block, so just check it in a rough way.
2522  *
2523  * Caller must hold refcount tree lock.
2524  */
2525 int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
2526                                           struct buffer_head *di_bh,
2527                                           u64 phys_blkno,
2528                                           u32 clusters,
2529                                           int *credits,
2530                                           struct ocfs2_alloc_context **meta_ac)
2531 {
2532         int ret, ref_blocks = 0;
2533         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2534         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2535         struct buffer_head *ref_root_bh = NULL;
2536         struct ocfs2_refcount_tree *tree;
2537         u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno);
2538
2539         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2540                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2541                             "tree, but the feature bit is not set in the "
2542                             "super block.", inode->i_ino);
2543                 ret = -EROFS;
2544                 goto out;
2545         }
2546
2547         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2548
2549         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb),
2550                                       le64_to_cpu(di->i_refcount_loc), &tree);
2551         if (ret) {
2552                 mlog_errno(ret);
2553                 goto out;
2554         }
2555
2556         ret = ocfs2_read_refcount_block(&tree->rf_ci,
2557                                         le64_to_cpu(di->i_refcount_loc),
2558                                         &ref_root_bh);
2559         if (ret) {
2560                 mlog_errno(ret);
2561                 goto out;
2562         }
2563
2564         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
2565                                                &tree->rf_ci,
2566                                                ref_root_bh,
2567                                                start_cpos, clusters,
2568                                                &ref_blocks, credits);
2569         if (ret) {
2570                 mlog_errno(ret);
2571                 goto out;
2572         }
2573
2574         mlog(0, "reserve new metadata %d, credits = %d\n",
2575              ref_blocks, *credits);
2576
2577         if (ref_blocks) {
2578                 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
2579                                                         ref_blocks, meta_ac);
2580                 if (ret)
2581                         mlog_errno(ret);
2582         }
2583
2584 out:
2585         brelse(ref_root_bh);
2586         return ret;
2587 }
2588
2589 #define MAX_CONTIG_BYTES        1048576
2590
2591 static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb)
2592 {
2593         return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES);
2594 }
2595
2596 static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb)
2597 {
2598         return ~(ocfs2_cow_contig_clusters(sb) - 1);
2599 }
2600
2601 /*
2602  * Given an extent that starts at 'start' and an I/O that starts at 'cpos',
2603  * find an offset (start + (n * contig_clusters)) that is closest to cpos
2604  * while still being less than or equal to it.
2605  *
2606  * The goal is to break the extent at a multiple of contig_clusters.
2607  */
2608 static inline unsigned int ocfs2_cow_align_start(struct super_block *sb,
2609                                                  unsigned int start,
2610                                                  unsigned int cpos)
2611 {
2612         BUG_ON(start > cpos);
2613
2614         return start + ((cpos - start) & ocfs2_cow_contig_mask(sb));
2615 }
2616
2617 /*
2618  * Given a cluster count of len, pad it out so that it is a multiple
2619  * of contig_clusters.
2620  */
2621 static inline unsigned int ocfs2_cow_align_length(struct super_block *sb,
2622                                                   unsigned int len)
2623 {
2624         unsigned int padded =
2625                 (len + (ocfs2_cow_contig_clusters(sb) - 1)) &
2626                 ocfs2_cow_contig_mask(sb);
2627
2628         /* Did we wrap? */
2629         if (padded < len)
2630                 padded = UINT_MAX;
2631
2632         return padded;
2633 }
2634
2635 /*
2636  * Calculate out the start and number of virtual clusters we need to to CoW.
2637  *
2638  * cpos is vitual start cluster position we want to do CoW in a
2639  * file and write_len is the cluster length.
2640  * max_cpos is the place where we want to stop CoW intentionally.
2641  *
2642  * Normal we will start CoW from the beginning of extent record cotaining cpos.
2643  * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we
2644  * get good I/O from the resulting extent tree.
2645  */
2646 static int ocfs2_refcount_cal_cow_clusters(struct inode *inode,
2647                                            struct ocfs2_extent_list *el,
2648                                            u32 cpos,
2649                                            u32 write_len,
2650                                            u32 max_cpos,
2651                                            u32 *cow_start,
2652                                            u32 *cow_len)
2653 {
2654         int ret = 0;
2655         int tree_height = le16_to_cpu(el->l_tree_depth), i;
2656         struct buffer_head *eb_bh = NULL;
2657         struct ocfs2_extent_block *eb = NULL;
2658         struct ocfs2_extent_rec *rec;
2659         unsigned int want_clusters, rec_end = 0;
2660         int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb);
2661         int leaf_clusters;
2662
2663         BUG_ON(cpos + write_len > max_cpos);
2664
2665         if (tree_height > 0) {
2666                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh);
2667                 if (ret) {
2668                         mlog_errno(ret);
2669                         goto out;
2670                 }
2671
2672                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2673                 el = &eb->h_list;
2674
2675                 if (el->l_tree_depth) {
2676                         ocfs2_error(inode->i_sb,
2677                                     "Inode %lu has non zero tree depth in "
2678                                     "leaf block %llu\n", inode->i_ino,
2679                                     (unsigned long long)eb_bh->b_blocknr);
2680                         ret = -EROFS;
2681                         goto out;
2682                 }
2683         }
2684
2685         *cow_len = 0;
2686         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
2687                 rec = &el->l_recs[i];
2688
2689                 if (ocfs2_is_empty_extent(rec)) {
2690                         mlog_bug_on_msg(i != 0, "Inode %lu has empty record in "
2691                                         "index %d\n", inode->i_ino, i);
2692                         continue;
2693                 }
2694
2695                 if (le32_to_cpu(rec->e_cpos) +
2696                     le16_to_cpu(rec->e_leaf_clusters) <= cpos)
2697                         continue;
2698
2699                 if (*cow_len == 0) {
2700                         /*
2701                          * We should find a refcounted record in the
2702                          * first pass.
2703                          */
2704                         BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED));
2705                         *cow_start = le32_to_cpu(rec->e_cpos);
2706                 }
2707
2708                 /*
2709                  * If we encounter a hole, a non-refcounted record or
2710                  * pass the max_cpos, stop the search.
2711                  */
2712                 if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) ||
2713                     (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) ||
2714                     (max_cpos <= le32_to_cpu(rec->e_cpos)))
2715                         break;
2716
2717                 leaf_clusters = le16_to_cpu(rec->e_leaf_clusters);
2718                 rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters;
2719                 if (rec_end > max_cpos) {
2720                         rec_end = max_cpos;
2721                         leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos);
2722                 }
2723
2724                 /*
2725                  * How many clusters do we actually need from
2726                  * this extent?  First we see how many we actually
2727                  * need to complete the write.  If that's smaller
2728                  * than contig_clusters, we try for contig_clusters.
2729                  */
2730                 if (!*cow_len)
2731                         want_clusters = write_len;
2732                 else
2733                         want_clusters = (cpos + write_len) -
2734                                 (*cow_start + *cow_len);
2735                 if (want_clusters < contig_clusters)
2736                         want_clusters = contig_clusters;
2737
2738                 /*
2739                  * If the write does not cover the whole extent, we
2740                  * need to calculate how we're going to split the extent.
2741                  * We try to do it on contig_clusters boundaries.
2742                  *
2743                  * Any extent smaller than contig_clusters will be
2744                  * CoWed in its entirety.
2745                  */
2746                 if (leaf_clusters <= contig_clusters)
2747                         *cow_len += leaf_clusters;
2748                 else if (*cow_len || (*cow_start == cpos)) {
2749                         /*
2750                          * This extent needs to be CoW'd from its
2751                          * beginning, so all we have to do is compute
2752                          * how many clusters to grab.  We align
2753                          * want_clusters to the edge of contig_clusters
2754                          * to get better I/O.
2755                          */
2756                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2757                                                                want_clusters);
2758
2759                         if (leaf_clusters < want_clusters)
2760                                 *cow_len += leaf_clusters;
2761                         else
2762                                 *cow_len += want_clusters;
2763                 } else if ((*cow_start + contig_clusters) >=
2764                            (cpos + write_len)) {
2765                         /*
2766                          * Breaking off contig_clusters at the front
2767                          * of the extent will cover our write.  That's
2768                          * easy.
2769                          */
2770                         *cow_len = contig_clusters;
2771                 } else if ((rec_end - cpos) <= contig_clusters) {
2772                         /*
2773                          * Breaking off contig_clusters at the tail of
2774                          * this extent will cover cpos.
2775                          */
2776                         *cow_start = rec_end - contig_clusters;
2777                         *cow_len = contig_clusters;
2778                 } else if ((rec_end - cpos) <= want_clusters) {
2779                         /*
2780                          * While we can't fit the entire write in this
2781                          * extent, we know that the write goes from cpos
2782                          * to the end of the extent.  Break that off.
2783                          * We try to break it at some multiple of
2784                          * contig_clusters from the front of the extent.
2785                          * Failing that (ie, cpos is within
2786                          * contig_clusters of the front), we'll CoW the
2787                          * entire extent.
2788                          */
2789                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2790                                                            *cow_start, cpos);
2791                         *cow_len = rec_end - *cow_start;
2792                 } else {
2793                         /*
2794                          * Ok, the entire write lives in the middle of
2795                          * this extent.  Let's try to slice the extent up
2796                          * nicely.  Optimally, our CoW region starts at
2797                          * m*contig_clusters from the beginning of the
2798                          * extent and goes for n*contig_clusters,
2799                          * covering the entire write.
2800                          */
2801                         *cow_start = ocfs2_cow_align_start(inode->i_sb,
2802                                                            *cow_start, cpos);
2803
2804                         want_clusters = (cpos + write_len) - *cow_start;
2805                         want_clusters = ocfs2_cow_align_length(inode->i_sb,
2806                                                                want_clusters);
2807                         if (*cow_start + want_clusters <= rec_end)
2808                                 *cow_len = want_clusters;
2809                         else
2810                                 *cow_len = rec_end - *cow_start;
2811                 }
2812
2813                 /* Have we covered our entire write yet? */
2814                 if ((*cow_start + *cow_len) >= (cpos + write_len))
2815                         break;
2816
2817                 /*
2818                  * If we reach the end of the extent block and don't get enough
2819                  * clusters, continue with the next extent block if possible.
2820                  */
2821                 if (i + 1 == le16_to_cpu(el->l_next_free_rec) &&
2822                     eb && eb->h_next_leaf_blk) {
2823                         brelse(eb_bh);
2824                         eb_bh = NULL;
2825
2826                         ret = ocfs2_read_extent_block(INODE_CACHE(inode),
2827                                                le64_to_cpu(eb->h_next_leaf_blk),
2828                                                &eb_bh);
2829                         if (ret) {
2830                                 mlog_errno(ret);
2831                                 goto out;
2832                         }
2833
2834                         eb = (struct ocfs2_extent_block *) eb_bh->b_data;
2835                         el = &eb->h_list;
2836                         i = -1;
2837                 }
2838         }
2839
2840 out:
2841         brelse(eb_bh);
2842         return ret;
2843 }
2844
2845 /*
2846  * Prepare meta_ac, data_ac and calculate credits when we want to add some
2847  * num_clusters in data_tree "et" and change the refcount for the old
2848  * clusters(starting form p_cluster) in the refcount tree.
2849  *
2850  * Note:
2851  * 1. since we may split the old tree, so we at most will need num_clusters + 2
2852  *    more new leaf records.
2853  * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so
2854  *    just give data_ac = NULL.
2855  */
2856 static int ocfs2_lock_refcount_allocators(struct super_block *sb,
2857                                         u32 p_cluster, u32 num_clusters,
2858                                         struct ocfs2_extent_tree *et,
2859                                         struct ocfs2_caching_info *ref_ci,
2860                                         struct buffer_head *ref_root_bh,
2861                                         struct ocfs2_alloc_context **meta_ac,
2862                                         struct ocfs2_alloc_context **data_ac,
2863                                         int *credits)
2864 {
2865         int ret = 0, meta_add = 0;
2866         int num_free_extents = ocfs2_num_free_extents(OCFS2_SB(sb), et);
2867
2868         if (num_free_extents < 0) {
2869                 ret = num_free_extents;
2870                 mlog_errno(ret);
2871                 goto out;
2872         }
2873
2874         if (num_free_extents < num_clusters + 2)
2875                 meta_add =
2876                         ocfs2_extend_meta_needed(et->et_root_el);
2877
2878         *credits += ocfs2_calc_extend_credits(sb, et->et_root_el,
2879                                               num_clusters + 2);
2880
2881         ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh,
2882                                                p_cluster, num_clusters,
2883                                                &meta_add, credits);
2884         if (ret) {
2885                 mlog_errno(ret);
2886                 goto out;
2887         }
2888
2889         mlog(0, "reserve new metadata %d, clusters %u, credits = %d\n",
2890              meta_add, num_clusters, *credits);
2891         ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add,
2892                                                 meta_ac);
2893         if (ret) {
2894                 mlog_errno(ret);
2895                 goto out;
2896         }
2897
2898         if (data_ac) {
2899                 ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters,
2900                                              data_ac);
2901                 if (ret)
2902                         mlog_errno(ret);
2903         }
2904
2905 out:
2906         if (ret) {
2907                 if (*meta_ac) {
2908                         ocfs2_free_alloc_context(*meta_ac);
2909                         *meta_ac = NULL;
2910                 }
2911         }
2912
2913         return ret;
2914 }
2915
2916 static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh)
2917 {
2918         BUG_ON(buffer_dirty(bh));
2919
2920         clear_buffer_mapped(bh);
2921
2922         return 0;
2923 }
2924
2925 static int ocfs2_duplicate_clusters_by_page(handle_t *handle,
2926                                             struct ocfs2_cow_context *context,
2927                                             u32 cpos, u32 old_cluster,
2928                                             u32 new_cluster, u32 new_len)
2929 {
2930         int ret = 0, partial;
2931         struct ocfs2_caching_info *ci = context->data_et.et_ci;
2932         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2933         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
2934         struct page *page;
2935         pgoff_t page_index;
2936         unsigned int from, to;
2937         loff_t offset, end, map_end;
2938         struct address_space *mapping = context->inode->i_mapping;
2939
2940         mlog(0, "old_cluster %u, new %u, len %u at offset %u\n", old_cluster,
2941              new_cluster, new_len, cpos);
2942
2943         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
2944         end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
2945
2946         while (offset < end) {
2947                 page_index = offset >> PAGE_CACHE_SHIFT;
2948                 map_end = ((loff_t)page_index + 1) << PAGE_CACHE_SHIFT;
2949                 if (map_end > end)
2950                         map_end = end;
2951
2952                 /* from, to is the offset within the page. */
2953                 from = offset & (PAGE_CACHE_SIZE - 1);
2954                 to = PAGE_CACHE_SIZE;
2955                 if (map_end & (PAGE_CACHE_SIZE - 1))
2956                         to = map_end & (PAGE_CACHE_SIZE - 1);
2957
2958                 page = grab_cache_page(mapping, page_index);
2959
2960                 /*
2961                  * In case PAGE_CACHE_SIZE <= CLUSTER_SIZE, This page
2962                  * can't be dirtied before we CoW it out.
2963                  */
2964                 if (PAGE_CACHE_SIZE <= OCFS2_SB(sb)->s_clustersize)
2965                         BUG_ON(PageDirty(page));
2966
2967                 if (!PageUptodate(page)) {
2968                         ret = block_read_full_page(page, ocfs2_get_block);
2969                         if (ret) {
2970                                 mlog_errno(ret);
2971                                 goto unlock;
2972                         }
2973                         lock_page(page);
2974                 }
2975
2976                 if (page_has_buffers(page)) {
2977                         ret = walk_page_buffers(handle, page_buffers(page),
2978                                                 from, to, &partial,
2979                                                 ocfs2_clear_cow_buffer);
2980                         if (ret) {
2981                                 mlog_errno(ret);
2982                                 goto unlock;
2983                         }
2984                 }
2985
2986                 ocfs2_map_and_dirty_page(context->inode,
2987                                          handle, from, to,
2988                                          page, 0, &new_block);
2989                 mark_page_accessed(page);
2990 unlock:
2991                 unlock_page(page);
2992                 page_cache_release(page);
2993                 page = NULL;
2994                 offset = map_end;
2995                 if (ret)
2996                         break;
2997         }
2998
2999         return ret;
3000 }
3001
3002 static int ocfs2_duplicate_clusters_by_jbd(handle_t *handle,
3003                                            struct ocfs2_cow_context *context,
3004                                            u32 cpos, u32 old_cluster,
3005                                            u32 new_cluster, u32 new_len)
3006 {
3007         int ret = 0;
3008         struct super_block *sb = context->inode->i_sb;
3009         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3010         int i, blocks = ocfs2_clusters_to_blocks(sb, new_len);
3011         u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster);
3012         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
3013         struct ocfs2_super *osb = OCFS2_SB(sb);
3014         struct buffer_head *old_bh = NULL;
3015         struct buffer_head *new_bh = NULL;
3016
3017         mlog(0, "old_cluster %u, new %u, len %u\n", old_cluster,
3018              new_cluster, new_len);
3019
3020         for (i = 0; i < blocks; i++, old_block++, new_block++) {
3021                 new_bh = sb_getblk(osb->sb, new_block);
3022                 if (new_bh == NULL) {
3023                         ret = -EIO;
3024                         mlog_errno(ret);
3025                         break;
3026                 }
3027
3028                 ocfs2_set_new_buffer_uptodate(ci, new_bh);
3029
3030                 ret = ocfs2_read_block(ci, old_block, &old_bh, NULL);
3031                 if (ret) {
3032                         mlog_errno(ret);
3033                         break;
3034                 }
3035
3036                 ret = ocfs2_journal_access(handle, ci, new_bh,
3037                                            OCFS2_JOURNAL_ACCESS_CREATE);
3038                 if (ret) {
3039                         mlog_errno(ret);
3040                         break;
3041                 }
3042
3043                 memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize);
3044                 ret = ocfs2_journal_dirty(handle, new_bh);
3045                 if (ret) {
3046                         mlog_errno(ret);
3047                         break;
3048                 }
3049
3050                 brelse(new_bh);
3051                 brelse(old_bh);
3052                 new_bh = NULL;
3053                 old_bh = NULL;
3054         }
3055
3056         brelse(new_bh);
3057         brelse(old_bh);
3058         return ret;
3059 }
3060
3061 static int ocfs2_clear_ext_refcount(handle_t *handle,
3062                                     struct ocfs2_extent_tree *et,
3063                                     u32 cpos, u32 p_cluster, u32 len,
3064                                     unsigned int ext_flags,
3065                                     struct ocfs2_alloc_context *meta_ac,
3066                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
3067 {
3068         int ret, index;
3069         struct ocfs2_extent_rec replace_rec;
3070         struct ocfs2_path *path = NULL;
3071         struct ocfs2_extent_list *el;
3072         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
3073         u64 ino = ocfs2_metadata_cache_owner(et->et_ci);
3074
3075         mlog(0, "inode %llu cpos %u, len %u, p_cluster %u, ext_flags %u\n",
3076              (unsigned long long)ino, cpos, len, p_cluster, ext_flags);
3077
3078         memset(&replace_rec, 0, sizeof(replace_rec));
3079         replace_rec.e_cpos = cpu_to_le32(cpos);
3080         replace_rec.e_leaf_clusters = cpu_to_le16(len);
3081         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb,
3082                                                                    p_cluster));
3083         replace_rec.e_flags = ext_flags;
3084         replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED;
3085
3086         path = ocfs2_new_path_from_et(et);
3087         if (!path) {
3088                 ret = -ENOMEM;
3089                 mlog_errno(ret);
3090                 goto out;
3091         }
3092
3093         ret = ocfs2_find_path(et->et_ci, path, cpos);
3094         if (ret) {
3095                 mlog_errno(ret);
3096                 goto out;
3097         }
3098
3099         el = path_leaf_el(path);
3100
3101         index = ocfs2_search_extent_list(el, cpos);
3102         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
3103                 ocfs2_error(sb,
3104                             "Inode %llu has an extent at cpos %u which can no "
3105                             "longer be found.\n",
3106                             (unsigned long long)ino, cpos);
3107                 ret = -EROFS;
3108                 goto out;
3109         }
3110
3111         ret = ocfs2_split_extent(handle, et, path, index,
3112                                  &replace_rec, meta_ac, dealloc);
3113         if (ret)
3114                 mlog_errno(ret);
3115
3116 out:
3117         ocfs2_free_path(path);
3118         return ret;
3119 }
3120
3121 static int ocfs2_replace_clusters(handle_t *handle,
3122                                   struct ocfs2_cow_context *context,
3123                                   u32 cpos, u32 old,
3124                                   u32 new, u32 len,
3125                                   unsigned int ext_flags)
3126 {
3127         int ret;
3128         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3129         u64 ino = ocfs2_metadata_cache_owner(ci);
3130
3131         mlog(0, "inode %llu, cpos %u, old %u, new %u, len %u, ext_flags %u\n",
3132              (unsigned long long)ino, cpos, old, new, len, ext_flags);
3133
3134         /*If the old clusters is unwritten, no need to duplicate. */
3135         if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) {
3136                 ret = context->cow_duplicate_clusters(handle, context, cpos,
3137                                                       old, new, len);
3138                 if (ret) {
3139                         mlog_errno(ret);
3140                         goto out;
3141                 }
3142         }
3143
3144         ret = ocfs2_clear_ext_refcount(handle, &context->data_et,
3145                                        cpos, new, len, ext_flags,
3146                                        context->meta_ac, &context->dealloc);
3147         if (ret)
3148                 mlog_errno(ret);
3149 out:
3150         return ret;
3151 }
3152
3153 static int ocfs2_cow_sync_writeback(struct super_block *sb,
3154                                     struct ocfs2_cow_context *context,
3155                                     u32 cpos, u32 num_clusters)
3156 {
3157         int ret = 0;
3158         loff_t offset, end, map_end;
3159         pgoff_t page_index;
3160         struct page *page;
3161
3162         if (ocfs2_should_order_data(context->inode))
3163                 return 0;
3164
3165         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
3166         end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits);
3167
3168         ret = filemap_fdatawrite_range(context->inode->i_mapping,
3169                                        offset, end - 1);
3170         if (ret < 0) {
3171                 mlog_errno(ret);
3172                 return ret;
3173         }
3174
3175         while (offset < end) {
3176                 page_index = offset >> PAGE_CACHE_SHIFT;
3177                 map_end = ((loff_t)page_index + 1) << PAGE_CACHE_SHIFT;
3178                 if (map_end > end)
3179                         map_end = end;
3180
3181                 page = grab_cache_page(context->inode->i_mapping, page_index);
3182                 BUG_ON(!page);
3183
3184                 wait_on_page_writeback(page);
3185                 if (PageError(page)) {
3186                         ret = -EIO;
3187                         mlog_errno(ret);
3188                 } else
3189                         mark_page_accessed(page);
3190
3191                 unlock_page(page);
3192                 page_cache_release(page);
3193                 page = NULL;
3194                 offset = map_end;
3195                 if (ret)
3196                         break;
3197         }
3198
3199         return ret;
3200 }
3201
3202 static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context,
3203                                  u32 v_cluster, u32 *p_cluster,
3204                                  u32 *num_clusters,
3205                                  unsigned int *extent_flags)
3206 {
3207         return ocfs2_get_clusters(context->inode, v_cluster, p_cluster,
3208                                   num_clusters, extent_flags);
3209 }
3210
3211 static int ocfs2_make_clusters_writable(struct super_block *sb,
3212                                         struct ocfs2_cow_context *context,
3213                                         u32 cpos, u32 p_cluster,
3214                                         u32 num_clusters, unsigned int e_flags)
3215 {
3216         int ret, delete, index, credits =  0;
3217         u32 new_bit, new_len;
3218         unsigned int set_len;
3219         struct ocfs2_super *osb = OCFS2_SB(sb);
3220         handle_t *handle;
3221         struct buffer_head *ref_leaf_bh = NULL;
3222         struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci;
3223         struct ocfs2_refcount_rec rec;
3224
3225         mlog(0, "cpos %u, p_cluster %u, num_clusters %u, e_flags %u\n",
3226              cpos, p_cluster, num_clusters, e_flags);
3227
3228         ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters,
3229                                              &context->data_et,
3230                                              ref_ci,
3231                                              context->ref_root_bh,
3232                                              &context->meta_ac,
3233                                              &context->data_ac, &credits);
3234         if (ret) {
3235                 mlog_errno(ret);
3236                 return ret;
3237         }
3238
3239         if (context->post_refcount)
3240                 credits += context->post_refcount->credits;
3241
3242         credits += context->extra_credits;
3243         handle = ocfs2_start_trans(osb, credits);
3244         if (IS_ERR(handle)) {
3245                 ret = PTR_ERR(handle);
3246                 mlog_errno(ret);
3247                 goto out;
3248         }
3249
3250         while (num_clusters) {
3251                 ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh,
3252                                              p_cluster, num_clusters,
3253                                              &rec, &index, &ref_leaf_bh);
3254                 if (ret) {
3255                         mlog_errno(ret);
3256                         goto out_commit;
3257                 }
3258
3259                 BUG_ON(!rec.r_refcount);
3260                 set_len = min((u64)p_cluster + num_clusters,
3261                               le64_to_cpu(rec.r_cpos) +
3262                               le32_to_cpu(rec.r_clusters)) - p_cluster;
3263
3264                 /*
3265                  * There are many different situation here.
3266                  * 1. If refcount == 1, remove the flag and don't COW.
3267                  * 2. If refcount > 1, allocate clusters.
3268                  *    Here we may not allocate r_len once at a time, so continue
3269                  *    until we reach num_clusters.
3270                  */
3271                 if (le32_to_cpu(rec.r_refcount) == 1) {
3272                         delete = 0;
3273                         ret = ocfs2_clear_ext_refcount(handle,
3274                                                        &context->data_et,
3275                                                        cpos, p_cluster,
3276                                                        set_len, e_flags,
3277                                                        context->meta_ac,
3278                                                        &context->dealloc);
3279                         if (ret) {
3280                                 mlog_errno(ret);
3281                                 goto out_commit;
3282                         }
3283                 } else {
3284                         delete = 1;
3285
3286                         ret = __ocfs2_claim_clusters(osb, handle,
3287                                                      context->data_ac,
3288                                                      1, set_len,
3289                                                      &new_bit, &new_len);
3290                         if (ret) {
3291                                 mlog_errno(ret);
3292                                 goto out_commit;
3293                         }
3294
3295                         ret = ocfs2_replace_clusters(handle, context,
3296                                                      cpos, p_cluster, new_bit,
3297                                                      new_len, e_flags);
3298                         if (ret) {
3299                                 mlog_errno(ret);
3300                                 goto out_commit;
3301                         }
3302                         set_len = new_len;
3303                 }
3304
3305                 ret = __ocfs2_decrease_refcount(handle, ref_ci,
3306                                                 context->ref_root_bh,
3307                                                 p_cluster, set_len,
3308                                                 context->meta_ac,
3309                                                 &context->dealloc, delete);
3310                 if (ret) {
3311                         mlog_errno(ret);
3312                         goto out_commit;
3313                 }
3314
3315                 cpos += set_len;
3316                 p_cluster += set_len;
3317                 num_clusters -= set_len;
3318                 brelse(ref_leaf_bh);
3319                 ref_leaf_bh = NULL;
3320         }
3321
3322         /* handle any post_cow action. */
3323         if (context->post_refcount && context->post_refcount->func) {
3324                 ret = context->post_refcount->func(context->inode, handle,
3325                                                 context->post_refcount->para);
3326                 if (ret) {
3327                         mlog_errno(ret);
3328                         goto out_commit;
3329                 }
3330         }
3331
3332         /*
3333          * Here we should write the new page out first if we are
3334          * in write-back mode.
3335          */
3336         if (context->get_clusters == ocfs2_di_get_clusters) {
3337                 ret = ocfs2_cow_sync_writeback(sb, context, cpos, num_clusters);
3338                 if (ret)
3339                         mlog_errno(ret);
3340         }
3341
3342 out_commit:
3343         ocfs2_commit_trans(osb, handle);
3344
3345 out:
3346         if (context->data_ac) {
3347                 ocfs2_free_alloc_context(context->data_ac);
3348                 context->data_ac = NULL;
3349         }
3350         if (context->meta_ac) {
3351                 ocfs2_free_alloc_context(context->meta_ac);
3352                 context->meta_ac = NULL;
3353         }
3354         brelse(ref_leaf_bh);
3355
3356         return ret;
3357 }
3358
3359 static int ocfs2_replace_cow(struct ocfs2_cow_context *context)
3360 {
3361         int ret = 0;
3362         struct inode *inode = context->inode;
3363         u32 cow_start = context->cow_start, cow_len = context->cow_len;
3364         u32 p_cluster, num_clusters;
3365         unsigned int ext_flags;
3366         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3367
3368         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
3369                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
3370                             "tree, but the feature bit is not set in the "
3371                             "super block.", inode->i_ino);
3372                 return -EROFS;
3373         }
3374
3375         ocfs2_init_dealloc_ctxt(&context->dealloc);
3376
3377         while (cow_len) {
3378                 ret = context->get_clusters(context, cow_start, &p_cluster,
3379                                             &num_clusters, &ext_flags);
3380                 if (ret) {
3381                         mlog_errno(ret);
3382                         break;
3383                 }
3384
3385                 BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED));
3386
3387                 if (cow_len < num_clusters)
3388                         num_clusters = cow_len;
3389
3390                 ret = ocfs2_make_clusters_writable(inode->i_sb, context,
3391                                                    cow_start, p_cluster,
3392                                                    num_clusters, ext_flags);
3393                 if (ret) {
3394                         mlog_errno(ret);
3395                         break;
3396                 }
3397
3398                 cow_len -= num_clusters;
3399                 cow_start += num_clusters;
3400         }
3401
3402         if (ocfs2_dealloc_has_cluster(&context->dealloc)) {
3403                 ocfs2_schedule_truncate_log_flush(osb, 1);
3404                 ocfs2_run_deallocs(osb, &context->dealloc);
3405         }
3406
3407         return ret;
3408 }
3409
3410 /*
3411  * Starting at cpos, try to CoW write_len clusters.  Don't CoW
3412  * past max_cpos.  This will stop when it runs into a hole or an
3413  * unrefcounted extent.
3414  */
3415 static int ocfs2_refcount_cow_hunk(struct inode *inode,
3416                                    struct buffer_head *di_bh,
3417                                    u32 cpos, u32 write_len, u32 max_cpos)
3418 {
3419         int ret;
3420         u32 cow_start = 0, cow_len = 0;
3421         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3422         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3423         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3424         struct buffer_head *ref_root_bh = NULL;
3425         struct ocfs2_refcount_tree *ref_tree;
3426         struct ocfs2_cow_context *context = NULL;
3427
3428         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3429
3430         ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list,
3431                                               cpos, write_len, max_cpos,
3432                                               &cow_start, &cow_len);
3433         if (ret) {
3434                 mlog_errno(ret);
3435                 goto out;
3436         }
3437
3438         mlog(0, "CoW inode %lu, cpos %u, write_len %u, cow_start %u, "
3439              "cow_len %u\n", inode->i_ino,
3440              cpos, write_len, cow_start, cow_len);
3441
3442         BUG_ON(cow_len == 0);
3443
3444         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3445         if (!context) {
3446                 ret = -ENOMEM;
3447                 mlog_errno(ret);
3448                 goto out;
3449         }
3450
3451         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
3452                                        1, &ref_tree, &ref_root_bh);
3453         if (ret) {
3454                 mlog_errno(ret);
3455                 goto out;
3456         }
3457
3458         context->inode = inode;
3459         context->cow_start = cow_start;
3460         context->cow_len = cow_len;
3461         context->ref_tree = ref_tree;
3462         context->ref_root_bh = ref_root_bh;
3463         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
3464         context->get_clusters = ocfs2_di_get_clusters;
3465
3466         ocfs2_init_dinode_extent_tree(&context->data_et,
3467                                       INODE_CACHE(inode), di_bh);
3468
3469         ret = ocfs2_replace_cow(context);
3470         if (ret)
3471                 mlog_errno(ret);
3472
3473         /*
3474          * truncate the extent map here since no matter whether we meet with
3475          * any error during the action, we shouldn't trust cached extent map
3476          * any more.
3477          */
3478         ocfs2_extent_map_trunc(inode, cow_start);
3479
3480         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3481         brelse(ref_root_bh);
3482 out:
3483         kfree(context);
3484         return ret;
3485 }
3486
3487 /*
3488  * CoW any and all clusters between cpos and cpos+write_len.
3489  * Don't CoW past max_cpos.  If this returns successfully, all
3490  * clusters between cpos and cpos+write_len are safe to modify.
3491  */
3492 int ocfs2_refcount_cow(struct inode *inode,
3493                        struct buffer_head *di_bh,
3494                        u32 cpos, u32 write_len, u32 max_cpos)
3495 {
3496         int ret = 0;
3497         u32 p_cluster, num_clusters;
3498         unsigned int ext_flags;
3499
3500         while (write_len) {
3501                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3502                                          &num_clusters, &ext_flags);
3503                 if (ret) {
3504                         mlog_errno(ret);
3505                         break;
3506                 }
3507
3508                 if (write_len < num_clusters)
3509                         num_clusters = write_len;
3510
3511                 if (ext_flags & OCFS2_EXT_REFCOUNTED) {
3512                         ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos,
3513                                                       num_clusters, max_cpos);
3514                         if (ret) {
3515                                 mlog_errno(ret);
3516                                 break;
3517                         }
3518                 }
3519
3520                 write_len -= num_clusters;
3521                 cpos += num_clusters;
3522         }
3523
3524         return ret;
3525 }
3526
3527 static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context,
3528                                           u32 v_cluster, u32 *p_cluster,
3529                                           u32 *num_clusters,
3530                                           unsigned int *extent_flags)
3531 {
3532         struct inode *inode = context->inode;
3533         struct ocfs2_xattr_value_root *xv = context->cow_object;
3534
3535         return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster,
3536                                         num_clusters, &xv->xr_list,
3537                                         extent_flags);
3538 }
3539
3540 /*
3541  * Given a xattr value root, calculate the most meta/credits we need for
3542  * refcount tree change if we truncate it to 0.
3543  */
3544 int ocfs2_refcounted_xattr_delete_need(struct inode *inode,
3545                                        struct ocfs2_caching_info *ref_ci,
3546                                        struct buffer_head *ref_root_bh,
3547                                        struct ocfs2_xattr_value_root *xv,
3548                                        int *meta_add, int *credits)
3549 {
3550         int ret = 0, index, ref_blocks = 0;
3551         u32 p_cluster, num_clusters;
3552         u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters);
3553         struct ocfs2_refcount_block *rb;
3554         struct ocfs2_refcount_rec rec;
3555         struct buffer_head *ref_leaf_bh = NULL;
3556
3557         while (cpos < clusters) {
3558                 ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster,
3559                                                &num_clusters, &xv->xr_list,
3560                                                NULL);
3561                 if (ret) {
3562                         mlog_errno(ret);
3563                         goto out;
3564                 }
3565
3566                 cpos += num_clusters;
3567
3568                 while (num_clusters) {
3569                         ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh,
3570                                                      p_cluster, num_clusters,
3571                                                      &rec, &index,
3572                                                      &ref_leaf_bh);
3573                         if (ret) {
3574                                 mlog_errno(ret);
3575                                 goto out;
3576                         }
3577
3578                         BUG_ON(!rec.r_refcount);
3579
3580                         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
3581
3582                         /*
3583                          * We really don't know whether the other clusters is in
3584                          * this refcount block or not, so just take the worst
3585                          * case that all the clusters are in this block and each
3586                          * one will split a refcount rec, so totally we need
3587                          * clusters * 2 new refcount rec.
3588                          */
3589                         if (le64_to_cpu(rb->rf_records.rl_used) + clusters * 2 >
3590                             le16_to_cpu(rb->rf_records.rl_count))
3591                                 ref_blocks++;
3592
3593                         *credits += 1;
3594                         brelse(ref_leaf_bh);
3595                         ref_leaf_bh = NULL;
3596
3597                         if (num_clusters <= le32_to_cpu(rec.r_clusters))
3598                                 break;
3599                         else
3600                                 num_clusters -= le32_to_cpu(rec.r_clusters);
3601                         p_cluster += num_clusters;
3602                 }
3603         }
3604
3605         *meta_add += ref_blocks;
3606         if (!ref_blocks)
3607                 goto out;
3608
3609         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
3610         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)
3611                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
3612         else {
3613                 struct ocfs2_extent_tree et;
3614
3615                 ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh);
3616                 *credits += ocfs2_calc_extend_credits(inode->i_sb,
3617                                                       et.et_root_el,
3618                                                       ref_blocks);
3619         }
3620
3621 out:
3622         brelse(ref_leaf_bh);
3623         return ret;
3624 }
3625
3626 /*
3627  * Do CoW for xattr.
3628  */
3629 int ocfs2_refcount_cow_xattr(struct inode *inode,
3630                              struct ocfs2_dinode *di,
3631                              struct ocfs2_xattr_value_buf *vb,
3632                              struct ocfs2_refcount_tree *ref_tree,
3633                              struct buffer_head *ref_root_bh,
3634                              u32 cpos, u32 write_len,
3635                              struct ocfs2_post_refcount *post)
3636 {
3637         int ret;
3638         struct ocfs2_xattr_value_root *xv = vb->vb_xv;
3639         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3640         struct ocfs2_cow_context *context = NULL;
3641         u32 cow_start, cow_len;
3642
3643         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3644
3645         ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list,
3646                                               cpos, write_len, UINT_MAX,
3647                                               &cow_start, &cow_len);
3648         if (ret) {
3649                 mlog_errno(ret);
3650                 goto out;
3651         }
3652
3653         BUG_ON(cow_len == 0);
3654
3655         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3656         if (!context) {
3657                 ret = -ENOMEM;
3658                 mlog_errno(ret);
3659                 goto out;
3660         }
3661
3662         context->inode = inode;
3663         context->cow_start = cow_start;
3664         context->cow_len = cow_len;
3665         context->ref_tree = ref_tree;
3666         context->ref_root_bh = ref_root_bh;;
3667         context->cow_object = xv;
3668
3669         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd;
3670         /* We need the extra credits for duplicate_clusters by jbd. */
3671         context->extra_credits =
3672                 ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len;
3673         context->get_clusters = ocfs2_xattr_value_get_clusters;
3674         context->post_refcount = post;
3675
3676         ocfs2_init_xattr_value_extent_tree(&context->data_et,
3677                                            INODE_CACHE(inode), vb);
3678
3679         ret = ocfs2_replace_cow(context);
3680         if (ret)
3681                 mlog_errno(ret);
3682
3683 out:
3684         kfree(context);
3685         return ret;
3686 }
3687
3688 /*
3689  * Insert a new extent into refcount tree and mark a extent rec
3690  * as refcounted in the dinode tree.
3691  */
3692 int ocfs2_add_refcount_flag(struct inode *inode,
3693                             struct ocfs2_extent_tree *data_et,
3694                             struct ocfs2_caching_info *ref_ci,
3695                             struct buffer_head *ref_root_bh,
3696                             u32 cpos, u32 p_cluster, u32 num_clusters,
3697                             struct ocfs2_cached_dealloc_ctxt *dealloc,
3698                             struct ocfs2_post_refcount *post)
3699 {
3700         int ret;
3701         handle_t *handle;
3702         int credits = 1, ref_blocks = 0;
3703         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3704         struct ocfs2_alloc_context *meta_ac = NULL;
3705
3706         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
3707                                                ref_ci, ref_root_bh,
3708                                                p_cluster, num_clusters,
3709                                                &ref_blocks, &credits);
3710         if (ret) {
3711                 mlog_errno(ret);
3712                 goto out;
3713         }
3714
3715         mlog(0, "reserve new metadata %d, credits = %d\n",
3716              ref_blocks, credits);
3717
3718         if (ref_blocks) {
3719                 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
3720                                                         ref_blocks, &meta_ac);
3721                 if (ret) {
3722                         mlog_errno(ret);
3723                         goto out;
3724                 }
3725         }
3726
3727         if (post)
3728                 credits += post->credits;
3729
3730         handle = ocfs2_start_trans(osb, credits);
3731         if (IS_ERR(handle)) {
3732                 ret = PTR_ERR(handle);
3733                 mlog_errno(ret);
3734                 goto out;
3735         }
3736
3737         ret = ocfs2_mark_extent_refcounted(inode, data_et, handle,
3738                                            cpos, num_clusters, p_cluster,
3739                                            meta_ac, dealloc);
3740         if (ret) {
3741                 mlog_errno(ret);
3742                 goto out_commit;
3743         }
3744
3745         ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3746                                         p_cluster, num_clusters, 0,
3747                                         meta_ac, dealloc);
3748         if (ret) {
3749                 mlog_errno(ret);
3750                 goto out_commit;
3751         }
3752
3753         if (post && post->func) {
3754                 ret = post->func(inode, handle, post->para);
3755                 if (ret)
3756                         mlog_errno(ret);
3757         }
3758
3759 out_commit:
3760         ocfs2_commit_trans(osb, handle);
3761 out:
3762         if (meta_ac)
3763                 ocfs2_free_alloc_context(meta_ac);
3764         return ret;
3765 }
3766
3767 static int ocfs2_change_ctime(struct inode *inode,
3768                               struct buffer_head *di_bh)
3769 {
3770         int ret;
3771         handle_t *handle;
3772         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3773
3774         handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
3775                                    OCFS2_INODE_UPDATE_CREDITS);
3776         if (IS_ERR(handle)) {
3777                 ret = PTR_ERR(handle);
3778                 mlog_errno(ret);
3779                 goto out;
3780         }
3781
3782         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
3783                                       OCFS2_JOURNAL_ACCESS_WRITE);
3784         if (ret) {
3785                 mlog_errno(ret);
3786                 goto out_commit;
3787         }
3788
3789         inode->i_ctime = CURRENT_TIME;
3790         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
3791         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
3792
3793         ocfs2_journal_dirty(handle, di_bh);
3794
3795 out_commit:
3796         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
3797 out:
3798         return ret;
3799 }
3800
3801 static int ocfs2_attach_refcount_tree(struct inode *inode,
3802                                       struct buffer_head *di_bh)
3803 {
3804         int ret, data_changed = 0;
3805         struct buffer_head *ref_root_bh = NULL;
3806         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3807         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3808         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3809         struct ocfs2_refcount_tree *ref_tree;
3810         unsigned int ext_flags;
3811         loff_t size;
3812         u32 cpos, num_clusters, clusters, p_cluster;
3813         struct ocfs2_cached_dealloc_ctxt dealloc;
3814         struct ocfs2_extent_tree di_et;
3815
3816         ocfs2_init_dealloc_ctxt(&dealloc);
3817
3818         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) {
3819                 ret = ocfs2_create_refcount_tree(inode, di_bh);
3820                 if (ret) {
3821                         mlog_errno(ret);
3822                         goto out;
3823                 }
3824         }
3825
3826         BUG_ON(!di->i_refcount_loc);
3827         ret = ocfs2_lock_refcount_tree(osb,
3828                                        le64_to_cpu(di->i_refcount_loc), 1,
3829                                        &ref_tree, &ref_root_bh);
3830         if (ret) {
3831                 mlog_errno(ret);
3832                 goto out;
3833         }
3834
3835         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
3836                 goto attach_xattr;
3837
3838         ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh);
3839
3840         size = i_size_read(inode);
3841         clusters = ocfs2_clusters_for_bytes(inode->i_sb, size);
3842
3843         cpos = 0;
3844         while (cpos < clusters) {
3845                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3846                                          &num_clusters, &ext_flags);
3847
3848                 if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) {
3849                         ret = ocfs2_add_refcount_flag(inode, &di_et,
3850                                                       &ref_tree->rf_ci,
3851                                                       ref_root_bh, cpos,
3852                                                       p_cluster, num_clusters,
3853                                                       &dealloc, NULL);
3854                         if (ret) {
3855                                 mlog_errno(ret);
3856                                 goto unlock;
3857                         }
3858
3859                         data_changed = 1;
3860                 }
3861                 cpos += num_clusters;
3862         }
3863
3864 attach_xattr:
3865         if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
3866                 ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh,
3867                                                        &ref_tree->rf_ci,
3868                                                        ref_root_bh,
3869                                                        &dealloc);
3870                 if (ret) {
3871                         mlog_errno(ret);
3872                         goto unlock;
3873                 }
3874         }
3875
3876         if (data_changed) {
3877                 ret = ocfs2_change_ctime(inode, di_bh);
3878                 if (ret)
3879                         mlog_errno(ret);
3880         }
3881
3882 unlock:
3883         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3884         brelse(ref_root_bh);
3885
3886         if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) {
3887                 ocfs2_schedule_truncate_log_flush(osb, 1);
3888                 ocfs2_run_deallocs(osb, &dealloc);
3889         }
3890 out:
3891         /*
3892          * Empty the extent map so that we may get the right extent
3893          * record from the disk.
3894          */
3895         ocfs2_extent_map_trunc(inode, 0);
3896
3897         return ret;
3898 }
3899
3900 static int ocfs2_add_refcounted_extent(struct inode *inode,
3901                                    struct ocfs2_extent_tree *et,
3902                                    struct ocfs2_caching_info *ref_ci,
3903                                    struct buffer_head *ref_root_bh,
3904                                    u32 cpos, u32 p_cluster, u32 num_clusters,
3905                                    unsigned int ext_flags,
3906                                    struct ocfs2_cached_dealloc_ctxt *dealloc)
3907 {
3908         int ret;
3909         handle_t *handle;
3910         int credits = 0;
3911         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3912         struct ocfs2_alloc_context *meta_ac = NULL;
3913
3914         ret = ocfs2_lock_refcount_allocators(inode->i_sb,
3915                                              p_cluster, num_clusters,
3916                                              et, ref_ci,
3917                                              ref_root_bh, &meta_ac,
3918                                              NULL, &credits);
3919         if (ret) {
3920                 mlog_errno(ret);
3921                 goto out;
3922         }
3923
3924         handle = ocfs2_start_trans(osb, credits);
3925         if (IS_ERR(handle)) {
3926                 ret = PTR_ERR(handle);
3927                 mlog_errno(ret);
3928                 goto out;
3929         }
3930
3931         ret = ocfs2_insert_extent(handle, et, cpos,
3932                         ocfs2_clusters_to_blocks(inode->i_sb, p_cluster),
3933                         num_clusters, ext_flags, meta_ac);
3934         if (ret) {
3935                 mlog_errno(ret);
3936                 goto out_commit;
3937         }
3938
3939         ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3940                                       p_cluster, num_clusters,
3941                                       meta_ac, dealloc);
3942         if (ret)
3943                 mlog_errno(ret);
3944
3945 out_commit:
3946         ocfs2_commit_trans(osb, handle);
3947 out:
3948         if (meta_ac)
3949                 ocfs2_free_alloc_context(meta_ac);
3950         return ret;
3951 }
3952
3953 static int ocfs2_duplicate_inline_data(struct inode *s_inode,
3954                                        struct buffer_head *s_bh,
3955                                        struct inode *t_inode,
3956                                        struct buffer_head *t_bh)
3957 {
3958         int ret;
3959         handle_t *handle;
3960         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
3961         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
3962         struct ocfs2_dinode *t_di = (struct ocfs2_dinode *)t_bh->b_data;
3963
3964         BUG_ON(!(OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
3965
3966         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
3967         if (IS_ERR(handle)) {
3968                 ret = PTR_ERR(handle);
3969                 mlog_errno(ret);
3970                 goto out;
3971         }
3972
3973         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
3974                                       OCFS2_JOURNAL_ACCESS_WRITE);
3975         if (ret) {
3976                 mlog_errno(ret);
3977                 goto out_commit;
3978         }
3979
3980         t_di->id2.i_data.id_count = s_di->id2.i_data.id_count;
3981         memcpy(t_di->id2.i_data.id_data, s_di->id2.i_data.id_data,
3982                le16_to_cpu(s_di->id2.i_data.id_count));
3983         spin_lock(&OCFS2_I(t_inode)->ip_lock);
3984         OCFS2_I(t_inode)->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
3985         t_di->i_dyn_features = cpu_to_le16(OCFS2_I(t_inode)->ip_dyn_features);
3986         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
3987
3988         ocfs2_journal_dirty(handle, t_bh);
3989
3990 out_commit:
3991         ocfs2_commit_trans(osb, handle);
3992 out:
3993         return ret;
3994 }
3995
3996 static int ocfs2_duplicate_extent_list(struct inode *s_inode,
3997                                 struct inode *t_inode,
3998                                 struct buffer_head *t_bh,
3999                                 struct ocfs2_caching_info *ref_ci,
4000                                 struct buffer_head *ref_root_bh,
4001                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
4002 {
4003         int ret = 0;
4004         u32 p_cluster, num_clusters, clusters, cpos;
4005         loff_t size;
4006         unsigned int ext_flags;
4007         struct ocfs2_extent_tree et;
4008
4009         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh);
4010
4011         size = i_size_read(s_inode);
4012         clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size);
4013
4014         cpos = 0;
4015         while (cpos < clusters) {
4016                 ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster,
4017                                          &num_clusters, &ext_flags);
4018
4019                 if (p_cluster) {
4020                         ret = ocfs2_add_refcounted_extent(t_inode, &et,
4021                                                           ref_ci, ref_root_bh,
4022                                                           cpos, p_cluster,
4023                                                           num_clusters,
4024                                                           ext_flags,
4025                                                           dealloc);
4026                         if (ret) {
4027                                 mlog_errno(ret);
4028                                 goto out;
4029                         }
4030                 }
4031
4032                 cpos += num_clusters;
4033         }
4034
4035 out:
4036         return ret;
4037 }
4038
4039 /*
4040  * change the new file's attributes to the src.
4041  *
4042  * reflink creates a snapshot of a file, that means the attributes
4043  * must be identical except for three exceptions - nlink, ino, and ctime.
4044  */
4045 static int ocfs2_complete_reflink(struct inode *s_inode,
4046                                   struct buffer_head *s_bh,
4047                                   struct inode *t_inode,
4048                                   struct buffer_head *t_bh,
4049                                   bool preserve)
4050 {
4051         int ret;
4052         handle_t *handle;
4053         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
4054         struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data;
4055         loff_t size = i_size_read(s_inode);
4056
4057         handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb),
4058                                    OCFS2_INODE_UPDATE_CREDITS);
4059         if (IS_ERR(handle)) {
4060                 ret = PTR_ERR(handle);
4061                 mlog_errno(ret);
4062                 return ret;
4063         }
4064
4065         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
4066                                       OCFS2_JOURNAL_ACCESS_WRITE);
4067         if (ret) {
4068                 mlog_errno(ret);
4069                 goto out_commit;
4070         }
4071
4072         spin_lock(&OCFS2_I(t_inode)->ip_lock);
4073         OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters;
4074         OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr;
4075         OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features;
4076         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
4077         i_size_write(t_inode, size);
4078
4079         di->i_xattr_inline_size = s_di->i_xattr_inline_size;
4080         di->i_clusters = s_di->i_clusters;
4081         di->i_size = s_di->i_size;
4082         di->i_dyn_features = s_di->i_dyn_features;
4083         di->i_attr = s_di->i_attr;
4084
4085         if (preserve) {
4086                 di->i_uid = s_di->i_uid;
4087                 di->i_gid = s_di->i_gid;
4088                 di->i_mode = s_di->i_mode;
4089
4090                 /*
4091                  * update time.
4092                  * we want mtime to appear identical to the source and
4093                  * update ctime.
4094                  */
4095                 t_inode->i_ctime = CURRENT_TIME;
4096
4097                 di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec);
4098                 di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec);
4099
4100                 t_inode->i_mtime = s_inode->i_mtime;
4101                 di->i_mtime = s_di->i_mtime;
4102                 di->i_mtime_nsec = s_di->i_mtime_nsec;
4103         }
4104
4105         ocfs2_journal_dirty(handle, t_bh);
4106
4107 out_commit:
4108         ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle);
4109         return ret;
4110 }
4111
4112 static int ocfs2_create_reflink_node(struct inode *s_inode,
4113                                      struct buffer_head *s_bh,
4114                                      struct inode *t_inode,
4115                                      struct buffer_head *t_bh,
4116                                      bool preserve)
4117 {
4118         int ret;
4119         struct buffer_head *ref_root_bh = NULL;
4120         struct ocfs2_cached_dealloc_ctxt dealloc;
4121         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
4122         struct ocfs2_refcount_block *rb;
4123         struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data;
4124         struct ocfs2_refcount_tree *ref_tree;
4125
4126         ocfs2_init_dealloc_ctxt(&dealloc);
4127
4128         ret = ocfs2_set_refcount_tree(t_inode, t_bh,
4129                                       le64_to_cpu(di->i_refcount_loc));
4130         if (ret) {
4131                 mlog_errno(ret);
4132                 goto out;
4133         }
4134
4135         if (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4136                 ret = ocfs2_duplicate_inline_data(s_inode, s_bh,
4137                                                   t_inode, t_bh);
4138                 if (ret)
4139                         mlog_errno(ret);
4140                 goto out;
4141         }
4142
4143         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
4144                                        1, &ref_tree, &ref_root_bh);
4145         if (ret) {
4146                 mlog_errno(ret);
4147                 goto out;
4148         }
4149         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
4150
4151         ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh,
4152                                           &ref_tree->rf_ci, ref_root_bh,
4153                                           &dealloc);
4154         if (ret) {
4155                 mlog_errno(ret);
4156                 goto out_unlock_refcount;
4157         }
4158
4159 out_unlock_refcount:
4160         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4161         brelse(ref_root_bh);
4162 out:
4163         if (ocfs2_dealloc_has_cluster(&dealloc)) {
4164                 ocfs2_schedule_truncate_log_flush(osb, 1);
4165                 ocfs2_run_deallocs(osb, &dealloc);
4166         }
4167
4168         return ret;
4169 }
4170
4171 static int __ocfs2_reflink(struct dentry *old_dentry,
4172                            struct buffer_head *old_bh,
4173                            struct inode *new_inode,
4174                            bool preserve)
4175 {
4176         int ret;
4177         struct inode *inode = old_dentry->d_inode;
4178         struct buffer_head *new_bh = NULL;
4179
4180         ret = filemap_fdatawrite(inode->i_mapping);
4181         if (ret) {
4182                 mlog_errno(ret);
4183                 goto out;
4184         }
4185
4186         ret = ocfs2_attach_refcount_tree(inode, old_bh);
4187         if (ret) {
4188                 mlog_errno(ret);
4189                 goto out;
4190         }
4191
4192         mutex_lock(&new_inode->i_mutex);
4193         ret = ocfs2_inode_lock(new_inode, &new_bh, 1);
4194         if (ret) {
4195                 mlog_errno(ret);
4196                 goto out_unlock;
4197         }
4198
4199         ret = ocfs2_create_reflink_node(inode, old_bh,
4200                                         new_inode, new_bh, preserve);
4201         if (ret) {
4202                 mlog_errno(ret);
4203                 goto inode_unlock;
4204         }
4205
4206         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
4207                 ret = ocfs2_reflink_xattrs(inode, old_bh,
4208                                            new_inode, new_bh,
4209                                            preserve);
4210                 if (ret) {
4211                         mlog_errno(ret);
4212                         goto inode_unlock;
4213                 }
4214         }
4215
4216         ret = ocfs2_complete_reflink(inode, old_bh,
4217                                      new_inode, new_bh, preserve);
4218         if (ret)
4219                 mlog_errno(ret);
4220
4221 inode_unlock:
4222         ocfs2_inode_unlock(new_inode, 1);
4223         brelse(new_bh);
4224 out_unlock:
4225         mutex_unlock(&new_inode->i_mutex);
4226 out:
4227         if (!ret) {
4228                 ret = filemap_fdatawait(inode->i_mapping);
4229                 if (ret)
4230                         mlog_errno(ret);
4231         }
4232         return ret;
4233 }
4234
4235 static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir,
4236                          struct dentry *new_dentry, bool preserve)
4237 {
4238         int error;
4239         struct inode *inode = old_dentry->d_inode;
4240         struct buffer_head *old_bh = NULL;
4241         struct inode *new_orphan_inode = NULL;
4242
4243         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4244                 return -EOPNOTSUPP;
4245
4246         error = ocfs2_create_inode_in_orphan(dir, inode->i_mode,
4247                                              &new_orphan_inode);
4248         if (error) {
4249                 mlog_errno(error);
4250                 goto out;
4251         }
4252
4253         error = ocfs2_inode_lock(inode, &old_bh, 1);
4254         if (error) {
4255                 mlog_errno(error);
4256                 goto out;
4257         }
4258
4259         down_write(&OCFS2_I(inode)->ip_xattr_sem);
4260         down_write(&OCFS2_I(inode)->ip_alloc_sem);
4261         error = __ocfs2_reflink(old_dentry, old_bh,
4262                                 new_orphan_inode, preserve);
4263         up_write(&OCFS2_I(inode)->ip_alloc_sem);
4264         up_write(&OCFS2_I(inode)->ip_xattr_sem);
4265
4266         ocfs2_inode_unlock(inode, 1);
4267         brelse(old_bh);
4268
4269         if (error) {
4270                 mlog_errno(error);
4271                 goto out;
4272         }
4273
4274         /* If the security isn't preserved, we need to re-initialize them. */
4275         if (!preserve) {
4276                 error = ocfs2_init_security_and_acl(dir, new_orphan_inode);
4277                 if (error)
4278                         mlog_errno(error);
4279         }
4280 out:
4281         if (!error) {
4282                 error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode,
4283                                                        new_dentry);
4284                 if (error)
4285                         mlog_errno(error);
4286         }
4287
4288         if (new_orphan_inode) {
4289                 /*
4290                  * We need to open_unlock the inode no matter whether we
4291                  * succeed or not, so that other nodes can delete it later.
4292                  */
4293                 ocfs2_open_unlock(new_orphan_inode);
4294                 if (error)
4295                         iput(new_orphan_inode);
4296         }
4297
4298         return error;
4299 }
4300
4301 /*
4302  * Below here are the bits used by OCFS2_IOC_REFLINK() to fake
4303  * sys_reflink().  This will go away when vfs_reflink() exists in
4304  * fs/namei.c.
4305  */
4306
4307 /* copied from may_create in VFS. */
4308 static inline int ocfs2_may_create(struct inode *dir, struct dentry *child)
4309 {
4310         if (child->d_inode)
4311                 return -EEXIST;
4312         if (IS_DEADDIR(dir))
4313                 return -ENOENT;
4314         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
4315 }
4316
4317 /* copied from user_path_parent. */
4318 static int ocfs2_user_path_parent(const char __user *path,
4319                                   struct nameidata *nd, char **name)
4320 {
4321         char *s = getname(path);
4322         int error;
4323
4324         if (IS_ERR(s))
4325                 return PTR_ERR(s);
4326
4327         error = path_lookup(s, LOOKUP_PARENT, nd);
4328         if (error)
4329                 putname(s);
4330         else
4331                 *name = s;
4332
4333         return error;
4334 }
4335
4336 /**
4337  * ocfs2_vfs_reflink - Create a reference-counted link
4338  *
4339  * @old_dentry:        source dentry + inode
4340  * @dir:       directory to create the target
4341  * @new_dentry:        target dentry
4342  * @preserve:  if true, preserve all file attributes
4343  */
4344 static int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir,
4345                              struct dentry *new_dentry, bool preserve)
4346 {
4347         struct inode *inode = old_dentry->d_inode;
4348         int error;
4349
4350         if (!inode)
4351                 return -ENOENT;
4352
4353         error = ocfs2_may_create(dir, new_dentry);
4354         if (error)
4355                 return error;
4356
4357         if (dir->i_sb != inode->i_sb)
4358                 return -EXDEV;
4359
4360         /*
4361          * A reflink to an append-only or immutable file cannot be created.
4362          */
4363         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4364                 return -EPERM;
4365
4366         /* Only regular files can be reflinked. */
4367         if (!S_ISREG(inode->i_mode))
4368                 return -EPERM;
4369
4370         /*
4371          * If the caller wants to preserve ownership, they require the
4372          * rights to do so.
4373          */
4374         if (preserve) {
4375                 if ((current_fsuid() != inode->i_uid) && !capable(CAP_CHOWN))
4376                         return -EPERM;
4377                 if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN))
4378                         return -EPERM;
4379         }
4380
4381         /*
4382          * If the caller is modifying any aspect of the attributes, they
4383          * are not creating a snapshot.  They need read permission on the
4384          * file.
4385          */
4386         if (!preserve) {
4387                 error = inode_permission(inode, MAY_READ);
4388                 if (error)
4389                         return error;
4390         }
4391
4392         mutex_lock(&inode->i_mutex);
4393         dquot_initialize(dir);
4394         error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve);
4395         mutex_unlock(&inode->i_mutex);
4396         if (!error)
4397                 fsnotify_create(dir, new_dentry);
4398         return error;
4399 }
4400 /*
4401  * Most codes are copied from sys_linkat.
4402  */
4403 int ocfs2_reflink_ioctl(struct inode *inode,
4404                         const char __user *oldname,
4405                         const char __user *newname,
4406                         bool preserve)
4407 {
4408         struct dentry *new_dentry;
4409         struct nameidata nd;
4410         struct path old_path;
4411         int error;
4412         char *to = NULL;
4413
4414         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4415                 return -EOPNOTSUPP;
4416
4417         error = user_path_at(AT_FDCWD, oldname, 0, &old_path);
4418         if (error) {
4419                 mlog_errno(error);
4420                 return error;
4421         }
4422
4423         error = ocfs2_user_path_parent(newname, &nd, &to);
4424         if (error) {
4425                 mlog_errno(error);
4426                 goto out;
4427         }
4428
4429         error = -EXDEV;
4430         if (old_path.mnt != nd.path.mnt)
4431                 goto out_release;
4432         new_dentry = lookup_create(&nd, 0);
4433         error = PTR_ERR(new_dentry);
4434         if (IS_ERR(new_dentry)) {
4435                 mlog_errno(error);
4436                 goto out_unlock;
4437         }
4438
4439         error = mnt_want_write(nd.path.mnt);
4440         if (error) {
4441                 mlog_errno(error);
4442                 goto out_dput;
4443         }
4444
4445         error = ocfs2_vfs_reflink(old_path.dentry,
4446                                   nd.path.dentry->d_inode,
4447                                   new_dentry, preserve);
4448         mnt_drop_write(nd.path.mnt);
4449 out_dput:
4450         dput(new_dentry);
4451 out_unlock:
4452         mutex_unlock(&nd.path.dentry->d_inode->i_mutex);
4453 out_release:
4454         path_put(&nd.path);
4455         putname(to);
4456 out:
4457         path_put(&old_path);
4458
4459         return error;
4460 }