fs/ocfs2/alloc.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* -*- mode: c; c-basic-offset: 8; -*-
   3  * vim: noexpandtab sw=8 ts=8 sts=0:
   4  *
   5  * alloc.c
   6  *
   7  * Extent allocs and frees
   8  *
   9  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
  10  */
  11
  12 #include <linux/fs.h>
  13 #include <linux/types.h>
  14 #include <linux/slab.h>
  15 #include <linux/highmem.h>
  16 #include <linux/swap.h>
  17 #include <linux/quotaops.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/sched/signal.h>
  20
  21 #include <cluster/masklog.h>
  22
  23 #include "ocfs2.h"
  24
  25 #include "alloc.h"
  26 #include "aops.h"
  27 #include "blockcheck.h"
  28 #include "dlmglue.h"
  29 #include "extent_map.h"
  30 #include "inode.h"
  31 #include "journal.h"
  32 #include "localalloc.h"
  33 #include "suballoc.h"
  34 #include "sysfile.h"
  35 #include "file.h"
  36 #include "super.h"
  37 #include "uptodate.h"
  38 #include "xattr.h"
  39 #include "refcounttree.h"
  40 #include "ocfs2_trace.h"
  41
  42 #include "buffer_head_io.h"
  43
  44 enum ocfs2_contig_type {
  45         CONTIG_NONE = 0,
  46         CONTIG_LEFT,
  47         CONTIG_RIGHT,
  48         CONTIG_LEFTRIGHT,
  49 };
  50
  51 static enum ocfs2_contig_type
  52         ocfs2_extent_rec_contig(struct super_block *sb,
  53                                 struct ocfs2_extent_rec *ext,
  54                                 struct ocfs2_extent_rec *insert_rec);
  55 /*
  56  * Operations for a specific extent tree type.
  57  *
  58  * To implement an on-disk btree (extent tree) type in ocfs2, add
  59  * an ocfs2_extent_tree_operations structure and the matching
  60  * ocfs2_init_<thingy>_extent_tree() function.  That's pretty much it
  61  * for the allocation portion of the extent tree.
  62  */
  63 struct ocfs2_extent_tree_operations {
  64         /*
  65          * last_eb_blk is the block number of the right most leaf extent
  66          * block.  Most on-disk structures containing an extent tree store
  67          * this value for fast access.  The ->eo_set_last_eb_blk() and
  68          * ->eo_get_last_eb_blk() operations access this value.  They are
  69          *  both required.
  70          */
  71         void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
  72                                    u64 blkno);
  73         u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
  74
  75         /*
  76          * The on-disk structure usually keeps track of how many total
  77          * clusters are stored in this extent tree.  This function updates
  78          * that value.  new_clusters is the delta, and must be
  79          * added to the total.  Required.
  80          */
  81         void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
  82                                    u32 new_clusters);
  83
  84         /*
  85          * If this extent tree is supported by an extent map, insert
  86          * a record into the map.
  87          */
  88         void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
  89                                      struct ocfs2_extent_rec *rec);
  90
  91         /*
  92          * If this extent tree is supported by an extent map, truncate the
  93          * map to clusters,
  94          */
  95         void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
  96                                        u32 clusters);
  97
  98         /*
  99          * If ->eo_insert_check() exists, it is called before rec is
 100          * inserted into the extent tree.  It is optional.
 101          */
 102         int (*eo_insert_check)(struct ocfs2_extent_tree *et,
 103                                struct ocfs2_extent_rec *rec);
 104         int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
 105
 106         /*
 107          * --------------------------------------------------------------
 108          * The remaining are internal to ocfs2_extent_tree and don't have
 109          * accessor functions
 110          */
 111
 112         /*
 113          * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
 114          * It is required.
 115          */
 116         void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
 117
 118         /*
 119          * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
 120          * it exists.  If it does not, et->et_max_leaf_clusters is set
 121          * to 0 (unlimited).  Optional.
 122          */
 123         void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
 124
 125         /*
 126          * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
 127          * are contiguous or not. Optional. Don't need to set it if use
 128          * ocfs2_extent_rec as the tree leaf.
 129          */
 130         enum ocfs2_contig_type
 131                 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
 132                                     struct ocfs2_extent_rec *ext,
 133                                     struct ocfs2_extent_rec *insert_rec);
 134 };
 135
 136
 137 /*
 138  * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
 139  * in the methods.
 140  */
 141 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
 142 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
 143                                          u64 blkno);
 144 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
 145                                          u32 clusters);
 146 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
 147                                            struct ocfs2_extent_rec *rec);
 148 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
 149                                              u32 clusters);
 150 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
 151                                      struct ocfs2_extent_rec *rec);
 152 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
 153 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
 154
 155 static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
 156                                         struct ocfs2_extent_tree *et,
 157                                         struct buffer_head **new_eb_bh,
 158                                         int blk_wanted, int *blk_given);
 159 static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et);
 160
 161 static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
 162         .eo_set_last_eb_blk     = ocfs2_dinode_set_last_eb_blk,
 163         .eo_get_last_eb_blk     = ocfs2_dinode_get_last_eb_blk,
 164         .eo_update_clusters     = ocfs2_dinode_update_clusters,
 165         .eo_extent_map_insert   = ocfs2_dinode_extent_map_insert,
 166         .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
 167         .eo_insert_check        = ocfs2_dinode_insert_check,
 168         .eo_sanity_check        = ocfs2_dinode_sanity_check,
 169         .eo_fill_root_el        = ocfs2_dinode_fill_root_el,
 170 };
 171
 172 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
 173                                          u64 blkno)
 174 {
 175         struct ocfs2_dinode *di = et->et_object;
 176
 177         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 178         di->i_last_eb_blk = cpu_to_le64(blkno);
 179 }
 180
 181 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
 182 {
 183         struct ocfs2_dinode *di = et->et_object;
 184
 185         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 186         return le64_to_cpu(di->i_last_eb_blk);
 187 }
 188
 189 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
 190                                          u32 clusters)
 191 {
 192         struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
 193         struct ocfs2_dinode *di = et->et_object;
 194
 195         le32_add_cpu(&di->i_clusters, clusters);
 196         spin_lock(&oi->ip_lock);
 197         oi->ip_clusters = le32_to_cpu(di->i_clusters);
 198         spin_unlock(&oi->ip_lock);
 199 }
 200
 201 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
 202                                            struct ocfs2_extent_rec *rec)
 203 {
 204         struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
 205
 206         ocfs2_extent_map_insert_rec(inode, rec);
 207 }
 208
 209 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
 210                                              u32 clusters)
 211 {
 212         struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
 213
 214         ocfs2_extent_map_trunc(inode, clusters);
 215 }
 216
 217 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
 218                                      struct ocfs2_extent_rec *rec)
 219 {
 220         struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
 221         struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
 222
 223         BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
 224         mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
 225                         (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
 226                         "Device %s, asking for sparse allocation: inode %llu, "
 227                         "cpos %u, clusters %u\n",
 228                         osb->dev_str,
 229                         (unsigned long long)oi->ip_blkno,
 230                         rec->e_cpos, oi->ip_clusters);
 231
 232         return 0;
 233 }
 234
 235 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
 236 {
 237         struct ocfs2_dinode *di = et->et_object;
 238
 239         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 240         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
 241
 242         return 0;
 243 }
 244
 245 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
 246 {
 247         struct ocfs2_dinode *di = et->et_object;
 248
 249         et->et_root_el = &di->id2.i_list;
 250 }
 251
 252
 253 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
 254 {
 255         struct ocfs2_xattr_value_buf *vb = et->et_object;
 256
 257         et->et_root_el = &vb->vb_xv->xr_list;
 258 }
 259
 260 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
 261                                               u64 blkno)
 262 {
 263         struct ocfs2_xattr_value_buf *vb = et->et_object;
 264
 265         vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
 266 }
 267
 268 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
 269 {
 270         struct ocfs2_xattr_value_buf *vb = et->et_object;
 271
 272         return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
 273 }
 274
 275 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
 276                                               u32 clusters)
 277 {
 278         struct ocfs2_xattr_value_buf *vb = et->et_object;
 279
 280         le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
 281 }
 282
 283 static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
 284         .eo_set_last_eb_blk     = ocfs2_xattr_value_set_last_eb_blk,
 285         .eo_get_last_eb_blk     = ocfs2_xattr_value_get_last_eb_blk,
 286         .eo_update_clusters     = ocfs2_xattr_value_update_clusters,
 287         .eo_fill_root_el        = ocfs2_xattr_value_fill_root_el,
 288 };
 289
 290 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
 291 {
 292         struct ocfs2_xattr_block *xb = et->et_object;
 293
 294         et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
 295 }
 296
 297 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
 298 {
 299         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
 300         et->et_max_leaf_clusters =
 301                 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
 302 }
 303
 304 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
 305                                              u64 blkno)
 306 {
 307         struct ocfs2_xattr_block *xb = et->et_object;
 308         struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
 309
 310         xt->xt_last_eb_blk = cpu_to_le64(blkno);
 311 }
 312
 313 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
 314 {
 315         struct ocfs2_xattr_block *xb = et->et_object;
 316         struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
 317
 318         return le64_to_cpu(xt->xt_last_eb_blk);
 319 }
 320
 321 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
 322                                              u32 clusters)
 323 {
 324         struct ocfs2_xattr_block *xb = et->et_object;
 325
 326         le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
 327 }
 328
 329 static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
 330         .eo_set_last_eb_blk     = ocfs2_xattr_tree_set_last_eb_blk,
 331         .eo_get_last_eb_blk     = ocfs2_xattr_tree_get_last_eb_blk,
 332         .eo_update_clusters     = ocfs2_xattr_tree_update_clusters,
 333         .eo_fill_root_el        = ocfs2_xattr_tree_fill_root_el,
 334         .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
 335 };
 336
 337 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
 338                                           u64 blkno)
 339 {
 340         struct ocfs2_dx_root_block *dx_root = et->et_object;
 341
 342         dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
 343 }
 344
 345 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
 346 {
 347         struct ocfs2_dx_root_block *dx_root = et->et_object;
 348
 349         return le64_to_cpu(dx_root->dr_last_eb_blk);
 350 }
 351
 352 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
 353                                           u32 clusters)
 354 {
 355         struct ocfs2_dx_root_block *dx_root = et->et_object;
 356
 357         le32_add_cpu(&dx_root->dr_clusters, clusters);
 358 }
 359
 360 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
 361 {
 362         struct ocfs2_dx_root_block *dx_root = et->et_object;
 363
 364         BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
 365
 366         return 0;
 367 }
 368
 369 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
 370 {
 371         struct ocfs2_dx_root_block *dx_root = et->et_object;
 372
 373         et->et_root_el = &dx_root->dr_list;
 374 }
 375
 376 static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
 377         .eo_set_last_eb_blk     = ocfs2_dx_root_set_last_eb_blk,
 378         .eo_get_last_eb_blk     = ocfs2_dx_root_get_last_eb_blk,
 379         .eo_update_clusters     = ocfs2_dx_root_update_clusters,
 380         .eo_sanity_check        = ocfs2_dx_root_sanity_check,
 381         .eo_fill_root_el        = ocfs2_dx_root_fill_root_el,
 382 };
 383
 384 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
 385 {
 386         struct ocfs2_refcount_block *rb = et->et_object;
 387
 388         et->et_root_el = &rb->rf_list;
 389 }
 390
 391 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
 392                                                 u64 blkno)
 393 {
 394         struct ocfs2_refcount_block *rb = et->et_object;
 395
 396         rb->rf_last_eb_blk = cpu_to_le64(blkno);
 397 }
 398
 399 static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
 400 {
 401         struct ocfs2_refcount_block *rb = et->et_object;
 402
 403         return le64_to_cpu(rb->rf_last_eb_blk);
 404 }
 405
 406 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
 407                                                 u32 clusters)
 408 {
 409         struct ocfs2_refcount_block *rb = et->et_object;
 410
 411         le32_add_cpu(&rb->rf_clusters, clusters);
 412 }
 413
 414 static enum ocfs2_contig_type
 415 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
 416                                   struct ocfs2_extent_rec *ext,
 417                                   struct ocfs2_extent_rec *insert_rec)
 418 {
 419         return CONTIG_NONE;
 420 }
 421
 422 static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
 423         .eo_set_last_eb_blk     = ocfs2_refcount_tree_set_last_eb_blk,
 424         .eo_get_last_eb_blk     = ocfs2_refcount_tree_get_last_eb_blk,
 425         .eo_update_clusters     = ocfs2_refcount_tree_update_clusters,
 426         .eo_fill_root_el        = ocfs2_refcount_tree_fill_root_el,
 427         .eo_extent_contig       = ocfs2_refcount_tree_extent_contig,
 428 };
 429
 430 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
 431                                      struct ocfs2_caching_info *ci,
 432                                      struct buffer_head *bh,
 433                                      ocfs2_journal_access_func access,
 434                                      void *obj,
 435                                      const struct ocfs2_extent_tree_operations *ops)
 436 {
 437         et->et_ops = ops;
 438         et->et_root_bh = bh;
 439         et->et_ci = ci;
 440         et->et_root_journal_access = access;
 441         if (!obj)
 442                 obj = (void *)bh->b_data;
 443         et->et_object = obj;
 444         et->et_dealloc = NULL;
 445
 446         et->et_ops->eo_fill_root_el(et);
 447         if (!et->et_ops->eo_fill_max_leaf_clusters)
 448                 et->et_max_leaf_clusters = 0;
 449         else
 450                 et->et_ops->eo_fill_max_leaf_clusters(et);
 451 }
 452
 453 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
 454                                    struct ocfs2_caching_info *ci,
 455                                    struct buffer_head *bh)
 456 {
 457         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
 458                                  NULL, &ocfs2_dinode_et_ops);
 459 }
 460
 461 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
 462                                        struct ocfs2_caching_info *ci,
 463                                        struct buffer_head *bh)
 464 {
 465         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
 466                                  NULL, &ocfs2_xattr_tree_et_ops);
 467 }
 468
 469 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
 470                                         struct ocfs2_caching_info *ci,
 471                                         struct ocfs2_xattr_value_buf *vb)
 472 {
 473         __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
 474                                  &ocfs2_xattr_value_et_ops);
 475 }
 476
 477 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
 478                                     struct ocfs2_caching_info *ci,
 479                                     struct buffer_head *bh)
 480 {
 481         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
 482                                  NULL, &ocfs2_dx_root_et_ops);
 483 }
 484
 485 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
 486                                      struct ocfs2_caching_info *ci,
 487                                      struct buffer_head *bh)
 488 {
 489         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
 490                                  NULL, &ocfs2_refcount_tree_et_ops);
 491 }
 492
 493 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
 494                                             u64 new_last_eb_blk)
 495 {
 496         et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
 497 }
 498
 499 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
 500 {
 501         return et->et_ops->eo_get_last_eb_blk(et);
 502 }
 503
 504 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
 505                                             u32 clusters)
 506 {
 507         et->et_ops->eo_update_clusters(et, clusters);
 508 }
 509
 510 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
 511                                               struct ocfs2_extent_rec *rec)
 512 {
 513         if (et->et_ops->eo_extent_map_insert)
 514                 et->et_ops->eo_extent_map_insert(et, rec);
 515 }
 516
 517 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
 518                                                 u32 clusters)
 519 {
 520         if (et->et_ops->eo_extent_map_truncate)
 521                 et->et_ops->eo_extent_map_truncate(et, clusters);
 522 }
 523
 524 static inline int ocfs2_et_root_journal_access(handle_t *handle,
 525                                                struct ocfs2_extent_tree *et,
 526                                                int type)
 527 {
 528         return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
 529                                           type);
 530 }
 531
 532 static inline enum ocfs2_contig_type
 533         ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
 534                                struct ocfs2_extent_rec *rec,
 535                                struct ocfs2_extent_rec *insert_rec)
 536 {
 537         if (et->et_ops->eo_extent_contig)
 538                 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
 539
 540         return ocfs2_extent_rec_contig(
 541                                 ocfs2_metadata_cache_get_super(et->et_ci),
 542                                 rec, insert_rec);
 543 }
 544
 545 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
 546                                         struct ocfs2_extent_rec *rec)
 547 {
 548         int ret = 0;
 549
 550         if (et->et_ops->eo_insert_check)
 551                 ret = et->et_ops->eo_insert_check(et, rec);
 552         return ret;
 553 }
 554
 555 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
 556 {
 557         int ret = 0;
 558
 559         if (et->et_ops->eo_sanity_check)
 560                 ret = et->et_ops->eo_sanity_check(et);
 561         return ret;
 562 }
 563
 564 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
 565                                          struct ocfs2_extent_block *eb);
 566 static void ocfs2_adjust_rightmost_records(handle_t *handle,
 567                                            struct ocfs2_extent_tree *et,
 568                                            struct ocfs2_path *path,
 569                                            struct ocfs2_extent_rec *insert_rec);
 570 /*
 571  * Reset the actual path elements so that we can re-use the structure
 572  * to build another path. Generally, this involves freeing the buffer
 573  * heads.
 574  */
 575 void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
 576 {
 577         int i, start = 0, depth = 0;
 578         struct ocfs2_path_item *node;
 579
 580         if (keep_root)
 581                 start = 1;
 582
 583         for(i = start; i < path_num_items(path); i++) {
 584                 node = &path->p_node[i];
 585
 586                 brelse(node->bh);
 587                 node->bh = NULL;
 588                 node->el = NULL;
 589         }
 590
 591         /*
 592          * Tree depth may change during truncate, or insert. If we're
 593          * keeping the root extent list, then make sure that our path
 594          * structure reflects the proper depth.
 595          */
 596         if (keep_root)
 597                 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
 598         else
 599                 path_root_access(path) = NULL;
 600
 601         path->p_tree_depth = depth;
 602 }
 603
 604 void ocfs2_free_path(struct ocfs2_path *path)
 605 {
 606         if (path) {
 607                 ocfs2_reinit_path(path, 0);
 608                 kfree(path);
 609         }
 610 }
 611
 612 /*
 613  * All the elements of src into dest. After this call, src could be freed
 614  * without affecting dest.
 615  *
 616  * Both paths should have the same root. Any non-root elements of dest
 617  * will be freed.
 618  */
 619 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
 620 {
 621         int i;
 622
 623         BUG_ON(path_root_bh(dest) != path_root_bh(src));
 624         BUG_ON(path_root_el(dest) != path_root_el(src));
 625         BUG_ON(path_root_access(dest) != path_root_access(src));
 626
 627         ocfs2_reinit_path(dest, 1);
 628
 629         for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
 630                 dest->p_node[i].bh = src->p_node[i].bh;
 631                 dest->p_node[i].el = src->p_node[i].el;
 632
 633                 if (dest->p_node[i].bh)
 634                         get_bh(dest->p_node[i].bh);
 635         }
 636 }
 637
 638 /*
 639  * Make the *dest path the same as src and re-initialize src path to
 640  * have a root only.
 641  */
 642 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
 643 {
 644         int i;
 645
 646         BUG_ON(path_root_bh(dest) != path_root_bh(src));
 647         BUG_ON(path_root_access(dest) != path_root_access(src));
 648
 649         for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
 650                 brelse(dest->p_node[i].bh);
 651
 652                 dest->p_node[i].bh = src->p_node[i].bh;
 653                 dest->p_node[i].el = src->p_node[i].el;
 654
 655                 src->p_node[i].bh = NULL;
 656                 src->p_node[i].el = NULL;
 657         }
 658 }
 659
 660 /*
 661  * Insert an extent block at given index.
 662  *
 663  * This will not take an additional reference on eb_bh.
 664  */
 665 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
 666                                         struct buffer_head *eb_bh)
 667 {
 668         struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
 669
 670         /*
 671          * Right now, no root bh is an extent block, so this helps
 672          * catch code errors with dinode trees. The assertion can be
 673          * safely removed if we ever need to insert extent block
 674          * structures at the root.
 675          */
 676         BUG_ON(index == 0);
 677
 678         path->p_node[index].bh = eb_bh;
 679         path->p_node[index].el = &eb->h_list;
 680 }
 681
 682 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
 683                                          struct ocfs2_extent_list *root_el,
 684                                          ocfs2_journal_access_func access)
 685 {
 686         struct ocfs2_path *path;
 687
 688         BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
 689
 690         path = kzalloc(sizeof(*path), GFP_NOFS);
 691         if (path) {
 692                 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
 693                 get_bh(root_bh);
 694                 path_root_bh(path) = root_bh;
 695                 path_root_el(path) = root_el;
 696                 path_root_access(path) = access;
 697         }
 698
 699         return path;
 700 }
 701
 702 struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
 703 {
 704         return ocfs2_new_path(path_root_bh(path), path_root_el(path),
 705                               path_root_access(path));
 706 }
 707
 708 struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
 709 {
 710         return ocfs2_new_path(et->et_root_bh, et->et_root_el,
 711                               et->et_root_journal_access);
 712 }
 713
 714 /*
 715  * Journal the buffer at depth idx.  All idx>0 are extent_blocks,
 716  * otherwise it's the root_access function.
 717  *
 718  * I don't like the way this function's name looks next to
 719  * ocfs2_journal_access_path(), but I don't have a better one.
 720  */
 721 int ocfs2_path_bh_journal_access(handle_t *handle,
 722                                  struct ocfs2_caching_info *ci,
 723                                  struct ocfs2_path *path,
 724                                  int idx)
 725 {
 726         ocfs2_journal_access_func access = path_root_access(path);
 727
 728         if (!access)
 729                 access = ocfs2_journal_access;
 730
 731         if (idx)
 732                 access = ocfs2_journal_access_eb;
 733
 734         return access(handle, ci, path->p_node[idx].bh,
 735                       OCFS2_JOURNAL_ACCESS_WRITE);
 736 }
 737
 738 /*
 739  * Convenience function to journal all components in a path.
 740  */
 741 int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
 742                               handle_t *handle,
 743                               struct ocfs2_path *path)
 744 {
 745         int i, ret = 0;
 746
 747         if (!path)
 748                 goto out;
 749
 750         for(i = 0; i < path_num_items(path); i++) {
 751                 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
 752                 if (ret < 0) {
 753                         mlog_errno(ret);
 754                         goto out;
 755                 }
 756         }
 757
 758 out:
 759         return ret;
 760 }
 761
 762 /*
 763  * Return the index of the extent record which contains cluster #v_cluster.
 764  * -1 is returned if it was not found.
 765  *
 766  * Should work fine on interior and exterior nodes.
 767  */
 768 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
 769 {
 770         int ret = -1;
 771         int i;
 772         struct ocfs2_extent_rec *rec;
 773         u32 rec_end, rec_start, clusters;
 774
 775         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
 776                 rec = &el->l_recs[i];
 777
 778                 rec_start = le32_to_cpu(rec->e_cpos);
 779                 clusters = ocfs2_rec_clusters(el, rec);
 780
 781                 rec_end = rec_start + clusters;
 782
 783                 if (v_cluster >= rec_start && v_cluster < rec_end) {
 784                         ret = i;
 785                         break;
 786                 }
 787         }
 788
 789         return ret;
 790 }
 791
 792 /*
 793  * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
 794  * ocfs2_extent_rec_contig only work properly against leaf nodes!
 795  */
 796 static int ocfs2_block_extent_contig(struct super_block *sb,
 797                                      struct ocfs2_extent_rec *ext,
 798                                      u64 blkno)
 799 {
 800         u64 blk_end = le64_to_cpu(ext->e_blkno);
 801
 802         blk_end += ocfs2_clusters_to_blocks(sb,
 803                                     le16_to_cpu(ext->e_leaf_clusters));
 804
 805         return blkno == blk_end;
 806 }
 807
 808 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
 809                                   struct ocfs2_extent_rec *right)
 810 {
 811         u32 left_range;
 812
 813         left_range = le32_to_cpu(left->e_cpos) +
 814                 le16_to_cpu(left->e_leaf_clusters);
 815
 816         return (left_range == le32_to_cpu(right->e_cpos));
 817 }
 818
 819 static enum ocfs2_contig_type
 820         ocfs2_extent_rec_contig(struct super_block *sb,
 821                                 struct ocfs2_extent_rec *ext,
 822                                 struct ocfs2_extent_rec *insert_rec)
 823 {
 824         u64 blkno = le64_to_cpu(insert_rec->e_blkno);
 825
 826         /*
 827          * Refuse to coalesce extent records with different flag
 828          * fields - we don't want to mix unwritten extents with user
 829          * data.
 830          */
 831         if (ext->e_flags != insert_rec->e_flags)
 832                 return CONTIG_NONE;
 833
 834         if (ocfs2_extents_adjacent(ext, insert_rec) &&
 835             ocfs2_block_extent_contig(sb, ext, blkno))
 836                         return CONTIG_RIGHT;
 837
 838         blkno = le64_to_cpu(ext->e_blkno);
 839         if (ocfs2_extents_adjacent(insert_rec, ext) &&
 840             ocfs2_block_extent_contig(sb, insert_rec, blkno))
 841                 return CONTIG_LEFT;
 842
 843         return CONTIG_NONE;
 844 }
 845
 846 /*
 847  * NOTE: We can have pretty much any combination of contiguousness and
 848  * appending.
 849  *
 850  * The usefulness of APPEND_TAIL is more in that it lets us know that
 851  * we'll have to update the path to that leaf.
 852  */
 853 enum ocfs2_append_type {
 854         APPEND_NONE = 0,
 855         APPEND_TAIL,
 856 };
 857
 858 enum ocfs2_split_type {
 859         SPLIT_NONE = 0,
 860         SPLIT_LEFT,
 861         SPLIT_RIGHT,
 862 };
 863
 864 struct ocfs2_insert_type {
 865         enum ocfs2_split_type   ins_split;
 866         enum ocfs2_append_type  ins_appending;
 867         enum ocfs2_contig_type  ins_contig;
 868         int                     ins_contig_index;
 869         int                     ins_tree_depth;
 870 };
 871
 872 struct ocfs2_merge_ctxt {
 873         enum ocfs2_contig_type  c_contig_type;
 874         int                     c_has_empty_extent;
 875         int                     c_split_covers_rec;
 876 };
 877
 878 static int ocfs2_validate_extent_block(struct super_block *sb,
 879                                        struct buffer_head *bh)
 880 {
 881         int rc;
 882         struct ocfs2_extent_block *eb =
 883                 (struct ocfs2_extent_block *)bh->b_data;
 884
 885         trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
 886
 887         BUG_ON(!buffer_uptodate(bh));
 888
 889         /*
 890          * If the ecc fails, we return the error but otherwise
 891          * leave the filesystem running.  We know any error is
 892          * local to this block.
 893          */
 894         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
 895         if (rc) {
 896                 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
 897                      (unsigned long long)bh->b_blocknr);
 898                 return rc;
 899         }
 900
 901         /*
 902          * Errors after here are fatal.
 903          */
 904
 905         if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
 906                 rc = ocfs2_error(sb,
 907                                  "Extent block #%llu has bad signature %.*s\n",
 908                                  (unsigned long long)bh->b_blocknr, 7,
 909                                  eb->h_signature);
 910                 goto bail;
 911         }
 912
 913         if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
 914                 rc = ocfs2_error(sb,
 915                                  "Extent block #%llu has an invalid h_blkno of %llu\n",
 916                                  (unsigned long long)bh->b_blocknr,
 917                                  (unsigned long long)le64_to_cpu(eb->h_blkno));
 918                 goto bail;
 919         }
 920
 921         if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation)
 922                 rc = ocfs2_error(sb,
 923                                  "Extent block #%llu has an invalid h_fs_generation of #%u\n",
 924                                  (unsigned long long)bh->b_blocknr,
 925                                  le32_to_cpu(eb->h_fs_generation));
 926 bail:
 927         return rc;
 928 }
 929
 930 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
 931                             struct buffer_head **bh)
 932 {
 933         int rc;
 934         struct buffer_head *tmp = *bh;
 935
 936         rc = ocfs2_read_block(ci, eb_blkno, &tmp,
 937                               ocfs2_validate_extent_block);
 938
 939         /* If ocfs2_read_block() got us a new bh, pass it up. */
 940         if (!rc && !*bh)
 941                 *bh = tmp;
 942
 943         return rc;
 944 }
 945
 946
 947 /*
 948  * How many free extents have we got before we need more meta data?
 949  */
 950 int ocfs2_num_free_extents(struct ocfs2_extent_tree *et)
 951 {
 952         int retval;
 953         struct ocfs2_extent_list *el = NULL;
 954         struct ocfs2_extent_block *eb;
 955         struct buffer_head *eb_bh = NULL;
 956         u64 last_eb_blk = 0;
 957
 958         el = et->et_root_el;
 959         last_eb_blk = ocfs2_et_get_last_eb_blk(et);
 960
 961         if (last_eb_blk) {
 962                 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
 963                                                  &eb_bh);
 964                 if (retval < 0) {
 965                         mlog_errno(retval);
 966                         goto bail;
 967                 }
 968                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 969                 el = &eb->h_list;
 970         }
 971
 972         BUG_ON(el->l_tree_depth != 0);
 973
 974         retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
 975 bail:
 976         brelse(eb_bh);
 977
 978         trace_ocfs2_num_free_extents(retval);
 979         return retval;
 980 }
 981
 982 /* expects array to already be allocated
 983  *
 984  * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
 985  * l_count for you
 986  */
 987 static int ocfs2_create_new_meta_bhs(handle_t *handle,
 988                                      struct ocfs2_extent_tree *et,
 989                                      int wanted,
 990                                      struct ocfs2_alloc_context *meta_ac,
 991                                      struct buffer_head *bhs[])
 992 {
 993         int count, status, i;
 994         u16 suballoc_bit_start;
 995         u32 num_got;
 996         u64 suballoc_loc, first_blkno;
 997         struct ocfs2_super *osb =
 998                 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
 999         struct ocfs2_extent_block *eb;
1000
1001         count = 0;
1002         while (count < wanted) {
1003                 status = ocfs2_claim_metadata(handle,
1004                                               meta_ac,
1005                                               wanted - count,
1006                                               &suballoc_loc,
1007                                               &suballoc_bit_start,
1008                                               &num_got,
1009                                               &first_blkno);
1010                 if (status < 0) {
1011                         mlog_errno(status);
1012                         goto bail;
1013                 }
1014
1015                 for(i = count;  i < (num_got + count); i++) {
1016                         bhs[i] = sb_getblk(osb->sb, first_blkno);
1017                         if (bhs[i] == NULL) {
1018                                 status = -ENOMEM;
1019                                 mlog_errno(status);
1020                                 goto bail;
1021                         }
1022                         ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1023
1024                         status = ocfs2_journal_access_eb(handle, et->et_ci,
1025                                                          bhs[i],
1026                                                          OCFS2_JOURNAL_ACCESS_CREATE);
1027                         if (status < 0) {
1028                                 mlog_errno(status);
1029                                 goto bail;
1030                         }
1031
1032                         memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1033                         eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1034                         /* Ok, setup the minimal stuff here. */
1035                         strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1036                         eb->h_blkno = cpu_to_le64(first_blkno);
1037                         eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1038                         eb->h_suballoc_slot =
1039                                 cpu_to_le16(meta_ac->ac_alloc_slot);
1040                         eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1041                         eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1042                         eb->h_list.l_count =
1043                                 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1044
1045                         suballoc_bit_start++;
1046                         first_blkno++;
1047
1048                         /* We'll also be dirtied by the caller, so
1049                          * this isn't absolutely necessary. */
1050                         ocfs2_journal_dirty(handle, bhs[i]);
1051                 }
1052
1053                 count += num_got;
1054         }
1055
1056         status = 0;
1057 bail:
1058         if (status < 0) {
1059                 for(i = 0; i < wanted; i++) {
1060                         brelse(bhs[i]);
1061                         bhs[i] = NULL;
1062                 }
1063                 mlog_errno(status);
1064         }
1065         return status;
1066 }
1067
1068 /*
1069  * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1070  *
1071  * Returns the sum of the rightmost extent rec logical offset and
1072  * cluster count.
1073  *
1074  * ocfs2_add_branch() uses this to determine what logical cluster
1075  * value should be populated into the leftmost new branch records.
1076  *
1077  * ocfs2_shift_tree_depth() uses this to determine the # clusters
1078  * value for the new topmost tree record.
1079  */
1080 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list  *el)
1081 {
1082         int i;
1083
1084         i = le16_to_cpu(el->l_next_free_rec) - 1;
1085
1086         return le32_to_cpu(el->l_recs[i].e_cpos) +
1087                 ocfs2_rec_clusters(el, &el->l_recs[i]);
1088 }
1089
1090 /*
1091  * Change range of the branches in the right most path according to the leaf
1092  * extent block's rightmost record.
1093  */
1094 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1095                                          struct ocfs2_extent_tree *et)
1096 {
1097         int status;
1098         struct ocfs2_path *path = NULL;
1099         struct ocfs2_extent_list *el;
1100         struct ocfs2_extent_rec *rec;
1101
1102         path = ocfs2_new_path_from_et(et);
1103         if (!path) {
1104                 status = -ENOMEM;
1105                 return status;
1106         }
1107
1108         status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1109         if (status < 0) {
1110                 mlog_errno(status);
1111                 goto out;
1112         }
1113
1114         status = ocfs2_extend_trans(handle, path_num_items(path));
1115         if (status < 0) {
1116                 mlog_errno(status);
1117                 goto out;
1118         }
1119
1120         status = ocfs2_journal_access_path(et->et_ci, handle, path);
1121         if (status < 0) {
1122                 mlog_errno(status);
1123                 goto out;
1124         }
1125
1126         el = path_leaf_el(path);
1127         rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1128
1129         ocfs2_adjust_rightmost_records(handle, et, path, rec);
1130
1131 out:
1132         ocfs2_free_path(path);
1133         return status;
1134 }
1135
1136 /*
1137  * Add an entire tree branch to our inode. eb_bh is the extent block
1138  * to start at, if we don't want to start the branch at the root
1139  * structure.
1140  *
1141  * last_eb_bh is required as we have to update it's next_leaf pointer
1142  * for the new last extent block.
1143  *
1144  * the new branch will be 'empty' in the sense that every block will
1145  * contain a single record with cluster count == 0.
1146  */
1147 static int ocfs2_add_branch(handle_t *handle,
1148                             struct ocfs2_extent_tree *et,
1149                             struct buffer_head *eb_bh,
1150                             struct buffer_head **last_eb_bh,
1151                             struct ocfs2_alloc_context *meta_ac)
1152 {
1153         int status, new_blocks, i, block_given = 0;
1154         u64 next_blkno, new_last_eb_blk;
1155         struct buffer_head *bh;
1156         struct buffer_head **new_eb_bhs = NULL;
1157         struct ocfs2_extent_block *eb;
1158         struct ocfs2_extent_list  *eb_el;
1159         struct ocfs2_extent_list  *el;
1160         u32 new_cpos, root_end;
1161
1162         BUG_ON(!last_eb_bh || !*last_eb_bh);
1163
1164         if (eb_bh) {
1165                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1166                 el = &eb->h_list;
1167         } else
1168                 el = et->et_root_el;
1169
1170         /* we never add a branch to a leaf. */
1171         BUG_ON(!el->l_tree_depth);
1172
1173         new_blocks = le16_to_cpu(el->l_tree_depth);
1174
1175         eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1176         new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1177         root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1178
1179         /*
1180          * If there is a gap before the root end and the real end
1181          * of the righmost leaf block, we need to remove the gap
1182          * between new_cpos and root_end first so that the tree
1183          * is consistent after we add a new branch(it will start
1184          * from new_cpos).
1185          */
1186         if (root_end > new_cpos) {
1187                 trace_ocfs2_adjust_rightmost_branch(
1188                         (unsigned long long)
1189                         ocfs2_metadata_cache_owner(et->et_ci),
1190                         root_end, new_cpos);
1191
1192                 status = ocfs2_adjust_rightmost_branch(handle, et);
1193                 if (status) {
1194                         mlog_errno(status);
1195                         goto bail;
1196                 }
1197         }
1198
1199         /* allocate the number of new eb blocks we need */
1200         new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1201                              GFP_KERNEL);
1202         if (!new_eb_bhs) {
1203                 status = -ENOMEM;
1204                 mlog_errno(status);
1205                 goto bail;
1206         }
1207
1208         /* Firstyly, try to reuse dealloc since we have already estimated how
1209          * many extent blocks we may use.
1210          */
1211         if (!ocfs2_is_dealloc_empty(et)) {
1212                 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1213                                                       new_eb_bhs, new_blocks,
1214                                                       &block_given);
1215                 if (status < 0) {
1216                         mlog_errno(status);
1217                         goto bail;
1218                 }
1219         }
1220
1221         BUG_ON(block_given > new_blocks);
1222
1223         if (block_given < new_blocks) {
1224                 BUG_ON(!meta_ac);
1225                 status = ocfs2_create_new_meta_bhs(handle, et,
1226                                                    new_blocks - block_given,
1227                                                    meta_ac,
1228                                                    &new_eb_bhs[block_given]);
1229                 if (status < 0) {
1230                         mlog_errno(status);
1231                         goto bail;
1232                 }
1233         }
1234
1235         /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1236          * linked with the rest of the tree.
1237          * conversly, new_eb_bhs[0] is the new bottommost leaf.
1238          *
1239          * when we leave the loop, new_last_eb_blk will point to the
1240          * newest leaf, and next_blkno will point to the topmost extent
1241          * block. */
1242         next_blkno = new_last_eb_blk = 0;
1243         for(i = 0; i < new_blocks; i++) {
1244                 bh = new_eb_bhs[i];
1245                 eb = (struct ocfs2_extent_block *) bh->b_data;
1246                 /* ocfs2_create_new_meta_bhs() should create it right! */
1247                 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1248                 eb_el = &eb->h_list;
1249
1250                 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1251                                                  OCFS2_JOURNAL_ACCESS_CREATE);
1252                 if (status < 0) {
1253                         mlog_errno(status);
1254                         goto bail;
1255                 }
1256
1257                 eb->h_next_leaf_blk = 0;
1258                 eb_el->l_tree_depth = cpu_to_le16(i);
1259                 eb_el->l_next_free_rec = cpu_to_le16(1);
1260                 /*
1261                  * This actually counts as an empty extent as
1262                  * c_clusters == 0
1263                  */
1264                 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1265                 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1266                 /*
1267                  * eb_el isn't always an interior node, but even leaf
1268                  * nodes want a zero'd flags and reserved field so
1269                  * this gets the whole 32 bits regardless of use.
1270                  */
1271                 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1272                 if (!eb_el->l_tree_depth)
1273                         new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1274
1275                 ocfs2_journal_dirty(handle, bh);
1276                 next_blkno = le64_to_cpu(eb->h_blkno);
1277         }
1278
1279         /* This is a bit hairy. We want to update up to three blocks
1280          * here without leaving any of them in an inconsistent state
1281          * in case of error. We don't have to worry about
1282          * journal_dirty erroring as it won't unless we've aborted the
1283          * handle (in which case we would never be here) so reserving
1284          * the write with journal_access is all we need to do. */
1285         status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1286                                          OCFS2_JOURNAL_ACCESS_WRITE);
1287         if (status < 0) {
1288                 mlog_errno(status);
1289                 goto bail;
1290         }
1291         status = ocfs2_et_root_journal_access(handle, et,
1292                                               OCFS2_JOURNAL_ACCESS_WRITE);
1293         if (status < 0) {
1294                 mlog_errno(status);
1295                 goto bail;
1296         }
1297         if (eb_bh) {
1298                 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1299                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1300                 if (status < 0) {
1301                         mlog_errno(status);
1302                         goto bail;
1303                 }
1304         }
1305
1306         /* Link the new branch into the rest of the tree (el will
1307          * either be on the root_bh, or the extent block passed in. */
1308         i = le16_to_cpu(el->l_next_free_rec);
1309         el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1310         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1311         el->l_recs[i].e_int_clusters = 0;
1312         le16_add_cpu(&el->l_next_free_rec, 1);
1313
1314         /* fe needs a new last extent block pointer, as does the
1315          * next_leaf on the previously last-extent-block. */
1316         ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1317
1318         eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1319         eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1320
1321         ocfs2_journal_dirty(handle, *last_eb_bh);
1322         ocfs2_journal_dirty(handle, et->et_root_bh);
1323         if (eb_bh)
1324                 ocfs2_journal_dirty(handle, eb_bh);
1325
1326         /*
1327          * Some callers want to track the rightmost leaf so pass it
1328          * back here.
1329          */
1330         brelse(*last_eb_bh);
1331         get_bh(new_eb_bhs[0]);
1332         *last_eb_bh = new_eb_bhs[0];
1333
1334         status = 0;
1335 bail:
1336         if (new_eb_bhs) {
1337                 for (i = 0; i < new_blocks; i++)
1338                         brelse(new_eb_bhs[i]);
1339                 kfree(new_eb_bhs);
1340         }
1341
1342         return status;
1343 }
1344
1345 /*
1346  * adds another level to the allocation tree.
1347  * returns back the new extent block so you can add a branch to it
1348  * after this call.
1349  */
1350 static int ocfs2_shift_tree_depth(handle_t *handle,
1351                                   struct ocfs2_extent_tree *et,
1352                                   struct ocfs2_alloc_context *meta_ac,
1353                                   struct buffer_head **ret_new_eb_bh)
1354 {
1355         int status, i, block_given = 0;
1356         u32 new_clusters;
1357         struct buffer_head *new_eb_bh = NULL;
1358         struct ocfs2_extent_block *eb;
1359         struct ocfs2_extent_list  *root_el;
1360         struct ocfs2_extent_list  *eb_el;
1361
1362         if (!ocfs2_is_dealloc_empty(et)) {
1363                 status = ocfs2_reuse_blk_from_dealloc(handle, et,
1364                                                       &new_eb_bh, 1,
1365                                                       &block_given);
1366         } else if (meta_ac) {
1367                 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1368                                                    &new_eb_bh);
1369
1370         } else {
1371                 BUG();
1372         }
1373
1374         if (status < 0) {
1375                 mlog_errno(status);
1376                 goto bail;
1377         }
1378
1379         eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1380         /* ocfs2_create_new_meta_bhs() should create it right! */
1381         BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1382
1383         eb_el = &eb->h_list;
1384         root_el = et->et_root_el;
1385
1386         status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1387                                          OCFS2_JOURNAL_ACCESS_CREATE);
1388         if (status < 0) {
1389                 mlog_errno(status);
1390                 goto bail;
1391         }
1392
1393         /* copy the root extent list data into the new extent block */
1394         eb_el->l_tree_depth = root_el->l_tree_depth;
1395         eb_el->l_next_free_rec = root_el->l_next_free_rec;
1396         for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1397                 eb_el->l_recs[i] = root_el->l_recs[i];
1398
1399         ocfs2_journal_dirty(handle, new_eb_bh);
1400
1401         status = ocfs2_et_root_journal_access(handle, et,
1402                                               OCFS2_JOURNAL_ACCESS_WRITE);
1403         if (status < 0) {
1404                 mlog_errno(status);
1405                 goto bail;
1406         }
1407
1408         new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1409
1410         /* update root_bh now */
1411         le16_add_cpu(&root_el->l_tree_depth, 1);
1412         root_el->l_recs[0].e_cpos = 0;
1413         root_el->l_recs[0].e_blkno = eb->h_blkno;
1414         root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1415         for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1416                 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1417         root_el->l_next_free_rec = cpu_to_le16(1);
1418
1419         /* If this is our 1st tree depth shift, then last_eb_blk
1420          * becomes the allocated extent block */
1421         if (root_el->l_tree_depth == cpu_to_le16(1))
1422                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1423
1424         ocfs2_journal_dirty(handle, et->et_root_bh);
1425
1426         *ret_new_eb_bh = new_eb_bh;
1427         new_eb_bh = NULL;
1428         status = 0;
1429 bail:
1430         brelse(new_eb_bh);
1431
1432         return status;
1433 }
1434
1435 /*
1436  * Should only be called when there is no space left in any of the
1437  * leaf nodes. What we want to do is find the lowest tree depth
1438  * non-leaf extent block with room for new records. There are three
1439  * valid results of this search:
1440  *
1441  * 1) a lowest extent block is found, then we pass it back in
1442  *    *lowest_eb_bh and return '0'
1443  *
1444  * 2) the search fails to find anything, but the root_el has room. We
1445  *    pass NULL back in *lowest_eb_bh, but still return '0'
1446  *
1447  * 3) the search fails to find anything AND the root_el is full, in
1448  *    which case we return > 0
1449  *
1450  * return status < 0 indicates an error.
1451  */
1452 static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1453                                     struct buffer_head **target_bh)
1454 {
1455         int status = 0, i;
1456         u64 blkno;
1457         struct ocfs2_extent_block *eb;
1458         struct ocfs2_extent_list  *el;
1459         struct buffer_head *bh = NULL;
1460         struct buffer_head *lowest_bh = NULL;
1461
1462         *target_bh = NULL;
1463
1464         el = et->et_root_el;
1465
1466         while(le16_to_cpu(el->l_tree_depth) > 1) {
1467                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1468                         status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1469                                         "Owner %llu has empty extent list (next_free_rec == 0)\n",
1470                                         (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1471                         goto bail;
1472                 }
1473                 i = le16_to_cpu(el->l_next_free_rec) - 1;
1474                 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1475                 if (!blkno) {
1476                         status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1477                                         "Owner %llu has extent list where extent # %d has no physical block start\n",
1478                                         (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1479                         goto bail;
1480                 }
1481
1482                 brelse(bh);
1483                 bh = NULL;
1484
1485                 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1486                 if (status < 0) {
1487                         mlog_errno(status);
1488                         goto bail;
1489                 }
1490
1491                 eb = (struct ocfs2_extent_block *) bh->b_data;
1492                 el = &eb->h_list;
1493
1494                 if (le16_to_cpu(el->l_next_free_rec) <
1495                     le16_to_cpu(el->l_count)) {
1496                         brelse(lowest_bh);
1497                         lowest_bh = bh;
1498                         get_bh(lowest_bh);
1499                 }
1500         }
1501
1502         /* If we didn't find one and the fe doesn't have any room,
1503          * then return '1' */
1504         el = et->et_root_el;
1505         if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1506                 status = 1;
1507
1508         *target_bh = lowest_bh;
1509 bail:
1510         brelse(bh);
1511
1512         return status;
1513 }
1514
1515 /*
1516  * Grow a b-tree so that it has more records.
1517  *
1518  * We might shift the tree depth in which case existing paths should
1519  * be considered invalid.
1520  *
1521  * Tree depth after the grow is returned via *final_depth.
1522  *
1523  * *last_eb_bh will be updated by ocfs2_add_branch().
1524  */
1525 static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1526                            int *final_depth, struct buffer_head **last_eb_bh,
1527                            struct ocfs2_alloc_context *meta_ac)
1528 {
1529         int ret, shift;
1530         struct ocfs2_extent_list *el = et->et_root_el;
1531         int depth = le16_to_cpu(el->l_tree_depth);
1532         struct buffer_head *bh = NULL;
1533
1534         BUG_ON(meta_ac == NULL && ocfs2_is_dealloc_empty(et));
1535
1536         shift = ocfs2_find_branch_target(et, &bh);
1537         if (shift < 0) {
1538                 ret = shift;
1539                 mlog_errno(ret);
1540                 goto out;
1541         }
1542
1543         /* We traveled all the way to the bottom of the allocation tree
1544          * and didn't find room for any more extents - we need to add
1545          * another tree level */
1546         if (shift) {
1547                 BUG_ON(bh);
1548                 trace_ocfs2_grow_tree(
1549                         (unsigned long long)
1550                         ocfs2_metadata_cache_owner(et->et_ci),
1551                         depth);
1552
1553                 /* ocfs2_shift_tree_depth will return us a buffer with
1554                  * the new extent block (so we can pass that to
1555                  * ocfs2_add_branch). */
1556                 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1557                 if (ret < 0) {
1558                         mlog_errno(ret);
1559                         goto out;
1560                 }
1561                 depth++;
1562                 if (depth == 1) {
1563                         /*
1564                          * Special case: we have room now if we shifted from
1565                          * tree_depth 0, so no more work needs to be done.
1566                          *
1567                          * We won't be calling add_branch, so pass
1568                          * back *last_eb_bh as the new leaf. At depth
1569                          * zero, it should always be null so there's
1570                          * no reason to brelse.
1571                          */
1572                         BUG_ON(*last_eb_bh);
1573                         get_bh(bh);
1574                         *last_eb_bh = bh;
1575                         goto out;
1576                 }
1577         }
1578
1579         /* call ocfs2_add_branch to add the final part of the tree with
1580          * the new data. */
1581         ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1582                                meta_ac);
1583         if (ret < 0)
1584                 mlog_errno(ret);
1585
1586 out:
1587         if (final_depth)
1588                 *final_depth = depth;
1589         brelse(bh);
1590         return ret;
1591 }
1592
1593 /*
1594  * This function will discard the rightmost extent record.
1595  */
1596 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1597 {
1598         int next_free = le16_to_cpu(el->l_next_free_rec);
1599         int count = le16_to_cpu(el->l_count);
1600         unsigned int num_bytes;
1601
1602         BUG_ON(!next_free);
1603         /* This will cause us to go off the end of our extent list. */
1604         BUG_ON(next_free >= count);
1605
1606         num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1607
1608         memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1609 }
1610
1611 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1612                               struct ocfs2_extent_rec *insert_rec)
1613 {
1614         int i, insert_index, next_free, has_empty, num_bytes;
1615         u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1616         struct ocfs2_extent_rec *rec;
1617
1618         next_free = le16_to_cpu(el->l_next_free_rec);
1619         has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1620
1621         BUG_ON(!next_free);
1622
1623         /* The tree code before us didn't allow enough room in the leaf. */
1624         BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1625
1626         /*
1627          * The easiest way to approach this is to just remove the
1628          * empty extent and temporarily decrement next_free.
1629          */
1630         if (has_empty) {
1631                 /*
1632                  * If next_free was 1 (only an empty extent), this
1633                  * loop won't execute, which is fine. We still want
1634                  * the decrement above to happen.
1635                  */
1636                 for(i = 0; i < (next_free - 1); i++)
1637                         el->l_recs[i] = el->l_recs[i+1];
1638
1639                 next_free--;
1640         }
1641
1642         /*
1643          * Figure out what the new record index should be.
1644          */
1645         for(i = 0; i < next_free; i++) {
1646                 rec = &el->l_recs[i];
1647
1648                 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1649                         break;
1650         }
1651         insert_index = i;
1652
1653         trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1654                                 has_empty, next_free,
1655                                 le16_to_cpu(el->l_count));
1656
1657         BUG_ON(insert_index < 0);
1658         BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1659         BUG_ON(insert_index > next_free);
1660
1661         /*
1662          * No need to memmove if we're just adding to the tail.
1663          */
1664         if (insert_index != next_free) {
1665                 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1666
1667                 num_bytes = next_free - insert_index;
1668                 num_bytes *= sizeof(struct ocfs2_extent_rec);
1669                 memmove(&el->l_recs[insert_index + 1],
1670                         &el->l_recs[insert_index],
1671                         num_bytes);
1672         }
1673
1674         /*
1675          * Either we had an empty extent, and need to re-increment or
1676          * there was no empty extent on a non full rightmost leaf node,
1677          * in which case we still need to increment.
1678          */
1679         next_free++;
1680         el->l_next_free_rec = cpu_to_le16(next_free);
1681         /*
1682          * Make sure none of the math above just messed up our tree.
1683          */
1684         BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1685
1686         el->l_recs[insert_index] = *insert_rec;
1687
1688 }
1689
1690 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1691 {
1692         int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1693
1694         BUG_ON(num_recs == 0);
1695
1696         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1697                 num_recs--;
1698                 size = num_recs * sizeof(struct ocfs2_extent_rec);
1699                 memmove(&el->l_recs[0], &el->l_recs[1], size);
1700                 memset(&el->l_recs[num_recs], 0,
1701                        sizeof(struct ocfs2_extent_rec));
1702                 el->l_next_free_rec = cpu_to_le16(num_recs);
1703         }
1704 }
1705
1706 /*
1707  * Create an empty extent record .
1708  *
1709  * l_next_free_rec may be updated.
1710  *
1711  * If an empty extent already exists do nothing.
1712  */
1713 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1714 {
1715         int next_free = le16_to_cpu(el->l_next_free_rec);
1716
1717         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1718
1719         if (next_free == 0)
1720                 goto set_and_inc;
1721
1722         if (ocfs2_is_empty_extent(&el->l_recs[0]))
1723                 return;
1724
1725         mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1726                         "Asked to create an empty extent in a full list:\n"
1727                         "count = %u, tree depth = %u",
1728                         le16_to_cpu(el->l_count),
1729                         le16_to_cpu(el->l_tree_depth));
1730
1731         ocfs2_shift_records_right(el);
1732
1733 set_and_inc:
1734         le16_add_cpu(&el->l_next_free_rec, 1);
1735         memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1736 }
1737
1738 /*
1739  * For a rotation which involves two leaf nodes, the "root node" is
1740  * the lowest level tree node which contains a path to both leafs. This
1741  * resulting set of information can be used to form a complete "subtree"
1742  *
1743  * This function is passed two full paths from the dinode down to a
1744  * pair of adjacent leaves. It's task is to figure out which path
1745  * index contains the subtree root - this can be the root index itself
1746  * in a worst-case rotation.
1747  *
1748  * The array index of the subtree root is passed back.
1749  */
1750 int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1751                             struct ocfs2_path *left,
1752                             struct ocfs2_path *right)
1753 {
1754         int i = 0;
1755
1756         /*
1757          * Check that the caller passed in two paths from the same tree.
1758          */
1759         BUG_ON(path_root_bh(left) != path_root_bh(right));
1760
1761         do {
1762                 i++;
1763
1764                 /*
1765                  * The caller didn't pass two adjacent paths.
1766                  */
1767                 mlog_bug_on_msg(i > left->p_tree_depth,
1768                                 "Owner %llu, left depth %u, right depth %u\n"
1769                                 "left leaf blk %llu, right leaf blk %llu\n",
1770                                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1771                                 left->p_tree_depth, right->p_tree_depth,
1772                                 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1773                                 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1774         } while (left->p_node[i].bh->b_blocknr ==
1775                  right->p_node[i].bh->b_blocknr);
1776
1777         return i - 1;
1778 }
1779
1780 typedef void (path_insert_t)(void *, struct buffer_head *);
1781
1782 /*
1783  * Traverse a btree path in search of cpos, starting at root_el.
1784  *
1785  * This code can be called with a cpos larger than the tree, in which
1786  * case it will return the rightmost path.
1787  */
1788 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1789                              struct ocfs2_extent_list *root_el, u32 cpos,
1790                              path_insert_t *func, void *data)
1791 {
1792         int i, ret = 0;
1793         u32 range;
1794         u64 blkno;
1795         struct buffer_head *bh = NULL;
1796         struct ocfs2_extent_block *eb;
1797         struct ocfs2_extent_list *el;
1798         struct ocfs2_extent_rec *rec;
1799
1800         el = root_el;
1801         while (el->l_tree_depth) {
1802                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1803                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1804                                     "Owner %llu has empty extent list at depth %u\n",
1805                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1806                                     le16_to_cpu(el->l_tree_depth));
1807                         ret = -EROFS;
1808                         goto out;
1809
1810                 }
1811
1812                 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1813                         rec = &el->l_recs[i];
1814
1815                         /*
1816                          * In the case that cpos is off the allocation
1817                          * tree, this should just wind up returning the
1818                          * rightmost record.
1819                          */
1820                         range = le32_to_cpu(rec->e_cpos) +
1821                                 ocfs2_rec_clusters(el, rec);
1822                         if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1823                             break;
1824                 }
1825
1826                 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1827                 if (blkno == 0) {
1828                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1829                                     "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1830                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1831                                     le16_to_cpu(el->l_tree_depth), i);
1832                         ret = -EROFS;
1833                         goto out;
1834                 }
1835
1836                 brelse(bh);
1837                 bh = NULL;
1838                 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1839                 if (ret) {
1840                         mlog_errno(ret);
1841                         goto out;
1842                 }
1843
1844                 eb = (struct ocfs2_extent_block *) bh->b_data;
1845                 el = &eb->h_list;
1846
1847                 if (le16_to_cpu(el->l_next_free_rec) >
1848                     le16_to_cpu(el->l_count)) {
1849                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1850                                     "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1851                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1852                                     (unsigned long long)bh->b_blocknr,
1853                                     le16_to_cpu(el->l_next_free_rec),
1854                                     le16_to_cpu(el->l_count));
1855                         ret = -EROFS;
1856                         goto out;
1857                 }
1858
1859                 if (func)
1860                         func(data, bh);
1861         }
1862
1863 out:
1864         /*
1865          * Catch any trailing bh that the loop didn't handle.
1866          */
1867         brelse(bh);
1868
1869         return ret;
1870 }
1871
1872 /*
1873  * Given an initialized path (that is, it has a valid root extent
1874  * list), this function will traverse the btree in search of the path
1875  * which would contain cpos.
1876  *
1877  * The path traveled is recorded in the path structure.
1878  *
1879  * Note that this will not do any comparisons on leaf node extent
1880  * records, so it will work fine in the case that we just added a tree
1881  * branch.
1882  */
1883 struct find_path_data {
1884         int index;
1885         struct ocfs2_path *path;
1886 };
1887 static void find_path_ins(void *data, struct buffer_head *bh)
1888 {
1889         struct find_path_data *fp = data;
1890
1891         get_bh(bh);
1892         ocfs2_path_insert_eb(fp->path, fp->index, bh);
1893         fp->index++;
1894 }
1895 int ocfs2_find_path(struct ocfs2_caching_info *ci,
1896                     struct ocfs2_path *path, u32 cpos)
1897 {
1898         struct find_path_data data;
1899
1900         data.index = 1;
1901         data.path = path;
1902         return __ocfs2_find_path(ci, path_root_el(path), cpos,
1903                                  find_path_ins, &data);
1904 }
1905
1906 static void find_leaf_ins(void *data, struct buffer_head *bh)
1907 {
1908         struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1909         struct ocfs2_extent_list *el = &eb->h_list;
1910         struct buffer_head **ret = data;
1911
1912         /* We want to retain only the leaf block. */
1913         if (le16_to_cpu(el->l_tree_depth) == 0) {
1914                 get_bh(bh);
1915                 *ret = bh;
1916         }
1917 }
1918 /*
1919  * Find the leaf block in the tree which would contain cpos. No
1920  * checking of the actual leaf is done.
1921  *
1922  * Some paths want to call this instead of allocating a path structure
1923  * and calling ocfs2_find_path().
1924  *
1925  * This function doesn't handle non btree extent lists.
1926  */
1927 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1928                     struct ocfs2_extent_list *root_el, u32 cpos,
1929                     struct buffer_head **leaf_bh)
1930 {
1931         int ret;
1932         struct buffer_head *bh = NULL;
1933
1934         ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1935         if (ret) {
1936                 mlog_errno(ret);
1937                 goto out;
1938         }
1939
1940         *leaf_bh = bh;
1941 out:
1942         return ret;
1943 }
1944
1945 /*
1946  * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1947  *
1948  * Basically, we've moved stuff around at the bottom of the tree and
1949  * we need to fix up the extent records above the changes to reflect
1950  * the new changes.
1951  *
1952  * left_rec: the record on the left.
1953  * right_rec: the record to the right of left_rec
1954  * right_child_el: is the child list pointed to by right_rec
1955  *
1956  * By definition, this only works on interior nodes.
1957  */
1958 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1959                                   struct ocfs2_extent_rec *right_rec,
1960                                   struct ocfs2_extent_list *right_child_el)
1961 {
1962         u32 left_clusters, right_end;
1963
1964         /*
1965          * Interior nodes never have holes. Their cpos is the cpos of
1966          * the leftmost record in their child list. Their cluster
1967          * count covers the full theoretical range of their child list
1968          * - the range between their cpos and the cpos of the record
1969          * immediately to their right.
1970          */
1971         left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1972         if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1973                 BUG_ON(right_child_el->l_tree_depth);
1974                 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1975                 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1976         }
1977         left_clusters -= le32_to_cpu(left_rec->e_cpos);
1978         left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1979
1980         /*
1981          * Calculate the rightmost cluster count boundary before
1982          * moving cpos - we will need to adjust clusters after
1983          * updating e_cpos to keep the same highest cluster count.
1984          */
1985         right_end = le32_to_cpu(right_rec->e_cpos);
1986         right_end += le32_to_cpu(right_rec->e_int_clusters);
1987
1988         right_rec->e_cpos = left_rec->e_cpos;
1989         le32_add_cpu(&right_rec->e_cpos, left_clusters);
1990
1991         right_end -= le32_to_cpu(right_rec->e_cpos);
1992         right_rec->e_int_clusters = cpu_to_le32(right_end);
1993 }
1994
1995 /*
1996  * Adjust the adjacent root node records involved in a
1997  * rotation. left_el_blkno is passed in as a key so that we can easily
1998  * find it's index in the root list.
1999  */
2000 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
2001                                       struct ocfs2_extent_list *left_el,
2002                                       struct ocfs2_extent_list *right_el,
2003                                       u64 left_el_blkno)
2004 {
2005         int i;
2006
2007         BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2008                le16_to_cpu(left_el->l_tree_depth));
2009
2010         for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2011                 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2012                         break;
2013         }
2014
2015         /*
2016          * The path walking code should have never returned a root and
2017          * two paths which are not adjacent.
2018          */
2019         BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2020
2021         ocfs2_adjust_adjacent_records(&root_el->l_recs[i],
2022                                       &root_el->l_recs[i + 1], right_el);
2023 }
2024
2025 /*
2026  * We've changed a leaf block (in right_path) and need to reflect that
2027  * change back up the subtree.
2028  *
2029  * This happens in multiple places:
2030  *   - When we've moved an extent record from the left path leaf to the right
2031  *     path leaf to make room for an empty extent in the left path leaf.
2032  *   - When our insert into the right path leaf is at the leftmost edge
2033  *     and requires an update of the path immediately to it's left. This
2034  *     can occur at the end of some types of rotation and appending inserts.
2035  *   - When we've adjusted the last extent record in the left path leaf and the
2036  *     1st extent record in the right path leaf during cross extent block merge.
2037  */
2038 static void ocfs2_complete_edge_insert(handle_t *handle,
2039                                        struct ocfs2_path *left_path,
2040                                        struct ocfs2_path *right_path,
2041                                        int subtree_index)
2042 {
2043         int i, idx;
2044         struct ocfs2_extent_list *el, *left_el, *right_el;
2045         struct ocfs2_extent_rec *left_rec, *right_rec;
2046         struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2047
2048         /*
2049          * Update the counts and position values within all the
2050          * interior nodes to reflect the leaf rotation we just did.
2051          *
2052          * The root node is handled below the loop.
2053          *
2054          * We begin the loop with right_el and left_el pointing to the
2055          * leaf lists and work our way up.
2056          *
2057          * NOTE: within this loop, left_el and right_el always refer
2058          * to the *child* lists.
2059          */
2060         left_el = path_leaf_el(left_path);
2061         right_el = path_leaf_el(right_path);
2062         for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2063                 trace_ocfs2_complete_edge_insert(i);
2064
2065                 /*
2066                  * One nice property of knowing that all of these
2067                  * nodes are below the root is that we only deal with
2068                  * the leftmost right node record and the rightmost
2069                  * left node record.
2070                  */
2071                 el = left_path->p_node[i].el;
2072                 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2073                 left_rec = &el->l_recs[idx];
2074
2075                 el = right_path->p_node[i].el;
2076                 right_rec = &el->l_recs[0];
2077
2078                 ocfs2_adjust_adjacent_records(left_rec, right_rec, right_el);
2079
2080                 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2081                 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2082
2083                 /*
2084                  * Setup our list pointers now so that the current
2085                  * parents become children in the next iteration.
2086                  */
2087                 left_el = left_path->p_node[i].el;
2088                 right_el = right_path->p_node[i].el;
2089         }
2090
2091         /*
2092          * At the root node, adjust the two adjacent records which
2093          * begin our path to the leaves.
2094          */
2095
2096         el = left_path->p_node[subtree_index].el;
2097         left_el = left_path->p_node[subtree_index + 1].el;
2098         right_el = right_path->p_node[subtree_index + 1].el;
2099
2100         ocfs2_adjust_root_records(el, left_el, right_el,
2101                                   left_path->p_node[subtree_index + 1].bh->b_blocknr);
2102
2103         root_bh = left_path->p_node[subtree_index].bh;
2104
2105         ocfs2_journal_dirty(handle, root_bh);
2106 }
2107
2108 static int ocfs2_rotate_subtree_right(handle_t *handle,
2109                                       struct ocfs2_extent_tree *et,
2110                                       struct ocfs2_path *left_path,
2111                                       struct ocfs2_path *right_path,
2112                                       int subtree_index)
2113 {
2114         int ret, i;
2115         struct buffer_head *right_leaf_bh;
2116         struct buffer_head *left_leaf_bh = NULL;
2117         struct buffer_head *root_bh;
2118         struct ocfs2_extent_list *right_el, *left_el;
2119         struct ocfs2_extent_rec move_rec;
2120
2121         left_leaf_bh = path_leaf_bh(left_path);
2122         left_el = path_leaf_el(left_path);
2123
2124         if (left_el->l_next_free_rec != left_el->l_count) {
2125                 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2126                             "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2127                             (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2128                             (unsigned long long)left_leaf_bh->b_blocknr,
2129                             le16_to_cpu(left_el->l_next_free_rec));
2130                 return -EROFS;
2131         }
2132
2133         /*
2134          * This extent block may already have an empty record, so we
2135          * return early if so.
2136          */
2137         if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2138                 return 0;
2139
2140         root_bh = left_path->p_node[subtree_index].bh;
2141         BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2142
2143         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2144                                            subtree_index);
2145         if (ret) {
2146                 mlog_errno(ret);
2147                 goto out;
2148         }
2149
2150         for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2151                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2152                                                    right_path, i);
2153                 if (ret) {
2154                         mlog_errno(ret);
2155                         goto out;
2156                 }
2157
2158                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2159                                                    left_path, i);
2160                 if (ret) {
2161                         mlog_errno(ret);
2162                         goto out;
2163                 }
2164         }
2165
2166         right_leaf_bh = path_leaf_bh(right_path);
2167         right_el = path_leaf_el(right_path);
2168
2169         /* This is a code error, not a disk corruption. */
2170         mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2171                         "because rightmost leaf block %llu is empty\n",
2172                         (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2173                         (unsigned long long)right_leaf_bh->b_blocknr);
2174
2175         ocfs2_create_empty_extent(right_el);
2176
2177         ocfs2_journal_dirty(handle, right_leaf_bh);
2178
2179         /* Do the copy now. */
2180         i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2181         move_rec = left_el->l_recs[i];
2182         right_el->l_recs[0] = move_rec;
2183
2184         /*
2185          * Clear out the record we just copied and shift everything
2186          * over, leaving an empty extent in the left leaf.
2187          *
2188          * We temporarily subtract from next_free_rec so that the
2189          * shift will lose the tail record (which is now defunct).
2190          */
2191         le16_add_cpu(&left_el->l_next_free_rec, -1);
2192         ocfs2_shift_records_right(left_el);
2193         memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2194         le16_add_cpu(&left_el->l_next_free_rec, 1);
2195
2196         ocfs2_journal_dirty(handle, left_leaf_bh);
2197
2198         ocfs2_complete_edge_insert(handle, left_path, right_path,
2199                                    subtree_index);
2200
2201 out:
2202         return ret;
2203 }
2204
2205 /*
2206  * Given a full path, determine what cpos value would return us a path
2207  * containing the leaf immediately to the left of the current one.
2208  *
2209  * Will return zero if the path passed in is already the leftmost path.
2210  */
2211 int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2212                                   struct ocfs2_path *path, u32 *cpos)
2213 {
2214         int i, j, ret = 0;
2215         u64 blkno;
2216         struct ocfs2_extent_list *el;
2217
2218         BUG_ON(path->p_tree_depth == 0);
2219
2220         *cpos = 0;
2221
2222         blkno = path_leaf_bh(path)->b_blocknr;
2223
2224         /* Start at the tree node just above the leaf and work our way up. */
2225         i = path->p_tree_depth - 1;
2226         while (i >= 0) {
2227                 el = path->p_node[i].el;
2228
2229                 /*
2230                  * Find the extent record just before the one in our
2231                  * path.
2232                  */
2233                 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2234                         if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2235                                 if (j == 0) {
2236                                         if (i == 0) {
2237                                                 /*
2238                                                  * We've determined that the
2239                                                  * path specified is already
2240                                                  * the leftmost one - return a
2241                                                  * cpos of zero.
2242                                                  */
2243                                                 goto out;
2244                                         }
2245                                         /*
2246                                          * The leftmost record points to our
2247                                          * leaf - we need to travel up the
2248                                          * tree one level.
2249                                          */
2250                                         goto next_node;
2251                                 }
2252
2253                                 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2254                                 *cpos = *cpos + ocfs2_rec_clusters(el,
2255                                                            &el->l_recs[j - 1]);
2256                                 *cpos = *cpos - 1;
2257                                 goto out;
2258                         }
2259                 }
2260
2261                 /*
2262                  * If we got here, we never found a valid node where
2263                  * the tree indicated one should be.
2264                  */
2265                 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2266                             (unsigned long long)blkno);
2267                 ret = -EROFS;
2268                 goto out;
2269
2270 next_node:
2271                 blkno = path->p_node[i].bh->b_blocknr;
2272                 i--;
2273         }
2274
2275 out:
2276         return ret;
2277 }
2278
2279 /*
2280  * Extend the transaction by enough credits to complete the rotation,
2281  * and still leave at least the original number of credits allocated
2282  * to this transaction.
2283  */
2284 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2285                                            int op_credits,
2286                                            struct ocfs2_path *path)
2287 {
2288         int ret = 0;
2289         int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2290
2291         if (handle->h_buffer_credits < credits)
2292                 ret = ocfs2_extend_trans(handle,
2293                                          credits - handle->h_buffer_credits);
2294
2295         return ret;
2296 }
2297
2298 /*
2299  * Trap the case where we're inserting into the theoretical range past
2300  * the _actual_ left leaf range. Otherwise, we'll rotate a record
2301  * whose cpos is less than ours into the right leaf.
2302  *
2303  * It's only necessary to look at the rightmost record of the left
2304  * leaf because the logic that calls us should ensure that the
2305  * theoretical ranges in the path components above the leaves are
2306  * correct.
2307  */
2308 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2309                                                  u32 insert_cpos)
2310 {
2311         struct ocfs2_extent_list *left_el;
2312         struct ocfs2_extent_rec *rec;
2313         int next_free;
2314
2315         left_el = path_leaf_el(left_path);
2316         next_free = le16_to_cpu(left_el->l_next_free_rec);
2317         rec = &left_el->l_recs[next_free - 1];
2318
2319         if (insert_cpos > le32_to_cpu(rec->e_cpos))
2320                 return 1;
2321         return 0;
2322 }
2323
2324 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2325 {
2326         int next_free = le16_to_cpu(el->l_next_free_rec);
2327         unsigned int range;
2328         struct ocfs2_extent_rec *rec;
2329
2330         if (next_free == 0)
2331                 return 0;
2332
2333         rec = &el->l_recs[0];
2334         if (ocfs2_is_empty_extent(rec)) {
2335                 /* Empty list. */
2336                 if (next_free == 1)
2337                         return 0;
2338                 rec = &el->l_recs[1];
2339         }
2340
2341         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2342         if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2343                 return 1;
2344         return 0;
2345 }
2346
2347 /*
2348  * Rotate all the records in a btree right one record, starting at insert_cpos.
2349  *
2350  * The path to the rightmost leaf should be passed in.
2351  *
2352  * The array is assumed to be large enough to hold an entire path (tree depth).
2353  *
2354  * Upon successful return from this function:
2355  *
2356  * - The 'right_path' array will contain a path to the leaf block
2357  *   whose range contains e_cpos.
2358  * - That leaf block will have a single empty extent in list index 0.
2359  * - In the case that the rotation requires a post-insert update,
2360  *   *ret_left_path will contain a valid path which can be passed to
2361  *   ocfs2_insert_path().
2362  */
2363 static int ocfs2_rotate_tree_right(handle_t *handle,
2364                                    struct ocfs2_extent_tree *et,
2365                                    enum ocfs2_split_type split,
2366                                    u32 insert_cpos,
2367                                    struct ocfs2_path *right_path,
2368                                    struct ocfs2_path **ret_left_path)
2369 {
2370         int ret, start, orig_credits = handle->h_buffer_credits;
2371         u32 cpos;
2372         struct ocfs2_path *left_path = NULL;
2373         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2374
2375         *ret_left_path = NULL;
2376
2377         left_path = ocfs2_new_path_from_path(right_path);
2378         if (!left_path) {
2379                 ret = -ENOMEM;
2380                 mlog_errno(ret);
2381                 goto out;
2382         }
2383
2384         ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2385         if (ret) {
2386                 mlog_errno(ret);
2387                 goto out;
2388         }
2389
2390         trace_ocfs2_rotate_tree_right(
2391                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2392                 insert_cpos, cpos);
2393
2394         /*
2395          * What we want to do here is:
2396          *
2397          * 1) Start with the rightmost path.
2398          *
2399          * 2) Determine a path to the leaf block directly to the left
2400          *    of that leaf.
2401          *
2402          * 3) Determine the 'subtree root' - the lowest level tree node
2403          *    which contains a path to both leaves.
2404          *
2405          * 4) Rotate the subtree.
2406          *
2407          * 5) Find the next subtree by considering the left path to be
2408          *    the new right path.
2409          *
2410          * The check at the top of this while loop also accepts
2411          * insert_cpos == cpos because cpos is only a _theoretical_
2412          * value to get us the left path - insert_cpos might very well
2413          * be filling that hole.
2414          *
2415          * Stop at a cpos of '0' because we either started at the
2416          * leftmost branch (i.e., a tree with one branch and a
2417          * rotation inside of it), or we've gone as far as we can in
2418          * rotating subtrees.
2419          */
2420         while (cpos && insert_cpos <= cpos) {
2421                 trace_ocfs2_rotate_tree_right(
2422                         (unsigned long long)
2423                         ocfs2_metadata_cache_owner(et->et_ci),
2424                         insert_cpos, cpos);
2425
2426                 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2427                 if (ret) {
2428                         mlog_errno(ret);
2429                         goto out;
2430                 }
2431
2432                 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2433                                 path_leaf_bh(right_path),
2434                                 "Owner %llu: error during insert of %u "
2435                                 "(left path cpos %u) results in two identical "
2436                                 "paths ending at %llu\n",
2437                                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2438                                 insert_cpos, cpos,
2439                                 (unsigned long long)
2440                                 path_leaf_bh(left_path)->b_blocknr);
2441
2442                 if (split == SPLIT_NONE &&
2443                     ocfs2_rotate_requires_path_adjustment(left_path,
2444                                                           insert_cpos)) {
2445
2446                         /*
2447                          * We've rotated the tree as much as we
2448                          * should. The rest is up to
2449                          * ocfs2_insert_path() to complete, after the
2450                          * record insertion. We indicate this
2451                          * situation by returning the left path.
2452                          *
2453                          * The reason we don't adjust the records here
2454                          * before the record insert is that an error
2455                          * later might break the rule where a parent
2456                          * record e_cpos will reflect the actual
2457                          * e_cpos of the 1st nonempty record of the
2458                          * child list.
2459                          */
2460                         *ret_left_path = left_path;
2461                         goto out_ret_path;
2462                 }
2463
2464                 start = ocfs2_find_subtree_root(et, left_path, right_path);
2465
2466                 trace_ocfs2_rotate_subtree(start,
2467                         (unsigned long long)
2468                         right_path->p_node[start].bh->b_blocknr,
2469                         right_path->p_tree_depth);
2470
2471                 ret = ocfs2_extend_rotate_transaction(handle, start,
2472                                                       orig_credits, right_path);
2473                 if (ret) {
2474                         mlog_errno(ret);
2475                         goto out;
2476                 }
2477
2478                 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2479                                                  right_path, start);
2480                 if (ret) {
2481                         mlog_errno(ret);
2482                         goto out;
2483                 }
2484
2485                 if (split != SPLIT_NONE &&
2486                     ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2487                                                 insert_cpos)) {
2488                         /*
2489                          * A rotate moves the rightmost left leaf
2490                          * record over to the leftmost right leaf
2491                          * slot. If we're doing an extent split
2492                          * instead of a real insert, then we have to
2493                          * check that the extent to be split wasn't
2494                          * just moved over. If it was, then we can
2495                          * exit here, passing left_path back -
2496                          * ocfs2_split_extent() is smart enough to
2497                          * search both leaves.
2498                          */
2499                         *ret_left_path = left_path;
2500                         goto out_ret_path;
2501                 }
2502
2503                 /*
2504                  * There is no need to re-read the next right path
2505                  * as we know that it'll be our current left
2506                  * path. Optimize by copying values instead.
2507                  */
2508                 ocfs2_mv_path(right_path, left_path);
2509
2510                 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2511                 if (ret) {
2512                         mlog_errno(ret);
2513                         goto out;
2514                 }
2515         }
2516
2517 out:
2518         ocfs2_free_path(left_path);
2519
2520 out_ret_path:
2521         return ret;
2522 }
2523
2524 static int ocfs2_update_edge_lengths(handle_t *handle,
2525                                      struct ocfs2_extent_tree *et,
2526                                      struct ocfs2_path *path)
2527 {
2528         int i, idx, ret;
2529         struct ocfs2_extent_rec *rec;
2530         struct ocfs2_extent_list *el;
2531         struct ocfs2_extent_block *eb;
2532         u32 range;
2533
2534         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2535         if (ret) {
2536                 mlog_errno(ret);
2537                 goto out;
2538         }
2539
2540         /* Path should always be rightmost. */
2541         eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2542         BUG_ON(eb->h_next_leaf_blk != 0ULL);
2543
2544         el = &eb->h_list;
2545         BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2546         idx = le16_to_cpu(el->l_next_free_rec) - 1;
2547         rec = &el->l_recs[idx];
2548         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2549
2550         for (i = 0; i < path->p_tree_depth; i++) {
2551                 el = path->p_node[i].el;
2552                 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2553                 rec = &el->l_recs[idx];
2554
2555                 rec->e_int_clusters = cpu_to_le32(range);
2556                 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2557
2558                 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2559         }
2560 out:
2561         return ret;
2562 }
2563
2564 static void ocfs2_unlink_path(handle_t *handle,
2565                               struct ocfs2_extent_tree *et,
2566                               struct ocfs2_cached_dealloc_ctxt *dealloc,
2567                               struct ocfs2_path *path, int unlink_start)
2568 {
2569         int ret, i;
2570         struct ocfs2_extent_block *eb;
2571         struct ocfs2_extent_list *el;
2572         struct buffer_head *bh;
2573
2574         for(i = unlink_start; i < path_num_items(path); i++) {
2575                 bh = path->p_node[i].bh;
2576
2577                 eb = (struct ocfs2_extent_block *)bh->b_data;
2578                 /*
2579                  * Not all nodes might have had their final count
2580                  * decremented by the caller - handle this here.
2581                  */
2582                 el = &eb->h_list;
2583                 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2584                         mlog(ML_ERROR,
2585                              "Inode %llu, attempted to remove extent block "
2586                              "%llu with %u records\n",
2587                              (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2588                              (unsigned long long)le64_to_cpu(eb->h_blkno),
2589                              le16_to_cpu(el->l_next_free_rec));
2590
2591                         ocfs2_journal_dirty(handle, bh);
2592                         ocfs2_remove_from_cache(et->et_ci, bh);
2593                         continue;
2594                 }
2595
2596                 el->l_next_free_rec = 0;
2597                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2598
2599                 ocfs2_journal_dirty(handle, bh);
2600
2601                 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2602                 if (ret)
2603                         mlog_errno(ret);
2604
2605                 ocfs2_remove_from_cache(et->et_ci, bh);
2606         }
2607 }
2608
2609 static void ocfs2_unlink_subtree(handle_t *handle,
2610                                  struct ocfs2_extent_tree *et,
2611                                  struct ocfs2_path *left_path,
2612                                  struct ocfs2_path *right_path,
2613                                  int subtree_index,
2614                                  struct ocfs2_cached_dealloc_ctxt *dealloc)
2615 {
2616         int i;
2617         struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2618         struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2619         struct ocfs2_extent_block *eb;
2620
2621         eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2622
2623         for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2624                 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2625                         break;
2626
2627         BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2628
2629         memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2630         le16_add_cpu(&root_el->l_next_free_rec, -1);
2631
2632         eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2633         eb->h_next_leaf_blk = 0;
2634
2635         ocfs2_journal_dirty(handle, root_bh);
2636         ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2637
2638         ocfs2_unlink_path(handle, et, dealloc, right_path,
2639                           subtree_index + 1);
2640 }
2641
2642 static int ocfs2_rotate_subtree_left(handle_t *handle,
2643                                      struct ocfs2_extent_tree *et,
2644                                      struct ocfs2_path *left_path,
2645                                      struct ocfs2_path *right_path,
2646                                      int subtree_index,
2647                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2648                                      int *deleted)
2649 {
2650         int ret, i, del_right_subtree = 0, right_has_empty = 0;
2651         struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2652         struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2653         struct ocfs2_extent_block *eb;
2654
2655         *deleted = 0;
2656
2657         right_leaf_el = path_leaf_el(right_path);
2658         left_leaf_el = path_leaf_el(left_path);
2659         root_bh = left_path->p_node[subtree_index].bh;
2660         BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2661
2662         if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2663                 return 0;
2664
2665         eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2666         if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2667                 /*
2668                  * It's legal for us to proceed if the right leaf is
2669                  * the rightmost one and it has an empty extent. There
2670                  * are two cases to handle - whether the leaf will be
2671                  * empty after removal or not. If the leaf isn't empty
2672                  * then just remove the empty extent up front. The
2673                  * next block will handle empty leaves by flagging
2674                  * them for unlink.
2675                  *
2676                  * Non rightmost leaves will throw -EAGAIN and the
2677                  * caller can manually move the subtree and retry.
2678                  */
2679
2680                 if (eb->h_next_leaf_blk != 0ULL)
2681                         return -EAGAIN;
2682
2683                 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2684                         ret = ocfs2_journal_access_eb(handle, et->et_ci,
2685                                                       path_leaf_bh(right_path),
2686                                                       OCFS2_JOURNAL_ACCESS_WRITE);
2687                         if (ret) {
2688                                 mlog_errno(ret);
2689                                 goto out;
2690                         }
2691
2692                         ocfs2_remove_empty_extent(right_leaf_el);
2693                 } else
2694                         right_has_empty = 1;
2695         }
2696
2697         if (eb->h_next_leaf_blk == 0ULL &&
2698             le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2699                 /*
2700                  * We have to update i_last_eb_blk during the meta
2701                  * data delete.
2702                  */
2703                 ret = ocfs2_et_root_journal_access(handle, et,
2704                                                    OCFS2_JOURNAL_ACCESS_WRITE);
2705                 if (ret) {
2706                         mlog_errno(ret);
2707                         goto out;
2708                 }
2709
2710                 del_right_subtree = 1;
2711         }
2712
2713         /*
2714          * Getting here with an empty extent in the right path implies
2715          * that it's the rightmost path and will be deleted.
2716          */
2717         BUG_ON(right_has_empty && !del_right_subtree);
2718
2719         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2720                                            subtree_index);
2721         if (ret) {
2722                 mlog_errno(ret);
2723                 goto out;
2724         }
2725
2726         for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2727                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2728                                                    right_path, i);
2729                 if (ret) {
2730                         mlog_errno(ret);
2731                         goto out;
2732                 }
2733
2734                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2735                                                    left_path, i);
2736                 if (ret) {
2737                         mlog_errno(ret);
2738                         goto out;
2739                 }
2740         }
2741
2742         if (!right_has_empty) {
2743                 /*
2744                  * Only do this if we're moving a real
2745                  * record. Otherwise, the action is delayed until
2746                  * after removal of the right path in which case we
2747                  * can do a simple shift to remove the empty extent.
2748                  */
2749                 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2750                 memset(&right_leaf_el->l_recs[0], 0,
2751                        sizeof(struct ocfs2_extent_rec));
2752         }
2753         if (eb->h_next_leaf_blk == 0ULL) {
2754                 /*
2755                  * Move recs over to get rid of empty extent, decrease
2756                  * next_free. This is allowed to remove the last
2757                  * extent in our leaf (setting l_next_free_rec to
2758                  * zero) - the delete code below won't care.
2759                  */
2760                 ocfs2_remove_empty_extent(right_leaf_el);
2761         }
2762
2763         ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2764         ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2765
2766         if (del_right_subtree) {
2767                 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2768                                      subtree_index, dealloc);
2769                 ret = ocfs2_update_edge_lengths(handle, et, left_path);
2770                 if (ret) {
2771                         mlog_errno(ret);
2772                         goto out;
2773                 }
2774
2775                 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2776                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2777
2778                 /*
2779                  * Removal of the extent in the left leaf was skipped
2780                  * above so we could delete the right path
2781                  * 1st.
2782                  */
2783                 if (right_has_empty)
2784                         ocfs2_remove_empty_extent(left_leaf_el);
2785
2786                 ocfs2_journal_dirty(handle, et_root_bh);
2787
2788                 *deleted = 1;
2789         } else
2790                 ocfs2_complete_edge_insert(handle, left_path, right_path,
2791                                            subtree_index);
2792
2793 out:
2794         return ret;
2795 }
2796
2797 /*
2798  * Given a full path, determine what cpos value would return us a path
2799  * containing the leaf immediately to the right of the current one.
2800  *
2801  * Will return zero if the path passed in is already the rightmost path.
2802  *
2803  * This looks similar, but is subtly different to
2804  * ocfs2_find_cpos_for_left_leaf().
2805  */
2806 int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2807                                    struct ocfs2_path *path, u32 *cpos)
2808 {
2809         int i, j, ret = 0;
2810         u64 blkno;
2811         struct ocfs2_extent_list *el;
2812
2813         *cpos = 0;
2814
2815         if (path->p_tree_depth == 0)
2816                 return 0;
2817
2818         blkno = path_leaf_bh(path)->b_blocknr;
2819
2820         /* Start at the tree node just above the leaf and work our way up. */
2821         i = path->p_tree_depth - 1;
2822         while (i >= 0) {
2823                 int next_free;
2824
2825                 el = path->p_node[i].el;
2826
2827                 /*
2828                  * Find the extent record just after the one in our
2829                  * path.
2830                  */
2831                 next_free = le16_to_cpu(el->l_next_free_rec);
2832                 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2833                         if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2834                                 if (j == (next_free - 1)) {
2835                                         if (i == 0) {
2836                                                 /*
2837                                                  * We've determined that the
2838                                                  * path specified is already
2839                                                  * the rightmost one - return a
2840                                                  * cpos of zero.
2841                                                  */
2842                                                 goto out;
2843                                         }
2844                                         /*
2845                                          * The rightmost record points to our
2846                                          * leaf - we need to travel up the
2847                                          * tree one level.
2848                                          */
2849                                         goto next_node;
2850                                 }
2851
2852                                 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2853                                 goto out;
2854                         }
2855                 }
2856
2857                 /*
2858                  * If we got here, we never found a valid node where
2859                  * the tree indicated one should be.
2860                  */
2861                 ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2862                             (unsigned long long)blkno);
2863                 ret = -EROFS;
2864                 goto out;
2865
2866 next_node:
2867                 blkno = path->p_node[i].bh->b_blocknr;
2868                 i--;
2869         }
2870
2871 out:
2872         return ret;
2873 }
2874
2875 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2876                                             struct ocfs2_extent_tree *et,
2877                                             struct ocfs2_path *path)
2878 {
2879         int ret;
2880         struct buffer_head *bh = path_leaf_bh(path);
2881         struct ocfs2_extent_list *el = path_leaf_el(path);
2882
2883         if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2884                 return 0;
2885
2886         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2887                                            path_num_items(path) - 1);
2888         if (ret) {
2889                 mlog_errno(ret);
2890                 goto out;
2891         }
2892
2893         ocfs2_remove_empty_extent(el);
2894         ocfs2_journal_dirty(handle, bh);
2895
2896 out:
2897         return ret;
2898 }
2899
2900 static int __ocfs2_rotate_tree_left(handle_t *handle,
2901                                     struct ocfs2_extent_tree *et,
2902                                     int orig_credits,
2903                                     struct ocfs2_path *path,
2904                                     struct ocfs2_cached_dealloc_ctxt *dealloc,
2905                                     struct ocfs2_path **empty_extent_path)
2906 {
2907         int ret, subtree_root, deleted;
2908         u32 right_cpos;
2909         struct ocfs2_path *left_path = NULL;
2910         struct ocfs2_path *right_path = NULL;
2911         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2912
2913         if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2914                 return 0;
2915
2916         *empty_extent_path = NULL;
2917
2918         ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2919         if (ret) {
2920                 mlog_errno(ret);
2921                 goto out;
2922         }
2923
2924         left_path = ocfs2_new_path_from_path(path);
2925         if (!left_path) {
2926                 ret = -ENOMEM;
2927                 mlog_errno(ret);
2928                 goto out;
2929         }
2930
2931         ocfs2_cp_path(left_path, path);
2932
2933         right_path = ocfs2_new_path_from_path(path);
2934         if (!right_path) {
2935                 ret = -ENOMEM;
2936                 mlog_errno(ret);
2937                 goto out;
2938         }
2939
2940         while (right_cpos) {
2941                 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2942                 if (ret) {
2943                         mlog_errno(ret);
2944                         goto out;
2945                 }
2946
2947                 subtree_root = ocfs2_find_subtree_root(et, left_path,
2948                                                        right_path);
2949
2950                 trace_ocfs2_rotate_subtree(subtree_root,
2951                      (unsigned long long)
2952                      right_path->p_node[subtree_root].bh->b_blocknr,
2953                      right_path->p_tree_depth);
2954
2955                 ret = ocfs2_extend_rotate_transaction(handle, 0,
2956                                                       orig_credits, left_path);
2957                 if (ret) {
2958                         mlog_errno(ret);
2959                         goto out;
2960                 }
2961
2962                 /*
2963                  * Caller might still want to make changes to the
2964                  * tree root, so re-add it to the journal here.
2965                  */
2966                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2967                                                    left_path, 0);
2968                 if (ret) {
2969                         mlog_errno(ret);
2970                         goto out;
2971                 }
2972
2973                 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2974                                                 right_path, subtree_root,
2975                                                 dealloc, &deleted);
2976                 if (ret == -EAGAIN) {
2977                         /*
2978                          * The rotation has to temporarily stop due to
2979                          * the right subtree having an empty
2980                          * extent. Pass it back to the caller for a
2981                          * fixup.
2982                          */
2983                         *empty_extent_path = right_path;
2984                         right_path = NULL;
2985                         goto out;
2986                 }
2987                 if (ret) {
2988                         mlog_errno(ret);
2989                         goto out;
2990                 }
2991
2992                 /*
2993                  * The subtree rotate might have removed records on
2994                  * the rightmost edge. If so, then rotation is
2995                  * complete.
2996                  */
2997                 if (deleted)
2998                         break;
2999
3000                 ocfs2_mv_path(left_path, right_path);
3001
3002                 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3003                                                      &right_cpos);
3004                 if (ret) {
3005                         mlog_errno(ret);
3006                         goto out;
3007                 }
3008         }
3009
3010 out:
3011         ocfs2_free_path(right_path);
3012         ocfs2_free_path(left_path);
3013
3014         return ret;
3015 }
3016
3017 static int ocfs2_remove_rightmost_path(handle_t *handle,
3018                                 struct ocfs2_extent_tree *et,
3019                                 struct ocfs2_path *path,
3020                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
3021 {
3022         int ret, subtree_index;
3023         u32 cpos;
3024         struct ocfs2_path *left_path = NULL;
3025         struct ocfs2_extent_block *eb;
3026         struct ocfs2_extent_list *el;
3027
3028         ret = ocfs2_et_sanity_check(et);
3029         if (ret)
3030                 goto out;
3031
3032         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3033         if (ret) {
3034                 mlog_errno(ret);
3035                 goto out;
3036         }
3037
3038         ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3039                                             path, &cpos);
3040         if (ret) {
3041                 mlog_errno(ret);
3042                 goto out;
3043         }
3044
3045         if (cpos) {
3046                 /*
3047                  * We have a path to the left of this one - it needs
3048                  * an update too.
3049                  */
3050                 left_path = ocfs2_new_path_from_path(path);
3051                 if (!left_path) {
3052                         ret = -ENOMEM;
3053                         mlog_errno(ret);
3054                         goto out;
3055                 }
3056
3057                 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3058                 if (ret) {
3059                         mlog_errno(ret);
3060                         goto out;
3061                 }
3062
3063                 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3064                 if (ret) {
3065                         mlog_errno(ret);
3066                         goto out;
3067                 }
3068
3069                 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3070
3071                 ocfs2_unlink_subtree(handle, et, left_path, path,
3072                                      subtree_index, dealloc);
3073                 ret = ocfs2_update_edge_lengths(handle, et, left_path);
3074                 if (ret) {
3075                         mlog_errno(ret);
3076                         goto out;
3077                 }
3078
3079                 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3080                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3081         } else {
3082                 /*
3083                  * 'path' is also the leftmost path which
3084                  * means it must be the only one. This gets
3085                  * handled differently because we want to
3086                  * revert the root back to having extents
3087                  * in-line.
3088                  */
3089                 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3090
3091                 el = et->et_root_el;
3092                 el->l_tree_depth = 0;
3093                 el->l_next_free_rec = 0;
3094                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3095
3096                 ocfs2_et_set_last_eb_blk(et, 0);
3097         }
3098
3099         ocfs2_journal_dirty(handle, path_root_bh(path));
3100
3101 out:
3102         ocfs2_free_path(left_path);
3103         return ret;
3104 }
3105
3106 static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3107                                 struct ocfs2_extent_tree *et,
3108                                 struct ocfs2_path *path,
3109                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
3110 {
3111         handle_t *handle;
3112         int ret;
3113         int credits = path->p_tree_depth * 2 + 1;
3114
3115         handle = ocfs2_start_trans(osb, credits);
3116         if (IS_ERR(handle)) {
3117                 ret = PTR_ERR(handle);
3118                 mlog_errno(ret);
3119                 return ret;
3120         }
3121
3122         ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3123         if (ret)
3124                 mlog_errno(ret);
3125
3126         ocfs2_commit_trans(osb, handle);
3127         return ret;
3128 }
3129
3130 /*
3131  * Left rotation of btree records.
3132  *
3133  * In many ways, this is (unsurprisingly) the opposite of right
3134  * rotation. We start at some non-rightmost path containing an empty
3135  * extent in the leaf block. The code works its way to the rightmost
3136  * path by rotating records to the left in every subtree.
3137  *
3138  * This is used by any code which reduces the number of extent records
3139  * in a leaf. After removal, an empty record should be placed in the
3140  * leftmost list position.
3141  *
3142  * This won't handle a length update of the rightmost path records if
3143  * the rightmost tree leaf record is removed so the caller is
3144  * responsible for detecting and correcting that.
3145  */
3146 static int ocfs2_rotate_tree_left(handle_t *handle,
3147                                   struct ocfs2_extent_tree *et,
3148                                   struct ocfs2_path *path,
3149                                   struct ocfs2_cached_dealloc_ctxt *dealloc)
3150 {
3151         int ret, orig_credits = handle->h_buffer_credits;
3152         struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3153         struct ocfs2_extent_block *eb;
3154         struct ocfs2_extent_list *el;
3155
3156         el = path_leaf_el(path);
3157         if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3158                 return 0;
3159
3160         if (path->p_tree_depth == 0) {
3161 rightmost_no_delete:
3162                 /*
3163                  * Inline extents. This is trivially handled, so do
3164                  * it up front.
3165                  */
3166                 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3167                 if (ret)
3168                         mlog_errno(ret);
3169                 goto out;
3170         }
3171
3172         /*
3173          * Handle rightmost branch now. There's several cases:
3174          *  1) simple rotation leaving records in there. That's trivial.
3175          *  2) rotation requiring a branch delete - there's no more
3176          *     records left. Two cases of this:
3177          *     a) There are branches to the left.
3178          *     b) This is also the leftmost (the only) branch.
3179          *
3180          *  1) is handled via ocfs2_rotate_rightmost_leaf_left()
3181          *  2a) we need the left branch so that we can update it with the unlink
3182          *  2b) we need to bring the root back to inline extents.
3183          */
3184
3185         eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3186         el = &eb->h_list;
3187         if (eb->h_next_leaf_blk == 0) {
3188                 /*
3189                  * This gets a bit tricky if we're going to delete the
3190                  * rightmost path. Get the other cases out of the way
3191                  * 1st.
3192                  */
3193                 if (le16_to_cpu(el->l_next_free_rec) > 1)
3194                         goto rightmost_no_delete;
3195
3196                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3197                         ret = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3198                                         "Owner %llu has empty extent block at %llu\n",
3199                                         (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3200                                         (unsigned long long)le64_to_cpu(eb->h_blkno));
3201                         goto out;
3202                 }
3203
3204                 /*
3205                  * XXX: The caller can not trust "path" any more after
3206                  * this as it will have been deleted. What do we do?
3207                  *
3208                  * In theory the rotate-for-merge code will never get
3209                  * here because it'll always ask for a rotate in a
3210                  * nonempty list.
3211                  */
3212
3213                 ret = ocfs2_remove_rightmost_path(handle, et, path,
3214                                                   dealloc);
3215                 if (ret)
3216                         mlog_errno(ret);
3217                 goto out;
3218         }
3219
3220         /*
3221          * Now we can loop, remembering the path we get from -EAGAIN
3222          * and restarting from there.
3223          */
3224 try_rotate:
3225         ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3226                                        dealloc, &restart_path);
3227         if (ret && ret != -EAGAIN) {
3228                 mlog_errno(ret);
3229                 goto out;
3230         }
3231
3232         while (ret == -EAGAIN) {
3233                 tmp_path = restart_path;
3234                 restart_path = NULL;
3235
3236                 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3237                                                tmp_path, dealloc,
3238                                                &restart_path);
3239                 if (ret && ret != -EAGAIN) {
3240                         mlog_errno(ret);
3241                         goto out;
3242                 }
3243
3244                 ocfs2_free_path(tmp_path);
3245                 tmp_path = NULL;
3246
3247                 if (ret == 0)
3248                         goto try_rotate;
3249         }
3250
3251 out:
3252         ocfs2_free_path(tmp_path);
3253         ocfs2_free_path(restart_path);
3254         return ret;
3255 }
3256
3257 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3258                                 int index)
3259 {
3260         struct ocfs2_extent_rec *rec = &el->l_recs[index];
3261         unsigned int size;
3262
3263         if (rec->e_leaf_clusters == 0) {
3264                 /*
3265                  * We consumed all of the merged-from record. An empty
3266                  * extent cannot exist anywhere but the 1st array
3267                  * position, so move things over if the merged-from
3268                  * record doesn't occupy that position.
3269                  *
3270                  * This creates a new empty extent so the caller
3271                  * should be smart enough to have removed any existing
3272                  * ones.
3273                  */
3274                 if (index > 0) {
3275                         BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3276                         size = index * sizeof(struct ocfs2_extent_rec);
3277                         memmove(&el->l_recs[1], &el->l_recs[0], size);
3278                 }
3279
3280                 /*
3281                  * Always memset - the caller doesn't check whether it
3282                  * created an empty extent, so there could be junk in
3283                  * the other fields.
3284                  */
3285                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3286         }
3287 }
3288
3289 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3290                                 struct ocfs2_path *left_path,
3291                                 struct ocfs2_path **ret_right_path)
3292 {
3293         int ret;
3294         u32 right_cpos;
3295         struct ocfs2_path *right_path = NULL;
3296         struct ocfs2_extent_list *left_el;
3297
3298         *ret_right_path = NULL;
3299
3300         /* This function shouldn't be called for non-trees. */
3301         BUG_ON(left_path->p_tree_depth == 0);
3302
3303         left_el = path_leaf_el(left_path);
3304         BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3305
3306         ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3307                                              left_path, &right_cpos);
3308         if (ret) {
3309                 mlog_errno(ret);
3310                 goto out;
3311         }
3312
3313         /* This function shouldn't be called for the rightmost leaf. */
3314         BUG_ON(right_cpos == 0);
3315
3316         right_path = ocfs2_new_path_from_path(left_path);
3317         if (!right_path) {
3318                 ret = -ENOMEM;
3319                 mlog_errno(ret);
3320                 goto out;
3321         }
3322
3323         ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3324         if (ret) {
3325                 mlog_errno(ret);
3326                 goto out;
3327         }
3328
3329         *ret_right_path = right_path;
3330 out:
3331         if (ret)
3332                 ocfs2_free_path(right_path);
3333         return ret;
3334 }
3335
3336 /*
3337  * Remove split_rec clusters from the record at index and merge them
3338  * onto the beginning of the record "next" to it.
3339  * For index < l_count - 1, the next means the extent rec at index + 1.
3340  * For index == l_count - 1, the "next" means the 1st extent rec of the
3341  * next extent block.
3342  */
3343 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3344                                  handle_t *handle,
3345                                  struct ocfs2_extent_tree *et,
3346                                  struct ocfs2_extent_rec *split_rec,
3347                                  int index)
3348 {
3349         int ret, next_free, i;
3350         unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3351         struct ocfs2_extent_rec *left_rec;
3352         struct ocfs2_extent_rec *right_rec;
3353         struct ocfs2_extent_list *right_el;
3354         struct ocfs2_path *right_path = NULL;
3355         int subtree_index = 0;
3356         struct ocfs2_extent_list *el = path_leaf_el(left_path);
3357         struct buffer_head *bh = path_leaf_bh(left_path);
3358         struct buffer_head *root_bh = NULL;
3359
3360         BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3361         left_rec = &el->l_recs[index];
3362
3363         if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3364             le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3365                 /* we meet with a cross extent block merge. */
3366                 ret = ocfs2_get_right_path(et, left_path, &right_path);
3367                 if (ret) {
3368                         mlog_errno(ret);
3369                         return ret;
3370                 }
3371
3372                 right_el = path_leaf_el(right_path);
3373                 next_free = le16_to_cpu(right_el->l_next_free_rec);
3374                 BUG_ON(next_free <= 0);
3375                 right_rec = &right_el->l_recs[0];
3376                 if (ocfs2_is_empty_extent(right_rec)) {
3377                         BUG_ON(next_free <= 1);
3378                         right_rec = &right_el->l_recs[1];
3379                 }
3380
3381                 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3382                        le16_to_cpu(left_rec->e_leaf_clusters) !=
3383                        le32_to_cpu(right_rec->e_cpos));
3384
3385                 subtree_index = ocfs2_find_subtree_root(et, left_path,
3386                                                         right_path);
3387
3388                 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3389                                                       handle->h_buffer_credits,
3390                                                       right_path);
3391                 if (ret) {
3392                         mlog_errno(ret);
3393                         goto out;
3394                 }
3395
3396                 root_bh = left_path->p_node[subtree_index].bh;
3397                 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3398
3399                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3400                                                    subtree_index);
3401                 if (ret) {
3402                         mlog_errno(ret);
3403                         goto out;
3404                 }
3405
3406                 for (i = subtree_index + 1;
3407                      i < path_num_items(right_path); i++) {
3408                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3409                                                            right_path, i);
3410                         if (ret) {
3411                                 mlog_errno(ret);
3412                                 goto out;
3413                         }
3414
3415                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3416                                                            left_path, i);
3417                         if (ret) {
3418                                 mlog_errno(ret);
3419                                 goto out;
3420                         }
3421                 }
3422
3423         } else {
3424                 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3425                 right_rec = &el->l_recs[index + 1];
3426         }
3427
3428         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3429                                            path_num_items(left_path) - 1);
3430         if (ret) {
3431                 mlog_errno(ret);
3432                 goto out;
3433         }
3434
3435         le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3436
3437         le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3438         le64_add_cpu(&right_rec->e_blkno,
3439                      -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3440                                                split_clusters));
3441         le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3442
3443         ocfs2_cleanup_merge(el, index);
3444
3445         ocfs2_journal_dirty(handle, bh);
3446         if (right_path) {
3447                 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3448                 ocfs2_complete_edge_insert(handle, left_path, right_path,
3449                                            subtree_index);
3450         }
3451 out:
3452         ocfs2_free_path(right_path);
3453         return ret;
3454 }
3455
3456 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3457                                struct ocfs2_path *right_path,
3458                                struct ocfs2_path **ret_left_path)
3459 {
3460         int ret;
3461         u32 left_cpos;
3462         struct ocfs2_path *left_path = NULL;
3463
3464         *ret_left_path = NULL;
3465
3466         /* This function shouldn't be called for non-trees. */
3467         BUG_ON(right_path->p_tree_depth == 0);
3468
3469         ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3470                                             right_path, &left_cpos);
3471         if (ret) {
3472                 mlog_errno(ret);
3473                 goto out;
3474         }
3475
3476         /* This function shouldn't be called for the leftmost leaf. */
3477         BUG_ON(left_cpos == 0);
3478
3479         left_path = ocfs2_new_path_from_path(right_path);
3480         if (!left_path) {
3481                 ret = -ENOMEM;
3482                 mlog_errno(ret);
3483                 goto out;
3484         }
3485
3486         ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3487         if (ret) {
3488                 mlog_errno(ret);
3489                 goto out;
3490         }
3491
3492         *ret_left_path = left_path;
3493 out:
3494         if (ret)
3495                 ocfs2_free_path(left_path);
3496         return ret;
3497 }
3498
3499 /*
3500  * Remove split_rec clusters from the record at index and merge them
3501  * onto the tail of the record "before" it.
3502  * For index > 0, the "before" means the extent rec at index - 1.
3503  *
3504  * For index == 0, the "before" means the last record of the previous
3505  * extent block. And there is also a situation that we may need to
3506  * remove the rightmost leaf extent block in the right_path and change
3507  * the right path to indicate the new rightmost path.
3508  */
3509 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3510                                 handle_t *handle,
3511                                 struct ocfs2_extent_tree *et,
3512                                 struct ocfs2_extent_rec *split_rec,
3513                                 struct ocfs2_cached_dealloc_ctxt *dealloc,
3514                                 int index)
3515 {
3516         int ret, i, subtree_index = 0, has_empty_extent = 0;
3517         unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3518         struct ocfs2_extent_rec *left_rec;
3519         struct ocfs2_extent_rec *right_rec;
3520         struct ocfs2_extent_list *el = path_leaf_el(right_path);
3521         struct buffer_head *bh = path_leaf_bh(right_path);
3522         struct buffer_head *root_bh = NULL;
3523         struct ocfs2_path *left_path = NULL;
3524         struct ocfs2_extent_list *left_el;
3525
3526         BUG_ON(index < 0);
3527
3528         right_rec = &el->l_recs[index];
3529         if (index == 0) {
3530                 /* we meet with a cross extent block merge. */
3531                 ret = ocfs2_get_left_path(et, right_path, &left_path);
3532                 if (ret) {
3533                         mlog_errno(ret);
3534                         return ret;
3535                 }
3536
3537                 left_el = path_leaf_el(left_path);
3538                 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3539                        le16_to_cpu(left_el->l_count));
3540
3541                 left_rec = &left_el->l_recs[
3542                                 le16_to_cpu(left_el->l_next_free_rec) - 1];
3543                 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3544                        le16_to_cpu(left_rec->e_leaf_clusters) !=
3545                        le32_to_cpu(split_rec->e_cpos));
3546
3547                 subtree_index = ocfs2_find_subtree_root(et, left_path,
3548                                                         right_path);
3549
3550                 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3551                                                       handle->h_buffer_credits,
3552                                                       left_path);
3553                 if (ret) {
3554                         mlog_errno(ret);
3555                         goto out;
3556                 }
3557
3558                 root_bh = left_path->p_node[subtree_index].bh;
3559                 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3560
3561                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3562                                                    subtree_index);
3563                 if (ret) {
3564                         mlog_errno(ret);
3565                         goto out;
3566                 }
3567
3568                 for (i = subtree_index + 1;
3569                      i < path_num_items(right_path); i++) {
3570                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3571                                                            right_path, i);
3572                         if (ret) {
3573                                 mlog_errno(ret);
3574                                 goto out;
3575                         }
3576
3577                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3578                                                            left_path, i);
3579                         if (ret) {
3580                                 mlog_errno(ret);
3581                                 goto out;
3582                         }
3583                 }
3584         } else {
3585                 left_rec = &el->l_recs[index - 1];
3586                 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3587                         has_empty_extent = 1;
3588         }
3589
3590         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3591                                            path_num_items(right_path) - 1);
3592         if (ret) {
3593                 mlog_errno(ret);
3594                 goto out;
3595         }
3596
3597         if (has_empty_extent && index == 1) {
3598                 /*
3599                  * The easy case - we can just plop the record right in.
3600                  */
3601                 *left_rec = *split_rec;
3602         } else
3603                 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3604
3605         le32_add_cpu(&right_rec->e_cpos, split_clusters);
3606         le64_add_cpu(&right_rec->e_blkno,
3607                      ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3608                                               split_clusters));
3609         le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3610
3611         ocfs2_cleanup_merge(el, index);
3612
3613         ocfs2_journal_dirty(handle, bh);
3614         if (left_path) {
3615                 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3616
3617                 /*
3618                  * In the situation that the right_rec is empty and the extent
3619                  * block is empty also,  ocfs2_complete_edge_insert can't handle
3620                  * it and we need to delete the right extent block.
3621                  */
3622                 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3623                     le16_to_cpu(el->l_next_free_rec) == 1) {
3624                         /* extend credit for ocfs2_remove_rightmost_path */
3625                         ret = ocfs2_extend_rotate_transaction(handle, 0,
3626                                         handle->h_buffer_credits,
3627                                         right_path);
3628                         if (ret) {
3629                                 mlog_errno(ret);
3630                                 goto out;
3631                         }
3632
3633                         ret = ocfs2_remove_rightmost_path(handle, et,
3634                                                           right_path,
3635                                                           dealloc);
3636                         if (ret) {
3637                                 mlog_errno(ret);
3638                                 goto out;
3639                         }
3640
3641                         /* Now the rightmost extent block has been deleted.
3642                          * So we use the new rightmost path.
3643                          */
3644                         ocfs2_mv_path(right_path, left_path);
3645                         left_path = NULL;
3646                 } else
3647                         ocfs2_complete_edge_insert(handle, left_path,
3648                                                    right_path, subtree_index);
3649         }
3650 out:
3651         ocfs2_free_path(left_path);
3652         return ret;
3653 }
3654
3655 static int ocfs2_try_to_merge_extent(handle_t *handle,
3656                                      struct ocfs2_extent_tree *et,
3657                                      struct ocfs2_path *path,
3658                                      int split_index,
3659                                      struct ocfs2_extent_rec *split_rec,
3660                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
3661                                      struct ocfs2_merge_ctxt *ctxt)
3662 {
3663         int ret = 0;
3664         struct ocfs2_extent_list *el = path_leaf_el(path);
3665         struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3666
3667         BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3668
3669         if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3670                 /* extend credit for ocfs2_remove_rightmost_path */
3671                 ret = ocfs2_extend_rotate_transaction(handle, 0,
3672                                 handle->h_buffer_credits,
3673                                 path);
3674                 if (ret) {
3675                         mlog_errno(ret);
3676                         goto out;
3677                 }
3678                 /*
3679                  * The merge code will need to create an empty
3680                  * extent to take the place of the newly
3681                  * emptied slot. Remove any pre-existing empty
3682                  * extents - having more than one in a leaf is
3683                  * illegal.
3684                  */
3685                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3686                 if (ret) {
3687                         mlog_errno(ret);
3688                         goto out;
3689                 }
3690                 split_index--;
3691                 rec = &el->l_recs[split_index];
3692         }
3693
3694         if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3695                 /*
3696                  * Left-right contig implies this.
3697                  */
3698                 BUG_ON(!ctxt->c_split_covers_rec);
3699
3700                 /*
3701                  * Since the leftright insert always covers the entire
3702                  * extent, this call will delete the insert record
3703                  * entirely, resulting in an empty extent record added to
3704                  * the extent block.
3705                  *
3706                  * Since the adding of an empty extent shifts
3707                  * everything back to the right, there's no need to
3708                  * update split_index here.
3709                  *
3710                  * When the split_index is zero, we need to merge it to the
3711                  * prevoius extent block. It is more efficient and easier
3712                  * if we do merge_right first and merge_left later.
3713                  */
3714                 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3715                                             split_index);
3716                 if (ret) {
3717                         mlog_errno(ret);
3718                         goto out;
3719                 }
3720
3721                 /*
3722                  * We can only get this from logic error above.
3723                  */
3724                 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3725
3726                 /* extend credit for ocfs2_remove_rightmost_path */
3727                 ret = ocfs2_extend_rotate_transaction(handle, 0,
3728                                         handle->h_buffer_credits,
3729                                         path);
3730                 if (ret) {
3731                         mlog_errno(ret);
3732                         goto out;
3733                 }
3734
3735                 /* The merge left us with an empty extent, remove it. */
3736                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3737                 if (ret) {
3738                         mlog_errno(ret);
3739                         goto out;
3740                 }
3741
3742                 rec = &el->l_recs[split_index];
3743
3744                 /*
3745                  * Note that we don't pass split_rec here on purpose -
3746                  * we've merged it into the rec already.
3747                  */
3748                 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3749                                            dealloc, split_index);
3750
3751                 if (ret) {
3752                         mlog_errno(ret);
3753                         goto out;
3754                 }
3755
3756                 /* extend credit for ocfs2_remove_rightmost_path */
3757                 ret = ocfs2_extend_rotate_transaction(handle, 0,
3758                                 handle->h_buffer_credits,
3759                                 path);
3760                 if (ret) {
3761                         mlog_errno(ret);
3762                         goto out;
3763                 }
3764
3765                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3766                 /*
3767                  * Error from this last rotate is not critical, so
3768                  * print but don't bubble it up.
3769                  */
3770                 if (ret)
3771                         mlog_errno(ret);
3772                 ret = 0;
3773         } else {
3774                 /*
3775                  * Merge a record to the left or right.
3776                  *
3777                  * 'contig_type' is relative to the existing record,
3778                  * so for example, if we're "right contig", it's to
3779                  * the record on the left (hence the left merge).
3780                  */
3781                 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3782                         ret = ocfs2_merge_rec_left(path, handle, et,
3783                                                    split_rec, dealloc,
3784                                                    split_index);
3785                         if (ret) {
3786                                 mlog_errno(ret);
3787                                 goto out;
3788                         }
3789                 } else {
3790                         ret = ocfs2_merge_rec_right(path, handle,
3791                                                     et, split_rec,
3792                                                     split_index);
3793                         if (ret) {
3794                                 mlog_errno(ret);
3795                                 goto out;
3796                         }
3797                 }
3798
3799                 if (ctxt->c_split_covers_rec) {
3800                         /* extend credit for ocfs2_remove_rightmost_path */
3801                         ret = ocfs2_extend_rotate_transaction(handle, 0,
3802                                         handle->h_buffer_credits,
3803                                         path);
3804                         if (ret) {
3805                                 mlog_errno(ret);
3806                                 ret = 0;
3807                                 goto out;
3808                         }
3809
3810                         /*
3811                          * The merge may have left an empty extent in
3812                          * our leaf. Try to rotate it away.
3813                          */
3814                         ret = ocfs2_rotate_tree_left(handle, et, path,
3815                                                      dealloc);
3816                         if (ret)
3817                                 mlog_errno(ret);
3818                         ret = 0;
3819                 }
3820         }
3821
3822 out:
3823         return ret;
3824 }
3825
3826 static void ocfs2_subtract_from_rec(struct super_block *sb,
3827                                     enum ocfs2_split_type split,
3828                                     struct ocfs2_extent_rec *rec,
3829                                     struct ocfs2_extent_rec *split_rec)
3830 {
3831         u64 len_blocks;
3832
3833         len_blocks = ocfs2_clusters_to_blocks(sb,
3834                                 le16_to_cpu(split_rec->e_leaf_clusters));
3835
3836         if (split == SPLIT_LEFT) {
3837                 /*
3838                  * Region is on the left edge of the existing
3839                  * record.
3840                  */
3841                 le32_add_cpu(&rec->e_cpos,
3842                              le16_to_cpu(split_rec->e_leaf_clusters));
3843                 le64_add_cpu(&rec->e_blkno, len_blocks);
3844                 le16_add_cpu(&rec->e_leaf_clusters,
3845                              -le16_to_cpu(split_rec->e_leaf_clusters));
3846         } else {
3847                 /*
3848                  * Region is on the right edge of the existing
3849                  * record.
3850                  */
3851                 le16_add_cpu(&rec->e_leaf_clusters,
3852                              -le16_to_cpu(split_rec->e_leaf_clusters));
3853         }
3854 }
3855
3856 /*
3857  * Do the final bits of extent record insertion at the target leaf
3858  * list. If this leaf is part of an allocation tree, it is assumed
3859  * that the tree above has been prepared.
3860  */
3861 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3862                                  struct ocfs2_extent_rec *insert_rec,
3863                                  struct ocfs2_extent_list *el,
3864                                  struct ocfs2_insert_type *insert)
3865 {
3866         int i = insert->ins_contig_index;
3867         unsigned int range;
3868         struct ocfs2_extent_rec *rec;
3869
3870         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3871
3872         if (insert->ins_split != SPLIT_NONE) {
3873                 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3874                 BUG_ON(i == -1);
3875                 rec = &el->l_recs[i];
3876                 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3877                                         insert->ins_split, rec,
3878                                         insert_rec);
3879                 goto rotate;
3880         }
3881
3882         /*
3883          * Contiguous insert - either left or right.
3884          */
3885         if (insert->ins_contig != CONTIG_NONE) {
3886                 rec = &el->l_recs[i];
3887                 if (insert->ins_contig == CONTIG_LEFT) {
3888                         rec->e_blkno = insert_rec->e_blkno;
3889                         rec->e_cpos = insert_rec->e_cpos;
3890                 }
3891                 le16_add_cpu(&rec->e_leaf_clusters,
3892                              le16_to_cpu(insert_rec->e_leaf_clusters));
3893                 return;
3894         }
3895
3896         /*
3897          * Handle insert into an empty leaf.
3898          */
3899         if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3900             ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3901              ocfs2_is_empty_extent(&el->l_recs[0]))) {
3902                 el->l_recs[0] = *insert_rec;
3903                 el->l_next_free_rec = cpu_to_le16(1);
3904                 return;
3905         }
3906
3907         /*
3908          * Appending insert.
3909          */
3910         if (insert->ins_appending == APPEND_TAIL) {
3911                 i = le16_to_cpu(el->l_next_free_rec) - 1;
3912                 rec = &el->l_recs[i];
3913                 range = le32_to_cpu(rec->e_cpos)
3914                         + le16_to_cpu(rec->e_leaf_clusters);
3915                 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3916
3917                 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3918                                 le16_to_cpu(el->l_count),
3919                                 "owner %llu, depth %u, count %u, next free %u, "
3920                                 "rec.cpos %u, rec.clusters %u, "
3921                                 "insert.cpos %u, insert.clusters %u\n",
3922                                 ocfs2_metadata_cache_owner(et->et_ci),
3923                                 le16_to_cpu(el->l_tree_depth),
3924                                 le16_to_cpu(el->l_count),
3925                                 le16_to_cpu(el->l_next_free_rec),
3926                                 le32_to_cpu(el->l_recs[i].e_cpos),
3927                                 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3928                                 le32_to_cpu(insert_rec->e_cpos),
3929                                 le16_to_cpu(insert_rec->e_leaf_clusters));
3930                 i++;
3931                 el->l_recs[i] = *insert_rec;
3932                 le16_add_cpu(&el->l_next_free_rec, 1);
3933                 return;
3934         }
3935
3936 rotate:
3937         /*
3938          * Ok, we have to rotate.
3939          *
3940          * At this point, it is safe to assume that inserting into an
3941          * empty leaf and appending to a leaf have both been handled
3942          * above.
3943          *
3944          * This leaf needs to have space, either by the empty 1st
3945          * extent record, or by virtue of an l_next_rec < l_count.
3946          */
3947         ocfs2_rotate_leaf(el, insert_rec);
3948 }
3949
3950 static void ocfs2_adjust_rightmost_records(handle_t *handle,
3951                                            struct ocfs2_extent_tree *et,
3952                                            struct ocfs2_path *path,
3953                                            struct ocfs2_extent_rec *insert_rec)
3954 {
3955         int i, next_free;
3956         struct buffer_head *bh;
3957         struct ocfs2_extent_list *el;
3958         struct ocfs2_extent_rec *rec;
3959
3960         /*
3961          * Update everything except the leaf block.
3962          */
3963         for (i = 0; i < path->p_tree_depth; i++) {
3964                 bh = path->p_node[i].bh;
3965                 el = path->p_node[i].el;
3966
3967                 next_free = le16_to_cpu(el->l_next_free_rec);
3968                 if (next_free == 0) {
3969                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3970                                     "Owner %llu has a bad extent list\n",
3971                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3972                         return;
3973                 }
3974
3975                 rec = &el->l_recs[next_free - 1];
3976
3977                 rec->e_int_clusters = insert_rec->e_cpos;
3978                 le32_add_cpu(&rec->e_int_clusters,
3979                              le16_to_cpu(insert_rec->e_leaf_clusters));
3980                 le32_add_cpu(&rec->e_int_clusters,
3981                              -le32_to_cpu(rec->e_cpos));
3982
3983                 ocfs2_journal_dirty(handle, bh);
3984         }
3985 }
3986
3987 static int ocfs2_append_rec_to_path(handle_t *handle,
3988                                     struct ocfs2_extent_tree *et,
3989                                     struct ocfs2_extent_rec *insert_rec,
3990                                     struct ocfs2_path *right_path,
3991                                     struct ocfs2_path **ret_left_path)
3992 {
3993         int ret, next_free;
3994         struct ocfs2_extent_list *el;
3995         struct ocfs2_path *left_path = NULL;
3996
3997         *ret_left_path = NULL;
3998
3999         /*
4000          * This shouldn't happen for non-trees. The extent rec cluster
4001          * count manipulation below only works for interior nodes.
4002          */
4003         BUG_ON(right_path->p_tree_depth == 0);
4004
4005         /*
4006          * If our appending insert is at the leftmost edge of a leaf,
4007          * then we might need to update the rightmost records of the
4008          * neighboring path.
4009          */
4010         el = path_leaf_el(right_path);
4011         next_free = le16_to_cpu(el->l_next_free_rec);
4012         if (next_free == 0 ||
4013             (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4014                 u32 left_cpos;
4015
4016                 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4017                                                     right_path, &left_cpos);
4018                 if (ret) {
4019                         mlog_errno(ret);
4020                         goto out;
4021                 }
4022
4023                 trace_ocfs2_append_rec_to_path(
4024                         (unsigned long long)
4025                         ocfs2_metadata_cache_owner(et->et_ci),
4026                         le32_to_cpu(insert_rec->e_cpos),
4027                         left_cpos);
4028
4029                 /*
4030                  * No need to worry if the append is already in the
4031                  * leftmost leaf.
4032                  */
4033                 if (left_cpos) {
4034                         left_path = ocfs2_new_path_from_path(right_path);
4035                         if (!left_path) {
4036                                 ret = -ENOMEM;
4037                                 mlog_errno(ret);
4038                                 goto out;
4039                         }
4040
4041                         ret = ocfs2_find_path(et->et_ci, left_path,
4042                                               left_cpos);
4043                         if (ret) {
4044                                 mlog_errno(ret);
4045                                 goto out;
4046                         }
4047
4048                         /*
4049                          * ocfs2_insert_path() will pass the left_path to the
4050                          * journal for us.
4051                          */
4052                 }
4053         }
4054
4055         ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4056         if (ret) {
4057                 mlog_errno(ret);
4058                 goto out;
4059         }
4060
4061         ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4062
4063         *ret_left_path = left_path;
4064         ret = 0;
4065 out:
4066         if (ret != 0)
4067                 ocfs2_free_path(left_path);
4068
4069         return ret;
4070 }
4071
4072 static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4073                                struct ocfs2_path *left_path,
4074                                struct ocfs2_path *right_path,
4075                                struct ocfs2_extent_rec *split_rec,
4076                                enum ocfs2_split_type split)
4077 {
4078         int index;
4079         u32 cpos = le32_to_cpu(split_rec->e_cpos);
4080         struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4081         struct ocfs2_extent_rec *rec, *tmprec;
4082
4083         right_el = path_leaf_el(right_path);
4084         if (left_path)
4085                 left_el = path_leaf_el(left_path);
4086
4087         el = right_el;
4088         insert_el = right_el;
4089         index = ocfs2_search_extent_list(el, cpos);
4090         if (index != -1) {
4091                 if (index == 0 && left_path) {
4092                         BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4093
4094                         /*
4095                          * This typically means that the record
4096                          * started in the left path but moved to the
4097                          * right as a result of rotation. We either
4098                          * move the existing record to the left, or we
4099                          * do the later insert there.
4100                          *
4101                          * In this case, the left path should always
4102                          * exist as the rotate code will have passed
4103                          * it back for a post-insert update.
4104                          */
4105
4106                         if (split == SPLIT_LEFT) {
4107                                 /*
4108                                  * It's a left split. Since we know
4109                                  * that the rotate code gave us an
4110                                  * empty extent in the left path, we
4111                                  * can just do the insert there.
4112                                  */
4113                                 insert_el = left_el;
4114                         } else {
4115                                 /*
4116                                  * Right split - we have to move the
4117                                  * existing record over to the left
4118                                  * leaf. The insert will be into the
4119                                  * newly created empty extent in the
4120                                  * right leaf.
4121                                  */
4122                                 tmprec = &right_el->l_recs[index];
4123                                 ocfs2_rotate_leaf(left_el, tmprec);
4124                                 el = left_el;
4125
4126                                 memset(tmprec, 0, sizeof(*tmprec));
4127                                 index = ocfs2_search_extent_list(left_el, cpos);
4128                                 BUG_ON(index == -1);
4129                         }
4130                 }
4131         } else {
4132                 BUG_ON(!left_path);
4133                 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4134                 /*
4135                  * Left path is easy - we can just allow the insert to
4136                  * happen.
4137                  */
4138                 el = left_el;
4139                 insert_el = left_el;
4140                 index = ocfs2_search_extent_list(el, cpos);
4141                 BUG_ON(index == -1);
4142         }
4143
4144         rec = &el->l_recs[index];
4145         ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4146                                 split, rec, split_rec);
4147         ocfs2_rotate_leaf(insert_el, split_rec);
4148 }
4149
4150 /*
4151  * This function only does inserts on an allocation b-tree. For tree
4152  * depth = 0, ocfs2_insert_at_leaf() is called directly.
4153  *
4154  * right_path is the path we want to do the actual insert
4155  * in. left_path should only be passed in if we need to update that
4156  * portion of the tree after an edge insert.
4157  */
4158 static int ocfs2_insert_path(handle_t *handle,
4159                              struct ocfs2_extent_tree *et,
4160                              struct ocfs2_path *left_path,
4161                              struct ocfs2_path *right_path,
4162                              struct ocfs2_extent_rec *insert_rec,
4163                              struct ocfs2_insert_type *insert)
4164 {
4165         int ret, subtree_index;
4166         struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4167
4168         if (left_path) {
4169                 /*
4170                  * There's a chance that left_path got passed back to
4171                  * us without being accounted for in the
4172                  * journal. Extend our transaction here to be sure we
4173                  * can change those blocks.
4174                  */
4175                 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4176                 if (ret < 0) {
4177                         mlog_errno(ret);
4178                         goto out;
4179                 }
4180
4181                 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4182                 if (ret < 0) {
4183                         mlog_errno(ret);
4184                         goto out;
4185                 }
4186         }
4187
4188         /*
4189          * Pass both paths to the journal. The majority of inserts
4190          * will be touching all components anyway.
4191          */
4192         ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4193         if (ret < 0) {
4194                 mlog_errno(ret);
4195                 goto out;
4196         }
4197
4198         if (insert->ins_split != SPLIT_NONE) {
4199                 /*
4200                  * We could call ocfs2_insert_at_leaf() for some types
4201                  * of splits, but it's easier to just let one separate
4202                  * function sort it all out.
4203                  */
4204                 ocfs2_split_record(et, left_path, right_path,
4205                                    insert_rec, insert->ins_split);
4206
4207                 /*
4208                  * Split might have modified either leaf and we don't
4209                  * have a guarantee that the later edge insert will
4210                  * dirty this for us.
4211                  */
4212                 if (left_path)
4213                         ocfs2_journal_dirty(handle,
4214                                             path_leaf_bh(left_path));
4215         } else
4216                 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4217                                      insert);
4218
4219         ocfs2_journal_dirty(handle, leaf_bh);
4220
4221         if (left_path) {
4222                 /*
4223                  * The rotate code has indicated that we need to fix
4224                  * up portions of the tree after the insert.
4225                  *
4226                  * XXX: Should we extend the transaction here?
4227                  */
4228                 subtree_index = ocfs2_find_subtree_root(et, left_path,
4229                                                         right_path);
4230                 ocfs2_complete_edge_insert(handle, left_path, right_path,
4231                                            subtree_index);
4232         }
4233
4234         ret = 0;
4235 out:
4236         return ret;
4237 }
4238
4239 static int ocfs2_do_insert_extent(handle_t *handle,
4240                                   struct ocfs2_extent_tree *et,
4241                                   struct ocfs2_extent_rec *insert_rec,
4242                                   struct ocfs2_insert_type *type)
4243 {
4244         int ret, rotate = 0;
4245         u32 cpos;
4246         struct ocfs2_path *right_path = NULL;
4247         struct ocfs2_path *left_path = NULL;
4248         struct ocfs2_extent_list *el;
4249
4250         el = et->et_root_el;
4251
4252         ret = ocfs2_et_root_journal_access(handle, et,
4253                                            OCFS2_JOURNAL_ACCESS_WRITE);
4254         if (ret) {
4255                 mlog_errno(ret);
4256                 goto out;
4257         }
4258
4259         if (le16_to_cpu(el->l_tree_depth) == 0) {
4260                 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4261                 goto out_update_clusters;
4262         }
4263
4264         right_path = ocfs2_new_path_from_et(et);
4265         if (!right_path) {
4266                 ret = -ENOMEM;
4267                 mlog_errno(ret);
4268                 goto out;
4269         }
4270
4271         /*
4272          * Determine the path to start with. Rotations need the
4273          * rightmost path, everything else can go directly to the
4274          * target leaf.
4275          */
4276         cpos = le32_to_cpu(insert_rec->e_cpos);
4277         if (type->ins_appending == APPEND_NONE &&
4278             type->ins_contig == CONTIG_NONE) {
4279                 rotate = 1;
4280                 cpos = UINT_MAX;
4281         }
4282
4283         ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4284         if (ret) {
4285                 mlog_errno(ret);
4286                 goto out;
4287         }
4288
4289         /*
4290          * Rotations and appends need special treatment - they modify
4291          * parts of the tree's above them.
4292          *
4293          * Both might pass back a path immediate to the left of the
4294          * one being inserted to. This will be cause
4295          * ocfs2_insert_path() to modify the rightmost records of
4296          * left_path to account for an edge insert.
4297          *
4298          * XXX: When modifying this code, keep in mind that an insert
4299          * can wind up skipping both of these two special cases...
4300          */
4301         if (rotate) {
4302                 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4303                                               le32_to_cpu(insert_rec->e_cpos),
4304                                               right_path, &left_path);
4305                 if (ret) {
4306                         mlog_errno(ret);
4307                         goto out;
4308                 }
4309
4310                 /*
4311                  * ocfs2_rotate_tree_right() might have extended the
4312                  * transaction without re-journaling our tree root.
4313                  */
4314                 ret = ocfs2_et_root_journal_access(handle, et,
4315                                                    OCFS2_JOURNAL_ACCESS_WRITE);
4316                 if (ret) {
4317                         mlog_errno(ret);
4318                         goto out;
4319                 }
4320         } else if (type->ins_appending == APPEND_TAIL
4321                    && type->ins_contig != CONTIG_LEFT) {
4322                 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4323                                                right_path, &left_path);
4324                 if (ret) {
4325                         mlog_errno(ret);
4326                         goto out;
4327                 }
4328         }
4329
4330         ret = ocfs2_insert_path(handle, et, left_path, right_path,
4331                                 insert_rec, type);
4332         if (ret) {
4333                 mlog_errno(ret);
4334                 goto out;
4335         }
4336
4337 out_update_clusters:
4338         if (type->ins_split == SPLIT_NONE)
4339                 ocfs2_et_update_clusters(et,
4340                                          le16_to_cpu(insert_rec->e_leaf_clusters));
4341
4342         ocfs2_journal_dirty(handle, et->et_root_bh);
4343
4344 out:
4345         ocfs2_free_path(left_path);
4346         ocfs2_free_path(right_path);
4347
4348         return ret;
4349 }
4350
4351 static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4352                                struct ocfs2_path *path,
4353                                struct ocfs2_extent_list *el, int index,
4354                                struct ocfs2_extent_rec *split_rec,
4355                                struct ocfs2_merge_ctxt *ctxt)
4356 {
4357         int status = 0;
4358         enum ocfs2_contig_type ret = CONTIG_NONE;
4359         u32 left_cpos, right_cpos;
4360         struct ocfs2_extent_rec *rec = NULL;
4361         struct ocfs2_extent_list *new_el;
4362         struct ocfs2_path *left_path = NULL, *right_path = NULL;
4363         struct buffer_head *bh;
4364         struct ocfs2_extent_block *eb;
4365         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4366
4367         if (index > 0) {
4368                 rec = &el->l_recs[index - 1];
4369         } else if (path->p_tree_depth > 0) {
4370                 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4371                 if (status)
4372                         goto exit;
4373
4374                 if (left_cpos != 0) {
4375                         left_path = ocfs2_new_path_from_path(path);
4376                         if (!left_path) {
4377                                 status = -ENOMEM;
4378                                 mlog_errno(status);
4379                                 goto exit;
4380                         }
4381
4382                         status = ocfs2_find_path(et->et_ci, left_path,
4383                                                  left_cpos);
4384                         if (status)
4385                                 goto free_left_path;
4386
4387                         new_el = path_leaf_el(left_path);
4388
4389                         if (le16_to_cpu(new_el->l_next_free_rec) !=
4390                             le16_to_cpu(new_el->l_count)) {
4391                                 bh = path_leaf_bh(left_path);
4392                                 eb = (struct ocfs2_extent_block *)bh->b_data;
4393                                 status = ocfs2_error(sb,
4394                                                 "Extent block #%llu has an invalid l_next_free_rec of %d.  It should have matched the l_count of %d\n",
4395                                                 (unsigned long long)le64_to_cpu(eb->h_blkno),
4396                                                 le16_to_cpu(new_el->l_next_free_rec),
4397                                                 le16_to_cpu(new_el->l_count));
4398                                 goto free_left_path;
4399                         }
4400                         rec = &new_el->l_recs[
4401                                 le16_to_cpu(new_el->l_next_free_rec) - 1];
4402                 }
4403         }
4404
4405         /*
4406          * We're careful to check for an empty extent record here -
4407          * the merge code will know what to do if it sees one.
4408          */
4409         if (rec) {
4410                 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4411                         if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4412                                 ret = CONTIG_RIGHT;
4413                 } else {
4414                         ret = ocfs2_et_extent_contig(et, rec, split_rec);
4415                 }
4416         }
4417
4418         rec = NULL;
4419         if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4420                 rec = &el->l_recs[index + 1];
4421         else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4422                  path->p_tree_depth > 0) {
4423                 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4424                 if (status)
4425                         goto free_left_path;
4426
4427                 if (right_cpos == 0)
4428                         goto free_left_path;
4429
4430                 right_path = ocfs2_new_path_from_path(path);
4431                 if (!right_path) {
4432                         status = -ENOMEM;
4433                         mlog_errno(status);
4434                         goto free_left_path;
4435                 }
4436
4437                 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4438                 if (status)
4439                         goto free_right_path;
4440
4441                 new_el = path_leaf_el(right_path);
4442                 rec = &new_el->l_recs[0];
4443                 if (ocfs2_is_empty_extent(rec)) {
4444                         if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4445                                 bh = path_leaf_bh(right_path);
4446                                 eb = (struct ocfs2_extent_block *)bh->b_data;
4447                                 status = ocfs2_error(sb,
4448                                                 "Extent block #%llu has an invalid l_next_free_rec of %d\n",
4449                                                 (unsigned long long)le64_to_cpu(eb->h_blkno),
4450                                                 le16_to_cpu(new_el->l_next_free_rec));
4451                                 goto free_right_path;
4452                         }
4453                         rec = &new_el->l_recs[1];
4454                 }
4455         }
4456
4457         if (rec) {
4458                 enum ocfs2_contig_type contig_type;
4459
4460                 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4461
4462                 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4463                         ret = CONTIG_LEFTRIGHT;
4464                 else if (ret == CONTIG_NONE)
4465                         ret = contig_type;
4466         }
4467
4468 free_right_path:
4469         ocfs2_free_path(right_path);
4470 free_left_path:
4471         ocfs2_free_path(left_path);
4472 exit:
4473         if (status == 0)
4474                 ctxt->c_contig_type = ret;
4475
4476         return status;
4477 }
4478
4479 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4480                                      struct ocfs2_insert_type *insert,
4481                                      struct ocfs2_extent_list *el,
4482                                      struct ocfs2_extent_rec *insert_rec)
4483 {
4484         int i;
4485         enum ocfs2_contig_type contig_type = CONTIG_NONE;
4486
4487         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4488
4489         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4490                 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4491                                                      insert_rec);
4492                 if (contig_type != CONTIG_NONE) {
4493                         insert->ins_contig_index = i;
4494                         break;
4495                 }
4496         }
4497         insert->ins_contig = contig_type;
4498
4499         if (insert->ins_contig != CONTIG_NONE) {
4500                 struct ocfs2_extent_rec *rec =
4501                                 &el->l_recs[insert->ins_contig_index];
4502                 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4503                                    le16_to_cpu(insert_rec->e_leaf_clusters);
4504
4505                 /*
4506                  * Caller might want us to limit the size of extents, don't
4507                  * calculate contiguousness if we might exceed that limit.
4508                  */
4509                 if (et->et_max_leaf_clusters &&
4510                     (len > et->et_max_leaf_clusters))
4511                         insert->ins_contig = CONTIG_NONE;
4512         }
4513 }
4514
4515 /*
4516  * This should only be called against the righmost leaf extent list.
4517  *
4518  * ocfs2_figure_appending_type() will figure out whether we'll have to
4519  * insert at the tail of the rightmost leaf.
4520  *
4521  * This should also work against the root extent list for tree's with 0
4522  * depth. If we consider the root extent list to be the rightmost leaf node
4523  * then the logic here makes sense.
4524  */
4525 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4526                                         struct ocfs2_extent_list *el,
4527                                         struct ocfs2_extent_rec *insert_rec)
4528 {
4529         int i;
4530         u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4531         struct ocfs2_extent_rec *rec;
4532
4533         insert->ins_appending = APPEND_NONE;
4534
4535         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4536
4537         if (!el->l_next_free_rec)
4538                 goto set_tail_append;
4539
4540         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4541                 /* Were all records empty? */
4542                 if (le16_to_cpu(el->l_next_free_rec) == 1)
4543                         goto set_tail_append;
4544         }
4545
4546         i = le16_to_cpu(el->l_next_free_rec) - 1;
4547         rec = &el->l_recs[i];
4548
4549         if (cpos >=
4550             (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4551                 goto set_tail_append;
4552
4553         return;
4554
4555 set_tail_append:
4556         insert->ins_appending = APPEND_TAIL;
4557 }
4558
4559 /*
4560  * Helper function called at the beginning of an insert.
4561  *
4562  * This computes a few things that are commonly used in the process of
4563  * inserting into the btree:
4564  *   - Whether the new extent is contiguous with an existing one.
4565  *   - The current tree depth.
4566  *   - Whether the insert is an appending one.
4567  *   - The total # of free records in the tree.
4568  *
4569  * All of the information is stored on the ocfs2_insert_type
4570  * structure.
4571  */
4572 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4573                                     struct buffer_head **last_eb_bh,
4574                                     struct ocfs2_extent_rec *insert_rec,
4575                                     int *free_records,
4576                                     struct ocfs2_insert_type *insert)
4577 {
4578         int ret;
4579         struct ocfs2_extent_block *eb;
4580         struct ocfs2_extent_list *el;
4581         struct ocfs2_path *path = NULL;
4582         struct buffer_head *bh = NULL;
4583
4584         insert->ins_split = SPLIT_NONE;
4585
4586         el = et->et_root_el;
4587         insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4588
4589         if (el->l_tree_depth) {
4590                 /*
4591                  * If we have tree depth, we read in the
4592                  * rightmost extent block ahead of time as
4593                  * ocfs2_figure_insert_type() and ocfs2_add_branch()
4594                  * may want it later.
4595                  */
4596                 ret = ocfs2_read_extent_block(et->et_ci,
4597                                               ocfs2_et_get_last_eb_blk(et),
4598                                               &bh);
4599                 if (ret) {
4600                         mlog_errno(ret);
4601                         goto out;
4602                 }
4603                 eb = (struct ocfs2_extent_block *) bh->b_data;
4604                 el = &eb->h_list;
4605         }
4606
4607         /*
4608          * Unless we have a contiguous insert, we'll need to know if
4609          * there is room left in our allocation tree for another
4610          * extent record.
4611          *
4612          * XXX: This test is simplistic, we can search for empty
4613          * extent records too.
4614          */
4615         *free_records = le16_to_cpu(el->l_count) -
4616                 le16_to_cpu(el->l_next_free_rec);
4617
4618         if (!insert->ins_tree_depth) {
4619                 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4620                 ocfs2_figure_appending_type(insert, el, insert_rec);
4621                 return 0;
4622         }
4623
4624         path = ocfs2_new_path_from_et(et);
4625         if (!path) {
4626                 ret = -ENOMEM;
4627                 mlog_errno(ret);
4628                 goto out;
4629         }
4630
4631         /*
4632          * In the case that we're inserting past what the tree
4633          * currently accounts for, ocfs2_find_path() will return for
4634          * us the rightmost tree path. This is accounted for below in
4635          * the appending code.
4636          */
4637         ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4638         if (ret) {
4639                 mlog_errno(ret);
4640                 goto out;
4641         }
4642
4643         el = path_leaf_el(path);
4644
4645         /*
4646          * Now that we have the path, there's two things we want to determine:
4647          * 1) Contiguousness (also set contig_index if this is so)
4648          *
4649          * 2) Are we doing an append? We can trivially break this up
4650          *     into two types of appends: simple record append, or a
4651          *     rotate inside the tail leaf.
4652          */
4653         ocfs2_figure_contig_type(et, insert, el, insert_rec);
4654
4655         /*
4656          * The insert code isn't quite ready to deal with all cases of
4657          * left contiguousness. Specifically, if it's an insert into
4658          * the 1st record in a leaf, it will require the adjustment of
4659          * cluster count on the last record of the path directly to it's
4660          * left. For now, just catch that case and fool the layers
4661          * above us. This works just fine for tree_depth == 0, which
4662          * is why we allow that above.
4663          */
4664         if (insert->ins_contig == CONTIG_LEFT &&
4665             insert->ins_contig_index == 0)
4666                 insert->ins_contig = CONTIG_NONE;
4667
4668         /*
4669          * Ok, so we can simply compare against last_eb to figure out
4670          * whether the path doesn't exist. This will only happen in
4671          * the case that we're doing a tail append, so maybe we can
4672          * take advantage of that information somehow.
4673          */
4674         if (ocfs2_et_get_last_eb_blk(et) ==
4675             path_leaf_bh(path)->b_blocknr) {
4676                 /*
4677                  * Ok, ocfs2_find_path() returned us the rightmost
4678                  * tree path. This might be an appending insert. There are
4679                  * two cases:
4680                  *    1) We're doing a true append at the tail:
4681                  *      -This might even be off the end of the leaf
4682                  *    2) We're "appending" by rotating in the tail
4683                  */
4684                 ocfs2_figure_appending_type(insert, el, insert_rec);
4685         }
4686
4687 out:
4688         ocfs2_free_path(path);
4689
4690         if (ret == 0)
4691                 *last_eb_bh = bh;
4692         else
4693                 brelse(bh);
4694         return ret;
4695 }
4696
4697 /*
4698  * Insert an extent into a btree.
4699  *
4700  * The caller needs to update the owning btree's cluster count.
4701  */
4702 int ocfs2_insert_extent(handle_t *handle,
4703                         struct ocfs2_extent_tree *et,
4704                         u32 cpos,
4705                         u64 start_blk,
4706                         u32 new_clusters,
4707                         u8 flags,
4708                         struct ocfs2_alloc_context *meta_ac)
4709 {
4710         int status;
4711         int uninitialized_var(free_records);
4712         struct buffer_head *last_eb_bh = NULL;
4713         struct ocfs2_insert_type insert = {0, };
4714         struct ocfs2_extent_rec rec;
4715
4716         trace_ocfs2_insert_extent_start(
4717                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4718                 cpos, new_clusters);
4719
4720         memset(&rec, 0, sizeof(rec));
4721         rec.e_cpos = cpu_to_le32(cpos);
4722         rec.e_blkno = cpu_to_le64(start_blk);
4723         rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4724         rec.e_flags = flags;
4725         status = ocfs2_et_insert_check(et, &rec);
4726         if (status) {
4727                 mlog_errno(status);
4728                 goto bail;
4729         }
4730
4731         status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4732                                           &free_records, &insert);
4733         if (status < 0) {
4734                 mlog_errno(status);
4735                 goto bail;
4736         }
4737
4738         trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4739                                   insert.ins_contig_index, free_records,
4740                                   insert.ins_tree_depth);
4741
4742         if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4743                 status = ocfs2_grow_tree(handle, et,
4744                                          &insert.ins_tree_depth, &last_eb_bh,
4745                                          meta_ac);
4746                 if (status) {
4747                         mlog_errno(status);
4748                         goto bail;
4749                 }
4750         }
4751
4752         /* Finally, we can add clusters. This might rotate the tree for us. */
4753         status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4754         if (status < 0)
4755                 mlog_errno(status);
4756         else
4757                 ocfs2_et_extent_map_insert(et, &rec);
4758
4759 bail:
4760         brelse(last_eb_bh);
4761
4762         return status;
4763 }
4764
4765 /*
4766  * Allcate and add clusters into the extent b-tree.
4767  * The new clusters(clusters_to_add) will be inserted at logical_offset.
4768  * The extent b-tree's root is specified by et, and
4769  * it is not limited to the file storage. Any extent tree can use this
4770  * function if it implements the proper ocfs2_extent_tree.
4771  */
4772 int ocfs2_add_clusters_in_btree(handle_t *handle,
4773                                 struct ocfs2_extent_tree *et,
4774                                 u32 *logical_offset,
4775                                 u32 clusters_to_add,
4776                                 int mark_unwritten,
4777                                 struct ocfs2_alloc_context *data_ac,
4778                                 struct ocfs2_alloc_context *meta_ac,
4779                                 enum ocfs2_alloc_restarted *reason_ret)
4780 {
4781         int status = 0, err = 0;
4782         int need_free = 0;
4783         int free_extents;
4784         enum ocfs2_alloc_restarted reason = RESTART_NONE;
4785         u32 bit_off, num_bits;
4786         u64 block;
4787         u8 flags = 0;
4788         struct ocfs2_super *osb =
4789                 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4790
4791         BUG_ON(!clusters_to_add);
4792
4793         if (mark_unwritten)
4794                 flags = OCFS2_EXT_UNWRITTEN;
4795
4796         free_extents = ocfs2_num_free_extents(et);
4797         if (free_extents < 0) {
4798                 status = free_extents;
4799                 mlog_errno(status);
4800                 goto leave;
4801         }
4802
4803         /* there are two cases which could cause us to EAGAIN in the
4804          * we-need-more-metadata case:
4805          * 1) we haven't reserved *any*
4806          * 2) we are so fragmented, we've needed to add metadata too
4807          *    many times. */
4808         if (!free_extents && !meta_ac) {
4809                 err = -1;
4810                 status = -EAGAIN;
4811                 reason = RESTART_META;
4812                 goto leave;
4813         } else if ((!free_extents)
4814                    && (ocfs2_alloc_context_bits_left(meta_ac)
4815                        < ocfs2_extend_meta_needed(et->et_root_el))) {
4816                 err = -2;
4817                 status = -EAGAIN;
4818                 reason = RESTART_META;
4819                 goto leave;
4820         }
4821
4822         status = __ocfs2_claim_clusters(handle, data_ac, 1,
4823                                         clusters_to_add, &bit_off, &num_bits);
4824         if (status < 0) {
4825                 if (status != -ENOSPC)
4826                         mlog_errno(status);
4827                 goto leave;
4828         }
4829
4830         BUG_ON(num_bits > clusters_to_add);
4831
4832         /* reserve our write early -- insert_extent may update the tree root */
4833         status = ocfs2_et_root_journal_access(handle, et,
4834                                               OCFS2_JOURNAL_ACCESS_WRITE);
4835         if (status < 0) {
4836                 mlog_errno(status);
4837                 need_free = 1;
4838                 goto bail;
4839         }
4840
4841         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4842         trace_ocfs2_add_clusters_in_btree(
4843              (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4844              bit_off, num_bits);
4845         status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4846                                      num_bits, flags, meta_ac);
4847         if (status < 0) {
4848                 mlog_errno(status);
4849                 need_free = 1;
4850                 goto bail;
4851         }
4852
4853         ocfs2_journal_dirty(handle, et->et_root_bh);
4854
4855         clusters_to_add -= num_bits;
4856         *logical_offset += num_bits;
4857
4858         if (clusters_to_add) {
4859                 err = clusters_to_add;
4860                 status = -EAGAIN;
4861                 reason = RESTART_TRANS;
4862         }
4863
4864 bail:
4865         if (need_free) {
4866                 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4867                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4868                                         bit_off, num_bits);
4869                 else
4870                         ocfs2_free_clusters(handle,
4871                                         data_ac->ac_inode,
4872                                         data_ac->ac_bh,
4873                                         ocfs2_clusters_to_blocks(osb->sb, bit_off),
4874                                         num_bits);
4875         }
4876
4877 leave:
4878         if (reason_ret)
4879                 *reason_ret = reason;
4880         trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4881         return status;
4882 }
4883
4884 static void ocfs2_make_right_split_rec(struct super_block *sb,
4885                                        struct ocfs2_extent_rec *split_rec,
4886                                        u32 cpos,
4887                                        struct ocfs2_extent_rec *rec)
4888 {
4889         u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4890         u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4891
4892         memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4893
4894         split_rec->e_cpos = cpu_to_le32(cpos);
4895         split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4896
4897         split_rec->e_blkno = rec->e_blkno;
4898         le64_add_cpu(&split_rec->e_blkno,
4899                      ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4900
4901         split_rec->e_flags = rec->e_flags;
4902 }
4903
4904 static int ocfs2_split_and_insert(handle_t *handle,
4905                                   struct ocfs2_extent_tree *et,
4906                                   struct ocfs2_path *path,
4907                                   struct buffer_head **last_eb_bh,
4908                                   int split_index,
4909                                   struct ocfs2_extent_rec *orig_split_rec,
4910                                   struct ocfs2_alloc_context *meta_ac)
4911 {
4912         int ret = 0, depth;
4913         unsigned int insert_range, rec_range, do_leftright = 0;
4914         struct ocfs2_extent_rec tmprec;
4915         struct ocfs2_extent_list *rightmost_el;
4916         struct ocfs2_extent_rec rec;
4917         struct ocfs2_extent_rec split_rec = *orig_split_rec;
4918         struct ocfs2_insert_type insert;
4919         struct ocfs2_extent_block *eb;
4920
4921 leftright:
4922         /*
4923          * Store a copy of the record on the stack - it might move
4924          * around as the tree is manipulated below.
4925          */
4926         rec = path_leaf_el(path)->l_recs[split_index];
4927
4928         rightmost_el = et->et_root_el;
4929
4930         depth = le16_to_cpu(rightmost_el->l_tree_depth);
4931         if (depth) {
4932                 BUG_ON(!(*last_eb_bh));
4933                 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4934                 rightmost_el = &eb->h_list;
4935         }
4936
4937         if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4938             le16_to_cpu(rightmost_el->l_count)) {
4939                 ret = ocfs2_grow_tree(handle, et,
4940                                       &depth, last_eb_bh, meta_ac);
4941                 if (ret) {
4942                         mlog_errno(ret);
4943                         goto out;
4944                 }
4945         }
4946
4947         memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4948         insert.ins_appending = APPEND_NONE;
4949         insert.ins_contig = CONTIG_NONE;
4950         insert.ins_tree_depth = depth;
4951
4952         insert_range = le32_to_cpu(split_rec.e_cpos) +
4953                 le16_to_cpu(split_rec.e_leaf_clusters);
4954         rec_range = le32_to_cpu(rec.e_cpos) +
4955                 le16_to_cpu(rec.e_leaf_clusters);
4956
4957         if (split_rec.e_cpos == rec.e_cpos) {
4958                 insert.ins_split = SPLIT_LEFT;
4959         } else if (insert_range == rec_range) {
4960                 insert.ins_split = SPLIT_RIGHT;
4961         } else {
4962                 /*
4963                  * Left/right split. We fake this as a right split
4964                  * first and then make a second pass as a left split.
4965                  */
4966                 insert.ins_split = SPLIT_RIGHT;
4967
4968                 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4969                                            &tmprec, insert_range, &rec);
4970
4971                 split_rec = tmprec;
4972
4973                 BUG_ON(do_leftright);
4974                 do_leftright = 1;
4975         }
4976
4977         ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4978         if (ret) {
4979                 mlog_errno(ret);
4980                 goto out;
4981         }
4982
4983         if (do_leftright == 1) {
4984                 u32 cpos;
4985                 struct ocfs2_extent_list *el;
4986
4987                 do_leftright++;
4988                 split_rec = *orig_split_rec;
4989
4990                 ocfs2_reinit_path(path, 1);
4991
4992                 cpos = le32_to_cpu(split_rec.e_cpos);
4993                 ret = ocfs2_find_path(et->et_ci, path, cpos);
4994                 if (ret) {
4995                         mlog_errno(ret);
4996                         goto out;
4997                 }
4998
4999                 el = path_leaf_el(path);
5000                 split_index = ocfs2_search_extent_list(el, cpos);
5001                 if (split_index == -1) {
5002                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5003                                     "Owner %llu has an extent at cpos %u which can no longer be found\n",
5004                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5005                                     cpos);
5006                         ret = -EROFS;
5007                         goto out;
5008                 }
5009                 goto leftright;
5010         }
5011 out:
5012
5013         return ret;
5014 }
5015
5016 static int ocfs2_replace_extent_rec(handle_t *handle,
5017                                     struct ocfs2_extent_tree *et,
5018                                     struct ocfs2_path *path,
5019                                     struct ocfs2_extent_list *el,
5020                                     int split_index,
5021                                     struct ocfs2_extent_rec *split_rec)
5022 {
5023         int ret;
5024
5025         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5026                                            path_num_items(path) - 1);
5027         if (ret) {
5028                 mlog_errno(ret);
5029                 goto out;
5030         }
5031
5032         el->l_recs[split_index] = *split_rec;
5033
5034         ocfs2_journal_dirty(handle, path_leaf_bh(path));
5035 out:
5036         return ret;
5037 }
5038
5039 /*
5040  * Split part or all of the extent record at split_index in the leaf
5041  * pointed to by path. Merge with the contiguous extent record if needed.
5042  *
5043  * Care is taken to handle contiguousness so as to not grow the tree.
5044  *
5045  * meta_ac is not strictly necessary - we only truly need it if growth
5046  * of the tree is required. All other cases will degrade into a less
5047  * optimal tree layout.
5048  *
5049  * last_eb_bh should be the rightmost leaf block for any extent
5050  * btree. Since a split may grow the tree or a merge might shrink it,
5051  * the caller cannot trust the contents of that buffer after this call.
5052  *
5053  * This code is optimized for readability - several passes might be
5054  * made over certain portions of the tree. All of those blocks will
5055  * have been brought into cache (and pinned via the journal), so the
5056  * extra overhead is not expressed in terms of disk reads.
5057  */
5058 int ocfs2_split_extent(handle_t *handle,
5059                        struct ocfs2_extent_tree *et,
5060                        struct ocfs2_path *path,
5061                        int split_index,
5062                        struct ocfs2_extent_rec *split_rec,
5063                        struct ocfs2_alloc_context *meta_ac,
5064                        struct ocfs2_cached_dealloc_ctxt *dealloc)
5065 {
5066         int ret = 0;
5067         struct ocfs2_extent_list *el = path_leaf_el(path);
5068         struct buffer_head *last_eb_bh = NULL;
5069         struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5070         struct ocfs2_merge_ctxt ctxt;
5071
5072         if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5073             ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5074              (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5075                 ret = -EIO;
5076                 mlog_errno(ret);
5077                 goto out;
5078         }
5079
5080         ret = ocfs2_figure_merge_contig_type(et, path, el,
5081                                              split_index,
5082                                              split_rec,
5083                                              &ctxt);
5084         if (ret) {
5085                 mlog_errno(ret);
5086                 goto out;
5087         }
5088
5089         /*
5090          * The core merge / split code wants to know how much room is
5091          * left in this allocation tree, so we pass the
5092          * rightmost extent list.
5093          */
5094         if (path->p_tree_depth) {
5095                 ret = ocfs2_read_extent_block(et->et_ci,
5096                                               ocfs2_et_get_last_eb_blk(et),
5097                                               &last_eb_bh);
5098                 if (ret) {
5099                         mlog_errno(ret);
5100                         goto out;
5101                 }
5102         }
5103
5104         if (rec->e_cpos == split_rec->e_cpos &&
5105             rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5106                 ctxt.c_split_covers_rec = 1;
5107         else
5108                 ctxt.c_split_covers_rec = 0;
5109
5110         ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5111
5112         trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5113                                  ctxt.c_has_empty_extent,
5114                                  ctxt.c_split_covers_rec);
5115
5116         if (ctxt.c_contig_type == CONTIG_NONE) {
5117                 if (ctxt.c_split_covers_rec)
5118                         ret = ocfs2_replace_extent_rec(handle, et, path, el,
5119                                                        split_index, split_rec);
5120                 else
5121                         ret = ocfs2_split_and_insert(handle, et, path,
5122                                                      &last_eb_bh, split_index,
5123                                                      split_rec, meta_ac);
5124                 if (ret)
5125                         mlog_errno(ret);
5126         } else {
5127                 ret = ocfs2_try_to_merge_extent(handle, et, path,
5128                                                 split_index, split_rec,
5129                                                 dealloc, &ctxt);
5130                 if (ret)
5131                         mlog_errno(ret);
5132         }
5133
5134 out:
5135         brelse(last_eb_bh);
5136         return ret;
5137 }
5138
5139 /*
5140  * Change the flags of the already-existing extent at cpos for len clusters.
5141  *
5142  * new_flags: the flags we want to set.
5143  * clear_flags: the flags we want to clear.
5144  * phys: the new physical offset we want this new extent starts from.
5145  *
5146  * If the existing extent is larger than the request, initiate a
5147  * split. An attempt will be made at merging with adjacent extents.
5148  *
5149  * The caller is responsible for passing down meta_ac if we'll need it.
5150  */
5151 int ocfs2_change_extent_flag(handle_t *handle,
5152                              struct ocfs2_extent_tree *et,
5153                              u32 cpos, u32 len, u32 phys,
5154                              struct ocfs2_alloc_context *meta_ac,
5155                              struct ocfs2_cached_dealloc_ctxt *dealloc,
5156                              int new_flags, int clear_flags)
5157 {
5158         int ret, index;
5159         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5160         u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5161         struct ocfs2_extent_rec split_rec;
5162         struct ocfs2_path *left_path = NULL;
5163         struct ocfs2_extent_list *el;
5164         struct ocfs2_extent_rec *rec;
5165
5166         left_path = ocfs2_new_path_from_et(et);
5167         if (!left_path) {
5168                 ret = -ENOMEM;
5169                 mlog_errno(ret);
5170                 goto out;
5171         }
5172
5173         ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5174         if (ret) {
5175                 mlog_errno(ret);
5176                 goto out;
5177         }
5178         el = path_leaf_el(left_path);
5179
5180         index = ocfs2_search_extent_list(el, cpos);
5181         if (index == -1) {
5182                 ocfs2_error(sb,
5183                             "Owner %llu has an extent at cpos %u which can no longer be found\n",
5184                             (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5185                             cpos);
5186                 ret = -EROFS;
5187                 goto out;
5188         }
5189
5190         ret = -EIO;
5191         rec = &el->l_recs[index];
5192         if (new_flags && (rec->e_flags & new_flags)) {
5193                 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5194                      "extent that already had them\n",
5195                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5196                      new_flags);
5197                 goto out;
5198         }
5199
5200         if (clear_flags && !(rec->e_flags & clear_flags)) {
5201                 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5202                      "extent that didn't have them\n",
5203                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5204                      clear_flags);
5205                 goto out;
5206         }
5207
5208         memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5209         split_rec.e_cpos = cpu_to_le32(cpos);
5210         split_rec.e_leaf_clusters = cpu_to_le16(len);
5211         split_rec.e_blkno = cpu_to_le64(start_blkno);
5212         split_rec.e_flags = rec->e_flags;
5213         if (new_flags)
5214                 split_rec.e_flags |= new_flags;
5215         if (clear_flags)
5216                 split_rec.e_flags &= ~clear_flags;
5217
5218         ret = ocfs2_split_extent(handle, et, left_path,
5219                                  index, &split_rec, meta_ac,
5220                                  dealloc);
5221         if (ret)
5222                 mlog_errno(ret);
5223
5224 out:
5225         ocfs2_free_path(left_path);
5226         return ret;
5227
5228 }
5229
5230 /*
5231  * Mark the already-existing extent at cpos as written for len clusters.
5232  * This removes the unwritten extent flag.
5233  *
5234  * If the existing extent is larger than the request, initiate a
5235  * split. An attempt will be made at merging with adjacent extents.
5236  *
5237  * The caller is responsible for passing down meta_ac if we'll need it.
5238  */
5239 int ocfs2_mark_extent_written(struct inode *inode,
5240                               struct ocfs2_extent_tree *et,
5241                               handle_t *handle, u32 cpos, u32 len, u32 phys,
5242                               struct ocfs2_alloc_context *meta_ac,
5243                               struct ocfs2_cached_dealloc_ctxt *dealloc)
5244 {
5245         int ret;
5246
5247         trace_ocfs2_mark_extent_written(
5248                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5249                 cpos, len, phys);
5250
5251         if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5252                 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents that are being written to, but the feature bit is not set in the super block\n",
5253                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
5254                 ret = -EROFS;
5255                 goto out;
5256         }
5257
5258         /*
5259          * XXX: This should be fixed up so that we just re-insert the
5260          * next extent records.
5261          */
5262         ocfs2_et_extent_map_truncate(et, 0);
5263
5264         ret = ocfs2_change_extent_flag(handle, et, cpos,
5265                                        len, phys, meta_ac, dealloc,
5266                                        0, OCFS2_EXT_UNWRITTEN);
5267         if (ret)
5268                 mlog_errno(ret);
5269
5270 out:
5271         return ret;
5272 }
5273
5274 static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5275                             struct ocfs2_path *path,
5276                             int index, u32 new_range,
5277                             struct ocfs2_alloc_context *meta_ac)
5278 {
5279         int ret, depth, credits;
5280         struct buffer_head *last_eb_bh = NULL;
5281         struct ocfs2_extent_block *eb;
5282         struct ocfs2_extent_list *rightmost_el, *el;
5283         struct ocfs2_extent_rec split_rec;
5284         struct ocfs2_extent_rec *rec;
5285         struct ocfs2_insert_type insert;
5286
5287         /*
5288          * Setup the record to split before we grow the tree.
5289          */
5290         el = path_leaf_el(path);
5291         rec = &el->l_recs[index];
5292         ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5293                                    &split_rec, new_range, rec);
5294
5295         depth = path->p_tree_depth;
5296         if (depth > 0) {
5297                 ret = ocfs2_read_extent_block(et->et_ci,
5298                                               ocfs2_et_get_last_eb_blk(et),
5299                                               &last_eb_bh);
5300                 if (ret < 0) {
5301                         mlog_errno(ret);
5302                         goto out;
5303                 }
5304
5305                 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5306                 rightmost_el = &eb->h_list;
5307         } else
5308                 rightmost_el = path_leaf_el(path);
5309
5310         credits = path->p_tree_depth +
5311                   ocfs2_extend_meta_needed(et->et_root_el);
5312         ret = ocfs2_extend_trans(handle, credits);
5313         if (ret) {
5314                 mlog_errno(ret);
5315                 goto out;
5316         }
5317
5318         if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5319             le16_to_cpu(rightmost_el->l_count)) {
5320                 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5321                                       meta_ac);
5322                 if (ret) {
5323                         mlog_errno(ret);
5324                         goto out;
5325                 }
5326         }
5327
5328         memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5329         insert.ins_appending = APPEND_NONE;
5330         insert.ins_contig = CONTIG_NONE;
5331         insert.ins_split = SPLIT_RIGHT;
5332         insert.ins_tree_depth = depth;
5333
5334         ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5335         if (ret)
5336                 mlog_errno(ret);
5337
5338 out:
5339         brelse(last_eb_bh);
5340         return ret;
5341 }
5342
5343 static int ocfs2_truncate_rec(handle_t *handle,
5344                               struct ocfs2_extent_tree *et,
5345                               struct ocfs2_path *path, int index,
5346                               struct ocfs2_cached_dealloc_ctxt *dealloc,
5347                               u32 cpos, u32 len)
5348 {
5349         int ret;
5350         u32 left_cpos, rec_range, trunc_range;
5351         int is_rightmost_tree_rec = 0;
5352         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5353         struct ocfs2_path *left_path = NULL;
5354         struct ocfs2_extent_list *el = path_leaf_el(path);
5355         struct ocfs2_extent_rec *rec;
5356         struct ocfs2_extent_block *eb;
5357
5358         if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5359                 /* extend credit for ocfs2_remove_rightmost_path */
5360                 ret = ocfs2_extend_rotate_transaction(handle, 0,
5361                                 handle->h_buffer_credits,
5362                                 path);
5363                 if (ret) {
5364                         mlog_errno(ret);
5365                         goto out;
5366                 }
5367
5368                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5369                 if (ret) {
5370                         mlog_errno(ret);
5371                         goto out;
5372                 }
5373
5374                 index--;
5375         }
5376
5377         if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5378             path->p_tree_depth) {
5379                 /*
5380                  * Check whether this is the rightmost tree record. If
5381                  * we remove all of this record or part of its right
5382                  * edge then an update of the record lengths above it
5383                  * will be required.
5384                  */
5385                 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5386                 if (eb->h_next_leaf_blk == 0)
5387                         is_rightmost_tree_rec = 1;
5388         }
5389
5390         rec = &el->l_recs[index];
5391         if (index == 0 && path->p_tree_depth &&
5392             le32_to_cpu(rec->e_cpos) == cpos) {
5393                 /*
5394                  * Changing the leftmost offset (via partial or whole
5395                  * record truncate) of an interior (or rightmost) path
5396                  * means we have to update the subtree that is formed
5397                  * by this leaf and the one to it's left.
5398                  *
5399                  * There are two cases we can skip:
5400                  *   1) Path is the leftmost one in our btree.
5401                  *   2) The leaf is rightmost and will be empty after
5402                  *      we remove the extent record - the rotate code
5403                  *      knows how to update the newly formed edge.
5404                  */
5405
5406                 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5407                 if (ret) {
5408                         mlog_errno(ret);
5409                         goto out;
5410                 }
5411
5412                 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5413                         left_path = ocfs2_new_path_from_path(path);
5414                         if (!left_path) {
5415                                 ret = -ENOMEM;
5416                                 mlog_errno(ret);
5417                                 goto out;
5418                         }
5419
5420                         ret = ocfs2_find_path(et->et_ci, left_path,
5421                                               left_cpos);
5422                         if (ret) {
5423                                 mlog_errno(ret);
5424                                 goto out;
5425                         }
5426                 }
5427         }
5428
5429         ret = ocfs2_extend_rotate_transaction(handle, 0,
5430                                               handle->h_buffer_credits,
5431                                               path);
5432         if (ret) {
5433                 mlog_errno(ret);
5434                 goto out;
5435         }
5436
5437         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5438         if (ret) {
5439                 mlog_errno(ret);
5440                 goto out;
5441         }
5442
5443         ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5444         if (ret) {
5445                 mlog_errno(ret);
5446                 goto out;
5447         }
5448
5449         rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5450         trunc_range = cpos + len;
5451
5452         if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5453                 int next_free;
5454
5455                 memset(rec, 0, sizeof(*rec));
5456                 ocfs2_cleanup_merge(el, index);
5457
5458                 next_free = le16_to_cpu(el->l_next_free_rec);
5459                 if (is_rightmost_tree_rec && next_free > 1) {
5460                         /*
5461                          * We skip the edge update if this path will
5462                          * be deleted by the rotate code.
5463                          */
5464                         rec = &el->l_recs[next_free - 1];
5465                         ocfs2_adjust_rightmost_records(handle, et, path,
5466                                                        rec);
5467                 }
5468         } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5469                 /* Remove leftmost portion of the record. */
5470                 le32_add_cpu(&rec->e_cpos, len);
5471                 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5472                 le16_add_cpu(&rec->e_leaf_clusters, -len);
5473         } else if (rec_range == trunc_range) {
5474                 /* Remove rightmost portion of the record */
5475                 le16_add_cpu(&rec->e_leaf_clusters, -len);
5476                 if (is_rightmost_tree_rec)
5477                         ocfs2_adjust_rightmost_records(handle, et, path, rec);
5478         } else {
5479                 /* Caller should have trapped this. */
5480                 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5481                      "(%u, %u)\n",
5482                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5483                      le32_to_cpu(rec->e_cpos),
5484                      le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5485                 BUG();
5486         }
5487
5488         if (left_path) {
5489                 int subtree_index;
5490
5491                 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5492                 ocfs2_complete_edge_insert(handle, left_path, path,
5493                                            subtree_index);
5494         }
5495
5496         ocfs2_journal_dirty(handle, path_leaf_bh(path));
5497
5498         ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5499         if (ret)
5500                 mlog_errno(ret);
5501
5502 out:
5503         ocfs2_free_path(left_path);
5504         return ret;
5505 }
5506
5507 int ocfs2_remove_extent(handle_t *handle,
5508                         struct ocfs2_extent_tree *et,
5509                         u32 cpos, u32 len,
5510                         struct ocfs2_alloc_context *meta_ac,
5511                         struct ocfs2_cached_dealloc_ctxt *dealloc)
5512 {
5513         int ret, index;
5514         u32 rec_range, trunc_range;
5515         struct ocfs2_extent_rec *rec;
5516         struct ocfs2_extent_list *el;
5517         struct ocfs2_path *path = NULL;
5518
5519         /*
5520          * XXX: Why are we truncating to 0 instead of wherever this
5521          * affects us?
5522          */
5523         ocfs2_et_extent_map_truncate(et, 0);
5524
5525         path = ocfs2_new_path_from_et(et);
5526         if (!path) {
5527                 ret = -ENOMEM;
5528                 mlog_errno(ret);
5529                 goto out;
5530         }
5531
5532         ret = ocfs2_find_path(et->et_ci, path, cpos);
5533         if (ret) {
5534                 mlog_errno(ret);
5535                 goto out;
5536         }
5537
5538         el = path_leaf_el(path);
5539         index = ocfs2_search_extent_list(el, cpos);
5540         if (index == -1) {
5541                 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5542                             "Owner %llu has an extent at cpos %u which can no longer be found\n",
5543                             (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5544                             cpos);
5545                 ret = -EROFS;
5546                 goto out;
5547         }
5548
5549         /*
5550          * We have 3 cases of extent removal:
5551          *   1) Range covers the entire extent rec
5552          *   2) Range begins or ends on one edge of the extent rec
5553          *   3) Range is in the middle of the extent rec (no shared edges)
5554          *
5555          * For case 1 we remove the extent rec and left rotate to
5556          * fill the hole.
5557          *
5558          * For case 2 we just shrink the existing extent rec, with a
5559          * tree update if the shrinking edge is also the edge of an
5560          * extent block.
5561          *
5562          * For case 3 we do a right split to turn the extent rec into
5563          * something case 2 can handle.
5564          */
5565         rec = &el->l_recs[index];
5566         rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5567         trunc_range = cpos + len;
5568
5569         BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5570
5571         trace_ocfs2_remove_extent(
5572                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5573                 cpos, len, index, le32_to_cpu(rec->e_cpos),
5574                 ocfs2_rec_clusters(el, rec));
5575
5576         if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5577                 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5578                                          cpos, len);
5579                 if (ret) {
5580                         mlog_errno(ret);
5581                         goto out;
5582                 }
5583         } else {
5584                 ret = ocfs2_split_tree(handle, et, path, index,
5585                                        trunc_range, meta_ac);
5586                 if (ret) {
5587                         mlog_errno(ret);
5588                         goto out;
5589                 }
5590
5591                 /*
5592                  * The split could have manipulated the tree enough to
5593                  * move the record location, so we have to look for it again.
5594                  */
5595                 ocfs2_reinit_path(path, 1);
5596
5597                 ret = ocfs2_find_path(et->et_ci, path, cpos);
5598                 if (ret) {
5599                         mlog_errno(ret);
5600                         goto out;
5601                 }
5602
5603                 el = path_leaf_el(path);
5604                 index = ocfs2_search_extent_list(el, cpos);
5605                 if (index == -1) {
5606                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5607                                     "Owner %llu: split at cpos %u lost record\n",
5608                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5609                                     cpos);
5610                         ret = -EROFS;
5611                         goto out;
5612                 }
5613
5614                 /*
5615                  * Double check our values here. If anything is fishy,
5616                  * it's easier to catch it at the top level.
5617                  */
5618                 rec = &el->l_recs[index];
5619                 rec_range = le32_to_cpu(rec->e_cpos) +
5620                         ocfs2_rec_clusters(el, rec);
5621                 if (rec_range != trunc_range) {
5622                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5623                                     "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5624                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5625                                     cpos, len, le32_to_cpu(rec->e_cpos),
5626                                     ocfs2_rec_clusters(el, rec));
5627                         ret = -EROFS;
5628                         goto out;
5629                 }
5630
5631                 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5632                                          cpos, len);
5633                 if (ret)
5634                         mlog_errno(ret);
5635         }
5636
5637 out:
5638         ocfs2_free_path(path);
5639         return ret;
5640 }
5641
5642 /*
5643  * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5644  * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5645  * number to reserve some extra blocks, and it only handles meta
5646  * data allocations.
5647  *
5648  * Currently, only ocfs2_remove_btree_range() uses it for truncating
5649  * and punching holes.
5650  */
5651 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5652                                               struct ocfs2_extent_tree *et,
5653                                               u32 extents_to_split,
5654                                               struct ocfs2_alloc_context **ac,
5655                                               int extra_blocks)
5656 {
5657         int ret = 0, num_free_extents;
5658         unsigned int max_recs_needed = 2 * extents_to_split;
5659         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5660
5661         *ac = NULL;
5662
5663         num_free_extents = ocfs2_num_free_extents(et);
5664         if (num_free_extents < 0) {
5665                 ret = num_free_extents;
5666                 mlog_errno(ret);
5667                 goto out;
5668         }
5669
5670         if (!num_free_extents ||
5671             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5672                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5673
5674         if (extra_blocks) {
5675                 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5676                 if (ret < 0) {
5677                         if (ret != -ENOSPC)
5678                                 mlog_errno(ret);
5679                 }
5680         }
5681
5682 out:
5683         if (ret) {
5684                 if (*ac) {
5685                         ocfs2_free_alloc_context(*ac);
5686                         *ac = NULL;
5687                 }
5688         }
5689
5690         return ret;
5691 }
5692
5693 int ocfs2_remove_btree_range(struct inode *inode,
5694                              struct ocfs2_extent_tree *et,
5695                              u32 cpos, u32 phys_cpos, u32 len, int flags,
5696                              struct ocfs2_cached_dealloc_ctxt *dealloc,
5697                              u64 refcount_loc, bool refcount_tree_locked)
5698 {
5699         int ret, credits = 0, extra_blocks = 0;
5700         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5701         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5702         struct inode *tl_inode = osb->osb_tl_inode;
5703         handle_t *handle;
5704         struct ocfs2_alloc_context *meta_ac = NULL;
5705         struct ocfs2_refcount_tree *ref_tree = NULL;
5706
5707         if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5708                 BUG_ON(!ocfs2_is_refcount_inode(inode));
5709
5710                 if (!refcount_tree_locked) {
5711                         ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5712                                                        &ref_tree, NULL);
5713                         if (ret) {
5714                                 mlog_errno(ret);
5715                                 goto bail;
5716                         }
5717                 }
5718
5719                 ret = ocfs2_prepare_refcount_change_for_del(inode,
5720                                                             refcount_loc,
5721                                                             phys_blkno,
5722                                                             len,
5723                                                             &credits,
5724                                                             &extra_blocks);
5725                 if (ret < 0) {
5726                         mlog_errno(ret);
5727                         goto bail;
5728                 }
5729         }
5730
5731         ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5732                                                  extra_blocks);
5733         if (ret) {
5734                 mlog_errno(ret);
5735                 goto bail;
5736         }
5737
5738         inode_lock(tl_inode);
5739
5740         if (ocfs2_truncate_log_needs_flush(osb)) {
5741                 ret = __ocfs2_flush_truncate_log(osb);
5742                 if (ret < 0) {
5743                         mlog_errno(ret);
5744                         goto out;
5745                 }
5746         }
5747
5748         handle = ocfs2_start_trans(osb,
5749                         ocfs2_remove_extent_credits(osb->sb) + credits);
5750         if (IS_ERR(handle)) {
5751                 ret = PTR_ERR(handle);
5752                 mlog_errno(ret);
5753                 goto out;
5754         }
5755
5756         ret = ocfs2_et_root_journal_access(handle, et,
5757                                            OCFS2_JOURNAL_ACCESS_WRITE);
5758         if (ret) {
5759                 mlog_errno(ret);
5760                 goto out_commit;
5761         }
5762
5763         dquot_free_space_nodirty(inode,
5764                                   ocfs2_clusters_to_bytes(inode->i_sb, len));
5765
5766         ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5767         if (ret) {
5768                 mlog_errno(ret);
5769                 goto out_commit;
5770         }
5771
5772         ocfs2_et_update_clusters(et, -len);
5773         ocfs2_update_inode_fsync_trans(handle, inode, 1);
5774
5775         ocfs2_journal_dirty(handle, et->et_root_bh);
5776
5777         if (phys_blkno) {
5778                 if (flags & OCFS2_EXT_REFCOUNTED)
5779                         ret = ocfs2_decrease_refcount(inode, handle,
5780                                         ocfs2_blocks_to_clusters(osb->sb,
5781                                                                  phys_blkno),
5782                                         len, meta_ac,
5783                                         dealloc, 1);
5784                 else
5785                         ret = ocfs2_truncate_log_append(osb, handle,
5786                                                         phys_blkno, len);
5787                 if (ret)
5788                         mlog_errno(ret);
5789
5790         }
5791
5792 out_commit:
5793         ocfs2_commit_trans(osb, handle);
5794 out:
5795         inode_unlock(tl_inode);
5796 bail:
5797         if (meta_ac)
5798                 ocfs2_free_alloc_context(meta_ac);
5799
5800         if (ref_tree)
5801                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5802
5803         return ret;
5804 }
5805
5806 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5807 {
5808         struct buffer_head *tl_bh = osb->osb_tl_bh;
5809         struct ocfs2_dinode *di;
5810         struct ocfs2_truncate_log *tl;
5811
5812         di = (struct ocfs2_dinode *) tl_bh->b_data;
5813         tl = &di->id2.i_dealloc;
5814
5815         mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5816                         "slot %d, invalid truncate log parameters: used = "
5817                         "%u, count = %u\n", osb->slot_num,
5818                         le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5819         return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5820 }
5821
5822 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5823                                            unsigned int new_start)
5824 {
5825         unsigned int tail_index;
5826         unsigned int current_tail;
5827
5828         /* No records, nothing to coalesce */
5829         if (!le16_to_cpu(tl->tl_used))
5830                 return 0;
5831
5832         tail_index = le16_to_cpu(tl->tl_used) - 1;
5833         current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5834         current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5835
5836         return current_tail == new_start;
5837 }
5838
5839 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5840                               handle_t *handle,
5841                               u64 start_blk,
5842                               unsigned int num_clusters)
5843 {
5844         int status, index;
5845         unsigned int start_cluster, tl_count;
5846         struct inode *tl_inode = osb->osb_tl_inode;
5847         struct buffer_head *tl_bh = osb->osb_tl_bh;
5848         struct ocfs2_dinode *di;
5849         struct ocfs2_truncate_log *tl;
5850
5851         BUG_ON(inode_trylock(tl_inode));
5852
5853         start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5854
5855         di = (struct ocfs2_dinode *) tl_bh->b_data;
5856
5857         /* tl_bh is loaded from ocfs2_truncate_log_init().  It's validated
5858          * by the underlying call to ocfs2_read_inode_block(), so any
5859          * corruption is a code bug */
5860         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5861
5862         tl = &di->id2.i_dealloc;
5863         tl_count = le16_to_cpu(tl->tl_count);
5864         mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5865                         tl_count == 0,
5866                         "Truncate record count on #%llu invalid "
5867                         "wanted %u, actual %u\n",
5868                         (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5869                         ocfs2_truncate_recs_per_inode(osb->sb),
5870                         le16_to_cpu(tl->tl_count));
5871
5872         /* Caller should have known to flush before calling us. */
5873         index = le16_to_cpu(tl->tl_used);
5874         if (index >= tl_count) {
5875                 status = -ENOSPC;
5876                 mlog_errno(status);
5877                 goto bail;
5878         }
5879
5880         status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5881                                          OCFS2_JOURNAL_ACCESS_WRITE);
5882         if (status < 0) {
5883                 mlog_errno(status);
5884                 goto bail;
5885         }
5886
5887         trace_ocfs2_truncate_log_append(
5888                 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5889                 start_cluster, num_clusters);
5890         if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5891                 /*
5892                  * Move index back to the record we are coalescing with.
5893                  * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5894                  */
5895                 index--;
5896
5897                 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5898                 trace_ocfs2_truncate_log_append(
5899                         (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5900                         index, le32_to_cpu(tl->tl_recs[index].t_start),
5901                         num_clusters);
5902         } else {
5903                 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5904                 tl->tl_used = cpu_to_le16(index + 1);
5905         }
5906         tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5907
5908         ocfs2_journal_dirty(handle, tl_bh);
5909
5910         osb->truncated_clusters += num_clusters;
5911 bail:
5912         return status;
5913 }
5914
5915 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5916                                          struct inode *data_alloc_inode,
5917                                          struct buffer_head *data_alloc_bh)
5918 {
5919         int status = 0;
5920         int i;
5921         unsigned int num_clusters;
5922         u64 start_blk;
5923         struct ocfs2_truncate_rec rec;
5924         struct ocfs2_dinode *di;
5925         struct ocfs2_truncate_log *tl;
5926         struct inode *tl_inode = osb->osb_tl_inode;
5927         struct buffer_head *tl_bh = osb->osb_tl_bh;
5928         handle_t *handle;
5929
5930         di = (struct ocfs2_dinode *) tl_bh->b_data;
5931         tl = &di->id2.i_dealloc;
5932         i = le16_to_cpu(tl->tl_used) - 1;
5933         while (i >= 0) {
5934                 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5935                 if (IS_ERR(handle)) {
5936                         status = PTR_ERR(handle);
5937                         mlog_errno(status);
5938                         goto bail;
5939                 }
5940
5941                 /* Caller has given us at least enough credits to
5942                  * update the truncate log dinode */
5943                 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5944                                                  OCFS2_JOURNAL_ACCESS_WRITE);
5945                 if (status < 0) {
5946                         mlog_errno(status);
5947                         goto bail;
5948                 }
5949
5950                 tl->tl_used = cpu_to_le16(i);
5951
5952                 ocfs2_journal_dirty(handle, tl_bh);
5953
5954                 rec = tl->tl_recs[i];
5955                 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5956                                                     le32_to_cpu(rec.t_start));
5957                 num_clusters = le32_to_cpu(rec.t_clusters);
5958
5959                 /* if start_blk is not set, we ignore the record as
5960                  * invalid. */
5961                 if (start_blk) {
5962                         trace_ocfs2_replay_truncate_records(
5963                                 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5964                                 i, le32_to_cpu(rec.t_start), num_clusters);
5965
5966                         status = ocfs2_free_clusters(handle, data_alloc_inode,
5967                                                      data_alloc_bh, start_blk,
5968                                                      num_clusters);
5969                         if (status < 0) {
5970                                 mlog_errno(status);
5971                                 goto bail;
5972                         }
5973                 }
5974
5975                 ocfs2_commit_trans(osb, handle);
5976                 i--;
5977         }
5978
5979         osb->truncated_clusters = 0;
5980
5981 bail:
5982         return status;
5983 }
5984
5985 /* Expects you to already be holding tl_inode->i_mutex */
5986 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5987 {
5988         int status;
5989         unsigned int num_to_flush;
5990         struct inode *tl_inode = osb->osb_tl_inode;
5991         struct inode *data_alloc_inode = NULL;
5992         struct buffer_head *tl_bh = osb->osb_tl_bh;
5993         struct buffer_head *data_alloc_bh = NULL;
5994         struct ocfs2_dinode *di;
5995         struct ocfs2_truncate_log *tl;
5996
5997         BUG_ON(inode_trylock(tl_inode));
5998
5999         di = (struct ocfs2_dinode *) tl_bh->b_data;
6000
6001         /* tl_bh is loaded from ocfs2_truncate_log_init().  It's validated
6002          * by the underlying call to ocfs2_read_inode_block(), so any
6003          * corruption is a code bug */
6004         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6005
6006         tl = &di->id2.i_dealloc;
6007         num_to_flush = le16_to_cpu(tl->tl_used);
6008         trace_ocfs2_flush_truncate_log(
6009                 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6010                 num_to_flush);
6011         if (!num_to_flush) {
6012                 status = 0;
6013                 goto out;
6014         }
6015
6016         data_alloc_inode = ocfs2_get_system_file_inode(osb,
6017                                                        GLOBAL_BITMAP_SYSTEM_INODE,
6018                                                        OCFS2_INVALID_SLOT);
6019         if (!data_alloc_inode) {
6020                 status = -EINVAL;
6021                 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6022                 goto out;
6023         }
6024
6025         inode_lock(data_alloc_inode);
6026
6027         status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6028         if (status < 0) {
6029                 mlog_errno(status);
6030                 goto out_mutex;
6031         }
6032
6033         status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6034                                                data_alloc_bh);
6035         if (status < 0)
6036                 mlog_errno(status);
6037
6038         brelse(data_alloc_bh);
6039         ocfs2_inode_unlock(data_alloc_inode, 1);
6040
6041 out_mutex:
6042         inode_unlock(data_alloc_inode);
6043         iput(data_alloc_inode);
6044
6045 out:
6046         return status;
6047 }
6048
6049 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6050 {
6051         int status;
6052         struct inode *tl_inode = osb->osb_tl_inode;
6053
6054         inode_lock(tl_inode);
6055         status = __ocfs2_flush_truncate_log(osb);
6056         inode_unlock(tl_inode);
6057
6058         return status;
6059 }
6060
6061 static void ocfs2_truncate_log_worker(struct work_struct *work)
6062 {
6063         int status;
6064         struct ocfs2_super *osb =
6065                 container_of(work, struct ocfs2_super,
6066                              osb_truncate_log_wq.work);
6067
6068         status = ocfs2_flush_truncate_log(osb);
6069         if (status < 0)
6070                 mlog_errno(status);
6071         else
6072                 ocfs2_init_steal_slots(osb);
6073 }
6074
6075 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6076 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6077                                        int cancel)
6078 {
6079         if (osb->osb_tl_inode &&
6080                         atomic_read(&osb->osb_tl_disable) == 0) {
6081                 /* We want to push off log flushes while truncates are
6082                  * still running. */
6083                 if (cancel)
6084                         cancel_delayed_work(&osb->osb_truncate_log_wq);
6085
6086                 queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6087                                    OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6088         }
6089 }
6090
6091 /*
6092  * Try to flush truncate logs if we can free enough clusters from it.
6093  * As for return value, "< 0" means error, "0" no space and "1" means
6094  * we have freed enough spaces and let the caller try to allocate again.
6095  */
6096 int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6097                                         unsigned int needed)
6098 {
6099         tid_t target;
6100         int ret = 0;
6101         unsigned int truncated_clusters;
6102
6103         inode_lock(osb->osb_tl_inode);
6104         truncated_clusters = osb->truncated_clusters;
6105         inode_unlock(osb->osb_tl_inode);
6106
6107         /*
6108          * Check whether we can succeed in allocating if we free
6109          * the truncate log.
6110          */
6111         if (truncated_clusters < needed)
6112                 goto out;
6113
6114         ret = ocfs2_flush_truncate_log(osb);
6115         if (ret) {
6116                 mlog_errno(ret);
6117                 goto out;
6118         }
6119
6120         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6121                 jbd2_log_wait_commit(osb->journal->j_journal, target);
6122                 ret = 1;
6123         }
6124 out:
6125         return ret;
6126 }
6127
6128 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6129                                        int slot_num,
6130                                        struct inode **tl_inode,
6131                                        struct buffer_head **tl_bh)
6132 {
6133         int status;
6134         struct inode *inode = NULL;
6135         struct buffer_head *bh = NULL;
6136
6137         inode = ocfs2_get_system_file_inode(osb,
6138                                            TRUNCATE_LOG_SYSTEM_INODE,
6139                                            slot_num);
6140         if (!inode) {
6141                 status = -EINVAL;
6142                 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6143                 goto bail;
6144         }
6145
6146         status = ocfs2_read_inode_block(inode, &bh);
6147         if (status < 0) {
6148                 iput(inode);
6149                 mlog_errno(status);
6150                 goto bail;
6151         }
6152
6153         *tl_inode = inode;
6154         *tl_bh    = bh;
6155 bail:
6156         return status;
6157 }
6158
6159 /* called during the 1st stage of node recovery. we stamp a clean
6160  * truncate log and pass back a copy for processing later. if the
6161  * truncate log does not require processing, a *tl_copy is set to
6162  * NULL. */
6163 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6164                                       int slot_num,
6165                                       struct ocfs2_dinode **tl_copy)
6166 {
6167         int status;
6168         struct inode *tl_inode = NULL;
6169         struct buffer_head *tl_bh = NULL;
6170         struct ocfs2_dinode *di;
6171         struct ocfs2_truncate_log *tl;
6172
6173         *tl_copy = NULL;
6174
6175         trace_ocfs2_begin_truncate_log_recovery(slot_num);
6176
6177         status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6178         if (status < 0) {
6179                 mlog_errno(status);
6180                 goto bail;
6181         }
6182
6183         di = (struct ocfs2_dinode *) tl_bh->b_data;
6184
6185         /* tl_bh is loaded from ocfs2_get_truncate_log_info().  It's
6186          * validated by the underlying call to ocfs2_read_inode_block(),
6187          * so any corruption is a code bug */
6188         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6189
6190         tl = &di->id2.i_dealloc;
6191         if (le16_to_cpu(tl->tl_used)) {
6192                 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6193
6194                 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6195                 if (!(*tl_copy)) {
6196                         status = -ENOMEM;
6197                         mlog_errno(status);
6198                         goto bail;
6199                 }
6200
6201                 /* Assuming the write-out below goes well, this copy
6202                  * will be passed back to recovery for processing. */
6203                 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6204
6205                 /* All we need to do to clear the truncate log is set
6206                  * tl_used. */
6207                 tl->tl_used = 0;
6208
6209                 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6210                 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6211                 if (status < 0) {
6212                         mlog_errno(status);
6213                         goto bail;
6214                 }
6215         }
6216
6217 bail:
6218         iput(tl_inode);
6219         brelse(tl_bh);
6220
6221         if (status < 0) {
6222                 kfree(*tl_copy);
6223                 *tl_copy = NULL;
6224                 mlog_errno(status);
6225         }
6226
6227         return status;
6228 }
6229
6230 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6231                                          struct ocfs2_dinode *tl_copy)
6232 {
6233         int status = 0;
6234         int i;
6235         unsigned int clusters, num_recs, start_cluster;
6236         u64 start_blk;
6237         handle_t *handle;
6238         struct inode *tl_inode = osb->osb_tl_inode;
6239         struct ocfs2_truncate_log *tl;
6240
6241         if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6242                 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6243                 return -EINVAL;
6244         }
6245
6246         tl = &tl_copy->id2.i_dealloc;
6247         num_recs = le16_to_cpu(tl->tl_used);
6248         trace_ocfs2_complete_truncate_log_recovery(
6249                 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6250                 num_recs);
6251
6252         inode_lock(tl_inode);
6253         for(i = 0; i < num_recs; i++) {
6254                 if (ocfs2_truncate_log_needs_flush(osb)) {
6255                         status = __ocfs2_flush_truncate_log(osb);
6256                         if (status < 0) {
6257                                 mlog_errno(status);
6258                                 goto bail_up;
6259                         }
6260                 }
6261
6262                 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6263                 if (IS_ERR(handle)) {
6264                         status = PTR_ERR(handle);
6265                         mlog_errno(status);
6266                         goto bail_up;
6267                 }
6268
6269                 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6270                 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6271                 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6272
6273                 status = ocfs2_truncate_log_append(osb, handle,
6274                                                    start_blk, clusters);
6275                 ocfs2_commit_trans(osb, handle);
6276                 if (status < 0) {
6277                         mlog_errno(status);
6278                         goto bail_up;
6279                 }
6280         }
6281
6282 bail_up:
6283         inode_unlock(tl_inode);
6284
6285         return status;
6286 }
6287
6288 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6289 {
6290         int status;
6291         struct inode *tl_inode = osb->osb_tl_inode;
6292
6293         atomic_set(&osb->osb_tl_disable, 1);
6294
6295         if (tl_inode) {
6296                 cancel_delayed_work(&osb->osb_truncate_log_wq);
6297                 flush_workqueue(osb->ocfs2_wq);
6298
6299                 status = ocfs2_flush_truncate_log(osb);
6300                 if (status < 0)
6301                         mlog_errno(status);
6302
6303                 brelse(osb->osb_tl_bh);
6304                 iput(osb->osb_tl_inode);
6305         }
6306 }
6307
6308 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6309 {
6310         int status;
6311         struct inode *tl_inode = NULL;
6312         struct buffer_head *tl_bh = NULL;
6313
6314         status = ocfs2_get_truncate_log_info(osb,
6315                                              osb->slot_num,
6316                                              &tl_inode,
6317                                              &tl_bh);
6318         if (status < 0)
6319                 mlog_errno(status);
6320
6321         /* ocfs2_truncate_log_shutdown keys on the existence of
6322          * osb->osb_tl_inode so we don't set any of the osb variables
6323          * until we're sure all is well. */
6324         INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6325                           ocfs2_truncate_log_worker);
6326         atomic_set(&osb->osb_tl_disable, 0);
6327         osb->osb_tl_bh    = tl_bh;
6328         osb->osb_tl_inode = tl_inode;
6329
6330         return status;
6331 }
6332
6333 /*
6334  * Delayed de-allocation of suballocator blocks.
6335  *
6336  * Some sets of block de-allocations might involve multiple suballocator inodes.
6337  *
6338  * The locking for this can get extremely complicated, especially when
6339  * the suballocator inodes to delete from aren't known until deep
6340  * within an unrelated codepath.
6341  *
6342  * ocfs2_extent_block structures are a good example of this - an inode
6343  * btree could have been grown by any number of nodes each allocating
6344  * out of their own suballoc inode.
6345  *
6346  * These structures allow the delay of block de-allocation until a
6347  * later time, when locking of multiple cluster inodes won't cause
6348  * deadlock.
6349  */
6350
6351 /*
6352  * Describe a single bit freed from a suballocator.  For the block
6353  * suballocators, it represents one block.  For the global cluster
6354  * allocator, it represents some clusters and free_bit indicates
6355  * clusters number.
6356  */
6357 struct ocfs2_cached_block_free {
6358         struct ocfs2_cached_block_free          *free_next;
6359         u64                                     free_bg;
6360         u64                                     free_blk;
6361         unsigned int                            free_bit;
6362 };
6363
6364 struct ocfs2_per_slot_free_list {
6365         struct ocfs2_per_slot_free_list         *f_next_suballocator;
6366         int                                     f_inode_type;
6367         int                                     f_slot;
6368         struct ocfs2_cached_block_free          *f_first;
6369 };
6370
6371 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6372                                     int sysfile_type,
6373                                     int slot,
6374                                     struct ocfs2_cached_block_free *head)
6375 {
6376         int ret;
6377         u64 bg_blkno;
6378         handle_t *handle;
6379         struct inode *inode;
6380         struct buffer_head *di_bh = NULL;
6381         struct ocfs2_cached_block_free *tmp;
6382
6383         inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6384         if (!inode) {
6385                 ret = -EINVAL;
6386                 mlog_errno(ret);
6387                 goto out;
6388         }
6389
6390         inode_lock(inode);
6391
6392         ret = ocfs2_inode_lock(inode, &di_bh, 1);
6393         if (ret) {
6394                 mlog_errno(ret);
6395                 goto out_mutex;
6396         }
6397
6398         while (head) {
6399                 if (head->free_bg)
6400                         bg_blkno = head->free_bg;
6401                 else
6402                         bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6403                                                               head->free_bit);
6404                 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6405                 if (IS_ERR(handle)) {
6406                         ret = PTR_ERR(handle);
6407                         mlog_errno(ret);
6408                         goto out_unlock;
6409                 }
6410
6411                 trace_ocfs2_free_cached_blocks(
6412                      (unsigned long long)head->free_blk, head->free_bit);
6413
6414                 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6415                                                head->free_bit, bg_blkno, 1);
6416                 if (ret)
6417                         mlog_errno(ret);
6418
6419                 ocfs2_commit_trans(osb, handle);
6420
6421                 tmp = head;
6422                 head = head->free_next;
6423                 kfree(tmp);
6424         }
6425
6426 out_unlock:
6427         ocfs2_inode_unlock(inode, 1);
6428         brelse(di_bh);
6429 out_mutex:
6430         inode_unlock(inode);
6431         iput(inode);
6432 out:
6433         while(head) {
6434                 /* Premature exit may have left some dangling items. */
6435                 tmp = head;
6436                 head = head->free_next;
6437                 kfree(tmp);
6438         }
6439
6440         return ret;
6441 }
6442
6443 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6444                                 u64 blkno, unsigned int bit)
6445 {
6446         int ret = 0;
6447         struct ocfs2_cached_block_free *item;
6448
6449         item = kzalloc(sizeof(*item), GFP_NOFS);
6450         if (item == NULL) {
6451                 ret = -ENOMEM;
6452                 mlog_errno(ret);
6453                 return ret;
6454         }
6455
6456         trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6457
6458         item->free_blk = blkno;
6459         item->free_bit = bit;
6460         item->free_next = ctxt->c_global_allocator;
6461
6462         ctxt->c_global_allocator = item;
6463         return ret;
6464 }
6465
6466 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6467                                       struct ocfs2_cached_block_free *head)
6468 {
6469         struct ocfs2_cached_block_free *tmp;
6470         struct inode *tl_inode = osb->osb_tl_inode;
6471         handle_t *handle;
6472         int ret = 0;
6473
6474         inode_lock(tl_inode);
6475
6476         while (head) {
6477                 if (ocfs2_truncate_log_needs_flush(osb)) {
6478                         ret = __ocfs2_flush_truncate_log(osb);
6479                         if (ret < 0) {
6480                                 mlog_errno(ret);
6481                                 break;
6482                         }
6483                 }
6484
6485                 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6486                 if (IS_ERR(handle)) {
6487                         ret = PTR_ERR(handle);
6488                         mlog_errno(ret);
6489                         break;
6490                 }
6491
6492                 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6493                                                 head->free_bit);
6494
6495                 ocfs2_commit_trans(osb, handle);
6496                 tmp = head;
6497                 head = head->free_next;
6498                 kfree(tmp);
6499
6500                 if (ret < 0) {
6501                         mlog_errno(ret);
6502                         break;
6503                 }
6504         }
6505
6506         inode_unlock(tl_inode);
6507
6508         while (head) {
6509                 /* Premature exit may have left some dangling items. */
6510                 tmp = head;
6511                 head = head->free_next;
6512                 kfree(tmp);
6513         }
6514
6515         return ret;
6516 }
6517
6518 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6519                        struct ocfs2_cached_dealloc_ctxt *ctxt)
6520 {
6521         int ret = 0, ret2;
6522         struct ocfs2_per_slot_free_list *fl;
6523
6524         if (!ctxt)
6525                 return 0;
6526
6527         while (ctxt->c_first_suballocator) {
6528                 fl = ctxt->c_first_suballocator;
6529
6530                 if (fl->f_first) {
6531                         trace_ocfs2_run_deallocs(fl->f_inode_type,
6532                                                  fl->f_slot);
6533                         ret2 = ocfs2_free_cached_blocks(osb,
6534                                                         fl->f_inode_type,
6535                                                         fl->f_slot,
6536                                                         fl->f_first);
6537                         if (ret2)
6538                                 mlog_errno(ret2);
6539                         if (!ret)
6540                                 ret = ret2;
6541                 }
6542
6543                 ctxt->c_first_suballocator = fl->f_next_suballocator;
6544                 kfree(fl);
6545         }
6546
6547         if (ctxt->c_global_allocator) {
6548                 ret2 = ocfs2_free_cached_clusters(osb,
6549                                                   ctxt->c_global_allocator);
6550                 if (ret2)
6551                         mlog_errno(ret2);
6552                 if (!ret)
6553                         ret = ret2;
6554
6555                 ctxt->c_global_allocator = NULL;
6556         }
6557
6558         return ret;
6559 }
6560
6561 static struct ocfs2_per_slot_free_list *
6562 ocfs2_find_per_slot_free_list(int type,
6563                               int slot,
6564                               struct ocfs2_cached_dealloc_ctxt *ctxt)
6565 {
6566         struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6567
6568         while (fl) {
6569                 if (fl->f_inode_type == type && fl->f_slot == slot)
6570                         return fl;
6571
6572                 fl = fl->f_next_suballocator;
6573         }
6574
6575         fl = kmalloc(sizeof(*fl), GFP_NOFS);
6576         if (fl) {
6577                 fl->f_inode_type = type;
6578                 fl->f_slot = slot;
6579                 fl->f_first = NULL;
6580                 fl->f_next_suballocator = ctxt->c_first_suballocator;
6581
6582                 ctxt->c_first_suballocator = fl;
6583         }
6584         return fl;
6585 }
6586
6587 static struct ocfs2_per_slot_free_list *
6588 ocfs2_find_preferred_free_list(int type,
6589                                int preferred_slot,
6590                                int *real_slot,
6591                                struct ocfs2_cached_dealloc_ctxt *ctxt)
6592 {
6593         struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6594
6595         while (fl) {
6596                 if (fl->f_inode_type == type && fl->f_slot == preferred_slot) {
6597                         *real_slot = fl->f_slot;
6598                         return fl;
6599                 }
6600
6601                 fl = fl->f_next_suballocator;
6602         }
6603
6604         /* If we can't find any free list matching preferred slot, just use
6605          * the first one.
6606          */
6607         fl = ctxt->c_first_suballocator;
6608         *real_slot = fl->f_slot;
6609
6610         return fl;
6611 }
6612
6613 /* Return Value 1 indicates empty */
6614 static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et)
6615 {
6616         struct ocfs2_per_slot_free_list *fl = NULL;
6617
6618         if (!et->et_dealloc)
6619                 return 1;
6620
6621         fl = et->et_dealloc->c_first_suballocator;
6622         if (!fl)
6623                 return 1;
6624
6625         if (!fl->f_first)
6626                 return 1;
6627
6628         return 0;
6629 }
6630
6631 /* If extent was deleted from tree due to extent rotation and merging, and
6632  * no metadata is reserved ahead of time. Try to reuse some extents
6633  * just deleted. This is only used to reuse extent blocks.
6634  * It is supposed to find enough extent blocks in dealloc if our estimation
6635  * on metadata is accurate.
6636  */
6637 static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
6638                                         struct ocfs2_extent_tree *et,
6639                                         struct buffer_head **new_eb_bh,
6640                                         int blk_wanted, int *blk_given)
6641 {
6642         int i, status = 0, real_slot;
6643         struct ocfs2_cached_dealloc_ctxt *dealloc;
6644         struct ocfs2_per_slot_free_list *fl;
6645         struct ocfs2_cached_block_free *bf;
6646         struct ocfs2_extent_block *eb;
6647         struct ocfs2_super *osb =
6648                 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
6649
6650         *blk_given = 0;
6651
6652         /* If extent tree doesn't have a dealloc, this is not faulty. Just
6653          * tell upper caller dealloc can't provide any block and it should
6654          * ask for alloc to claim more space.
6655          */
6656         dealloc = et->et_dealloc;
6657         if (!dealloc)
6658                 goto bail;
6659
6660         for (i = 0; i < blk_wanted; i++) {
6661                 /* Prefer to use local slot */
6662                 fl = ocfs2_find_preferred_free_list(EXTENT_ALLOC_SYSTEM_INODE,
6663                                                     osb->slot_num, &real_slot,
6664                                                     dealloc);
6665                 /* If no more block can be reused, we should claim more
6666                  * from alloc. Just return here normally.
6667                  */
6668                 if (!fl) {
6669                         status = 0;
6670                         break;
6671                 }
6672
6673                 bf = fl->f_first;
6674                 fl->f_first = bf->free_next;
6675
6676                 new_eb_bh[i] = sb_getblk(osb->sb, bf->free_blk);
6677                 if (new_eb_bh[i] == NULL) {
6678                         status = -ENOMEM;
6679                         mlog_errno(status);
6680                         goto bail;
6681                 }
6682
6683                 mlog(0, "Reusing block(%llu) from "
6684                      "dealloc(local slot:%d, real slot:%d)\n",
6685                      bf->free_blk, osb->slot_num, real_slot);
6686
6687                 ocfs2_set_new_buffer_uptodate(et->et_ci, new_eb_bh[i]);
6688
6689                 status = ocfs2_journal_access_eb(handle, et->et_ci,
6690                                                  new_eb_bh[i],
6691                                                  OCFS2_JOURNAL_ACCESS_CREATE);
6692                 if (status < 0) {
6693                         mlog_errno(status);
6694                         goto bail;
6695                 }
6696
6697                 memset(new_eb_bh[i]->b_data, 0, osb->sb->s_blocksize);
6698                 eb = (struct ocfs2_extent_block *) new_eb_bh[i]->b_data;
6699
6700                 /* We can't guarantee that buffer head is still cached, so
6701                  * polutlate the extent block again.
6702                  */
6703                 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
6704                 eb->h_blkno = cpu_to_le64(bf->free_blk);
6705                 eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
6706                 eb->h_suballoc_slot = cpu_to_le16(real_slot);
6707                 eb->h_suballoc_loc = cpu_to_le64(bf->free_bg);
6708                 eb->h_suballoc_bit = cpu_to_le16(bf->free_bit);
6709                 eb->h_list.l_count =
6710                         cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
6711
6712                 /* We'll also be dirtied by the caller, so
6713                  * this isn't absolutely necessary.
6714                  */
6715                 ocfs2_journal_dirty(handle, new_eb_bh[i]);
6716
6717                 if (!fl->f_first) {
6718                         dealloc->c_first_suballocator = fl->f_next_suballocator;
6719                         kfree(fl);
6720                 }
6721                 kfree(bf);
6722         }
6723
6724         *blk_given = i;
6725
6726 bail:
6727         if (unlikely(status < 0)) {
6728                 for (i = 0; i < blk_wanted; i++)
6729                         brelse(new_eb_bh[i]);
6730         }
6731
6732         return status;
6733 }
6734
6735 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6736                               int type, int slot, u64 suballoc,
6737                               u64 blkno, unsigned int bit)
6738 {
6739         int ret;
6740         struct ocfs2_per_slot_free_list *fl;
6741         struct ocfs2_cached_block_free *item;
6742
6743         fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6744         if (fl == NULL) {
6745                 ret = -ENOMEM;
6746                 mlog_errno(ret);
6747                 goto out;
6748         }
6749
6750         item = kzalloc(sizeof(*item), GFP_NOFS);
6751         if (item == NULL) {
6752                 ret = -ENOMEM;
6753                 mlog_errno(ret);
6754                 goto out;
6755         }
6756
6757         trace_ocfs2_cache_block_dealloc(type, slot,
6758                                         (unsigned long long)suballoc,
6759                                         (unsigned long long)blkno, bit);
6760
6761         item->free_bg = suballoc;
6762         item->free_blk = blkno;
6763         item->free_bit = bit;
6764         item->free_next = fl->f_first;
6765
6766         fl->f_first = item;
6767
6768         ret = 0;
6769 out:
6770         return ret;
6771 }
6772
6773 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6774                                          struct ocfs2_extent_block *eb)
6775 {
6776         return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6777                                          le16_to_cpu(eb->h_suballoc_slot),
6778                                          le64_to_cpu(eb->h_suballoc_loc),
6779                                          le64_to_cpu(eb->h_blkno),
6780                                          le16_to_cpu(eb->h_suballoc_bit));
6781 }
6782
6783 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6784 {
6785         set_buffer_uptodate(bh);
6786         mark_buffer_dirty(bh);
6787         return 0;
6788 }
6789
6790 void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6791                               unsigned int from, unsigned int to,
6792                               struct page *page, int zero, u64 *phys)
6793 {
6794         int ret, partial = 0;
6795
6796         ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6797         if (ret)
6798                 mlog_errno(ret);
6799
6800         if (zero)
6801                 zero_user_segment(page, from, to);
6802
6803         /*
6804          * Need to set the buffers we zero'd into uptodate
6805          * here if they aren't - ocfs2_map_page_blocks()
6806          * might've skipped some
6807          */
6808         ret = walk_page_buffers(handle, page_buffers(page),
6809                                 from, to, &partial,
6810                                 ocfs2_zero_func);
6811         if (ret < 0)
6812                 mlog_errno(ret);
6813         else if (ocfs2_should_order_data(inode)) {
6814                 ret = ocfs2_jbd2_file_inode(handle, inode);
6815                 if (ret < 0)
6816                         mlog_errno(ret);
6817         }
6818
6819         if (!partial)
6820                 SetPageUptodate(page);
6821
6822         flush_dcache_page(page);
6823 }
6824
6825 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6826                                      loff_t end, struct page **pages,
6827                                      int numpages, u64 phys, handle_t *handle)
6828 {
6829         int i;
6830         struct page *page;
6831         unsigned int from, to = PAGE_SIZE;
6832         struct super_block *sb = inode->i_sb;
6833
6834         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6835
6836         if (numpages == 0)
6837                 goto out;
6838
6839         to = PAGE_SIZE;
6840         for(i = 0; i < numpages; i++) {
6841                 page = pages[i];
6842
6843                 from = start & (PAGE_SIZE - 1);
6844                 if ((end >> PAGE_SHIFT) == page->index)
6845                         to = end & (PAGE_SIZE - 1);
6846
6847                 BUG_ON(from > PAGE_SIZE);
6848                 BUG_ON(to > PAGE_SIZE);
6849
6850                 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6851                                          &phys);
6852
6853                 start = (page->index + 1) << PAGE_SHIFT;
6854         }
6855 out:
6856         if (pages)
6857                 ocfs2_unlock_and_free_pages(pages, numpages);
6858 }
6859
6860 int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6861                      struct page **pages, int *num)
6862 {
6863         int numpages, ret = 0;
6864         struct address_space *mapping = inode->i_mapping;
6865         unsigned long index;
6866         loff_t last_page_bytes;
6867
6868         BUG_ON(start > end);
6869
6870         numpages = 0;
6871         last_page_bytes = PAGE_ALIGN(end);
6872         index = start >> PAGE_SHIFT;
6873         do {
6874                 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6875                 if (!pages[numpages]) {
6876                         ret = -ENOMEM;
6877                         mlog_errno(ret);
6878                         goto out;
6879                 }
6880
6881                 numpages++;
6882                 index++;
6883         } while (index < (last_page_bytes >> PAGE_SHIFT));
6884
6885 out:
6886         if (ret != 0) {
6887                 if (pages)
6888                         ocfs2_unlock_and_free_pages(pages, numpages);
6889                 numpages = 0;
6890         }
6891
6892         *num = numpages;
6893
6894         return ret;
6895 }
6896
6897 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6898                                 struct page **pages, int *num)
6899 {
6900         struct super_block *sb = inode->i_sb;
6901
6902         BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6903                (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6904
6905         return ocfs2_grab_pages(inode, start, end, pages, num);
6906 }
6907
6908 /*
6909  * Zero the area past i_size but still within an allocated
6910  * cluster. This avoids exposing nonzero data on subsequent file
6911  * extends.
6912  *
6913  * We need to call this before i_size is updated on the inode because
6914  * otherwise block_write_full_page() will skip writeout of pages past
6915  * i_size. The new_i_size parameter is passed for this reason.
6916  */
6917 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6918                                   u64 range_start, u64 range_end)
6919 {
6920         int ret = 0, numpages;
6921         struct page **pages = NULL;
6922         u64 phys;
6923         unsigned int ext_flags;
6924         struct super_block *sb = inode->i_sb;
6925
6926         /*
6927          * File systems which don't support sparse files zero on every
6928          * extend.
6929          */
6930         if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6931                 return 0;
6932
6933         pages = kcalloc(ocfs2_pages_per_cluster(sb),
6934                         sizeof(struct page *), GFP_NOFS);
6935         if (pages == NULL) {
6936                 ret = -ENOMEM;
6937                 mlog_errno(ret);
6938                 goto out;
6939         }
6940
6941         if (range_start == range_end)
6942                 goto out;
6943
6944         ret = ocfs2_extent_map_get_blocks(inode,
6945                                           range_start >> sb->s_blocksize_bits,
6946                                           &phys, NULL, &ext_flags);
6947         if (ret) {
6948                 mlog_errno(ret);
6949                 goto out;
6950         }
6951
6952         /*
6953          * Tail is a hole, or is marked unwritten. In either case, we
6954          * can count on read and write to return/push zero's.
6955          */
6956         if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6957                 goto out;
6958
6959         ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6960                                    &numpages);
6961         if (ret) {
6962                 mlog_errno(ret);
6963                 goto out;
6964         }
6965
6966         ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6967                                  numpages, phys, handle);
6968
6969         /*
6970          * Initiate writeout of the pages we zero'd here. We don't
6971          * wait on them - the truncate_inode_pages() call later will
6972          * do that for us.
6973          */
6974         ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6975                                        range_end - 1);
6976         if (ret)
6977                 mlog_errno(ret);
6978
6979 out:
6980         kfree(pages);
6981
6982         return ret;
6983 }
6984
6985 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6986                                              struct ocfs2_dinode *di)
6987 {
6988         unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6989         unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6990
6991         if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6992                 memset(&di->id2, 0, blocksize -
6993                                     offsetof(struct ocfs2_dinode, id2) -
6994                                     xattrsize);
6995         else
6996                 memset(&di->id2, 0, blocksize -
6997                                     offsetof(struct ocfs2_dinode, id2));
6998 }
6999
7000 void ocfs2_dinode_new_extent_list(struct inode *inode,
7001                                   struct ocfs2_dinode *di)
7002 {
7003         ocfs2_zero_dinode_id2_with_xattr(inode, di);
7004         di->id2.i_list.l_tree_depth = 0;
7005         di->id2.i_list.l_next_free_rec = 0;
7006         di->id2.i_list.l_count = cpu_to_le16(
7007                 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7008 }
7009
7010 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7011 {
7012         struct ocfs2_inode_info *oi = OCFS2_I(inode);
7013         struct ocfs2_inline_data *idata = &di->id2.i_data;
7014
7015         spin_lock(&oi->ip_lock);
7016         oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7017         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7018         spin_unlock(&oi->ip_lock);
7019
7020         /*
7021          * We clear the entire i_data structure here so that all
7022          * fields can be properly initialized.
7023          */
7024         ocfs2_zero_dinode_id2_with_xattr(inode, di);
7025
7026         idata->id_count = cpu_to_le16(
7027                         ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7028 }
7029
7030 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7031                                          struct buffer_head *di_bh)
7032 {
7033         int ret, i, has_data, num_pages = 0;
7034         int need_free = 0;
7035         u32 bit_off, num;
7036         handle_t *handle;
7037         u64 uninitialized_var(block);
7038         struct ocfs2_inode_info *oi = OCFS2_I(inode);
7039         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7040         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7041         struct ocfs2_alloc_context *data_ac = NULL;
7042         struct page **pages = NULL;
7043         loff_t end = osb->s_clustersize;
7044         struct ocfs2_extent_tree et;
7045         int did_quota = 0;
7046
7047         has_data = i_size_read(inode) ? 1 : 0;
7048
7049         if (has_data) {
7050                 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
7051                                 sizeof(struct page *), GFP_NOFS);
7052                 if (pages == NULL) {
7053                         ret = -ENOMEM;
7054                         mlog_errno(ret);
7055                         return ret;
7056                 }
7057
7058                 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7059                 if (ret) {
7060                         mlog_errno(ret);
7061                         goto free_pages;
7062                 }
7063         }
7064
7065         handle = ocfs2_start_trans(osb,
7066                                    ocfs2_inline_to_extents_credits(osb->sb));
7067         if (IS_ERR(handle)) {
7068                 ret = PTR_ERR(handle);
7069                 mlog_errno(ret);
7070                 goto out;
7071         }
7072
7073         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7074                                       OCFS2_JOURNAL_ACCESS_WRITE);
7075         if (ret) {
7076                 mlog_errno(ret);
7077                 goto out_commit;
7078         }
7079
7080         if (has_data) {
7081                 unsigned int page_end;
7082                 u64 phys;
7083
7084                 ret = dquot_alloc_space_nodirty(inode,
7085                                        ocfs2_clusters_to_bytes(osb->sb, 1));
7086                 if (ret)
7087                         goto out_commit;
7088                 did_quota = 1;
7089
7090                 data_ac->ac_resv = &oi->ip_la_data_resv;
7091
7092                 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
7093                                            &num);
7094                 if (ret) {
7095                         mlog_errno(ret);
7096                         goto out_commit;
7097                 }
7098
7099                 /*
7100                  * Save two copies, one for insert, and one that can
7101                  * be changed by ocfs2_map_and_dirty_page() below.
7102                  */
7103                 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7104
7105                 /*
7106                  * Non sparse file systems zero on extend, so no need
7107                  * to do that now.
7108                  */
7109                 if (!ocfs2_sparse_alloc(osb) &&
7110                     PAGE_SIZE < osb->s_clustersize)
7111                         end = PAGE_SIZE;
7112
7113                 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
7114                 if (ret) {
7115                         mlog_errno(ret);
7116                         need_free = 1;
7117                         goto out_commit;
7118                 }
7119
7120                 /*
7121                  * This should populate the 1st page for us and mark
7122                  * it up to date.
7123                  */
7124                 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
7125                 if (ret) {
7126                         mlog_errno(ret);
7127                         need_free = 1;
7128                         goto out_unlock;
7129                 }
7130
7131                 page_end = PAGE_SIZE;
7132                 if (PAGE_SIZE > osb->s_clustersize)
7133                         page_end = osb->s_clustersize;
7134
7135                 for (i = 0; i < num_pages; i++)
7136                         ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
7137                                                  pages[i], i > 0, &phys);
7138         }
7139
7140         spin_lock(&oi->ip_lock);
7141         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7142         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7143         spin_unlock(&oi->ip_lock);
7144
7145         ocfs2_update_inode_fsync_trans(handle, inode, 1);
7146         ocfs2_dinode_new_extent_list(inode, di);
7147
7148         ocfs2_journal_dirty(handle, di_bh);
7149
7150         if (has_data) {
7151                 /*
7152                  * An error at this point should be extremely rare. If
7153                  * this proves to be false, we could always re-build
7154                  * the in-inode data from our pages.
7155                  */
7156                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7157                 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7158                 if (ret) {
7159                         mlog_errno(ret);
7160                         need_free = 1;
7161                         goto out_unlock;
7162                 }
7163
7164                 inode->i_blocks = ocfs2_inode_sector_count(inode);
7165         }
7166
7167 out_unlock:
7168         if (pages)
7169                 ocfs2_unlock_and_free_pages(pages, num_pages);
7170
7171 out_commit:
7172         if (ret < 0 && did_quota)
7173                 dquot_free_space_nodirty(inode,
7174                                           ocfs2_clusters_to_bytes(osb->sb, 1));
7175
7176         if (need_free) {
7177                 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7178                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7179                                         bit_off, num);
7180                 else
7181                         ocfs2_free_clusters(handle,
7182                                         data_ac->ac_inode,
7183                                         data_ac->ac_bh,
7184                                         ocfs2_clusters_to_blocks(osb->sb, bit_off),
7185                                         num);
7186         }
7187
7188         ocfs2_commit_trans(osb, handle);
7189
7190 out:
7191         if (data_ac)
7192                 ocfs2_free_alloc_context(data_ac);
7193 free_pages:
7194         kfree(pages);
7195         return ret;
7196 }
7197
7198 /*
7199  * It is expected, that by the time you call this function,
7200  * inode->i_size and fe->i_size have been adjusted.
7201  *
7202  * WARNING: This will kfree the truncate context
7203  */
7204 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7205                           struct inode *inode,
7206                           struct buffer_head *di_bh)
7207 {
7208         int status = 0, i, flags = 0;
7209         u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7210         u64 blkno = 0;
7211         struct ocfs2_extent_list *el;
7212         struct ocfs2_extent_rec *rec;
7213         struct ocfs2_path *path = NULL;
7214         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7215         struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7216         u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7217         struct ocfs2_extent_tree et;
7218         struct ocfs2_cached_dealloc_ctxt dealloc;
7219         struct ocfs2_refcount_tree *ref_tree = NULL;
7220
7221         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7222         ocfs2_init_dealloc_ctxt(&dealloc);
7223
7224         new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7225                                                      i_size_read(inode));
7226
7227         path = ocfs2_new_path(di_bh, &di->id2.i_list,
7228                               ocfs2_journal_access_di);
7229         if (!path) {
7230                 status = -ENOMEM;
7231                 mlog_errno(status);
7232                 goto bail;
7233         }
7234
7235         ocfs2_extent_map_trunc(inode, new_highest_cpos);
7236
7237 start:
7238         /*
7239          * Check that we still have allocation to delete.
7240          */
7241         if (OCFS2_I(inode)->ip_clusters == 0) {
7242                 status = 0;
7243                 goto bail;
7244         }
7245
7246         /*
7247          * Truncate always works against the rightmost tree branch.
7248          */
7249         status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7250         if (status) {
7251                 mlog_errno(status);
7252                 goto bail;
7253         }
7254
7255         trace_ocfs2_commit_truncate(
7256                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7257                 new_highest_cpos,
7258                 OCFS2_I(inode)->ip_clusters,
7259                 path->p_tree_depth);
7260
7261         /*
7262          * By now, el will point to the extent list on the bottom most
7263          * portion of this tree. Only the tail record is considered in
7264          * each pass.
7265          *
7266          * We handle the following cases, in order:
7267          * - empty extent: delete the remaining branch
7268          * - remove the entire record
7269          * - remove a partial record
7270          * - no record needs to be removed (truncate has completed)
7271          */
7272         el = path_leaf_el(path);
7273         if (le16_to_cpu(el->l_next_free_rec) == 0) {
7274                 ocfs2_error(inode->i_sb,
7275                             "Inode %llu has empty extent block at %llu\n",
7276                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
7277                             (unsigned long long)path_leaf_bh(path)->b_blocknr);
7278                 status = -EROFS;
7279                 goto bail;
7280         }
7281
7282         i = le16_to_cpu(el->l_next_free_rec) - 1;
7283         rec = &el->l_recs[i];
7284         flags = rec->e_flags;
7285         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7286
7287         if (i == 0 && ocfs2_is_empty_extent(rec)) {
7288                 /*
7289                  * Lower levels depend on this never happening, but it's best
7290                  * to check it up here before changing the tree.
7291                 */
7292                 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7293                         mlog(ML_ERROR, "Inode %lu has an empty "
7294                                     "extent record, depth %u\n", inode->i_ino,
7295                                     le16_to_cpu(root_el->l_tree_depth));
7296                         status = ocfs2_remove_rightmost_empty_extent(osb,
7297                                         &et, path, &dealloc);
7298                         if (status) {
7299                                 mlog_errno(status);
7300                                 goto bail;
7301                         }
7302
7303                         ocfs2_reinit_path(path, 1);
7304                         goto start;
7305                 } else {
7306                         trunc_cpos = le32_to_cpu(rec->e_cpos);
7307                         trunc_len = 0;
7308                         blkno = 0;
7309                 }
7310         } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7311                 /*
7312                  * Truncate entire record.
7313                  */
7314                 trunc_cpos = le32_to_cpu(rec->e_cpos);
7315                 trunc_len = ocfs2_rec_clusters(el, rec);
7316                 blkno = le64_to_cpu(rec->e_blkno);
7317         } else if (range > new_highest_cpos) {
7318                 /*
7319                  * Partial truncate. it also should be
7320                  * the last truncate we're doing.
7321                  */
7322                 trunc_cpos = new_highest_cpos;
7323                 trunc_len = range - new_highest_cpos;
7324                 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7325                 blkno = le64_to_cpu(rec->e_blkno) +
7326                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7327         } else {
7328                 /*
7329                  * Truncate completed, leave happily.
7330                  */
7331                 status = 0;
7332                 goto bail;
7333         }
7334
7335         phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7336
7337         if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7338                 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7339                                 &ref_tree, NULL);
7340                 if (status) {
7341                         mlog_errno(status);
7342                         goto bail;
7343                 }
7344         }
7345
7346         status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7347                                           phys_cpos, trunc_len, flags, &dealloc,
7348                                           refcount_loc, true);
7349         if (status < 0) {
7350                 mlog_errno(status);
7351                 goto bail;
7352         }
7353
7354         ocfs2_reinit_path(path, 1);
7355
7356         /*
7357          * The check above will catch the case where we've truncated
7358          * away all allocation.
7359          */
7360         goto start;
7361
7362 bail:
7363         if (ref_tree)
7364                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7365
7366         ocfs2_schedule_truncate_log_flush(osb, 1);
7367
7368         ocfs2_run_deallocs(osb, &dealloc);
7369
7370         ocfs2_free_path(path);
7371
7372         return status;
7373 }
7374
7375 /*
7376  * 'start' is inclusive, 'end' is not.
7377  */
7378 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7379                           unsigned int start, unsigned int end, int trunc)
7380 {
7381         int ret;
7382         unsigned int numbytes;
7383         handle_t *handle;
7384         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7385         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7386         struct ocfs2_inline_data *idata = &di->id2.i_data;
7387
7388         if (end > i_size_read(inode))
7389                 end = i_size_read(inode);
7390
7391         BUG_ON(start > end);
7392
7393         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7394             !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7395             !ocfs2_supports_inline_data(osb)) {
7396                 ocfs2_error(inode->i_sb,
7397                             "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7398                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
7399                             le16_to_cpu(di->i_dyn_features),
7400                             OCFS2_I(inode)->ip_dyn_features,
7401                             osb->s_feature_incompat);
7402                 ret = -EROFS;
7403                 goto out;
7404         }
7405
7406         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7407         if (IS_ERR(handle)) {
7408                 ret = PTR_ERR(handle);
7409                 mlog_errno(ret);
7410                 goto out;
7411         }
7412
7413         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7414                                       OCFS2_JOURNAL_ACCESS_WRITE);
7415         if (ret) {
7416                 mlog_errno(ret);
7417                 goto out_commit;
7418         }
7419
7420         numbytes = end - start;
7421         memset(idata->id_data + start, 0, numbytes);
7422
7423         /*
7424          * No need to worry about the data page here - it's been
7425          * truncated already and inline data doesn't need it for
7426          * pushing zero's to disk, so we'll let readpage pick it up
7427          * later.
7428          */
7429         if (trunc) {
7430                 i_size_write(inode, start);
7431                 di->i_size = cpu_to_le64(start);
7432         }
7433
7434         inode->i_blocks = ocfs2_inode_sector_count(inode);
7435         inode->i_ctime = inode->i_mtime = current_time(inode);
7436
7437         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7438         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7439
7440         ocfs2_update_inode_fsync_trans(handle, inode, 1);
7441         ocfs2_journal_dirty(handle, di_bh);
7442
7443 out_commit:
7444         ocfs2_commit_trans(osb, handle);
7445
7446 out:
7447         return ret;
7448 }
7449
7450 static int ocfs2_trim_extent(struct super_block *sb,
7451                              struct ocfs2_group_desc *gd,
7452                              u64 group, u32 start, u32 count)
7453 {
7454         u64 discard, bcount;
7455         struct ocfs2_super *osb = OCFS2_SB(sb);
7456
7457         bcount = ocfs2_clusters_to_blocks(sb, count);
7458         discard = ocfs2_clusters_to_blocks(sb, start);
7459
7460         /*
7461          * For the first cluster group, the gd->bg_blkno is not at the start
7462          * of the group, but at an offset from the start. If we add it while
7463          * calculating discard for first group, we will wrongly start fstrim a
7464          * few blocks after the desried start block and the range can cross
7465          * over into the next cluster group. So, add it only if this is not
7466          * the first cluster group.
7467          */
7468         if (group != osb->first_cluster_group_blkno)
7469                 discard += le64_to_cpu(gd->bg_blkno);
7470
7471         trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7472
7473         return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7474 }
7475
7476 static int ocfs2_trim_group(struct super_block *sb,
7477                             struct ocfs2_group_desc *gd, u64 group,
7478                             u32 start, u32 max, u32 minbits)
7479 {
7480         int ret = 0, count = 0, next;
7481         void *bitmap = gd->bg_bitmap;
7482
7483         if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7484                 return 0;
7485
7486         trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7487                                start, max, minbits);
7488
7489         while (start < max) {
7490                 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7491                 if (start >= max)
7492                         break;
7493                 next = ocfs2_find_next_bit(bitmap, max, start);
7494
7495                 if ((next - start) >= minbits) {
7496                         ret = ocfs2_trim_extent(sb, gd, group,
7497                                                 start, next - start);
7498                         if (ret < 0) {
7499                                 mlog_errno(ret);
7500                                 break;
7501                         }
7502                         count += next - start;
7503                 }
7504                 start = next + 1;
7505
7506                 if (fatal_signal_pending(current)) {
7507                         count = -ERESTARTSYS;
7508                         break;
7509                 }
7510
7511                 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7512                         break;
7513         }
7514
7515         if (ret < 0)
7516                 count = ret;
7517
7518         return count;
7519 }
7520
7521 static
7522 int ocfs2_trim_mainbm(struct super_block *sb, struct fstrim_range *range)
7523 {
7524         struct ocfs2_super *osb = OCFS2_SB(sb);
7525         u64 start, len, trimmed = 0, first_group, last_group = 0, group = 0;
7526         int ret, cnt;
7527         u32 first_bit, last_bit, minlen;
7528         struct buffer_head *main_bm_bh = NULL;
7529         struct inode *main_bm_inode = NULL;
7530         struct buffer_head *gd_bh = NULL;
7531         struct ocfs2_dinode *main_bm;
7532         struct ocfs2_group_desc *gd = NULL;
7533
7534         start = range->start >> osb->s_clustersize_bits;
7535         len = range->len >> osb->s_clustersize_bits;
7536         minlen = range->minlen >> osb->s_clustersize_bits;
7537
7538         if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7539                 return -EINVAL;
7540
7541         trace_ocfs2_trim_mainbm(start, len, minlen);
7542
7543 next_group:
7544         main_bm_inode = ocfs2_get_system_file_inode(osb,
7545                                                     GLOBAL_BITMAP_SYSTEM_INODE,
7546                                                     OCFS2_INVALID_SLOT);
7547         if (!main_bm_inode) {
7548                 ret = -EIO;
7549                 mlog_errno(ret);
7550                 goto out;
7551         }
7552
7553         inode_lock(main_bm_inode);
7554
7555         ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7556         if (ret < 0) {
7557                 mlog_errno(ret);
7558                 goto out_mutex;
7559         }
7560         main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7561
7562         /*
7563          * Do some check before trim the first group.
7564          */
7565         if (!group) {
7566                 if (start >= le32_to_cpu(main_bm->i_clusters)) {
7567                         ret = -EINVAL;
7568                         goto out_unlock;
7569                 }
7570
7571                 if (start + len > le32_to_cpu(main_bm->i_clusters))
7572                         len = le32_to_cpu(main_bm->i_clusters) - start;
7573
7574                 /*
7575                  * Determine first and last group to examine based on
7576                  * start and len
7577                  */
7578                 first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7579                 if (first_group == osb->first_cluster_group_blkno)
7580                         first_bit = start;
7581                 else
7582                         first_bit = start - ocfs2_blocks_to_clusters(sb,
7583                                                                 first_group);
7584                 last_group = ocfs2_which_cluster_group(main_bm_inode,
7585                                                        start + len - 1);
7586                 group = first_group;
7587         }
7588
7589         do {
7590                 if (first_bit + len >= osb->bitmap_cpg)
7591                         last_bit = osb->bitmap_cpg;
7592                 else
7593                         last_bit = first_bit + len;
7594
7595                 ret = ocfs2_read_group_descriptor(main_bm_inode,
7596                                                   main_bm, group,
7597                                                   &gd_bh);
7598                 if (ret < 0) {
7599                         mlog_errno(ret);
7600                         break;
7601                 }
7602
7603                 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7604                 cnt = ocfs2_trim_group(sb, gd, group,
7605                                        first_bit, last_bit, minlen);
7606                 brelse(gd_bh);
7607                 gd_bh = NULL;
7608                 if (cnt < 0) {
7609                         ret = cnt;
7610                         mlog_errno(ret);
7611                         break;
7612                 }
7613
7614                 trimmed += cnt;
7615                 len -= osb->bitmap_cpg - first_bit;
7616                 first_bit = 0;
7617                 if (group == osb->first_cluster_group_blkno)
7618                         group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7619                 else
7620                         group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7621         } while (0);
7622
7623 out_unlock:
7624         ocfs2_inode_unlock(main_bm_inode, 0);
7625         brelse(main_bm_bh);
7626         main_bm_bh = NULL;
7627 out_mutex:
7628         inode_unlock(main_bm_inode);
7629         iput(main_bm_inode);
7630
7631         /*
7632          * If all the groups trim are not done or failed, but we should release
7633          * main_bm related locks for avoiding the current IO starve, then go to
7634          * trim the next group
7635          */
7636         if (ret >= 0 && group <= last_group)
7637                 goto next_group;
7638 out:
7639         range->len = trimmed * sb->s_blocksize;
7640         return ret;
7641 }
7642
7643 int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7644 {
7645         int ret;
7646         struct ocfs2_super *osb = OCFS2_SB(sb);
7647         struct ocfs2_trim_fs_info info, *pinfo = NULL;
7648
7649         ocfs2_trim_fs_lock_res_init(osb);
7650
7651         trace_ocfs2_trim_fs(range->start, range->len, range->minlen);
7652
7653         ret = ocfs2_trim_fs_lock(osb, NULL, 1);
7654         if (ret < 0) {
7655                 if (ret != -EAGAIN) {
7656                         mlog_errno(ret);
7657                         ocfs2_trim_fs_lock_res_uninit(osb);
7658                         return ret;
7659                 }
7660
7661                 mlog(ML_NOTICE, "Wait for trim on device (%s) to "
7662                      "finish, which is running from another node.\n",
7663                      osb->dev_str);
7664                 ret = ocfs2_trim_fs_lock(osb, &info, 0);
7665                 if (ret < 0) {
7666                         mlog_errno(ret);
7667                         ocfs2_trim_fs_lock_res_uninit(osb);
7668                         return ret;
7669                 }
7670
7671                 if (info.tf_valid && info.tf_success &&
7672                     info.tf_start == range->start &&
7673                     info.tf_len == range->len &&
7674                     info.tf_minlen == range->minlen) {
7675                         /* Avoid sending duplicated trim to a shared device */
7676                         mlog(ML_NOTICE, "The same trim on device (%s) was "
7677                              "just done from node (%u), return.\n",
7678                              osb->dev_str, info.tf_nodenum);
7679                         range->len = info.tf_trimlen;
7680                         goto out;
7681                 }
7682         }
7683
7684         info.tf_nodenum = osb->node_num;
7685         info.tf_start = range->start;
7686         info.tf_len = range->len;
7687         info.tf_minlen = range->minlen;
7688
7689         ret = ocfs2_trim_mainbm(sb, range);
7690
7691         info.tf_trimlen = range->len;
7692         info.tf_success = (ret < 0 ? 0 : 1);
7693         pinfo = &info;
7694 out:
7695         ocfs2_trim_fs_unlock(osb, pinfo);
7696         ocfs2_trim_fs_lock_res_uninit(osb);
7697         return ret;
7698 }