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/slab.h>
  41 #include <linux/writeback.h>
  42 #include <linux/pagevec.h>
  43 #include <linux/swap.h>
  44 #include <linux/security.h>
  45 #include <linux/fsnotify.h>
  46 #include <linux/quotaops.h>
  47 #include <linux/namei.h>
  48 #include <linux/mount.h>
  49
  50 struct ocfs2_cow_context {
  51         struct inode *inode;
  52         struct file *file;
  53         u32 cow_start;
  54         u32 cow_len;
  55         struct ocfs2_extent_tree data_et;
  56         struct ocfs2_refcount_tree *ref_tree;
  57         struct buffer_head *ref_root_bh;
  58         struct ocfs2_alloc_context *meta_ac;
  59         struct ocfs2_alloc_context *data_ac;
  60         struct ocfs2_cached_dealloc_ctxt dealloc;
  61         void *cow_object;
  62         struct ocfs2_post_refcount *post_refcount;
  63         int extra_credits;
  64         int (*get_clusters)(struct ocfs2_cow_context *context,
  65                             u32 v_cluster, u32 *p_cluster,
  66                             u32 *num_clusters,
  67                             unsigned int *extent_flags);
  68         int (*cow_duplicate_clusters)(handle_t *handle,
  69                                       struct ocfs2_cow_context *context,
  70                                       u32 cpos, u32 old_cluster,
  71                                       u32 new_cluster, u32 new_len);
  72 };
  73
  74 static inline struct ocfs2_refcount_tree *
  75 cache_info_to_refcount(struct ocfs2_caching_info *ci)
  76 {
  77         return container_of(ci, struct ocfs2_refcount_tree, rf_ci);
  78 }
  79
  80 static int ocfs2_validate_refcount_block(struct super_block *sb,
  81                                          struct buffer_head *bh)
  82 {
  83         int rc;
  84         struct ocfs2_refcount_block *rb =
  85                 (struct ocfs2_refcount_block *)bh->b_data;
  86
  87         mlog(0, "Validating refcount block %llu\n",
  88              (unsigned long long)bh->b_blocknr);
  89
  90         BUG_ON(!buffer_uptodate(bh));
  91
  92         /*
  93          * If the ecc fails, we return the error but otherwise
  94          * leave the filesystem running.  We know any error is
  95          * local to this block.
  96          */
  97         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check);
  98         if (rc) {
  99                 mlog(ML_ERROR, "Checksum failed for refcount block %llu\n",
 100                      (unsigned long long)bh->b_blocknr);
 101                 return rc;
 102         }
 103
 104
 105         if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) {
 106                 ocfs2_error(sb,
 107                             "Refcount block #%llu has bad signature %.*s",
 108                             (unsigned long long)bh->b_blocknr, 7,
 109                             rb->rf_signature);
 110                 return -EINVAL;
 111         }
 112
 113         if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) {
 114                 ocfs2_error(sb,
 115                             "Refcount block #%llu has an invalid rf_blkno "
 116                             "of %llu",
 117                             (unsigned long long)bh->b_blocknr,
 118                             (unsigned long long)le64_to_cpu(rb->rf_blkno));
 119                 return -EINVAL;
 120         }
 121
 122         if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) {
 123                 ocfs2_error(sb,
 124                             "Refcount block #%llu has an invalid "
 125                             "rf_fs_generation of #%u",
 126                             (unsigned long long)bh->b_blocknr,
 127                             le32_to_cpu(rb->rf_fs_generation));
 128                 return -EINVAL;
 129         }
 130
 131         return 0;
 132 }
 133
 134 static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci,
 135                                      u64 rb_blkno,
 136                                      struct buffer_head **bh)
 137 {
 138         int rc;
 139         struct buffer_head *tmp = *bh;
 140
 141         rc = ocfs2_read_block(ci, rb_blkno, &tmp,
 142                               ocfs2_validate_refcount_block);
 143
 144         /* If ocfs2_read_block() got us a new bh, pass it up. */
 145         if (!rc && !*bh)
 146                 *bh = tmp;
 147
 148         return rc;
 149 }
 150
 151 static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci)
 152 {
 153         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 154
 155         return rf->rf_blkno;
 156 }
 157
 158 static struct super_block *
 159 ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci)
 160 {
 161         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 162
 163         return rf->rf_sb;
 164 }
 165
 166 static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci)
 167 {
 168         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 169
 170         spin_lock(&rf->rf_lock);
 171 }
 172
 173 static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci)
 174 {
 175         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 176
 177         spin_unlock(&rf->rf_lock);
 178 }
 179
 180 static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci)
 181 {
 182         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 183
 184         mutex_lock(&rf->rf_io_mutex);
 185 }
 186
 187 static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci)
 188 {
 189         struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci);
 190
 191         mutex_unlock(&rf->rf_io_mutex);
 192 }
 193
 194 static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = {
 195         .co_owner               = ocfs2_refcount_cache_owner,
 196         .co_get_super           = ocfs2_refcount_cache_get_super,
 197         .co_cache_lock          = ocfs2_refcount_cache_lock,
 198         .co_cache_unlock        = ocfs2_refcount_cache_unlock,
 199         .co_io_lock             = ocfs2_refcount_cache_io_lock,
 200         .co_io_unlock           = ocfs2_refcount_cache_io_unlock,
 201 };
 202
 203 static struct ocfs2_refcount_tree *
 204 ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno)
 205 {
 206         struct rb_node *n = osb->osb_rf_lock_tree.rb_node;
 207         struct ocfs2_refcount_tree *tree = NULL;
 208
 209         while (n) {
 210                 tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node);
 211
 212                 if (blkno < tree->rf_blkno)
 213                         n = n->rb_left;
 214                 else if (blkno > tree->rf_blkno)
 215                         n = n->rb_right;
 216                 else
 217                         return tree;
 218         }
 219
 220         return NULL;
 221 }
 222
 223 /* osb_lock is already locked. */
 224 static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb,
 225                                        struct ocfs2_refcount_tree *new)
 226 {
 227         u64 rf_blkno = new->rf_blkno;
 228         struct rb_node *parent = NULL;
 229         struct rb_node **p = &osb->osb_rf_lock_tree.rb_node;
 230         struct ocfs2_refcount_tree *tmp;
 231
 232         while (*p) {
 233                 parent = *p;
 234
 235                 tmp = rb_entry(parent, struct ocfs2_refcount_tree,
 236                                rf_node);
 237
 238                 if (rf_blkno < tmp->rf_blkno)
 239                         p = &(*p)->rb_left;
 240                 else if (rf_blkno > tmp->rf_blkno)
 241                         p = &(*p)->rb_right;
 242                 else {
 243                         /* This should never happen! */
 244                         mlog(ML_ERROR, "Duplicate refcount block %llu found!\n",
 245                              (unsigned long long)rf_blkno);
 246                         BUG();
 247                 }
 248         }
 249
 250         rb_link_node(&new->rf_node, parent, p);
 251         rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree);
 252 }
 253
 254 static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree)
 255 {
 256         ocfs2_metadata_cache_exit(&tree->rf_ci);
 257         ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres);
 258         ocfs2_lock_res_free(&tree->rf_lockres);
 259         kfree(tree);
 260 }
 261
 262 static inline void
 263 ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb,
 264                                         struct ocfs2_refcount_tree *tree)
 265 {
 266         rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree);
 267         if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree)
 268                 osb->osb_ref_tree_lru = NULL;
 269 }
 270
 271 static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb,
 272                                         struct ocfs2_refcount_tree *tree)
 273 {
 274         spin_lock(&osb->osb_lock);
 275         ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
 276         spin_unlock(&osb->osb_lock);
 277 }
 278
 279 static void ocfs2_kref_remove_refcount_tree(struct kref *kref)
 280 {
 281         struct ocfs2_refcount_tree *tree =
 282                 container_of(kref, struct ocfs2_refcount_tree, rf_getcnt);
 283
 284         ocfs2_free_refcount_tree(tree);
 285 }
 286
 287 static inline void
 288 ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree)
 289 {
 290         kref_get(&tree->rf_getcnt);
 291 }
 292
 293 static inline void
 294 ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree)
 295 {
 296         kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree);
 297 }
 298
 299 static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new,
 300                                                struct super_block *sb)
 301 {
 302         ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops);
 303         mutex_init(&new->rf_io_mutex);
 304         new->rf_sb = sb;
 305         spin_lock_init(&new->rf_lock);
 306 }
 307
 308 static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb,
 309                                         struct ocfs2_refcount_tree *new,
 310                                         u64 rf_blkno, u32 generation)
 311 {
 312         init_rwsem(&new->rf_sem);
 313         ocfs2_refcount_lock_res_init(&new->rf_lockres, osb,
 314                                      rf_blkno, generation);
 315 }
 316
 317 static struct ocfs2_refcount_tree*
 318 ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno)
 319 {
 320         struct ocfs2_refcount_tree *new;
 321
 322         new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS);
 323         if (!new)
 324                 return NULL;
 325
 326         new->rf_blkno = rf_blkno;
 327         kref_init(&new->rf_getcnt);
 328         ocfs2_init_refcount_tree_ci(new, osb->sb);
 329
 330         return new;
 331 }
 332
 333 static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno,
 334                                    struct ocfs2_refcount_tree **ret_tree)
 335 {
 336         int ret = 0;
 337         struct ocfs2_refcount_tree *tree, *new = NULL;
 338         struct buffer_head *ref_root_bh = NULL;
 339         struct ocfs2_refcount_block *ref_rb;
 340
 341         spin_lock(&osb->osb_lock);
 342         if (osb->osb_ref_tree_lru &&
 343             osb->osb_ref_tree_lru->rf_blkno == rf_blkno)
 344                 tree = osb->osb_ref_tree_lru;
 345         else
 346                 tree = ocfs2_find_refcount_tree(osb, rf_blkno);
 347         if (tree)
 348                 goto out;
 349
 350         spin_unlock(&osb->osb_lock);
 351
 352         new = ocfs2_allocate_refcount_tree(osb, rf_blkno);
 353         if (!new) {
 354                 ret = -ENOMEM;
 355                 mlog_errno(ret);
 356                 return ret;
 357         }
 358         /*
 359          * We need the generation to create the refcount tree lock and since
 360          * it isn't changed during the tree modification, we are safe here to
 361          * read without protection.
 362          * We also have to purge the cache after we create the lock since the
 363          * refcount block may have the stale data. It can only be trusted when
 364          * we hold the refcount lock.
 365          */
 366         ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh);
 367         if (ret) {
 368                 mlog_errno(ret);
 369                 ocfs2_metadata_cache_exit(&new->rf_ci);
 370                 kfree(new);
 371                 return ret;
 372         }
 373
 374         ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 375         new->rf_generation = le32_to_cpu(ref_rb->rf_generation);
 376         ocfs2_init_refcount_tree_lock(osb, new, rf_blkno,
 377                                       new->rf_generation);
 378         ocfs2_metadata_cache_purge(&new->rf_ci);
 379
 380         spin_lock(&osb->osb_lock);
 381         tree = ocfs2_find_refcount_tree(osb, rf_blkno);
 382         if (tree)
 383                 goto out;
 384
 385         ocfs2_insert_refcount_tree(osb, new);
 386
 387         tree = new;
 388         new = NULL;
 389
 390 out:
 391         *ret_tree = tree;
 392
 393         osb->osb_ref_tree_lru = tree;
 394
 395         spin_unlock(&osb->osb_lock);
 396
 397         if (new)
 398                 ocfs2_free_refcount_tree(new);
 399
 400         brelse(ref_root_bh);
 401         return ret;
 402 }
 403
 404 static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno)
 405 {
 406         int ret;
 407         struct buffer_head *di_bh = NULL;
 408         struct ocfs2_dinode *di;
 409
 410         ret = ocfs2_read_inode_block(inode, &di_bh);
 411         if (ret) {
 412                 mlog_errno(ret);
 413                 goto out;
 414         }
 415
 416         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
 417
 418         di = (struct ocfs2_dinode *)di_bh->b_data;
 419         *ref_blkno = le64_to_cpu(di->i_refcount_loc);
 420         brelse(di_bh);
 421 out:
 422         return ret;
 423 }
 424
 425 static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
 426                                       struct ocfs2_refcount_tree *tree, int rw)
 427 {
 428         int ret;
 429
 430         ret = ocfs2_refcount_lock(tree, rw);
 431         if (ret) {
 432                 mlog_errno(ret);
 433                 goto out;
 434         }
 435
 436         if (rw)
 437                 down_write(&tree->rf_sem);
 438         else
 439                 down_read(&tree->rf_sem);
 440
 441 out:
 442         return ret;
 443 }
 444
 445 /*
 446  * Lock the refcount tree pointed by ref_blkno and return the tree.
 447  * In most case, we lock the tree and read the refcount block.
 448  * So read it here if the caller really needs it.
 449  *
 450  * If the tree has been re-created by other node, it will free the
 451  * old one and re-create it.
 452  */
 453 int ocfs2_lock_refcount_tree(struct ocfs2_super *osb,
 454                              u64 ref_blkno, int rw,
 455                              struct ocfs2_refcount_tree **ret_tree,
 456                              struct buffer_head **ref_bh)
 457 {
 458         int ret, delete_tree = 0;
 459         struct ocfs2_refcount_tree *tree = NULL;
 460         struct buffer_head *ref_root_bh = NULL;
 461         struct ocfs2_refcount_block *rb;
 462
 463 again:
 464         ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree);
 465         if (ret) {
 466                 mlog_errno(ret);
 467                 return ret;
 468         }
 469
 470         ocfs2_refcount_tree_get(tree);
 471
 472         ret = __ocfs2_lock_refcount_tree(osb, tree, rw);
 473         if (ret) {
 474                 mlog_errno(ret);
 475                 ocfs2_refcount_tree_put(tree);
 476                 goto out;
 477         }
 478
 479         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
 480                                         &ref_root_bh);
 481         if (ret) {
 482                 mlog_errno(ret);
 483                 ocfs2_unlock_refcount_tree(osb, tree, rw);
 484                 ocfs2_refcount_tree_put(tree);
 485                 goto out;
 486         }
 487
 488         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 489         /*
 490          * If the refcount block has been freed and re-created, we may need
 491          * to recreate the refcount tree also.
 492          *
 493          * Here we just remove the tree from the rb-tree, and the last
 494          * kref holder will unlock and delete this refcount_tree.
 495          * Then we goto "again" and ocfs2_get_refcount_tree will create
 496          * the new refcount tree for us.
 497          */
 498         if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) {
 499                 if (!tree->rf_removed) {
 500                         ocfs2_erase_refcount_tree_from_list(osb, tree);
 501                         tree->rf_removed = 1;
 502                         delete_tree = 1;
 503                 }
 504
 505                 ocfs2_unlock_refcount_tree(osb, tree, rw);
 506                 /*
 507                  * We get an extra reference when we create the refcount
 508                  * tree, so another put will destroy it.
 509                  */
 510                 if (delete_tree)
 511                         ocfs2_refcount_tree_put(tree);
 512                 brelse(ref_root_bh);
 513                 ref_root_bh = NULL;
 514                 goto again;
 515         }
 516
 517         *ret_tree = tree;
 518         if (ref_bh) {
 519                 *ref_bh = ref_root_bh;
 520                 ref_root_bh = NULL;
 521         }
 522 out:
 523         brelse(ref_root_bh);
 524         return ret;
 525 }
 526
 527 void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb,
 528                                 struct ocfs2_refcount_tree *tree, int rw)
 529 {
 530         if (rw)
 531                 up_write(&tree->rf_sem);
 532         else
 533                 up_read(&tree->rf_sem);
 534
 535         ocfs2_refcount_unlock(tree, rw);
 536         ocfs2_refcount_tree_put(tree);
 537 }
 538
 539 void ocfs2_purge_refcount_trees(struct ocfs2_super *osb)
 540 {
 541         struct rb_node *node;
 542         struct ocfs2_refcount_tree *tree;
 543         struct rb_root *root = &osb->osb_rf_lock_tree;
 544
 545         while ((node = rb_last(root)) != NULL) {
 546                 tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node);
 547
 548                 mlog(0, "Purge tree %llu\n",
 549                      (unsigned long long) tree->rf_blkno);
 550
 551                 rb_erase(&tree->rf_node, root);
 552                 ocfs2_free_refcount_tree(tree);
 553         }
 554 }
 555
 556 /*
 557  * Create a refcount tree for an inode.
 558  * We take for granted that the inode is already locked.
 559  */
 560 static int ocfs2_create_refcount_tree(struct inode *inode,
 561                                       struct buffer_head *di_bh)
 562 {
 563         int ret;
 564         handle_t *handle = NULL;
 565         struct ocfs2_alloc_context *meta_ac = NULL;
 566         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 567         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 568         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 569         struct buffer_head *new_bh = NULL;
 570         struct ocfs2_refcount_block *rb;
 571         struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL;
 572         u16 suballoc_bit_start;
 573         u32 num_got;
 574         u64 suballoc_loc, first_blkno;
 575
 576         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
 577
 578         mlog(0, "create tree for inode %lu\n", inode->i_ino);
 579
 580         ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
 581         if (ret) {
 582                 mlog_errno(ret);
 583                 goto out;
 584         }
 585
 586         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS);
 587         if (IS_ERR(handle)) {
 588                 ret = PTR_ERR(handle);
 589                 mlog_errno(ret);
 590                 goto out;
 591         }
 592
 593         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 594                                       OCFS2_JOURNAL_ACCESS_WRITE);
 595         if (ret) {
 596                 mlog_errno(ret);
 597                 goto out_commit;
 598         }
 599
 600         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
 601                                    &suballoc_bit_start, &num_got,
 602                                    &first_blkno);
 603         if (ret) {
 604                 mlog_errno(ret);
 605                 goto out_commit;
 606         }
 607
 608         new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno);
 609         if (!new_tree) {
 610                 ret = -ENOMEM;
 611                 mlog_errno(ret);
 612                 goto out_commit;
 613         }
 614
 615         new_bh = sb_getblk(inode->i_sb, first_blkno);
 616         ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh);
 617
 618         ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh,
 619                                       OCFS2_JOURNAL_ACCESS_CREATE);
 620         if (ret) {
 621                 mlog_errno(ret);
 622                 goto out_commit;
 623         }
 624
 625         /* Initialize ocfs2_refcount_block. */
 626         rb = (struct ocfs2_refcount_block *)new_bh->b_data;
 627         memset(rb, 0, inode->i_sb->s_blocksize);
 628         strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
 629         rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
 630         rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
 631         rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
 632         rb->rf_fs_generation = cpu_to_le32(osb->fs_generation);
 633         rb->rf_blkno = cpu_to_le64(first_blkno);
 634         rb->rf_count = cpu_to_le32(1);
 635         rb->rf_records.rl_count =
 636                         cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb));
 637         spin_lock(&osb->osb_lock);
 638         rb->rf_generation = osb->s_next_generation++;
 639         spin_unlock(&osb->osb_lock);
 640
 641         ocfs2_journal_dirty(handle, new_bh);
 642
 643         spin_lock(&oi->ip_lock);
 644         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
 645         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 646         di->i_refcount_loc = cpu_to_le64(first_blkno);
 647         spin_unlock(&oi->ip_lock);
 648
 649         mlog(0, "created tree for inode %lu, refblock %llu\n",
 650              inode->i_ino, (unsigned long long)first_blkno);
 651
 652         ocfs2_journal_dirty(handle, di_bh);
 653
 654         /*
 655          * We have to init the tree lock here since it will use
 656          * the generation number to create it.
 657          */
 658         new_tree->rf_generation = le32_to_cpu(rb->rf_generation);
 659         ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno,
 660                                       new_tree->rf_generation);
 661
 662         spin_lock(&osb->osb_lock);
 663         tree = ocfs2_find_refcount_tree(osb, first_blkno);
 664
 665         /*
 666          * We've just created a new refcount tree in this block.  If
 667          * we found a refcount tree on the ocfs2_super, it must be
 668          * one we just deleted.  We free the old tree before
 669          * inserting the new tree.
 670          */
 671         BUG_ON(tree && tree->rf_generation == new_tree->rf_generation);
 672         if (tree)
 673                 ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree);
 674         ocfs2_insert_refcount_tree(osb, new_tree);
 675         spin_unlock(&osb->osb_lock);
 676         new_tree = NULL;
 677         if (tree)
 678                 ocfs2_refcount_tree_put(tree);
 679
 680 out_commit:
 681         ocfs2_commit_trans(osb, handle);
 682
 683 out:
 684         if (new_tree) {
 685                 ocfs2_metadata_cache_exit(&new_tree->rf_ci);
 686                 kfree(new_tree);
 687         }
 688
 689         brelse(new_bh);
 690         if (meta_ac)
 691                 ocfs2_free_alloc_context(meta_ac);
 692
 693         return ret;
 694 }
 695
 696 static int ocfs2_set_refcount_tree(struct inode *inode,
 697                                    struct buffer_head *di_bh,
 698                                    u64 refcount_loc)
 699 {
 700         int ret;
 701         handle_t *handle = NULL;
 702         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 703         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 704         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 705         struct buffer_head *ref_root_bh = NULL;
 706         struct ocfs2_refcount_block *rb;
 707         struct ocfs2_refcount_tree *ref_tree;
 708
 709         BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL);
 710
 711         ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
 712                                        &ref_tree, &ref_root_bh);
 713         if (ret) {
 714                 mlog_errno(ret);
 715                 return ret;
 716         }
 717
 718         handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS);
 719         if (IS_ERR(handle)) {
 720                 ret = PTR_ERR(handle);
 721                 mlog_errno(ret);
 722                 goto out;
 723         }
 724
 725         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 726                                       OCFS2_JOURNAL_ACCESS_WRITE);
 727         if (ret) {
 728                 mlog_errno(ret);
 729                 goto out_commit;
 730         }
 731
 732         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh,
 733                                       OCFS2_JOURNAL_ACCESS_WRITE);
 734         if (ret) {
 735                 mlog_errno(ret);
 736                 goto out_commit;
 737         }
 738
 739         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
 740         le32_add_cpu(&rb->rf_count, 1);
 741
 742         ocfs2_journal_dirty(handle, ref_root_bh);
 743
 744         spin_lock(&oi->ip_lock);
 745         oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL;
 746         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 747         di->i_refcount_loc = cpu_to_le64(refcount_loc);
 748         spin_unlock(&oi->ip_lock);
 749         ocfs2_journal_dirty(handle, di_bh);
 750
 751 out_commit:
 752         ocfs2_commit_trans(osb, handle);
 753 out:
 754         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 755         brelse(ref_root_bh);
 756
 757         return ret;
 758 }
 759
 760 int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh)
 761 {
 762         int ret, delete_tree = 0;
 763         handle_t *handle = NULL;
 764         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 765         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 766         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 767         struct ocfs2_refcount_block *rb;
 768         struct inode *alloc_inode = NULL;
 769         struct buffer_head *alloc_bh = NULL;
 770         struct buffer_head *blk_bh = NULL;
 771         struct ocfs2_refcount_tree *ref_tree;
 772         int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS;
 773         u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc);
 774         u16 bit = 0;
 775
 776         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
 777                 return 0;
 778
 779         BUG_ON(!ref_blkno);
 780         ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh);
 781         if (ret) {
 782                 mlog_errno(ret);
 783                 return ret;
 784         }
 785
 786         rb = (struct ocfs2_refcount_block *)blk_bh->b_data;
 787
 788         /*
 789          * If we are the last user, we need to free the block.
 790          * So lock the allocator ahead.
 791          */
 792         if (le32_to_cpu(rb->rf_count) == 1) {
 793                 blk = le64_to_cpu(rb->rf_blkno);
 794                 bit = le16_to_cpu(rb->rf_suballoc_bit);
 795                 if (rb->rf_suballoc_loc)
 796                         bg_blkno = le64_to_cpu(rb->rf_suballoc_loc);
 797                 else
 798                         bg_blkno = ocfs2_which_suballoc_group(blk, bit);
 799
 800                 alloc_inode = ocfs2_get_system_file_inode(osb,
 801                                         EXTENT_ALLOC_SYSTEM_INODE,
 802                                         le16_to_cpu(rb->rf_suballoc_slot));
 803                 if (!alloc_inode) {
 804                         ret = -ENOMEM;
 805                         mlog_errno(ret);
 806                         goto out;
 807                 }
 808                 mutex_lock(&alloc_inode->i_mutex);
 809
 810                 ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1);
 811                 if (ret) {
 812                         mlog_errno(ret);
 813                         goto out_mutex;
 814                 }
 815
 816                 credits += OCFS2_SUBALLOC_FREE;
 817         }
 818
 819         handle = ocfs2_start_trans(osb, credits);
 820         if (IS_ERR(handle)) {
 821                 ret = PTR_ERR(handle);
 822                 mlog_errno(ret);
 823                 goto out_unlock;
 824         }
 825
 826         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 827                                       OCFS2_JOURNAL_ACCESS_WRITE);
 828         if (ret) {
 829                 mlog_errno(ret);
 830                 goto out_commit;
 831         }
 832
 833         ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh,
 834                                       OCFS2_JOURNAL_ACCESS_WRITE);
 835         if (ret) {
 836                 mlog_errno(ret);
 837                 goto out_commit;
 838         }
 839
 840         spin_lock(&oi->ip_lock);
 841         oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL;
 842         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
 843         di->i_refcount_loc = 0;
 844         spin_unlock(&oi->ip_lock);
 845         ocfs2_journal_dirty(handle, di_bh);
 846
 847         le32_add_cpu(&rb->rf_count , -1);
 848         ocfs2_journal_dirty(handle, blk_bh);
 849
 850         if (!rb->rf_count) {
 851                 delete_tree = 1;
 852                 ocfs2_erase_refcount_tree_from_list(osb, ref_tree);
 853                 ret = ocfs2_free_suballoc_bits(handle, alloc_inode,
 854                                                alloc_bh, bit, bg_blkno, 1);
 855                 if (ret)
 856                         mlog_errno(ret);
 857         }
 858
 859 out_commit:
 860         ocfs2_commit_trans(osb, handle);
 861 out_unlock:
 862         if (alloc_inode) {
 863                 ocfs2_inode_unlock(alloc_inode, 1);
 864                 brelse(alloc_bh);
 865         }
 866 out_mutex:
 867         if (alloc_inode) {
 868                 mutex_unlock(&alloc_inode->i_mutex);
 869                 iput(alloc_inode);
 870         }
 871 out:
 872         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 873         if (delete_tree)
 874                 ocfs2_refcount_tree_put(ref_tree);
 875         brelse(blk_bh);
 876
 877         return ret;
 878 }
 879
 880 static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci,
 881                                           struct buffer_head *ref_leaf_bh,
 882                                           u64 cpos, unsigned int len,
 883                                           struct ocfs2_refcount_rec *ret_rec,
 884                                           int *index)
 885 {
 886         int i = 0;
 887         struct ocfs2_refcount_block *rb =
 888                 (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
 889         struct ocfs2_refcount_rec *rec = NULL;
 890
 891         for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) {
 892                 rec = &rb->rf_records.rl_recs[i];
 893
 894                 if (le64_to_cpu(rec->r_cpos) +
 895                     le32_to_cpu(rec->r_clusters) <= cpos)
 896                         continue;
 897                 else if (le64_to_cpu(rec->r_cpos) > cpos)
 898                         break;
 899
 900                 /* ok, cpos fail in this rec. Just return. */
 901                 if (ret_rec)
 902                         *ret_rec = *rec;
 903                 goto out;
 904         }
 905
 906         if (ret_rec) {
 907                 /* We meet with a hole here, so fake the rec. */
 908                 ret_rec->r_cpos = cpu_to_le64(cpos);
 909                 ret_rec->r_refcount = 0;
 910                 if (i < le16_to_cpu(rb->rf_records.rl_used) &&
 911                     le64_to_cpu(rec->r_cpos) < cpos + len)
 912                         ret_rec->r_clusters =
 913                                 cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos);
 914                 else
 915                         ret_rec->r_clusters = cpu_to_le32(len);
 916         }
 917
 918 out:
 919         *index = i;
 920 }
 921
 922 /*
 923  * Try to remove refcount tree. The mechanism is:
 924  * 1) Check whether i_clusters == 0, if no, exit.
 925  * 2) check whether we have i_xattr_loc in dinode. if yes, exit.
 926  * 3) Check whether we have inline xattr stored outside, if yes, exit.
 927  * 4) Remove the tree.
 928  */
 929 int ocfs2_try_remove_refcount_tree(struct inode *inode,
 930                                    struct buffer_head *di_bh)
 931 {
 932         int ret;
 933         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 934         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 935
 936         down_write(&oi->ip_xattr_sem);
 937         down_write(&oi->ip_alloc_sem);
 938
 939         if (oi->ip_clusters)
 940                 goto out;
 941
 942         if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc)
 943                 goto out;
 944
 945         if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL &&
 946             ocfs2_has_inline_xattr_value_outside(inode, di))
 947                 goto out;
 948
 949         ret = ocfs2_remove_refcount_tree(inode, di_bh);
 950         if (ret)
 951                 mlog_errno(ret);
 952 out:
 953         up_write(&oi->ip_alloc_sem);
 954         up_write(&oi->ip_xattr_sem);
 955         return 0;
 956 }
 957
 958 /*
 959  * Find the end range for a leaf refcount block indicated by
 960  * el->l_recs[index].e_blkno.
 961  */
 962 static int ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info *ci,
 963                                        struct buffer_head *ref_root_bh,
 964                                        struct ocfs2_extent_block *eb,
 965                                        struct ocfs2_extent_list *el,
 966                                        int index,  u32 *cpos_end)
 967 {
 968         int ret, i, subtree_root;
 969         u32 cpos;
 970         u64 blkno;
 971         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
 972         struct ocfs2_path *left_path = NULL, *right_path = NULL;
 973         struct ocfs2_extent_tree et;
 974         struct ocfs2_extent_list *tmp_el;
 975
 976         if (index < le16_to_cpu(el->l_next_free_rec) - 1) {
 977                 /*
 978                  * We have a extent rec after index, so just use the e_cpos
 979                  * of the next extent rec.
 980                  */
 981                 *cpos_end = le32_to_cpu(el->l_recs[index+1].e_cpos);
 982                 return 0;
 983         }
 984
 985         if (!eb || (eb && !eb->h_next_leaf_blk)) {
 986                 /*
 987                  * We are the last extent rec, so any high cpos should
 988                  * be stored in this leaf refcount block.
 989                  */
 990                 *cpos_end = UINT_MAX;
 991                 return 0;
 992         }
 993
 994         /*
 995          * If the extent block isn't the last one, we have to find
 996          * the subtree root between this extent block and the next
 997          * leaf extent block and get the corresponding e_cpos from
 998          * the subroot. Otherwise we may corrupt the b-tree.
 999          */
1000         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1001
1002         left_path = ocfs2_new_path_from_et(&et);
1003         if (!left_path) {
1004                 ret = -ENOMEM;
1005                 mlog_errno(ret);
1006                 goto out;
1007         }
1008
1009         cpos = le32_to_cpu(eb->h_list.l_recs[index].e_cpos);
1010         ret = ocfs2_find_path(ci, left_path, cpos);
1011         if (ret) {
1012                 mlog_errno(ret);
1013                 goto out;
1014         }
1015
1016         right_path = ocfs2_new_path_from_path(left_path);
1017         if (!right_path) {
1018                 ret = -ENOMEM;
1019                 mlog_errno(ret);
1020                 goto out;
1021         }
1022
1023         ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, &cpos);
1024         if (ret) {
1025                 mlog_errno(ret);
1026                 goto out;
1027         }
1028
1029         ret = ocfs2_find_path(ci, right_path, cpos);
1030         if (ret) {
1031                 mlog_errno(ret);
1032                 goto out;
1033         }
1034
1035         subtree_root = ocfs2_find_subtree_root(&et, left_path,
1036                                                right_path);
1037
1038         tmp_el = left_path->p_node[subtree_root].el;
1039         blkno = left_path->p_node[subtree_root+1].bh->b_blocknr;
1040         for (i = 0; i < le32_to_cpu(tmp_el->l_next_free_rec); i++) {
1041                 if (le64_to_cpu(tmp_el->l_recs[i].e_blkno) == blkno) {
1042                         *cpos_end = le32_to_cpu(tmp_el->l_recs[i+1].e_cpos);
1043                         break;
1044                 }
1045         }
1046
1047         BUG_ON(i == le32_to_cpu(tmp_el->l_next_free_rec));
1048
1049 out:
1050         ocfs2_free_path(left_path);
1051         ocfs2_free_path(right_path);
1052         return ret;
1053 }
1054
1055 /*
1056  * Given a cpos and len, try to find the refcount record which contains cpos.
1057  * 1. If cpos can be found in one refcount record, return the record.
1058  * 2. If cpos can't be found, return a fake record which start from cpos
1059  *    and end at a small value between cpos+len and start of the next record.
1060  *    This fake record has r_refcount = 0.
1061  */
1062 static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci,
1063                                   struct buffer_head *ref_root_bh,
1064                                   u64 cpos, unsigned int len,
1065                                   struct ocfs2_refcount_rec *ret_rec,
1066                                   int *index,
1067                                   struct buffer_head **ret_bh)
1068 {
1069         int ret = 0, i, found;
1070         u32 low_cpos, uninitialized_var(cpos_end);
1071         struct ocfs2_extent_list *el;
1072         struct ocfs2_extent_rec *rec = NULL;
1073         struct ocfs2_extent_block *eb = NULL;
1074         struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL;
1075         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1076         struct ocfs2_refcount_block *rb =
1077                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1078
1079         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) {
1080                 ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len,
1081                                               ret_rec, index);
1082                 *ret_bh = ref_root_bh;
1083                 get_bh(ref_root_bh);
1084                 return 0;
1085         }
1086
1087         el = &rb->rf_list;
1088         low_cpos = cpos & OCFS2_32BIT_POS_MASK;
1089
1090         if (el->l_tree_depth) {
1091                 ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh);
1092                 if (ret) {
1093                         mlog_errno(ret);
1094                         goto out;
1095                 }
1096
1097                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1098                 el = &eb->h_list;
1099
1100                 if (el->l_tree_depth) {
1101                         ocfs2_error(sb,
1102                         "refcount tree %llu has non zero tree "
1103                         "depth in leaf btree tree block %llu\n",
1104                         (unsigned long long)ocfs2_metadata_cache_owner(ci),
1105                         (unsigned long long)eb_bh->b_blocknr);
1106                         ret = -EROFS;
1107                         goto out;
1108                 }
1109         }
1110
1111         found = 0;
1112         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1113                 rec = &el->l_recs[i];
1114
1115                 if (le32_to_cpu(rec->e_cpos) <= low_cpos) {
1116                         found = 1;
1117                         break;
1118                 }
1119         }
1120
1121         if (found) {
1122                 ret = ocfs2_get_refcount_cpos_end(ci, ref_root_bh,
1123                                                   eb, el, i, &cpos_end);
1124                 if (ret) {
1125                         mlog_errno(ret);
1126                         goto out;
1127                 }
1128
1129                 if (cpos_end < low_cpos + len)
1130                         len = cpos_end - low_cpos;
1131         }
1132
1133         ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno),
1134                                         &ref_leaf_bh);
1135         if (ret) {
1136                 mlog_errno(ret);
1137                 goto out;
1138         }
1139
1140         ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len,
1141                                       ret_rec, index);
1142         *ret_bh = ref_leaf_bh;
1143 out:
1144         brelse(eb_bh);
1145         return ret;
1146 }
1147
1148 enum ocfs2_ref_rec_contig {
1149         REF_CONTIG_NONE = 0,
1150         REF_CONTIG_LEFT,
1151         REF_CONTIG_RIGHT,
1152         REF_CONTIG_LEFTRIGHT,
1153 };
1154
1155 static enum ocfs2_ref_rec_contig
1156         ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb,
1157                                     int index)
1158 {
1159         if ((rb->rf_records.rl_recs[index].r_refcount ==
1160             rb->rf_records.rl_recs[index + 1].r_refcount) &&
1161             (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) +
1162             le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) ==
1163             le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos)))
1164                 return REF_CONTIG_RIGHT;
1165
1166         return REF_CONTIG_NONE;
1167 }
1168
1169 static enum ocfs2_ref_rec_contig
1170         ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb,
1171                                   int index)
1172 {
1173         enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE;
1174
1175         if (index < le16_to_cpu(rb->rf_records.rl_used) - 1)
1176                 ret = ocfs2_refcount_rec_adjacent(rb, index);
1177
1178         if (index > 0) {
1179                 enum ocfs2_ref_rec_contig tmp;
1180
1181                 tmp = ocfs2_refcount_rec_adjacent(rb, index - 1);
1182
1183                 if (tmp == REF_CONTIG_RIGHT) {
1184                         if (ret == REF_CONTIG_RIGHT)
1185                                 ret = REF_CONTIG_LEFTRIGHT;
1186                         else
1187                                 ret = REF_CONTIG_LEFT;
1188                 }
1189         }
1190
1191         return ret;
1192 }
1193
1194 static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb,
1195                                            int index)
1196 {
1197         BUG_ON(rb->rf_records.rl_recs[index].r_refcount !=
1198                rb->rf_records.rl_recs[index+1].r_refcount);
1199
1200         le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters,
1201                      le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters));
1202
1203         if (index < le16_to_cpu(rb->rf_records.rl_used) - 2)
1204                 memmove(&rb->rf_records.rl_recs[index + 1],
1205                         &rb->rf_records.rl_recs[index + 2],
1206                         sizeof(struct ocfs2_refcount_rec) *
1207                         (le16_to_cpu(rb->rf_records.rl_used) - index - 2));
1208
1209         memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1],
1210                0, sizeof(struct ocfs2_refcount_rec));
1211         le16_add_cpu(&rb->rf_records.rl_used, -1);
1212 }
1213
1214 /*
1215  * Merge the refcount rec if we are contiguous with the adjacent recs.
1216  */
1217 static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb,
1218                                      int index)
1219 {
1220         enum ocfs2_ref_rec_contig contig =
1221                                 ocfs2_refcount_rec_contig(rb, index);
1222
1223         if (contig == REF_CONTIG_NONE)
1224                 return;
1225
1226         if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) {
1227                 BUG_ON(index == 0);
1228                 index--;
1229         }
1230
1231         ocfs2_rotate_refcount_rec_left(rb, index);
1232
1233         if (contig == REF_CONTIG_LEFTRIGHT)
1234                 ocfs2_rotate_refcount_rec_left(rb, index);
1235 }
1236
1237 /*
1238  * Change the refcount indexed by "index" in ref_bh.
1239  * If refcount reaches 0, remove it.
1240  */
1241 static int ocfs2_change_refcount_rec(handle_t *handle,
1242                                      struct ocfs2_caching_info *ci,
1243                                      struct buffer_head *ref_leaf_bh,
1244                                      int index, int merge, int change)
1245 {
1246         int ret;
1247         struct ocfs2_refcount_block *rb =
1248                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1249         struct ocfs2_refcount_list *rl = &rb->rf_records;
1250         struct ocfs2_refcount_rec *rec = &rl->rl_recs[index];
1251
1252         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1253                                       OCFS2_JOURNAL_ACCESS_WRITE);
1254         if (ret) {
1255                 mlog_errno(ret);
1256                 goto out;
1257         }
1258
1259         mlog(0, "change index %d, old count %u, change %d\n", index,
1260              le32_to_cpu(rec->r_refcount), change);
1261         le32_add_cpu(&rec->r_refcount, change);
1262
1263         if (!rec->r_refcount) {
1264                 if (index != le16_to_cpu(rl->rl_used) - 1) {
1265                         memmove(rec, rec + 1,
1266                                 (le16_to_cpu(rl->rl_used) - index - 1) *
1267                                 sizeof(struct ocfs2_refcount_rec));
1268                         memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1],
1269                                0, sizeof(struct ocfs2_refcount_rec));
1270                 }
1271
1272                 le16_add_cpu(&rl->rl_used, -1);
1273         } else if (merge)
1274                 ocfs2_refcount_rec_merge(rb, index);
1275
1276         ocfs2_journal_dirty(handle, ref_leaf_bh);
1277 out:
1278         return ret;
1279 }
1280
1281 static int ocfs2_expand_inline_ref_root(handle_t *handle,
1282                                         struct ocfs2_caching_info *ci,
1283                                         struct buffer_head *ref_root_bh,
1284                                         struct buffer_head **ref_leaf_bh,
1285                                         struct ocfs2_alloc_context *meta_ac)
1286 {
1287         int ret;
1288         u16 suballoc_bit_start;
1289         u32 num_got;
1290         u64 suballoc_loc, blkno;
1291         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1292         struct buffer_head *new_bh = NULL;
1293         struct ocfs2_refcount_block *new_rb;
1294         struct ocfs2_refcount_block *root_rb =
1295                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1296
1297         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1298                                       OCFS2_JOURNAL_ACCESS_WRITE);
1299         if (ret) {
1300                 mlog_errno(ret);
1301                 goto out;
1302         }
1303
1304         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
1305                                    &suballoc_bit_start, &num_got,
1306                                    &blkno);
1307         if (ret) {
1308                 mlog_errno(ret);
1309                 goto out;
1310         }
1311
1312         new_bh = sb_getblk(sb, blkno);
1313         if (new_bh == NULL) {
1314                 ret = -EIO;
1315                 mlog_errno(ret);
1316                 goto out;
1317         }
1318         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1319
1320         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1321                                       OCFS2_JOURNAL_ACCESS_CREATE);
1322         if (ret) {
1323                 mlog_errno(ret);
1324                 goto out;
1325         }
1326
1327         /*
1328          * Initialize ocfs2_refcount_block.
1329          * It should contain the same information as the old root.
1330          * so just memcpy it and change the corresponding field.
1331          */
1332         memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize);
1333
1334         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1335         new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
1336         new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
1337         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1338         new_rb->rf_blkno = cpu_to_le64(blkno);
1339         new_rb->rf_cpos = cpu_to_le32(0);
1340         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1341         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1342         ocfs2_journal_dirty(handle, new_bh);
1343
1344         /* Now change the root. */
1345         memset(&root_rb->rf_list, 0, sb->s_blocksize -
1346                offsetof(struct ocfs2_refcount_block, rf_list));
1347         root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb));
1348         root_rb->rf_clusters = cpu_to_le32(1);
1349         root_rb->rf_list.l_next_free_rec = cpu_to_le16(1);
1350         root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno);
1351         root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1);
1352         root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL);
1353
1354         ocfs2_journal_dirty(handle, ref_root_bh);
1355
1356         mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno,
1357              le16_to_cpu(new_rb->rf_records.rl_used));
1358
1359         *ref_leaf_bh = new_bh;
1360         new_bh = NULL;
1361 out:
1362         brelse(new_bh);
1363         return ret;
1364 }
1365
1366 static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev,
1367                                            struct ocfs2_refcount_rec *next)
1368 {
1369         if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <=
1370                 ocfs2_get_ref_rec_low_cpos(next))
1371                 return 1;
1372
1373         return 0;
1374 }
1375
1376 static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b)
1377 {
1378         const struct ocfs2_refcount_rec *l = a, *r = b;
1379         u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l);
1380         u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r);
1381
1382         if (l_cpos > r_cpos)
1383                 return 1;
1384         if (l_cpos < r_cpos)
1385                 return -1;
1386         return 0;
1387 }
1388
1389 static int cmp_refcount_rec_by_cpos(const void *a, const void *b)
1390 {
1391         const struct ocfs2_refcount_rec *l = a, *r = b;
1392         u64 l_cpos = le64_to_cpu(l->r_cpos);
1393         u64 r_cpos = le64_to_cpu(r->r_cpos);
1394
1395         if (l_cpos > r_cpos)
1396                 return 1;
1397         if (l_cpos < r_cpos)
1398                 return -1;
1399         return 0;
1400 }
1401
1402 static void swap_refcount_rec(void *a, void *b, int size)
1403 {
1404         struct ocfs2_refcount_rec *l = a, *r = b, tmp;
1405
1406         tmp = *(struct ocfs2_refcount_rec *)l;
1407         *(struct ocfs2_refcount_rec *)l =
1408                         *(struct ocfs2_refcount_rec *)r;
1409         *(struct ocfs2_refcount_rec *)r = tmp;
1410 }
1411
1412 /*
1413  * The refcount cpos are ordered by their 64bit cpos,
1414  * But we will use the low 32 bit to be the e_cpos in the b-tree.
1415  * So we need to make sure that this pos isn't intersected with others.
1416  *
1417  * Note: The refcount block is already sorted by their low 32 bit cpos,
1418  *       So just try the middle pos first, and we will exit when we find
1419  *       the good position.
1420  */
1421 static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl,
1422                                          u32 *split_pos, int *split_index)
1423 {
1424         int num_used = le16_to_cpu(rl->rl_used);
1425         int delta, middle = num_used / 2;
1426
1427         for (delta = 0; delta < middle; delta++) {
1428                 /* Let's check delta earlier than middle */
1429                 if (ocfs2_refcount_rec_no_intersect(
1430                                         &rl->rl_recs[middle - delta - 1],
1431                                         &rl->rl_recs[middle - delta])) {
1432                         *split_index = middle - delta;
1433                         break;
1434                 }
1435
1436                 /* For even counts, don't walk off the end */
1437                 if ((middle + delta + 1) == num_used)
1438                         continue;
1439
1440                 /* Now try delta past middle */
1441                 if (ocfs2_refcount_rec_no_intersect(
1442                                         &rl->rl_recs[middle + delta],
1443                                         &rl->rl_recs[middle + delta + 1])) {
1444                         *split_index = middle + delta + 1;
1445                         break;
1446                 }
1447         }
1448
1449         if (delta >= middle)
1450                 return -ENOSPC;
1451
1452         *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]);
1453         return 0;
1454 }
1455
1456 static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh,
1457                                             struct buffer_head *new_bh,
1458                                             u32 *split_cpos)
1459 {
1460         int split_index = 0, num_moved, ret;
1461         u32 cpos = 0;
1462         struct ocfs2_refcount_block *rb =
1463                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1464         struct ocfs2_refcount_list *rl = &rb->rf_records;
1465         struct ocfs2_refcount_block *new_rb =
1466                         (struct ocfs2_refcount_block *)new_bh->b_data;
1467         struct ocfs2_refcount_list *new_rl = &new_rb->rf_records;
1468
1469         mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n",
1470              (unsigned long long)ref_leaf_bh->b_blocknr,
1471              le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used));
1472
1473         /*
1474          * XXX: Improvement later.
1475          * If we know all the high 32 bit cpos is the same, no need to sort.
1476          *
1477          * In order to make the whole process safe, we do:
1478          * 1. sort the entries by their low 32 bit cpos first so that we can
1479          *    find the split cpos easily.
1480          * 2. call ocfs2_insert_extent to insert the new refcount block.
1481          * 3. move the refcount rec to the new block.
1482          * 4. sort the entries by their 64 bit cpos.
1483          * 5. dirty the new_rb and rb.
1484          */
1485         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1486              sizeof(struct ocfs2_refcount_rec),
1487              cmp_refcount_rec_by_low_cpos, swap_refcount_rec);
1488
1489         ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index);
1490         if (ret) {
1491                 mlog_errno(ret);
1492                 return ret;
1493         }
1494
1495         new_rb->rf_cpos = cpu_to_le32(cpos);
1496
1497         /* move refcount records starting from split_index to the new block. */
1498         num_moved = le16_to_cpu(rl->rl_used) - split_index;
1499         memcpy(new_rl->rl_recs, &rl->rl_recs[split_index],
1500                num_moved * sizeof(struct ocfs2_refcount_rec));
1501
1502         /*ok, remove the entries we just moved over to the other block. */
1503         memset(&rl->rl_recs[split_index], 0,
1504                num_moved * sizeof(struct ocfs2_refcount_rec));
1505
1506         /* change old and new rl_used accordingly. */
1507         le16_add_cpu(&rl->rl_used, -num_moved);
1508         new_rl->rl_used = cpu_to_le16(num_moved);
1509
1510         sort(&rl->rl_recs, le16_to_cpu(rl->rl_used),
1511              sizeof(struct ocfs2_refcount_rec),
1512              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1513
1514         sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used),
1515              sizeof(struct ocfs2_refcount_rec),
1516              cmp_refcount_rec_by_cpos, swap_refcount_rec);
1517
1518         *split_cpos = cpos;
1519         return 0;
1520 }
1521
1522 static int ocfs2_new_leaf_refcount_block(handle_t *handle,
1523                                          struct ocfs2_caching_info *ci,
1524                                          struct buffer_head *ref_root_bh,
1525                                          struct buffer_head *ref_leaf_bh,
1526                                          struct ocfs2_alloc_context *meta_ac)
1527 {
1528         int ret;
1529         u16 suballoc_bit_start;
1530         u32 num_got, new_cpos;
1531         u64 suballoc_loc, blkno;
1532         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
1533         struct ocfs2_refcount_block *root_rb =
1534                         (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1535         struct buffer_head *new_bh = NULL;
1536         struct ocfs2_refcount_block *new_rb;
1537         struct ocfs2_extent_tree ref_et;
1538
1539         BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL));
1540
1541         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
1542                                       OCFS2_JOURNAL_ACCESS_WRITE);
1543         if (ret) {
1544                 mlog_errno(ret);
1545                 goto out;
1546         }
1547
1548         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1549                                       OCFS2_JOURNAL_ACCESS_WRITE);
1550         if (ret) {
1551                 mlog_errno(ret);
1552                 goto out;
1553         }
1554
1555         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
1556                                    &suballoc_bit_start, &num_got,
1557                                    &blkno);
1558         if (ret) {
1559                 mlog_errno(ret);
1560                 goto out;
1561         }
1562
1563         new_bh = sb_getblk(sb, blkno);
1564         if (new_bh == NULL) {
1565                 ret = -EIO;
1566                 mlog_errno(ret);
1567                 goto out;
1568         }
1569         ocfs2_set_new_buffer_uptodate(ci, new_bh);
1570
1571         ret = ocfs2_journal_access_rb(handle, ci, new_bh,
1572                                       OCFS2_JOURNAL_ACCESS_CREATE);
1573         if (ret) {
1574                 mlog_errno(ret);
1575                 goto out;
1576         }
1577
1578         /* Initialize ocfs2_refcount_block. */
1579         new_rb = (struct ocfs2_refcount_block *)new_bh->b_data;
1580         memset(new_rb, 0, sb->s_blocksize);
1581         strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE);
1582         new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
1583         new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc);
1584         new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1585         new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
1586         new_rb->rf_blkno = cpu_to_le64(blkno);
1587         new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr);
1588         new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL);
1589         new_rb->rf_records.rl_count =
1590                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
1591         new_rb->rf_generation = root_rb->rf_generation;
1592
1593         ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos);
1594         if (ret) {
1595                 mlog_errno(ret);
1596                 goto out;
1597         }
1598
1599         ocfs2_journal_dirty(handle, ref_leaf_bh);
1600         ocfs2_journal_dirty(handle, new_bh);
1601
1602         ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh);
1603
1604         mlog(0, "insert new leaf block %llu at %u\n",
1605              (unsigned long long)new_bh->b_blocknr, new_cpos);
1606
1607         /* Insert the new leaf block with the specific offset cpos. */
1608         ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr,
1609                                   1, 0, meta_ac);
1610         if (ret)
1611                 mlog_errno(ret);
1612
1613 out:
1614         brelse(new_bh);
1615         return ret;
1616 }
1617
1618 static int ocfs2_expand_refcount_tree(handle_t *handle,
1619                                       struct ocfs2_caching_info *ci,
1620                                       struct buffer_head *ref_root_bh,
1621                                       struct buffer_head *ref_leaf_bh,
1622                                       struct ocfs2_alloc_context *meta_ac)
1623 {
1624         int ret;
1625         struct buffer_head *expand_bh = NULL;
1626
1627         if (ref_root_bh == ref_leaf_bh) {
1628                 /*
1629                  * the old root bh hasn't been expanded to a b-tree,
1630                  * so expand it first.
1631                  */
1632                 ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh,
1633                                                    &expand_bh, meta_ac);
1634                 if (ret) {
1635                         mlog_errno(ret);
1636                         goto out;
1637                 }
1638         } else {
1639                 expand_bh = ref_leaf_bh;
1640                 get_bh(expand_bh);
1641         }
1642
1643
1644         /* Now add a new refcount block into the tree.*/
1645         ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh,
1646                                             expand_bh, meta_ac);
1647         if (ret)
1648                 mlog_errno(ret);
1649 out:
1650         brelse(expand_bh);
1651         return ret;
1652 }
1653
1654 /*
1655  * Adjust the extent rec in b-tree representing ref_leaf_bh.
1656  *
1657  * Only called when we have inserted a new refcount rec at index 0
1658  * which means ocfs2_extent_rec.e_cpos may need some change.
1659  */
1660 static int ocfs2_adjust_refcount_rec(handle_t *handle,
1661                                      struct ocfs2_caching_info *ci,
1662                                      struct buffer_head *ref_root_bh,
1663                                      struct buffer_head *ref_leaf_bh,
1664                                      struct ocfs2_refcount_rec *rec)
1665 {
1666         int ret = 0, i;
1667         u32 new_cpos, old_cpos;
1668         struct ocfs2_path *path = NULL;
1669         struct ocfs2_extent_tree et;
1670         struct ocfs2_refcount_block *rb =
1671                 (struct ocfs2_refcount_block *)ref_root_bh->b_data;
1672         struct ocfs2_extent_list *el;
1673
1674         if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL))
1675                 goto out;
1676
1677         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1678         old_cpos = le32_to_cpu(rb->rf_cpos);
1679         new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK;
1680         if (old_cpos <= new_cpos)
1681                 goto out;
1682
1683         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
1684
1685         path = ocfs2_new_path_from_et(&et);
1686         if (!path) {
1687                 ret = -ENOMEM;
1688                 mlog_errno(ret);
1689                 goto out;
1690         }
1691
1692         ret = ocfs2_find_path(ci, path, old_cpos);
1693         if (ret) {
1694                 mlog_errno(ret);
1695                 goto out;
1696         }
1697
1698         /*
1699          * 2 more credits, one for the leaf refcount block, one for
1700          * the extent block contains the extent rec.
1701          */
1702         ret = ocfs2_extend_trans(handle, 2);
1703         if (ret < 0) {
1704                 mlog_errno(ret);
1705                 goto out;
1706         }
1707
1708         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1709                                       OCFS2_JOURNAL_ACCESS_WRITE);
1710         if (ret < 0) {
1711                 mlog_errno(ret);
1712                 goto out;
1713         }
1714
1715         ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path),
1716                                       OCFS2_JOURNAL_ACCESS_WRITE);
1717         if (ret < 0) {
1718                 mlog_errno(ret);
1719                 goto out;
1720         }
1721
1722         /* change the leaf extent block first. */
1723         el = path_leaf_el(path);
1724
1725         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++)
1726                 if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos)
1727                         break;
1728
1729         BUG_ON(i == le16_to_cpu(el->l_next_free_rec));
1730
1731         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1732
1733         /* change the r_cpos in the leaf block. */
1734         rb->rf_cpos = cpu_to_le32(new_cpos);
1735
1736         ocfs2_journal_dirty(handle, path_leaf_bh(path));
1737         ocfs2_journal_dirty(handle, ref_leaf_bh);
1738
1739 out:
1740         ocfs2_free_path(path);
1741         return ret;
1742 }
1743
1744 static int ocfs2_insert_refcount_rec(handle_t *handle,
1745                                      struct ocfs2_caching_info *ci,
1746                                      struct buffer_head *ref_root_bh,
1747                                      struct buffer_head *ref_leaf_bh,
1748                                      struct ocfs2_refcount_rec *rec,
1749                                      int index, int merge,
1750                                      struct ocfs2_alloc_context *meta_ac)
1751 {
1752         int ret;
1753         struct ocfs2_refcount_block *rb =
1754                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1755         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1756         struct buffer_head *new_bh = NULL;
1757
1758         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1759
1760         if (rf_list->rl_used == rf_list->rl_count) {
1761                 u64 cpos = le64_to_cpu(rec->r_cpos);
1762                 u32 len = le32_to_cpu(rec->r_clusters);
1763
1764                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1765                                                  ref_leaf_bh, meta_ac);
1766                 if (ret) {
1767                         mlog_errno(ret);
1768                         goto out;
1769                 }
1770
1771                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1772                                              cpos, len, NULL, &index,
1773                                              &new_bh);
1774                 if (ret) {
1775                         mlog_errno(ret);
1776                         goto out;
1777                 }
1778
1779                 ref_leaf_bh = new_bh;
1780                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1781                 rf_list = &rb->rf_records;
1782         }
1783
1784         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1785                                       OCFS2_JOURNAL_ACCESS_WRITE);
1786         if (ret) {
1787                 mlog_errno(ret);
1788                 goto out;
1789         }
1790
1791         if (index < le16_to_cpu(rf_list->rl_used))
1792                 memmove(&rf_list->rl_recs[index + 1],
1793                         &rf_list->rl_recs[index],
1794                         (le16_to_cpu(rf_list->rl_used) - index) *
1795                          sizeof(struct ocfs2_refcount_rec));
1796
1797         mlog(0, "insert refcount record start %llu, len %u, count %u "
1798              "to leaf block %llu at index %d\n",
1799              (unsigned long long)le64_to_cpu(rec->r_cpos),
1800              le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount),
1801              (unsigned long long)ref_leaf_bh->b_blocknr, index);
1802
1803         rf_list->rl_recs[index] = *rec;
1804
1805         le16_add_cpu(&rf_list->rl_used, 1);
1806
1807         if (merge)
1808                 ocfs2_refcount_rec_merge(rb, index);
1809
1810         ocfs2_journal_dirty(handle, ref_leaf_bh);
1811
1812         if (index == 0) {
1813                 ret = ocfs2_adjust_refcount_rec(handle, ci,
1814                                                 ref_root_bh,
1815                                                 ref_leaf_bh, rec);
1816                 if (ret)
1817                         mlog_errno(ret);
1818         }
1819 out:
1820         brelse(new_bh);
1821         return ret;
1822 }
1823
1824 /*
1825  * Split the refcount_rec indexed by "index" in ref_leaf_bh.
1826  * This is much simple than our b-tree code.
1827  * split_rec is the new refcount rec we want to insert.
1828  * If split_rec->r_refcount > 0, we are changing the refcount(in case we
1829  * increase refcount or decrease a refcount to non-zero).
1830  * If split_rec->r_refcount == 0, we are punching a hole in current refcount
1831  * rec( in case we decrease a refcount to zero).
1832  */
1833 static int ocfs2_split_refcount_rec(handle_t *handle,
1834                                     struct ocfs2_caching_info *ci,
1835                                     struct buffer_head *ref_root_bh,
1836                                     struct buffer_head *ref_leaf_bh,
1837                                     struct ocfs2_refcount_rec *split_rec,
1838                                     int index, int merge,
1839                                     struct ocfs2_alloc_context *meta_ac,
1840                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
1841 {
1842         int ret, recs_need;
1843         u32 len;
1844         struct ocfs2_refcount_block *rb =
1845                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1846         struct ocfs2_refcount_list *rf_list = &rb->rf_records;
1847         struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index];
1848         struct ocfs2_refcount_rec *tail_rec = NULL;
1849         struct buffer_head *new_bh = NULL;
1850
1851         BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL);
1852
1853         mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n",
1854              le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters),
1855              le64_to_cpu(split_rec->r_cpos),
1856              le32_to_cpu(split_rec->r_clusters));
1857
1858         /*
1859          * If we just need to split the header or tail clusters,
1860          * no more recs are needed, just split is OK.
1861          * Otherwise we at least need one new recs.
1862          */
1863         if (!split_rec->r_refcount &&
1864             (split_rec->r_cpos == orig_rec->r_cpos ||
1865              le64_to_cpu(split_rec->r_cpos) +
1866              le32_to_cpu(split_rec->r_clusters) ==
1867              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1868                 recs_need = 0;
1869         else
1870                 recs_need = 1;
1871
1872         /*
1873          * We need one more rec if we split in the middle and the new rec have
1874          * some refcount in it.
1875          */
1876         if (split_rec->r_refcount &&
1877             (split_rec->r_cpos != orig_rec->r_cpos &&
1878              le64_to_cpu(split_rec->r_cpos) +
1879              le32_to_cpu(split_rec->r_clusters) !=
1880              le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters)))
1881                 recs_need++;
1882
1883         /* If the leaf block don't have enough record, expand it. */
1884         if (le16_to_cpu(rf_list->rl_used) + recs_need >
1885                                          le16_to_cpu(rf_list->rl_count)) {
1886                 struct ocfs2_refcount_rec tmp_rec;
1887                 u64 cpos = le64_to_cpu(orig_rec->r_cpos);
1888                 len = le32_to_cpu(orig_rec->r_clusters);
1889                 ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh,
1890                                                  ref_leaf_bh, meta_ac);
1891                 if (ret) {
1892                         mlog_errno(ret);
1893                         goto out;
1894                 }
1895
1896                 /*
1897                  * We have to re-get it since now cpos may be moved to
1898                  * another leaf block.
1899                  */
1900                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
1901                                              cpos, len, &tmp_rec, &index,
1902                                              &new_bh);
1903                 if (ret) {
1904                         mlog_errno(ret);
1905                         goto out;
1906                 }
1907
1908                 ref_leaf_bh = new_bh;
1909                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
1910                 rf_list = &rb->rf_records;
1911                 orig_rec = &rf_list->rl_recs[index];
1912         }
1913
1914         ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh,
1915                                       OCFS2_JOURNAL_ACCESS_WRITE);
1916         if (ret) {
1917                 mlog_errno(ret);
1918                 goto out;
1919         }
1920
1921         /*
1922          * We have calculated out how many new records we need and store
1923          * in recs_need, so spare enough space first by moving the records
1924          * after "index" to the end.
1925          */
1926         if (index != le16_to_cpu(rf_list->rl_used) - 1)
1927                 memmove(&rf_list->rl_recs[index + 1 + recs_need],
1928                         &rf_list->rl_recs[index + 1],
1929                         (le16_to_cpu(rf_list->rl_used) - index - 1) *
1930                          sizeof(struct ocfs2_refcount_rec));
1931
1932         len = (le64_to_cpu(orig_rec->r_cpos) +
1933               le32_to_cpu(orig_rec->r_clusters)) -
1934               (le64_to_cpu(split_rec->r_cpos) +
1935               le32_to_cpu(split_rec->r_clusters));
1936
1937         /*
1938          * If we have "len", the we will split in the tail and move it
1939          * to the end of the space we have just spared.
1940          */
1941         if (len) {
1942                 tail_rec = &rf_list->rl_recs[index + recs_need];
1943
1944                 memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec));
1945                 le64_add_cpu(&tail_rec->r_cpos,
1946                              le32_to_cpu(tail_rec->r_clusters) - len);
1947                 tail_rec->r_clusters = cpu_to_le32(len);
1948         }
1949
1950         /*
1951          * If the split pos isn't the same as the original one, we need to
1952          * split in the head.
1953          *
1954          * Note: We have the chance that split_rec.r_refcount = 0,
1955          * recs_need = 0 and len > 0, which means we just cut the head from
1956          * the orig_rec and in that case we have done some modification in
1957          * orig_rec above, so the check for r_cpos is faked.
1958          */
1959         if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) {
1960                 len = le64_to_cpu(split_rec->r_cpos) -
1961                       le64_to_cpu(orig_rec->r_cpos);
1962                 orig_rec->r_clusters = cpu_to_le32(len);
1963                 index++;
1964         }
1965
1966         le16_add_cpu(&rf_list->rl_used, recs_need);
1967
1968         if (split_rec->r_refcount) {
1969                 rf_list->rl_recs[index] = *split_rec;
1970                 mlog(0, "insert refcount record start %llu, len %u, count %u "
1971                      "to leaf block %llu at index %d\n",
1972                      (unsigned long long)le64_to_cpu(split_rec->r_cpos),
1973                      le32_to_cpu(split_rec->r_clusters),
1974                      le32_to_cpu(split_rec->r_refcount),
1975                      (unsigned long long)ref_leaf_bh->b_blocknr, index);
1976
1977                 if (merge)
1978                         ocfs2_refcount_rec_merge(rb, index);
1979         }
1980
1981         ocfs2_journal_dirty(handle, ref_leaf_bh);
1982
1983 out:
1984         brelse(new_bh);
1985         return ret;
1986 }
1987
1988 static int __ocfs2_increase_refcount(handle_t *handle,
1989                                      struct ocfs2_caching_info *ci,
1990                                      struct buffer_head *ref_root_bh,
1991                                      u64 cpos, u32 len, int merge,
1992                                      struct ocfs2_alloc_context *meta_ac,
1993                                      struct ocfs2_cached_dealloc_ctxt *dealloc)
1994 {
1995         int ret = 0, index;
1996         struct buffer_head *ref_leaf_bh = NULL;
1997         struct ocfs2_refcount_rec rec;
1998         unsigned int set_len = 0;
1999
2000         mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n",
2001              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2002              (unsigned long long)cpos, len);
2003
2004         while (len) {
2005                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2006                                              cpos, len, &rec, &index,
2007                                              &ref_leaf_bh);
2008                 if (ret) {
2009                         mlog_errno(ret);
2010                         goto out;
2011                 }
2012
2013                 set_len = le32_to_cpu(rec.r_clusters);
2014
2015                 /*
2016                  * Here we may meet with 3 situations:
2017                  *
2018                  * 1. If we find an already existing record, and the length
2019                  *    is the same, cool, we just need to increase the r_refcount
2020                  *    and it is OK.
2021                  * 2. If we find a hole, just insert it with r_refcount = 1.
2022                  * 3. If we are in the middle of one extent record, split
2023                  *    it.
2024                  */
2025                 if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos &&
2026                     set_len <= len) {
2027                         mlog(0, "increase refcount rec, start %llu, len %u, "
2028                              "count %u\n", (unsigned long long)cpos, set_len,
2029                              le32_to_cpu(rec.r_refcount));
2030                         ret = ocfs2_change_refcount_rec(handle, ci,
2031                                                         ref_leaf_bh, index,
2032                                                         merge, 1);
2033                         if (ret) {
2034                                 mlog_errno(ret);
2035                                 goto out;
2036                         }
2037                 } else if (!rec.r_refcount) {
2038                         rec.r_refcount = cpu_to_le32(1);
2039
2040                         mlog(0, "insert refcount rec, start %llu, len %u\n",
2041                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2042                              set_len);
2043                         ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh,
2044                                                         ref_leaf_bh,
2045                                                         &rec, index,
2046                                                         merge, meta_ac);
2047                         if (ret) {
2048                                 mlog_errno(ret);
2049                                 goto out;
2050                         }
2051                 } else  {
2052                         set_len = min((u64)(cpos + len),
2053                                       le64_to_cpu(rec.r_cpos) + set_len) - cpos;
2054                         rec.r_cpos = cpu_to_le64(cpos);
2055                         rec.r_clusters = cpu_to_le32(set_len);
2056                         le32_add_cpu(&rec.r_refcount, 1);
2057
2058                         mlog(0, "split refcount rec, start %llu, "
2059                              "len %u, count %u\n",
2060                              (unsigned long long)le64_to_cpu(rec.r_cpos),
2061                              set_len, le32_to_cpu(rec.r_refcount));
2062                         ret = ocfs2_split_refcount_rec(handle, ci,
2063                                                        ref_root_bh, ref_leaf_bh,
2064                                                        &rec, index, merge,
2065                                                        meta_ac, dealloc);
2066                         if (ret) {
2067                                 mlog_errno(ret);
2068                                 goto out;
2069                         }
2070                 }
2071
2072                 cpos += set_len;
2073                 len -= set_len;
2074                 brelse(ref_leaf_bh);
2075                 ref_leaf_bh = NULL;
2076         }
2077
2078 out:
2079         brelse(ref_leaf_bh);
2080         return ret;
2081 }
2082
2083 static int ocfs2_remove_refcount_extent(handle_t *handle,
2084                                 struct ocfs2_caching_info *ci,
2085                                 struct buffer_head *ref_root_bh,
2086                                 struct buffer_head *ref_leaf_bh,
2087                                 struct ocfs2_alloc_context *meta_ac,
2088                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2089 {
2090         int ret;
2091         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2092         struct ocfs2_refcount_block *rb =
2093                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2094         struct ocfs2_extent_tree et;
2095
2096         BUG_ON(rb->rf_records.rl_used);
2097
2098         ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2099         ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos),
2100                                   1, meta_ac, dealloc);
2101         if (ret) {
2102                 mlog_errno(ret);
2103                 goto out;
2104         }
2105
2106         ocfs2_remove_from_cache(ci, ref_leaf_bh);
2107
2108         /*
2109          * add the freed block to the dealloc so that it will be freed
2110          * when we run dealloc.
2111          */
2112         ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE,
2113                                         le16_to_cpu(rb->rf_suballoc_slot),
2114                                         le64_to_cpu(rb->rf_suballoc_loc),
2115                                         le64_to_cpu(rb->rf_blkno),
2116                                         le16_to_cpu(rb->rf_suballoc_bit));
2117         if (ret) {
2118                 mlog_errno(ret);
2119                 goto out;
2120         }
2121
2122         ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh,
2123                                       OCFS2_JOURNAL_ACCESS_WRITE);
2124         if (ret) {
2125                 mlog_errno(ret);
2126                 goto out;
2127         }
2128
2129         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2130
2131         le32_add_cpu(&rb->rf_clusters, -1);
2132
2133         /*
2134          * check whether we need to restore the root refcount block if
2135          * there is no leaf extent block at atll.
2136          */
2137         if (!rb->rf_list.l_next_free_rec) {
2138                 BUG_ON(rb->rf_clusters);
2139
2140                 mlog(0, "reset refcount tree root %llu to be a record block.\n",
2141                      (unsigned long long)ref_root_bh->b_blocknr);
2142
2143                 rb->rf_flags = 0;
2144                 rb->rf_parent = 0;
2145                 rb->rf_cpos = 0;
2146                 memset(&rb->rf_records, 0, sb->s_blocksize -
2147                        offsetof(struct ocfs2_refcount_block, rf_records));
2148                 rb->rf_records.rl_count =
2149                                 cpu_to_le16(ocfs2_refcount_recs_per_rb(sb));
2150         }
2151
2152         ocfs2_journal_dirty(handle, ref_root_bh);
2153
2154 out:
2155         return ret;
2156 }
2157
2158 int ocfs2_increase_refcount(handle_t *handle,
2159                             struct ocfs2_caching_info *ci,
2160                             struct buffer_head *ref_root_bh,
2161                             u64 cpos, u32 len,
2162                             struct ocfs2_alloc_context *meta_ac,
2163                             struct ocfs2_cached_dealloc_ctxt *dealloc)
2164 {
2165         return __ocfs2_increase_refcount(handle, ci, ref_root_bh,
2166                                          cpos, len, 1,
2167                                          meta_ac, dealloc);
2168 }
2169
2170 static int ocfs2_decrease_refcount_rec(handle_t *handle,
2171                                 struct ocfs2_caching_info *ci,
2172                                 struct buffer_head *ref_root_bh,
2173                                 struct buffer_head *ref_leaf_bh,
2174                                 int index, u64 cpos, unsigned int len,
2175                                 struct ocfs2_alloc_context *meta_ac,
2176                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2177 {
2178         int ret;
2179         struct ocfs2_refcount_block *rb =
2180                         (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2181         struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index];
2182
2183         BUG_ON(cpos < le64_to_cpu(rec->r_cpos));
2184         BUG_ON(cpos + len >
2185                le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters));
2186
2187         if (cpos == le64_to_cpu(rec->r_cpos) &&
2188             len == le32_to_cpu(rec->r_clusters))
2189                 ret = ocfs2_change_refcount_rec(handle, ci,
2190                                                 ref_leaf_bh, index, 1, -1);
2191         else {
2192                 struct ocfs2_refcount_rec split = *rec;
2193                 split.r_cpos = cpu_to_le64(cpos);
2194                 split.r_clusters = cpu_to_le32(len);
2195
2196                 le32_add_cpu(&split.r_refcount, -1);
2197
2198                 mlog(0, "split refcount rec, start %llu, "
2199                      "len %u, count %u, original start %llu, len %u\n",
2200                      (unsigned long long)le64_to_cpu(split.r_cpos),
2201                      len, le32_to_cpu(split.r_refcount),
2202                      (unsigned long long)le64_to_cpu(rec->r_cpos),
2203                      le32_to_cpu(rec->r_clusters));
2204                 ret = ocfs2_split_refcount_rec(handle, ci,
2205                                                ref_root_bh, ref_leaf_bh,
2206                                                &split, index, 1,
2207                                                meta_ac, dealloc);
2208         }
2209
2210         if (ret) {
2211                 mlog_errno(ret);
2212                 goto out;
2213         }
2214
2215         /* Remove the leaf refcount block if it contains no refcount record. */
2216         if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) {
2217                 ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh,
2218                                                    ref_leaf_bh, meta_ac,
2219                                                    dealloc);
2220                 if (ret)
2221                         mlog_errno(ret);
2222         }
2223
2224 out:
2225         return ret;
2226 }
2227
2228 static int __ocfs2_decrease_refcount(handle_t *handle,
2229                                      struct ocfs2_caching_info *ci,
2230                                      struct buffer_head *ref_root_bh,
2231                                      u64 cpos, u32 len,
2232                                      struct ocfs2_alloc_context *meta_ac,
2233                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2234                                      int delete)
2235 {
2236         int ret = 0, index = 0;
2237         struct ocfs2_refcount_rec rec;
2238         unsigned int r_count = 0, r_len;
2239         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
2240         struct buffer_head *ref_leaf_bh = NULL;
2241
2242         mlog(0, "Tree owner %llu, decrease refcount start %llu, "
2243              "len %u, delete %u\n",
2244              (unsigned long long)ocfs2_metadata_cache_owner(ci),
2245              (unsigned long long)cpos, len, delete);
2246
2247         while (len) {
2248                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2249                                              cpos, len, &rec, &index,
2250                                              &ref_leaf_bh);
2251                 if (ret) {
2252                         mlog_errno(ret);
2253                         goto out;
2254                 }
2255
2256                 r_count = le32_to_cpu(rec.r_refcount);
2257                 BUG_ON(r_count == 0);
2258                 if (!delete)
2259                         BUG_ON(r_count > 1);
2260
2261                 r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) +
2262                               le32_to_cpu(rec.r_clusters)) - cpos;
2263
2264                 ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh,
2265                                                   ref_leaf_bh, index,
2266                                                   cpos, r_len,
2267                                                   meta_ac, dealloc);
2268                 if (ret) {
2269                         mlog_errno(ret);
2270                         goto out;
2271                 }
2272
2273                 if (le32_to_cpu(rec.r_refcount) == 1 && delete) {
2274                         ret = ocfs2_cache_cluster_dealloc(dealloc,
2275                                           ocfs2_clusters_to_blocks(sb, cpos),
2276                                                           r_len);
2277                         if (ret) {
2278                                 mlog_errno(ret);
2279                                 goto out;
2280                         }
2281                 }
2282
2283                 cpos += r_len;
2284                 len -= r_len;
2285                 brelse(ref_leaf_bh);
2286                 ref_leaf_bh = NULL;
2287         }
2288
2289 out:
2290         brelse(ref_leaf_bh);
2291         return ret;
2292 }
2293
2294 /* Caller must hold refcount tree lock. */
2295 int ocfs2_decrease_refcount(struct inode *inode,
2296                             handle_t *handle, u32 cpos, u32 len,
2297                             struct ocfs2_alloc_context *meta_ac,
2298                             struct ocfs2_cached_dealloc_ctxt *dealloc,
2299                             int delete)
2300 {
2301         int ret;
2302         u64 ref_blkno;
2303         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2304         struct buffer_head *ref_root_bh = NULL;
2305         struct ocfs2_refcount_tree *tree;
2306
2307         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2308
2309         ret = ocfs2_get_refcount_block(inode, &ref_blkno);
2310         if (ret) {
2311                 mlog_errno(ret);
2312                 goto out;
2313         }
2314
2315         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree);
2316         if (ret) {
2317                 mlog_errno(ret);
2318                 goto out;
2319         }
2320
2321         ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno,
2322                                         &ref_root_bh);
2323         if (ret) {
2324                 mlog_errno(ret);
2325                 goto out;
2326         }
2327
2328         ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh,
2329                                         cpos, len, meta_ac, dealloc, delete);
2330         if (ret)
2331                 mlog_errno(ret);
2332 out:
2333         brelse(ref_root_bh);
2334         return ret;
2335 }
2336
2337 /*
2338  * Mark the already-existing extent at cpos as refcounted for len clusters.
2339  * This adds the refcount extent flag.
2340  *
2341  * If the existing extent is larger than the request, initiate a
2342  * split. An attempt will be made at merging with adjacent extents.
2343  *
2344  * The caller is responsible for passing down meta_ac if we'll need it.
2345  */
2346 static int ocfs2_mark_extent_refcounted(struct inode *inode,
2347                                 struct ocfs2_extent_tree *et,
2348                                 handle_t *handle, u32 cpos,
2349                                 u32 len, u32 phys,
2350                                 struct ocfs2_alloc_context *meta_ac,
2351                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
2352 {
2353         int ret;
2354
2355         mlog(0, "Inode %lu refcount tree cpos %u, len %u, phys cluster %u\n",
2356              inode->i_ino, cpos, len, phys);
2357
2358         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2359                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2360                             "tree, but the feature bit is not set in the "
2361                             "super block.", inode->i_ino);
2362                 ret = -EROFS;
2363                 goto out;
2364         }
2365
2366         ret = ocfs2_change_extent_flag(handle, et, cpos,
2367                                        len, phys, meta_ac, dealloc,
2368                                        OCFS2_EXT_REFCOUNTED, 0);
2369         if (ret)
2370                 mlog_errno(ret);
2371
2372 out:
2373         return ret;
2374 }
2375
2376 /*
2377  * Given some contiguous physical clusters, calculate what we need
2378  * for modifying their refcount.
2379  */
2380 static int ocfs2_calc_refcount_meta_credits(struct super_block *sb,
2381                                             struct ocfs2_caching_info *ci,
2382                                             struct buffer_head *ref_root_bh,
2383                                             u64 start_cpos,
2384                                             u32 clusters,
2385                                             int *meta_add,
2386                                             int *credits)
2387 {
2388         int ret = 0, index, ref_blocks = 0, recs_add = 0;
2389         u64 cpos = start_cpos;
2390         struct ocfs2_refcount_block *rb;
2391         struct ocfs2_refcount_rec rec;
2392         struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL;
2393         u32 len;
2394
2395         mlog(0, "start_cpos %llu, clusters %u\n",
2396              (unsigned long long)start_cpos, clusters);
2397         while (clusters) {
2398                 ret = ocfs2_get_refcount_rec(ci, ref_root_bh,
2399                                              cpos, clusters, &rec,
2400                                              &index, &ref_leaf_bh);
2401                 if (ret) {
2402                         mlog_errno(ret);
2403                         goto out;
2404                 }
2405
2406                 if (ref_leaf_bh != prev_bh) {
2407                         /*
2408                          * Now we encounter a new leaf block, so calculate
2409                          * whether we need to extend the old leaf.
2410                          */
2411                         if (prev_bh) {
2412                                 rb = (struct ocfs2_refcount_block *)
2413                                                         prev_bh->b_data;
2414
2415                                 if (le64_to_cpu(rb->rf_records.rl_used) +
2416                                     recs_add >
2417                                     le16_to_cpu(rb->rf_records.rl_count))
2418                                         ref_blocks++;
2419                         }
2420
2421                         recs_add = 0;
2422                         *credits += 1;
2423                         brelse(prev_bh);
2424                         prev_bh = ref_leaf_bh;
2425                         get_bh(prev_bh);
2426                 }
2427
2428                 rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
2429
2430                 mlog(0, "recs_add %d,cpos %llu, clusters %u, rec->r_cpos %llu,"
2431                      "rec->r_clusters %u, rec->r_refcount %u, index %d\n",
2432                      recs_add, (unsigned long long)cpos, clusters,
2433                      (unsigned long long)le64_to_cpu(rec.r_cpos),
2434                      le32_to_cpu(rec.r_clusters),
2435                      le32_to_cpu(rec.r_refcount), index);
2436
2437                 len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) +
2438                           le32_to_cpu(rec.r_clusters)) - cpos;
2439                 /*
2440                  * We record all the records which will be inserted to the
2441                  * same refcount block, so that we can tell exactly whether
2442                  * we need a new refcount block or not.
2443                  *
2444                  * If we will insert a new one, this is easy and only happens
2445                  * during adding refcounted flag to the extent, so we don't
2446                  * have a chance of spliting. We just need one record.
2447                  *
2448                  * If the refcount rec already exists, that would be a little
2449                  * complicated. we may have to:
2450                  * 1) split at the beginning if the start pos isn't aligned.
2451                  *    we need 1 more record in this case.
2452                  * 2) split int the end if the end pos isn't aligned.
2453                  *    we need 1 more record in this case.
2454                  * 3) split in the middle because of file system fragmentation.
2455                  *    we need 2 more records in this case(we can't detect this
2456                  *    beforehand, so always think of the worst case).
2457                  */
2458                 if (rec.r_refcount) {
2459                         recs_add += 2;
2460                         /* Check whether we need a split at the beginning. */
2461                         if (cpos == start_cpos &&
2462                             cpos != le64_to_cpu(rec.r_cpos))
2463                                 recs_add++;
2464
2465                         /* Check whether we need a split in the end. */
2466                         if (cpos + clusters < le64_to_cpu(rec.r_cpos) +
2467                             le32_to_cpu(rec.r_clusters))
2468                                 recs_add++;
2469                 } else
2470                         recs_add++;
2471
2472                 brelse(ref_leaf_bh);
2473                 ref_leaf_bh = NULL;
2474                 clusters -= len;
2475                 cpos += len;
2476         }
2477
2478         if (prev_bh) {
2479                 rb = (struct ocfs2_refcount_block *)prev_bh->b_data;
2480
2481                 if (le64_to_cpu(rb->rf_records.rl_used) + recs_add >
2482                     le16_to_cpu(rb->rf_records.rl_count))
2483                         ref_blocks++;
2484
2485                 *credits += 1;
2486         }
2487
2488         if (!ref_blocks)
2489                 goto out;
2490
2491         mlog(0, "we need ref_blocks %d\n", ref_blocks);
2492         *meta_add += ref_blocks;
2493         *credits += ref_blocks;
2494
2495         /*
2496          * So we may need ref_blocks to insert into the tree.
2497          * That also means we need to change the b-tree and add that number
2498          * of records since we never merge them.
2499          * We need one more block for expansion since the new created leaf
2500          * block is also full and needs split.
2501          */
2502         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
2503         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) {
2504                 struct ocfs2_extent_tree et;
2505
2506                 ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh);
2507                 *meta_add += ocfs2_extend_meta_needed(et.et_root_el);
2508                 *credits += ocfs2_calc_extend_credits(sb,
2509                                                       et.et_root_el,
2510                                                       ref_blocks);
2511         } else {
2512                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
2513                 *meta_add += 1;
2514         }
2515
2516 out:
2517         brelse(ref_leaf_bh);
2518         brelse(prev_bh);
2519         return ret;
2520 }
2521
2522 /*
2523  * For refcount tree, we will decrease some contiguous clusters
2524  * refcount count, so just go through it to see how many blocks
2525  * we gonna touch and whether we need to create new blocks.
2526  *
2527  * Normally the refcount blocks store these refcount should be
2528  * contiguous also, so that we can get the number easily.
2529  * We will at most add split 2 refcount records and 2 more
2530  * refcount blocks, so just check it in a rough way.
2531  *
2532  * Caller must hold refcount tree lock.
2533  */
2534 int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
2535                                           u64 refcount_loc,
2536                                           u64 phys_blkno,
2537                                           u32 clusters,
2538                                           int *credits,
2539                                           int *ref_blocks)
2540 {
2541         int ret;
2542         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2543         struct buffer_head *ref_root_bh = NULL;
2544         struct ocfs2_refcount_tree *tree;
2545         u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno);
2546
2547         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
2548                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
2549                             "tree, but the feature bit is not set in the "
2550                             "super block.", inode->i_ino);
2551                 ret = -EROFS;
2552                 goto out;
2553         }
2554
2555         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
2556
2557         ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb),
2558                                       refcount_loc, &tree);
2559         if (ret) {
2560                 mlog_errno(ret);
2561                 goto out;
2562         }
2563
2564         ret = ocfs2_read_refcount_block(&tree->rf_ci, refcount_loc,
2565                                         &ref_root_bh);
2566         if (ret) {
2567                 mlog_errno(ret);
2568                 goto out;
2569         }
2570
2571         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
2572                                                &tree->rf_ci,
2573                                                ref_root_bh,
2574                                                start_cpos, clusters,
2575                                                ref_blocks, credits);
2576         if (ret) {
2577                 mlog_errno(ret);
2578                 goto out;
2579         }
2580
2581         mlog(0, "reserve new metadata %d blocks, credits = %d\n",
2582              *ref_blocks, *credits);
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, readahead_pages;
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         readahead_pages =
2944                 (ocfs2_cow_contig_clusters(sb) <<
2945                  OCFS2_SB(sb)->s_clustersize_bits) >> PAGE_CACHE_SHIFT;
2946         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
2947         end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
2948         /*
2949          * We only duplicate pages until we reach the page contains i_size - 1.
2950          * So trim 'end' to i_size.
2951          */
2952         if (end > i_size_read(context->inode))
2953                 end = i_size_read(context->inode);
2954
2955         while (offset < end) {
2956                 page_index = offset >> PAGE_CACHE_SHIFT;
2957                 map_end = ((loff_t)page_index + 1) << PAGE_CACHE_SHIFT;
2958                 if (map_end > end)
2959                         map_end = end;
2960
2961                 /* from, to is the offset within the page. */
2962                 from = offset & (PAGE_CACHE_SIZE - 1);
2963                 to = PAGE_CACHE_SIZE;
2964                 if (map_end & (PAGE_CACHE_SIZE - 1))
2965                         to = map_end & (PAGE_CACHE_SIZE - 1);
2966
2967                 page = find_or_create_page(mapping, page_index, GFP_NOFS);
2968
2969                 /*
2970                  * In case PAGE_CACHE_SIZE <= CLUSTER_SIZE, This page
2971                  * can't be dirtied before we CoW it out.
2972                  */
2973                 if (PAGE_CACHE_SIZE <= OCFS2_SB(sb)->s_clustersize)
2974                         BUG_ON(PageDirty(page));
2975
2976                 if (PageReadahead(page) && context->file) {
2977                         page_cache_async_readahead(mapping,
2978                                                    &context->file->f_ra,
2979                                                    context->file,
2980                                                    page, page_index,
2981                                                    readahead_pages);
2982                 }
2983
2984                 if (!PageUptodate(page)) {
2985                         ret = block_read_full_page(page, ocfs2_get_block);
2986                         if (ret) {
2987                                 mlog_errno(ret);
2988                                 goto unlock;
2989                         }
2990                         lock_page(page);
2991                 }
2992
2993                 if (page_has_buffers(page)) {
2994                         ret = walk_page_buffers(handle, page_buffers(page),
2995                                                 from, to, &partial,
2996                                                 ocfs2_clear_cow_buffer);
2997                         if (ret) {
2998                                 mlog_errno(ret);
2999                                 goto unlock;
3000                         }
3001                 }
3002
3003                 ocfs2_map_and_dirty_page(context->inode,
3004                                          handle, from, to,
3005                                          page, 0, &new_block);
3006                 mark_page_accessed(page);
3007 unlock:
3008                 unlock_page(page);
3009                 page_cache_release(page);
3010                 page = NULL;
3011                 offset = map_end;
3012                 if (ret)
3013                         break;
3014         }
3015
3016         return ret;
3017 }
3018
3019 static int ocfs2_duplicate_clusters_by_jbd(handle_t *handle,
3020                                            struct ocfs2_cow_context *context,
3021                                            u32 cpos, u32 old_cluster,
3022                                            u32 new_cluster, u32 new_len)
3023 {
3024         int ret = 0;
3025         struct super_block *sb = context->inode->i_sb;
3026         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3027         int i, blocks = ocfs2_clusters_to_blocks(sb, new_len);
3028         u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster);
3029         u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
3030         struct ocfs2_super *osb = OCFS2_SB(sb);
3031         struct buffer_head *old_bh = NULL;
3032         struct buffer_head *new_bh = NULL;
3033
3034         mlog(0, "old_cluster %u, new %u, len %u\n", old_cluster,
3035              new_cluster, new_len);
3036
3037         for (i = 0; i < blocks; i++, old_block++, new_block++) {
3038                 new_bh = sb_getblk(osb->sb, new_block);
3039                 if (new_bh == NULL) {
3040                         ret = -EIO;
3041                         mlog_errno(ret);
3042                         break;
3043                 }
3044
3045                 ocfs2_set_new_buffer_uptodate(ci, new_bh);
3046
3047                 ret = ocfs2_read_block(ci, old_block, &old_bh, NULL);
3048                 if (ret) {
3049                         mlog_errno(ret);
3050                         break;
3051                 }
3052
3053                 ret = ocfs2_journal_access(handle, ci, new_bh,
3054                                            OCFS2_JOURNAL_ACCESS_CREATE);
3055                 if (ret) {
3056                         mlog_errno(ret);
3057                         break;
3058                 }
3059
3060                 memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize);
3061                 ocfs2_journal_dirty(handle, new_bh);
3062
3063                 brelse(new_bh);
3064                 brelse(old_bh);
3065                 new_bh = NULL;
3066                 old_bh = NULL;
3067         }
3068
3069         brelse(new_bh);
3070         brelse(old_bh);
3071         return ret;
3072 }
3073
3074 static int ocfs2_clear_ext_refcount(handle_t *handle,
3075                                     struct ocfs2_extent_tree *et,
3076                                     u32 cpos, u32 p_cluster, u32 len,
3077                                     unsigned int ext_flags,
3078                                     struct ocfs2_alloc_context *meta_ac,
3079                                     struct ocfs2_cached_dealloc_ctxt *dealloc)
3080 {
3081         int ret, index;
3082         struct ocfs2_extent_rec replace_rec;
3083         struct ocfs2_path *path = NULL;
3084         struct ocfs2_extent_list *el;
3085         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
3086         u64 ino = ocfs2_metadata_cache_owner(et->et_ci);
3087
3088         mlog(0, "inode %llu cpos %u, len %u, p_cluster %u, ext_flags %u\n",
3089              (unsigned long long)ino, cpos, len, p_cluster, ext_flags);
3090
3091         memset(&replace_rec, 0, sizeof(replace_rec));
3092         replace_rec.e_cpos = cpu_to_le32(cpos);
3093         replace_rec.e_leaf_clusters = cpu_to_le16(len);
3094         replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb,
3095                                                                    p_cluster));
3096         replace_rec.e_flags = ext_flags;
3097         replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED;
3098
3099         path = ocfs2_new_path_from_et(et);
3100         if (!path) {
3101                 ret = -ENOMEM;
3102                 mlog_errno(ret);
3103                 goto out;
3104         }
3105
3106         ret = ocfs2_find_path(et->et_ci, path, cpos);
3107         if (ret) {
3108                 mlog_errno(ret);
3109                 goto out;
3110         }
3111
3112         el = path_leaf_el(path);
3113
3114         index = ocfs2_search_extent_list(el, cpos);
3115         if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) {
3116                 ocfs2_error(sb,
3117                             "Inode %llu has an extent at cpos %u which can no "
3118                             "longer be found.\n",
3119                             (unsigned long long)ino, cpos);
3120                 ret = -EROFS;
3121                 goto out;
3122         }
3123
3124         ret = ocfs2_split_extent(handle, et, path, index,
3125                                  &replace_rec, meta_ac, dealloc);
3126         if (ret)
3127                 mlog_errno(ret);
3128
3129 out:
3130         ocfs2_free_path(path);
3131         return ret;
3132 }
3133
3134 static int ocfs2_replace_clusters(handle_t *handle,
3135                                   struct ocfs2_cow_context *context,
3136                                   u32 cpos, u32 old,
3137                                   u32 new, u32 len,
3138                                   unsigned int ext_flags)
3139 {
3140         int ret;
3141         struct ocfs2_caching_info *ci = context->data_et.et_ci;
3142         u64 ino = ocfs2_metadata_cache_owner(ci);
3143
3144         mlog(0, "inode %llu, cpos %u, old %u, new %u, len %u, ext_flags %u\n",
3145              (unsigned long long)ino, cpos, old, new, len, ext_flags);
3146
3147         /*If the old clusters is unwritten, no need to duplicate. */
3148         if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) {
3149                 ret = context->cow_duplicate_clusters(handle, context, cpos,
3150                                                       old, new, len);
3151                 if (ret) {
3152                         mlog_errno(ret);
3153                         goto out;
3154                 }
3155         }
3156
3157         ret = ocfs2_clear_ext_refcount(handle, &context->data_et,
3158                                        cpos, new, len, ext_flags,
3159                                        context->meta_ac, &context->dealloc);
3160         if (ret)
3161                 mlog_errno(ret);
3162 out:
3163         return ret;
3164 }
3165
3166 static int ocfs2_cow_sync_writeback(struct super_block *sb,
3167                                     struct ocfs2_cow_context *context,
3168                                     u32 cpos, u32 num_clusters)
3169 {
3170         int ret = 0;
3171         loff_t offset, end, map_end;
3172         pgoff_t page_index;
3173         struct page *page;
3174
3175         if (ocfs2_should_order_data(context->inode))
3176                 return 0;
3177
3178         offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
3179         end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits);
3180
3181         ret = filemap_fdatawrite_range(context->inode->i_mapping,
3182                                        offset, end - 1);
3183         if (ret < 0) {
3184                 mlog_errno(ret);
3185                 return ret;
3186         }
3187
3188         while (offset < end) {
3189                 page_index = offset >> PAGE_CACHE_SHIFT;
3190                 map_end = ((loff_t)page_index + 1) << PAGE_CACHE_SHIFT;
3191                 if (map_end > end)
3192                         map_end = end;
3193
3194                 page = find_or_create_page(context->inode->i_mapping,
3195                                            page_index, GFP_NOFS);
3196                 BUG_ON(!page);
3197
3198                 wait_on_page_writeback(page);
3199                 if (PageError(page)) {
3200                         ret = -EIO;
3201                         mlog_errno(ret);
3202                 } else
3203                         mark_page_accessed(page);
3204
3205                 unlock_page(page);
3206                 page_cache_release(page);
3207                 page = NULL;
3208                 offset = map_end;
3209                 if (ret)
3210                         break;
3211         }
3212
3213         return ret;
3214 }
3215
3216 static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context,
3217                                  u32 v_cluster, u32 *p_cluster,
3218                                  u32 *num_clusters,
3219                                  unsigned int *extent_flags)
3220 {
3221         return ocfs2_get_clusters(context->inode, v_cluster, p_cluster,
3222                                   num_clusters, extent_flags);
3223 }
3224
3225 static int ocfs2_make_clusters_writable(struct super_block *sb,
3226                                         struct ocfs2_cow_context *context,
3227                                         u32 cpos, u32 p_cluster,
3228                                         u32 num_clusters, unsigned int e_flags)
3229 {
3230         int ret, delete, index, credits =  0;
3231         u32 new_bit, new_len, orig_num_clusters;
3232         unsigned int set_len;
3233         struct ocfs2_super *osb = OCFS2_SB(sb);
3234         handle_t *handle;
3235         struct buffer_head *ref_leaf_bh = NULL;
3236         struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci;
3237         struct ocfs2_refcount_rec rec;
3238
3239         mlog(0, "cpos %u, p_cluster %u, num_clusters %u, e_flags %u\n",
3240              cpos, p_cluster, num_clusters, e_flags);
3241
3242         ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters,
3243                                              &context->data_et,
3244                                              ref_ci,
3245                                              context->ref_root_bh,
3246                                              &context->meta_ac,
3247                                              &context->data_ac, &credits);
3248         if (ret) {
3249                 mlog_errno(ret);
3250                 return ret;
3251         }
3252
3253         if (context->post_refcount)
3254                 credits += context->post_refcount->credits;
3255
3256         credits += context->extra_credits;
3257         handle = ocfs2_start_trans(osb, credits);
3258         if (IS_ERR(handle)) {
3259                 ret = PTR_ERR(handle);
3260                 mlog_errno(ret);
3261                 goto out;
3262         }
3263
3264         orig_num_clusters = num_clusters;
3265
3266         while (num_clusters) {
3267                 ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh,
3268                                              p_cluster, num_clusters,
3269                                              &rec, &index, &ref_leaf_bh);
3270                 if (ret) {
3271                         mlog_errno(ret);
3272                         goto out_commit;
3273                 }
3274
3275                 BUG_ON(!rec.r_refcount);
3276                 set_len = min((u64)p_cluster + num_clusters,
3277                               le64_to_cpu(rec.r_cpos) +
3278                               le32_to_cpu(rec.r_clusters)) - p_cluster;
3279
3280                 /*
3281                  * There are many different situation here.
3282                  * 1. If refcount == 1, remove the flag and don't COW.
3283                  * 2. If refcount > 1, allocate clusters.
3284                  *    Here we may not allocate r_len once at a time, so continue
3285                  *    until we reach num_clusters.
3286                  */
3287                 if (le32_to_cpu(rec.r_refcount) == 1) {
3288                         delete = 0;
3289                         ret = ocfs2_clear_ext_refcount(handle,
3290                                                        &context->data_et,
3291                                                        cpos, p_cluster,
3292                                                        set_len, e_flags,
3293                                                        context->meta_ac,
3294                                                        &context->dealloc);
3295                         if (ret) {
3296                                 mlog_errno(ret);
3297                                 goto out_commit;
3298                         }
3299                 } else {
3300                         delete = 1;
3301
3302                         ret = __ocfs2_claim_clusters(handle,
3303                                                      context->data_ac,
3304                                                      1, set_len,
3305                                                      &new_bit, &new_len);
3306                         if (ret) {
3307                                 mlog_errno(ret);
3308                                 goto out_commit;
3309                         }
3310
3311                         ret = ocfs2_replace_clusters(handle, context,
3312                                                      cpos, p_cluster, new_bit,
3313                                                      new_len, e_flags);
3314                         if (ret) {
3315                                 mlog_errno(ret);
3316                                 goto out_commit;
3317                         }
3318                         set_len = new_len;
3319                 }
3320
3321                 ret = __ocfs2_decrease_refcount(handle, ref_ci,
3322                                                 context->ref_root_bh,
3323                                                 p_cluster, set_len,
3324                                                 context->meta_ac,
3325                                                 &context->dealloc, delete);
3326                 if (ret) {
3327                         mlog_errno(ret);
3328                         goto out_commit;
3329                 }
3330
3331                 cpos += set_len;
3332                 p_cluster += set_len;
3333                 num_clusters -= set_len;
3334                 brelse(ref_leaf_bh);
3335                 ref_leaf_bh = NULL;
3336         }
3337
3338         /* handle any post_cow action. */
3339         if (context->post_refcount && context->post_refcount->func) {
3340                 ret = context->post_refcount->func(context->inode, handle,
3341                                                 context->post_refcount->para);
3342                 if (ret) {
3343                         mlog_errno(ret);
3344                         goto out_commit;
3345                 }
3346         }
3347
3348         /*
3349          * Here we should write the new page out first if we are
3350          * in write-back mode.
3351          */
3352         if (context->get_clusters == ocfs2_di_get_clusters) {
3353                 ret = ocfs2_cow_sync_writeback(sb, context, cpos,
3354                                                orig_num_clusters);
3355                 if (ret)
3356                         mlog_errno(ret);
3357         }
3358
3359 out_commit:
3360         ocfs2_commit_trans(osb, handle);
3361
3362 out:
3363         if (context->data_ac) {
3364                 ocfs2_free_alloc_context(context->data_ac);
3365                 context->data_ac = NULL;
3366         }
3367         if (context->meta_ac) {
3368                 ocfs2_free_alloc_context(context->meta_ac);
3369                 context->meta_ac = NULL;
3370         }
3371         brelse(ref_leaf_bh);
3372
3373         return ret;
3374 }
3375
3376 static int ocfs2_replace_cow(struct ocfs2_cow_context *context)
3377 {
3378         int ret = 0;
3379         struct inode *inode = context->inode;
3380         u32 cow_start = context->cow_start, cow_len = context->cow_len;
3381         u32 p_cluster, num_clusters;
3382         unsigned int ext_flags;
3383         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3384
3385         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) {
3386                 ocfs2_error(inode->i_sb, "Inode %lu want to use refcount "
3387                             "tree, but the feature bit is not set in the "
3388                             "super block.", inode->i_ino);
3389                 return -EROFS;
3390         }
3391
3392         ocfs2_init_dealloc_ctxt(&context->dealloc);
3393
3394         while (cow_len) {
3395                 ret = context->get_clusters(context, cow_start, &p_cluster,
3396                                             &num_clusters, &ext_flags);
3397                 if (ret) {
3398                         mlog_errno(ret);
3399                         break;
3400                 }
3401
3402                 BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED));
3403
3404                 if (cow_len < num_clusters)
3405                         num_clusters = cow_len;
3406
3407                 ret = ocfs2_make_clusters_writable(inode->i_sb, context,
3408                                                    cow_start, p_cluster,
3409                                                    num_clusters, ext_flags);
3410                 if (ret) {
3411                         mlog_errno(ret);
3412                         break;
3413                 }
3414
3415                 cow_len -= num_clusters;
3416                 cow_start += num_clusters;
3417         }
3418
3419         if (ocfs2_dealloc_has_cluster(&context->dealloc)) {
3420                 ocfs2_schedule_truncate_log_flush(osb, 1);
3421                 ocfs2_run_deallocs(osb, &context->dealloc);
3422         }
3423
3424         return ret;
3425 }
3426
3427 static void ocfs2_readahead_for_cow(struct inode *inode,
3428                                     struct file *file,
3429                                     u32 start, u32 len)
3430 {
3431         struct address_space *mapping;
3432         pgoff_t index;
3433         unsigned long num_pages;
3434         int cs_bits = OCFS2_SB(inode->i_sb)->s_clustersize_bits;
3435
3436         if (!file)
3437                 return;
3438
3439         mapping = file->f_mapping;
3440         num_pages = (len << cs_bits) >> PAGE_CACHE_SHIFT;
3441         if (!num_pages)
3442                 num_pages = 1;
3443
3444         index = ((loff_t)start << cs_bits) >> PAGE_CACHE_SHIFT;
3445         page_cache_sync_readahead(mapping, &file->f_ra, file,
3446                                   index, num_pages);
3447 }
3448
3449 /*
3450  * Starting at cpos, try to CoW write_len clusters.  Don't CoW
3451  * past max_cpos.  This will stop when it runs into a hole or an
3452  * unrefcounted extent.
3453  */
3454 static int ocfs2_refcount_cow_hunk(struct inode *inode,
3455                                    struct file *file,
3456                                    struct buffer_head *di_bh,
3457                                    u32 cpos, u32 write_len, u32 max_cpos)
3458 {
3459         int ret;
3460         u32 cow_start = 0, cow_len = 0;
3461         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3462         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3463         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3464         struct buffer_head *ref_root_bh = NULL;
3465         struct ocfs2_refcount_tree *ref_tree;
3466         struct ocfs2_cow_context *context = NULL;
3467
3468         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3469
3470         ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list,
3471                                               cpos, write_len, max_cpos,
3472                                               &cow_start, &cow_len);
3473         if (ret) {
3474                 mlog_errno(ret);
3475                 goto out;
3476         }
3477
3478         mlog(0, "CoW inode %lu, cpos %u, write_len %u, cow_start %u, "
3479              "cow_len %u\n", inode->i_ino,
3480              cpos, write_len, cow_start, cow_len);
3481
3482         BUG_ON(cow_len == 0);
3483
3484         ocfs2_readahead_for_cow(inode, file, cow_start, cow_len);
3485
3486         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3487         if (!context) {
3488                 ret = -ENOMEM;
3489                 mlog_errno(ret);
3490                 goto out;
3491         }
3492
3493         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
3494                                        1, &ref_tree, &ref_root_bh);
3495         if (ret) {
3496                 mlog_errno(ret);
3497                 goto out;
3498         }
3499
3500         context->inode = inode;
3501         context->cow_start = cow_start;
3502         context->cow_len = cow_len;
3503         context->ref_tree = ref_tree;
3504         context->ref_root_bh = ref_root_bh;
3505         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
3506         context->get_clusters = ocfs2_di_get_clusters;
3507         context->file = file;
3508
3509         ocfs2_init_dinode_extent_tree(&context->data_et,
3510                                       INODE_CACHE(inode), di_bh);
3511
3512         ret = ocfs2_replace_cow(context);
3513         if (ret)
3514                 mlog_errno(ret);
3515
3516         /*
3517          * truncate the extent map here since no matter whether we meet with
3518          * any error during the action, we shouldn't trust cached extent map
3519          * any more.
3520          */
3521         ocfs2_extent_map_trunc(inode, cow_start);
3522
3523         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3524         brelse(ref_root_bh);
3525 out:
3526         kfree(context);
3527         return ret;
3528 }
3529
3530 /*
3531  * CoW any and all clusters between cpos and cpos+write_len.
3532  * Don't CoW past max_cpos.  If this returns successfully, all
3533  * clusters between cpos and cpos+write_len are safe to modify.
3534  */
3535 int ocfs2_refcount_cow(struct inode *inode,
3536                        struct file *file,
3537                        struct buffer_head *di_bh,
3538                        u32 cpos, u32 write_len, u32 max_cpos)
3539 {
3540         int ret = 0;
3541         u32 p_cluster, num_clusters;
3542         unsigned int ext_flags;
3543
3544         while (write_len) {
3545                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3546                                          &num_clusters, &ext_flags);
3547                 if (ret) {
3548                         mlog_errno(ret);
3549                         break;
3550                 }
3551
3552                 if (write_len < num_clusters)
3553                         num_clusters = write_len;
3554
3555                 if (ext_flags & OCFS2_EXT_REFCOUNTED) {
3556                         ret = ocfs2_refcount_cow_hunk(inode, file, di_bh, cpos,
3557                                                       num_clusters, max_cpos);
3558                         if (ret) {
3559                                 mlog_errno(ret);
3560                                 break;
3561                         }
3562                 }
3563
3564                 write_len -= num_clusters;
3565                 cpos += num_clusters;
3566         }
3567
3568         return ret;
3569 }
3570
3571 static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context,
3572                                           u32 v_cluster, u32 *p_cluster,
3573                                           u32 *num_clusters,
3574                                           unsigned int *extent_flags)
3575 {
3576         struct inode *inode = context->inode;
3577         struct ocfs2_xattr_value_root *xv = context->cow_object;
3578
3579         return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster,
3580                                         num_clusters, &xv->xr_list,
3581                                         extent_flags);
3582 }
3583
3584 /*
3585  * Given a xattr value root, calculate the most meta/credits we need for
3586  * refcount tree change if we truncate it to 0.
3587  */
3588 int ocfs2_refcounted_xattr_delete_need(struct inode *inode,
3589                                        struct ocfs2_caching_info *ref_ci,
3590                                        struct buffer_head *ref_root_bh,
3591                                        struct ocfs2_xattr_value_root *xv,
3592                                        int *meta_add, int *credits)
3593 {
3594         int ret = 0, index, ref_blocks = 0;
3595         u32 p_cluster, num_clusters;
3596         u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters);
3597         struct ocfs2_refcount_block *rb;
3598         struct ocfs2_refcount_rec rec;
3599         struct buffer_head *ref_leaf_bh = NULL;
3600
3601         while (cpos < clusters) {
3602                 ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster,
3603                                                &num_clusters, &xv->xr_list,
3604                                                NULL);
3605                 if (ret) {
3606                         mlog_errno(ret);
3607                         goto out;
3608                 }
3609
3610                 cpos += num_clusters;
3611
3612                 while (num_clusters) {
3613                         ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh,
3614                                                      p_cluster, num_clusters,
3615                                                      &rec, &index,
3616                                                      &ref_leaf_bh);
3617                         if (ret) {
3618                                 mlog_errno(ret);
3619                                 goto out;
3620                         }
3621
3622                         BUG_ON(!rec.r_refcount);
3623
3624                         rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data;
3625
3626                         /*
3627                          * We really don't know whether the other clusters is in
3628                          * this refcount block or not, so just take the worst
3629                          * case that all the clusters are in this block and each
3630                          * one will split a refcount rec, so totally we need
3631                          * clusters * 2 new refcount rec.
3632                          */
3633                         if (le64_to_cpu(rb->rf_records.rl_used) + clusters * 2 >
3634                             le16_to_cpu(rb->rf_records.rl_count))
3635                                 ref_blocks++;
3636
3637                         *credits += 1;
3638                         brelse(ref_leaf_bh);
3639                         ref_leaf_bh = NULL;
3640
3641                         if (num_clusters <= le32_to_cpu(rec.r_clusters))
3642                                 break;
3643                         else
3644                                 num_clusters -= le32_to_cpu(rec.r_clusters);
3645                         p_cluster += num_clusters;
3646                 }
3647         }
3648
3649         *meta_add += ref_blocks;
3650         if (!ref_blocks)
3651                 goto out;
3652
3653         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
3654         if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)
3655                 *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS;
3656         else {
3657                 struct ocfs2_extent_tree et;
3658
3659                 ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh);
3660                 *credits += ocfs2_calc_extend_credits(inode->i_sb,
3661                                                       et.et_root_el,
3662                                                       ref_blocks);
3663         }
3664
3665 out:
3666         brelse(ref_leaf_bh);
3667         return ret;
3668 }
3669
3670 /*
3671  * Do CoW for xattr.
3672  */
3673 int ocfs2_refcount_cow_xattr(struct inode *inode,
3674                              struct ocfs2_dinode *di,
3675                              struct ocfs2_xattr_value_buf *vb,
3676                              struct ocfs2_refcount_tree *ref_tree,
3677                              struct buffer_head *ref_root_bh,
3678                              u32 cpos, u32 write_len,
3679                              struct ocfs2_post_refcount *post)
3680 {
3681         int ret;
3682         struct ocfs2_xattr_value_root *xv = vb->vb_xv;
3683         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3684         struct ocfs2_cow_context *context = NULL;
3685         u32 cow_start, cow_len;
3686
3687         BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
3688
3689         ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list,
3690                                               cpos, write_len, UINT_MAX,
3691                                               &cow_start, &cow_len);
3692         if (ret) {
3693                 mlog_errno(ret);
3694                 goto out;
3695         }
3696
3697         BUG_ON(cow_len == 0);
3698
3699         context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
3700         if (!context) {
3701                 ret = -ENOMEM;
3702                 mlog_errno(ret);
3703                 goto out;
3704         }
3705
3706         context->inode = inode;
3707         context->cow_start = cow_start;
3708         context->cow_len = cow_len;
3709         context->ref_tree = ref_tree;
3710         context->ref_root_bh = ref_root_bh;;
3711         context->cow_object = xv;
3712
3713         context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd;
3714         /* We need the extra credits for duplicate_clusters by jbd. */
3715         context->extra_credits =
3716                 ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len;
3717         context->get_clusters = ocfs2_xattr_value_get_clusters;
3718         context->post_refcount = post;
3719
3720         ocfs2_init_xattr_value_extent_tree(&context->data_et,
3721                                            INODE_CACHE(inode), vb);
3722
3723         ret = ocfs2_replace_cow(context);
3724         if (ret)
3725                 mlog_errno(ret);
3726
3727 out:
3728         kfree(context);
3729         return ret;
3730 }
3731
3732 /*
3733  * Insert a new extent into refcount tree and mark a extent rec
3734  * as refcounted in the dinode tree.
3735  */
3736 int ocfs2_add_refcount_flag(struct inode *inode,
3737                             struct ocfs2_extent_tree *data_et,
3738                             struct ocfs2_caching_info *ref_ci,
3739                             struct buffer_head *ref_root_bh,
3740                             u32 cpos, u32 p_cluster, u32 num_clusters,
3741                             struct ocfs2_cached_dealloc_ctxt *dealloc,
3742                             struct ocfs2_post_refcount *post)
3743 {
3744         int ret;
3745         handle_t *handle;
3746         int credits = 1, ref_blocks = 0;
3747         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3748         struct ocfs2_alloc_context *meta_ac = NULL;
3749
3750         ret = ocfs2_calc_refcount_meta_credits(inode->i_sb,
3751                                                ref_ci, ref_root_bh,
3752                                                p_cluster, num_clusters,
3753                                                &ref_blocks, &credits);
3754         if (ret) {
3755                 mlog_errno(ret);
3756                 goto out;
3757         }
3758
3759         mlog(0, "reserve new metadata %d, credits = %d\n",
3760              ref_blocks, credits);
3761
3762         if (ref_blocks) {
3763                 ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb),
3764                                                         ref_blocks, &meta_ac);
3765                 if (ret) {
3766                         mlog_errno(ret);
3767                         goto out;
3768                 }
3769         }
3770
3771         if (post)
3772                 credits += post->credits;
3773
3774         handle = ocfs2_start_trans(osb, credits);
3775         if (IS_ERR(handle)) {
3776                 ret = PTR_ERR(handle);
3777                 mlog_errno(ret);
3778                 goto out;
3779         }
3780
3781         ret = ocfs2_mark_extent_refcounted(inode, data_et, handle,
3782                                            cpos, num_clusters, p_cluster,
3783                                            meta_ac, dealloc);
3784         if (ret) {
3785                 mlog_errno(ret);
3786                 goto out_commit;
3787         }
3788
3789         ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3790                                         p_cluster, num_clusters, 0,
3791                                         meta_ac, dealloc);
3792         if (ret) {
3793                 mlog_errno(ret);
3794                 goto out_commit;
3795         }
3796
3797         if (post && post->func) {
3798                 ret = post->func(inode, handle, post->para);
3799                 if (ret)
3800                         mlog_errno(ret);
3801         }
3802
3803 out_commit:
3804         ocfs2_commit_trans(osb, handle);
3805 out:
3806         if (meta_ac)
3807                 ocfs2_free_alloc_context(meta_ac);
3808         return ret;
3809 }
3810
3811 static int ocfs2_change_ctime(struct inode *inode,
3812                               struct buffer_head *di_bh)
3813 {
3814         int ret;
3815         handle_t *handle;
3816         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3817
3818         handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb),
3819                                    OCFS2_INODE_UPDATE_CREDITS);
3820         if (IS_ERR(handle)) {
3821                 ret = PTR_ERR(handle);
3822                 mlog_errno(ret);
3823                 goto out;
3824         }
3825
3826         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
3827                                       OCFS2_JOURNAL_ACCESS_WRITE);
3828         if (ret) {
3829                 mlog_errno(ret);
3830                 goto out_commit;
3831         }
3832
3833         inode->i_ctime = CURRENT_TIME;
3834         di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec);
3835         di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
3836
3837         ocfs2_journal_dirty(handle, di_bh);
3838
3839 out_commit:
3840         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
3841 out:
3842         return ret;
3843 }
3844
3845 static int ocfs2_attach_refcount_tree(struct inode *inode,
3846                                       struct buffer_head *di_bh)
3847 {
3848         int ret, data_changed = 0;
3849         struct buffer_head *ref_root_bh = NULL;
3850         struct ocfs2_inode_info *oi = OCFS2_I(inode);
3851         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3852         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3853         struct ocfs2_refcount_tree *ref_tree;
3854         unsigned int ext_flags;
3855         loff_t size;
3856         u32 cpos, num_clusters, clusters, p_cluster;
3857         struct ocfs2_cached_dealloc_ctxt dealloc;
3858         struct ocfs2_extent_tree di_et;
3859
3860         ocfs2_init_dealloc_ctxt(&dealloc);
3861
3862         if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) {
3863                 ret = ocfs2_create_refcount_tree(inode, di_bh);
3864                 if (ret) {
3865                         mlog_errno(ret);
3866                         goto out;
3867                 }
3868         }
3869
3870         BUG_ON(!di->i_refcount_loc);
3871         ret = ocfs2_lock_refcount_tree(osb,
3872                                        le64_to_cpu(di->i_refcount_loc), 1,
3873                                        &ref_tree, &ref_root_bh);
3874         if (ret) {
3875                 mlog_errno(ret);
3876                 goto out;
3877         }
3878
3879         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
3880                 goto attach_xattr;
3881
3882         ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh);
3883
3884         size = i_size_read(inode);
3885         clusters = ocfs2_clusters_for_bytes(inode->i_sb, size);
3886
3887         cpos = 0;
3888         while (cpos < clusters) {
3889                 ret = ocfs2_get_clusters(inode, cpos, &p_cluster,
3890                                          &num_clusters, &ext_flags);
3891
3892                 if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) {
3893                         ret = ocfs2_add_refcount_flag(inode, &di_et,
3894                                                       &ref_tree->rf_ci,
3895                                                       ref_root_bh, cpos,
3896                                                       p_cluster, num_clusters,
3897                                                       &dealloc, NULL);
3898                         if (ret) {
3899                                 mlog_errno(ret);
3900                                 goto unlock;
3901                         }
3902
3903                         data_changed = 1;
3904                 }
3905                 cpos += num_clusters;
3906         }
3907
3908 attach_xattr:
3909         if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
3910                 ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh,
3911                                                        &ref_tree->rf_ci,
3912                                                        ref_root_bh,
3913                                                        &dealloc);
3914                 if (ret) {
3915                         mlog_errno(ret);
3916                         goto unlock;
3917                 }
3918         }
3919
3920         if (data_changed) {
3921                 ret = ocfs2_change_ctime(inode, di_bh);
3922                 if (ret)
3923                         mlog_errno(ret);
3924         }
3925
3926 unlock:
3927         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
3928         brelse(ref_root_bh);
3929
3930         if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) {
3931                 ocfs2_schedule_truncate_log_flush(osb, 1);
3932                 ocfs2_run_deallocs(osb, &dealloc);
3933         }
3934 out:
3935         /*
3936          * Empty the extent map so that we may get the right extent
3937          * record from the disk.
3938          */
3939         ocfs2_extent_map_trunc(inode, 0);
3940
3941         return ret;
3942 }
3943
3944 static int ocfs2_add_refcounted_extent(struct inode *inode,
3945                                    struct ocfs2_extent_tree *et,
3946                                    struct ocfs2_caching_info *ref_ci,
3947                                    struct buffer_head *ref_root_bh,
3948                                    u32 cpos, u32 p_cluster, u32 num_clusters,
3949                                    unsigned int ext_flags,
3950                                    struct ocfs2_cached_dealloc_ctxt *dealloc)
3951 {
3952         int ret;
3953         handle_t *handle;
3954         int credits = 0;
3955         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
3956         struct ocfs2_alloc_context *meta_ac = NULL;
3957
3958         ret = ocfs2_lock_refcount_allocators(inode->i_sb,
3959                                              p_cluster, num_clusters,
3960                                              et, ref_ci,
3961                                              ref_root_bh, &meta_ac,
3962                                              NULL, &credits);
3963         if (ret) {
3964                 mlog_errno(ret);
3965                 goto out;
3966         }
3967
3968         handle = ocfs2_start_trans(osb, credits);
3969         if (IS_ERR(handle)) {
3970                 ret = PTR_ERR(handle);
3971                 mlog_errno(ret);
3972                 goto out;
3973         }
3974
3975         ret = ocfs2_insert_extent(handle, et, cpos,
3976                         ocfs2_clusters_to_blocks(inode->i_sb, p_cluster),
3977                         num_clusters, ext_flags, meta_ac);
3978         if (ret) {
3979                 mlog_errno(ret);
3980                 goto out_commit;
3981         }
3982
3983         ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh,
3984                                       p_cluster, num_clusters,
3985                                       meta_ac, dealloc);
3986         if (ret)
3987                 mlog_errno(ret);
3988
3989 out_commit:
3990         ocfs2_commit_trans(osb, handle);
3991 out:
3992         if (meta_ac)
3993                 ocfs2_free_alloc_context(meta_ac);
3994         return ret;
3995 }
3996
3997 static int ocfs2_duplicate_inline_data(struct inode *s_inode,
3998                                        struct buffer_head *s_bh,
3999                                        struct inode *t_inode,
4000                                        struct buffer_head *t_bh)
4001 {
4002         int ret;
4003         handle_t *handle;
4004         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
4005         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
4006         struct ocfs2_dinode *t_di = (struct ocfs2_dinode *)t_bh->b_data;
4007
4008         BUG_ON(!(OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
4009
4010         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
4011         if (IS_ERR(handle)) {
4012                 ret = PTR_ERR(handle);
4013                 mlog_errno(ret);
4014                 goto out;
4015         }
4016
4017         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
4018                                       OCFS2_JOURNAL_ACCESS_WRITE);
4019         if (ret) {
4020                 mlog_errno(ret);
4021                 goto out_commit;
4022         }
4023
4024         t_di->id2.i_data.id_count = s_di->id2.i_data.id_count;
4025         memcpy(t_di->id2.i_data.id_data, s_di->id2.i_data.id_data,
4026                le16_to_cpu(s_di->id2.i_data.id_count));
4027         spin_lock(&OCFS2_I(t_inode)->ip_lock);
4028         OCFS2_I(t_inode)->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
4029         t_di->i_dyn_features = cpu_to_le16(OCFS2_I(t_inode)->ip_dyn_features);
4030         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
4031
4032         ocfs2_journal_dirty(handle, t_bh);
4033
4034 out_commit:
4035         ocfs2_commit_trans(osb, handle);
4036 out:
4037         return ret;
4038 }
4039
4040 static int ocfs2_duplicate_extent_list(struct inode *s_inode,
4041                                 struct inode *t_inode,
4042                                 struct buffer_head *t_bh,
4043                                 struct ocfs2_caching_info *ref_ci,
4044                                 struct buffer_head *ref_root_bh,
4045                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
4046 {
4047         int ret = 0;
4048         u32 p_cluster, num_clusters, clusters, cpos;
4049         loff_t size;
4050         unsigned int ext_flags;
4051         struct ocfs2_extent_tree et;
4052
4053         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh);
4054
4055         size = i_size_read(s_inode);
4056         clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size);
4057
4058         cpos = 0;
4059         while (cpos < clusters) {
4060                 ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster,
4061                                          &num_clusters, &ext_flags);
4062
4063                 if (p_cluster) {
4064                         ret = ocfs2_add_refcounted_extent(t_inode, &et,
4065                                                           ref_ci, ref_root_bh,
4066                                                           cpos, p_cluster,
4067                                                           num_clusters,
4068                                                           ext_flags,
4069                                                           dealloc);
4070                         if (ret) {
4071                                 mlog_errno(ret);
4072                                 goto out;
4073                         }
4074                 }
4075
4076                 cpos += num_clusters;
4077         }
4078
4079 out:
4080         return ret;
4081 }
4082
4083 /*
4084  * change the new file's attributes to the src.
4085  *
4086  * reflink creates a snapshot of a file, that means the attributes
4087  * must be identical except for three exceptions - nlink, ino, and ctime.
4088  */
4089 static int ocfs2_complete_reflink(struct inode *s_inode,
4090                                   struct buffer_head *s_bh,
4091                                   struct inode *t_inode,
4092                                   struct buffer_head *t_bh,
4093                                   bool preserve)
4094 {
4095         int ret;
4096         handle_t *handle;
4097         struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data;
4098         struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data;
4099         loff_t size = i_size_read(s_inode);
4100
4101         handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb),
4102                                    OCFS2_INODE_UPDATE_CREDITS);
4103         if (IS_ERR(handle)) {
4104                 ret = PTR_ERR(handle);
4105                 mlog_errno(ret);
4106                 return ret;
4107         }
4108
4109         ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh,
4110                                       OCFS2_JOURNAL_ACCESS_WRITE);
4111         if (ret) {
4112                 mlog_errno(ret);
4113                 goto out_commit;
4114         }
4115
4116         spin_lock(&OCFS2_I(t_inode)->ip_lock);
4117         OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters;
4118         OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr;
4119         OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features;
4120         spin_unlock(&OCFS2_I(t_inode)->ip_lock);
4121         i_size_write(t_inode, size);
4122         t_inode->i_blocks = s_inode->i_blocks;
4123
4124         di->i_xattr_inline_size = s_di->i_xattr_inline_size;
4125         di->i_clusters = s_di->i_clusters;
4126         di->i_size = s_di->i_size;
4127         di->i_dyn_features = s_di->i_dyn_features;
4128         di->i_attr = s_di->i_attr;
4129
4130         if (preserve) {
4131                 t_inode->i_uid = s_inode->i_uid;
4132                 t_inode->i_gid = s_inode->i_gid;
4133                 t_inode->i_mode = s_inode->i_mode;
4134                 di->i_uid = s_di->i_uid;
4135                 di->i_gid = s_di->i_gid;
4136                 di->i_mode = s_di->i_mode;
4137
4138                 /*
4139                  * update time.
4140                  * we want mtime to appear identical to the source and
4141                  * update ctime.
4142                  */
4143                 t_inode->i_ctime = CURRENT_TIME;
4144
4145                 di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec);
4146                 di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec);
4147
4148                 t_inode->i_mtime = s_inode->i_mtime;
4149                 di->i_mtime = s_di->i_mtime;
4150                 di->i_mtime_nsec = s_di->i_mtime_nsec;
4151         }
4152
4153         ocfs2_journal_dirty(handle, t_bh);
4154
4155 out_commit:
4156         ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle);
4157         return ret;
4158 }
4159
4160 static int ocfs2_create_reflink_node(struct inode *s_inode,
4161                                      struct buffer_head *s_bh,
4162                                      struct inode *t_inode,
4163                                      struct buffer_head *t_bh,
4164                                      bool preserve)
4165 {
4166         int ret;
4167         struct buffer_head *ref_root_bh = NULL;
4168         struct ocfs2_cached_dealloc_ctxt dealloc;
4169         struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb);
4170         struct ocfs2_refcount_block *rb;
4171         struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data;
4172         struct ocfs2_refcount_tree *ref_tree;
4173
4174         ocfs2_init_dealloc_ctxt(&dealloc);
4175
4176         ret = ocfs2_set_refcount_tree(t_inode, t_bh,
4177                                       le64_to_cpu(di->i_refcount_loc));
4178         if (ret) {
4179                 mlog_errno(ret);
4180                 goto out;
4181         }
4182
4183         if (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4184                 ret = ocfs2_duplicate_inline_data(s_inode, s_bh,
4185                                                   t_inode, t_bh);
4186                 if (ret)
4187                         mlog_errno(ret);
4188                 goto out;
4189         }
4190
4191         ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc),
4192                                        1, &ref_tree, &ref_root_bh);
4193         if (ret) {
4194                 mlog_errno(ret);
4195                 goto out;
4196         }
4197         rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data;
4198
4199         ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh,
4200                                           &ref_tree->rf_ci, ref_root_bh,
4201                                           &dealloc);
4202         if (ret) {
4203                 mlog_errno(ret);
4204                 goto out_unlock_refcount;
4205         }
4206
4207 out_unlock_refcount:
4208         ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
4209         brelse(ref_root_bh);
4210 out:
4211         if (ocfs2_dealloc_has_cluster(&dealloc)) {
4212                 ocfs2_schedule_truncate_log_flush(osb, 1);
4213                 ocfs2_run_deallocs(osb, &dealloc);
4214         }
4215
4216         return ret;
4217 }
4218
4219 static int __ocfs2_reflink(struct dentry *old_dentry,
4220                            struct buffer_head *old_bh,
4221                            struct inode *new_inode,
4222                            bool preserve)
4223 {
4224         int ret;
4225         struct inode *inode = old_dentry->d_inode;
4226         struct buffer_head *new_bh = NULL;
4227
4228         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) {
4229                 ret = -EINVAL;
4230                 mlog_errno(ret);
4231                 goto out;
4232         }
4233
4234         ret = filemap_fdatawrite(inode->i_mapping);
4235         if (ret) {
4236                 mlog_errno(ret);
4237                 goto out;
4238         }
4239
4240         ret = ocfs2_attach_refcount_tree(inode, old_bh);
4241         if (ret) {
4242                 mlog_errno(ret);
4243                 goto out;
4244         }
4245
4246         mutex_lock_nested(&new_inode->i_mutex, I_MUTEX_CHILD);
4247         ret = ocfs2_inode_lock_nested(new_inode, &new_bh, 1,
4248                                       OI_LS_REFLINK_TARGET);
4249         if (ret) {
4250                 mlog_errno(ret);
4251                 goto out_unlock;
4252         }
4253
4254         ret = ocfs2_create_reflink_node(inode, old_bh,
4255                                         new_inode, new_bh, preserve);
4256         if (ret) {
4257                 mlog_errno(ret);
4258                 goto inode_unlock;
4259         }
4260
4261         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) {
4262                 ret = ocfs2_reflink_xattrs(inode, old_bh,
4263                                            new_inode, new_bh,
4264                                            preserve);
4265                 if (ret) {
4266                         mlog_errno(ret);
4267                         goto inode_unlock;
4268                 }
4269         }
4270
4271         ret = ocfs2_complete_reflink(inode, old_bh,
4272                                      new_inode, new_bh, preserve);
4273         if (ret)
4274                 mlog_errno(ret);
4275
4276 inode_unlock:
4277         ocfs2_inode_unlock(new_inode, 1);
4278         brelse(new_bh);
4279 out_unlock:
4280         mutex_unlock(&new_inode->i_mutex);
4281 out:
4282         if (!ret) {
4283                 ret = filemap_fdatawait(inode->i_mapping);
4284                 if (ret)
4285                         mlog_errno(ret);
4286         }
4287         return ret;
4288 }
4289
4290 static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir,
4291                          struct dentry *new_dentry, bool preserve)
4292 {
4293         int error;
4294         struct inode *inode = old_dentry->d_inode;
4295         struct buffer_head *old_bh = NULL;
4296         struct inode *new_orphan_inode = NULL;
4297
4298         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4299                 return -EOPNOTSUPP;
4300
4301         error = ocfs2_create_inode_in_orphan(dir, inode->i_mode,
4302                                              &new_orphan_inode);
4303         if (error) {
4304                 mlog_errno(error);
4305                 goto out;
4306         }
4307
4308         error = ocfs2_inode_lock(inode, &old_bh, 1);
4309         if (error) {
4310                 mlog_errno(error);
4311                 goto out;
4312         }
4313
4314         down_write(&OCFS2_I(inode)->ip_xattr_sem);
4315         down_write(&OCFS2_I(inode)->ip_alloc_sem);
4316         error = __ocfs2_reflink(old_dentry, old_bh,
4317                                 new_orphan_inode, preserve);
4318         up_write(&OCFS2_I(inode)->ip_alloc_sem);
4319         up_write(&OCFS2_I(inode)->ip_xattr_sem);
4320
4321         ocfs2_inode_unlock(inode, 1);
4322         brelse(old_bh);
4323
4324         if (error) {
4325                 mlog_errno(error);
4326                 goto out;
4327         }
4328
4329         /* If the security isn't preserved, we need to re-initialize them. */
4330         if (!preserve) {
4331                 error = ocfs2_init_security_and_acl(dir, new_orphan_inode,
4332                                                     &new_dentry->d_name);
4333                 if (error)
4334                         mlog_errno(error);
4335         }
4336 out:
4337         if (!error) {
4338                 error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode,
4339                                                        new_dentry);
4340                 if (error)
4341                         mlog_errno(error);
4342         }
4343
4344         if (new_orphan_inode) {
4345                 /*
4346                  * We need to open_unlock the inode no matter whether we
4347                  * succeed or not, so that other nodes can delete it later.
4348                  */
4349                 ocfs2_open_unlock(new_orphan_inode);
4350                 if (error)
4351                         iput(new_orphan_inode);
4352         }
4353
4354         return error;
4355 }
4356
4357 /*
4358  * Below here are the bits used by OCFS2_IOC_REFLINK() to fake
4359  * sys_reflink().  This will go away when vfs_reflink() exists in
4360  * fs/namei.c.
4361  */
4362
4363 /* copied from may_create in VFS. */
4364 static inline int ocfs2_may_create(struct inode *dir, struct dentry *child)
4365 {
4366         if (child->d_inode)
4367                 return -EEXIST;
4368         if (IS_DEADDIR(dir))
4369                 return -ENOENT;
4370         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
4371 }
4372
4373 /* copied from user_path_parent. */
4374 static int ocfs2_user_path_parent(const char __user *path,
4375                                   struct nameidata *nd, char **name)
4376 {
4377         char *s = getname(path);
4378         int error;
4379
4380         if (IS_ERR(s))
4381                 return PTR_ERR(s);
4382
4383         error = path_lookup(s, LOOKUP_PARENT, nd);
4384         if (error)
4385                 putname(s);
4386         else
4387                 *name = s;
4388
4389         return error;
4390 }
4391
4392 /**
4393  * ocfs2_vfs_reflink - Create a reference-counted link
4394  *
4395  * @old_dentry:        source dentry + inode
4396  * @dir:       directory to create the target
4397  * @new_dentry:        target dentry
4398  * @preserve:  if true, preserve all file attributes
4399  */
4400 static int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir,
4401                              struct dentry *new_dentry, bool preserve)
4402 {
4403         struct inode *inode = old_dentry->d_inode;
4404         int error;
4405
4406         if (!inode)
4407                 return -ENOENT;
4408
4409         error = ocfs2_may_create(dir, new_dentry);
4410         if (error)
4411                 return error;
4412
4413         if (dir->i_sb != inode->i_sb)
4414                 return -EXDEV;
4415
4416         /*
4417          * A reflink to an append-only or immutable file cannot be created.
4418          */
4419         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4420                 return -EPERM;
4421
4422         /* Only regular files can be reflinked. */
4423         if (!S_ISREG(inode->i_mode))
4424                 return -EPERM;
4425
4426         /*
4427          * If the caller wants to preserve ownership, they require the
4428          * rights to do so.
4429          */
4430         if (preserve) {
4431                 if ((current_fsuid() != inode->i_uid) && !capable(CAP_CHOWN))
4432                         return -EPERM;
4433                 if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN))
4434                         return -EPERM;
4435         }
4436
4437         /*
4438          * If the caller is modifying any aspect of the attributes, they
4439          * are not creating a snapshot.  They need read permission on the
4440          * file.
4441          */
4442         if (!preserve) {
4443                 error = inode_permission(inode, MAY_READ);
4444                 if (error)
4445                         return error;
4446         }
4447
4448         mutex_lock(&inode->i_mutex);
4449         dquot_initialize(dir);
4450         error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve);
4451         mutex_unlock(&inode->i_mutex);
4452         if (!error)
4453                 fsnotify_create(dir, new_dentry);
4454         return error;
4455 }
4456 /*
4457  * Most codes are copied from sys_linkat.
4458  */
4459 int ocfs2_reflink_ioctl(struct inode *inode,
4460                         const char __user *oldname,
4461                         const char __user *newname,
4462                         bool preserve)
4463 {
4464         struct dentry *new_dentry;
4465         struct nameidata nd;
4466         struct path old_path;
4467         int error;
4468         char *to = NULL;
4469
4470         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)))
4471                 return -EOPNOTSUPP;
4472
4473         error = user_path_at(AT_FDCWD, oldname, 0, &old_path);
4474         if (error) {
4475                 mlog_errno(error);
4476                 return error;
4477         }
4478
4479         error = ocfs2_user_path_parent(newname, &nd, &to);
4480         if (error) {
4481                 mlog_errno(error);
4482                 goto out;
4483         }
4484
4485         error = -EXDEV;
4486         if (old_path.mnt != nd.path.mnt)
4487                 goto out_release;
4488         new_dentry = lookup_create(&nd, 0);
4489         error = PTR_ERR(new_dentry);
4490         if (IS_ERR(new_dentry)) {
4491                 mlog_errno(error);
4492                 goto out_unlock;
4493         }
4494
4495         error = mnt_want_write(nd.path.mnt);
4496         if (error) {
4497                 mlog_errno(error);
4498                 goto out_dput;
4499         }
4500
4501         error = ocfs2_vfs_reflink(old_path.dentry,
4502                                   nd.path.dentry->d_inode,
4503                                   new_dentry, preserve);
4504         mnt_drop_write(nd.path.mnt);
4505 out_dput:
4506         dput(new_dentry);
4507 out_unlock:
4508         mutex_unlock(&nd.path.dentry->d_inode->i_mutex);
4509 out_release:
4510         path_put(&nd.path);
4511         putname(to);
4512 out:
4513         path_put(&old_path);
4514
4515         return error;
4516 }