fs/ocfs2/alloc.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * alloc.c
   5  *
   6  * Extent allocs and frees
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/fs.h>
  27 #include <linux/types.h>
  28 #include <linux/slab.h>
  29 #include <linux/highmem.h>
  30 #include <linux/swap.h>
  31 #include <linux/quotaops.h>
  32 #include <linux/blkdev.h>
  33
  34 #include <cluster/masklog.h>
  35
  36 #include "ocfs2.h"
  37
  38 #include "alloc.h"
  39 #include "aops.h"
  40 #include "blockcheck.h"
  41 #include "dlmglue.h"
  42 #include "extent_map.h"
  43 #include "inode.h"
  44 #include "journal.h"
  45 #include "localalloc.h"
  46 #include "suballoc.h"
  47 #include "sysfile.h"
  48 #include "file.h"
  49 #include "super.h"
  50 #include "uptodate.h"
  51 #include "xattr.h"
  52 #include "refcounttree.h"
  53 #include "ocfs2_trace.h"
  54
  55 #include "buffer_head_io.h"
  56
  57 enum ocfs2_contig_type {
  58         CONTIG_NONE = 0,
  59         CONTIG_LEFT,
  60         CONTIG_RIGHT,
  61         CONTIG_LEFTRIGHT,
  62 };
  63
  64 static enum ocfs2_contig_type
  65         ocfs2_extent_rec_contig(struct super_block *sb,
  66                                 struct ocfs2_extent_rec *ext,
  67                                 struct ocfs2_extent_rec *insert_rec);
  68 /*
  69  * Operations for a specific extent tree type.
  70  *
  71  * To implement an on-disk btree (extent tree) type in ocfs2, add
  72  * an ocfs2_extent_tree_operations structure and the matching
  73  * ocfs2_init_<thingy>_extent_tree() function.  That's pretty much it
  74  * for the allocation portion of the extent tree.
  75  */
  76 struct ocfs2_extent_tree_operations {
  77         /*
  78          * last_eb_blk is the block number of the right most leaf extent
  79          * block.  Most on-disk structures containing an extent tree store
  80          * this value for fast access.  The ->eo_set_last_eb_blk() and
  81          * ->eo_get_last_eb_blk() operations access this value.  They are
  82          *  both required.
  83          */
  84         void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
  85                                    u64 blkno);
  86         u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
  87
  88         /*
  89          * The on-disk structure usually keeps track of how many total
  90          * clusters are stored in this extent tree.  This function updates
  91          * that value.  new_clusters is the delta, and must be
  92          * added to the total.  Required.
  93          */
  94         void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
  95                                    u32 new_clusters);
  96
  97         /*
  98          * If this extent tree is supported by an extent map, insert
  99          * a record into the map.
 100          */
 101         void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
 102                                      struct ocfs2_extent_rec *rec);
 103
 104         /*
 105          * If this extent tree is supported by an extent map, truncate the
 106          * map to clusters,
 107          */
 108         void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
 109                                        u32 clusters);
 110
 111         /*
 112          * If ->eo_insert_check() exists, it is called before rec is
 113          * inserted into the extent tree.  It is optional.
 114          */
 115         int (*eo_insert_check)(struct ocfs2_extent_tree *et,
 116                                struct ocfs2_extent_rec *rec);
 117         int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
 118
 119         /*
 120          * --------------------------------------------------------------
 121          * The remaining are internal to ocfs2_extent_tree and don't have
 122          * accessor functions
 123          */
 124
 125         /*
 126          * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
 127          * It is required.
 128          */
 129         void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
 130
 131         /*
 132          * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
 133          * it exists.  If it does not, et->et_max_leaf_clusters is set
 134          * to 0 (unlimited).  Optional.
 135          */
 136         void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
 137
 138         /*
 139          * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
 140          * are contiguous or not. Optional. Don't need to set it if use
 141          * ocfs2_extent_rec as the tree leaf.
 142          */
 143         enum ocfs2_contig_type
 144                 (*eo_extent_contig)(struct ocfs2_extent_tree *et,
 145                                     struct ocfs2_extent_rec *ext,
 146                                     struct ocfs2_extent_rec *insert_rec);
 147 };
 148
 149
 150 /*
 151  * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
 152  * in the methods.
 153  */
 154 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
 155 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
 156                                          u64 blkno);
 157 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
 158                                          u32 clusters);
 159 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
 160                                            struct ocfs2_extent_rec *rec);
 161 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
 162                                              u32 clusters);
 163 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
 164                                      struct ocfs2_extent_rec *rec);
 165 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
 166 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
 167 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
 168         .eo_set_last_eb_blk     = ocfs2_dinode_set_last_eb_blk,
 169         .eo_get_last_eb_blk     = ocfs2_dinode_get_last_eb_blk,
 170         .eo_update_clusters     = ocfs2_dinode_update_clusters,
 171         .eo_extent_map_insert   = ocfs2_dinode_extent_map_insert,
 172         .eo_extent_map_truncate = ocfs2_dinode_extent_map_truncate,
 173         .eo_insert_check        = ocfs2_dinode_insert_check,
 174         .eo_sanity_check        = ocfs2_dinode_sanity_check,
 175         .eo_fill_root_el        = ocfs2_dinode_fill_root_el,
 176 };
 177
 178 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
 179                                          u64 blkno)
 180 {
 181         struct ocfs2_dinode *di = et->et_object;
 182
 183         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 184         di->i_last_eb_blk = cpu_to_le64(blkno);
 185 }
 186
 187 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
 188 {
 189         struct ocfs2_dinode *di = et->et_object;
 190
 191         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 192         return le64_to_cpu(di->i_last_eb_blk);
 193 }
 194
 195 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
 196                                          u32 clusters)
 197 {
 198         struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
 199         struct ocfs2_dinode *di = et->et_object;
 200
 201         le32_add_cpu(&di->i_clusters, clusters);
 202         spin_lock(&oi->ip_lock);
 203         oi->ip_clusters = le32_to_cpu(di->i_clusters);
 204         spin_unlock(&oi->ip_lock);
 205 }
 206
 207 static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
 208                                            struct ocfs2_extent_rec *rec)
 209 {
 210         struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
 211
 212         ocfs2_extent_map_insert_rec(inode, rec);
 213 }
 214
 215 static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
 216                                              u32 clusters)
 217 {
 218         struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
 219
 220         ocfs2_extent_map_trunc(inode, clusters);
 221 }
 222
 223 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
 224                                      struct ocfs2_extent_rec *rec)
 225 {
 226         struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
 227         struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
 228
 229         BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
 230         mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
 231                         (oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
 232                         "Device %s, asking for sparse allocation: inode %llu, "
 233                         "cpos %u, clusters %u\n",
 234                         osb->dev_str,
 235                         (unsigned long long)oi->ip_blkno,
 236                         rec->e_cpos, oi->ip_clusters);
 237
 238         return 0;
 239 }
 240
 241 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
 242 {
 243         struct ocfs2_dinode *di = et->et_object;
 244
 245         BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 246         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
 247
 248         return 0;
 249 }
 250
 251 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
 252 {
 253         struct ocfs2_dinode *di = et->et_object;
 254
 255         et->et_root_el = &di->id2.i_list;
 256 }
 257
 258
 259 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
 260 {
 261         struct ocfs2_xattr_value_buf *vb = et->et_object;
 262
 263         et->et_root_el = &vb->vb_xv->xr_list;
 264 }
 265
 266 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
 267                                               u64 blkno)
 268 {
 269         struct ocfs2_xattr_value_buf *vb = et->et_object;
 270
 271         vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
 272 }
 273
 274 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
 275 {
 276         struct ocfs2_xattr_value_buf *vb = et->et_object;
 277
 278         return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
 279 }
 280
 281 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
 282                                               u32 clusters)
 283 {
 284         struct ocfs2_xattr_value_buf *vb = et->et_object;
 285
 286         le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
 287 }
 288
 289 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
 290         .eo_set_last_eb_blk     = ocfs2_xattr_value_set_last_eb_blk,
 291         .eo_get_last_eb_blk     = ocfs2_xattr_value_get_last_eb_blk,
 292         .eo_update_clusters     = ocfs2_xattr_value_update_clusters,
 293         .eo_fill_root_el        = ocfs2_xattr_value_fill_root_el,
 294 };
 295
 296 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
 297 {
 298         struct ocfs2_xattr_block *xb = et->et_object;
 299
 300         et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
 301 }
 302
 303 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
 304 {
 305         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
 306         et->et_max_leaf_clusters =
 307                 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
 308 }
 309
 310 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
 311                                              u64 blkno)
 312 {
 313         struct ocfs2_xattr_block *xb = et->et_object;
 314         struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
 315
 316         xt->xt_last_eb_blk = cpu_to_le64(blkno);
 317 }
 318
 319 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
 320 {
 321         struct ocfs2_xattr_block *xb = et->et_object;
 322         struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
 323
 324         return le64_to_cpu(xt->xt_last_eb_blk);
 325 }
 326
 327 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
 328                                              u32 clusters)
 329 {
 330         struct ocfs2_xattr_block *xb = et->et_object;
 331
 332         le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
 333 }
 334
 335 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
 336         .eo_set_last_eb_blk     = ocfs2_xattr_tree_set_last_eb_blk,
 337         .eo_get_last_eb_blk     = ocfs2_xattr_tree_get_last_eb_blk,
 338         .eo_update_clusters     = ocfs2_xattr_tree_update_clusters,
 339         .eo_fill_root_el        = ocfs2_xattr_tree_fill_root_el,
 340         .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
 341 };
 342
 343 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
 344                                           u64 blkno)
 345 {
 346         struct ocfs2_dx_root_block *dx_root = et->et_object;
 347
 348         dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
 349 }
 350
 351 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
 352 {
 353         struct ocfs2_dx_root_block *dx_root = et->et_object;
 354
 355         return le64_to_cpu(dx_root->dr_last_eb_blk);
 356 }
 357
 358 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
 359                                           u32 clusters)
 360 {
 361         struct ocfs2_dx_root_block *dx_root = et->et_object;
 362
 363         le32_add_cpu(&dx_root->dr_clusters, clusters);
 364 }
 365
 366 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
 367 {
 368         struct ocfs2_dx_root_block *dx_root = et->et_object;
 369
 370         BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
 371
 372         return 0;
 373 }
 374
 375 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
 376 {
 377         struct ocfs2_dx_root_block *dx_root = et->et_object;
 378
 379         et->et_root_el = &dx_root->dr_list;
 380 }
 381
 382 static struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
 383         .eo_set_last_eb_blk     = ocfs2_dx_root_set_last_eb_blk,
 384         .eo_get_last_eb_blk     = ocfs2_dx_root_get_last_eb_blk,
 385         .eo_update_clusters     = ocfs2_dx_root_update_clusters,
 386         .eo_sanity_check        = ocfs2_dx_root_sanity_check,
 387         .eo_fill_root_el        = ocfs2_dx_root_fill_root_el,
 388 };
 389
 390 static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
 391 {
 392         struct ocfs2_refcount_block *rb = et->et_object;
 393
 394         et->et_root_el = &rb->rf_list;
 395 }
 396
 397 static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
 398                                                 u64 blkno)
 399 {
 400         struct ocfs2_refcount_block *rb = et->et_object;
 401
 402         rb->rf_last_eb_blk = cpu_to_le64(blkno);
 403 }
 404
 405 static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
 406 {
 407         struct ocfs2_refcount_block *rb = et->et_object;
 408
 409         return le64_to_cpu(rb->rf_last_eb_blk);
 410 }
 411
 412 static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
 413                                                 u32 clusters)
 414 {
 415         struct ocfs2_refcount_block *rb = et->et_object;
 416
 417         le32_add_cpu(&rb->rf_clusters, clusters);
 418 }
 419
 420 static enum ocfs2_contig_type
 421 ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
 422                                   struct ocfs2_extent_rec *ext,
 423                                   struct ocfs2_extent_rec *insert_rec)
 424 {
 425         return CONTIG_NONE;
 426 }
 427
 428 static struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
 429         .eo_set_last_eb_blk     = ocfs2_refcount_tree_set_last_eb_blk,
 430         .eo_get_last_eb_blk     = ocfs2_refcount_tree_get_last_eb_blk,
 431         .eo_update_clusters     = ocfs2_refcount_tree_update_clusters,
 432         .eo_fill_root_el        = ocfs2_refcount_tree_fill_root_el,
 433         .eo_extent_contig       = ocfs2_refcount_tree_extent_contig,
 434 };
 435
 436 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
 437                                      struct ocfs2_caching_info *ci,
 438                                      struct buffer_head *bh,
 439                                      ocfs2_journal_access_func access,
 440                                      void *obj,
 441                                      struct ocfs2_extent_tree_operations *ops)
 442 {
 443         et->et_ops = ops;
 444         et->et_root_bh = bh;
 445         et->et_ci = ci;
 446         et->et_root_journal_access = access;
 447         if (!obj)
 448                 obj = (void *)bh->b_data;
 449         et->et_object = obj;
 450
 451         et->et_ops->eo_fill_root_el(et);
 452         if (!et->et_ops->eo_fill_max_leaf_clusters)
 453                 et->et_max_leaf_clusters = 0;
 454         else
 455                 et->et_ops->eo_fill_max_leaf_clusters(et);
 456 }
 457
 458 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
 459                                    struct ocfs2_caching_info *ci,
 460                                    struct buffer_head *bh)
 461 {
 462         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
 463                                  NULL, &ocfs2_dinode_et_ops);
 464 }
 465
 466 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
 467                                        struct ocfs2_caching_info *ci,
 468                                        struct buffer_head *bh)
 469 {
 470         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
 471                                  NULL, &ocfs2_xattr_tree_et_ops);
 472 }
 473
 474 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
 475                                         struct ocfs2_caching_info *ci,
 476                                         struct ocfs2_xattr_value_buf *vb)
 477 {
 478         __ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
 479                                  &ocfs2_xattr_value_et_ops);
 480 }
 481
 482 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
 483                                     struct ocfs2_caching_info *ci,
 484                                     struct buffer_head *bh)
 485 {
 486         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
 487                                  NULL, &ocfs2_dx_root_et_ops);
 488 }
 489
 490 void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
 491                                      struct ocfs2_caching_info *ci,
 492                                      struct buffer_head *bh)
 493 {
 494         __ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
 495                                  NULL, &ocfs2_refcount_tree_et_ops);
 496 }
 497
 498 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
 499                                             u64 new_last_eb_blk)
 500 {
 501         et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
 502 }
 503
 504 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
 505 {
 506         return et->et_ops->eo_get_last_eb_blk(et);
 507 }
 508
 509 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
 510                                             u32 clusters)
 511 {
 512         et->et_ops->eo_update_clusters(et, clusters);
 513 }
 514
 515 static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
 516                                               struct ocfs2_extent_rec *rec)
 517 {
 518         if (et->et_ops->eo_extent_map_insert)
 519                 et->et_ops->eo_extent_map_insert(et, rec);
 520 }
 521
 522 static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
 523                                                 u32 clusters)
 524 {
 525         if (et->et_ops->eo_extent_map_truncate)
 526                 et->et_ops->eo_extent_map_truncate(et, clusters);
 527 }
 528
 529 static inline int ocfs2_et_root_journal_access(handle_t *handle,
 530                                                struct ocfs2_extent_tree *et,
 531                                                int type)
 532 {
 533         return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
 534                                           type);
 535 }
 536
 537 static inline enum ocfs2_contig_type
 538         ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
 539                                struct ocfs2_extent_rec *rec,
 540                                struct ocfs2_extent_rec *insert_rec)
 541 {
 542         if (et->et_ops->eo_extent_contig)
 543                 return et->et_ops->eo_extent_contig(et, rec, insert_rec);
 544
 545         return ocfs2_extent_rec_contig(
 546                                 ocfs2_metadata_cache_get_super(et->et_ci),
 547                                 rec, insert_rec);
 548 }
 549
 550 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
 551                                         struct ocfs2_extent_rec *rec)
 552 {
 553         int ret = 0;
 554
 555         if (et->et_ops->eo_insert_check)
 556                 ret = et->et_ops->eo_insert_check(et, rec);
 557         return ret;
 558 }
 559
 560 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
 561 {
 562         int ret = 0;
 563
 564         if (et->et_ops->eo_sanity_check)
 565                 ret = et->et_ops->eo_sanity_check(et);
 566         return ret;
 567 }
 568
 569 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
 570                                          struct ocfs2_extent_block *eb);
 571 static void ocfs2_adjust_rightmost_records(handle_t *handle,
 572                                            struct ocfs2_extent_tree *et,
 573                                            struct ocfs2_path *path,
 574                                            struct ocfs2_extent_rec *insert_rec);
 575 /*
 576  * Reset the actual path elements so that we can re-use the structure
 577  * to build another path. Generally, this involves freeing the buffer
 578  * heads.
 579  */
 580 void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
 581 {
 582         int i, start = 0, depth = 0;
 583         struct ocfs2_path_item *node;
 584
 585         if (keep_root)
 586                 start = 1;
 587
 588         for(i = start; i < path_num_items(path); i++) {
 589                 node = &path->p_node[i];
 590
 591                 brelse(node->bh);
 592                 node->bh = NULL;
 593                 node->el = NULL;
 594         }
 595
 596         /*
 597          * Tree depth may change during truncate, or insert. If we're
 598          * keeping the root extent list, then make sure that our path
 599          * structure reflects the proper depth.
 600          */
 601         if (keep_root)
 602                 depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
 603         else
 604                 path_root_access(path) = NULL;
 605
 606         path->p_tree_depth = depth;
 607 }
 608
 609 void ocfs2_free_path(struct ocfs2_path *path)
 610 {
 611         if (path) {
 612                 ocfs2_reinit_path(path, 0);
 613                 kfree(path);
 614         }
 615 }
 616
 617 /*
 618  * All the elements of src into dest. After this call, src could be freed
 619  * without affecting dest.
 620  *
 621  * Both paths should have the same root. Any non-root elements of dest
 622  * will be freed.
 623  */
 624 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
 625 {
 626         int i;
 627
 628         BUG_ON(path_root_bh(dest) != path_root_bh(src));
 629         BUG_ON(path_root_el(dest) != path_root_el(src));
 630         BUG_ON(path_root_access(dest) != path_root_access(src));
 631
 632         ocfs2_reinit_path(dest, 1);
 633
 634         for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
 635                 dest->p_node[i].bh = src->p_node[i].bh;
 636                 dest->p_node[i].el = src->p_node[i].el;
 637
 638                 if (dest->p_node[i].bh)
 639                         get_bh(dest->p_node[i].bh);
 640         }
 641 }
 642
 643 /*
 644  * Make the *dest path the same as src and re-initialize src path to
 645  * have a root only.
 646  */
 647 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
 648 {
 649         int i;
 650
 651         BUG_ON(path_root_bh(dest) != path_root_bh(src));
 652         BUG_ON(path_root_access(dest) != path_root_access(src));
 653
 654         for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
 655                 brelse(dest->p_node[i].bh);
 656
 657                 dest->p_node[i].bh = src->p_node[i].bh;
 658                 dest->p_node[i].el = src->p_node[i].el;
 659
 660                 src->p_node[i].bh = NULL;
 661                 src->p_node[i].el = NULL;
 662         }
 663 }
 664
 665 /*
 666  * Insert an extent block at given index.
 667  *
 668  * This will not take an additional reference on eb_bh.
 669  */
 670 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
 671                                         struct buffer_head *eb_bh)
 672 {
 673         struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
 674
 675         /*
 676          * Right now, no root bh is an extent block, so this helps
 677          * catch code errors with dinode trees. The assertion can be
 678          * safely removed if we ever need to insert extent block
 679          * structures at the root.
 680          */
 681         BUG_ON(index == 0);
 682
 683         path->p_node[index].bh = eb_bh;
 684         path->p_node[index].el = &eb->h_list;
 685 }
 686
 687 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
 688                                          struct ocfs2_extent_list *root_el,
 689                                          ocfs2_journal_access_func access)
 690 {
 691         struct ocfs2_path *path;
 692
 693         BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
 694
 695         path = kzalloc(sizeof(*path), GFP_NOFS);
 696         if (path) {
 697                 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
 698                 get_bh(root_bh);
 699                 path_root_bh(path) = root_bh;
 700                 path_root_el(path) = root_el;
 701                 path_root_access(path) = access;
 702         }
 703
 704         return path;
 705 }
 706
 707 struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
 708 {
 709         return ocfs2_new_path(path_root_bh(path), path_root_el(path),
 710                               path_root_access(path));
 711 }
 712
 713 struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
 714 {
 715         return ocfs2_new_path(et->et_root_bh, et->et_root_el,
 716                               et->et_root_journal_access);
 717 }
 718
 719 /*
 720  * Journal the buffer at depth idx.  All idx>0 are extent_blocks,
 721  * otherwise it's the root_access function.
 722  *
 723  * I don't like the way this function's name looks next to
 724  * ocfs2_journal_access_path(), but I don't have a better one.
 725  */
 726 int ocfs2_path_bh_journal_access(handle_t *handle,
 727                                  struct ocfs2_caching_info *ci,
 728                                  struct ocfs2_path *path,
 729                                  int idx)
 730 {
 731         ocfs2_journal_access_func access = path_root_access(path);
 732
 733         if (!access)
 734                 access = ocfs2_journal_access;
 735
 736         if (idx)
 737                 access = ocfs2_journal_access_eb;
 738
 739         return access(handle, ci, path->p_node[idx].bh,
 740                       OCFS2_JOURNAL_ACCESS_WRITE);
 741 }
 742
 743 /*
 744  * Convenience function to journal all components in a path.
 745  */
 746 int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
 747                               handle_t *handle,
 748                               struct ocfs2_path *path)
 749 {
 750         int i, ret = 0;
 751
 752         if (!path)
 753                 goto out;
 754
 755         for(i = 0; i < path_num_items(path); i++) {
 756                 ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
 757                 if (ret < 0) {
 758                         mlog_errno(ret);
 759                         goto out;
 760                 }
 761         }
 762
 763 out:
 764         return ret;
 765 }
 766
 767 /*
 768  * Return the index of the extent record which contains cluster #v_cluster.
 769  * -1 is returned if it was not found.
 770  *
 771  * Should work fine on interior and exterior nodes.
 772  */
 773 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
 774 {
 775         int ret = -1;
 776         int i;
 777         struct ocfs2_extent_rec *rec;
 778         u32 rec_end, rec_start, clusters;
 779
 780         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
 781                 rec = &el->l_recs[i];
 782
 783                 rec_start = le32_to_cpu(rec->e_cpos);
 784                 clusters = ocfs2_rec_clusters(el, rec);
 785
 786                 rec_end = rec_start + clusters;
 787
 788                 if (v_cluster >= rec_start && v_cluster < rec_end) {
 789                         ret = i;
 790                         break;
 791                 }
 792         }
 793
 794         return ret;
 795 }
 796
 797 /*
 798  * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
 799  * ocfs2_extent_rec_contig only work properly against leaf nodes!
 800  */
 801 static int ocfs2_block_extent_contig(struct super_block *sb,
 802                                      struct ocfs2_extent_rec *ext,
 803                                      u64 blkno)
 804 {
 805         u64 blk_end = le64_to_cpu(ext->e_blkno);
 806
 807         blk_end += ocfs2_clusters_to_blocks(sb,
 808                                     le16_to_cpu(ext->e_leaf_clusters));
 809
 810         return blkno == blk_end;
 811 }
 812
 813 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
 814                                   struct ocfs2_extent_rec *right)
 815 {
 816         u32 left_range;
 817
 818         left_range = le32_to_cpu(left->e_cpos) +
 819                 le16_to_cpu(left->e_leaf_clusters);
 820
 821         return (left_range == le32_to_cpu(right->e_cpos));
 822 }
 823
 824 static enum ocfs2_contig_type
 825         ocfs2_extent_rec_contig(struct super_block *sb,
 826                                 struct ocfs2_extent_rec *ext,
 827                                 struct ocfs2_extent_rec *insert_rec)
 828 {
 829         u64 blkno = le64_to_cpu(insert_rec->e_blkno);
 830
 831         /*
 832          * Refuse to coalesce extent records with different flag
 833          * fields - we don't want to mix unwritten extents with user
 834          * data.
 835          */
 836         if (ext->e_flags != insert_rec->e_flags)
 837                 return CONTIG_NONE;
 838
 839         if (ocfs2_extents_adjacent(ext, insert_rec) &&
 840             ocfs2_block_extent_contig(sb, ext, blkno))
 841                         return CONTIG_RIGHT;
 842
 843         blkno = le64_to_cpu(ext->e_blkno);
 844         if (ocfs2_extents_adjacent(insert_rec, ext) &&
 845             ocfs2_block_extent_contig(sb, insert_rec, blkno))
 846                 return CONTIG_LEFT;
 847
 848         return CONTIG_NONE;
 849 }
 850
 851 /*
 852  * NOTE: We can have pretty much any combination of contiguousness and
 853  * appending.
 854  *
 855  * The usefulness of APPEND_TAIL is more in that it lets us know that
 856  * we'll have to update the path to that leaf.
 857  */
 858 enum ocfs2_append_type {
 859         APPEND_NONE = 0,
 860         APPEND_TAIL,
 861 };
 862
 863 enum ocfs2_split_type {
 864         SPLIT_NONE = 0,
 865         SPLIT_LEFT,
 866         SPLIT_RIGHT,
 867 };
 868
 869 struct ocfs2_insert_type {
 870         enum ocfs2_split_type   ins_split;
 871         enum ocfs2_append_type  ins_appending;
 872         enum ocfs2_contig_type  ins_contig;
 873         int                     ins_contig_index;
 874         int                     ins_tree_depth;
 875 };
 876
 877 struct ocfs2_merge_ctxt {
 878         enum ocfs2_contig_type  c_contig_type;
 879         int                     c_has_empty_extent;
 880         int                     c_split_covers_rec;
 881 };
 882
 883 static int ocfs2_validate_extent_block(struct super_block *sb,
 884                                        struct buffer_head *bh)
 885 {
 886         int rc;
 887         struct ocfs2_extent_block *eb =
 888                 (struct ocfs2_extent_block *)bh->b_data;
 889
 890         trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
 891
 892         BUG_ON(!buffer_uptodate(bh));
 893
 894         /*
 895          * If the ecc fails, we return the error but otherwise
 896          * leave the filesystem running.  We know any error is
 897          * local to this block.
 898          */
 899         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
 900         if (rc) {
 901                 mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
 902                      (unsigned long long)bh->b_blocknr);
 903                 return rc;
 904         }
 905
 906         /*
 907          * Errors after here are fatal.
 908          */
 909
 910         if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
 911                 ocfs2_error(sb,
 912                             "Extent block #%llu has bad signature %.*s",
 913                             (unsigned long long)bh->b_blocknr, 7,
 914                             eb->h_signature);
 915                 return -EINVAL;
 916         }
 917
 918         if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
 919                 ocfs2_error(sb,
 920                             "Extent block #%llu has an invalid h_blkno "
 921                             "of %llu",
 922                             (unsigned long long)bh->b_blocknr,
 923                             (unsigned long long)le64_to_cpu(eb->h_blkno));
 924                 return -EINVAL;
 925         }
 926
 927         if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) {
 928                 ocfs2_error(sb,
 929                             "Extent block #%llu has an invalid "
 930                             "h_fs_generation of #%u",
 931                             (unsigned long long)bh->b_blocknr,
 932                             le32_to_cpu(eb->h_fs_generation));
 933                 return -EINVAL;
 934         }
 935
 936         return 0;
 937 }
 938
 939 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
 940                             struct buffer_head **bh)
 941 {
 942         int rc;
 943         struct buffer_head *tmp = *bh;
 944
 945         rc = ocfs2_read_block(ci, eb_blkno, &tmp,
 946                               ocfs2_validate_extent_block);
 947
 948         /* If ocfs2_read_block() got us a new bh, pass it up. */
 949         if (!rc && !*bh)
 950                 *bh = tmp;
 951
 952         return rc;
 953 }
 954
 955
 956 /*
 957  * How many free extents have we got before we need more meta data?
 958  */
 959 int ocfs2_num_free_extents(struct ocfs2_super *osb,
 960                            struct ocfs2_extent_tree *et)
 961 {
 962         int retval;
 963         struct ocfs2_extent_list *el = NULL;
 964         struct ocfs2_extent_block *eb;
 965         struct buffer_head *eb_bh = NULL;
 966         u64 last_eb_blk = 0;
 967
 968         el = et->et_root_el;
 969         last_eb_blk = ocfs2_et_get_last_eb_blk(et);
 970
 971         if (last_eb_blk) {
 972                 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
 973                                                  &eb_bh);
 974                 if (retval < 0) {
 975                         mlog_errno(retval);
 976                         goto bail;
 977                 }
 978                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 979                 el = &eb->h_list;
 980         }
 981
 982         BUG_ON(el->l_tree_depth != 0);
 983
 984         retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
 985 bail:
 986         brelse(eb_bh);
 987
 988         trace_ocfs2_num_free_extents(retval);
 989         return retval;
 990 }
 991
 992 /* expects array to already be allocated
 993  *
 994  * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
 995  * l_count for you
 996  */
 997 static int ocfs2_create_new_meta_bhs(handle_t *handle,
 998                                      struct ocfs2_extent_tree *et,
 999                                      int wanted,
1000                                      struct ocfs2_alloc_context *meta_ac,
1001                                      struct buffer_head *bhs[])
1002 {
1003         int count, status, i;
1004         u16 suballoc_bit_start;
1005         u32 num_got;
1006         u64 suballoc_loc, first_blkno;
1007         struct ocfs2_super *osb =
1008                 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
1009         struct ocfs2_extent_block *eb;
1010
1011         count = 0;
1012         while (count < wanted) {
1013                 status = ocfs2_claim_metadata(handle,
1014                                               meta_ac,
1015                                               wanted - count,
1016                                               &suballoc_loc,
1017                                               &suballoc_bit_start,
1018                                               &num_got,
1019                                               &first_blkno);
1020                 if (status < 0) {
1021                         mlog_errno(status);
1022                         goto bail;
1023                 }
1024
1025                 for(i = count;  i < (num_got + count); i++) {
1026                         bhs[i] = sb_getblk(osb->sb, first_blkno);
1027                         if (bhs[i] == NULL) {
1028                                 status = -ENOMEM;
1029                                 mlog_errno(status);
1030                                 goto bail;
1031                         }
1032                         ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1033
1034                         status = ocfs2_journal_access_eb(handle, et->et_ci,
1035                                                          bhs[i],
1036                                                          OCFS2_JOURNAL_ACCESS_CREATE);
1037                         if (status < 0) {
1038                                 mlog_errno(status);
1039                                 goto bail;
1040                         }
1041
1042                         memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1043                         eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1044                         /* Ok, setup the minimal stuff here. */
1045                         strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1046                         eb->h_blkno = cpu_to_le64(first_blkno);
1047                         eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1048                         eb->h_suballoc_slot =
1049                                 cpu_to_le16(meta_ac->ac_alloc_slot);
1050                         eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1051                         eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1052                         eb->h_list.l_count =
1053                                 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1054
1055                         suballoc_bit_start++;
1056                         first_blkno++;
1057
1058                         /* We'll also be dirtied by the caller, so
1059                          * this isn't absolutely necessary. */
1060                         ocfs2_journal_dirty(handle, bhs[i]);
1061                 }
1062
1063                 count += num_got;
1064         }
1065
1066         status = 0;
1067 bail:
1068         if (status < 0) {
1069                 for(i = 0; i < wanted; i++) {
1070                         brelse(bhs[i]);
1071                         bhs[i] = NULL;
1072                 }
1073                 mlog_errno(status);
1074         }
1075         return status;
1076 }
1077
1078 /*
1079  * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1080  *
1081  * Returns the sum of the rightmost extent rec logical offset and
1082  * cluster count.
1083  *
1084  * ocfs2_add_branch() uses this to determine what logical cluster
1085  * value should be populated into the leftmost new branch records.
1086  *
1087  * ocfs2_shift_tree_depth() uses this to determine the # clusters
1088  * value for the new topmost tree record.
1089  */
1090 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list  *el)
1091 {
1092         int i;
1093
1094         i = le16_to_cpu(el->l_next_free_rec) - 1;
1095
1096         return le32_to_cpu(el->l_recs[i].e_cpos) +
1097                 ocfs2_rec_clusters(el, &el->l_recs[i]);
1098 }
1099
1100 /*
1101  * Change range of the branches in the right most path according to the leaf
1102  * extent block's rightmost record.
1103  */
1104 static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1105                                          struct ocfs2_extent_tree *et)
1106 {
1107         int status;
1108         struct ocfs2_path *path = NULL;
1109         struct ocfs2_extent_list *el;
1110         struct ocfs2_extent_rec *rec;
1111
1112         path = ocfs2_new_path_from_et(et);
1113         if (!path) {
1114                 status = -ENOMEM;
1115                 return status;
1116         }
1117
1118         status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1119         if (status < 0) {
1120                 mlog_errno(status);
1121                 goto out;
1122         }
1123
1124         status = ocfs2_extend_trans(handle, path_num_items(path));
1125         if (status < 0) {
1126                 mlog_errno(status);
1127                 goto out;
1128         }
1129
1130         status = ocfs2_journal_access_path(et->et_ci, handle, path);
1131         if (status < 0) {
1132                 mlog_errno(status);
1133                 goto out;
1134         }
1135
1136         el = path_leaf_el(path);
1137         rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1138
1139         ocfs2_adjust_rightmost_records(handle, et, path, rec);
1140
1141 out:
1142         ocfs2_free_path(path);
1143         return status;
1144 }
1145
1146 /*
1147  * Add an entire tree branch to our inode. eb_bh is the extent block
1148  * to start at, if we don't want to start the branch at the root
1149  * structure.
1150  *
1151  * last_eb_bh is required as we have to update it's next_leaf pointer
1152  * for the new last extent block.
1153  *
1154  * the new branch will be 'empty' in the sense that every block will
1155  * contain a single record with cluster count == 0.
1156  */
1157 static int ocfs2_add_branch(handle_t *handle,
1158                             struct ocfs2_extent_tree *et,
1159                             struct buffer_head *eb_bh,
1160                             struct buffer_head **last_eb_bh,
1161                             struct ocfs2_alloc_context *meta_ac)
1162 {
1163         int status, new_blocks, i;
1164         u64 next_blkno, new_last_eb_blk;
1165         struct buffer_head *bh;
1166         struct buffer_head **new_eb_bhs = NULL;
1167         struct ocfs2_extent_block *eb;
1168         struct ocfs2_extent_list  *eb_el;
1169         struct ocfs2_extent_list  *el;
1170         u32 new_cpos, root_end;
1171
1172         BUG_ON(!last_eb_bh || !*last_eb_bh);
1173
1174         if (eb_bh) {
1175                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1176                 el = &eb->h_list;
1177         } else
1178                 el = et->et_root_el;
1179
1180         /* we never add a branch to a leaf. */
1181         BUG_ON(!el->l_tree_depth);
1182
1183         new_blocks = le16_to_cpu(el->l_tree_depth);
1184
1185         eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1186         new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1187         root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1188
1189         /*
1190          * If there is a gap before the root end and the real end
1191          * of the righmost leaf block, we need to remove the gap
1192          * between new_cpos and root_end first so that the tree
1193          * is consistent after we add a new branch(it will start
1194          * from new_cpos).
1195          */
1196         if (root_end > new_cpos) {
1197                 trace_ocfs2_adjust_rightmost_branch(
1198                         (unsigned long long)
1199                         ocfs2_metadata_cache_owner(et->et_ci),
1200                         root_end, new_cpos);
1201
1202                 status = ocfs2_adjust_rightmost_branch(handle, et);
1203                 if (status) {
1204                         mlog_errno(status);
1205                         goto bail;
1206                 }
1207         }
1208
1209         /* allocate the number of new eb blocks we need */
1210         new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1211                              GFP_KERNEL);
1212         if (!new_eb_bhs) {
1213                 status = -ENOMEM;
1214                 mlog_errno(status);
1215                 goto bail;
1216         }
1217
1218         status = ocfs2_create_new_meta_bhs(handle, et, new_blocks,
1219                                            meta_ac, new_eb_bhs);
1220         if (status < 0) {
1221                 mlog_errno(status);
1222                 goto bail;
1223         }
1224
1225         /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1226          * linked with the rest of the tree.
1227          * conversly, new_eb_bhs[0] is the new bottommost leaf.
1228          *
1229          * when we leave the loop, new_last_eb_blk will point to the
1230          * newest leaf, and next_blkno will point to the topmost extent
1231          * block. */
1232         next_blkno = new_last_eb_blk = 0;
1233         for(i = 0; i < new_blocks; i++) {
1234                 bh = new_eb_bhs[i];
1235                 eb = (struct ocfs2_extent_block *) bh->b_data;
1236                 /* ocfs2_create_new_meta_bhs() should create it right! */
1237                 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1238                 eb_el = &eb->h_list;
1239
1240                 status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1241                                                  OCFS2_JOURNAL_ACCESS_CREATE);
1242                 if (status < 0) {
1243                         mlog_errno(status);
1244                         goto bail;
1245                 }
1246
1247                 eb->h_next_leaf_blk = 0;
1248                 eb_el->l_tree_depth = cpu_to_le16(i);
1249                 eb_el->l_next_free_rec = cpu_to_le16(1);
1250                 /*
1251                  * This actually counts as an empty extent as
1252                  * c_clusters == 0
1253                  */
1254                 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1255                 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1256                 /*
1257                  * eb_el isn't always an interior node, but even leaf
1258                  * nodes want a zero'd flags and reserved field so
1259                  * this gets the whole 32 bits regardless of use.
1260                  */
1261                 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1262                 if (!eb_el->l_tree_depth)
1263                         new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1264
1265                 ocfs2_journal_dirty(handle, bh);
1266                 next_blkno = le64_to_cpu(eb->h_blkno);
1267         }
1268
1269         /* This is a bit hairy. We want to update up to three blocks
1270          * here without leaving any of them in an inconsistent state
1271          * in case of error. We don't have to worry about
1272          * journal_dirty erroring as it won't unless we've aborted the
1273          * handle (in which case we would never be here) so reserving
1274          * the write with journal_access is all we need to do. */
1275         status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1276                                          OCFS2_JOURNAL_ACCESS_WRITE);
1277         if (status < 0) {
1278                 mlog_errno(status);
1279                 goto bail;
1280         }
1281         status = ocfs2_et_root_journal_access(handle, et,
1282                                               OCFS2_JOURNAL_ACCESS_WRITE);
1283         if (status < 0) {
1284                 mlog_errno(status);
1285                 goto bail;
1286         }
1287         if (eb_bh) {
1288                 status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1289                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1290                 if (status < 0) {
1291                         mlog_errno(status);
1292                         goto bail;
1293                 }
1294         }
1295
1296         /* Link the new branch into the rest of the tree (el will
1297          * either be on the root_bh, or the extent block passed in. */
1298         i = le16_to_cpu(el->l_next_free_rec);
1299         el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1300         el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1301         el->l_recs[i].e_int_clusters = 0;
1302         le16_add_cpu(&el->l_next_free_rec, 1);
1303
1304         /* fe needs a new last extent block pointer, as does the
1305          * next_leaf on the previously last-extent-block. */
1306         ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1307
1308         eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1309         eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1310
1311         ocfs2_journal_dirty(handle, *last_eb_bh);
1312         ocfs2_journal_dirty(handle, et->et_root_bh);
1313         if (eb_bh)
1314                 ocfs2_journal_dirty(handle, eb_bh);
1315
1316         /*
1317          * Some callers want to track the rightmost leaf so pass it
1318          * back here.
1319          */
1320         brelse(*last_eb_bh);
1321         get_bh(new_eb_bhs[0]);
1322         *last_eb_bh = new_eb_bhs[0];
1323
1324         status = 0;
1325 bail:
1326         if (new_eb_bhs) {
1327                 for (i = 0; i < new_blocks; i++)
1328                         brelse(new_eb_bhs[i]);
1329                 kfree(new_eb_bhs);
1330         }
1331
1332         return status;
1333 }
1334
1335 /*
1336  * adds another level to the allocation tree.
1337  * returns back the new extent block so you can add a branch to it
1338  * after this call.
1339  */
1340 static int ocfs2_shift_tree_depth(handle_t *handle,
1341                                   struct ocfs2_extent_tree *et,
1342                                   struct ocfs2_alloc_context *meta_ac,
1343                                   struct buffer_head **ret_new_eb_bh)
1344 {
1345         int status, i;
1346         u32 new_clusters;
1347         struct buffer_head *new_eb_bh = NULL;
1348         struct ocfs2_extent_block *eb;
1349         struct ocfs2_extent_list  *root_el;
1350         struct ocfs2_extent_list  *eb_el;
1351
1352         status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1353                                            &new_eb_bh);
1354         if (status < 0) {
1355                 mlog_errno(status);
1356                 goto bail;
1357         }
1358
1359         eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1360         /* ocfs2_create_new_meta_bhs() should create it right! */
1361         BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1362
1363         eb_el = &eb->h_list;
1364         root_el = et->et_root_el;
1365
1366         status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1367                                          OCFS2_JOURNAL_ACCESS_CREATE);
1368         if (status < 0) {
1369                 mlog_errno(status);
1370                 goto bail;
1371         }
1372
1373         /* copy the root extent list data into the new extent block */
1374         eb_el->l_tree_depth = root_el->l_tree_depth;
1375         eb_el->l_next_free_rec = root_el->l_next_free_rec;
1376         for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1377                 eb_el->l_recs[i] = root_el->l_recs[i];
1378
1379         ocfs2_journal_dirty(handle, new_eb_bh);
1380
1381         status = ocfs2_et_root_journal_access(handle, et,
1382                                               OCFS2_JOURNAL_ACCESS_WRITE);
1383         if (status < 0) {
1384                 mlog_errno(status);
1385                 goto bail;
1386         }
1387
1388         new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1389
1390         /* update root_bh now */
1391         le16_add_cpu(&root_el->l_tree_depth, 1);
1392         root_el->l_recs[0].e_cpos = 0;
1393         root_el->l_recs[0].e_blkno = eb->h_blkno;
1394         root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1395         for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1396                 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1397         root_el->l_next_free_rec = cpu_to_le16(1);
1398
1399         /* If this is our 1st tree depth shift, then last_eb_blk
1400          * becomes the allocated extent block */
1401         if (root_el->l_tree_depth == cpu_to_le16(1))
1402                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1403
1404         ocfs2_journal_dirty(handle, et->et_root_bh);
1405
1406         *ret_new_eb_bh = new_eb_bh;
1407         new_eb_bh = NULL;
1408         status = 0;
1409 bail:
1410         brelse(new_eb_bh);
1411
1412         return status;
1413 }
1414
1415 /*
1416  * Should only be called when there is no space left in any of the
1417  * leaf nodes. What we want to do is find the lowest tree depth
1418  * non-leaf extent block with room for new records. There are three
1419  * valid results of this search:
1420  *
1421  * 1) a lowest extent block is found, then we pass it back in
1422  *    *lowest_eb_bh and return '0'
1423  *
1424  * 2) the search fails to find anything, but the root_el has room. We
1425  *    pass NULL back in *lowest_eb_bh, but still return '0'
1426  *
1427  * 3) the search fails to find anything AND the root_el is full, in
1428  *    which case we return > 0
1429  *
1430  * return status < 0 indicates an error.
1431  */
1432 static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1433                                     struct buffer_head **target_bh)
1434 {
1435         int status = 0, i;
1436         u64 blkno;
1437         struct ocfs2_extent_block *eb;
1438         struct ocfs2_extent_list  *el;
1439         struct buffer_head *bh = NULL;
1440         struct buffer_head *lowest_bh = NULL;
1441
1442         *target_bh = NULL;
1443
1444         el = et->et_root_el;
1445
1446         while(le16_to_cpu(el->l_tree_depth) > 1) {
1447                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1448                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1449                                     "Owner %llu has empty "
1450                                     "extent list (next_free_rec == 0)",
1451                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1452                         status = -EIO;
1453                         goto bail;
1454                 }
1455                 i = le16_to_cpu(el->l_next_free_rec) - 1;
1456                 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1457                 if (!blkno) {
1458                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1459                                     "Owner %llu has extent "
1460                                     "list where extent # %d has no physical "
1461                                     "block start",
1462                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1463                         status = -EIO;
1464                         goto bail;
1465                 }
1466
1467                 brelse(bh);
1468                 bh = NULL;
1469
1470                 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1471                 if (status < 0) {
1472                         mlog_errno(status);
1473                         goto bail;
1474                 }
1475
1476                 eb = (struct ocfs2_extent_block *) bh->b_data;
1477                 el = &eb->h_list;
1478
1479                 if (le16_to_cpu(el->l_next_free_rec) <
1480                     le16_to_cpu(el->l_count)) {
1481                         brelse(lowest_bh);
1482                         lowest_bh = bh;
1483                         get_bh(lowest_bh);
1484                 }
1485         }
1486
1487         /* If we didn't find one and the fe doesn't have any room,
1488          * then return '1' */
1489         el = et->et_root_el;
1490         if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1491                 status = 1;
1492
1493         *target_bh = lowest_bh;
1494 bail:
1495         brelse(bh);
1496
1497         return status;
1498 }
1499
1500 /*
1501  * Grow a b-tree so that it has more records.
1502  *
1503  * We might shift the tree depth in which case existing paths should
1504  * be considered invalid.
1505  *
1506  * Tree depth after the grow is returned via *final_depth.
1507  *
1508  * *last_eb_bh will be updated by ocfs2_add_branch().
1509  */
1510 static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1511                            int *final_depth, struct buffer_head **last_eb_bh,
1512                            struct ocfs2_alloc_context *meta_ac)
1513 {
1514         int ret, shift;
1515         struct ocfs2_extent_list *el = et->et_root_el;
1516         int depth = le16_to_cpu(el->l_tree_depth);
1517         struct buffer_head *bh = NULL;
1518
1519         BUG_ON(meta_ac == NULL);
1520
1521         shift = ocfs2_find_branch_target(et, &bh);
1522         if (shift < 0) {
1523                 ret = shift;
1524                 mlog_errno(ret);
1525                 goto out;
1526         }
1527
1528         /* We traveled all the way to the bottom of the allocation tree
1529          * and didn't find room for any more extents - we need to add
1530          * another tree level */
1531         if (shift) {
1532                 BUG_ON(bh);
1533                 trace_ocfs2_grow_tree(
1534                         (unsigned long long)
1535                         ocfs2_metadata_cache_owner(et->et_ci),
1536                         depth);
1537
1538                 /* ocfs2_shift_tree_depth will return us a buffer with
1539                  * the new extent block (so we can pass that to
1540                  * ocfs2_add_branch). */
1541                 ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1542                 if (ret < 0) {
1543                         mlog_errno(ret);
1544                         goto out;
1545                 }
1546                 depth++;
1547                 if (depth == 1) {
1548                         /*
1549                          * Special case: we have room now if we shifted from
1550                          * tree_depth 0, so no more work needs to be done.
1551                          *
1552                          * We won't be calling add_branch, so pass
1553                          * back *last_eb_bh as the new leaf. At depth
1554                          * zero, it should always be null so there's
1555                          * no reason to brelse.
1556                          */
1557                         BUG_ON(*last_eb_bh);
1558                         get_bh(bh);
1559                         *last_eb_bh = bh;
1560                         goto out;
1561                 }
1562         }
1563
1564         /* call ocfs2_add_branch to add the final part of the tree with
1565          * the new data. */
1566         ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1567                                meta_ac);
1568         if (ret < 0) {
1569                 mlog_errno(ret);
1570                 goto out;
1571         }
1572
1573 out:
1574         if (final_depth)
1575                 *final_depth = depth;
1576         brelse(bh);
1577         return ret;
1578 }
1579
1580 /*
1581  * This function will discard the rightmost extent record.
1582  */
1583 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1584 {
1585         int next_free = le16_to_cpu(el->l_next_free_rec);
1586         int count = le16_to_cpu(el->l_count);
1587         unsigned int num_bytes;
1588
1589         BUG_ON(!next_free);
1590         /* This will cause us to go off the end of our extent list. */
1591         BUG_ON(next_free >= count);
1592
1593         num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1594
1595         memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1596 }
1597
1598 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1599                               struct ocfs2_extent_rec *insert_rec)
1600 {
1601         int i, insert_index, next_free, has_empty, num_bytes;
1602         u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1603         struct ocfs2_extent_rec *rec;
1604
1605         next_free = le16_to_cpu(el->l_next_free_rec);
1606         has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1607
1608         BUG_ON(!next_free);
1609
1610         /* The tree code before us didn't allow enough room in the leaf. */
1611         BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1612
1613         /*
1614          * The easiest way to approach this is to just remove the
1615          * empty extent and temporarily decrement next_free.
1616          */
1617         if (has_empty) {
1618                 /*
1619                  * If next_free was 1 (only an empty extent), this
1620                  * loop won't execute, which is fine. We still want
1621                  * the decrement above to happen.
1622                  */
1623                 for(i = 0; i < (next_free - 1); i++)
1624                         el->l_recs[i] = el->l_recs[i+1];
1625
1626                 next_free--;
1627         }
1628
1629         /*
1630          * Figure out what the new record index should be.
1631          */
1632         for(i = 0; i < next_free; i++) {
1633                 rec = &el->l_recs[i];
1634
1635                 if (insert_cpos < le32_to_cpu(rec->e_cpos))
1636                         break;
1637         }
1638         insert_index = i;
1639
1640         trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1641                                 has_empty, next_free,
1642                                 le16_to_cpu(el->l_count));
1643
1644         BUG_ON(insert_index < 0);
1645         BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1646         BUG_ON(insert_index > next_free);
1647
1648         /*
1649          * No need to memmove if we're just adding to the tail.
1650          */
1651         if (insert_index != next_free) {
1652                 BUG_ON(next_free >= le16_to_cpu(el->l_count));
1653
1654                 num_bytes = next_free - insert_index;
1655                 num_bytes *= sizeof(struct ocfs2_extent_rec);
1656                 memmove(&el->l_recs[insert_index + 1],
1657                         &el->l_recs[insert_index],
1658                         num_bytes);
1659         }
1660
1661         /*
1662          * Either we had an empty extent, and need to re-increment or
1663          * there was no empty extent on a non full rightmost leaf node,
1664          * in which case we still need to increment.
1665          */
1666         next_free++;
1667         el->l_next_free_rec = cpu_to_le16(next_free);
1668         /*
1669          * Make sure none of the math above just messed up our tree.
1670          */
1671         BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1672
1673         el->l_recs[insert_index] = *insert_rec;
1674
1675 }
1676
1677 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1678 {
1679         int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1680
1681         BUG_ON(num_recs == 0);
1682
1683         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1684                 num_recs--;
1685                 size = num_recs * sizeof(struct ocfs2_extent_rec);
1686                 memmove(&el->l_recs[0], &el->l_recs[1], size);
1687                 memset(&el->l_recs[num_recs], 0,
1688                        sizeof(struct ocfs2_extent_rec));
1689                 el->l_next_free_rec = cpu_to_le16(num_recs);
1690         }
1691 }
1692
1693 /*
1694  * Create an empty extent record .
1695  *
1696  * l_next_free_rec may be updated.
1697  *
1698  * If an empty extent already exists do nothing.
1699  */
1700 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1701 {
1702         int next_free = le16_to_cpu(el->l_next_free_rec);
1703
1704         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1705
1706         if (next_free == 0)
1707                 goto set_and_inc;
1708
1709         if (ocfs2_is_empty_extent(&el->l_recs[0]))
1710                 return;
1711
1712         mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1713                         "Asked to create an empty extent in a full list:\n"
1714                         "count = %u, tree depth = %u",
1715                         le16_to_cpu(el->l_count),
1716                         le16_to_cpu(el->l_tree_depth));
1717
1718         ocfs2_shift_records_right(el);
1719
1720 set_and_inc:
1721         le16_add_cpu(&el->l_next_free_rec, 1);
1722         memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1723 }
1724
1725 /*
1726  * For a rotation which involves two leaf nodes, the "root node" is
1727  * the lowest level tree node which contains a path to both leafs. This
1728  * resulting set of information can be used to form a complete "subtree"
1729  *
1730  * This function is passed two full paths from the dinode down to a
1731  * pair of adjacent leaves. It's task is to figure out which path
1732  * index contains the subtree root - this can be the root index itself
1733  * in a worst-case rotation.
1734  *
1735  * The array index of the subtree root is passed back.
1736  */
1737 int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1738                             struct ocfs2_path *left,
1739                             struct ocfs2_path *right)
1740 {
1741         int i = 0;
1742
1743         /*
1744          * Check that the caller passed in two paths from the same tree.
1745          */
1746         BUG_ON(path_root_bh(left) != path_root_bh(right));
1747
1748         do {
1749                 i++;
1750
1751                 /*
1752                  * The caller didn't pass two adjacent paths.
1753                  */
1754                 mlog_bug_on_msg(i > left->p_tree_depth,
1755                                 "Owner %llu, left depth %u, right depth %u\n"
1756                                 "left leaf blk %llu, right leaf blk %llu\n",
1757                                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1758                                 left->p_tree_depth, right->p_tree_depth,
1759                                 (unsigned long long)path_leaf_bh(left)->b_blocknr,
1760                                 (unsigned long long)path_leaf_bh(right)->b_blocknr);
1761         } while (left->p_node[i].bh->b_blocknr ==
1762                  right->p_node[i].bh->b_blocknr);
1763
1764         return i - 1;
1765 }
1766
1767 typedef void (path_insert_t)(void *, struct buffer_head *);
1768
1769 /*
1770  * Traverse a btree path in search of cpos, starting at root_el.
1771  *
1772  * This code can be called with a cpos larger than the tree, in which
1773  * case it will return the rightmost path.
1774  */
1775 static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1776                              struct ocfs2_extent_list *root_el, u32 cpos,
1777                              path_insert_t *func, void *data)
1778 {
1779         int i, ret = 0;
1780         u32 range;
1781         u64 blkno;
1782         struct buffer_head *bh = NULL;
1783         struct ocfs2_extent_block *eb;
1784         struct ocfs2_extent_list *el;
1785         struct ocfs2_extent_rec *rec;
1786
1787         el = root_el;
1788         while (el->l_tree_depth) {
1789                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
1790                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1791                                     "Owner %llu has empty extent list at "
1792                                     "depth %u\n",
1793                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1794                                     le16_to_cpu(el->l_tree_depth));
1795                         ret = -EROFS;
1796                         goto out;
1797
1798                 }
1799
1800                 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1801                         rec = &el->l_recs[i];
1802
1803                         /*
1804                          * In the case that cpos is off the allocation
1805                          * tree, this should just wind up returning the
1806                          * rightmost record.
1807                          */
1808                         range = le32_to_cpu(rec->e_cpos) +
1809                                 ocfs2_rec_clusters(el, rec);
1810                         if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1811                             break;
1812                 }
1813
1814                 blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1815                 if (blkno == 0) {
1816                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1817                                     "Owner %llu has bad blkno in extent list "
1818                                     "at depth %u (index %d)\n",
1819                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1820                                     le16_to_cpu(el->l_tree_depth), i);
1821                         ret = -EROFS;
1822                         goto out;
1823                 }
1824
1825                 brelse(bh);
1826                 bh = NULL;
1827                 ret = ocfs2_read_extent_block(ci, blkno, &bh);
1828                 if (ret) {
1829                         mlog_errno(ret);
1830                         goto out;
1831                 }
1832
1833                 eb = (struct ocfs2_extent_block *) bh->b_data;
1834                 el = &eb->h_list;
1835
1836                 if (le16_to_cpu(el->l_next_free_rec) >
1837                     le16_to_cpu(el->l_count)) {
1838                         ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1839                                     "Owner %llu has bad count in extent list "
1840                                     "at block %llu (next free=%u, count=%u)\n",
1841                                     (unsigned long long)ocfs2_metadata_cache_owner(ci),
1842                                     (unsigned long long)bh->b_blocknr,
1843                                     le16_to_cpu(el->l_next_free_rec),
1844                                     le16_to_cpu(el->l_count));
1845                         ret = -EROFS;
1846                         goto out;
1847                 }
1848
1849                 if (func)
1850                         func(data, bh);
1851         }
1852
1853 out:
1854         /*
1855          * Catch any trailing bh that the loop didn't handle.
1856          */
1857         brelse(bh);
1858
1859         return ret;
1860 }
1861
1862 /*
1863  * Given an initialized path (that is, it has a valid root extent
1864  * list), this function will traverse the btree in search of the path
1865  * which would contain cpos.
1866  *
1867  * The path traveled is recorded in the path structure.
1868  *
1869  * Note that this will not do any comparisons on leaf node extent
1870  * records, so it will work fine in the case that we just added a tree
1871  * branch.
1872  */
1873 struct find_path_data {
1874         int index;
1875         struct ocfs2_path *path;
1876 };
1877 static void find_path_ins(void *data, struct buffer_head *bh)
1878 {
1879         struct find_path_data *fp = data;
1880
1881         get_bh(bh);
1882         ocfs2_path_insert_eb(fp->path, fp->index, bh);
1883         fp->index++;
1884 }
1885 int ocfs2_find_path(struct ocfs2_caching_info *ci,
1886                     struct ocfs2_path *path, u32 cpos)
1887 {
1888         struct find_path_data data;
1889
1890         data.index = 1;
1891         data.path = path;
1892         return __ocfs2_find_path(ci, path_root_el(path), cpos,
1893                                  find_path_ins, &data);
1894 }
1895
1896 static void find_leaf_ins(void *data, struct buffer_head *bh)
1897 {
1898         struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1899         struct ocfs2_extent_list *el = &eb->h_list;
1900         struct buffer_head **ret = data;
1901
1902         /* We want to retain only the leaf block. */
1903         if (le16_to_cpu(el->l_tree_depth) == 0) {
1904                 get_bh(bh);
1905                 *ret = bh;
1906         }
1907 }
1908 /*
1909  * Find the leaf block in the tree which would contain cpos. No
1910  * checking of the actual leaf is done.
1911  *
1912  * Some paths want to call this instead of allocating a path structure
1913  * and calling ocfs2_find_path().
1914  *
1915  * This function doesn't handle non btree extent lists.
1916  */
1917 int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1918                     struct ocfs2_extent_list *root_el, u32 cpos,
1919                     struct buffer_head **leaf_bh)
1920 {
1921         int ret;
1922         struct buffer_head *bh = NULL;
1923
1924         ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1925         if (ret) {
1926                 mlog_errno(ret);
1927                 goto out;
1928         }
1929
1930         *leaf_bh = bh;
1931 out:
1932         return ret;
1933 }
1934
1935 /*
1936  * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1937  *
1938  * Basically, we've moved stuff around at the bottom of the tree and
1939  * we need to fix up the extent records above the changes to reflect
1940  * the new changes.
1941  *
1942  * left_rec: the record on the left.
1943  * left_child_el: is the child list pointed to by left_rec
1944  * right_rec: the record to the right of left_rec
1945  * right_child_el: is the child list pointed to by right_rec
1946  *
1947  * By definition, this only works on interior nodes.
1948  */
1949 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1950                                   struct ocfs2_extent_list *left_child_el,
1951                                   struct ocfs2_extent_rec *right_rec,
1952                                   struct ocfs2_extent_list *right_child_el)
1953 {
1954         u32 left_clusters, right_end;
1955
1956         /*
1957          * Interior nodes never have holes. Their cpos is the cpos of
1958          * the leftmost record in their child list. Their cluster
1959          * count covers the full theoretical range of their child list
1960          * - the range between their cpos and the cpos of the record
1961          * immediately to their right.
1962          */
1963         left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1964         if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1965                 BUG_ON(right_child_el->l_tree_depth);
1966                 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1967                 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1968         }
1969         left_clusters -= le32_to_cpu(left_rec->e_cpos);
1970         left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1971
1972         /*
1973          * Calculate the rightmost cluster count boundary before
1974          * moving cpos - we will need to adjust clusters after
1975          * updating e_cpos to keep the same highest cluster count.
1976          */
1977         right_end = le32_to_cpu(right_rec->e_cpos);
1978         right_end += le32_to_cpu(right_rec->e_int_clusters);
1979
1980         right_rec->e_cpos = left_rec->e_cpos;
1981         le32_add_cpu(&right_rec->e_cpos, left_clusters);
1982
1983         right_end -= le32_to_cpu(right_rec->e_cpos);
1984         right_rec->e_int_clusters = cpu_to_le32(right_end);
1985 }
1986
1987 /*
1988  * Adjust the adjacent root node records involved in a
1989  * rotation. left_el_blkno is passed in as a key so that we can easily
1990  * find it's index in the root list.
1991  */
1992 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1993                                       struct ocfs2_extent_list *left_el,
1994                                       struct ocfs2_extent_list *right_el,
1995                                       u64 left_el_blkno)
1996 {
1997         int i;
1998
1999         BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2000                le16_to_cpu(left_el->l_tree_depth));
2001
2002         for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2003                 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2004                         break;
2005         }
2006
2007         /*
2008          * The path walking code should have never returned a root and
2009          * two paths which are not adjacent.
2010          */
2011         BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2012
2013         ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el,
2014                                       &root_el->l_recs[i + 1], right_el);
2015 }
2016
2017 /*
2018  * We've changed a leaf block (in right_path) and need to reflect that
2019  * change back up the subtree.
2020  *
2021  * This happens in multiple places:
2022  *   - When we've moved an extent record from the left path leaf to the right
2023  *     path leaf to make room for an empty extent in the left path leaf.
2024  *   - When our insert into the right path leaf is at the leftmost edge
2025  *     and requires an update of the path immediately to it's left. This
2026  *     can occur at the end of some types of rotation and appending inserts.
2027  *   - When we've adjusted the last extent record in the left path leaf and the
2028  *     1st extent record in the right path leaf during cross extent block merge.
2029  */
2030 static void ocfs2_complete_edge_insert(handle_t *handle,
2031                                        struct ocfs2_path *left_path,
2032                                        struct ocfs2_path *right_path,
2033                                        int subtree_index)
2034 {
2035         int i, idx;
2036         struct ocfs2_extent_list *el, *left_el, *right_el;
2037         struct ocfs2_extent_rec *left_rec, *right_rec;
2038         struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2039
2040         /*
2041          * Update the counts and position values within all the
2042          * interior nodes to reflect the leaf rotation we just did.
2043          *
2044          * The root node is handled below the loop.
2045          *
2046          * We begin the loop with right_el and left_el pointing to the
2047          * leaf lists and work our way up.
2048          *
2049          * NOTE: within this loop, left_el and right_el always refer
2050          * to the *child* lists.
2051          */
2052         left_el = path_leaf_el(left_path);
2053         right_el = path_leaf_el(right_path);
2054         for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2055                 trace_ocfs2_complete_edge_insert(i);
2056
2057                 /*
2058                  * One nice property of knowing that all of these
2059                  * nodes are below the root is that we only deal with
2060                  * the leftmost right node record and the rightmost
2061                  * left node record.
2062                  */
2063                 el = left_path->p_node[i].el;
2064                 idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2065                 left_rec = &el->l_recs[idx];
2066
2067                 el = right_path->p_node[i].el;
2068                 right_rec = &el->l_recs[0];
2069
2070                 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec,
2071                                               right_el);
2072
2073                 ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2074                 ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2075
2076                 /*
2077                  * Setup our list pointers now so that the current
2078                  * parents become children in the next iteration.
2079                  */
2080                 left_el = left_path->p_node[i].el;
2081                 right_el = right_path->p_node[i].el;
2082         }
2083
2084         /*
2085          * At the root node, adjust the two adjacent records which
2086          * begin our path to the leaves.
2087          */
2088
2089         el = left_path->p_node[subtree_index].el;
2090         left_el = left_path->p_node[subtree_index + 1].el;
2091         right_el = right_path->p_node[subtree_index + 1].el;
2092
2093         ocfs2_adjust_root_records(el, left_el, right_el,
2094                                   left_path->p_node[subtree_index + 1].bh->b_blocknr);
2095
2096         root_bh = left_path->p_node[subtree_index].bh;
2097
2098         ocfs2_journal_dirty(handle, root_bh);
2099 }
2100
2101 static int ocfs2_rotate_subtree_right(handle_t *handle,
2102                                       struct ocfs2_extent_tree *et,
2103                                       struct ocfs2_path *left_path,
2104                                       struct ocfs2_path *right_path,
2105                                       int subtree_index)
2106 {
2107         int ret, i;
2108         struct buffer_head *right_leaf_bh;
2109         struct buffer_head *left_leaf_bh = NULL;
2110         struct buffer_head *root_bh;
2111         struct ocfs2_extent_list *right_el, *left_el;
2112         struct ocfs2_extent_rec move_rec;
2113
2114         left_leaf_bh = path_leaf_bh(left_path);
2115         left_el = path_leaf_el(left_path);
2116
2117         if (left_el->l_next_free_rec != left_el->l_count) {
2118                 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2119                             "Inode %llu has non-full interior leaf node %llu"
2120                             "(next free = %u)",
2121                             (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2122                             (unsigned long long)left_leaf_bh->b_blocknr,
2123                             le16_to_cpu(left_el->l_next_free_rec));
2124                 return -EROFS;
2125         }
2126
2127         /*
2128          * This extent block may already have an empty record, so we
2129          * return early if so.
2130          */
2131         if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2132                 return 0;
2133
2134         root_bh = left_path->p_node[subtree_index].bh;
2135         BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2136
2137         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2138                                            subtree_index);
2139         if (ret) {
2140                 mlog_errno(ret);
2141                 goto out;
2142         }
2143
2144         for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2145                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2146                                                    right_path, i);
2147                 if (ret) {
2148                         mlog_errno(ret);
2149                         goto out;
2150                 }
2151
2152                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2153                                                    left_path, i);
2154                 if (ret) {
2155                         mlog_errno(ret);
2156                         goto out;
2157                 }
2158         }
2159
2160         right_leaf_bh = path_leaf_bh(right_path);
2161         right_el = path_leaf_el(right_path);
2162
2163         /* This is a code error, not a disk corruption. */
2164         mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2165                         "because rightmost leaf block %llu is empty\n",
2166                         (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2167                         (unsigned long long)right_leaf_bh->b_blocknr);
2168
2169         ocfs2_create_empty_extent(right_el);
2170
2171         ocfs2_journal_dirty(handle, right_leaf_bh);
2172
2173         /* Do the copy now. */
2174         i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2175         move_rec = left_el->l_recs[i];
2176         right_el->l_recs[0] = move_rec;
2177
2178         /*
2179          * Clear out the record we just copied and shift everything
2180          * over, leaving an empty extent in the left leaf.
2181          *
2182          * We temporarily subtract from next_free_rec so that the
2183          * shift will lose the tail record (which is now defunct).
2184          */
2185         le16_add_cpu(&left_el->l_next_free_rec, -1);
2186         ocfs2_shift_records_right(left_el);
2187         memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2188         le16_add_cpu(&left_el->l_next_free_rec, 1);
2189
2190         ocfs2_journal_dirty(handle, left_leaf_bh);
2191
2192         ocfs2_complete_edge_insert(handle, left_path, right_path,
2193                                    subtree_index);
2194
2195 out:
2196         return ret;
2197 }
2198
2199 /*
2200  * Given a full path, determine what cpos value would return us a path
2201  * containing the leaf immediately to the left of the current one.
2202  *
2203  * Will return zero if the path passed in is already the leftmost path.
2204  */
2205 int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2206                                   struct ocfs2_path *path, u32 *cpos)
2207 {
2208         int i, j, ret = 0;
2209         u64 blkno;
2210         struct ocfs2_extent_list *el;
2211
2212         BUG_ON(path->p_tree_depth == 0);
2213
2214         *cpos = 0;
2215
2216         blkno = path_leaf_bh(path)->b_blocknr;
2217
2218         /* Start at the tree node just above the leaf and work our way up. */
2219         i = path->p_tree_depth - 1;
2220         while (i >= 0) {
2221                 el = path->p_node[i].el;
2222
2223                 /*
2224                  * Find the extent record just before the one in our
2225                  * path.
2226                  */
2227                 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2228                         if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2229                                 if (j == 0) {
2230                                         if (i == 0) {
2231                                                 /*
2232                                                  * We've determined that the
2233                                                  * path specified is already
2234                                                  * the leftmost one - return a
2235                                                  * cpos of zero.
2236                                                  */
2237                                                 goto out;
2238                                         }
2239                                         /*
2240                                          * The leftmost record points to our
2241                                          * leaf - we need to travel up the
2242                                          * tree one level.
2243                                          */
2244                                         goto next_node;
2245                                 }
2246
2247                                 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2248                                 *cpos = *cpos + ocfs2_rec_clusters(el,
2249                                                            &el->l_recs[j - 1]);
2250                                 *cpos = *cpos - 1;
2251                                 goto out;
2252                         }
2253                 }
2254
2255                 /*
2256                  * If we got here, we never found a valid node where
2257                  * the tree indicated one should be.
2258                  */
2259                 ocfs2_error(sb,
2260                             "Invalid extent tree at extent block %llu\n",
2261                             (unsigned long long)blkno);
2262                 ret = -EROFS;
2263                 goto out;
2264
2265 next_node:
2266                 blkno = path->p_node[i].bh->b_blocknr;
2267                 i--;
2268         }
2269
2270 out:
2271         return ret;
2272 }
2273
2274 /*
2275  * Extend the transaction by enough credits to complete the rotation,
2276  * and still leave at least the original number of credits allocated
2277  * to this transaction.
2278  */
2279 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2280                                            int op_credits,
2281                                            struct ocfs2_path *path)
2282 {
2283         int ret = 0;
2284         int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2285
2286         if (handle->h_buffer_credits < credits)
2287                 ret = ocfs2_extend_trans(handle,
2288                                          credits - handle->h_buffer_credits);
2289
2290         return ret;
2291 }
2292
2293 /*
2294  * Trap the case where we're inserting into the theoretical range past
2295  * the _actual_ left leaf range. Otherwise, we'll rotate a record
2296  * whose cpos is less than ours into the right leaf.
2297  *
2298  * It's only necessary to look at the rightmost record of the left
2299  * leaf because the logic that calls us should ensure that the
2300  * theoretical ranges in the path components above the leaves are
2301  * correct.
2302  */
2303 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2304                                                  u32 insert_cpos)
2305 {
2306         struct ocfs2_extent_list *left_el;
2307         struct ocfs2_extent_rec *rec;
2308         int next_free;
2309
2310         left_el = path_leaf_el(left_path);
2311         next_free = le16_to_cpu(left_el->l_next_free_rec);
2312         rec = &left_el->l_recs[next_free - 1];
2313
2314         if (insert_cpos > le32_to_cpu(rec->e_cpos))
2315                 return 1;
2316         return 0;
2317 }
2318
2319 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2320 {
2321         int next_free = le16_to_cpu(el->l_next_free_rec);
2322         unsigned int range;
2323         struct ocfs2_extent_rec *rec;
2324
2325         if (next_free == 0)
2326                 return 0;
2327
2328         rec = &el->l_recs[0];
2329         if (ocfs2_is_empty_extent(rec)) {
2330                 /* Empty list. */
2331                 if (next_free == 1)
2332                         return 0;
2333                 rec = &el->l_recs[1];
2334         }
2335
2336         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2337         if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2338                 return 1;
2339         return 0;
2340 }
2341
2342 /*
2343  * Rotate all the records in a btree right one record, starting at insert_cpos.
2344  *
2345  * The path to the rightmost leaf should be passed in.
2346  *
2347  * The array is assumed to be large enough to hold an entire path (tree depth).
2348  *
2349  * Upon successful return from this function:
2350  *
2351  * - The 'right_path' array will contain a path to the leaf block
2352  *   whose range contains e_cpos.
2353  * - That leaf block will have a single empty extent in list index 0.
2354  * - In the case that the rotation requires a post-insert update,
2355  *   *ret_left_path will contain a valid path which can be passed to
2356  *   ocfs2_insert_path().
2357  */
2358 static int ocfs2_rotate_tree_right(handle_t *handle,
2359                                    struct ocfs2_extent_tree *et,
2360                                    enum ocfs2_split_type split,
2361                                    u32 insert_cpos,
2362                                    struct ocfs2_path *right_path,
2363                                    struct ocfs2_path **ret_left_path)
2364 {
2365         int ret, start, orig_credits = handle->h_buffer_credits;
2366         u32 cpos;
2367         struct ocfs2_path *left_path = NULL;
2368         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2369
2370         *ret_left_path = NULL;
2371
2372         left_path = ocfs2_new_path_from_path(right_path);
2373         if (!left_path) {
2374                 ret = -ENOMEM;
2375                 mlog_errno(ret);
2376                 goto out;
2377         }
2378
2379         ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2380         if (ret) {
2381                 mlog_errno(ret);
2382                 goto out;
2383         }
2384
2385         trace_ocfs2_rotate_tree_right(
2386                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2387                 insert_cpos, cpos);
2388
2389         /*
2390          * What we want to do here is:
2391          *
2392          * 1) Start with the rightmost path.
2393          *
2394          * 2) Determine a path to the leaf block directly to the left
2395          *    of that leaf.
2396          *
2397          * 3) Determine the 'subtree root' - the lowest level tree node
2398          *    which contains a path to both leaves.
2399          *
2400          * 4) Rotate the subtree.
2401          *
2402          * 5) Find the next subtree by considering the left path to be
2403          *    the new right path.
2404          *
2405          * The check at the top of this while loop also accepts
2406          * insert_cpos == cpos because cpos is only a _theoretical_
2407          * value to get us the left path - insert_cpos might very well
2408          * be filling that hole.
2409          *
2410          * Stop at a cpos of '0' because we either started at the
2411          * leftmost branch (i.e., a tree with one branch and a
2412          * rotation inside of it), or we've gone as far as we can in
2413          * rotating subtrees.
2414          */
2415         while (cpos && insert_cpos <= cpos) {
2416                 trace_ocfs2_rotate_tree_right(
2417                         (unsigned long long)
2418                         ocfs2_metadata_cache_owner(et->et_ci),
2419                         insert_cpos, cpos);
2420
2421                 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2422                 if (ret) {
2423                         mlog_errno(ret);
2424                         goto out;
2425                 }
2426
2427                 mlog_bug_on_msg(path_leaf_bh(left_path) ==
2428                                 path_leaf_bh(right_path),
2429                                 "Owner %llu: error during insert of %u "
2430                                 "(left path cpos %u) results in two identical "
2431                                 "paths ending at %llu\n",
2432                                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2433                                 insert_cpos, cpos,
2434                                 (unsigned long long)
2435                                 path_leaf_bh(left_path)->b_blocknr);
2436
2437                 if (split == SPLIT_NONE &&
2438                     ocfs2_rotate_requires_path_adjustment(left_path,
2439                                                           insert_cpos)) {
2440
2441                         /*
2442                          * We've rotated the tree as much as we
2443                          * should. The rest is up to
2444                          * ocfs2_insert_path() to complete, after the
2445                          * record insertion. We indicate this
2446                          * situation by returning the left path.
2447                          *
2448                          * The reason we don't adjust the records here
2449                          * before the record insert is that an error
2450                          * later might break the rule where a parent
2451                          * record e_cpos will reflect the actual
2452                          * e_cpos of the 1st nonempty record of the
2453                          * child list.
2454                          */
2455                         *ret_left_path = left_path;
2456                         goto out_ret_path;
2457                 }
2458
2459                 start = ocfs2_find_subtree_root(et, left_path, right_path);
2460
2461                 trace_ocfs2_rotate_subtree(start,
2462                         (unsigned long long)
2463                         right_path->p_node[start].bh->b_blocknr,
2464                         right_path->p_tree_depth);
2465
2466                 ret = ocfs2_extend_rotate_transaction(handle, start,
2467                                                       orig_credits, right_path);
2468                 if (ret) {
2469                         mlog_errno(ret);
2470                         goto out;
2471                 }
2472
2473                 ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2474                                                  right_path, start);
2475                 if (ret) {
2476                         mlog_errno(ret);
2477                         goto out;
2478                 }
2479
2480                 if (split != SPLIT_NONE &&
2481                     ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2482                                                 insert_cpos)) {
2483                         /*
2484                          * A rotate moves the rightmost left leaf
2485                          * record over to the leftmost right leaf
2486                          * slot. If we're doing an extent split
2487                          * instead of a real insert, then we have to
2488                          * check that the extent to be split wasn't
2489                          * just moved over. If it was, then we can
2490                          * exit here, passing left_path back -
2491                          * ocfs2_split_extent() is smart enough to
2492                          * search both leaves.
2493                          */
2494                         *ret_left_path = left_path;
2495                         goto out_ret_path;
2496                 }
2497
2498                 /*
2499                  * There is no need to re-read the next right path
2500                  * as we know that it'll be our current left
2501                  * path. Optimize by copying values instead.
2502                  */
2503                 ocfs2_mv_path(right_path, left_path);
2504
2505                 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2506                 if (ret) {
2507                         mlog_errno(ret);
2508                         goto out;
2509                 }
2510         }
2511
2512 out:
2513         ocfs2_free_path(left_path);
2514
2515 out_ret_path:
2516         return ret;
2517 }
2518
2519 static int ocfs2_update_edge_lengths(handle_t *handle,
2520                                      struct ocfs2_extent_tree *et,
2521                                      int subtree_index, struct ocfs2_path *path)
2522 {
2523         int i, idx, ret;
2524         struct ocfs2_extent_rec *rec;
2525         struct ocfs2_extent_list *el;
2526         struct ocfs2_extent_block *eb;
2527         u32 range;
2528
2529         /*
2530          * In normal tree rotation process, we will never touch the
2531          * tree branch above subtree_index and ocfs2_extend_rotate_transaction
2532          * doesn't reserve the credits for them either.
2533          *
2534          * But we do have a special case here which will update the rightmost
2535          * records for all the bh in the path.
2536          * So we have to allocate extra credits and access them.
2537          */
2538         ret = ocfs2_extend_trans(handle, subtree_index);
2539         if (ret) {
2540                 mlog_errno(ret);
2541                 goto out;
2542         }
2543
2544         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2545         if (ret) {
2546                 mlog_errno(ret);
2547                 goto out;
2548         }
2549
2550         /* Path should always be rightmost. */
2551         eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2552         BUG_ON(eb->h_next_leaf_blk != 0ULL);
2553
2554         el = &eb->h_list;
2555         BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2556         idx = le16_to_cpu(el->l_next_free_rec) - 1;
2557         rec = &el->l_recs[idx];
2558         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2559
2560         for (i = 0; i < path->p_tree_depth; i++) {
2561                 el = path->p_node[i].el;
2562                 idx = le16_to_cpu(el->l_next_free_rec) - 1;
2563                 rec = &el->l_recs[idx];
2564
2565                 rec->e_int_clusters = cpu_to_le32(range);
2566                 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2567
2568                 ocfs2_journal_dirty(handle, path->p_node[i].bh);
2569         }
2570 out:
2571         return ret;
2572 }
2573
2574 static void ocfs2_unlink_path(handle_t *handle,
2575                               struct ocfs2_extent_tree *et,
2576                               struct ocfs2_cached_dealloc_ctxt *dealloc,
2577                               struct ocfs2_path *path, int unlink_start)
2578 {
2579         int ret, i;
2580         struct ocfs2_extent_block *eb;
2581         struct ocfs2_extent_list *el;
2582         struct buffer_head *bh;
2583
2584         for(i = unlink_start; i < path_num_items(path); i++) {
2585                 bh = path->p_node[i].bh;
2586
2587                 eb = (struct ocfs2_extent_block *)bh->b_data;
2588                 /*
2589                  * Not all nodes might have had their final count
2590                  * decremented by the caller - handle this here.
2591                  */
2592                 el = &eb->h_list;
2593                 if (le16_to_cpu(el->l_next_free_rec) > 1) {
2594                         mlog(ML_ERROR,
2595                              "Inode %llu, attempted to remove extent block "
2596                              "%llu with %u records\n",
2597                              (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2598                              (unsigned long long)le64_to_cpu(eb->h_blkno),
2599                              le16_to_cpu(el->l_next_free_rec));
2600
2601                         ocfs2_journal_dirty(handle, bh);
2602                         ocfs2_remove_from_cache(et->et_ci, bh);
2603                         continue;
2604                 }
2605
2606                 el->l_next_free_rec = 0;
2607                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2608
2609                 ocfs2_journal_dirty(handle, bh);
2610
2611                 ret = ocfs2_cache_extent_block_free(dealloc, eb);
2612                 if (ret)
2613                         mlog_errno(ret);
2614
2615                 ocfs2_remove_from_cache(et->et_ci, bh);
2616         }
2617 }
2618
2619 static void ocfs2_unlink_subtree(handle_t *handle,
2620                                  struct ocfs2_extent_tree *et,
2621                                  struct ocfs2_path *left_path,
2622                                  struct ocfs2_path *right_path,
2623                                  int subtree_index,
2624                                  struct ocfs2_cached_dealloc_ctxt *dealloc)
2625 {
2626         int i;
2627         struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2628         struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2629         struct ocfs2_extent_list *el;
2630         struct ocfs2_extent_block *eb;
2631
2632         el = path_leaf_el(left_path);
2633
2634         eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2635
2636         for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2637                 if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2638                         break;
2639
2640         BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2641
2642         memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2643         le16_add_cpu(&root_el->l_next_free_rec, -1);
2644
2645         eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2646         eb->h_next_leaf_blk = 0;
2647
2648         ocfs2_journal_dirty(handle, root_bh);
2649         ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2650
2651         ocfs2_unlink_path(handle, et, dealloc, right_path,
2652                           subtree_index + 1);
2653 }
2654
2655 static int ocfs2_rotate_subtree_left(handle_t *handle,
2656                                      struct ocfs2_extent_tree *et,
2657                                      struct ocfs2_path *left_path,
2658                                      struct ocfs2_path *right_path,
2659                                      int subtree_index,
2660                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
2661                                      int *deleted)
2662 {
2663         int ret, i, del_right_subtree = 0, right_has_empty = 0;
2664         struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2665         struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2666         struct ocfs2_extent_block *eb;
2667
2668         *deleted = 0;
2669
2670         right_leaf_el = path_leaf_el(right_path);
2671         left_leaf_el = path_leaf_el(left_path);
2672         root_bh = left_path->p_node[subtree_index].bh;
2673         BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2674
2675         if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2676                 return 0;
2677
2678         eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2679         if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2680                 /*
2681                  * It's legal for us to proceed if the right leaf is
2682                  * the rightmost one and it has an empty extent. There
2683                  * are two cases to handle - whether the leaf will be
2684                  * empty after removal or not. If the leaf isn't empty
2685                  * then just remove the empty extent up front. The
2686                  * next block will handle empty leaves by flagging
2687                  * them for unlink.
2688                  *
2689                  * Non rightmost leaves will throw -EAGAIN and the
2690                  * caller can manually move the subtree and retry.
2691                  */
2692
2693                 if (eb->h_next_leaf_blk != 0ULL)
2694                         return -EAGAIN;
2695
2696                 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2697                         ret = ocfs2_journal_access_eb(handle, et->et_ci,
2698                                                       path_leaf_bh(right_path),
2699                                                       OCFS2_JOURNAL_ACCESS_WRITE);
2700                         if (ret) {
2701                                 mlog_errno(ret);
2702                                 goto out;
2703                         }
2704
2705                         ocfs2_remove_empty_extent(right_leaf_el);
2706                 } else
2707                         right_has_empty = 1;
2708         }
2709
2710         if (eb->h_next_leaf_blk == 0ULL &&
2711             le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2712                 /*
2713                  * We have to update i_last_eb_blk during the meta
2714                  * data delete.
2715                  */
2716                 ret = ocfs2_et_root_journal_access(handle, et,
2717                                                    OCFS2_JOURNAL_ACCESS_WRITE);
2718                 if (ret) {
2719                         mlog_errno(ret);
2720                         goto out;
2721                 }
2722
2723                 del_right_subtree = 1;
2724         }
2725
2726         /*
2727          * Getting here with an empty extent in the right path implies
2728          * that it's the rightmost path and will be deleted.
2729          */
2730         BUG_ON(right_has_empty && !del_right_subtree);
2731
2732         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2733                                            subtree_index);
2734         if (ret) {
2735                 mlog_errno(ret);
2736                 goto out;
2737         }
2738
2739         for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2740                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2741                                                    right_path, i);
2742                 if (ret) {
2743                         mlog_errno(ret);
2744                         goto out;
2745                 }
2746
2747                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2748                                                    left_path, i);
2749                 if (ret) {
2750                         mlog_errno(ret);
2751                         goto out;
2752                 }
2753         }
2754
2755         if (!right_has_empty) {
2756                 /*
2757                  * Only do this if we're moving a real
2758                  * record. Otherwise, the action is delayed until
2759                  * after removal of the right path in which case we
2760                  * can do a simple shift to remove the empty extent.
2761                  */
2762                 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2763                 memset(&right_leaf_el->l_recs[0], 0,
2764                        sizeof(struct ocfs2_extent_rec));
2765         }
2766         if (eb->h_next_leaf_blk == 0ULL) {
2767                 /*
2768                  * Move recs over to get rid of empty extent, decrease
2769                  * next_free. This is allowed to remove the last
2770                  * extent in our leaf (setting l_next_free_rec to
2771                  * zero) - the delete code below won't care.
2772                  */
2773                 ocfs2_remove_empty_extent(right_leaf_el);
2774         }
2775
2776         ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2777         ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2778
2779         if (del_right_subtree) {
2780                 ocfs2_unlink_subtree(handle, et, left_path, right_path,
2781                                      subtree_index, dealloc);
2782                 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
2783                                                 left_path);
2784                 if (ret) {
2785                         mlog_errno(ret);
2786                         goto out;
2787                 }
2788
2789                 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2790                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2791
2792                 /*
2793                  * Removal of the extent in the left leaf was skipped
2794                  * above so we could delete the right path
2795                  * 1st.
2796                  */
2797                 if (right_has_empty)
2798                         ocfs2_remove_empty_extent(left_leaf_el);
2799
2800                 ocfs2_journal_dirty(handle, et_root_bh);
2801
2802                 *deleted = 1;
2803         } else
2804                 ocfs2_complete_edge_insert(handle, left_path, right_path,
2805                                            subtree_index);
2806
2807 out:
2808         return ret;
2809 }
2810
2811 /*
2812  * Given a full path, determine what cpos value would return us a path
2813  * containing the leaf immediately to the right of the current one.
2814  *
2815  * Will return zero if the path passed in is already the rightmost path.
2816  *
2817  * This looks similar, but is subtly different to
2818  * ocfs2_find_cpos_for_left_leaf().
2819  */
2820 int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2821                                    struct ocfs2_path *path, u32 *cpos)
2822 {
2823         int i, j, ret = 0;
2824         u64 blkno;
2825         struct ocfs2_extent_list *el;
2826
2827         *cpos = 0;
2828
2829         if (path->p_tree_depth == 0)
2830                 return 0;
2831
2832         blkno = path_leaf_bh(path)->b_blocknr;
2833
2834         /* Start at the tree node just above the leaf and work our way up. */
2835         i = path->p_tree_depth - 1;
2836         while (i >= 0) {
2837                 int next_free;
2838
2839                 el = path->p_node[i].el;
2840
2841                 /*
2842                  * Find the extent record just after the one in our
2843                  * path.
2844                  */
2845                 next_free = le16_to_cpu(el->l_next_free_rec);
2846                 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2847                         if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2848                                 if (j == (next_free - 1)) {
2849                                         if (i == 0) {
2850                                                 /*
2851                                                  * We've determined that the
2852                                                  * path specified is already
2853                                                  * the rightmost one - return a
2854                                                  * cpos of zero.
2855                                                  */
2856                                                 goto out;
2857                                         }
2858                                         /*
2859                                          * The rightmost record points to our
2860                                          * leaf - we need to travel up the
2861                                          * tree one level.
2862                                          */
2863                                         goto next_node;
2864                                 }
2865
2866                                 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2867                                 goto out;
2868                         }
2869                 }
2870
2871                 /*
2872                  * If we got here, we never found a valid node where
2873                  * the tree indicated one should be.
2874                  */
2875                 ocfs2_error(sb,
2876                             "Invalid extent tree at extent block %llu\n",
2877                             (unsigned long long)blkno);
2878                 ret = -EROFS;
2879                 goto out;
2880
2881 next_node:
2882                 blkno = path->p_node[i].bh->b_blocknr;
2883                 i--;
2884         }
2885
2886 out:
2887         return ret;
2888 }
2889
2890 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2891                                             struct ocfs2_extent_tree *et,
2892                                             struct ocfs2_path *path)
2893 {
2894         int ret;
2895         struct buffer_head *bh = path_leaf_bh(path);
2896         struct ocfs2_extent_list *el = path_leaf_el(path);
2897
2898         if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2899                 return 0;
2900
2901         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2902                                            path_num_items(path) - 1);
2903         if (ret) {
2904                 mlog_errno(ret);
2905                 goto out;
2906         }
2907
2908         ocfs2_remove_empty_extent(el);
2909         ocfs2_journal_dirty(handle, bh);
2910
2911 out:
2912         return ret;
2913 }
2914
2915 static int __ocfs2_rotate_tree_left(handle_t *handle,
2916                                     struct ocfs2_extent_tree *et,
2917                                     int orig_credits,
2918                                     struct ocfs2_path *path,
2919                                     struct ocfs2_cached_dealloc_ctxt *dealloc,
2920                                     struct ocfs2_path **empty_extent_path)
2921 {
2922         int ret, subtree_root, deleted;
2923         u32 right_cpos;
2924         struct ocfs2_path *left_path = NULL;
2925         struct ocfs2_path *right_path = NULL;
2926         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2927
2928         if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2929                 return 0;
2930
2931         *empty_extent_path = NULL;
2932
2933         ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2934         if (ret) {
2935                 mlog_errno(ret);
2936                 goto out;
2937         }
2938
2939         left_path = ocfs2_new_path_from_path(path);
2940         if (!left_path) {
2941                 ret = -ENOMEM;
2942                 mlog_errno(ret);
2943                 goto out;
2944         }
2945
2946         ocfs2_cp_path(left_path, path);
2947
2948         right_path = ocfs2_new_path_from_path(path);
2949         if (!right_path) {
2950                 ret = -ENOMEM;
2951                 mlog_errno(ret);
2952                 goto out;
2953         }
2954
2955         while (right_cpos) {
2956                 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2957                 if (ret) {
2958                         mlog_errno(ret);
2959                         goto out;
2960                 }
2961
2962                 subtree_root = ocfs2_find_subtree_root(et, left_path,
2963                                                        right_path);
2964
2965                 trace_ocfs2_rotate_subtree(subtree_root,
2966                      (unsigned long long)
2967                      right_path->p_node[subtree_root].bh->b_blocknr,
2968                      right_path->p_tree_depth);
2969
2970                 ret = ocfs2_extend_rotate_transaction(handle, subtree_root,
2971                                                       orig_credits, left_path);
2972                 if (ret) {
2973                         mlog_errno(ret);
2974                         goto out;
2975                 }
2976
2977                 /*
2978                  * Caller might still want to make changes to the
2979                  * tree root, so re-add it to the journal here.
2980                  */
2981                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2982                                                    left_path, 0);
2983                 if (ret) {
2984                         mlog_errno(ret);
2985                         goto out;
2986                 }
2987
2988                 ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2989                                                 right_path, subtree_root,
2990                                                 dealloc, &deleted);
2991                 if (ret == -EAGAIN) {
2992                         /*
2993                          * The rotation has to temporarily stop due to
2994                          * the right subtree having an empty
2995                          * extent. Pass it back to the caller for a
2996                          * fixup.
2997                          */
2998                         *empty_extent_path = right_path;
2999                         right_path = NULL;
3000                         goto out;
3001                 }
3002                 if (ret) {
3003                         mlog_errno(ret);
3004                         goto out;
3005                 }
3006
3007                 /*
3008                  * The subtree rotate might have removed records on
3009                  * the rightmost edge. If so, then rotation is
3010                  * complete.
3011                  */
3012                 if (deleted)
3013                         break;
3014
3015                 ocfs2_mv_path(left_path, right_path);
3016
3017                 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3018                                                      &right_cpos);
3019                 if (ret) {
3020                         mlog_errno(ret);
3021                         goto out;
3022                 }
3023         }
3024
3025 out:
3026         ocfs2_free_path(right_path);
3027         ocfs2_free_path(left_path);
3028
3029         return ret;
3030 }
3031
3032 static int ocfs2_remove_rightmost_path(handle_t *handle,
3033                                 struct ocfs2_extent_tree *et,
3034                                 struct ocfs2_path *path,
3035                                 struct ocfs2_cached_dealloc_ctxt *dealloc)
3036 {
3037         int ret, subtree_index;
3038         u32 cpos;
3039         struct ocfs2_path *left_path = NULL;
3040         struct ocfs2_extent_block *eb;
3041         struct ocfs2_extent_list *el;
3042
3043
3044         ret = ocfs2_et_sanity_check(et);
3045         if (ret)
3046                 goto out;
3047         /*
3048          * There's two ways we handle this depending on
3049          * whether path is the only existing one.
3050          */
3051         ret = ocfs2_extend_rotate_transaction(handle, 0,
3052                                               handle->h_buffer_credits,
3053                                               path);
3054         if (ret) {
3055                 mlog_errno(ret);
3056                 goto out;
3057         }
3058
3059         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3060         if (ret) {
3061                 mlog_errno(ret);
3062                 goto out;
3063         }
3064
3065         ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3066                                             path, &cpos);
3067         if (ret) {
3068                 mlog_errno(ret);
3069                 goto out;
3070         }
3071
3072         if (cpos) {
3073                 /*
3074                  * We have a path to the left of this one - it needs
3075                  * an update too.
3076                  */
3077                 left_path = ocfs2_new_path_from_path(path);
3078                 if (!left_path) {
3079                         ret = -ENOMEM;
3080                         mlog_errno(ret);
3081                         goto out;
3082                 }
3083
3084                 ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3085                 if (ret) {
3086                         mlog_errno(ret);
3087                         goto out;
3088                 }
3089
3090                 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3091                 if (ret) {
3092                         mlog_errno(ret);
3093                         goto out;
3094                 }
3095
3096                 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3097
3098                 ocfs2_unlink_subtree(handle, et, left_path, path,
3099                                      subtree_index, dealloc);
3100                 ret = ocfs2_update_edge_lengths(handle, et, subtree_index,
3101                                                 left_path);
3102                 if (ret) {
3103                         mlog_errno(ret);
3104                         goto out;
3105                 }
3106
3107                 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3108                 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3109         } else {
3110                 /*
3111                  * 'path' is also the leftmost path which
3112                  * means it must be the only one. This gets
3113                  * handled differently because we want to
3114                  * revert the root back to having extents
3115                  * in-line.
3116                  */
3117                 ocfs2_unlink_path(handle, et, dealloc, path, 1);
3118
3119                 el = et->et_root_el;
3120                 el->l_tree_depth = 0;
3121                 el->l_next_free_rec = 0;
3122                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3123
3124                 ocfs2_et_set_last_eb_blk(et, 0);
3125         }
3126
3127         ocfs2_journal_dirty(handle, path_root_bh(path));
3128
3129 out:
3130         ocfs2_free_path(left_path);
3131         return ret;
3132 }
3133
3134 /*
3135  * Left rotation of btree records.
3136  *
3137  * In many ways, this is (unsurprisingly) the opposite of right
3138  * rotation. We start at some non-rightmost path containing an empty
3139  * extent in the leaf block. The code works its way to the rightmost
3140  * path by rotating records to the left in every subtree.
3141  *
3142  * This is used by any code which reduces the number of extent records
3143  * in a leaf. After removal, an empty record should be placed in the
3144  * leftmost list position.
3145  *
3146  * This won't handle a length update of the rightmost path records if
3147  * the rightmost tree leaf record is removed so the caller is
3148  * responsible for detecting and correcting that.
3149  */
3150 static int ocfs2_rotate_tree_left(handle_t *handle,
3151                                   struct ocfs2_extent_tree *et,
3152                                   struct ocfs2_path *path,
3153                                   struct ocfs2_cached_dealloc_ctxt *dealloc)
3154 {
3155         int ret, orig_credits = handle->h_buffer_credits;
3156         struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3157         struct ocfs2_extent_block *eb;
3158         struct ocfs2_extent_list *el;
3159
3160         el = path_leaf_el(path);
3161         if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3162                 return 0;
3163
3164         if (path->p_tree_depth == 0) {
3165 rightmost_no_delete:
3166                 /*
3167                  * Inline extents. This is trivially handled, so do
3168                  * it up front.
3169                  */
3170                 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3171                 if (ret)
3172                         mlog_errno(ret);
3173                 goto out;
3174         }
3175
3176         /*
3177          * Handle rightmost branch now. There's several cases:
3178          *  1) simple rotation leaving records in there. That's trivial.
3179          *  2) rotation requiring a branch delete - there's no more
3180          *     records left. Two cases of this:
3181          *     a) There are branches to the left.
3182          *     b) This is also the leftmost (the only) branch.
3183          *
3184          *  1) is handled via ocfs2_rotate_rightmost_leaf_left()
3185          *  2a) we need the left branch so that we can update it with the unlink
3186          *  2b) we need to bring the root back to inline extents.
3187          */
3188
3189         eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3190         el = &eb->h_list;
3191         if (eb->h_next_leaf_blk == 0) {
3192                 /*
3193                  * This gets a bit tricky if we're going to delete the
3194                  * rightmost path. Get the other cases out of the way
3195                  * 1st.
3196                  */
3197                 if (le16_to_cpu(el->l_next_free_rec) > 1)
3198                         goto rightmost_no_delete;
3199
3200                 if (le16_to_cpu(el->l_next_free_rec) == 0) {
3201                         ret = -EIO;
3202                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3203                                     "Owner %llu has empty extent block at %llu",
3204                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3205                                     (unsigned long long)le64_to_cpu(eb->h_blkno));
3206                         goto out;
3207                 }
3208
3209                 /*
3210                  * XXX: The caller can not trust "path" any more after
3211                  * this as it will have been deleted. What do we do?
3212                  *
3213                  * In theory the rotate-for-merge code will never get
3214                  * here because it'll always ask for a rotate in a
3215                  * nonempty list.
3216                  */
3217
3218                 ret = ocfs2_remove_rightmost_path(handle, et, path,
3219                                                   dealloc);
3220                 if (ret)
3221                         mlog_errno(ret);
3222                 goto out;
3223         }
3224
3225         /*
3226          * Now we can loop, remembering the path we get from -EAGAIN
3227          * and restarting from there.
3228          */
3229 try_rotate:
3230         ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3231                                        dealloc, &restart_path);
3232         if (ret && ret != -EAGAIN) {
3233                 mlog_errno(ret);
3234                 goto out;
3235         }
3236
3237         while (ret == -EAGAIN) {
3238                 tmp_path = restart_path;
3239                 restart_path = NULL;
3240
3241                 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3242                                                tmp_path, dealloc,
3243                                                &restart_path);
3244                 if (ret && ret != -EAGAIN) {
3245                         mlog_errno(ret);
3246                         goto out;
3247                 }
3248
3249                 ocfs2_free_path(tmp_path);
3250                 tmp_path = NULL;
3251
3252                 if (ret == 0)
3253                         goto try_rotate;
3254         }
3255
3256 out:
3257         ocfs2_free_path(tmp_path);
3258         ocfs2_free_path(restart_path);
3259         return ret;
3260 }
3261
3262 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3263                                 int index)
3264 {
3265         struct ocfs2_extent_rec *rec = &el->l_recs[index];
3266         unsigned int size;
3267
3268         if (rec->e_leaf_clusters == 0) {
3269                 /*
3270                  * We consumed all of the merged-from record. An empty
3271                  * extent cannot exist anywhere but the 1st array
3272                  * position, so move things over if the merged-from
3273                  * record doesn't occupy that position.
3274                  *
3275                  * This creates a new empty extent so the caller
3276                  * should be smart enough to have removed any existing
3277                  * ones.
3278                  */
3279                 if (index > 0) {
3280                         BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3281                         size = index * sizeof(struct ocfs2_extent_rec);
3282                         memmove(&el->l_recs[1], &el->l_recs[0], size);
3283                 }
3284
3285                 /*
3286                  * Always memset - the caller doesn't check whether it
3287                  * created an empty extent, so there could be junk in
3288                  * the other fields.
3289                  */
3290                 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3291         }
3292 }
3293
3294 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3295                                 struct ocfs2_path *left_path,
3296                                 struct ocfs2_path **ret_right_path)
3297 {
3298         int ret;
3299         u32 right_cpos;
3300         struct ocfs2_path *right_path = NULL;
3301         struct ocfs2_extent_list *left_el;
3302
3303         *ret_right_path = NULL;
3304
3305         /* This function shouldn't be called for non-trees. */
3306         BUG_ON(left_path->p_tree_depth == 0);
3307
3308         left_el = path_leaf_el(left_path);
3309         BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3310
3311         ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3312                                              left_path, &right_cpos);
3313         if (ret) {
3314                 mlog_errno(ret);
3315                 goto out;
3316         }
3317
3318         /* This function shouldn't be called for the rightmost leaf. */
3319         BUG_ON(right_cpos == 0);
3320
3321         right_path = ocfs2_new_path_from_path(left_path);
3322         if (!right_path) {
3323                 ret = -ENOMEM;
3324                 mlog_errno(ret);
3325                 goto out;
3326         }
3327
3328         ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3329         if (ret) {
3330                 mlog_errno(ret);
3331                 goto out;
3332         }
3333
3334         *ret_right_path = right_path;
3335 out:
3336         if (ret)
3337                 ocfs2_free_path(right_path);
3338         return ret;
3339 }
3340
3341 /*
3342  * Remove split_rec clusters from the record at index and merge them
3343  * onto the beginning of the record "next" to it.
3344  * For index < l_count - 1, the next means the extent rec at index + 1.
3345  * For index == l_count - 1, the "next" means the 1st extent rec of the
3346  * next extent block.
3347  */
3348 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3349                                  handle_t *handle,
3350                                  struct ocfs2_extent_tree *et,
3351                                  struct ocfs2_extent_rec *split_rec,
3352                                  int index)
3353 {
3354         int ret, next_free, i;
3355         unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3356         struct ocfs2_extent_rec *left_rec;
3357         struct ocfs2_extent_rec *right_rec;
3358         struct ocfs2_extent_list *right_el;
3359         struct ocfs2_path *right_path = NULL;
3360         int subtree_index = 0;
3361         struct ocfs2_extent_list *el = path_leaf_el(left_path);
3362         struct buffer_head *bh = path_leaf_bh(left_path);
3363         struct buffer_head *root_bh = NULL;
3364
3365         BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3366         left_rec = &el->l_recs[index];
3367
3368         if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3369             le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3370                 /* we meet with a cross extent block merge. */
3371                 ret = ocfs2_get_right_path(et, left_path, &right_path);
3372                 if (ret) {
3373                         mlog_errno(ret);
3374                         return ret;
3375                 }
3376
3377                 right_el = path_leaf_el(right_path);
3378                 next_free = le16_to_cpu(right_el->l_next_free_rec);
3379                 BUG_ON(next_free <= 0);
3380                 right_rec = &right_el->l_recs[0];
3381                 if (ocfs2_is_empty_extent(right_rec)) {
3382                         BUG_ON(next_free <= 1);
3383                         right_rec = &right_el->l_recs[1];
3384                 }
3385
3386                 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3387                        le16_to_cpu(left_rec->e_leaf_clusters) !=
3388                        le32_to_cpu(right_rec->e_cpos));
3389
3390                 subtree_index = ocfs2_find_subtree_root(et, left_path,
3391                                                         right_path);
3392
3393                 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3394                                                       handle->h_buffer_credits,
3395                                                       right_path);
3396                 if (ret) {
3397                         mlog_errno(ret);
3398                         goto out;
3399                 }
3400
3401                 root_bh = left_path->p_node[subtree_index].bh;
3402                 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3403
3404                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3405                                                    subtree_index);
3406                 if (ret) {
3407                         mlog_errno(ret);
3408                         goto out;
3409                 }
3410
3411                 for (i = subtree_index + 1;
3412                      i < path_num_items(right_path); i++) {
3413                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3414                                                            right_path, i);
3415                         if (ret) {
3416                                 mlog_errno(ret);
3417                                 goto out;
3418                         }
3419
3420                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3421                                                            left_path, i);
3422                         if (ret) {
3423                                 mlog_errno(ret);
3424                                 goto out;
3425                         }
3426                 }
3427
3428         } else {
3429                 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3430                 right_rec = &el->l_recs[index + 1];
3431         }
3432
3433         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3434                                            path_num_items(left_path) - 1);
3435         if (ret) {
3436                 mlog_errno(ret);
3437                 goto out;
3438         }
3439
3440         le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3441
3442         le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3443         le64_add_cpu(&right_rec->e_blkno,
3444                      -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3445                                                split_clusters));
3446         le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3447
3448         ocfs2_cleanup_merge(el, index);
3449
3450         ocfs2_journal_dirty(handle, bh);
3451         if (right_path) {
3452                 ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3453                 ocfs2_complete_edge_insert(handle, left_path, right_path,
3454                                            subtree_index);
3455         }
3456 out:
3457         ocfs2_free_path(right_path);
3458         return ret;
3459 }
3460
3461 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3462                                struct ocfs2_path *right_path,
3463                                struct ocfs2_path **ret_left_path)
3464 {
3465         int ret;
3466         u32 left_cpos;
3467         struct ocfs2_path *left_path = NULL;
3468
3469         *ret_left_path = NULL;
3470
3471         /* This function shouldn't be called for non-trees. */
3472         BUG_ON(right_path->p_tree_depth == 0);
3473
3474         ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3475                                             right_path, &left_cpos);
3476         if (ret) {
3477                 mlog_errno(ret);
3478                 goto out;
3479         }
3480
3481         /* This function shouldn't be called for the leftmost leaf. */
3482         BUG_ON(left_cpos == 0);
3483
3484         left_path = ocfs2_new_path_from_path(right_path);
3485         if (!left_path) {
3486                 ret = -ENOMEM;
3487                 mlog_errno(ret);
3488                 goto out;
3489         }
3490
3491         ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3492         if (ret) {
3493                 mlog_errno(ret);
3494                 goto out;
3495         }
3496
3497         *ret_left_path = left_path;
3498 out:
3499         if (ret)
3500                 ocfs2_free_path(left_path);
3501         return ret;
3502 }
3503
3504 /*
3505  * Remove split_rec clusters from the record at index and merge them
3506  * onto the tail of the record "before" it.
3507  * For index > 0, the "before" means the extent rec at index - 1.
3508  *
3509  * For index == 0, the "before" means the last record of the previous
3510  * extent block. And there is also a situation that we may need to
3511  * remove the rightmost leaf extent block in the right_path and change
3512  * the right path to indicate the new rightmost path.
3513  */
3514 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3515                                 handle_t *handle,
3516                                 struct ocfs2_extent_tree *et,
3517                                 struct ocfs2_extent_rec *split_rec,
3518                                 struct ocfs2_cached_dealloc_ctxt *dealloc,
3519                                 int index)
3520 {
3521         int ret, i, subtree_index = 0, has_empty_extent = 0;
3522         unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3523         struct ocfs2_extent_rec *left_rec;
3524         struct ocfs2_extent_rec *right_rec;
3525         struct ocfs2_extent_list *el = path_leaf_el(right_path);
3526         struct buffer_head *bh = path_leaf_bh(right_path);
3527         struct buffer_head *root_bh = NULL;
3528         struct ocfs2_path *left_path = NULL;
3529         struct ocfs2_extent_list *left_el;
3530
3531         BUG_ON(index < 0);
3532
3533         right_rec = &el->l_recs[index];
3534         if (index == 0) {
3535                 /* we meet with a cross extent block merge. */
3536                 ret = ocfs2_get_left_path(et, right_path, &left_path);
3537                 if (ret) {
3538                         mlog_errno(ret);
3539                         return ret;
3540                 }
3541
3542                 left_el = path_leaf_el(left_path);
3543                 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3544                        le16_to_cpu(left_el->l_count));
3545
3546                 left_rec = &left_el->l_recs[
3547                                 le16_to_cpu(left_el->l_next_free_rec) - 1];
3548                 BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3549                        le16_to_cpu(left_rec->e_leaf_clusters) !=
3550                        le32_to_cpu(split_rec->e_cpos));
3551
3552                 subtree_index = ocfs2_find_subtree_root(et, left_path,
3553                                                         right_path);
3554
3555                 ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3556                                                       handle->h_buffer_credits,
3557                                                       left_path);
3558                 if (ret) {
3559                         mlog_errno(ret);
3560                         goto out;
3561                 }
3562
3563                 root_bh = left_path->p_node[subtree_index].bh;
3564                 BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3565
3566                 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3567                                                    subtree_index);
3568                 if (ret) {
3569                         mlog_errno(ret);
3570                         goto out;
3571                 }
3572
3573                 for (i = subtree_index + 1;
3574                      i < path_num_items(right_path); i++) {
3575                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3576                                                            right_path, i);
3577                         if (ret) {
3578                                 mlog_errno(ret);
3579                                 goto out;
3580                         }
3581
3582                         ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3583                                                            left_path, i);
3584                         if (ret) {
3585                                 mlog_errno(ret);
3586                                 goto out;
3587                         }
3588                 }
3589         } else {
3590                 left_rec = &el->l_recs[index - 1];
3591                 if (ocfs2_is_empty_extent(&el->l_recs[0]))
3592                         has_empty_extent = 1;
3593         }
3594
3595         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3596                                            path_num_items(right_path) - 1);
3597         if (ret) {
3598                 mlog_errno(ret);
3599                 goto out;
3600         }
3601
3602         if (has_empty_extent && index == 1) {
3603                 /*
3604                  * The easy case - we can just plop the record right in.
3605                  */
3606                 *left_rec = *split_rec;
3607
3608                 has_empty_extent = 0;
3609         } else
3610                 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3611
3612         le32_add_cpu(&right_rec->e_cpos, split_clusters);
3613         le64_add_cpu(&right_rec->e_blkno,
3614                      ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3615                                               split_clusters));
3616         le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3617
3618         ocfs2_cleanup_merge(el, index);
3619
3620         ocfs2_journal_dirty(handle, bh);
3621         if (left_path) {
3622                 ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3623
3624                 /*
3625                  * In the situation that the right_rec is empty and the extent
3626                  * block is empty also,  ocfs2_complete_edge_insert can't handle
3627                  * it and we need to delete the right extent block.
3628                  */
3629                 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3630                     le16_to_cpu(el->l_next_free_rec) == 1) {
3631
3632                         ret = ocfs2_remove_rightmost_path(handle, et,
3633                                                           right_path,
3634                                                           dealloc);
3635                         if (ret) {
3636                                 mlog_errno(ret);
3637                                 goto out;
3638                         }
3639
3640                         /* Now the rightmost extent block has been deleted.
3641                          * So we use the new rightmost path.
3642                          */
3643                         ocfs2_mv_path(right_path, left_path);
3644                         left_path = NULL;
3645                 } else
3646                         ocfs2_complete_edge_insert(handle, left_path,
3647                                                    right_path, subtree_index);
3648         }
3649 out:
3650         ocfs2_free_path(left_path);
3651         return ret;
3652 }
3653
3654 static int ocfs2_try_to_merge_extent(handle_t *handle,
3655                                      struct ocfs2_extent_tree *et,
3656                                      struct ocfs2_path *path,
3657                                      int split_index,
3658                                      struct ocfs2_extent_rec *split_rec,
3659                                      struct ocfs2_cached_dealloc_ctxt *dealloc,
3660                                      struct ocfs2_merge_ctxt *ctxt)
3661 {
3662         int ret = 0;
3663         struct ocfs2_extent_list *el = path_leaf_el(path);
3664         struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3665
3666         BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3667
3668         if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3669                 /*
3670                  * The merge code will need to create an empty
3671                  * extent to take the place of the newly
3672                  * emptied slot. Remove any pre-existing empty
3673                  * extents - having more than one in a leaf is
3674                  * illegal.
3675                  */
3676                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3677                 if (ret) {
3678                         mlog_errno(ret);
3679                         goto out;
3680                 }
3681                 split_index--;
3682                 rec = &el->l_recs[split_index];
3683         }
3684
3685         if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3686                 /*
3687                  * Left-right contig implies this.
3688                  */
3689                 BUG_ON(!ctxt->c_split_covers_rec);
3690
3691                 /*
3692                  * Since the leftright insert always covers the entire
3693                  * extent, this call will delete the insert record
3694                  * entirely, resulting in an empty extent record added to
3695                  * the extent block.
3696                  *
3697                  * Since the adding of an empty extent shifts
3698                  * everything back to the right, there's no need to
3699                  * update split_index here.
3700                  *
3701                  * When the split_index is zero, we need to merge it to the
3702                  * prevoius extent block. It is more efficient and easier
3703                  * if we do merge_right first and merge_left later.
3704                  */
3705                 ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3706                                             split_index);
3707                 if (ret) {
3708                         mlog_errno(ret);
3709                         goto out;
3710                 }
3711
3712                 /*
3713                  * We can only get this from logic error above.
3714                  */
3715                 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3716
3717                 /* The merge left us with an empty extent, remove it. */
3718                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3719                 if (ret) {
3720                         mlog_errno(ret);
3721                         goto out;
3722                 }
3723
3724                 rec = &el->l_recs[split_index];
3725
3726                 /*
3727                  * Note that we don't pass split_rec here on purpose -
3728                  * we've merged it into the rec already.
3729                  */
3730                 ret = ocfs2_merge_rec_left(path, handle, et, rec,
3731                                            dealloc, split_index);
3732
3733                 if (ret) {
3734                         mlog_errno(ret);
3735                         goto out;
3736                 }
3737
3738                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3739                 /*
3740                  * Error from this last rotate is not critical, so
3741                  * print but don't bubble it up.
3742                  */
3743                 if (ret)
3744                         mlog_errno(ret);
3745                 ret = 0;
3746         } else {
3747                 /*
3748                  * Merge a record to the left or right.
3749                  *
3750                  * 'contig_type' is relative to the existing record,
3751                  * so for example, if we're "right contig", it's to
3752                  * the record on the left (hence the left merge).
3753                  */
3754                 if (ctxt->c_contig_type == CONTIG_RIGHT) {
3755                         ret = ocfs2_merge_rec_left(path, handle, et,
3756                                                    split_rec, dealloc,
3757                                                    split_index);
3758                         if (ret) {
3759                                 mlog_errno(ret);
3760                                 goto out;
3761                         }
3762                 } else {
3763                         ret = ocfs2_merge_rec_right(path, handle,
3764                                                     et, split_rec,
3765                                                     split_index);
3766                         if (ret) {
3767                                 mlog_errno(ret);
3768                                 goto out;
3769                         }
3770                 }
3771
3772                 if (ctxt->c_split_covers_rec) {
3773                         /*
3774                          * The merge may have left an empty extent in
3775                          * our leaf. Try to rotate it away.
3776                          */
3777                         ret = ocfs2_rotate_tree_left(handle, et, path,
3778                                                      dealloc);
3779                         if (ret)
3780                                 mlog_errno(ret);
3781                         ret = 0;
3782                 }
3783         }
3784
3785 out:
3786         return ret;
3787 }
3788
3789 static void ocfs2_subtract_from_rec(struct super_block *sb,
3790                                     enum ocfs2_split_type split,
3791                                     struct ocfs2_extent_rec *rec,
3792                                     struct ocfs2_extent_rec *split_rec)
3793 {
3794         u64 len_blocks;
3795
3796         len_blocks = ocfs2_clusters_to_blocks(sb,
3797                                 le16_to_cpu(split_rec->e_leaf_clusters));
3798
3799         if (split == SPLIT_LEFT) {
3800                 /*
3801                  * Region is on the left edge of the existing
3802                  * record.
3803                  */
3804                 le32_add_cpu(&rec->e_cpos,
3805                              le16_to_cpu(split_rec->e_leaf_clusters));
3806                 le64_add_cpu(&rec->e_blkno, len_blocks);
3807                 le16_add_cpu(&rec->e_leaf_clusters,
3808                              -le16_to_cpu(split_rec->e_leaf_clusters));
3809         } else {
3810                 /*
3811                  * Region is on the right edge of the existing
3812                  * record.
3813                  */
3814                 le16_add_cpu(&rec->e_leaf_clusters,
3815                              -le16_to_cpu(split_rec->e_leaf_clusters));
3816         }
3817 }
3818
3819 /*
3820  * Do the final bits of extent record insertion at the target leaf
3821  * list. If this leaf is part of an allocation tree, it is assumed
3822  * that the tree above has been prepared.
3823  */
3824 static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3825                                  struct ocfs2_extent_rec *insert_rec,
3826                                  struct ocfs2_extent_list *el,
3827                                  struct ocfs2_insert_type *insert)
3828 {
3829         int i = insert->ins_contig_index;
3830         unsigned int range;
3831         struct ocfs2_extent_rec *rec;
3832
3833         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3834
3835         if (insert->ins_split != SPLIT_NONE) {
3836                 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3837                 BUG_ON(i == -1);
3838                 rec = &el->l_recs[i];
3839                 ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3840                                         insert->ins_split, rec,
3841                                         insert_rec);
3842                 goto rotate;
3843         }
3844
3845         /*
3846          * Contiguous insert - either left or right.
3847          */
3848         if (insert->ins_contig != CONTIG_NONE) {
3849                 rec = &el->l_recs[i];
3850                 if (insert->ins_contig == CONTIG_LEFT) {
3851                         rec->e_blkno = insert_rec->e_blkno;
3852                         rec->e_cpos = insert_rec->e_cpos;
3853                 }
3854                 le16_add_cpu(&rec->e_leaf_clusters,
3855                              le16_to_cpu(insert_rec->e_leaf_clusters));
3856                 return;
3857         }
3858
3859         /*
3860          * Handle insert into an empty leaf.
3861          */
3862         if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3863             ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3864              ocfs2_is_empty_extent(&el->l_recs[0]))) {
3865                 el->l_recs[0] = *insert_rec;
3866                 el->l_next_free_rec = cpu_to_le16(1);
3867                 return;
3868         }
3869
3870         /*
3871          * Appending insert.
3872          */
3873         if (insert->ins_appending == APPEND_TAIL) {
3874                 i = le16_to_cpu(el->l_next_free_rec) - 1;
3875                 rec = &el->l_recs[i];
3876                 range = le32_to_cpu(rec->e_cpos)
3877                         + le16_to_cpu(rec->e_leaf_clusters);
3878                 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3879
3880                 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3881                                 le16_to_cpu(el->l_count),
3882                                 "owner %llu, depth %u, count %u, next free %u, "
3883                                 "rec.cpos %u, rec.clusters %u, "
3884                                 "insert.cpos %u, insert.clusters %u\n",
3885                                 ocfs2_metadata_cache_owner(et->et_ci),
3886                                 le16_to_cpu(el->l_tree_depth),
3887                                 le16_to_cpu(el->l_count),
3888                                 le16_to_cpu(el->l_next_free_rec),
3889                                 le32_to_cpu(el->l_recs[i].e_cpos),
3890                                 le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3891                                 le32_to_cpu(insert_rec->e_cpos),
3892                                 le16_to_cpu(insert_rec->e_leaf_clusters));
3893                 i++;
3894                 el->l_recs[i] = *insert_rec;
3895                 le16_add_cpu(&el->l_next_free_rec, 1);
3896                 return;
3897         }
3898
3899 rotate:
3900         /*
3901          * Ok, we have to rotate.
3902          *
3903          * At this point, it is safe to assume that inserting into an
3904          * empty leaf and appending to a leaf have both been handled
3905          * above.
3906          *
3907          * This leaf needs to have space, either by the empty 1st
3908          * extent record, or by virtue of an l_next_rec < l_count.
3909          */
3910         ocfs2_rotate_leaf(el, insert_rec);
3911 }
3912
3913 static void ocfs2_adjust_rightmost_records(handle_t *handle,
3914                                            struct ocfs2_extent_tree *et,
3915                                            struct ocfs2_path *path,
3916                                            struct ocfs2_extent_rec *insert_rec)
3917 {
3918         int ret, i, next_free;
3919         struct buffer_head *bh;
3920         struct ocfs2_extent_list *el;
3921         struct ocfs2_extent_rec *rec;
3922
3923         /*
3924          * Update everything except the leaf block.
3925          */
3926         for (i = 0; i < path->p_tree_depth; i++) {
3927                 bh = path->p_node[i].bh;
3928                 el = path->p_node[i].el;
3929
3930                 next_free = le16_to_cpu(el->l_next_free_rec);
3931                 if (next_free == 0) {
3932                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3933                                     "Owner %llu has a bad extent list",
3934                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3935                         ret = -EIO;
3936                         return;
3937                 }
3938
3939                 rec = &el->l_recs[next_free - 1];
3940
3941                 rec->e_int_clusters = insert_rec->e_cpos;
3942                 le32_add_cpu(&rec->e_int_clusters,
3943                              le16_to_cpu(insert_rec->e_leaf_clusters));
3944                 le32_add_cpu(&rec->e_int_clusters,
3945                              -le32_to_cpu(rec->e_cpos));
3946
3947                 ocfs2_journal_dirty(handle, bh);
3948         }
3949 }
3950
3951 static int ocfs2_append_rec_to_path(handle_t *handle,
3952                                     struct ocfs2_extent_tree *et,
3953                                     struct ocfs2_extent_rec *insert_rec,
3954                                     struct ocfs2_path *right_path,
3955                                     struct ocfs2_path **ret_left_path)
3956 {
3957         int ret, next_free;
3958         struct ocfs2_extent_list *el;
3959         struct ocfs2_path *left_path = NULL;
3960
3961         *ret_left_path = NULL;
3962
3963         /*
3964          * This shouldn't happen for non-trees. The extent rec cluster
3965          * count manipulation below only works for interior nodes.
3966          */
3967         BUG_ON(right_path->p_tree_depth == 0);
3968
3969         /*
3970          * If our appending insert is at the leftmost edge of a leaf,
3971          * then we might need to update the rightmost records of the
3972          * neighboring path.
3973          */
3974         el = path_leaf_el(right_path);
3975         next_free = le16_to_cpu(el->l_next_free_rec);
3976         if (next_free == 0 ||
3977             (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
3978                 u32 left_cpos;
3979
3980                 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3981                                                     right_path, &left_cpos);
3982                 if (ret) {
3983                         mlog_errno(ret);
3984                         goto out;
3985                 }
3986
3987                 trace_ocfs2_append_rec_to_path(
3988                         (unsigned long long)
3989                         ocfs2_metadata_cache_owner(et->et_ci),
3990                         le32_to_cpu(insert_rec->e_cpos),
3991                         left_cpos);
3992
3993                 /*
3994                  * No need to worry if the append is already in the
3995                  * leftmost leaf.
3996                  */
3997                 if (left_cpos) {
3998                         left_path = ocfs2_new_path_from_path(right_path);
3999                         if (!left_path) {
4000                                 ret = -ENOMEM;
4001                                 mlog_errno(ret);
4002                                 goto out;
4003                         }
4004
4005                         ret = ocfs2_find_path(et->et_ci, left_path,
4006                                               left_cpos);
4007                         if (ret) {
4008                                 mlog_errno(ret);
4009                                 goto out;
4010                         }
4011
4012                         /*
4013                          * ocfs2_insert_path() will pass the left_path to the
4014                          * journal for us.
4015                          */
4016                 }
4017         }
4018
4019         ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4020         if (ret) {
4021                 mlog_errno(ret);
4022                 goto out;
4023         }
4024
4025         ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4026
4027         *ret_left_path = left_path;
4028         ret = 0;
4029 out:
4030         if (ret != 0)
4031                 ocfs2_free_path(left_path);
4032
4033         return ret;
4034 }
4035
4036 static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4037                                struct ocfs2_path *left_path,
4038                                struct ocfs2_path *right_path,
4039                                struct ocfs2_extent_rec *split_rec,
4040                                enum ocfs2_split_type split)
4041 {
4042         int index;
4043         u32 cpos = le32_to_cpu(split_rec->e_cpos);
4044         struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4045         struct ocfs2_extent_rec *rec, *tmprec;
4046
4047         right_el = path_leaf_el(right_path);
4048         if (left_path)
4049                 left_el = path_leaf_el(left_path);
4050
4051         el = right_el;
4052         insert_el = right_el;
4053         index = ocfs2_search_extent_list(el, cpos);
4054         if (index != -1) {
4055                 if (index == 0 && left_path) {
4056                         BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4057
4058                         /*
4059                          * This typically means that the record
4060                          * started in the left path but moved to the
4061                          * right as a result of rotation. We either
4062                          * move the existing record to the left, or we
4063                          * do the later insert there.
4064                          *
4065                          * In this case, the left path should always
4066                          * exist as the rotate code will have passed
4067                          * it back for a post-insert update.
4068                          */
4069
4070                         if (split == SPLIT_LEFT) {
4071                                 /*
4072                                  * It's a left split. Since we know
4073                                  * that the rotate code gave us an
4074                                  * empty extent in the left path, we
4075                                  * can just do the insert there.
4076                                  */
4077                                 insert_el = left_el;
4078                         } else {
4079                                 /*
4080                                  * Right split - we have to move the
4081                                  * existing record over to the left
4082                                  * leaf. The insert will be into the
4083                                  * newly created empty extent in the
4084                                  * right leaf.
4085                                  */
4086                                 tmprec = &right_el->l_recs[index];
4087                                 ocfs2_rotate_leaf(left_el, tmprec);
4088                                 el = left_el;
4089
4090                                 memset(tmprec, 0, sizeof(*tmprec));
4091                                 index = ocfs2_search_extent_list(left_el, cpos);
4092                                 BUG_ON(index == -1);
4093                         }
4094                 }
4095         } else {
4096                 BUG_ON(!left_path);
4097                 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4098                 /*
4099                  * Left path is easy - we can just allow the insert to
4100                  * happen.
4101                  */
4102                 el = left_el;
4103                 insert_el = left_el;
4104                 index = ocfs2_search_extent_list(el, cpos);
4105                 BUG_ON(index == -1);
4106         }
4107
4108         rec = &el->l_recs[index];
4109         ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4110                                 split, rec, split_rec);
4111         ocfs2_rotate_leaf(insert_el, split_rec);
4112 }
4113
4114 /*
4115  * This function only does inserts on an allocation b-tree. For tree
4116  * depth = 0, ocfs2_insert_at_leaf() is called directly.
4117  *
4118  * right_path is the path we want to do the actual insert
4119  * in. left_path should only be passed in if we need to update that
4120  * portion of the tree after an edge insert.
4121  */
4122 static int ocfs2_insert_path(handle_t *handle,
4123                              struct ocfs2_extent_tree *et,
4124                              struct ocfs2_path *left_path,
4125                              struct ocfs2_path *right_path,
4126                              struct ocfs2_extent_rec *insert_rec,
4127                              struct ocfs2_insert_type *insert)
4128 {
4129         int ret, subtree_index;
4130         struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4131
4132         if (left_path) {
4133                 /*
4134                  * There's a chance that left_path got passed back to
4135                  * us without being accounted for in the
4136                  * journal. Extend our transaction here to be sure we
4137                  * can change those blocks.
4138                  */
4139                 ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4140                 if (ret < 0) {
4141                         mlog_errno(ret);
4142                         goto out;
4143                 }
4144
4145                 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4146                 if (ret < 0) {
4147                         mlog_errno(ret);
4148                         goto out;
4149                 }
4150         }
4151
4152         /*
4153          * Pass both paths to the journal. The majority of inserts
4154          * will be touching all components anyway.
4155          */
4156         ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4157         if (ret < 0) {
4158                 mlog_errno(ret);
4159                 goto out;
4160         }
4161
4162         if (insert->ins_split != SPLIT_NONE) {
4163                 /*
4164                  * We could call ocfs2_insert_at_leaf() for some types
4165                  * of splits, but it's easier to just let one separate
4166                  * function sort it all out.
4167                  */
4168                 ocfs2_split_record(et, left_path, right_path,
4169                                    insert_rec, insert->ins_split);
4170
4171                 /*
4172                  * Split might have modified either leaf and we don't
4173                  * have a guarantee that the later edge insert will
4174                  * dirty this for us.
4175                  */
4176                 if (left_path)
4177                         ocfs2_journal_dirty(handle,
4178                                             path_leaf_bh(left_path));
4179         } else
4180                 ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4181                                      insert);
4182
4183         ocfs2_journal_dirty(handle, leaf_bh);
4184
4185         if (left_path) {
4186                 /*
4187                  * The rotate code has indicated that we need to fix
4188                  * up portions of the tree after the insert.
4189                  *
4190                  * XXX: Should we extend the transaction here?
4191                  */
4192                 subtree_index = ocfs2_find_subtree_root(et, left_path,
4193                                                         right_path);
4194                 ocfs2_complete_edge_insert(handle, left_path, right_path,
4195                                            subtree_index);
4196         }
4197
4198         ret = 0;
4199 out:
4200         return ret;
4201 }
4202
4203 static int ocfs2_do_insert_extent(handle_t *handle,
4204                                   struct ocfs2_extent_tree *et,
4205                                   struct ocfs2_extent_rec *insert_rec,
4206                                   struct ocfs2_insert_type *type)
4207 {
4208         int ret, rotate = 0;
4209         u32 cpos;
4210         struct ocfs2_path *right_path = NULL;
4211         struct ocfs2_path *left_path = NULL;
4212         struct ocfs2_extent_list *el;
4213
4214         el = et->et_root_el;
4215
4216         ret = ocfs2_et_root_journal_access(handle, et,
4217                                            OCFS2_JOURNAL_ACCESS_WRITE);
4218         if (ret) {
4219                 mlog_errno(ret);
4220                 goto out;
4221         }
4222
4223         if (le16_to_cpu(el->l_tree_depth) == 0) {
4224                 ocfs2_insert_at_leaf(et, insert_rec, el, type);
4225                 goto out_update_clusters;
4226         }
4227
4228         right_path = ocfs2_new_path_from_et(et);
4229         if (!right_path) {
4230                 ret = -ENOMEM;
4231                 mlog_errno(ret);
4232                 goto out;
4233         }
4234
4235         /*
4236          * Determine the path to start with. Rotations need the
4237          * rightmost path, everything else can go directly to the
4238          * target leaf.
4239          */
4240         cpos = le32_to_cpu(insert_rec->e_cpos);
4241         if (type->ins_appending == APPEND_NONE &&
4242             type->ins_contig == CONTIG_NONE) {
4243                 rotate = 1;
4244                 cpos = UINT_MAX;
4245         }
4246
4247         ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4248         if (ret) {
4249                 mlog_errno(ret);
4250                 goto out;
4251         }
4252
4253         /*
4254          * Rotations and appends need special treatment - they modify
4255          * parts of the tree's above them.
4256          *
4257          * Both might pass back a path immediate to the left of the
4258          * one being inserted to. This will be cause
4259          * ocfs2_insert_path() to modify the rightmost records of
4260          * left_path to account for an edge insert.
4261          *
4262          * XXX: When modifying this code, keep in mind that an insert
4263          * can wind up skipping both of these two special cases...
4264          */
4265         if (rotate) {
4266                 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4267                                               le32_to_cpu(insert_rec->e_cpos),
4268                                               right_path, &left_path);
4269                 if (ret) {
4270                         mlog_errno(ret);
4271                         goto out;
4272                 }
4273
4274                 /*
4275                  * ocfs2_rotate_tree_right() might have extended the
4276                  * transaction without re-journaling our tree root.
4277                  */
4278                 ret = ocfs2_et_root_journal_access(handle, et,
4279                                                    OCFS2_JOURNAL_ACCESS_WRITE);
4280                 if (ret) {
4281                         mlog_errno(ret);
4282                         goto out;
4283                 }
4284         } else if (type->ins_appending == APPEND_TAIL
4285                    && type->ins_contig != CONTIG_LEFT) {
4286                 ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4287                                                right_path, &left_path);
4288                 if (ret) {
4289                         mlog_errno(ret);
4290                         goto out;
4291                 }
4292         }
4293
4294         ret = ocfs2_insert_path(handle, et, left_path, right_path,
4295                                 insert_rec, type);
4296         if (ret) {
4297                 mlog_errno(ret);
4298                 goto out;
4299         }
4300
4301 out_update_clusters:
4302         if (type->ins_split == SPLIT_NONE)
4303                 ocfs2_et_update_clusters(et,
4304                                          le16_to_cpu(insert_rec->e_leaf_clusters));
4305
4306         ocfs2_journal_dirty(handle, et->et_root_bh);
4307
4308 out:
4309         ocfs2_free_path(left_path);
4310         ocfs2_free_path(right_path);
4311
4312         return ret;
4313 }
4314
4315 static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4316                                struct ocfs2_path *path,
4317                                struct ocfs2_extent_list *el, int index,
4318                                struct ocfs2_extent_rec *split_rec,
4319                                struct ocfs2_merge_ctxt *ctxt)
4320 {
4321         int status = 0;
4322         enum ocfs2_contig_type ret = CONTIG_NONE;
4323         u32 left_cpos, right_cpos;
4324         struct ocfs2_extent_rec *rec = NULL;
4325         struct ocfs2_extent_list *new_el;
4326         struct ocfs2_path *left_path = NULL, *right_path = NULL;
4327         struct buffer_head *bh;
4328         struct ocfs2_extent_block *eb;
4329         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4330
4331         if (index > 0) {
4332                 rec = &el->l_recs[index - 1];
4333         } else if (path->p_tree_depth > 0) {
4334                 status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4335                 if (status)
4336                         goto exit;
4337
4338                 if (left_cpos != 0) {
4339                         left_path = ocfs2_new_path_from_path(path);
4340                         if (!left_path) {
4341                                 status = -ENOMEM;
4342                                 mlog_errno(status);
4343                                 goto exit;
4344                         }
4345
4346                         status = ocfs2_find_path(et->et_ci, left_path,
4347                                                  left_cpos);
4348                         if (status)
4349                                 goto free_left_path;
4350
4351                         new_el = path_leaf_el(left_path);
4352
4353                         if (le16_to_cpu(new_el->l_next_free_rec) !=
4354                             le16_to_cpu(new_el->l_count)) {
4355                                 bh = path_leaf_bh(left_path);
4356                                 eb = (struct ocfs2_extent_block *)bh->b_data;
4357                                 ocfs2_error(sb,
4358                                             "Extent block #%llu has an "
4359                                             "invalid l_next_free_rec of "
4360                                             "%d.  It should have "
4361                                             "matched the l_count of %d",
4362                                             (unsigned long long)le64_to_cpu(eb->h_blkno),
4363                                             le16_to_cpu(new_el->l_next_free_rec),
4364                                             le16_to_cpu(new_el->l_count));
4365                                 status = -EINVAL;
4366                                 goto free_left_path;
4367                         }
4368                         rec = &new_el->l_recs[
4369                                 le16_to_cpu(new_el->l_next_free_rec) - 1];
4370                 }
4371         }
4372
4373         /*
4374          * We're careful to check for an empty extent record here -
4375          * the merge code will know what to do if it sees one.
4376          */
4377         if (rec) {
4378                 if (index == 1 && ocfs2_is_empty_extent(rec)) {
4379                         if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4380                                 ret = CONTIG_RIGHT;
4381                 } else {
4382                         ret = ocfs2_et_extent_contig(et, rec, split_rec);
4383                 }
4384         }
4385
4386         rec = NULL;
4387         if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4388                 rec = &el->l_recs[index + 1];
4389         else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4390                  path->p_tree_depth > 0) {
4391                 status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4392                 if (status)
4393                         goto free_left_path;
4394
4395                 if (right_cpos == 0)
4396                         goto free_left_path;
4397
4398                 right_path = ocfs2_new_path_from_path(path);
4399                 if (!right_path) {
4400                         status = -ENOMEM;
4401                         mlog_errno(status);
4402                         goto free_left_path;
4403                 }
4404
4405                 status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4406                 if (status)
4407                         goto free_right_path;
4408
4409                 new_el = path_leaf_el(right_path);
4410                 rec = &new_el->l_recs[0];
4411                 if (ocfs2_is_empty_extent(rec)) {
4412                         if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4413                                 bh = path_leaf_bh(right_path);
4414                                 eb = (struct ocfs2_extent_block *)bh->b_data;
4415                                 ocfs2_error(sb,
4416                                             "Extent block #%llu has an "
4417                                             "invalid l_next_free_rec of %d",
4418                                             (unsigned long long)le64_to_cpu(eb->h_blkno),
4419                                             le16_to_cpu(new_el->l_next_free_rec));
4420                                 status = -EINVAL;
4421                                 goto free_right_path;
4422                         }
4423                         rec = &new_el->l_recs[1];
4424                 }
4425         }
4426
4427         if (rec) {
4428                 enum ocfs2_contig_type contig_type;
4429
4430                 contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4431
4432                 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4433                         ret = CONTIG_LEFTRIGHT;
4434                 else if (ret == CONTIG_NONE)
4435                         ret = contig_type;
4436         }
4437
4438 free_right_path:
4439         ocfs2_free_path(right_path);
4440 free_left_path:
4441         ocfs2_free_path(left_path);
4442 exit:
4443         if (status == 0)
4444                 ctxt->c_contig_type = ret;
4445
4446         return status;
4447 }
4448
4449 static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4450                                      struct ocfs2_insert_type *insert,
4451                                      struct ocfs2_extent_list *el,
4452                                      struct ocfs2_extent_rec *insert_rec)
4453 {
4454         int i;
4455         enum ocfs2_contig_type contig_type = CONTIG_NONE;
4456
4457         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4458
4459         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4460                 contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4461                                                      insert_rec);
4462                 if (contig_type != CONTIG_NONE) {
4463                         insert->ins_contig_index = i;
4464                         break;
4465                 }
4466         }
4467         insert->ins_contig = contig_type;
4468
4469         if (insert->ins_contig != CONTIG_NONE) {
4470                 struct ocfs2_extent_rec *rec =
4471                                 &el->l_recs[insert->ins_contig_index];
4472                 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4473                                    le16_to_cpu(insert_rec->e_leaf_clusters);
4474
4475                 /*
4476                  * Caller might want us to limit the size of extents, don't
4477                  * calculate contiguousness if we might exceed that limit.
4478                  */
4479                 if (et->et_max_leaf_clusters &&
4480                     (len > et->et_max_leaf_clusters))
4481                         insert->ins_contig = CONTIG_NONE;
4482         }
4483 }
4484
4485 /*
4486  * This should only be called against the righmost leaf extent list.
4487  *
4488  * ocfs2_figure_appending_type() will figure out whether we'll have to
4489  * insert at the tail of the rightmost leaf.
4490  *
4491  * This should also work against the root extent list for tree's with 0
4492  * depth. If we consider the root extent list to be the rightmost leaf node
4493  * then the logic here makes sense.
4494  */
4495 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4496                                         struct ocfs2_extent_list *el,
4497                                         struct ocfs2_extent_rec *insert_rec)
4498 {
4499         int i;
4500         u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4501         struct ocfs2_extent_rec *rec;
4502
4503         insert->ins_appending = APPEND_NONE;
4504
4505         BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4506
4507         if (!el->l_next_free_rec)
4508                 goto set_tail_append;
4509
4510         if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4511                 /* Were all records empty? */
4512                 if (le16_to_cpu(el->l_next_free_rec) == 1)
4513                         goto set_tail_append;
4514         }
4515
4516         i = le16_to_cpu(el->l_next_free_rec) - 1;
4517         rec = &el->l_recs[i];
4518
4519         if (cpos >=
4520             (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4521                 goto set_tail_append;
4522
4523         return;
4524
4525 set_tail_append:
4526         insert->ins_appending = APPEND_TAIL;
4527 }
4528
4529 /*
4530  * Helper function called at the beginning of an insert.
4531  *
4532  * This computes a few things that are commonly used in the process of
4533  * inserting into the btree:
4534  *   - Whether the new extent is contiguous with an existing one.
4535  *   - The current tree depth.
4536  *   - Whether the insert is an appending one.
4537  *   - The total # of free records in the tree.
4538  *
4539  * All of the information is stored on the ocfs2_insert_type
4540  * structure.
4541  */
4542 static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4543                                     struct buffer_head **last_eb_bh,
4544                                     struct ocfs2_extent_rec *insert_rec,
4545                                     int *free_records,
4546                                     struct ocfs2_insert_type *insert)
4547 {
4548         int ret;
4549         struct ocfs2_extent_block *eb;
4550         struct ocfs2_extent_list *el;
4551         struct ocfs2_path *path = NULL;
4552         struct buffer_head *bh = NULL;
4553
4554         insert->ins_split = SPLIT_NONE;
4555
4556         el = et->et_root_el;
4557         insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4558
4559         if (el->l_tree_depth) {
4560                 /*
4561                  * If we have tree depth, we read in the
4562                  * rightmost extent block ahead of time as
4563                  * ocfs2_figure_insert_type() and ocfs2_add_branch()
4564                  * may want it later.
4565                  */
4566                 ret = ocfs2_read_extent_block(et->et_ci,
4567                                               ocfs2_et_get_last_eb_blk(et),
4568                                               &bh);
4569                 if (ret) {
4570                         mlog_errno(ret);
4571                         goto out;
4572                 }
4573                 eb = (struct ocfs2_extent_block *) bh->b_data;
4574                 el = &eb->h_list;
4575         }
4576
4577         /*
4578          * Unless we have a contiguous insert, we'll need to know if
4579          * there is room left in our allocation tree for another
4580          * extent record.
4581          *
4582          * XXX: This test is simplistic, we can search for empty
4583          * extent records too.
4584          */
4585         *free_records = le16_to_cpu(el->l_count) -
4586                 le16_to_cpu(el->l_next_free_rec);
4587
4588         if (!insert->ins_tree_depth) {
4589                 ocfs2_figure_contig_type(et, insert, el, insert_rec);
4590                 ocfs2_figure_appending_type(insert, el, insert_rec);
4591                 return 0;
4592         }
4593
4594         path = ocfs2_new_path_from_et(et);
4595         if (!path) {
4596                 ret = -ENOMEM;
4597                 mlog_errno(ret);
4598                 goto out;
4599         }
4600
4601         /*
4602          * In the case that we're inserting past what the tree
4603          * currently accounts for, ocfs2_find_path() will return for
4604          * us the rightmost tree path. This is accounted for below in
4605          * the appending code.
4606          */
4607         ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4608         if (ret) {
4609                 mlog_errno(ret);
4610                 goto out;
4611         }
4612
4613         el = path_leaf_el(path);
4614
4615         /*
4616          * Now that we have the path, there's two things we want to determine:
4617          * 1) Contiguousness (also set contig_index if this is so)
4618          *
4619          * 2) Are we doing an append? We can trivially break this up
4620          *     into two types of appends: simple record append, or a
4621          *     rotate inside the tail leaf.
4622          */
4623         ocfs2_figure_contig_type(et, insert, el, insert_rec);
4624
4625         /*
4626          * The insert code isn't quite ready to deal with all cases of
4627          * left contiguousness. Specifically, if it's an insert into
4628          * the 1st record in a leaf, it will require the adjustment of
4629          * cluster count on the last record of the path directly to it's
4630          * left. For now, just catch that case and fool the layers
4631          * above us. This works just fine for tree_depth == 0, which
4632          * is why we allow that above.
4633          */
4634         if (insert->ins_contig == CONTIG_LEFT &&
4635             insert->ins_contig_index == 0)
4636                 insert->ins_contig = CONTIG_NONE;
4637
4638         /*
4639          * Ok, so we can simply compare against last_eb to figure out
4640          * whether the path doesn't exist. This will only happen in
4641          * the case that we're doing a tail append, so maybe we can
4642          * take advantage of that information somehow.
4643          */
4644         if (ocfs2_et_get_last_eb_blk(et) ==
4645             path_leaf_bh(path)->b_blocknr) {
4646                 /*
4647                  * Ok, ocfs2_find_path() returned us the rightmost
4648                  * tree path. This might be an appending insert. There are
4649                  * two cases:
4650                  *    1) We're doing a true append at the tail:
4651                  *      -This might even be off the end of the leaf
4652                  *    2) We're "appending" by rotating in the tail
4653                  */
4654                 ocfs2_figure_appending_type(insert, el, insert_rec);
4655         }
4656
4657 out:
4658         ocfs2_free_path(path);
4659
4660         if (ret == 0)
4661                 *last_eb_bh = bh;
4662         else
4663                 brelse(bh);
4664         return ret;
4665 }
4666
4667 /*
4668  * Insert an extent into a btree.
4669  *
4670  * The caller needs to update the owning btree's cluster count.
4671  */
4672 int ocfs2_insert_extent(handle_t *handle,
4673                         struct ocfs2_extent_tree *et,
4674                         u32 cpos,
4675                         u64 start_blk,
4676                         u32 new_clusters,
4677                         u8 flags,
4678                         struct ocfs2_alloc_context *meta_ac)
4679 {
4680         int status;
4681         int uninitialized_var(free_records);
4682         struct buffer_head *last_eb_bh = NULL;
4683         struct ocfs2_insert_type insert = {0, };
4684         struct ocfs2_extent_rec rec;
4685
4686         trace_ocfs2_insert_extent_start(
4687                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4688                 cpos, new_clusters);
4689
4690         memset(&rec, 0, sizeof(rec));
4691         rec.e_cpos = cpu_to_le32(cpos);
4692         rec.e_blkno = cpu_to_le64(start_blk);
4693         rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4694         rec.e_flags = flags;
4695         status = ocfs2_et_insert_check(et, &rec);
4696         if (status) {
4697                 mlog_errno(status);
4698                 goto bail;
4699         }
4700
4701         status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4702                                           &free_records, &insert);
4703         if (status < 0) {
4704                 mlog_errno(status);
4705                 goto bail;
4706         }
4707
4708         trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4709                                   insert.ins_contig_index, free_records,
4710                                   insert.ins_tree_depth);
4711
4712         if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4713                 status = ocfs2_grow_tree(handle, et,
4714                                          &insert.ins_tree_depth, &last_eb_bh,
4715                                          meta_ac);
4716                 if (status) {
4717                         mlog_errno(status);
4718                         goto bail;
4719                 }
4720         }
4721
4722         /* Finally, we can add clusters. This might rotate the tree for us. */
4723         status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4724         if (status < 0)
4725                 mlog_errno(status);
4726         else
4727                 ocfs2_et_extent_map_insert(et, &rec);
4728
4729 bail:
4730         brelse(last_eb_bh);
4731
4732         return status;
4733 }
4734
4735 /*
4736  * Allcate and add clusters into the extent b-tree.
4737  * The new clusters(clusters_to_add) will be inserted at logical_offset.
4738  * The extent b-tree's root is specified by et, and
4739  * it is not limited to the file storage. Any extent tree can use this
4740  * function if it implements the proper ocfs2_extent_tree.
4741  */
4742 int ocfs2_add_clusters_in_btree(handle_t *handle,
4743                                 struct ocfs2_extent_tree *et,
4744                                 u32 *logical_offset,
4745                                 u32 clusters_to_add,
4746                                 int mark_unwritten,
4747                                 struct ocfs2_alloc_context *data_ac,
4748                                 struct ocfs2_alloc_context *meta_ac,
4749                                 enum ocfs2_alloc_restarted *reason_ret)
4750 {
4751         int status = 0, err = 0;
4752         int need_free = 0;
4753         int free_extents;
4754         enum ocfs2_alloc_restarted reason = RESTART_NONE;
4755         u32 bit_off, num_bits;
4756         u64 block;
4757         u8 flags = 0;
4758         struct ocfs2_super *osb =
4759                 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4760
4761         BUG_ON(!clusters_to_add);
4762
4763         if (mark_unwritten)
4764                 flags = OCFS2_EXT_UNWRITTEN;
4765
4766         free_extents = ocfs2_num_free_extents(osb, et);
4767         if (free_extents < 0) {
4768                 status = free_extents;
4769                 mlog_errno(status);
4770                 goto leave;
4771         }
4772
4773         /* there are two cases which could cause us to EAGAIN in the
4774          * we-need-more-metadata case:
4775          * 1) we haven't reserved *any*
4776          * 2) we are so fragmented, we've needed to add metadata too
4777          *    many times. */
4778         if (!free_extents && !meta_ac) {
4779                 err = -1;
4780                 status = -EAGAIN;
4781                 reason = RESTART_META;
4782                 goto leave;
4783         } else if ((!free_extents)
4784                    && (ocfs2_alloc_context_bits_left(meta_ac)
4785                        < ocfs2_extend_meta_needed(et->et_root_el))) {
4786                 err = -2;
4787                 status = -EAGAIN;
4788                 reason = RESTART_META;
4789                 goto leave;
4790         }
4791
4792         status = __ocfs2_claim_clusters(handle, data_ac, 1,
4793                                         clusters_to_add, &bit_off, &num_bits);
4794         if (status < 0) {
4795                 if (status != -ENOSPC)
4796                         mlog_errno(status);
4797                 goto leave;
4798         }
4799
4800         BUG_ON(num_bits > clusters_to_add);
4801
4802         /* reserve our write early -- insert_extent may update the tree root */
4803         status = ocfs2_et_root_journal_access(handle, et,
4804                                               OCFS2_JOURNAL_ACCESS_WRITE);
4805         if (status < 0) {
4806                 mlog_errno(status);
4807                 need_free = 1;
4808                 goto bail;
4809         }
4810
4811         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4812         trace_ocfs2_add_clusters_in_btree(
4813              (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4814              bit_off, num_bits);
4815         status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4816                                      num_bits, flags, meta_ac);
4817         if (status < 0) {
4818                 mlog_errno(status);
4819                 need_free = 1;
4820                 goto bail;
4821         }
4822
4823         ocfs2_journal_dirty(handle, et->et_root_bh);
4824
4825         clusters_to_add -= num_bits;
4826         *logical_offset += num_bits;
4827
4828         if (clusters_to_add) {
4829                 err = clusters_to_add;
4830                 status = -EAGAIN;
4831                 reason = RESTART_TRANS;
4832         }
4833
4834 bail:
4835         if (need_free) {
4836                 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4837                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4838                                         bit_off, num_bits);
4839                 else
4840                         ocfs2_free_clusters(handle,
4841                                         data_ac->ac_inode,
4842                                         data_ac->ac_bh,
4843                                         ocfs2_clusters_to_blocks(osb->sb, bit_off),
4844                                         num_bits);
4845         }
4846
4847 leave:
4848         if (reason_ret)
4849                 *reason_ret = reason;
4850         trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4851         return status;
4852 }
4853
4854 static void ocfs2_make_right_split_rec(struct super_block *sb,
4855                                        struct ocfs2_extent_rec *split_rec,
4856                                        u32 cpos,
4857                                        struct ocfs2_extent_rec *rec)
4858 {
4859         u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4860         u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4861
4862         memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4863
4864         split_rec->e_cpos = cpu_to_le32(cpos);
4865         split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4866
4867         split_rec->e_blkno = rec->e_blkno;
4868         le64_add_cpu(&split_rec->e_blkno,
4869                      ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4870
4871         split_rec->e_flags = rec->e_flags;
4872 }
4873
4874 static int ocfs2_split_and_insert(handle_t *handle,
4875                                   struct ocfs2_extent_tree *et,
4876                                   struct ocfs2_path *path,
4877                                   struct buffer_head **last_eb_bh,
4878                                   int split_index,
4879                                   struct ocfs2_extent_rec *orig_split_rec,
4880                                   struct ocfs2_alloc_context *meta_ac)
4881 {
4882         int ret = 0, depth;
4883         unsigned int insert_range, rec_range, do_leftright = 0;
4884         struct ocfs2_extent_rec tmprec;
4885         struct ocfs2_extent_list *rightmost_el;
4886         struct ocfs2_extent_rec rec;
4887         struct ocfs2_extent_rec split_rec = *orig_split_rec;
4888         struct ocfs2_insert_type insert;
4889         struct ocfs2_extent_block *eb;
4890
4891 leftright:
4892         /*
4893          * Store a copy of the record on the stack - it might move
4894          * around as the tree is manipulated below.
4895          */
4896         rec = path_leaf_el(path)->l_recs[split_index];
4897
4898         rightmost_el = et->et_root_el;
4899
4900         depth = le16_to_cpu(rightmost_el->l_tree_depth);
4901         if (depth) {
4902                 BUG_ON(!(*last_eb_bh));
4903                 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4904                 rightmost_el = &eb->h_list;
4905         }
4906
4907         if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4908             le16_to_cpu(rightmost_el->l_count)) {
4909                 ret = ocfs2_grow_tree(handle, et,
4910                                       &depth, last_eb_bh, meta_ac);
4911                 if (ret) {
4912                         mlog_errno(ret);
4913                         goto out;
4914                 }
4915         }
4916
4917         memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4918         insert.ins_appending = APPEND_NONE;
4919         insert.ins_contig = CONTIG_NONE;
4920         insert.ins_tree_depth = depth;
4921
4922         insert_range = le32_to_cpu(split_rec.e_cpos) +
4923                 le16_to_cpu(split_rec.e_leaf_clusters);
4924         rec_range = le32_to_cpu(rec.e_cpos) +
4925                 le16_to_cpu(rec.e_leaf_clusters);
4926
4927         if (split_rec.e_cpos == rec.e_cpos) {
4928                 insert.ins_split = SPLIT_LEFT;
4929         } else if (insert_range == rec_range) {
4930                 insert.ins_split = SPLIT_RIGHT;
4931         } else {
4932                 /*
4933                  * Left/right split. We fake this as a right split
4934                  * first and then make a second pass as a left split.
4935                  */
4936                 insert.ins_split = SPLIT_RIGHT;
4937
4938                 ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4939                                            &tmprec, insert_range, &rec);
4940
4941                 split_rec = tmprec;
4942
4943                 BUG_ON(do_leftright);
4944                 do_leftright = 1;
4945         }
4946
4947         ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4948         if (ret) {
4949                 mlog_errno(ret);
4950                 goto out;
4951         }
4952
4953         if (do_leftright == 1) {
4954                 u32 cpos;
4955                 struct ocfs2_extent_list *el;
4956
4957                 do_leftright++;
4958                 split_rec = *orig_split_rec;
4959
4960                 ocfs2_reinit_path(path, 1);
4961
4962                 cpos = le32_to_cpu(split_rec.e_cpos);
4963                 ret = ocfs2_find_path(et->et_ci, path, cpos);
4964                 if (ret) {
4965                         mlog_errno(ret);
4966                         goto out;
4967                 }
4968
4969                 el = path_leaf_el(path);
4970                 split_index = ocfs2_search_extent_list(el, cpos);
4971                 if (split_index == -1) {
4972                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
4973                                         "Owner %llu has an extent at cpos %u "
4974                                         "which can no longer be found.\n",
4975                                         (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4976                                         cpos);
4977                         ret = -EROFS;
4978                         goto out;
4979                 }
4980                 goto leftright;
4981         }
4982 out:
4983
4984         return ret;
4985 }
4986
4987 static int ocfs2_replace_extent_rec(handle_t *handle,
4988                                     struct ocfs2_extent_tree *et,
4989                                     struct ocfs2_path *path,
4990                                     struct ocfs2_extent_list *el,
4991                                     int split_index,
4992                                     struct ocfs2_extent_rec *split_rec)
4993 {
4994         int ret;
4995
4996         ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
4997                                            path_num_items(path) - 1);
4998         if (ret) {
4999                 mlog_errno(ret);
5000                 goto out;
5001         }
5002
5003         el->l_recs[split_index] = *split_rec;
5004
5005         ocfs2_journal_dirty(handle, path_leaf_bh(path));
5006 out:
5007         return ret;
5008 }
5009
5010 /*
5011  * Split part or all of the extent record at split_index in the leaf
5012  * pointed to by path. Merge with the contiguous extent record if needed.
5013  *
5014  * Care is taken to handle contiguousness so as to not grow the tree.
5015  *
5016  * meta_ac is not strictly necessary - we only truly need it if growth
5017  * of the tree is required. All other cases will degrade into a less
5018  * optimal tree layout.
5019  *
5020  * last_eb_bh should be the rightmost leaf block for any extent
5021  * btree. Since a split may grow the tree or a merge might shrink it,
5022  * the caller cannot trust the contents of that buffer after this call.
5023  *
5024  * This code is optimized for readability - several passes might be
5025  * made over certain portions of the tree. All of those blocks will
5026  * have been brought into cache (and pinned via the journal), so the
5027  * extra overhead is not expressed in terms of disk reads.
5028  */
5029 int ocfs2_split_extent(handle_t *handle,
5030                        struct ocfs2_extent_tree *et,
5031                        struct ocfs2_path *path,
5032                        int split_index,
5033                        struct ocfs2_extent_rec *split_rec,
5034                        struct ocfs2_alloc_context *meta_ac,
5035                        struct ocfs2_cached_dealloc_ctxt *dealloc)
5036 {
5037         int ret = 0;
5038         struct ocfs2_extent_list *el = path_leaf_el(path);
5039         struct buffer_head *last_eb_bh = NULL;
5040         struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5041         struct ocfs2_merge_ctxt ctxt;
5042         struct ocfs2_extent_list *rightmost_el;
5043
5044         if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5045             ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5046              (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5047                 ret = -EIO;
5048                 mlog_errno(ret);
5049                 goto out;
5050         }
5051
5052         ret = ocfs2_figure_merge_contig_type(et, path, el,
5053                                              split_index,
5054                                              split_rec,
5055                                              &ctxt);
5056         if (ret) {
5057                 mlog_errno(ret);
5058                 goto out;
5059         }
5060
5061         /*
5062          * The core merge / split code wants to know how much room is
5063          * left in this allocation tree, so we pass the
5064          * rightmost extent list.
5065          */
5066         if (path->p_tree_depth) {
5067                 struct ocfs2_extent_block *eb;
5068
5069                 ret = ocfs2_read_extent_block(et->et_ci,
5070                                               ocfs2_et_get_last_eb_blk(et),
5071                                               &last_eb_bh);
5072                 if (ret) {
5073                         mlog_errno(ret);
5074                         goto out;
5075                 }
5076
5077                 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5078                 rightmost_el = &eb->h_list;
5079         } else
5080                 rightmost_el = path_root_el(path);
5081
5082         if (rec->e_cpos == split_rec->e_cpos &&
5083             rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5084                 ctxt.c_split_covers_rec = 1;
5085         else
5086                 ctxt.c_split_covers_rec = 0;
5087
5088         ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5089
5090         trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5091                                  ctxt.c_has_empty_extent,
5092                                  ctxt.c_split_covers_rec);
5093
5094         if (ctxt.c_contig_type == CONTIG_NONE) {
5095                 if (ctxt.c_split_covers_rec)
5096                         ret = ocfs2_replace_extent_rec(handle, et, path, el,
5097                                                        split_index, split_rec);
5098                 else
5099                         ret = ocfs2_split_and_insert(handle, et, path,
5100                                                      &last_eb_bh, split_index,
5101                                                      split_rec, meta_ac);
5102                 if (ret)
5103                         mlog_errno(ret);
5104         } else {
5105                 ret = ocfs2_try_to_merge_extent(handle, et, path,
5106                                                 split_index, split_rec,
5107                                                 dealloc, &ctxt);
5108                 if (ret)
5109                         mlog_errno(ret);
5110         }
5111
5112 out:
5113         brelse(last_eb_bh);
5114         return ret;
5115 }
5116
5117 /*
5118  * Change the flags of the already-existing extent at cpos for len clusters.
5119  *
5120  * new_flags: the flags we want to set.
5121  * clear_flags: the flags we want to clear.
5122  * phys: the new physical offset we want this new extent starts from.
5123  *
5124  * If the existing extent is larger than the request, initiate a
5125  * split. An attempt will be made at merging with adjacent extents.
5126  *
5127  * The caller is responsible for passing down meta_ac if we'll need it.
5128  */
5129 int ocfs2_change_extent_flag(handle_t *handle,
5130                              struct ocfs2_extent_tree *et,
5131                              u32 cpos, u32 len, u32 phys,
5132                              struct ocfs2_alloc_context *meta_ac,
5133                              struct ocfs2_cached_dealloc_ctxt *dealloc,
5134                              int new_flags, int clear_flags)
5135 {
5136         int ret, index;
5137         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5138         u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5139         struct ocfs2_extent_rec split_rec;
5140         struct ocfs2_path *left_path = NULL;
5141         struct ocfs2_extent_list *el;
5142         struct ocfs2_extent_rec *rec;
5143
5144         left_path = ocfs2_new_path_from_et(et);
5145         if (!left_path) {
5146                 ret = -ENOMEM;
5147                 mlog_errno(ret);
5148                 goto out;
5149         }
5150
5151         ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5152         if (ret) {
5153                 mlog_errno(ret);
5154                 goto out;
5155         }
5156         el = path_leaf_el(left_path);
5157
5158         index = ocfs2_search_extent_list(el, cpos);
5159         if (index == -1) {
5160                 ocfs2_error(sb,
5161                             "Owner %llu has an extent at cpos %u which can no "
5162                             "longer be found.\n",
5163                              (unsigned long long)
5164                              ocfs2_metadata_cache_owner(et->et_ci), cpos);
5165                 ret = -EROFS;
5166                 goto out;
5167         }
5168
5169         ret = -EIO;
5170         rec = &el->l_recs[index];
5171         if (new_flags && (rec->e_flags & new_flags)) {
5172                 mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5173                      "extent that already had them",
5174                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5175                      new_flags);
5176                 goto out;
5177         }
5178
5179         if (clear_flags && !(rec->e_flags & clear_flags)) {
5180                 mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5181                      "extent that didn't have them",
5182                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5183                      clear_flags);
5184                 goto out;
5185         }
5186
5187         memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5188         split_rec.e_cpos = cpu_to_le32(cpos);
5189         split_rec.e_leaf_clusters = cpu_to_le16(len);
5190         split_rec.e_blkno = cpu_to_le64(start_blkno);
5191         split_rec.e_flags = rec->e_flags;
5192         if (new_flags)
5193                 split_rec.e_flags |= new_flags;
5194         if (clear_flags)
5195                 split_rec.e_flags &= ~clear_flags;
5196
5197         ret = ocfs2_split_extent(handle, et, left_path,
5198                                  index, &split_rec, meta_ac,
5199                                  dealloc);
5200         if (ret)
5201                 mlog_errno(ret);
5202
5203 out:
5204         ocfs2_free_path(left_path);
5205         return ret;
5206
5207 }
5208
5209 /*
5210  * Mark the already-existing extent at cpos as written for len clusters.
5211  * This removes the unwritten extent flag.
5212  *
5213  * If the existing extent is larger than the request, initiate a
5214  * split. An attempt will be made at merging with adjacent extents.
5215  *
5216  * The caller is responsible for passing down meta_ac if we'll need it.
5217  */
5218 int ocfs2_mark_extent_written(struct inode *inode,
5219                               struct ocfs2_extent_tree *et,
5220                               handle_t *handle, u32 cpos, u32 len, u32 phys,
5221                               struct ocfs2_alloc_context *meta_ac,
5222                               struct ocfs2_cached_dealloc_ctxt *dealloc)
5223 {
5224         int ret;
5225
5226         trace_ocfs2_mark_extent_written(
5227                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
5228                 cpos, len, phys);
5229
5230         if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5231                 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents "
5232                             "that are being written to, but the feature bit "
5233                             "is not set in the super block.",
5234                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
5235                 ret = -EROFS;
5236                 goto out;
5237         }
5238
5239         /*
5240          * XXX: This should be fixed up so that we just re-insert the
5241          * next extent records.
5242          */
5243         ocfs2_et_extent_map_truncate(et, 0);
5244
5245         ret = ocfs2_change_extent_flag(handle, et, cpos,
5246                                        len, phys, meta_ac, dealloc,
5247                                        0, OCFS2_EXT_UNWRITTEN);
5248         if (ret)
5249                 mlog_errno(ret);
5250
5251 out:
5252         return ret;
5253 }
5254
5255 static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5256                             struct ocfs2_path *path,
5257                             int index, u32 new_range,
5258                             struct ocfs2_alloc_context *meta_ac)
5259 {
5260         int ret, depth, credits;
5261         struct buffer_head *last_eb_bh = NULL;
5262         struct ocfs2_extent_block *eb;
5263         struct ocfs2_extent_list *rightmost_el, *el;
5264         struct ocfs2_extent_rec split_rec;
5265         struct ocfs2_extent_rec *rec;
5266         struct ocfs2_insert_type insert;
5267
5268         /*
5269          * Setup the record to split before we grow the tree.
5270          */
5271         el = path_leaf_el(path);
5272         rec = &el->l_recs[index];
5273         ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5274                                    &split_rec, new_range, rec);
5275
5276         depth = path->p_tree_depth;
5277         if (depth > 0) {
5278                 ret = ocfs2_read_extent_block(et->et_ci,
5279                                               ocfs2_et_get_last_eb_blk(et),
5280                                               &last_eb_bh);
5281                 if (ret < 0) {
5282                         mlog_errno(ret);
5283                         goto out;
5284                 }
5285
5286                 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5287                 rightmost_el = &eb->h_list;
5288         } else
5289                 rightmost_el = path_leaf_el(path);
5290
5291         credits = path->p_tree_depth +
5292                   ocfs2_extend_meta_needed(et->et_root_el);
5293         ret = ocfs2_extend_trans(handle, credits);
5294         if (ret) {
5295                 mlog_errno(ret);
5296                 goto out;
5297         }
5298
5299         if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5300             le16_to_cpu(rightmost_el->l_count)) {
5301                 ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5302                                       meta_ac);
5303                 if (ret) {
5304                         mlog_errno(ret);
5305                         goto out;
5306                 }
5307         }
5308
5309         memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5310         insert.ins_appending = APPEND_NONE;
5311         insert.ins_contig = CONTIG_NONE;
5312         insert.ins_split = SPLIT_RIGHT;
5313         insert.ins_tree_depth = depth;
5314
5315         ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5316         if (ret)
5317                 mlog_errno(ret);
5318
5319 out:
5320         brelse(last_eb_bh);
5321         return ret;
5322 }
5323
5324 static int ocfs2_truncate_rec(handle_t *handle,
5325                               struct ocfs2_extent_tree *et,
5326                               struct ocfs2_path *path, int index,
5327                               struct ocfs2_cached_dealloc_ctxt *dealloc,
5328                               u32 cpos, u32 len)
5329 {
5330         int ret;
5331         u32 left_cpos, rec_range, trunc_range;
5332         int wants_rotate = 0, is_rightmost_tree_rec = 0;
5333         struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5334         struct ocfs2_path *left_path = NULL;
5335         struct ocfs2_extent_list *el = path_leaf_el(path);
5336         struct ocfs2_extent_rec *rec;
5337         struct ocfs2_extent_block *eb;
5338
5339         if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5340                 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5341                 if (ret) {
5342                         mlog_errno(ret);
5343                         goto out;
5344                 }
5345
5346                 index--;
5347         }
5348
5349         if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5350             path->p_tree_depth) {
5351                 /*
5352                  * Check whether this is the rightmost tree record. If
5353                  * we remove all of this record or part of its right
5354                  * edge then an update of the record lengths above it
5355                  * will be required.
5356                  */
5357                 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5358                 if (eb->h_next_leaf_blk == 0)
5359                         is_rightmost_tree_rec = 1;
5360         }
5361
5362         rec = &el->l_recs[index];
5363         if (index == 0 && path->p_tree_depth &&
5364             le32_to_cpu(rec->e_cpos) == cpos) {
5365                 /*
5366                  * Changing the leftmost offset (via partial or whole
5367                  * record truncate) of an interior (or rightmost) path
5368                  * means we have to update the subtree that is formed
5369                  * by this leaf and the one to it's left.
5370                  *
5371                  * There are two cases we can skip:
5372                  *   1) Path is the leftmost one in our btree.
5373                  *   2) The leaf is rightmost and will be empty after
5374                  *      we remove the extent record - the rotate code
5375                  *      knows how to update the newly formed edge.
5376                  */
5377
5378                 ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5379                 if (ret) {
5380                         mlog_errno(ret);
5381                         goto out;
5382                 }
5383
5384                 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5385                         left_path = ocfs2_new_path_from_path(path);
5386                         if (!left_path) {
5387                                 ret = -ENOMEM;
5388                                 mlog_errno(ret);
5389                                 goto out;
5390                         }
5391
5392                         ret = ocfs2_find_path(et->et_ci, left_path,
5393                                               left_cpos);
5394                         if (ret) {
5395                                 mlog_errno(ret);
5396                                 goto out;
5397                         }
5398                 }
5399         }
5400
5401         ret = ocfs2_extend_rotate_transaction(handle, 0,
5402                                               handle->h_buffer_credits,
5403                                               path);
5404         if (ret) {
5405                 mlog_errno(ret);
5406                 goto out;
5407         }
5408
5409         ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5410         if (ret) {
5411                 mlog_errno(ret);
5412                 goto out;
5413         }
5414
5415         ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5416         if (ret) {
5417                 mlog_errno(ret);
5418                 goto out;
5419         }
5420
5421         rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5422         trunc_range = cpos + len;
5423
5424         if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5425                 int next_free;
5426
5427                 memset(rec, 0, sizeof(*rec));
5428                 ocfs2_cleanup_merge(el, index);
5429                 wants_rotate = 1;
5430
5431                 next_free = le16_to_cpu(el->l_next_free_rec);
5432                 if (is_rightmost_tree_rec && next_free > 1) {
5433                         /*
5434                          * We skip the edge update if this path will
5435                          * be deleted by the rotate code.
5436                          */
5437                         rec = &el->l_recs[next_free - 1];
5438                         ocfs2_adjust_rightmost_records(handle, et, path,
5439                                                        rec);
5440                 }
5441         } else if (le32_to_cpu(rec->e_cpos) == cpos) {
5442                 /* Remove leftmost portion of the record. */
5443                 le32_add_cpu(&rec->e_cpos, len);
5444                 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5445                 le16_add_cpu(&rec->e_leaf_clusters, -len);
5446         } else if (rec_range == trunc_range) {
5447                 /* Remove rightmost portion of the record */
5448                 le16_add_cpu(&rec->e_leaf_clusters, -len);
5449                 if (is_rightmost_tree_rec)
5450                         ocfs2_adjust_rightmost_records(handle, et, path, rec);
5451         } else {
5452                 /* Caller should have trapped this. */
5453                 mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5454                      "(%u, %u)\n",
5455                      (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5456                      le32_to_cpu(rec->e_cpos),
5457                      le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5458                 BUG();
5459         }
5460
5461         if (left_path) {
5462                 int subtree_index;
5463
5464                 subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5465                 ocfs2_complete_edge_insert(handle, left_path, path,
5466                                            subtree_index);
5467         }
5468
5469         ocfs2_journal_dirty(handle, path_leaf_bh(path));
5470
5471         ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5472         if (ret) {
5473                 mlog_errno(ret);
5474                 goto out;
5475         }
5476
5477 out:
5478         ocfs2_free_path(left_path);
5479         return ret;
5480 }
5481
5482 int ocfs2_remove_extent(handle_t *handle,
5483                         struct ocfs2_extent_tree *et,
5484                         u32 cpos, u32 len,
5485                         struct ocfs2_alloc_context *meta_ac,
5486                         struct ocfs2_cached_dealloc_ctxt *dealloc)
5487 {
5488         int ret, index;
5489         u32 rec_range, trunc_range;
5490         struct ocfs2_extent_rec *rec;
5491         struct ocfs2_extent_list *el;
5492         struct ocfs2_path *path = NULL;
5493
5494         /*
5495          * XXX: Why are we truncating to 0 instead of wherever this
5496          * affects us?
5497          */
5498         ocfs2_et_extent_map_truncate(et, 0);
5499
5500         path = ocfs2_new_path_from_et(et);
5501         if (!path) {
5502                 ret = -ENOMEM;
5503                 mlog_errno(ret);
5504                 goto out;
5505         }
5506
5507         ret = ocfs2_find_path(et->et_ci, path, cpos);
5508         if (ret) {
5509                 mlog_errno(ret);
5510                 goto out;
5511         }
5512
5513         el = path_leaf_el(path);
5514         index = ocfs2_search_extent_list(el, cpos);
5515         if (index == -1) {
5516                 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5517                             "Owner %llu has an extent at cpos %u which can no "
5518                             "longer be found.\n",
5519                             (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5520                             cpos);
5521                 ret = -EROFS;
5522                 goto out;
5523         }
5524
5525         /*
5526          * We have 3 cases of extent removal:
5527          *   1) Range covers the entire extent rec
5528          *   2) Range begins or ends on one edge of the extent rec
5529          *   3) Range is in the middle of the extent rec (no shared edges)
5530          *
5531          * For case 1 we remove the extent rec and left rotate to
5532          * fill the hole.
5533          *
5534          * For case 2 we just shrink the existing extent rec, with a
5535          * tree update if the shrinking edge is also the edge of an
5536          * extent block.
5537          *
5538          * For case 3 we do a right split to turn the extent rec into
5539          * something case 2 can handle.
5540          */
5541         rec = &el->l_recs[index];
5542         rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5543         trunc_range = cpos + len;
5544
5545         BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5546
5547         trace_ocfs2_remove_extent(
5548                 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5549                 cpos, len, index, le32_to_cpu(rec->e_cpos),
5550                 ocfs2_rec_clusters(el, rec));
5551
5552         if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5553                 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5554                                          cpos, len);
5555                 if (ret) {
5556                         mlog_errno(ret);
5557                         goto out;
5558                 }
5559         } else {
5560                 ret = ocfs2_split_tree(handle, et, path, index,
5561                                        trunc_range, meta_ac);
5562                 if (ret) {
5563                         mlog_errno(ret);
5564                         goto out;
5565                 }
5566
5567                 /*
5568                  * The split could have manipulated the tree enough to
5569                  * move the record location, so we have to look for it again.
5570                  */
5571                 ocfs2_reinit_path(path, 1);
5572
5573                 ret = ocfs2_find_path(et->et_ci, path, cpos);
5574                 if (ret) {
5575                         mlog_errno(ret);
5576                         goto out;
5577                 }
5578
5579                 el = path_leaf_el(path);
5580                 index = ocfs2_search_extent_list(el, cpos);
5581                 if (index == -1) {
5582                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5583                                     "Owner %llu: split at cpos %u lost record.",
5584                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5585                                     cpos);
5586                         ret = -EROFS;
5587                         goto out;
5588                 }
5589
5590                 /*
5591                  * Double check our values here. If anything is fishy,
5592                  * it's easier to catch it at the top level.
5593                  */
5594                 rec = &el->l_recs[index];
5595                 rec_range = le32_to_cpu(rec->e_cpos) +
5596                         ocfs2_rec_clusters(el, rec);
5597                 if (rec_range != trunc_range) {
5598                         ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5599                                     "Owner %llu: error after split at cpos %u"
5600                                     "trunc len %u, existing record is (%u,%u)",
5601                                     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5602                                     cpos, len, le32_to_cpu(rec->e_cpos),
5603                                     ocfs2_rec_clusters(el, rec));
5604                         ret = -EROFS;
5605                         goto out;
5606                 }
5607
5608                 ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5609                                          cpos, len);
5610                 if (ret) {
5611                         mlog_errno(ret);
5612                         goto out;
5613                 }
5614         }
5615
5616 out:
5617         ocfs2_free_path(path);
5618         return ret;
5619 }
5620
5621 /*
5622  * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5623  * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5624  * number to reserve some extra blocks, and it only handles meta
5625  * data allocations.
5626  *
5627  * Currently, only ocfs2_remove_btree_range() uses it for truncating
5628  * and punching holes.
5629  */
5630 static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5631                                               struct ocfs2_extent_tree *et,
5632                                               u32 extents_to_split,
5633                                               struct ocfs2_alloc_context **ac,
5634                                               int extra_blocks)
5635 {
5636         int ret = 0, num_free_extents;
5637         unsigned int max_recs_needed = 2 * extents_to_split;
5638         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5639
5640         *ac = NULL;
5641
5642         num_free_extents = ocfs2_num_free_extents(osb, et);
5643         if (num_free_extents < 0) {
5644                 ret = num_free_extents;
5645                 mlog_errno(ret);
5646                 goto out;
5647         }
5648
5649         if (!num_free_extents ||
5650             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5651                 extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5652
5653         if (extra_blocks) {
5654                 ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5655                 if (ret < 0) {
5656                         if (ret != -ENOSPC)
5657                                 mlog_errno(ret);
5658                         goto out;
5659                 }
5660         }
5661
5662 out:
5663         if (ret) {
5664                 if (*ac) {
5665                         ocfs2_free_alloc_context(*ac);
5666                         *ac = NULL;
5667                 }
5668         }
5669
5670         return ret;
5671 }
5672
5673 int ocfs2_remove_btree_range(struct inode *inode,
5674                              struct ocfs2_extent_tree *et,
5675                              u32 cpos, u32 phys_cpos, u32 len, int flags,
5676                              struct ocfs2_cached_dealloc_ctxt *dealloc,
5677                              u64 refcount_loc, bool refcount_tree_locked)
5678 {
5679         int ret, credits = 0, extra_blocks = 0;
5680         u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5681         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5682         struct inode *tl_inode = osb->osb_tl_inode;
5683         handle_t *handle;
5684         struct ocfs2_alloc_context *meta_ac = NULL;
5685         struct ocfs2_refcount_tree *ref_tree = NULL;
5686
5687         if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5688                 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
5689                          OCFS2_HAS_REFCOUNT_FL));
5690
5691                 if (!refcount_tree_locked) {
5692                         ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5693                                                        &ref_tree, NULL);
5694                         if (ret) {
5695                                 mlog_errno(ret);
5696                                 goto bail;
5697                         }
5698                 }
5699
5700                 ret = ocfs2_prepare_refcount_change_for_del(inode,
5701                                                             refcount_loc,
5702                                                             phys_blkno,
5703                                                             len,
5704                                                             &credits,
5705                                                             &extra_blocks);
5706                 if (ret < 0) {
5707                         mlog_errno(ret);
5708                         goto bail;
5709                 }
5710         }
5711
5712         ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5713                                                  extra_blocks);
5714         if (ret) {
5715                 mlog_errno(ret);
5716                 goto bail;
5717         }
5718
5719         mutex_lock(&tl_inode->i_mutex);
5720
5721         if (ocfs2_truncate_log_needs_flush(osb)) {
5722                 ret = __ocfs2_flush_truncate_log(osb);
5723                 if (ret < 0) {
5724                         mlog_errno(ret);
5725                         goto out;
5726                 }
5727         }
5728
5729         handle = ocfs2_start_trans(osb,
5730                         ocfs2_remove_extent_credits(osb->sb) + credits);
5731         if (IS_ERR(handle)) {
5732                 ret = PTR_ERR(handle);
5733                 mlog_errno(ret);
5734                 goto out;
5735         }
5736
5737         ret = ocfs2_et_root_journal_access(handle, et,
5738                                            OCFS2_JOURNAL_ACCESS_WRITE);
5739         if (ret) {
5740                 mlog_errno(ret);
5741                 goto out_commit;
5742         }
5743
5744         dquot_free_space_nodirty(inode,
5745                                   ocfs2_clusters_to_bytes(inode->i_sb, len));
5746
5747         ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5748         if (ret) {
5749                 mlog_errno(ret);
5750                 goto out_commit;
5751         }
5752
5753         ocfs2_et_update_clusters(et, -len);
5754         ocfs2_update_inode_fsync_trans(handle, inode, 1);
5755
5756         ocfs2_journal_dirty(handle, et->et_root_bh);
5757
5758         if (phys_blkno) {
5759                 if (flags & OCFS2_EXT_REFCOUNTED)
5760                         ret = ocfs2_decrease_refcount(inode, handle,
5761                                         ocfs2_blocks_to_clusters(osb->sb,
5762                                                                  phys_blkno),
5763                                         len, meta_ac,
5764                                         dealloc, 1);
5765                 else
5766                         ret = ocfs2_truncate_log_append(osb, handle,
5767                                                         phys_blkno, len);
5768                 if (ret)
5769                         mlog_errno(ret);
5770
5771         }
5772
5773 out_commit:
5774         ocfs2_commit_trans(osb, handle);
5775 out:
5776         mutex_unlock(&tl_inode->i_mutex);
5777 bail:
5778         if (meta_ac)
5779                 ocfs2_free_alloc_context(meta_ac);
5780
5781         if (ref_tree)
5782                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5783
5784         return ret;
5785 }
5786
5787 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5788 {
5789         struct buffer_head *tl_bh = osb->osb_tl_bh;
5790         struct ocfs2_dinode *di;
5791         struct ocfs2_truncate_log *tl;
5792
5793         di = (struct ocfs2_dinode *) tl_bh->b_data;
5794         tl = &di->id2.i_dealloc;
5795
5796         mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5797                         "slot %d, invalid truncate log parameters: used = "
5798                         "%u, count = %u\n", osb->slot_num,
5799                         le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5800         return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5801 }
5802
5803 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5804                                            unsigned int new_start)
5805 {
5806         unsigned int tail_index;
5807         unsigned int current_tail;
5808
5809         /* No records, nothing to coalesce */
5810         if (!le16_to_cpu(tl->tl_used))
5811                 return 0;
5812
5813         tail_index = le16_to_cpu(tl->tl_used) - 1;
5814         current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5815         current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5816
5817         return current_tail == new_start;
5818 }
5819
5820 int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5821                               handle_t *handle,
5822                               u64 start_blk,
5823                               unsigned int num_clusters)
5824 {
5825         int status, index;
5826         unsigned int start_cluster, tl_count;
5827         struct inode *tl_inode = osb->osb_tl_inode;
5828         struct buffer_head *tl_bh = osb->osb_tl_bh;
5829         struct ocfs2_dinode *di;
5830         struct ocfs2_truncate_log *tl;
5831
5832         BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5833
5834         start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5835
5836         di = (struct ocfs2_dinode *) tl_bh->b_data;
5837
5838         /* tl_bh is loaded from ocfs2_truncate_log_init().  It's validated
5839          * by the underlying call to ocfs2_read_inode_block(), so any
5840          * corruption is a code bug */
5841         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5842
5843         tl = &di->id2.i_dealloc;
5844         tl_count = le16_to_cpu(tl->tl_count);
5845         mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5846                         tl_count == 0,
5847                         "Truncate record count on #%llu invalid "
5848                         "wanted %u, actual %u\n",
5849                         (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5850                         ocfs2_truncate_recs_per_inode(osb->sb),
5851                         le16_to_cpu(tl->tl_count));
5852
5853         /* Caller should have known to flush before calling us. */
5854         index = le16_to_cpu(tl->tl_used);
5855         if (index >= tl_count) {
5856                 status = -ENOSPC;
5857                 mlog_errno(status);
5858                 goto bail;
5859         }
5860
5861         status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5862                                          OCFS2_JOURNAL_ACCESS_WRITE);
5863         if (status < 0) {
5864                 mlog_errno(status);
5865                 goto bail;
5866         }
5867
5868         trace_ocfs2_truncate_log_append(
5869                 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5870                 start_cluster, num_clusters);
5871         if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5872                 /*
5873                  * Move index back to the record we are coalescing with.
5874                  * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5875                  */
5876                 index--;
5877
5878                 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5879                 trace_ocfs2_truncate_log_append(
5880                         (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5881                         index, le32_to_cpu(tl->tl_recs[index].t_start),
5882                         num_clusters);
5883         } else {
5884                 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5885                 tl->tl_used = cpu_to_le16(index + 1);
5886         }
5887         tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5888
5889         ocfs2_journal_dirty(handle, tl_bh);
5890
5891         osb->truncated_clusters += num_clusters;
5892 bail:
5893         return status;
5894 }
5895
5896 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5897                                          handle_t *handle,
5898                                          struct inode *data_alloc_inode,
5899                                          struct buffer_head *data_alloc_bh)
5900 {
5901         int status = 0;
5902         int i;
5903         unsigned int num_clusters;
5904         u64 start_blk;
5905         struct ocfs2_truncate_rec rec;
5906         struct ocfs2_dinode *di;
5907         struct ocfs2_truncate_log *tl;
5908         struct inode *tl_inode = osb->osb_tl_inode;
5909         struct buffer_head *tl_bh = osb->osb_tl_bh;
5910
5911         di = (struct ocfs2_dinode *) tl_bh->b_data;
5912         tl = &di->id2.i_dealloc;
5913         i = le16_to_cpu(tl->tl_used) - 1;
5914         while (i >= 0) {
5915                 /* Caller has given us at least enough credits to
5916                  * update the truncate log dinode */
5917                 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5918                                                  OCFS2_JOURNAL_ACCESS_WRITE);
5919                 if (status < 0) {
5920                         mlog_errno(status);
5921                         goto bail;
5922                 }
5923
5924                 tl->tl_used = cpu_to_le16(i);
5925
5926                 ocfs2_journal_dirty(handle, tl_bh);
5927
5928                 /* TODO: Perhaps we can calculate the bulk of the
5929                  * credits up front rather than extending like
5930                  * this. */
5931                 status = ocfs2_extend_trans(handle,
5932                                             OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5933                 if (status < 0) {
5934                         mlog_errno(status);
5935                         goto bail;
5936                 }
5937
5938                 rec = tl->tl_recs[i];
5939                 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5940                                                     le32_to_cpu(rec.t_start));
5941                 num_clusters = le32_to_cpu(rec.t_clusters);
5942
5943                 /* if start_blk is not set, we ignore the record as
5944                  * invalid. */
5945                 if (start_blk) {
5946                         trace_ocfs2_replay_truncate_records(
5947                                 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5948                                 i, le32_to_cpu(rec.t_start), num_clusters);
5949
5950                         status = ocfs2_free_clusters(handle, data_alloc_inode,
5951                                                      data_alloc_bh, start_blk,
5952                                                      num_clusters);
5953                         if (status < 0) {
5954                                 mlog_errno(status);
5955                                 goto bail;
5956                         }
5957                 }
5958                 i--;
5959         }
5960
5961         osb->truncated_clusters = 0;
5962
5963 bail:
5964         return status;
5965 }
5966
5967 /* Expects you to already be holding tl_inode->i_mutex */
5968 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5969 {
5970         int status;
5971         unsigned int num_to_flush;
5972         handle_t *handle;
5973         struct inode *tl_inode = osb->osb_tl_inode;
5974         struct inode *data_alloc_inode = NULL;
5975         struct buffer_head *tl_bh = osb->osb_tl_bh;
5976         struct buffer_head *data_alloc_bh = NULL;
5977         struct ocfs2_dinode *di;
5978         struct ocfs2_truncate_log *tl;
5979
5980         BUG_ON(mutex_trylock(&tl_inode->i_mutex));
5981
5982         di = (struct ocfs2_dinode *) tl_bh->b_data;
5983
5984         /* tl_bh is loaded from ocfs2_truncate_log_init().  It's validated
5985          * by the underlying call to ocfs2_read_inode_block(), so any
5986          * corruption is a code bug */
5987         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5988
5989         tl = &di->id2.i_dealloc;
5990         num_to_flush = le16_to_cpu(tl->tl_used);
5991         trace_ocfs2_flush_truncate_log(
5992                 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5993                 num_to_flush);
5994         if (!num_to_flush) {
5995                 status = 0;
5996                 goto out;
5997         }
5998
5999         data_alloc_inode = ocfs2_get_system_file_inode(osb,
6000                                                        GLOBAL_BITMAP_SYSTEM_INODE,
6001                                                        OCFS2_INVALID_SLOT);
6002         if (!data_alloc_inode) {
6003                 status = -EINVAL;
6004                 mlog(ML_ERROR, "Could not get bitmap inode!\n");
6005                 goto out;
6006         }
6007
6008         mutex_lock(&data_alloc_inode->i_mutex);
6009
6010         status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6011         if (status < 0) {
6012                 mlog_errno(status);
6013                 goto out_mutex;
6014         }
6015
6016         handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6017         if (IS_ERR(handle)) {
6018                 status = PTR_ERR(handle);
6019                 mlog_errno(status);
6020                 goto out_unlock;
6021         }
6022
6023         status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode,
6024                                                data_alloc_bh);
6025         if (status < 0)
6026                 mlog_errno(status);
6027
6028         ocfs2_commit_trans(osb, handle);
6029
6030 out_unlock:
6031         brelse(data_alloc_bh);
6032         ocfs2_inode_unlock(data_alloc_inode, 1);
6033
6034 out_mutex:
6035         mutex_unlock(&data_alloc_inode->i_mutex);
6036         iput(data_alloc_inode);
6037
6038 out:
6039         return status;
6040 }
6041
6042 int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6043 {
6044         int status;
6045         struct inode *tl_inode = osb->osb_tl_inode;
6046
6047         mutex_lock(&tl_inode->i_mutex);
6048         status = __ocfs2_flush_truncate_log(osb);
6049         mutex_unlock(&tl_inode->i_mutex);
6050
6051         return status;
6052 }
6053
6054 static void ocfs2_truncate_log_worker(struct work_struct *work)
6055 {
6056         int status;
6057         struct ocfs2_super *osb =
6058                 container_of(work, struct ocfs2_super,
6059                              osb_truncate_log_wq.work);
6060
6061         status = ocfs2_flush_truncate_log(osb);
6062         if (status < 0)
6063                 mlog_errno(status);
6064         else
6065                 ocfs2_init_steal_slots(osb);
6066 }
6067
6068 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6069 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6070                                        int cancel)
6071 {
6072         if (osb->osb_tl_inode &&
6073                         atomic_read(&osb->osb_tl_disable) == 0) {
6074                 /* We want to push off log flushes while truncates are
6075                  * still running. */
6076                 if (cancel)
6077                         cancel_delayed_work(&osb->osb_truncate_log_wq);
6078
6079                 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq,
6080                                    OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6081         }
6082 }
6083
6084 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6085                                        int slot_num,
6086                                        struct inode **tl_inode,
6087                                        struct buffer_head **tl_bh)
6088 {
6089         int status;
6090         struct inode *inode = NULL;
6091         struct buffer_head *bh = NULL;
6092
6093         inode = ocfs2_get_system_file_inode(osb,
6094                                            TRUNCATE_LOG_SYSTEM_INODE,
6095                                            slot_num);
6096         if (!inode) {
6097                 status = -EINVAL;
6098                 mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6099                 goto bail;
6100         }
6101
6102         status = ocfs2_read_inode_block(inode, &bh);
6103         if (status < 0) {
6104                 iput(inode);
6105                 mlog_errno(status);
6106                 goto bail;
6107         }
6108
6109         *tl_inode = inode;
6110         *tl_bh    = bh;
6111 bail:
6112         return status;
6113 }
6114
6115 /* called during the 1st stage of node recovery. we stamp a clean
6116  * truncate log and pass back a copy for processing later. if the
6117  * truncate log does not require processing, a *tl_copy is set to
6118  * NULL. */
6119 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6120                                       int slot_num,
6121                                       struct ocfs2_dinode **tl_copy)
6122 {
6123         int status;
6124         struct inode *tl_inode = NULL;
6125         struct buffer_head *tl_bh = NULL;
6126         struct ocfs2_dinode *di;
6127         struct ocfs2_truncate_log *tl;
6128
6129         *tl_copy = NULL;
6130
6131         trace_ocfs2_begin_truncate_log_recovery(slot_num);
6132
6133         status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6134         if (status < 0) {
6135                 mlog_errno(status);
6136                 goto bail;
6137         }
6138
6139         di = (struct ocfs2_dinode *) tl_bh->b_data;
6140
6141         /* tl_bh is loaded from ocfs2_get_truncate_log_info().  It's
6142          * validated by the underlying call to ocfs2_read_inode_block(),
6143          * so any corruption is a code bug */
6144         BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6145
6146         tl = &di->id2.i_dealloc;
6147         if (le16_to_cpu(tl->tl_used)) {
6148                 trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6149
6150                 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL);
6151                 if (!(*tl_copy)) {
6152                         status = -ENOMEM;
6153                         mlog_errno(status);
6154                         goto bail;
6155                 }
6156
6157                 /* Assuming the write-out below goes well, this copy
6158                  * will be passed back to recovery for processing. */
6159                 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size);
6160
6161                 /* All we need to do to clear the truncate log is set
6162                  * tl_used. */
6163                 tl->tl_used = 0;
6164
6165                 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6166                 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6167                 if (status < 0) {
6168                         mlog_errno(status);
6169                         goto bail;
6170                 }
6171         }
6172
6173 bail:
6174         if (tl_inode)
6175                 iput(tl_inode);
6176         brelse(tl_bh);
6177
6178         if (status < 0 && (*tl_copy)) {
6179                 kfree(*tl_copy);
6180                 *tl_copy = NULL;
6181                 mlog_errno(status);
6182         }
6183
6184         return status;
6185 }
6186
6187 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6188                                          struct ocfs2_dinode *tl_copy)
6189 {
6190         int status = 0;
6191         int i;
6192         unsigned int clusters, num_recs, start_cluster;
6193         u64 start_blk;
6194         handle_t *handle;
6195         struct inode *tl_inode = osb->osb_tl_inode;
6196         struct ocfs2_truncate_log *tl;
6197
6198         if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6199                 mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6200                 return -EINVAL;
6201         }
6202
6203         tl = &tl_copy->id2.i_dealloc;
6204         num_recs = le16_to_cpu(tl->tl_used);
6205         trace_ocfs2_complete_truncate_log_recovery(
6206                 (unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6207                 num_recs);
6208
6209         mutex_lock(&tl_inode->i_mutex);
6210         for(i = 0; i < num_recs; i++) {
6211                 if (ocfs2_truncate_log_needs_flush(osb)) {
6212                         status = __ocfs2_flush_truncate_log(osb);
6213                         if (status < 0) {
6214                                 mlog_errno(status);
6215                                 goto bail_up;
6216                         }
6217                 }
6218
6219                 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6220                 if (IS_ERR(handle)) {
6221                         status = PTR_ERR(handle);
6222                         mlog_errno(status);
6223                         goto bail_up;
6224                 }
6225
6226                 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6227                 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6228                 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6229
6230                 status = ocfs2_truncate_log_append(osb, handle,
6231                                                    start_blk, clusters);
6232                 ocfs2_commit_trans(osb, handle);
6233                 if (status < 0) {
6234                         mlog_errno(status);
6235                         goto bail_up;
6236                 }
6237         }
6238
6239 bail_up:
6240         mutex_unlock(&tl_inode->i_mutex);
6241
6242         return status;
6243 }
6244
6245 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6246 {
6247         int status;
6248         struct inode *tl_inode = osb->osb_tl_inode;
6249
6250         atomic_set(&osb->osb_tl_disable, 1);
6251
6252         if (tl_inode) {
6253                 cancel_delayed_work(&osb->osb_truncate_log_wq);
6254                 flush_workqueue(ocfs2_wq);
6255
6256                 status = ocfs2_flush_truncate_log(osb);
6257                 if (status < 0)
6258                         mlog_errno(status);
6259
6260                 brelse(osb->osb_tl_bh);
6261                 iput(osb->osb_tl_inode);
6262         }
6263 }
6264
6265 int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6266 {
6267         int status;
6268         struct inode *tl_inode = NULL;
6269         struct buffer_head *tl_bh = NULL;
6270
6271         status = ocfs2_get_truncate_log_info(osb,
6272                                              osb->slot_num,
6273                                              &tl_inode,
6274                                              &tl_bh);
6275         if (status < 0)
6276                 mlog_errno(status);
6277
6278         /* ocfs2_truncate_log_shutdown keys on the existence of
6279          * osb->osb_tl_inode so we don't set any of the osb variables
6280          * until we're sure all is well. */
6281         INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6282                           ocfs2_truncate_log_worker);
6283         atomic_set(&osb->osb_tl_disable, 0);
6284         osb->osb_tl_bh    = tl_bh;
6285         osb->osb_tl_inode = tl_inode;
6286
6287         return status;
6288 }
6289
6290 /*
6291  * Delayed de-allocation of suballocator blocks.
6292  *
6293  * Some sets of block de-allocations might involve multiple suballocator inodes.
6294  *
6295  * The locking for this can get extremely complicated, especially when
6296  * the suballocator inodes to delete from aren't known until deep
6297  * within an unrelated codepath.
6298  *
6299  * ocfs2_extent_block structures are a good example of this - an inode
6300  * btree could have been grown by any number of nodes each allocating
6301  * out of their own suballoc inode.
6302  *
6303  * These structures allow the delay of block de-allocation until a
6304  * later time, when locking of multiple cluster inodes won't cause
6305  * deadlock.
6306  */
6307
6308 /*
6309  * Describe a single bit freed from a suballocator.  For the block
6310  * suballocators, it represents one block.  For the global cluster
6311  * allocator, it represents some clusters and free_bit indicates
6312  * clusters number.
6313  */
6314 struct ocfs2_cached_block_free {
6315         struct ocfs2_cached_block_free          *free_next;
6316         u64                                     free_bg;
6317         u64                                     free_blk;
6318         unsigned int                            free_bit;
6319 };
6320
6321 struct ocfs2_per_slot_free_list {
6322         struct ocfs2_per_slot_free_list         *f_next_suballocator;
6323         int                                     f_inode_type;
6324         int                                     f_slot;
6325         struct ocfs2_cached_block_free          *f_first;
6326 };
6327
6328 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6329                                     int sysfile_type,
6330                                     int slot,
6331                                     struct ocfs2_cached_block_free *head)
6332 {
6333         int ret;
6334         u64 bg_blkno;
6335         handle_t *handle;
6336         struct inode *inode;
6337         struct buffer_head *di_bh = NULL;
6338         struct ocfs2_cached_block_free *tmp;
6339
6340         inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6341         if (!inode) {
6342                 ret = -EINVAL;
6343                 mlog_errno(ret);
6344                 goto out;
6345         }
6346
6347         mutex_lock(&inode->i_mutex);
6348
6349         ret = ocfs2_inode_lock(inode, &di_bh, 1);
6350         if (ret) {
6351                 mlog_errno(ret);
6352                 goto out_mutex;
6353         }
6354
6355         handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6356         if (IS_ERR(handle)) {
6357                 ret = PTR_ERR(handle);
6358                 mlog_errno(ret);
6359                 goto out_unlock;
6360         }
6361
6362         while (head) {
6363                 if (head->free_bg)
6364                         bg_blkno = head->free_bg;
6365                 else
6366                         bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6367                                                               head->free_bit);
6368                 trace_ocfs2_free_cached_blocks(
6369                      (unsigned long long)head->free_blk, head->free_bit);
6370
6371                 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6372                                                head->free_bit, bg_blkno, 1);
6373                 if (ret) {
6374                         mlog_errno(ret);
6375                         goto out_journal;
6376                 }
6377
6378                 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE);
6379                 if (ret) {
6380                         mlog_errno(ret);
6381                         goto out_journal;
6382                 }
6383
6384                 tmp = head;
6385                 head = head->free_next;
6386                 kfree(tmp);
6387         }
6388
6389 out_journal:
6390         ocfs2_commit_trans(osb, handle);
6391
6392 out_unlock:
6393         ocfs2_inode_unlock(inode, 1);
6394         brelse(di_bh);
6395 out_mutex:
6396         mutex_unlock(&inode->i_mutex);
6397         iput(inode);
6398 out:
6399         while(head) {
6400                 /* Premature exit may have left some dangling items. */
6401                 tmp = head;
6402                 head = head->free_next;
6403                 kfree(tmp);
6404         }
6405
6406         return ret;
6407 }
6408
6409 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6410                                 u64 blkno, unsigned int bit)
6411 {
6412         int ret = 0;
6413         struct ocfs2_cached_block_free *item;
6414
6415         item = kzalloc(sizeof(*item), GFP_NOFS);
6416         if (item == NULL) {
6417                 ret = -ENOMEM;
6418                 mlog_errno(ret);
6419                 return ret;
6420         }
6421
6422         trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6423
6424         item->free_blk = blkno;
6425         item->free_bit = bit;
6426         item->free_next = ctxt->c_global_allocator;
6427
6428         ctxt->c_global_allocator = item;
6429         return ret;
6430 }
6431
6432 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6433                                       struct ocfs2_cached_block_free *head)
6434 {
6435         struct ocfs2_cached_block_free *tmp;
6436         struct inode *tl_inode = osb->osb_tl_inode;
6437         handle_t *handle;
6438         int ret = 0;
6439
6440         mutex_lock(&tl_inode->i_mutex);
6441
6442         while (head) {
6443                 if (ocfs2_truncate_log_needs_flush(osb)) {
6444                         ret = __ocfs2_flush_truncate_log(osb);
6445                         if (ret < 0) {
6446                                 mlog_errno(ret);
6447                                 break;
6448                         }
6449                 }
6450
6451                 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6452                 if (IS_ERR(handle)) {
6453                         ret = PTR_ERR(handle);
6454                         mlog_errno(ret);
6455                         break;
6456                 }
6457
6458                 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6459                                                 head->free_bit);
6460
6461                 ocfs2_commit_trans(osb, handle);
6462                 tmp = head;
6463                 head = head->free_next;
6464                 kfree(tmp);
6465
6466                 if (ret < 0) {
6467                         mlog_errno(ret);
6468                         break;
6469                 }
6470         }
6471
6472         mutex_unlock(&tl_inode->i_mutex);
6473
6474         while (head) {
6475                 /* Premature exit may have left some dangling items. */
6476                 tmp = head;
6477                 head = head->free_next;
6478                 kfree(tmp);
6479         }
6480
6481         return ret;
6482 }
6483
6484 int ocfs2_run_deallocs(struct ocfs2_super *osb,
6485                        struct ocfs2_cached_dealloc_ctxt *ctxt)
6486 {
6487         int ret = 0, ret2;
6488         struct ocfs2_per_slot_free_list *fl;
6489
6490         if (!ctxt)
6491                 return 0;
6492
6493         while (ctxt->c_first_suballocator) {
6494                 fl = ctxt->c_first_suballocator;
6495
6496                 if (fl->f_first) {
6497                         trace_ocfs2_run_deallocs(fl->f_inode_type,
6498                                                  fl->f_slot);
6499                         ret2 = ocfs2_free_cached_blocks(osb,
6500                                                         fl->f_inode_type,
6501                                                         fl->f_slot,
6502                                                         fl->f_first);
6503                         if (ret2)
6504                                 mlog_errno(ret2);
6505                         if (!ret)
6506                                 ret = ret2;
6507                 }
6508
6509                 ctxt->c_first_suballocator = fl->f_next_suballocator;
6510                 kfree(fl);
6511         }
6512
6513         if (ctxt->c_global_allocator) {
6514                 ret2 = ocfs2_free_cached_clusters(osb,
6515                                                   ctxt->c_global_allocator);
6516                 if (ret2)
6517                         mlog_errno(ret2);
6518                 if (!ret)
6519                         ret = ret2;
6520
6521                 ctxt->c_global_allocator = NULL;
6522         }
6523
6524         return ret;
6525 }
6526
6527 static struct ocfs2_per_slot_free_list *
6528 ocfs2_find_per_slot_free_list(int type,
6529                               int slot,
6530                               struct ocfs2_cached_dealloc_ctxt *ctxt)
6531 {
6532         struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6533
6534         while (fl) {
6535                 if (fl->f_inode_type == type && fl->f_slot == slot)
6536                         return fl;
6537
6538                 fl = fl->f_next_suballocator;
6539         }
6540
6541         fl = kmalloc(sizeof(*fl), GFP_NOFS);
6542         if (fl) {
6543                 fl->f_inode_type = type;
6544                 fl->f_slot = slot;
6545                 fl->f_first = NULL;
6546                 fl->f_next_suballocator = ctxt->c_first_suballocator;
6547
6548                 ctxt->c_first_suballocator = fl;
6549         }
6550         return fl;
6551 }
6552
6553 int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6554                               int type, int slot, u64 suballoc,
6555                               u64 blkno, unsigned int bit)
6556 {
6557         int ret;
6558         struct ocfs2_per_slot_free_list *fl;
6559         struct ocfs2_cached_block_free *item;
6560
6561         fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6562         if (fl == NULL) {
6563                 ret = -ENOMEM;
6564                 mlog_errno(ret);
6565                 goto out;
6566         }
6567
6568         item = kzalloc(sizeof(*item), GFP_NOFS);
6569         if (item == NULL) {
6570                 ret = -ENOMEM;
6571                 mlog_errno(ret);
6572                 goto out;
6573         }
6574
6575         trace_ocfs2_cache_block_dealloc(type, slot,
6576                                         (unsigned long long)suballoc,
6577                                         (unsigned long long)blkno, bit);
6578
6579         item->free_bg = suballoc;
6580         item->free_blk = blkno;
6581         item->free_bit = bit;
6582         item->free_next = fl->f_first;
6583
6584         fl->f_first = item;
6585
6586         ret = 0;
6587 out:
6588         return ret;
6589 }
6590
6591 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6592                                          struct ocfs2_extent_block *eb)
6593 {
6594         return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6595                                          le16_to_cpu(eb->h_suballoc_slot),
6596                                          le64_to_cpu(eb->h_suballoc_loc),
6597                                          le64_to_cpu(eb->h_blkno),
6598                                          le16_to_cpu(eb->h_suballoc_bit));
6599 }
6600
6601 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6602 {
6603         set_buffer_uptodate(bh);
6604         mark_buffer_dirty(bh);
6605         return 0;
6606 }
6607
6608 void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6609                               unsigned int from, unsigned int to,
6610                               struct page *page, int zero, u64 *phys)
6611 {
6612         int ret, partial = 0;
6613
6614         ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6615         if (ret)
6616                 mlog_errno(ret);
6617
6618         if (zero)
6619                 zero_user_segment(page, from, to);
6620
6621         /*
6622          * Need to set the buffers we zero'd into uptodate
6623          * here if they aren't - ocfs2_map_page_blocks()
6624          * might've skipped some
6625          */
6626         ret = walk_page_buffers(handle, page_buffers(page),
6627                                 from, to, &partial,
6628                                 ocfs2_zero_func);
6629         if (ret < 0)
6630                 mlog_errno(ret);
6631         else if (ocfs2_should_order_data(inode)) {
6632                 ret = ocfs2_jbd2_file_inode(handle, inode);
6633                 if (ret < 0)
6634                         mlog_errno(ret);
6635         }
6636
6637         if (!partial)
6638                 SetPageUptodate(page);
6639
6640         flush_dcache_page(page);
6641 }
6642
6643 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6644                                      loff_t end, struct page **pages,
6645                                      int numpages, u64 phys, handle_t *handle)
6646 {
6647         int i;
6648         struct page *page;
6649         unsigned int from, to = PAGE_CACHE_SIZE;
6650         struct super_block *sb = inode->i_sb;
6651
6652         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6653
6654         if (numpages == 0)
6655                 goto out;
6656
6657         to = PAGE_CACHE_SIZE;
6658         for(i = 0; i < numpages; i++) {
6659                 page = pages[i];
6660
6661                 from = start & (PAGE_CACHE_SIZE - 1);
6662                 if ((end >> PAGE_CACHE_SHIFT) == page->index)
6663                         to = end & (PAGE_CACHE_SIZE - 1);
6664
6665                 BUG_ON(from > PAGE_CACHE_SIZE);
6666                 BUG_ON(to > PAGE_CACHE_SIZE);
6667
6668                 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6669                                          &phys);
6670
6671                 start = (page->index + 1) << PAGE_CACHE_SHIFT;
6672         }
6673 out:
6674         if (pages)
6675                 ocfs2_unlock_and_free_pages(pages, numpages);
6676 }
6677
6678 int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6679                      struct page **pages, int *num)
6680 {
6681         int numpages, ret = 0;
6682         struct address_space *mapping = inode->i_mapping;
6683         unsigned long index;
6684         loff_t last_page_bytes;
6685
6686         BUG_ON(start > end);
6687
6688         numpages = 0;
6689         last_page_bytes = PAGE_ALIGN(end);
6690         index = start >> PAGE_CACHE_SHIFT;
6691         do {
6692                 pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6693                 if (!pages[numpages]) {
6694                         ret = -ENOMEM;
6695                         mlog_errno(ret);
6696                         goto out;
6697                 }
6698
6699                 numpages++;
6700                 index++;
6701         } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT));
6702
6703 out:
6704         if (ret != 0) {
6705                 if (pages)
6706                         ocfs2_unlock_and_free_pages(pages, numpages);
6707                 numpages = 0;
6708         }
6709
6710         *num = numpages;
6711
6712         return ret;
6713 }
6714
6715 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6716                                 struct page **pages, int *num)
6717 {
6718         struct super_block *sb = inode->i_sb;
6719
6720         BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6721                (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6722
6723         return ocfs2_grab_pages(inode, start, end, pages, num);
6724 }
6725
6726 /*
6727  * Zero the area past i_size but still within an allocated
6728  * cluster. This avoids exposing nonzero data on subsequent file
6729  * extends.
6730  *
6731  * We need to call this before i_size is updated on the inode because
6732  * otherwise block_write_full_page() will skip writeout of pages past
6733  * i_size. The new_i_size parameter is passed for this reason.
6734  */
6735 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6736                                   u64 range_start, u64 range_end)
6737 {
6738         int ret = 0, numpages;
6739         struct page **pages = NULL;
6740         u64 phys;
6741         unsigned int ext_flags;
6742         struct super_block *sb = inode->i_sb;
6743
6744         /*
6745          * File systems which don't support sparse files zero on every
6746          * extend.
6747          */
6748         if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6749                 return 0;
6750
6751         pages = kcalloc(ocfs2_pages_per_cluster(sb),
6752                         sizeof(struct page *), GFP_NOFS);
6753         if (pages == NULL) {
6754                 ret = -ENOMEM;
6755                 mlog_errno(ret);
6756                 goto out;
6757         }
6758
6759         if (range_start == range_end)
6760                 goto out;
6761
6762         ret = ocfs2_extent_map_get_blocks(inode,
6763                                           range_start >> sb->s_blocksize_bits,
6764                                           &phys, NULL, &ext_flags);
6765         if (ret) {
6766                 mlog_errno(ret);
6767                 goto out;
6768         }
6769
6770         /*
6771          * Tail is a hole, or is marked unwritten. In either case, we
6772          * can count on read and write to return/push zero's.
6773          */
6774         if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6775                 goto out;
6776
6777         ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6778                                    &numpages);
6779         if (ret) {
6780                 mlog_errno(ret);
6781                 goto out;
6782         }
6783
6784         ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6785                                  numpages, phys, handle);
6786
6787         /*
6788          * Initiate writeout of the pages we zero'd here. We don't
6789          * wait on them - the truncate_inode_pages() call later will
6790          * do that for us.
6791          */
6792         ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6793                                        range_end - 1);
6794         if (ret)
6795                 mlog_errno(ret);
6796
6797 out:
6798         kfree(pages);
6799
6800         return ret;
6801 }
6802
6803 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
6804                                              struct ocfs2_dinode *di)
6805 {
6806         unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
6807         unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
6808
6809         if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
6810                 memset(&di->id2, 0, blocksize -
6811                                     offsetof(struct ocfs2_dinode, id2) -
6812                                     xattrsize);
6813         else
6814                 memset(&di->id2, 0, blocksize -
6815                                     offsetof(struct ocfs2_dinode, id2));
6816 }
6817
6818 void ocfs2_dinode_new_extent_list(struct inode *inode,
6819                                   struct ocfs2_dinode *di)
6820 {
6821         ocfs2_zero_dinode_id2_with_xattr(inode, di);
6822         di->id2.i_list.l_tree_depth = 0;
6823         di->id2.i_list.l_next_free_rec = 0;
6824         di->id2.i_list.l_count = cpu_to_le16(
6825                 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
6826 }
6827
6828 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
6829 {
6830         struct ocfs2_inode_info *oi = OCFS2_I(inode);
6831         struct ocfs2_inline_data *idata = &di->id2.i_data;
6832
6833         spin_lock(&oi->ip_lock);
6834         oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
6835         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6836         spin_unlock(&oi->ip_lock);
6837
6838         /*
6839          * We clear the entire i_data structure here so that all
6840          * fields can be properly initialized.
6841          */
6842         ocfs2_zero_dinode_id2_with_xattr(inode, di);
6843
6844         idata->id_count = cpu_to_le16(
6845                         ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
6846 }
6847
6848 int ocfs2_convert_inline_data_to_extents(struct inode *inode,
6849                                          struct buffer_head *di_bh)
6850 {
6851         int ret, i, has_data, num_pages = 0;
6852         int need_free = 0;
6853         u32 bit_off, num;
6854         handle_t *handle;
6855         u64 uninitialized_var(block);
6856         struct ocfs2_inode_info *oi = OCFS2_I(inode);
6857         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
6858         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
6859         struct ocfs2_alloc_context *data_ac = NULL;
6860         struct page **pages = NULL;
6861         loff_t end = osb->s_clustersize;
6862         struct ocfs2_extent_tree et;
6863         int did_quota = 0;
6864
6865         has_data = i_size_read(inode) ? 1 : 0;
6866
6867         if (has_data) {
6868                 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
6869                                 sizeof(struct page *), GFP_NOFS);
6870                 if (pages == NULL) {
6871                         ret = -ENOMEM;
6872                         mlog_errno(ret);
6873                         return ret;
6874                 }
6875
6876                 ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
6877                 if (ret) {
6878                         mlog_errno(ret);
6879                         goto free_pages;
6880                 }
6881         }
6882
6883         handle = ocfs2_start_trans(osb,
6884                                    ocfs2_inline_to_extents_credits(osb->sb));
6885         if (IS_ERR(handle)) {
6886                 ret = PTR_ERR(handle);
6887                 mlog_errno(ret);
6888                 goto out;
6889         }
6890
6891         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
6892                                       OCFS2_JOURNAL_ACCESS_WRITE);
6893         if (ret) {
6894                 mlog_errno(ret);
6895                 goto out_commit;
6896         }
6897
6898         if (has_data) {
6899                 unsigned int page_end;
6900                 u64 phys;
6901
6902                 ret = dquot_alloc_space_nodirty(inode,
6903                                        ocfs2_clusters_to_bytes(osb->sb, 1));
6904                 if (ret)
6905                         goto out_commit;
6906                 did_quota = 1;
6907
6908                 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
6909
6910                 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
6911                                            &num);
6912                 if (ret) {
6913                         mlog_errno(ret);
6914                         goto out_commit;
6915                 }
6916
6917                 /*
6918                  * Save two copies, one for insert, and one that can
6919                  * be changed by ocfs2_map_and_dirty_page() below.
6920                  */
6921                 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
6922
6923                 /*
6924                  * Non sparse file systems zero on extend, so no need
6925                  * to do that now.
6926                  */
6927                 if (!ocfs2_sparse_alloc(osb) &&
6928                     PAGE_CACHE_SIZE < osb->s_clustersize)
6929                         end = PAGE_CACHE_SIZE;
6930
6931                 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
6932                 if (ret) {
6933                         mlog_errno(ret);
6934                         need_free = 1;
6935                         goto out_commit;
6936                 }
6937
6938                 /*
6939                  * This should populate the 1st page for us and mark
6940                  * it up to date.
6941                  */
6942                 ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
6943                 if (ret) {
6944                         mlog_errno(ret);
6945                         need_free = 1;
6946                         goto out_unlock;
6947                 }
6948
6949                 page_end = PAGE_CACHE_SIZE;
6950                 if (PAGE_CACHE_SIZE > osb->s_clustersize)
6951                         page_end = osb->s_clustersize;
6952
6953                 for (i = 0; i < num_pages; i++)
6954                         ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
6955                                                  pages[i], i > 0, &phys);
6956         }
6957
6958         spin_lock(&oi->ip_lock);
6959         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
6960         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
6961         spin_unlock(&oi->ip_lock);
6962
6963         ocfs2_update_inode_fsync_trans(handle, inode, 1);
6964         ocfs2_dinode_new_extent_list(inode, di);
6965
6966         ocfs2_journal_dirty(handle, di_bh);
6967
6968         if (has_data) {
6969                 /*
6970                  * An error at this point should be extremely rare. If
6971                  * this proves to be false, we could always re-build
6972                  * the in-inode data from our pages.
6973                  */
6974                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
6975                 ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
6976                 if (ret) {
6977                         mlog_errno(ret);
6978                         need_free = 1;
6979                         goto out_unlock;
6980                 }
6981
6982                 inode->i_blocks = ocfs2_inode_sector_count(inode);
6983         }
6984
6985 out_unlock:
6986         if (pages)
6987                 ocfs2_unlock_and_free_pages(pages, num_pages);
6988
6989 out_commit:
6990         if (ret < 0 && did_quota)
6991                 dquot_free_space_nodirty(inode,
6992                                           ocfs2_clusters_to_bytes(osb->sb, 1));
6993
6994         if (need_free) {
6995                 if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
6996                         ocfs2_free_local_alloc_bits(osb, handle, data_ac,
6997                                         bit_off, num);
6998                 else
6999                         ocfs2_free_clusters(handle,
7000                                         data_ac->ac_inode,
7001                                         data_ac->ac_bh,
7002                                         ocfs2_clusters_to_blocks(osb->sb, bit_off),
7003                                         num);
7004         }
7005
7006         ocfs2_commit_trans(osb, handle);
7007
7008 out:
7009         if (data_ac)
7010                 ocfs2_free_alloc_context(data_ac);
7011 free_pages:
7012         kfree(pages);
7013         return ret;
7014 }
7015
7016 /*
7017  * It is expected, that by the time you call this function,
7018  * inode->i_size and fe->i_size have been adjusted.
7019  *
7020  * WARNING: This will kfree the truncate context
7021  */
7022 int ocfs2_commit_truncate(struct ocfs2_super *osb,
7023                           struct inode *inode,
7024                           struct buffer_head *di_bh)
7025 {
7026         int status = 0, i, flags = 0;
7027         u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7028         u64 blkno = 0;
7029         struct ocfs2_extent_list *el;
7030         struct ocfs2_extent_rec *rec;
7031         struct ocfs2_path *path = NULL;
7032         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7033         struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7034         u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7035         struct ocfs2_extent_tree et;
7036         struct ocfs2_cached_dealloc_ctxt dealloc;
7037         struct ocfs2_refcount_tree *ref_tree = NULL;
7038
7039         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7040         ocfs2_init_dealloc_ctxt(&dealloc);
7041
7042         new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7043                                                      i_size_read(inode));
7044
7045         path = ocfs2_new_path(di_bh, &di->id2.i_list,
7046                               ocfs2_journal_access_di);
7047         if (!path) {
7048                 status = -ENOMEM;
7049                 mlog_errno(status);
7050                 goto bail;
7051         }
7052
7053         ocfs2_extent_map_trunc(inode, new_highest_cpos);
7054
7055 start:
7056         /*
7057          * Check that we still have allocation to delete.
7058          */
7059         if (OCFS2_I(inode)->ip_clusters == 0) {
7060                 status = 0;
7061                 goto bail;
7062         }
7063
7064         /*
7065          * Truncate always works against the rightmost tree branch.
7066          */
7067         status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7068         if (status) {
7069                 mlog_errno(status);
7070                 goto bail;
7071         }
7072
7073         trace_ocfs2_commit_truncate(
7074                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
7075                 new_highest_cpos,
7076                 OCFS2_I(inode)->ip_clusters,
7077                 path->p_tree_depth);
7078
7079         /*
7080          * By now, el will point to the extent list on the bottom most
7081          * portion of this tree. Only the tail record is considered in
7082          * each pass.
7083          *
7084          * We handle the following cases, in order:
7085          * - empty extent: delete the remaining branch
7086          * - remove the entire record
7087          * - remove a partial record
7088          * - no record needs to be removed (truncate has completed)
7089          */
7090         el = path_leaf_el(path);
7091         if (le16_to_cpu(el->l_next_free_rec) == 0) {
7092                 ocfs2_error(inode->i_sb,
7093                             "Inode %llu has empty extent block at %llu\n",
7094                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
7095                             (unsigned long long)path_leaf_bh(path)->b_blocknr);
7096                 status = -EROFS;
7097                 goto bail;
7098         }
7099
7100         i = le16_to_cpu(el->l_next_free_rec) - 1;
7101         rec = &el->l_recs[i];
7102         flags = rec->e_flags;
7103         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7104
7105         if (i == 0 && ocfs2_is_empty_extent(rec)) {
7106                 /*
7107                  * Lower levels depend on this never happening, but it's best
7108                  * to check it up here before changing the tree.
7109                 */
7110                 if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7111                         ocfs2_error(inode->i_sb, "Inode %lu has an empty "
7112                                     "extent record, depth %u\n", inode->i_ino,
7113                                     le16_to_cpu(root_el->l_tree_depth));
7114                         status = -EROFS;
7115                         goto bail;
7116                 }
7117                 trunc_cpos = le32_to_cpu(rec->e_cpos);
7118                 trunc_len = 0;
7119                 blkno = 0;
7120         } else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7121                 /*
7122                  * Truncate entire record.
7123                  */
7124                 trunc_cpos = le32_to_cpu(rec->e_cpos);
7125                 trunc_len = ocfs2_rec_clusters(el, rec);
7126                 blkno = le64_to_cpu(rec->e_blkno);
7127         } else if (range > new_highest_cpos) {
7128                 /*
7129                  * Partial truncate. it also should be
7130                  * the last truncate we're doing.
7131                  */
7132                 trunc_cpos = new_highest_cpos;
7133                 trunc_len = range - new_highest_cpos;
7134                 coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7135                 blkno = le64_to_cpu(rec->e_blkno) +
7136                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
7137         } else {
7138                 /*
7139                  * Truncate completed, leave happily.
7140                  */
7141                 status = 0;
7142                 goto bail;
7143         }
7144
7145         phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7146
7147         if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7148                 status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7149                                 &ref_tree, NULL);
7150                 if (status) {
7151                         mlog_errno(status);
7152                         goto bail;
7153                 }
7154         }
7155
7156         status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7157                                           phys_cpos, trunc_len, flags, &dealloc,
7158                                           refcount_loc, true);
7159         if (status < 0) {
7160                 mlog_errno(status);
7161                 goto bail;
7162         }
7163
7164         ocfs2_reinit_path(path, 1);
7165
7166         /*
7167          * The check above will catch the case where we've truncated
7168          * away all allocation.
7169          */
7170         goto start;
7171
7172 bail:
7173         if (ref_tree)
7174                 ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7175
7176         ocfs2_schedule_truncate_log_flush(osb, 1);
7177
7178         ocfs2_run_deallocs(osb, &dealloc);
7179
7180         ocfs2_free_path(path);
7181
7182         return status;
7183 }
7184
7185 /*
7186  * 'start' is inclusive, 'end' is not.
7187  */
7188 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7189                           unsigned int start, unsigned int end, int trunc)
7190 {
7191         int ret;
7192         unsigned int numbytes;
7193         handle_t *handle;
7194         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7195         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7196         struct ocfs2_inline_data *idata = &di->id2.i_data;
7197
7198         if (end > i_size_read(inode))
7199                 end = i_size_read(inode);
7200
7201         BUG_ON(start > end);
7202
7203         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7204             !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7205             !ocfs2_supports_inline_data(osb)) {
7206                 ocfs2_error(inode->i_sb,
7207                             "Inline data flags for inode %llu don't agree! "
7208                             "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7209                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
7210                             le16_to_cpu(di->i_dyn_features),
7211                             OCFS2_I(inode)->ip_dyn_features,
7212                             osb->s_feature_incompat);
7213                 ret = -EROFS;
7214                 goto out;
7215         }
7216
7217         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7218         if (IS_ERR(handle)) {
7219                 ret = PTR_ERR(handle);
7220                 mlog_errno(ret);
7221                 goto out;
7222         }
7223
7224         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7225                                       OCFS2_JOURNAL_ACCESS_WRITE);
7226         if (ret) {
7227                 mlog_errno(ret);
7228                 goto out_commit;
7229         }
7230
7231         numbytes = end - start;
7232         memset(idata->id_data + start, 0, numbytes);
7233
7234         /*
7235          * No need to worry about the data page here - it's been
7236          * truncated already and inline data doesn't need it for
7237          * pushing zero's to disk, so we'll let readpage pick it up
7238          * later.
7239          */
7240         if (trunc) {
7241                 i_size_write(inode, start);
7242                 di->i_size = cpu_to_le64(start);
7243         }
7244
7245         inode->i_blocks = ocfs2_inode_sector_count(inode);
7246         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
7247
7248         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7249         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7250
7251         ocfs2_update_inode_fsync_trans(handle, inode, 1);
7252         ocfs2_journal_dirty(handle, di_bh);
7253
7254 out_commit:
7255         ocfs2_commit_trans(osb, handle);
7256
7257 out:
7258         return ret;
7259 }
7260
7261 static int ocfs2_trim_extent(struct super_block *sb,
7262                              struct ocfs2_group_desc *gd,
7263                              u32 start, u32 count)
7264 {
7265         u64 discard, bcount;
7266
7267         bcount = ocfs2_clusters_to_blocks(sb, count);
7268         discard = le64_to_cpu(gd->bg_blkno) +
7269                         ocfs2_clusters_to_blocks(sb, start);
7270
7271         trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7272
7273         return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7274 }
7275
7276 static int ocfs2_trim_group(struct super_block *sb,
7277                             struct ocfs2_group_desc *gd,
7278                             u32 start, u32 max, u32 minbits)
7279 {
7280         int ret = 0, count = 0, next;
7281         void *bitmap = gd->bg_bitmap;
7282
7283         if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7284                 return 0;
7285
7286         trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7287                                start, max, minbits);
7288
7289         while (start < max) {
7290                 start = ocfs2_find_next_zero_bit(bitmap, max, start);
7291                 if (start >= max)
7292                         break;
7293                 next = ocfs2_find_next_bit(bitmap, max, start);
7294
7295                 if ((next - start) >= minbits) {
7296                         ret = ocfs2_trim_extent(sb, gd,
7297                                                 start, next - start);
7298                         if (ret < 0) {
7299                                 mlog_errno(ret);
7300                                 break;
7301                         }
7302                         count += next - start;
7303                 }
7304                 start = next + 1;
7305
7306                 if (fatal_signal_pending(current)) {
7307                         count = -ERESTARTSYS;
7308                         break;
7309                 }
7310
7311                 if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7312                         break;
7313         }
7314
7315         if (ret < 0)
7316                 count = ret;
7317
7318         return count;
7319 }
7320
7321 int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7322 {
7323         struct ocfs2_super *osb = OCFS2_SB(sb);
7324         u64 start, len, trimmed, first_group, last_group, group;
7325         int ret, cnt;
7326         u32 first_bit, last_bit, minlen;
7327         struct buffer_head *main_bm_bh = NULL;
7328         struct inode *main_bm_inode = NULL;
7329         struct buffer_head *gd_bh = NULL;
7330         struct ocfs2_dinode *main_bm;
7331         struct ocfs2_group_desc *gd = NULL;
7332
7333         start = range->start >> osb->s_clustersize_bits;
7334         len = range->len >> osb->s_clustersize_bits;
7335         minlen = range->minlen >> osb->s_clustersize_bits;
7336
7337         if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7338                 return -EINVAL;
7339
7340         main_bm_inode = ocfs2_get_system_file_inode(osb,
7341                                                     GLOBAL_BITMAP_SYSTEM_INODE,
7342                                                     OCFS2_INVALID_SLOT);
7343         if (!main_bm_inode) {
7344                 ret = -EIO;
7345                 mlog_errno(ret);
7346                 goto out;
7347         }
7348
7349         mutex_lock(&main_bm_inode->i_mutex);
7350
7351         ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7352         if (ret < 0) {
7353                 mlog_errno(ret);
7354                 goto out_mutex;
7355         }
7356         main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7357
7358         if (start >= le32_to_cpu(main_bm->i_clusters)) {
7359                 ret = -EINVAL;
7360                 goto out_unlock;
7361         }
7362
7363         len = range->len >> osb->s_clustersize_bits;
7364         if (start + len > le32_to_cpu(main_bm->i_clusters))
7365                 len = le32_to_cpu(main_bm->i_clusters) - start;
7366
7367         trace_ocfs2_trim_fs(start, len, minlen);
7368
7369         /* Determine first and last group to examine based on start and len */
7370         first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7371         if (first_group == osb->first_cluster_group_blkno)
7372                 first_bit = start;
7373         else
7374                 first_bit = start - ocfs2_blocks_to_clusters(sb, first_group);
7375         last_group = ocfs2_which_cluster_group(main_bm_inode, start + len - 1);
7376         last_bit = osb->bitmap_cpg;
7377
7378         trimmed = 0;
7379         for (group = first_group; group <= last_group;) {
7380                 if (first_bit + len >= osb->bitmap_cpg)
7381                         last_bit = osb->bitmap_cpg;
7382                 else
7383                         last_bit = first_bit + len;
7384
7385                 ret = ocfs2_read_group_descriptor(main_bm_inode,
7386                                                   main_bm, group,
7387                                                   &gd_bh);
7388                 if (ret < 0) {
7389                         mlog_errno(ret);
7390                         break;
7391                 }
7392
7393                 gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7394                 cnt = ocfs2_trim_group(sb, gd, first_bit, last_bit, minlen);
7395                 brelse(gd_bh);
7396                 gd_bh = NULL;
7397                 if (cnt < 0) {
7398                         ret = cnt;
7399                         mlog_errno(ret);
7400                         break;
7401                 }
7402
7403                 trimmed += cnt;
7404                 len -= osb->bitmap_cpg - first_bit;
7405                 first_bit = 0;
7406                 if (group == osb->first_cluster_group_blkno)
7407                         group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7408                 else
7409                         group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7410         }
7411         range->len = trimmed * sb->s_blocksize;
7412 out_unlock:
7413         ocfs2_inode_unlock(main_bm_inode, 0);
7414         brelse(main_bm_bh);
7415 out_mutex:
7416         mutex_unlock(&main_bm_inode->i_mutex);
7417         iput(main_bm_inode);
7418 out:
7419         return ret;
7420 }