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