Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/* -*- mode: c; c-basic-offset: 8; -*-
   3 * vim: noexpandtab sw=8 ts=8 sts=0:
   4 *
   5 * alloc.c
   6 *
   7 * Extent allocs and frees
   8 *
   9 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
  10 */
  11
  12#include <linux/fs.h>
  13#include <linux/types.h>
  14#include <linux/slab.h>
  15#include <linux/highmem.h>
  16#include <linux/swap.h>
  17#include <linux/quotaops.h>
  18#include <linux/blkdev.h>
  19#include <linux/sched/signal.h>
  20
  21#include <cluster/masklog.h>
  22
  23#include "ocfs2.h"
  24
  25#include "alloc.h"
  26#include "aops.h"
  27#include "blockcheck.h"
  28#include "dlmglue.h"
  29#include "extent_map.h"
  30#include "inode.h"
  31#include "journal.h"
  32#include "localalloc.h"
  33#include "suballoc.h"
  34#include "sysfile.h"
  35#include "file.h"
  36#include "super.h"
  37#include "uptodate.h"
  38#include "xattr.h"
  39#include "refcounttree.h"
  40#include "ocfs2_trace.h"
  41
  42#include "buffer_head_io.h"
  43
  44enum ocfs2_contig_type {
  45	CONTIG_NONE = 0,
  46	CONTIG_LEFT,
  47	CONTIG_RIGHT,
  48	CONTIG_LEFTRIGHT,
  49};
  50
  51static enum ocfs2_contig_type
  52	ocfs2_extent_rec_contig(struct super_block *sb,
  53				struct ocfs2_extent_rec *ext,
  54				struct ocfs2_extent_rec *insert_rec);
  55/*
  56 * Operations for a specific extent tree type.
  57 *
  58 * To implement an on-disk btree (extent tree) type in ocfs2, add
  59 * an ocfs2_extent_tree_operations structure and the matching
  60 * ocfs2_init_<thingy>_extent_tree() function.  That's pretty much it
  61 * for the allocation portion of the extent tree.
  62 */
  63struct ocfs2_extent_tree_operations {
  64	/*
  65	 * last_eb_blk is the block number of the right most leaf extent
  66	 * block.  Most on-disk structures containing an extent tree store
  67	 * this value for fast access.  The ->eo_set_last_eb_blk() and
  68	 * ->eo_get_last_eb_blk() operations access this value.  They are
  69	 *  both required.
  70	 */
  71	void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
  72				   u64 blkno);
  73	u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
  74
  75	/*
  76	 * The on-disk structure usually keeps track of how many total
  77	 * clusters are stored in this extent tree.  This function updates
  78	 * that value.  new_clusters is the delta, and must be
  79	 * added to the total.  Required.
  80	 */
  81	void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
  82				   u32 new_clusters);
  83
  84	/*
  85	 * If this extent tree is supported by an extent map, insert
  86	 * a record into the map.
  87	 */
  88	void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
  89				     struct ocfs2_extent_rec *rec);
  90
  91	/*
  92	 * If this extent tree is supported by an extent map, truncate the
  93	 * map to clusters,
  94	 */
  95	void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
  96				       u32 clusters);
  97
  98	/*
  99	 * If ->eo_insert_check() exists, it is called before rec is
 100	 * inserted into the extent tree.  It is optional.
 101	 */
 102	int (*eo_insert_check)(struct ocfs2_extent_tree *et,
 103			       struct ocfs2_extent_rec *rec);
 104	int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
 105
 106	/*
 107	 * --------------------------------------------------------------
 108	 * The remaining are internal to ocfs2_extent_tree and don't have
 109	 * accessor functions
 110	 */
 111
 112	/*
 113	 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
 114	 * It is required.
 115	 */
 116	void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
 117
 118	/*
 119	 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
 120	 * it exists.  If it does not, et->et_max_leaf_clusters is set
 121	 * to 0 (unlimited).  Optional.
 122	 */
 123	void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
 124
 125	/*
 126	 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
 127	 * are contiguous or not. Optional. Don't need to set it if use
 128	 * ocfs2_extent_rec as the tree leaf.
 129	 */
 130	enum ocfs2_contig_type
 131		(*eo_extent_contig)(struct ocfs2_extent_tree *et,
 132				    struct ocfs2_extent_rec *ext,
 133				    struct ocfs2_extent_rec *insert_rec);
 134};
 135
 136
 137/*
 138 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
 139 * in the methods.
 140 */
 141static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
 142static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
 143					 u64 blkno);
 144static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
 145					 u32 clusters);
 146static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
 147					   struct ocfs2_extent_rec *rec);
 148static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
 149					     u32 clusters);
 150static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
 151				     struct ocfs2_extent_rec *rec);
 152static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
 153static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
 154
 155static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
 156					struct ocfs2_extent_tree *et,
 157					struct buffer_head **new_eb_bh,
 158					int blk_wanted, int *blk_given);
 159static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et);
 160
 161static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
 162	.eo_set_last_eb_blk	= ocfs2_dinode_set_last_eb_blk,
 163	.eo_get_last_eb_blk	= ocfs2_dinode_get_last_eb_blk,
 164	.eo_update_clusters	= ocfs2_dinode_update_clusters,
 165	.eo_extent_map_insert	= ocfs2_dinode_extent_map_insert,
 166	.eo_extent_map_truncate	= ocfs2_dinode_extent_map_truncate,
 167	.eo_insert_check	= ocfs2_dinode_insert_check,
 168	.eo_sanity_check	= ocfs2_dinode_sanity_check,
 169	.eo_fill_root_el	= ocfs2_dinode_fill_root_el,
 170};
 171
 172static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
 173					 u64 blkno)
 174{
 175	struct ocfs2_dinode *di = et->et_object;
 176
 177	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 178	di->i_last_eb_blk = cpu_to_le64(blkno);
 179}
 180
 181static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
 182{
 183	struct ocfs2_dinode *di = et->et_object;
 184
 185	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 186	return le64_to_cpu(di->i_last_eb_blk);
 187}
 188
 189static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
 190					 u32 clusters)
 191{
 192	struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
 193	struct ocfs2_dinode *di = et->et_object;
 194
 195	le32_add_cpu(&di->i_clusters, clusters);
 196	spin_lock(&oi->ip_lock);
 197	oi->ip_clusters = le32_to_cpu(di->i_clusters);
 198	spin_unlock(&oi->ip_lock);
 199}
 200
 201static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
 202					   struct ocfs2_extent_rec *rec)
 203{
 204	struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
 205
 206	ocfs2_extent_map_insert_rec(inode, rec);
 207}
 208
 209static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
 210					     u32 clusters)
 211{
 212	struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
 213
 214	ocfs2_extent_map_trunc(inode, clusters);
 215}
 216
 217static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
 218				     struct ocfs2_extent_rec *rec)
 219{
 220	struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
 221	struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
 222
 223	BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
 224	mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
 225			(oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
 226			"Device %s, asking for sparse allocation: inode %llu, "
 227			"cpos %u, clusters %u\n",
 228			osb->dev_str,
 229			(unsigned long long)oi->ip_blkno,
 230			rec->e_cpos, oi->ip_clusters);
 231
 232	return 0;
 233}
 234
 235static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
 236{
 237	struct ocfs2_dinode *di = et->et_object;
 238
 239	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 240	BUG_ON(!OCFS2_IS_VALID_DINODE(di));
 241
 242	return 0;
 243}
 244
 245static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
 246{
 247	struct ocfs2_dinode *di = et->et_object;
 248
 249	et->et_root_el = &di->id2.i_list;
 250}
 251
 252
 253static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
 254{
 255	struct ocfs2_xattr_value_buf *vb = et->et_object;
 256
 257	et->et_root_el = &vb->vb_xv->xr_list;
 258}
 259
 260static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
 261					      u64 blkno)
 262{
 263	struct ocfs2_xattr_value_buf *vb = et->et_object;
 264
 265	vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
 266}
 267
 268static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
 269{
 270	struct ocfs2_xattr_value_buf *vb = et->et_object;
 271
 272	return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
 273}
 274
 275static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
 276					      u32 clusters)
 277{
 278	struct ocfs2_xattr_value_buf *vb = et->et_object;
 279
 280	le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
 281}
 282
 283static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
 284	.eo_set_last_eb_blk	= ocfs2_xattr_value_set_last_eb_blk,
 285	.eo_get_last_eb_blk	= ocfs2_xattr_value_get_last_eb_blk,
 286	.eo_update_clusters	= ocfs2_xattr_value_update_clusters,
 287	.eo_fill_root_el	= ocfs2_xattr_value_fill_root_el,
 288};
 289
 290static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
 291{
 292	struct ocfs2_xattr_block *xb = et->et_object;
 293
 294	et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
 295}
 296
 297static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
 298{
 299	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
 300	et->et_max_leaf_clusters =
 301		ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
 302}
 303
 304static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
 305					     u64 blkno)
 306{
 307	struct ocfs2_xattr_block *xb = et->et_object;
 308	struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
 309
 310	xt->xt_last_eb_blk = cpu_to_le64(blkno);
 311}
 312
 313static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
 314{
 315	struct ocfs2_xattr_block *xb = et->et_object;
 316	struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
 317
 318	return le64_to_cpu(xt->xt_last_eb_blk);
 319}
 320
 321static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
 322					     u32 clusters)
 323{
 324	struct ocfs2_xattr_block *xb = et->et_object;
 325
 326	le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
 327}
 328
 329static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
 330	.eo_set_last_eb_blk	= ocfs2_xattr_tree_set_last_eb_blk,
 331	.eo_get_last_eb_blk	= ocfs2_xattr_tree_get_last_eb_blk,
 332	.eo_update_clusters	= ocfs2_xattr_tree_update_clusters,
 333	.eo_fill_root_el	= ocfs2_xattr_tree_fill_root_el,
 334	.eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
 335};
 336
 337static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
 338					  u64 blkno)
 339{
 340	struct ocfs2_dx_root_block *dx_root = et->et_object;
 341
 342	dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
 343}
 344
 345static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
 346{
 347	struct ocfs2_dx_root_block *dx_root = et->et_object;
 348
 349	return le64_to_cpu(dx_root->dr_last_eb_blk);
 350}
 351
 352static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
 353					  u32 clusters)
 354{
 355	struct ocfs2_dx_root_block *dx_root = et->et_object;
 356
 357	le32_add_cpu(&dx_root->dr_clusters, clusters);
 358}
 359
 360static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
 361{
 362	struct ocfs2_dx_root_block *dx_root = et->et_object;
 363
 364	BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
 365
 366	return 0;
 367}
 368
 369static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
 370{
 371	struct ocfs2_dx_root_block *dx_root = et->et_object;
 372
 373	et->et_root_el = &dx_root->dr_list;
 374}
 375
 376static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
 377	.eo_set_last_eb_blk	= ocfs2_dx_root_set_last_eb_blk,
 378	.eo_get_last_eb_blk	= ocfs2_dx_root_get_last_eb_blk,
 379	.eo_update_clusters	= ocfs2_dx_root_update_clusters,
 380	.eo_sanity_check	= ocfs2_dx_root_sanity_check,
 381	.eo_fill_root_el	= ocfs2_dx_root_fill_root_el,
 382};
 383
 384static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
 385{
 386	struct ocfs2_refcount_block *rb = et->et_object;
 387
 388	et->et_root_el = &rb->rf_list;
 389}
 390
 391static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
 392						u64 blkno)
 393{
 394	struct ocfs2_refcount_block *rb = et->et_object;
 395
 396	rb->rf_last_eb_blk = cpu_to_le64(blkno);
 397}
 398
 399static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
 400{
 401	struct ocfs2_refcount_block *rb = et->et_object;
 402
 403	return le64_to_cpu(rb->rf_last_eb_blk);
 404}
 405
 406static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
 407						u32 clusters)
 408{
 409	struct ocfs2_refcount_block *rb = et->et_object;
 410
 411	le32_add_cpu(&rb->rf_clusters, clusters);
 412}
 413
 414static enum ocfs2_contig_type
 415ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
 416				  struct ocfs2_extent_rec *ext,
 417				  struct ocfs2_extent_rec *insert_rec)
 418{
 419	return CONTIG_NONE;
 420}
 421
 422static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
 423	.eo_set_last_eb_blk	= ocfs2_refcount_tree_set_last_eb_blk,
 424	.eo_get_last_eb_blk	= ocfs2_refcount_tree_get_last_eb_blk,
 425	.eo_update_clusters	= ocfs2_refcount_tree_update_clusters,
 426	.eo_fill_root_el	= ocfs2_refcount_tree_fill_root_el,
 427	.eo_extent_contig	= ocfs2_refcount_tree_extent_contig,
 428};
 429
 430static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
 431				     struct ocfs2_caching_info *ci,
 432				     struct buffer_head *bh,
 433				     ocfs2_journal_access_func access,
 434				     void *obj,
 435				     const struct ocfs2_extent_tree_operations *ops)
 436{
 437	et->et_ops = ops;
 438	et->et_root_bh = bh;
 439	et->et_ci = ci;
 440	et->et_root_journal_access = access;
 441	if (!obj)
 442		obj = (void *)bh->b_data;
 443	et->et_object = obj;
 444	et->et_dealloc = NULL;
 445
 446	et->et_ops->eo_fill_root_el(et);
 447	if (!et->et_ops->eo_fill_max_leaf_clusters)
 448		et->et_max_leaf_clusters = 0;
 449	else
 450		et->et_ops->eo_fill_max_leaf_clusters(et);
 451}
 452
 453void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
 454				   struct ocfs2_caching_info *ci,
 455				   struct buffer_head *bh)
 456{
 457	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
 458				 NULL, &ocfs2_dinode_et_ops);
 459}
 460
 461void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
 462				       struct ocfs2_caching_info *ci,
 463				       struct buffer_head *bh)
 464{
 465	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
 466				 NULL, &ocfs2_xattr_tree_et_ops);
 467}
 468
 469void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
 470					struct ocfs2_caching_info *ci,
 471					struct ocfs2_xattr_value_buf *vb)
 472{
 473	__ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
 474				 &ocfs2_xattr_value_et_ops);
 475}
 476
 477void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
 478				    struct ocfs2_caching_info *ci,
 479				    struct buffer_head *bh)
 480{
 481	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
 482				 NULL, &ocfs2_dx_root_et_ops);
 483}
 484
 485void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
 486				     struct ocfs2_caching_info *ci,
 487				     struct buffer_head *bh)
 488{
 489	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
 490				 NULL, &ocfs2_refcount_tree_et_ops);
 491}
 492
 493static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
 494					    u64 new_last_eb_blk)
 495{
 496	et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
 497}
 498
 499static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
 500{
 501	return et->et_ops->eo_get_last_eb_blk(et);
 502}
 503
 504static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
 505					    u32 clusters)
 506{
 507	et->et_ops->eo_update_clusters(et, clusters);
 508}
 509
 510static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
 511					      struct ocfs2_extent_rec *rec)
 512{
 513	if (et->et_ops->eo_extent_map_insert)
 514		et->et_ops->eo_extent_map_insert(et, rec);
 515}
 516
 517static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
 518						u32 clusters)
 519{
 520	if (et->et_ops->eo_extent_map_truncate)
 521		et->et_ops->eo_extent_map_truncate(et, clusters);
 522}
 523
 524static inline int ocfs2_et_root_journal_access(handle_t *handle,
 525					       struct ocfs2_extent_tree *et,
 526					       int type)
 527{
 528	return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
 529					  type);
 530}
 531
 532static inline enum ocfs2_contig_type
 533	ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
 534			       struct ocfs2_extent_rec *rec,
 535			       struct ocfs2_extent_rec *insert_rec)
 536{
 537	if (et->et_ops->eo_extent_contig)
 538		return et->et_ops->eo_extent_contig(et, rec, insert_rec);
 539
 540	return ocfs2_extent_rec_contig(
 541				ocfs2_metadata_cache_get_super(et->et_ci),
 542				rec, insert_rec);
 543}
 544
 545static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
 546					struct ocfs2_extent_rec *rec)
 547{
 548	int ret = 0;
 549
 550	if (et->et_ops->eo_insert_check)
 551		ret = et->et_ops->eo_insert_check(et, rec);
 552	return ret;
 553}
 554
 555static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
 556{
 557	int ret = 0;
 558
 559	if (et->et_ops->eo_sanity_check)
 560		ret = et->et_ops->eo_sanity_check(et);
 561	return ret;
 562}
 563
 564static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
 565					 struct ocfs2_extent_block *eb);
 566static void ocfs2_adjust_rightmost_records(handle_t *handle,
 567					   struct ocfs2_extent_tree *et,
 568					   struct ocfs2_path *path,
 569					   struct ocfs2_extent_rec *insert_rec);
 570/*
 571 * Reset the actual path elements so that we can re-use the structure
 572 * to build another path. Generally, this involves freeing the buffer
 573 * heads.
 574 */
 575void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
 576{
 577	int i, start = 0, depth = 0;
 578	struct ocfs2_path_item *node;
 579
 580	if (keep_root)
 581		start = 1;
 582
 583	for(i = start; i < path_num_items(path); i++) {
 584		node = &path->p_node[i];
 585
 586		brelse(node->bh);
 587		node->bh = NULL;
 588		node->el = NULL;
 589	}
 590
 591	/*
 592	 * Tree depth may change during truncate, or insert. If we're
 593	 * keeping the root extent list, then make sure that our path
 594	 * structure reflects the proper depth.
 595	 */
 596	if (keep_root)
 597		depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
 598	else
 599		path_root_access(path) = NULL;
 600
 601	path->p_tree_depth = depth;
 602}
 603
 604void ocfs2_free_path(struct ocfs2_path *path)
 605{
 606	if (path) {
 607		ocfs2_reinit_path(path, 0);
 608		kfree(path);
 609	}
 610}
 611
 612/*
 613 * All the elements of src into dest. After this call, src could be freed
 614 * without affecting dest.
 615 *
 616 * Both paths should have the same root. Any non-root elements of dest
 617 * will be freed.
 618 */
 619static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
 620{
 621	int i;
 622
 623	BUG_ON(path_root_bh(dest) != path_root_bh(src));
 624	BUG_ON(path_root_el(dest) != path_root_el(src));
 625	BUG_ON(path_root_access(dest) != path_root_access(src));
 626
 627	ocfs2_reinit_path(dest, 1);
 628
 629	for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
 630		dest->p_node[i].bh = src->p_node[i].bh;
 631		dest->p_node[i].el = src->p_node[i].el;
 632
 633		if (dest->p_node[i].bh)
 634			get_bh(dest->p_node[i].bh);
 635	}
 636}
 637
 638/*
 639 * Make the *dest path the same as src and re-initialize src path to
 640 * have a root only.
 641 */
 642static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
 643{
 644	int i;
 645
 646	BUG_ON(path_root_bh(dest) != path_root_bh(src));
 647	BUG_ON(path_root_access(dest) != path_root_access(src));
 648
 649	for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
 650		brelse(dest->p_node[i].bh);
 651
 652		dest->p_node[i].bh = src->p_node[i].bh;
 653		dest->p_node[i].el = src->p_node[i].el;
 654
 655		src->p_node[i].bh = NULL;
 656		src->p_node[i].el = NULL;
 657	}
 658}
 659
 660/*
 661 * Insert an extent block at given index.
 662 *
 663 * This will not take an additional reference on eb_bh.
 664 */
 665static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
 666					struct buffer_head *eb_bh)
 667{
 668	struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
 669
 670	/*
 671	 * Right now, no root bh is an extent block, so this helps
 672	 * catch code errors with dinode trees. The assertion can be
 673	 * safely removed if we ever need to insert extent block
 674	 * structures at the root.
 675	 */
 676	BUG_ON(index == 0);
 677
 678	path->p_node[index].bh = eb_bh;
 679	path->p_node[index].el = &eb->h_list;
 680}
 681
 682static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
 683					 struct ocfs2_extent_list *root_el,
 684					 ocfs2_journal_access_func access)
 685{
 686	struct ocfs2_path *path;
 687
 688	BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
 689
 690	path = kzalloc(sizeof(*path), GFP_NOFS);
 691	if (path) {
 692		path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
 693		get_bh(root_bh);
 694		path_root_bh(path) = root_bh;
 695		path_root_el(path) = root_el;
 696		path_root_access(path) = access;
 697	}
 698
 699	return path;
 700}
 701
 702struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
 703{
 704	return ocfs2_new_path(path_root_bh(path), path_root_el(path),
 705			      path_root_access(path));
 706}
 707
 708struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
 709{
 710	return ocfs2_new_path(et->et_root_bh, et->et_root_el,
 711			      et->et_root_journal_access);
 712}
 713
 714/*
 715 * Journal the buffer at depth idx.  All idx>0 are extent_blocks,
 716 * otherwise it's the root_access function.
 717 *
 718 * I don't like the way this function's name looks next to
 719 * ocfs2_journal_access_path(), but I don't have a better one.
 720 */
 721int ocfs2_path_bh_journal_access(handle_t *handle,
 722				 struct ocfs2_caching_info *ci,
 723				 struct ocfs2_path *path,
 724				 int idx)
 725{
 726	ocfs2_journal_access_func access = path_root_access(path);
 727
 728	if (!access)
 729		access = ocfs2_journal_access;
 730
 731	if (idx)
 732		access = ocfs2_journal_access_eb;
 733
 734	return access(handle, ci, path->p_node[idx].bh,
 735		      OCFS2_JOURNAL_ACCESS_WRITE);
 736}
 737
 738/*
 739 * Convenience function to journal all components in a path.
 740 */
 741int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
 742			      handle_t *handle,
 743			      struct ocfs2_path *path)
 744{
 745	int i, ret = 0;
 746
 747	if (!path)
 748		goto out;
 749
 750	for(i = 0; i < path_num_items(path); i++) {
 751		ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
 752		if (ret < 0) {
 753			mlog_errno(ret);
 754			goto out;
 755		}
 756	}
 757
 758out:
 759	return ret;
 760}
 761
 762/*
 763 * Return the index of the extent record which contains cluster #v_cluster.
 764 * -1 is returned if it was not found.
 765 *
 766 * Should work fine on interior and exterior nodes.
 767 */
 768int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
 769{
 770	int ret = -1;
 771	int i;
 772	struct ocfs2_extent_rec *rec;
 773	u32 rec_end, rec_start, clusters;
 774
 775	for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
 776		rec = &el->l_recs[i];
 777
 778		rec_start = le32_to_cpu(rec->e_cpos);
 779		clusters = ocfs2_rec_clusters(el, rec);
 780
 781		rec_end = rec_start + clusters;
 782
 783		if (v_cluster >= rec_start && v_cluster < rec_end) {
 784			ret = i;
 785			break;
 786		}
 787	}
 788
 789	return ret;
 790}
 791
 792/*
 793 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
 794 * ocfs2_extent_rec_contig only work properly against leaf nodes!
 795 */
 796static int ocfs2_block_extent_contig(struct super_block *sb,
 797				     struct ocfs2_extent_rec *ext,
 798				     u64 blkno)
 799{
 800	u64 blk_end = le64_to_cpu(ext->e_blkno);
 801
 802	blk_end += ocfs2_clusters_to_blocks(sb,
 803				    le16_to_cpu(ext->e_leaf_clusters));
 804
 805	return blkno == blk_end;
 806}
 807
 808static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
 809				  struct ocfs2_extent_rec *right)
 810{
 811	u32 left_range;
 812
 813	left_range = le32_to_cpu(left->e_cpos) +
 814		le16_to_cpu(left->e_leaf_clusters);
 815
 816	return (left_range == le32_to_cpu(right->e_cpos));
 817}
 818
 819static enum ocfs2_contig_type
 820	ocfs2_extent_rec_contig(struct super_block *sb,
 821				struct ocfs2_extent_rec *ext,
 822				struct ocfs2_extent_rec *insert_rec)
 823{
 824	u64 blkno = le64_to_cpu(insert_rec->e_blkno);
 825
 826	/*
 827	 * Refuse to coalesce extent records with different flag
 828	 * fields - we don't want to mix unwritten extents with user
 829	 * data.
 830	 */
 831	if (ext->e_flags != insert_rec->e_flags)
 832		return CONTIG_NONE;
 833
 834	if (ocfs2_extents_adjacent(ext, insert_rec) &&
 835	    ocfs2_block_extent_contig(sb, ext, blkno))
 836			return CONTIG_RIGHT;
 837
 838	blkno = le64_to_cpu(ext->e_blkno);
 839	if (ocfs2_extents_adjacent(insert_rec, ext) &&
 840	    ocfs2_block_extent_contig(sb, insert_rec, blkno))
 841		return CONTIG_LEFT;
 842
 843	return CONTIG_NONE;
 844}
 845
 846/*
 847 * NOTE: We can have pretty much any combination of contiguousness and
 848 * appending.
 849 *
 850 * The usefulness of APPEND_TAIL is more in that it lets us know that
 851 * we'll have to update the path to that leaf.
 852 */
 853enum ocfs2_append_type {
 854	APPEND_NONE = 0,
 855	APPEND_TAIL,
 856};
 857
 858enum ocfs2_split_type {
 859	SPLIT_NONE = 0,
 860	SPLIT_LEFT,
 861	SPLIT_RIGHT,
 862};
 863
 864struct ocfs2_insert_type {
 865	enum ocfs2_split_type	ins_split;
 866	enum ocfs2_append_type	ins_appending;
 867	enum ocfs2_contig_type	ins_contig;
 868	int			ins_contig_index;
 869	int			ins_tree_depth;
 870};
 871
 872struct ocfs2_merge_ctxt {
 873	enum ocfs2_contig_type	c_contig_type;
 874	int			c_has_empty_extent;
 875	int			c_split_covers_rec;
 876};
 877
 878static int ocfs2_validate_extent_block(struct super_block *sb,
 879				       struct buffer_head *bh)
 880{
 881	int rc;
 882	struct ocfs2_extent_block *eb =
 883		(struct ocfs2_extent_block *)bh->b_data;
 884
 885	trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
 886
 887	BUG_ON(!buffer_uptodate(bh));
 888
 889	/*
 890	 * If the ecc fails, we return the error but otherwise
 891	 * leave the filesystem running.  We know any error is
 892	 * local to this block.
 893	 */
 894	rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
 895	if (rc) {
 896		mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
 897		     (unsigned long long)bh->b_blocknr);
 898		return rc;
 899	}
 900
 901	/*
 902	 * Errors after here are fatal.
 903	 */
 904
 905	if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
 906		rc = ocfs2_error(sb,
 907				 "Extent block #%llu has bad signature %.*s\n",
 908				 (unsigned long long)bh->b_blocknr, 7,
 909				 eb->h_signature);
 910		goto bail;
 911	}
 912
 913	if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
 914		rc = ocfs2_error(sb,
 915				 "Extent block #%llu has an invalid h_blkno of %llu\n",
 916				 (unsigned long long)bh->b_blocknr,
 917				 (unsigned long long)le64_to_cpu(eb->h_blkno));
 918		goto bail;
 919	}
 920
 921	if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation)
 922		rc = ocfs2_error(sb,
 923				 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
 924				 (unsigned long long)bh->b_blocknr,
 925				 le32_to_cpu(eb->h_fs_generation));
 926bail:
 927	return rc;
 928}
 929
 930int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
 931			    struct buffer_head **bh)
 932{
 933	int rc;
 934	struct buffer_head *tmp = *bh;
 935
 936	rc = ocfs2_read_block(ci, eb_blkno, &tmp,
 937			      ocfs2_validate_extent_block);
 938
 939	/* If ocfs2_read_block() got us a new bh, pass it up. */
 940	if (!rc && !*bh)
 941		*bh = tmp;
 942
 943	return rc;
 944}
 945
 946
 947/*
 948 * How many free extents have we got before we need more meta data?
 949 */
 950int ocfs2_num_free_extents(struct ocfs2_extent_tree *et)
 951{
 952	int retval;
 953	struct ocfs2_extent_list *el = NULL;
 954	struct ocfs2_extent_block *eb;
 955	struct buffer_head *eb_bh = NULL;
 956	u64 last_eb_blk = 0;
 957
 958	el = et->et_root_el;
 959	last_eb_blk = ocfs2_et_get_last_eb_blk(et);
 960
 961	if (last_eb_blk) {
 962		retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
 963						 &eb_bh);
 964		if (retval < 0) {
 965			mlog_errno(retval);
 966			goto bail;
 967		}
 968		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 969		el = &eb->h_list;
 970	}
 971
 972	BUG_ON(el->l_tree_depth != 0);
 973
 974	retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
 975bail:
 976	brelse(eb_bh);
 977
 978	trace_ocfs2_num_free_extents(retval);
 979	return retval;
 980}
 981
 982/* expects array to already be allocated
 983 *
 984 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
 985 * l_count for you
 986 */
 987static int ocfs2_create_new_meta_bhs(handle_t *handle,
 988				     struct ocfs2_extent_tree *et,
 989				     int wanted,
 990				     struct ocfs2_alloc_context *meta_ac,
 991				     struct buffer_head *bhs[])
 992{
 993	int count, status, i;
 994	u16 suballoc_bit_start;
 995	u32 num_got;
 996	u64 suballoc_loc, first_blkno;
 997	struct ocfs2_super *osb =
 998		OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
 999	struct ocfs2_extent_block *eb;
1000
1001	count = 0;
1002	while (count < wanted) {
1003		status = ocfs2_claim_metadata(handle,
1004					      meta_ac,
1005					      wanted - count,
1006					      &suballoc_loc,
1007					      &suballoc_bit_start,
1008					      &num_got,
1009					      &first_blkno);
1010		if (status < 0) {
1011			mlog_errno(status);
1012			goto bail;
1013		}
1014
1015		for(i = count;  i < (num_got + count); i++) {
1016			bhs[i] = sb_getblk(osb->sb, first_blkno);
1017			if (bhs[i] == NULL) {
1018				status = -ENOMEM;
1019				mlog_errno(status);
1020				goto bail;
1021			}
1022			ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1023
1024			status = ocfs2_journal_access_eb(handle, et->et_ci,
1025							 bhs[i],
1026							 OCFS2_JOURNAL_ACCESS_CREATE);
1027			if (status < 0) {
1028				mlog_errno(status);
1029				goto bail;
1030			}
1031
1032			memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1033			eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1034			/* Ok, setup the minimal stuff here. */
1035			strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1036			eb->h_blkno = cpu_to_le64(first_blkno);
1037			eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1038			eb->h_suballoc_slot =
1039				cpu_to_le16(meta_ac->ac_alloc_slot);
1040			eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1041			eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1042			eb->h_list.l_count =
1043				cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1044
1045			suballoc_bit_start++;
1046			first_blkno++;
1047
1048			/* We'll also be dirtied by the caller, so
1049			 * this isn't absolutely necessary. */
1050			ocfs2_journal_dirty(handle, bhs[i]);
1051		}
1052
1053		count += num_got;
1054	}
1055
1056	status = 0;
1057bail:
1058	if (status < 0) {
1059		for(i = 0; i < wanted; i++) {
1060			brelse(bhs[i]);
1061			bhs[i] = NULL;
1062		}
1063		mlog_errno(status);
1064	}
1065	return status;
1066}
1067
1068/*
1069 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1070 *
1071 * Returns the sum of the rightmost extent rec logical offset and
1072 * cluster count.
1073 *
1074 * ocfs2_add_branch() uses this to determine what logical cluster
1075 * value should be populated into the leftmost new branch records.
1076 *
1077 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1078 * value for the new topmost tree record.
1079 */
1080static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list  *el)
1081{
1082	int i;
1083
1084	i = le16_to_cpu(el->l_next_free_rec) - 1;
1085
1086	return le32_to_cpu(el->l_recs[i].e_cpos) +
1087		ocfs2_rec_clusters(el, &el->l_recs[i]);
1088}
1089
1090/*
1091 * Change range of the branches in the right most path according to the leaf
1092 * extent block's rightmost record.
1093 */
1094static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1095					 struct ocfs2_extent_tree *et)
1096{
1097	int status;
1098	struct ocfs2_path *path = NULL;
1099	struct ocfs2_extent_list *el;
1100	struct ocfs2_extent_rec *rec;
1101
1102	path = ocfs2_new_path_from_et(et);
1103	if (!path) {
1104		status = -ENOMEM;
1105		return status;
1106	}
1107
1108	status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1109	if (status < 0) {
1110		mlog_errno(status);
1111		goto out;
1112	}
1113
1114	status = ocfs2_extend_trans(handle, path_num_items(path));
1115	if (status < 0) {
1116		mlog_errno(status);
1117		goto out;
1118	}
1119
1120	status = ocfs2_journal_access_path(et->et_ci, handle, path);
1121	if (status < 0) {
1122		mlog_errno(status);
1123		goto out;
1124	}
1125
1126	el = path_leaf_el(path);
1127	rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1128
1129	ocfs2_adjust_rightmost_records(handle, et, path, rec);
1130
1131out:
1132	ocfs2_free_path(path);
1133	return status;
1134}
1135
1136/*
1137 * Add an entire tree branch to our inode. eb_bh is the extent block
1138 * to start at, if we don't want to start the branch at the root
1139 * structure.
1140 *
1141 * last_eb_bh is required as we have to update it's next_leaf pointer
1142 * for the new last extent block.
1143 *
1144 * the new branch will be 'empty' in the sense that every block will
1145 * contain a single record with cluster count == 0.
1146 */
1147static int ocfs2_add_branch(handle_t *handle,
1148			    struct ocfs2_extent_tree *et,
1149			    struct buffer_head *eb_bh,
1150			    struct buffer_head **last_eb_bh,
1151			    struct ocfs2_alloc_context *meta_ac)
1152{
1153	int status, new_blocks, i, block_given = 0;
1154	u64 next_blkno, new_last_eb_blk;
1155	struct buffer_head *bh;
1156	struct buffer_head **new_eb_bhs = NULL;
1157	struct ocfs2_extent_block *eb;
1158	struct ocfs2_extent_list  *eb_el;
1159	struct ocfs2_extent_list  *el;
1160	u32 new_cpos, root_end;
1161
1162	BUG_ON(!last_eb_bh || !*last_eb_bh);
1163
1164	if (eb_bh) {
1165		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1166		el = &eb->h_list;
1167	} else
1168		el = et->et_root_el;
1169
1170	/* we never add a branch to a leaf. */
1171	BUG_ON(!el->l_tree_depth);
1172
1173	new_blocks = le16_to_cpu(el->l_tree_depth);
1174
1175	eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1176	new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1177	root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1178
1179	/*
1180	 * If there is a gap before the root end and the real end
1181	 * of the righmost leaf block, we need to remove the gap
1182	 * between new_cpos and root_end first so that the tree
1183	 * is consistent after we add a new branch(it will start
1184	 * from new_cpos).
1185	 */
1186	if (root_end > new_cpos) {
1187		trace_ocfs2_adjust_rightmost_branch(
1188			(unsigned long long)
1189			ocfs2_metadata_cache_owner(et->et_ci),
1190			root_end, new_cpos);
1191
1192		status = ocfs2_adjust_rightmost_branch(handle, et);
1193		if (status) {
1194			mlog_errno(status);
1195			goto bail;
1196		}
1197	}
1198
1199	/* allocate the number of new eb blocks we need */
1200	new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1201			     GFP_KERNEL);
1202	if (!new_eb_bhs) {
1203		status = -ENOMEM;
1204		mlog_errno(status);
1205		goto bail;
1206	}
1207
1208	/* Firstyly, try to reuse dealloc since we have already estimated how
1209	 * many extent blocks we may use.
1210	 */
1211	if (!ocfs2_is_dealloc_empty(et)) {
1212		status = ocfs2_reuse_blk_from_dealloc(handle, et,
1213						      new_eb_bhs, new_blocks,
1214						      &block_given);
1215		if (status < 0) {
1216			mlog_errno(status);
1217			goto bail;
1218		}
1219	}
1220
1221	BUG_ON(block_given > new_blocks);
1222
1223	if (block_given < new_blocks) {
1224		BUG_ON(!meta_ac);
1225		status = ocfs2_create_new_meta_bhs(handle, et,
1226						   new_blocks - block_given,
1227						   meta_ac,
1228						   &new_eb_bhs[block_given]);
1229		if (status < 0) {
1230			mlog_errno(status);
1231			goto bail;
1232		}
1233	}
1234
1235	/* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1236	 * linked with the rest of the tree.
1237	 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1238	 *
1239	 * when we leave the loop, new_last_eb_blk will point to the
1240	 * newest leaf, and next_blkno will point to the topmost extent
1241	 * block. */
1242	next_blkno = new_last_eb_blk = 0;
1243	for(i = 0; i < new_blocks; i++) {
1244		bh = new_eb_bhs[i];
1245		eb = (struct ocfs2_extent_block *) bh->b_data;
1246		/* ocfs2_create_new_meta_bhs() should create it right! */
1247		BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1248		eb_el = &eb->h_list;
1249
1250		status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1251						 OCFS2_JOURNAL_ACCESS_CREATE);
1252		if (status < 0) {
1253			mlog_errno(status);
1254			goto bail;
1255		}
1256
1257		eb->h_next_leaf_blk = 0;
1258		eb_el->l_tree_depth = cpu_to_le16(i);
1259		eb_el->l_next_free_rec = cpu_to_le16(1);
1260		/*
1261		 * This actually counts as an empty extent as
1262		 * c_clusters == 0
1263		 */
1264		eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1265		eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1266		/*
1267		 * eb_el isn't always an interior node, but even leaf
1268		 * nodes want a zero'd flags and reserved field so
1269		 * this gets the whole 32 bits regardless of use.
1270		 */
1271		eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1272		if (!eb_el->l_tree_depth)
1273			new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1274
1275		ocfs2_journal_dirty(handle, bh);
1276		next_blkno = le64_to_cpu(eb->h_blkno);
1277	}
1278
1279	/* This is a bit hairy. We want to update up to three blocks
1280	 * here without leaving any of them in an inconsistent state
1281	 * in case of error. We don't have to worry about
1282	 * journal_dirty erroring as it won't unless we've aborted the
1283	 * handle (in which case we would never be here) so reserving
1284	 * the write with journal_access is all we need to do. */
1285	status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1286					 OCFS2_JOURNAL_ACCESS_WRITE);
1287	if (status < 0) {
1288		mlog_errno(status);
1289		goto bail;
1290	}
1291	status = ocfs2_et_root_journal_access(handle, et,
1292					      OCFS2_JOURNAL_ACCESS_WRITE);
1293	if (status < 0) {
1294		mlog_errno(status);
1295		goto bail;
1296	}
1297	if (eb_bh) {
1298		status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1299						 OCFS2_JOURNAL_ACCESS_WRITE);
1300		if (status < 0) {
1301			mlog_errno(status);
1302			goto bail;
1303		}
1304	}
1305
1306	/* Link the new branch into the rest of the tree (el will
1307	 * either be on the root_bh, or the extent block passed in. */
1308	i = le16_to_cpu(el->l_next_free_rec);
1309	el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1310	el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1311	el->l_recs[i].e_int_clusters = 0;
1312	le16_add_cpu(&el->l_next_free_rec, 1);
1313
1314	/* fe needs a new last extent block pointer, as does the
1315	 * next_leaf on the previously last-extent-block. */
1316	ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1317
1318	eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1319	eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1320
1321	ocfs2_journal_dirty(handle, *last_eb_bh);
1322	ocfs2_journal_dirty(handle, et->et_root_bh);
1323	if (eb_bh)
1324		ocfs2_journal_dirty(handle, eb_bh);
1325
1326	/*
1327	 * Some callers want to track the rightmost leaf so pass it
1328	 * back here.
1329	 */
1330	brelse(*last_eb_bh);
1331	get_bh(new_eb_bhs[0]);
1332	*last_eb_bh = new_eb_bhs[0];
1333
1334	status = 0;
1335bail:
1336	if (new_eb_bhs) {
1337		for (i = 0; i < new_blocks; i++)
1338			brelse(new_eb_bhs[i]);
1339		kfree(new_eb_bhs);
1340	}
1341
1342	return status;
1343}
1344
1345/*
1346 * adds another level to the allocation tree.
1347 * returns back the new extent block so you can add a branch to it
1348 * after this call.
1349 */
1350static int ocfs2_shift_tree_depth(handle_t *handle,
1351				  struct ocfs2_extent_tree *et,
1352				  struct ocfs2_alloc_context *meta_ac,
1353				  struct buffer_head **ret_new_eb_bh)
1354{
1355	int status, i, block_given = 0;
1356	u32 new_clusters;
1357	struct buffer_head *new_eb_bh = NULL;
1358	struct ocfs2_extent_block *eb;
1359	struct ocfs2_extent_list  *root_el;
1360	struct ocfs2_extent_list  *eb_el;
1361
1362	if (!ocfs2_is_dealloc_empty(et)) {
1363		status = ocfs2_reuse_blk_from_dealloc(handle, et,
1364						      &new_eb_bh, 1,
1365						      &block_given);
1366	} else if (meta_ac) {
1367		status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1368						   &new_eb_bh);
1369
1370	} else {
1371		BUG();
1372	}
1373
1374	if (status < 0) {
1375		mlog_errno(status);
1376		goto bail;
1377	}
1378
1379	eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1380	/* ocfs2_create_new_meta_bhs() should create it right! */
1381	BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1382
1383	eb_el = &eb->h_list;
1384	root_el = et->et_root_el;
1385
1386	status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1387					 OCFS2_JOURNAL_ACCESS_CREATE);
1388	if (status < 0) {
1389		mlog_errno(status);
1390		goto bail;
1391	}
1392
1393	/* copy the root extent list data into the new extent block */
1394	eb_el->l_tree_depth = root_el->l_tree_depth;
1395	eb_el->l_next_free_rec = root_el->l_next_free_rec;
1396	for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1397		eb_el->l_recs[i] = root_el->l_recs[i];
1398
1399	ocfs2_journal_dirty(handle, new_eb_bh);
1400
1401	status = ocfs2_et_root_journal_access(handle, et,
1402					      OCFS2_JOURNAL_ACCESS_WRITE);
1403	if (status < 0) {
1404		mlog_errno(status);
1405		goto bail;
1406	}
1407
1408	new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1409
1410	/* update root_bh now */
1411	le16_add_cpu(&root_el->l_tree_depth, 1);
1412	root_el->l_recs[0].e_cpos = 0;
1413	root_el->l_recs[0].e_blkno = eb->h_blkno;
1414	root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1415	for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1416		memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1417	root_el->l_next_free_rec = cpu_to_le16(1);
1418
1419	/* If this is our 1st tree depth shift, then last_eb_blk
1420	 * becomes the allocated extent block */
1421	if (root_el->l_tree_depth == cpu_to_le16(1))
1422		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1423
1424	ocfs2_journal_dirty(handle, et->et_root_bh);
1425
1426	*ret_new_eb_bh = new_eb_bh;
1427	new_eb_bh = NULL;
1428	status = 0;
1429bail:
1430	brelse(new_eb_bh);
1431
1432	return status;
1433}
1434
1435/*
1436 * Should only be called when there is no space left in any of the
1437 * leaf nodes. What we want to do is find the lowest tree depth
1438 * non-leaf extent block with room for new records. There are three
1439 * valid results of this search:
1440 *
1441 * 1) a lowest extent block is found, then we pass it back in
1442 *    *lowest_eb_bh and return '0'
1443 *
1444 * 2) the search fails to find anything, but the root_el has room. We
1445 *    pass NULL back in *lowest_eb_bh, but still return '0'
1446 *
1447 * 3) the search fails to find anything AND the root_el is full, in
1448 *    which case we return > 0
1449 *
1450 * return status < 0 indicates an error.
1451 */
1452static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1453				    struct buffer_head **target_bh)
1454{
1455	int status = 0, i;
1456	u64 blkno;
1457	struct ocfs2_extent_block *eb;
1458	struct ocfs2_extent_list  *el;
1459	struct buffer_head *bh = NULL;
1460	struct buffer_head *lowest_bh = NULL;
1461
1462	*target_bh = NULL;
1463
1464	el = et->et_root_el;
1465
1466	while(le16_to_cpu(el->l_tree_depth) > 1) {
1467		if (le16_to_cpu(el->l_next_free_rec) == 0) {
1468			status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1469					"Owner %llu has empty extent list (next_free_rec == 0)\n",
1470					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1471			goto bail;
1472		}
1473		i = le16_to_cpu(el->l_next_free_rec) - 1;
1474		blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1475		if (!blkno) {
1476			status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1477					"Owner %llu has extent list where extent # %d has no physical block start\n",
1478					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1479			goto bail;
1480		}
1481
1482		brelse(bh);
1483		bh = NULL;
1484
1485		status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1486		if (status < 0) {
1487			mlog_errno(status);
1488			goto bail;
1489		}
1490
1491		eb = (struct ocfs2_extent_block *) bh->b_data;
1492		el = &eb->h_list;
1493
1494		if (le16_to_cpu(el->l_next_free_rec) <
1495		    le16_to_cpu(el->l_count)) {
1496			brelse(lowest_bh);
1497			lowest_bh = bh;
1498			get_bh(lowest_bh);
1499		}
1500	}
1501
1502	/* If we didn't find one and the fe doesn't have any room,
1503	 * then return '1' */
1504	el = et->et_root_el;
1505	if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1506		status = 1;
1507
1508	*target_bh = lowest_bh;
1509bail:
1510	brelse(bh);
1511
1512	return status;
1513}
1514
1515/*
1516 * Grow a b-tree so that it has more records.
1517 *
1518 * We might shift the tree depth in which case existing paths should
1519 * be considered invalid.
1520 *
1521 * Tree depth after the grow is returned via *final_depth.
1522 *
1523 * *last_eb_bh will be updated by ocfs2_add_branch().
1524 */
1525static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1526			   int *final_depth, struct buffer_head **last_eb_bh,
1527			   struct ocfs2_alloc_context *meta_ac)
1528{
1529	int ret, shift;
1530	struct ocfs2_extent_list *el = et->et_root_el;
1531	int depth = le16_to_cpu(el->l_tree_depth);
1532	struct buffer_head *bh = NULL;
1533
1534	BUG_ON(meta_ac == NULL && ocfs2_is_dealloc_empty(et));
1535
1536	shift = ocfs2_find_branch_target(et, &bh);
1537	if (shift < 0) {
1538		ret = shift;
1539		mlog_errno(ret);
1540		goto out;
1541	}
1542
1543	/* We traveled all the way to the bottom of the allocation tree
1544	 * and didn't find room for any more extents - we need to add
1545	 * another tree level */
1546	if (shift) {
1547		BUG_ON(bh);
1548		trace_ocfs2_grow_tree(
1549			(unsigned long long)
1550			ocfs2_metadata_cache_owner(et->et_ci),
1551			depth);
1552
1553		/* ocfs2_shift_tree_depth will return us a buffer with
1554		 * the new extent block (so we can pass that to
1555		 * ocfs2_add_branch). */
1556		ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1557		if (ret < 0) {
1558			mlog_errno(ret);
1559			goto out;
1560		}
1561		depth++;
1562		if (depth == 1) {
1563			/*
1564			 * Special case: we have room now if we shifted from
1565			 * tree_depth 0, so no more work needs to be done.
1566			 *
1567			 * We won't be calling add_branch, so pass
1568			 * back *last_eb_bh as the new leaf. At depth
1569			 * zero, it should always be null so there's
1570			 * no reason to brelse.
1571			 */
1572			BUG_ON(*last_eb_bh);
1573			get_bh(bh);
1574			*last_eb_bh = bh;
1575			goto out;
1576		}
1577	}
1578
1579	/* call ocfs2_add_branch to add the final part of the tree with
1580	 * the new data. */
1581	ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1582			       meta_ac);
1583	if (ret < 0)
1584		mlog_errno(ret);
1585
1586out:
1587	if (final_depth)
1588		*final_depth = depth;
1589	brelse(bh);
1590	return ret;
1591}
1592
1593/*
1594 * This function will discard the rightmost extent record.
1595 */
1596static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1597{
1598	int next_free = le16_to_cpu(el->l_next_free_rec);
1599	int count = le16_to_cpu(el->l_count);
1600	unsigned int num_bytes;
1601
1602	BUG_ON(!next_free);
1603	/* This will cause us to go off the end of our extent list. */
1604	BUG_ON(next_free >= count);
1605
1606	num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1607
1608	memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1609}
1610
1611static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1612			      struct ocfs2_extent_rec *insert_rec)
1613{
1614	int i, insert_index, next_free, has_empty, num_bytes;
1615	u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1616	struct ocfs2_extent_rec *rec;
1617
1618	next_free = le16_to_cpu(el->l_next_free_rec);
1619	has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1620
1621	BUG_ON(!next_free);
1622
1623	/* The tree code before us didn't allow enough room in the leaf. */
1624	BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1625
1626	/*
1627	 * The easiest way to approach this is to just remove the
1628	 * empty extent and temporarily decrement next_free.
1629	 */
1630	if (has_empty) {
1631		/*
1632		 * If next_free was 1 (only an empty extent), this
1633		 * loop won't execute, which is fine. We still want
1634		 * the decrement above to happen.
1635		 */
1636		for(i = 0; i < (next_free - 1); i++)
1637			el->l_recs[i] = el->l_recs[i+1];
1638
1639		next_free--;
1640	}
1641
1642	/*
1643	 * Figure out what the new record index should be.
1644	 */
1645	for(i = 0; i < next_free; i++) {
1646		rec = &el->l_recs[i];
1647
1648		if (insert_cpos < le32_to_cpu(rec->e_cpos))
1649			break;
1650	}
1651	insert_index = i;
1652
1653	trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1654				has_empty, next_free,
1655				le16_to_cpu(el->l_count));
1656
1657	BUG_ON(insert_index < 0);
1658	BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1659	BUG_ON(insert_index > next_free);
1660
1661	/*
1662	 * No need to memmove if we're just adding to the tail.
1663	 */
1664	if (insert_index != next_free) {
1665		BUG_ON(next_free >= le16_to_cpu(el->l_count));
1666
1667		num_bytes = next_free - insert_index;
1668		num_bytes *= sizeof(struct ocfs2_extent_rec);
1669		memmove(&el->l_recs[insert_index + 1],
1670			&el->l_recs[insert_index],
1671			num_bytes);
1672	}
1673
1674	/*
1675	 * Either we had an empty extent, and need to re-increment or
1676	 * there was no empty extent on a non full rightmost leaf node,
1677	 * in which case we still need to increment.
1678	 */
1679	next_free++;
1680	el->l_next_free_rec = cpu_to_le16(next_free);
1681	/*
1682	 * Make sure none of the math above just messed up our tree.
1683	 */
1684	BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1685
1686	el->l_recs[insert_index] = *insert_rec;
1687
1688}
1689
1690static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1691{
1692	int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1693
1694	BUG_ON(num_recs == 0);
1695
1696	if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1697		num_recs--;
1698		size = num_recs * sizeof(struct ocfs2_extent_rec);
1699		memmove(&el->l_recs[0], &el->l_recs[1], size);
1700		memset(&el->l_recs[num_recs], 0,
1701		       sizeof(struct ocfs2_extent_rec));
1702		el->l_next_free_rec = cpu_to_le16(num_recs);
1703	}
1704}
1705
1706/*
1707 * Create an empty extent record .
1708 *
1709 * l_next_free_rec may be updated.
1710 *
1711 * If an empty extent already exists do nothing.
1712 */
1713static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1714{
1715	int next_free = le16_to_cpu(el->l_next_free_rec);
1716
1717	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1718
1719	if (next_free == 0)
1720		goto set_and_inc;
1721
1722	if (ocfs2_is_empty_extent(&el->l_recs[0]))
1723		return;
1724
1725	mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1726			"Asked to create an empty extent in a full list:\n"
1727			"count = %u, tree depth = %u",
1728			le16_to_cpu(el->l_count),
1729			le16_to_cpu(el->l_tree_depth));
1730
1731	ocfs2_shift_records_right(el);
1732
1733set_and_inc:
1734	le16_add_cpu(&el->l_next_free_rec, 1);
1735	memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1736}
1737
1738/*
1739 * For a rotation which involves two leaf nodes, the "root node" is
1740 * the lowest level tree node which contains a path to both leafs. This
1741 * resulting set of information can be used to form a complete "subtree"
1742 *
1743 * This function is passed two full paths from the dinode down to a
1744 * pair of adjacent leaves. It's task is to figure out which path
1745 * index contains the subtree root - this can be the root index itself
1746 * in a worst-case rotation.
1747 *
1748 * The array index of the subtree root is passed back.
1749 */
1750int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1751			    struct ocfs2_path *left,
1752			    struct ocfs2_path *right)
1753{
1754	int i = 0;
1755
1756	/*
1757	 * Check that the caller passed in two paths from the same tree.
1758	 */
1759	BUG_ON(path_root_bh(left) != path_root_bh(right));
1760
1761	do {
1762		i++;
1763
1764		/*
1765		 * The caller didn't pass two adjacent paths.
1766		 */
1767		mlog_bug_on_msg(i > left->p_tree_depth,
1768				"Owner %llu, left depth %u, right depth %u\n"
1769				"left leaf blk %llu, right leaf blk %llu\n",
1770				(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1771				left->p_tree_depth, right->p_tree_depth,
1772				(unsigned long long)path_leaf_bh(left)->b_blocknr,
1773				(unsigned long long)path_leaf_bh(right)->b_blocknr);
1774	} while (left->p_node[i].bh->b_blocknr ==
1775		 right->p_node[i].bh->b_blocknr);
1776
1777	return i - 1;
1778}
1779
1780typedef void (path_insert_t)(void *, struct buffer_head *);
1781
1782/*
1783 * Traverse a btree path in search of cpos, starting at root_el.
1784 *
1785 * This code can be called with a cpos larger than the tree, in which
1786 * case it will return the rightmost path.
1787 */
1788static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1789			     struct ocfs2_extent_list *root_el, u32 cpos,
1790			     path_insert_t *func, void *data)
1791{
1792	int i, ret = 0;
1793	u32 range;
1794	u64 blkno;
1795	struct buffer_head *bh = NULL;
1796	struct ocfs2_extent_block *eb;
1797	struct ocfs2_extent_list *el;
1798	struct ocfs2_extent_rec *rec;
1799
1800	el = root_el;
1801	while (el->l_tree_depth) {
1802		if (le16_to_cpu(el->l_next_free_rec) == 0) {
1803			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1804				    "Owner %llu has empty extent list at depth %u\n",
1805				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
1806				    le16_to_cpu(el->l_tree_depth));
1807			ret = -EROFS;
1808			goto out;
1809
1810		}
1811
1812		for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1813			rec = &el->l_recs[i];
1814
1815			/*
1816			 * In the case that cpos is off the allocation
1817			 * tree, this should just wind up returning the
1818			 * rightmost record.
1819			 */
1820			range = le32_to_cpu(rec->e_cpos) +
1821				ocfs2_rec_clusters(el, rec);
1822			if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1823			    break;
1824		}
1825
1826		blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1827		if (blkno == 0) {
1828			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1829				    "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1830				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
1831				    le16_to_cpu(el->l_tree_depth), i);
1832			ret = -EROFS;
1833			goto out;
1834		}
1835
1836		brelse(bh);
1837		bh = NULL;
1838		ret = ocfs2_read_extent_block(ci, blkno, &bh);
1839		if (ret) {
1840			mlog_errno(ret);
1841			goto out;
1842		}
1843
1844		eb = (struct ocfs2_extent_block *) bh->b_data;
1845		el = &eb->h_list;
1846
1847		if (le16_to_cpu(el->l_next_free_rec) >
1848		    le16_to_cpu(el->l_count)) {
1849			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1850				    "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1851				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
1852				    (unsigned long long)bh->b_blocknr,
1853				    le16_to_cpu(el->l_next_free_rec),
1854				    le16_to_cpu(el->l_count));
1855			ret = -EROFS;
1856			goto out;
1857		}
1858
1859		if (func)
1860			func(data, bh);
1861	}
1862
1863out:
1864	/*
1865	 * Catch any trailing bh that the loop didn't handle.
1866	 */
1867	brelse(bh);
1868
1869	return ret;
1870}
1871
1872/*
1873 * Given an initialized path (that is, it has a valid root extent
1874 * list), this function will traverse the btree in search of the path
1875 * which would contain cpos.
1876 *
1877 * The path traveled is recorded in the path structure.
1878 *
1879 * Note that this will not do any comparisons on leaf node extent
1880 * records, so it will work fine in the case that we just added a tree
1881 * branch.
1882 */
1883struct find_path_data {
1884	int index;
1885	struct ocfs2_path *path;
1886};
1887static void find_path_ins(void *data, struct buffer_head *bh)
1888{
1889	struct find_path_data *fp = data;
1890
1891	get_bh(bh);
1892	ocfs2_path_insert_eb(fp->path, fp->index, bh);
1893	fp->index++;
1894}
1895int ocfs2_find_path(struct ocfs2_caching_info *ci,
1896		    struct ocfs2_path *path, u32 cpos)
1897{
1898	struct find_path_data data;
1899
1900	data.index = 1;
1901	data.path = path;
1902	return __ocfs2_find_path(ci, path_root_el(path), cpos,
1903				 find_path_ins, &data);
1904}
1905
1906static void find_leaf_ins(void *data, struct buffer_head *bh)
1907{
1908	struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1909	struct ocfs2_extent_list *el = &eb->h_list;
1910	struct buffer_head **ret = data;
1911
1912	/* We want to retain only the leaf block. */
1913	if (le16_to_cpu(el->l_tree_depth) == 0) {
1914		get_bh(bh);
1915		*ret = bh;
1916	}
1917}
1918/*
1919 * Find the leaf block in the tree which would contain cpos. No
1920 * checking of the actual leaf is done.
1921 *
1922 * Some paths want to call this instead of allocating a path structure
1923 * and calling ocfs2_find_path().
1924 *
1925 * This function doesn't handle non btree extent lists.
1926 */
1927int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1928		    struct ocfs2_extent_list *root_el, u32 cpos,
1929		    struct buffer_head **leaf_bh)
1930{
1931	int ret;
1932	struct buffer_head *bh = NULL;
1933
1934	ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1935	if (ret) {
1936		mlog_errno(ret);
1937		goto out;
1938	}
1939
1940	*leaf_bh = bh;
1941out:
1942	return ret;
1943}
1944
1945/*
1946 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1947 *
1948 * Basically, we've moved stuff around at the bottom of the tree and
1949 * we need to fix up the extent records above the changes to reflect
1950 * the new changes.
1951 *
1952 * left_rec: the record on the left.
1953 * right_rec: the record to the right of left_rec
1954 * right_child_el: is the child list pointed to by right_rec
1955 *
1956 * By definition, this only works on interior nodes.
1957 */
1958static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1959				  struct ocfs2_extent_rec *right_rec,
1960				  struct ocfs2_extent_list *right_child_el)
1961{
1962	u32 left_clusters, right_end;
1963
1964	/*
1965	 * Interior nodes never have holes. Their cpos is the cpos of
1966	 * the leftmost record in their child list. Their cluster
1967	 * count covers the full theoretical range of their child list
1968	 * - the range between their cpos and the cpos of the record
1969	 * immediately to their right.
1970	 */
1971	left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1972	if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1973		BUG_ON(right_child_el->l_tree_depth);
1974		BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1975		left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1976	}
1977	left_clusters -= le32_to_cpu(left_rec->e_cpos);
1978	left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1979
1980	/*
1981	 * Calculate the rightmost cluster count boundary before
1982	 * moving cpos - we will need to adjust clusters after
1983	 * updating e_cpos to keep the same highest cluster count.
1984	 */
1985	right_end = le32_to_cpu(right_rec->e_cpos);
1986	right_end += le32_to_cpu(right_rec->e_int_clusters);
1987
1988	right_rec->e_cpos = left_rec->e_cpos;
1989	le32_add_cpu(&right_rec->e_cpos, left_clusters);
1990
1991	right_end -= le32_to_cpu(right_rec->e_cpos);
1992	right_rec->e_int_clusters = cpu_to_le32(right_end);
1993}
1994
1995/*
1996 * Adjust the adjacent root node records involved in a
1997 * rotation. left_el_blkno is passed in as a key so that we can easily
1998 * find it's index in the root list.
1999 */
2000static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
2001				      struct ocfs2_extent_list *left_el,
2002				      struct ocfs2_extent_list *right_el,
2003				      u64 left_el_blkno)
2004{
2005	int i;
2006
2007	BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2008	       le16_to_cpu(left_el->l_tree_depth));
2009
2010	for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2011		if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2012			break;
2013	}
2014
2015	/*
2016	 * The path walking code should have never returned a root and
2017	 * two paths which are not adjacent.
2018	 */
2019	BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2020
2021	ocfs2_adjust_adjacent_records(&root_el->l_recs[i],
2022				      &root_el->l_recs[i + 1], right_el);
2023}
2024
2025/*
2026 * We've changed a leaf block (in right_path) and need to reflect that
2027 * change back up the subtree.
2028 *
2029 * This happens in multiple places:
2030 *   - When we've moved an extent record from the left path leaf to the right
2031 *     path leaf to make room for an empty extent in the left path leaf.
2032 *   - When our insert into the right path leaf is at the leftmost edge
2033 *     and requires an update of the path immediately to it's left. This
2034 *     can occur at the end of some types of rotation and appending inserts.
2035 *   - When we've adjusted the last extent record in the left path leaf and the
2036 *     1st extent record in the right path leaf during cross extent block merge.
2037 */
2038static void ocfs2_complete_edge_insert(handle_t *handle,
2039				       struct ocfs2_path *left_path,
2040				       struct ocfs2_path *right_path,
2041				       int subtree_index)
2042{
2043	int i, idx;
2044	struct ocfs2_extent_list *el, *left_el, *right_el;
2045	struct ocfs2_extent_rec *left_rec, *right_rec;
2046	struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2047
2048	/*
2049	 * Update the counts and position values within all the
2050	 * interior nodes to reflect the leaf rotation we just did.
2051	 *
2052	 * The root node is handled below the loop.
2053	 *
2054	 * We begin the loop with right_el and left_el pointing to the
2055	 * leaf lists and work our way up.
2056	 *
2057	 * NOTE: within this loop, left_el and right_el always refer
2058	 * to the *child* lists.
2059	 */
2060	left_el = path_leaf_el(left_path);
2061	right_el = path_leaf_el(right_path);
2062	for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2063		trace_ocfs2_complete_edge_insert(i);
2064
2065		/*
2066		 * One nice property of knowing that all of these
2067		 * nodes are below the root is that we only deal with
2068		 * the leftmost right node record and the rightmost
2069		 * left node record.
2070		 */
2071		el = left_path->p_node[i].el;
2072		idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2073		left_rec = &el->l_recs[idx];
2074
2075		el = right_path->p_node[i].el;
2076		right_rec = &el->l_recs[0];
2077
2078		ocfs2_adjust_adjacent_records(left_rec, right_rec, right_el);
2079
2080		ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2081		ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2082
2083		/*
2084		 * Setup our list pointers now so that the current
2085		 * parents become children in the next iteration.
2086		 */
2087		left_el = left_path->p_node[i].el;
2088		right_el = right_path->p_node[i].el;
2089	}
2090
2091	/*
2092	 * At the root node, adjust the two adjacent records which
2093	 * begin our path to the leaves.
2094	 */
2095
2096	el = left_path->p_node[subtree_index].el;
2097	left_el = left_path->p_node[subtree_index + 1].el;
2098	right_el = right_path->p_node[subtree_index + 1].el;
2099
2100	ocfs2_adjust_root_records(el, left_el, right_el,
2101				  left_path->p_node[subtree_index + 1].bh->b_blocknr);
2102
2103	root_bh = left_path->p_node[subtree_index].bh;
2104
2105	ocfs2_journal_dirty(handle, root_bh);
2106}
2107
2108static int ocfs2_rotate_subtree_right(handle_t *handle,
2109				      struct ocfs2_extent_tree *et,
2110				      struct ocfs2_path *left_path,
2111				      struct ocfs2_path *right_path,
2112				      int subtree_index)
2113{
2114	int ret, i;
2115	struct buffer_head *right_leaf_bh;
2116	struct buffer_head *left_leaf_bh = NULL;
2117	struct buffer_head *root_bh;
2118	struct ocfs2_extent_list *right_el, *left_el;
2119	struct ocfs2_extent_rec move_rec;
2120
2121	left_leaf_bh = path_leaf_bh(left_path);
2122	left_el = path_leaf_el(left_path);
2123
2124	if (left_el->l_next_free_rec != left_el->l_count) {
2125		ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2126			    "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2127			    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2128			    (unsigned long long)left_leaf_bh->b_blocknr,
2129			    le16_to_cpu(left_el->l_next_free_rec));
2130		return -EROFS;
2131	}
2132
2133	/*
2134	 * This extent block may already have an empty record, so we
2135	 * return early if so.
2136	 */
2137	if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2138		return 0;
2139
2140	root_bh = left_path->p_node[subtree_index].bh;
2141	BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2142
2143	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2144					   subtree_index);
2145	if (ret) {
2146		mlog_errno(ret);
2147		goto out;
2148	}
2149
2150	for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2151		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2152						   right_path, i);
2153		if (ret) {
2154			mlog_errno(ret);
2155			goto out;
2156		}
2157
2158		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2159						   left_path, i);
2160		if (ret) {
2161			mlog_errno(ret);
2162			goto out;
2163		}
2164	}
2165
2166	right_leaf_bh = path_leaf_bh(right_path);
2167	right_el = path_leaf_el(right_path);
2168
2169	/* This is a code error, not a disk corruption. */
2170	mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2171			"because rightmost leaf block %llu is empty\n",
2172			(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2173			(unsigned long long)right_leaf_bh->b_blocknr);
2174
2175	ocfs2_create_empty_extent(right_el);
2176
2177	ocfs2_journal_dirty(handle, right_leaf_bh);
2178
2179	/* Do the copy now. */
2180	i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2181	move_rec = left_el->l_recs[i];
2182	right_el->l_recs[0] = move_rec;
2183
2184	/*
2185	 * Clear out the record we just copied and shift everything
2186	 * over, leaving an empty extent in the left leaf.
2187	 *
2188	 * We temporarily subtract from next_free_rec so that the
2189	 * shift will lose the tail record (which is now defunct).
2190	 */
2191	le16_add_cpu(&left_el->l_next_free_rec, -1);
2192	ocfs2_shift_records_right(left_el);
2193	memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2194	le16_add_cpu(&left_el->l_next_free_rec, 1);
2195
2196	ocfs2_journal_dirty(handle, left_leaf_bh);
2197
2198	ocfs2_complete_edge_insert(handle, left_path, right_path,
2199				   subtree_index);
2200
2201out:
2202	return ret;
2203}
2204
2205/*
2206 * Given a full path, determine what cpos value would return us a path
2207 * containing the leaf immediately to the left of the current one.
2208 *
2209 * Will return zero if the path passed in is already the leftmost path.
2210 */
2211int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2212				  struct ocfs2_path *path, u32 *cpos)
2213{
2214	int i, j, ret = 0;
2215	u64 blkno;
2216	struct ocfs2_extent_list *el;
2217
2218	BUG_ON(path->p_tree_depth == 0);
2219
2220	*cpos = 0;
2221
2222	blkno = path_leaf_bh(path)->b_blocknr;
2223
2224	/* Start at the tree node just above the leaf and work our way up. */
2225	i = path->p_tree_depth - 1;
2226	while (i >= 0) {
2227		el = path->p_node[i].el;
2228
2229		/*
2230		 * Find the extent record just before the one in our
2231		 * path.
2232		 */
2233		for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2234			if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2235				if (j == 0) {
2236					if (i == 0) {
2237						/*
2238						 * We've determined that the
2239						 * path specified is already
2240						 * the leftmost one - return a
2241						 * cpos of zero.
2242						 */
2243						goto out;
2244					}
2245					/*
2246					 * The leftmost record points to our
2247					 * leaf - we need to travel up the
2248					 * tree one level.
2249					 */
2250					goto next_node;
2251				}
2252
2253				*cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2254				*cpos = *cpos + ocfs2_rec_clusters(el,
2255							   &el->l_recs[j - 1]);
2256				*cpos = *cpos - 1;
2257				goto out;
2258			}
2259		}
2260
2261		/*
2262		 * If we got here, we never found a valid node where
2263		 * the tree indicated one should be.
2264		 */
2265		ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2266			    (unsigned long long)blkno);
2267		ret = -EROFS;
2268		goto out;
2269
2270next_node:
2271		blkno = path->p_node[i].bh->b_blocknr;
2272		i--;
2273	}
2274
2275out:
2276	return ret;
2277}
2278
2279/*
2280 * Extend the transaction by enough credits to complete the rotation,
2281 * and still leave at least the original number of credits allocated
2282 * to this transaction.
2283 */
2284static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2285					   int op_credits,
2286					   struct ocfs2_path *path)
2287{
2288	int ret = 0;
2289	int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2290
2291	if (handle->h_buffer_credits < credits)
2292		ret = ocfs2_extend_trans(handle,
2293					 credits - handle->h_buffer_credits);
2294
2295	return ret;
2296}
2297
2298/*
2299 * Trap the case where we're inserting into the theoretical range past
2300 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2301 * whose cpos is less than ours into the right leaf.
2302 *
2303 * It's only necessary to look at the rightmost record of the left
2304 * leaf because the logic that calls us should ensure that the
2305 * theoretical ranges in the path components above the leaves are
2306 * correct.
2307 */
2308static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2309						 u32 insert_cpos)
2310{
2311	struct ocfs2_extent_list *left_el;
2312	struct ocfs2_extent_rec *rec;
2313	int next_free;
2314
2315	left_el = path_leaf_el(left_path);
2316	next_free = le16_to_cpu(left_el->l_next_free_rec);
2317	rec = &left_el->l_recs[next_free - 1];
2318
2319	if (insert_cpos > le32_to_cpu(rec->e_cpos))
2320		return 1;
2321	return 0;
2322}
2323
2324static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2325{
2326	int next_free = le16_to_cpu(el->l_next_free_rec);
2327	unsigned int range;
2328	struct ocfs2_extent_rec *rec;
2329
2330	if (next_free == 0)
2331		return 0;
2332
2333	rec = &el->l_recs[0];
2334	if (ocfs2_is_empty_extent(rec)) {
2335		/* Empty list. */
2336		if (next_free == 1)
2337			return 0;
2338		rec = &el->l_recs[1];
2339	}
2340
2341	range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2342	if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2343		return 1;
2344	return 0;
2345}
2346
2347/*
2348 * Rotate all the records in a btree right one record, starting at insert_cpos.
2349 *
2350 * The path to the rightmost leaf should be passed in.
2351 *
2352 * The array is assumed to be large enough to hold an entire path (tree depth).
2353 *
2354 * Upon successful return from this function:
2355 *
2356 * - The 'right_path' array will contain a path to the leaf block
2357 *   whose range contains e_cpos.
2358 * - That leaf block will have a single empty extent in list index 0.
2359 * - In the case that the rotation requires a post-insert update,
2360 *   *ret_left_path will contain a valid path which can be passed to
2361 *   ocfs2_insert_path().
2362 */
2363static int ocfs2_rotate_tree_right(handle_t *handle,
2364				   struct ocfs2_extent_tree *et,
2365				   enum ocfs2_split_type split,
2366				   u32 insert_cpos,
2367				   struct ocfs2_path *right_path,
2368				   struct ocfs2_path **ret_left_path)
2369{
2370	int ret, start, orig_credits = handle->h_buffer_credits;
2371	u32 cpos;
2372	struct ocfs2_path *left_path = NULL;
2373	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2374
2375	*ret_left_path = NULL;
2376
2377	left_path = ocfs2_new_path_from_path(right_path);
2378	if (!left_path) {
2379		ret = -ENOMEM;
2380		mlog_errno(ret);
2381		goto out;
2382	}
2383
2384	ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2385	if (ret) {
2386		mlog_errno(ret);
2387		goto out;
2388	}
2389
2390	trace_ocfs2_rotate_tree_right(
2391		(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2392		insert_cpos, cpos);
2393
2394	/*
2395	 * What we want to do here is:
2396	 *
2397	 * 1) Start with the rightmost path.
2398	 *
2399	 * 2) Determine a path to the leaf block directly to the left
2400	 *    of that leaf.
2401	 *
2402	 * 3) Determine the 'subtree root' - the lowest level tree node
2403	 *    which contains a path to both leaves.
2404	 *
2405	 * 4) Rotate the subtree.
2406	 *
2407	 * 5) Find the next subtree by considering the left path to be
2408	 *    the new right path.
2409	 *
2410	 * The check at the top of this while loop also accepts
2411	 * insert_cpos == cpos because cpos is only a _theoretical_
2412	 * value to get us the left path - insert_cpos might very well
2413	 * be filling that hole.
2414	 *
2415	 * Stop at a cpos of '0' because we either started at the
2416	 * leftmost branch (i.e., a tree with one branch and a
2417	 * rotation inside of it), or we've gone as far as we can in
2418	 * rotating subtrees.
2419	 */
2420	while (cpos && insert_cpos <= cpos) {
2421		trace_ocfs2_rotate_tree_right(
2422			(unsigned long long)
2423			ocfs2_metadata_cache_owner(et->et_ci),
2424			insert_cpos, cpos);
2425
2426		ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2427		if (ret) {
2428			mlog_errno(ret);
2429			goto out;
2430		}
2431
2432		mlog_bug_on_msg(path_leaf_bh(left_path) ==
2433				path_leaf_bh(right_path),
2434				"Owner %llu: error during insert of %u "
2435				"(left path cpos %u) results in two identical "
2436				"paths ending at %llu\n",
2437				(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2438				insert_cpos, cpos,
2439				(unsigned long long)
2440				path_leaf_bh(left_path)->b_blocknr);
2441
2442		if (split == SPLIT_NONE &&
2443		    ocfs2_rotate_requires_path_adjustment(left_path,
2444							  insert_cpos)) {
2445
2446			/*
2447			 * We've rotated the tree as much as we
2448			 * should. The rest is up to
2449			 * ocfs2_insert_path() to complete, after the
2450			 * record insertion. We indicate this
2451			 * situation by returning the left path.
2452			 *
2453			 * The reason we don't adjust the records here
2454			 * before the record insert is that an error
2455			 * later might break the rule where a parent
2456			 * record e_cpos will reflect the actual
2457			 * e_cpos of the 1st nonempty record of the
2458			 * child list.
2459			 */
2460			*ret_left_path = left_path;
2461			goto out_ret_path;
2462		}
2463
2464		start = ocfs2_find_subtree_root(et, left_path, right_path);
2465
2466		trace_ocfs2_rotate_subtree(start,
2467			(unsigned long long)
2468			right_path->p_node[start].bh->b_blocknr,
2469			right_path->p_tree_depth);
2470
2471		ret = ocfs2_extend_rotate_transaction(handle, start,
2472						      orig_credits, right_path);
2473		if (ret) {
2474			mlog_errno(ret);
2475			goto out;
2476		}
2477
2478		ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2479						 right_path, start);
2480		if (ret) {
2481			mlog_errno(ret);
2482			goto out;
2483		}
2484
2485		if (split != SPLIT_NONE &&
2486		    ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2487						insert_cpos)) {
2488			/*
2489			 * A rotate moves the rightmost left leaf
2490			 * record over to the leftmost right leaf
2491			 * slot. If we're doing an extent split
2492			 * instead of a real insert, then we have to
2493			 * check that the extent to be split wasn't
2494			 * just moved over. If it was, then we can
2495			 * exit here, passing left_path back -
2496			 * ocfs2_split_extent() is smart enough to
2497			 * search both leaves.
2498			 */
2499			*ret_left_path = left_path;
2500			goto out_ret_path;
2501		}
2502
2503		/*
2504		 * There is no need to re-read the next right path
2505		 * as we know that it'll be our current left
2506		 * path. Optimize by copying values instead.
2507		 */
2508		ocfs2_mv_path(right_path, left_path);
2509
2510		ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2511		if (ret) {
2512			mlog_errno(ret);
2513			goto out;
2514		}
2515	}
2516
2517out:
2518	ocfs2_free_path(left_path);
2519
2520out_ret_path:
2521	return ret;
2522}
2523
2524static int ocfs2_update_edge_lengths(handle_t *handle,
2525				     struct ocfs2_extent_tree *et,
2526				     struct ocfs2_path *path)
2527{
2528	int i, idx, ret;
2529	struct ocfs2_extent_rec *rec;
2530	struct ocfs2_extent_list *el;
2531	struct ocfs2_extent_block *eb;
2532	u32 range;
2533
2534	ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2535	if (ret) {
2536		mlog_errno(ret);
2537		goto out;
2538	}
2539
2540	/* Path should always be rightmost. */
2541	eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2542	BUG_ON(eb->h_next_leaf_blk != 0ULL);
2543
2544	el = &eb->h_list;
2545	BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2546	idx = le16_to_cpu(el->l_next_free_rec) - 1;
2547	rec = &el->l_recs[idx];
2548	range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2549
2550	for (i = 0; i < path->p_tree_depth; i++) {
2551		el = path->p_node[i].el;
2552		idx = le16_to_cpu(el->l_next_free_rec) - 1;
2553		rec = &el->l_recs[idx];
2554
2555		rec->e_int_clusters = cpu_to_le32(range);
2556		le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2557
2558		ocfs2_journal_dirty(handle, path->p_node[i].bh);
2559	}
2560out:
2561	return ret;
2562}
2563
2564static void ocfs2_unlink_path(handle_t *handle,
2565			      struct ocfs2_extent_tree *et,
2566			      struct ocfs2_cached_dealloc_ctxt *dealloc,
2567			      struct ocfs2_path *path, int unlink_start)
2568{
2569	int ret, i;
2570	struct ocfs2_extent_block *eb;
2571	struct ocfs2_extent_list *el;
2572	struct buffer_head *bh;
2573
2574	for(i = unlink_start; i < path_num_items(path); i++) {
2575		bh = path->p_node[i].bh;
2576
2577		eb = (struct ocfs2_extent_block *)bh->b_data;
2578		/*
2579		 * Not all nodes might have had their final count
2580		 * decremented by the caller - handle this here.
2581		 */
2582		el = &eb->h_list;
2583		if (le16_to_cpu(el->l_next_free_rec) > 1) {
2584			mlog(ML_ERROR,
2585			     "Inode %llu, attempted to remove extent block "
2586			     "%llu with %u records\n",
2587			     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2588			     (unsigned long long)le64_to_cpu(eb->h_blkno),
2589			     le16_to_cpu(el->l_next_free_rec));
2590
2591			ocfs2_journal_dirty(handle, bh);
2592			ocfs2_remove_from_cache(et->et_ci, bh);
2593			continue;
2594		}
2595
2596		el->l_next_free_rec = 0;
2597		memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2598
2599		ocfs2_journal_dirty(handle, bh);
2600
2601		ret = ocfs2_cache_extent_block_free(dealloc, eb);
2602		if (ret)
2603			mlog_errno(ret);
2604
2605		ocfs2_remove_from_cache(et->et_ci, bh);
2606	}
2607}
2608
2609static void ocfs2_unlink_subtree(handle_t *handle,
2610				 struct ocfs2_extent_tree *et,
2611				 struct ocfs2_path *left_path,
2612				 struct ocfs2_path *right_path,
2613				 int subtree_index,
2614				 struct ocfs2_cached_dealloc_ctxt *dealloc)
2615{
2616	int i;
2617	struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2618	struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2619	struct ocfs2_extent_block *eb;
2620
2621	eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2622
2623	for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2624		if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2625			break;
2626
2627	BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2628
2629	memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2630	le16_add_cpu(&root_el->l_next_free_rec, -1);
2631
2632	eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2633	eb->h_next_leaf_blk = 0;
2634
2635	ocfs2_journal_dirty(handle, root_bh);
2636	ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2637
2638	ocfs2_unlink_path(handle, et, dealloc, right_path,
2639			  subtree_index + 1);
2640}
2641
2642static int ocfs2_rotate_subtree_left(handle_t *handle,
2643				     struct ocfs2_extent_tree *et,
2644				     struct ocfs2_path *left_path,
2645				     struct ocfs2_path *right_path,
2646				     int subtree_index,
2647				     struct ocfs2_cached_dealloc_ctxt *dealloc,
2648				     int *deleted)
2649{
2650	int ret, i, del_right_subtree = 0, right_has_empty = 0;
2651	struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2652	struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2653	struct ocfs2_extent_block *eb;
2654
2655	*deleted = 0;
2656
2657	right_leaf_el = path_leaf_el(right_path);
2658	left_leaf_el = path_leaf_el(left_path);
2659	root_bh = left_path->p_node[subtree_index].bh;
2660	BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2661
2662	if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2663		return 0;
2664
2665	eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2666	if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2667		/*
2668		 * It's legal for us to proceed if the right leaf is
2669		 * the rightmost one and it has an empty extent. There
2670		 * are two cases to handle - whether the leaf will be
2671		 * empty after removal or not. If the leaf isn't empty
2672		 * then just remove the empty extent up front. The
2673		 * next block will handle empty leaves by flagging
2674		 * them for unlink.
2675		 *
2676		 * Non rightmost leaves will throw -EAGAIN and the
2677		 * caller can manually move the subtree and retry.
2678		 */
2679
2680		if (eb->h_next_leaf_blk != 0ULL)
2681			return -EAGAIN;
2682
2683		if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2684			ret = ocfs2_journal_access_eb(handle, et->et_ci,
2685						      path_leaf_bh(right_path),
2686						      OCFS2_JOURNAL_ACCESS_WRITE);
2687			if (ret) {
2688				mlog_errno(ret);
2689				goto out;
2690			}
2691
2692			ocfs2_remove_empty_extent(right_leaf_el);
2693		} else
2694			right_has_empty = 1;
2695	}
2696
2697	if (eb->h_next_leaf_blk == 0ULL &&
2698	    le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2699		/*
2700		 * We have to update i_last_eb_blk during the meta
2701		 * data delete.
2702		 */
2703		ret = ocfs2_et_root_journal_access(handle, et,
2704						   OCFS2_JOURNAL_ACCESS_WRITE);
2705		if (ret) {
2706			mlog_errno(ret);
2707			goto out;
2708		}
2709
2710		del_right_subtree = 1;
2711	}
2712
2713	/*
2714	 * Getting here with an empty extent in the right path implies
2715	 * that it's the rightmost path and will be deleted.
2716	 */
2717	BUG_ON(right_has_empty && !del_right_subtree);
2718
2719	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2720					   subtree_index);
2721	if (ret) {
2722		mlog_errno(ret);
2723		goto out;
2724	}
2725
2726	for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2727		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2728						   right_path, i);
2729		if (ret) {
2730			mlog_errno(ret);
2731			goto out;
2732		}
2733
2734		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2735						   left_path, i);
2736		if (ret) {
2737			mlog_errno(ret);
2738			goto out;
2739		}
2740	}
2741
2742	if (!right_has_empty) {
2743		/*
2744		 * Only do this if we're moving a real
2745		 * record. Otherwise, the action is delayed until
2746		 * after removal of the right path in which case we
2747		 * can do a simple shift to remove the empty extent.
2748		 */
2749		ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2750		memset(&right_leaf_el->l_recs[0], 0,
2751		       sizeof(struct ocfs2_extent_rec));
2752	}
2753	if (eb->h_next_leaf_blk == 0ULL) {
2754		/*
2755		 * Move recs over to get rid of empty extent, decrease
2756		 * next_free. This is allowed to remove the last
2757		 * extent in our leaf (setting l_next_free_rec to
2758		 * zero) - the delete code below won't care.
2759		 */
2760		ocfs2_remove_empty_extent(right_leaf_el);
2761	}
2762
2763	ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2764	ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2765
2766	if (del_right_subtree) {
2767		ocfs2_unlink_subtree(handle, et, left_path, right_path,
2768				     subtree_index, dealloc);
2769		ret = ocfs2_update_edge_lengths(handle, et, left_path);
2770		if (ret) {
2771			mlog_errno(ret);
2772			goto out;
2773		}
2774
2775		eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2776		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2777
2778		/*
2779		 * Removal of the extent in the left leaf was skipped
2780		 * above so we could delete the right path
2781		 * 1st.
2782		 */
2783		if (right_has_empty)
2784			ocfs2_remove_empty_extent(left_leaf_el);
2785
2786		ocfs2_journal_dirty(handle, et_root_bh);
2787
2788		*deleted = 1;
2789	} else
2790		ocfs2_complete_edge_insert(handle, left_path, right_path,
2791					   subtree_index);
2792
2793out:
2794	return ret;
2795}
2796
2797/*
2798 * Given a full path, determine what cpos value would return us a path
2799 * containing the leaf immediately to the right of the current one.
2800 *
2801 * Will return zero if the path passed in is already the rightmost path.
2802 *
2803 * This looks similar, but is subtly different to
2804 * ocfs2_find_cpos_for_left_leaf().
2805 */
2806int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2807				   struct ocfs2_path *path, u32 *cpos)
2808{
2809	int i, j, ret = 0;
2810	u64 blkno;
2811	struct ocfs2_extent_list *el;
2812
2813	*cpos = 0;
2814
2815	if (path->p_tree_depth == 0)
2816		return 0;
2817
2818	blkno = path_leaf_bh(path)->b_blocknr;
2819
2820	/* Start at the tree node just above the leaf and work our way up. */
2821	i = path->p_tree_depth - 1;
2822	while (i >= 0) {
2823		int next_free;
2824
2825		el = path->p_node[i].el;
2826
2827		/*
2828		 * Find the extent record just after the one in our
2829		 * path.
2830		 */
2831		next_free = le16_to_cpu(el->l_next_free_rec);
2832		for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2833			if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2834				if (j == (next_free - 1)) {
2835					if (i == 0) {
2836						/*
2837						 * We've determined that the
2838						 * path specified is already
2839						 * the rightmost one - return a
2840						 * cpos of zero.
2841						 */
2842						goto out;
2843					}
2844					/*
2845					 * The rightmost record points to our
2846					 * leaf - we need to travel up the
2847					 * tree one level.
2848					 */
2849					goto next_node;
2850				}
2851
2852				*cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2853				goto out;
2854			}
2855		}
2856
2857		/*
2858		 * If we got here, we never found a valid node where
2859		 * the tree indicated one should be.
2860		 */
2861		ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2862			    (unsigned long long)blkno);
2863		ret = -EROFS;
2864		goto out;
2865
2866next_node:
2867		blkno = path->p_node[i].bh->b_blocknr;
2868		i--;
2869	}
2870
2871out:
2872	return ret;
2873}
2874
2875static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2876					    struct ocfs2_extent_tree *et,
2877					    struct ocfs2_path *path)
2878{
2879	int ret;
2880	struct buffer_head *bh = path_leaf_bh(path);
2881	struct ocfs2_extent_list *el = path_leaf_el(path);
2882
2883	if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2884		return 0;
2885
2886	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2887					   path_num_items(path) - 1);
2888	if (ret) {
2889		mlog_errno(ret);
2890		goto out;
2891	}
2892
2893	ocfs2_remove_empty_extent(el);
2894	ocfs2_journal_dirty(handle, bh);
2895
2896out:
2897	return ret;
2898}
2899
2900static int __ocfs2_rotate_tree_left(handle_t *handle,
2901				    struct ocfs2_extent_tree *et,
2902				    int orig_credits,
2903				    struct ocfs2_path *path,
2904				    struct ocfs2_cached_dealloc_ctxt *dealloc,
2905				    struct ocfs2_path **empty_extent_path)
2906{
2907	int ret, subtree_root, deleted;
2908	u32 right_cpos;
2909	struct ocfs2_path *left_path = NULL;
2910	struct ocfs2_path *right_path = NULL;
2911	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2912
2913	if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2914		return 0;
2915
2916	*empty_extent_path = NULL;
2917
2918	ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2919	if (ret) {
2920		mlog_errno(ret);
2921		goto out;
2922	}
2923
2924	left_path = ocfs2_new_path_from_path(path);
2925	if (!left_path) {
2926		ret = -ENOMEM;
2927		mlog_errno(ret);
2928		goto out;
2929	}
2930
2931	ocfs2_cp_path(left_path, path);
2932
2933	right_path = ocfs2_new_path_from_path(path);
2934	if (!right_path) {
2935		ret = -ENOMEM;
2936		mlog_errno(ret);
2937		goto out;
2938	}
2939
2940	while (right_cpos) {
2941		ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2942		if (ret) {
2943			mlog_errno(ret);
2944			goto out;
2945		}
2946
2947		subtree_root = ocfs2_find_subtree_root(et, left_path,
2948						       right_path);
2949
2950		trace_ocfs2_rotate_subtree(subtree_root,
2951		     (unsigned long long)
2952		     right_path->p_node[subtree_root].bh->b_blocknr,
2953		     right_path->p_tree_depth);
2954
2955		ret = ocfs2_extend_rotate_transaction(handle, 0,
2956						      orig_credits, left_path);
2957		if (ret) {
2958			mlog_errno(ret);
2959			goto out;
2960		}
2961
2962		/*
2963		 * Caller might still want to make changes to the
2964		 * tree root, so re-add it to the journal here.
2965		 */
2966		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2967						   left_path, 0);
2968		if (ret) {
2969			mlog_errno(ret);
2970			goto out;
2971		}
2972
2973		ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2974						right_path, subtree_root,
2975						dealloc, &deleted);
2976		if (ret == -EAGAIN) {
2977			/*
2978			 * The rotation has to temporarily stop due to
2979			 * the right subtree having an empty
2980			 * extent. Pass it back to the caller for a
2981			 * fixup.
2982			 */
2983			*empty_extent_path = right_path;
2984			right_path = NULL;
2985			goto out;
2986		}
2987		if (ret) {
2988			mlog_errno(ret);
2989			goto out;
2990		}
2991
2992		/*
2993		 * The subtree rotate might have removed records on
2994		 * the rightmost edge. If so, then rotation is
2995		 * complete.
2996		 */
2997		if (deleted)
2998			break;
2999
3000		ocfs2_mv_path(left_path, right_path);
3001
3002		ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3003						     &right_cpos);
3004		if (ret) {
3005			mlog_errno(ret);
3006			goto out;
3007		}
3008	}
3009
3010out:
3011	ocfs2_free_path(right_path);
3012	ocfs2_free_path(left_path);
3013
3014	return ret;
3015}
3016
3017static int ocfs2_remove_rightmost_path(handle_t *handle,
3018				struct ocfs2_extent_tree *et,
3019				struct ocfs2_path *path,
3020				struct ocfs2_cached_dealloc_ctxt *dealloc)
3021{
3022	int ret, subtree_index;
3023	u32 cpos;
3024	struct ocfs2_path *left_path = NULL;
3025	struct ocfs2_extent_block *eb;
3026	struct ocfs2_extent_list *el;
3027
3028	ret = ocfs2_et_sanity_check(et);
3029	if (ret)
3030		goto out;
3031
3032	ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3033	if (ret) {
3034		mlog_errno(ret);
3035		goto out;
3036	}
3037
3038	ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3039					    path, &cpos);
3040	if (ret) {
3041		mlog_errno(ret);
3042		goto out;
3043	}
3044
3045	if (cpos) {
3046		/*
3047		 * We have a path to the left of this one - it needs
3048		 * an update too.
3049		 */
3050		left_path = ocfs2_new_path_from_path(path);
3051		if (!left_path) {
3052			ret = -ENOMEM;
3053			mlog_errno(ret);
3054			goto out;
3055		}
3056
3057		ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3058		if (ret) {
3059			mlog_errno(ret);
3060			goto out;
3061		}
3062
3063		ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3064		if (ret) {
3065			mlog_errno(ret);
3066			goto out;
3067		}
3068
3069		subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3070
3071		ocfs2_unlink_subtree(handle, et, left_path, path,
3072				     subtree_index, dealloc);
3073		ret = ocfs2_update_edge_lengths(handle, et, left_path);
3074		if (ret) {
3075			mlog_errno(ret);
3076			goto out;
3077		}
3078
3079		eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3080		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3081	} else {
3082		/*
3083		 * 'path' is also the leftmost path which
3084		 * means it must be the only one. This gets
3085		 * handled differently because we want to
3086		 * revert the root back to having extents
3087		 * in-line.
3088		 */
3089		ocfs2_unlink_path(handle, et, dealloc, path, 1);
3090
3091		el = et->et_root_el;
3092		el->l_tree_depth = 0;
3093		el->l_next_free_rec = 0;
3094		memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3095
3096		ocfs2_et_set_last_eb_blk(et, 0);
3097	}
3098
3099	ocfs2_journal_dirty(handle, path_root_bh(path));
3100
3101out:
3102	ocfs2_free_path(left_path);
3103	return ret;
3104}
3105
3106static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3107				struct ocfs2_extent_tree *et,
3108				struct ocfs2_path *path,
3109				struct ocfs2_cached_dealloc_ctxt *dealloc)
3110{
3111	handle_t *handle;
3112	int ret;
3113	int credits = path->p_tree_depth * 2 + 1;
3114
3115	handle = ocfs2_start_trans(osb, credits);
3116	if (IS_ERR(handle)) {
3117		ret = PTR_ERR(handle);
3118		mlog_errno(ret);
3119		return ret;
3120	}
3121
3122	ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3123	if (ret)
3124		mlog_errno(ret);
3125
3126	ocfs2_commit_trans(osb, handle);
3127	return ret;
3128}
3129
3130/*
3131 * Left rotation of btree records.
3132 *
3133 * In many ways, this is (unsurprisingly) the opposite of right
3134 * rotation. We start at some non-rightmost path containing an empty
3135 * extent in the leaf block. The code works its way to the rightmost
3136 * path by rotating records to the left in every subtree.
3137 *
3138 * This is used by any code which reduces the number of extent records
3139 * in a leaf. After removal, an empty record should be placed in the
3140 * leftmost list position.
3141 *
3142 * This won't handle a length update of the rightmost path records if
3143 * the rightmost tree leaf record is removed so the caller is
3144 * responsible for detecting and correcting that.
3145 */
3146static int ocfs2_rotate_tree_left(handle_t *handle,
3147				  struct ocfs2_extent_tree *et,
3148				  struct ocfs2_path *path,
3149				  struct ocfs2_cached_dealloc_ctxt *dealloc)
3150{
3151	int ret, orig_credits = handle->h_buffer_credits;
3152	struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3153	struct ocfs2_extent_block *eb;
3154	struct ocfs2_extent_list *el;
3155
3156	el = path_leaf_el(path);
3157	if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3158		return 0;
3159
3160	if (path->p_tree_depth == 0) {
3161rightmost_no_delete:
3162		/*
3163		 * Inline extents. This is trivially handled, so do
3164		 * it up front.
3165		 */
3166		ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3167		if (ret)
3168			mlog_errno(ret);
3169		goto out;
3170	}
3171
3172	/*
3173	 * Handle rightmost branch now. There's several cases:
3174	 *  1) simple rotation leaving records in there. That's trivial.
3175	 *  2) rotation requiring a branch delete - there's no more
3176	 *     records left. Two cases of this:
3177	 *     a) There are branches to the left.
3178	 *     b) This is also the leftmost (the only) branch.
3179	 *
3180	 *  1) is handled via ocfs2_rotate_rightmost_leaf_left()
3181	 *  2a) we need the left branch so that we can update it with the unlink
3182	 *  2b) we need to bring the root back to inline extents.
3183	 */
3184
3185	eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3186	el = &eb->h_list;
3187	if (eb->h_next_leaf_blk == 0) {
3188		/*
3189		 * This gets a bit tricky if we're going to delete the
3190		 * rightmost path. Get the other cases out of the way
3191		 * 1st.
3192		 */
3193		if (le16_to_cpu(el->l_next_free_rec) > 1)
3194			goto rightmost_no_delete;
3195
3196		if (le16_to_cpu(el->l_next_free_rec) == 0) {
3197			ret = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3198					"Owner %llu has empty extent block at %llu\n",
3199					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3200					(unsigned long long)le64_to_cpu(eb->h_blkno));
3201			goto out;
3202		}
3203
3204		/*
3205		 * XXX: The caller can not trust "path" any more after
3206		 * this as it will have been deleted. What do we do?
3207		 *
3208		 * In theory the rotate-for-merge code will never get
3209		 * here because it'll always ask for a rotate in a
3210		 * nonempty list.
3211		 */
3212
3213		ret = ocfs2_remove_rightmost_path(handle, et, path,
3214						  dealloc);
3215		if (ret)
3216			mlog_errno(ret);
3217		goto out;
3218	}
3219
3220	/*
3221	 * Now we can loop, remembering the path we get from -EAGAIN
3222	 * and restarting from there.
3223	 */
3224try_rotate:
3225	ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3226				       dealloc, &restart_path);
3227	if (ret && ret != -EAGAIN) {
3228		mlog_errno(ret);
3229		goto out;
3230	}
3231
3232	while (ret == -EAGAIN) {
3233		tmp_path = restart_path;
3234		restart_path = NULL;
3235
3236		ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3237					       tmp_path, dealloc,
3238					       &restart_path);
3239		if (ret && ret != -EAGAIN) {
3240			mlog_errno(ret);
3241			goto out;
3242		}
3243
3244		ocfs2_free_path(tmp_path);
3245		tmp_path = NULL;
3246
3247		if (ret == 0)
3248			goto try_rotate;
3249	}
3250
3251out:
3252	ocfs2_free_path(tmp_path);
3253	ocfs2_free_path(restart_path);
3254	return ret;
3255}
3256
3257static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3258				int index)
3259{
3260	struct ocfs2_extent_rec *rec = &el->l_recs[index];
3261	unsigned int size;
3262
3263	if (rec->e_leaf_clusters == 0) {
3264		/*
3265		 * We consumed all of the merged-from record. An empty
3266		 * extent cannot exist anywhere but the 1st array
3267		 * position, so move things over if the merged-from
3268		 * record doesn't occupy that position.
3269		 *
3270		 * This creates a new empty extent so the caller
3271		 * should be smart enough to have removed any existing
3272		 * ones.
3273		 */
3274		if (index > 0) {
3275			BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3276			size = index * sizeof(struct ocfs2_extent_rec);
3277			memmove(&el->l_recs[1], &el->l_recs[0], size);
3278		}
3279
3280		/*
3281		 * Always memset - the caller doesn't check whether it
3282		 * created an empty extent, so there could be junk in
3283		 * the other fields.
3284		 */
3285		memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3286	}
3287}
3288
3289static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3290				struct ocfs2_path *left_path,
3291				struct ocfs2_path **ret_right_path)
3292{
3293	int ret;
3294	u32 right_cpos;
3295	struct ocfs2_path *right_path = NULL;
3296	struct ocfs2_extent_list *left_el;
3297
3298	*ret_right_path = NULL;
3299
3300	/* This function shouldn't be called for non-trees. */
3301	BUG_ON(left_path->p_tree_depth == 0);
3302
3303	left_el = path_leaf_el(left_path);
3304	BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3305
3306	ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3307					     left_path, &right_cpos);
3308	if (ret) {
3309		mlog_errno(ret);
3310		goto out;
3311	}
3312
3313	/* This function shouldn't be called for the rightmost leaf. */
3314	BUG_ON(right_cpos == 0);
3315
3316	right_path = ocfs2_new_path_from_path(left_path);
3317	if (!right_path) {
3318		ret = -ENOMEM;
3319		mlog_errno(ret);
3320		goto out;
3321	}
3322
3323	ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3324	if (ret) {
3325		mlog_errno(ret);
3326		goto out;
3327	}
3328
3329	*ret_right_path = right_path;
3330out:
3331	if (ret)
3332		ocfs2_free_path(right_path);
3333	return ret;
3334}
3335
3336/*
3337 * Remove split_rec clusters from the record at index and merge them
3338 * onto the beginning of the record "next" to it.
3339 * For index < l_count - 1, the next means the extent rec at index + 1.
3340 * For index == l_count - 1, the "next" means the 1st extent rec of the
3341 * next extent block.
3342 */
3343static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3344				 handle_t *handle,
3345				 struct ocfs2_extent_tree *et,
3346				 struct ocfs2_extent_rec *split_rec,
3347				 int index)
3348{
3349	int ret, next_free, i;
3350	unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3351	struct ocfs2_extent_rec *left_rec;
3352	struct ocfs2_extent_rec *right_rec;
3353	struct ocfs2_extent_list *right_el;
3354	struct ocfs2_path *right_path = NULL;
3355	int subtree_index = 0;
3356	struct ocfs2_extent_list *el = path_leaf_el(left_path);
3357	struct buffer_head *bh = path_leaf_bh(left_path);
3358	struct buffer_head *root_bh = NULL;
3359
3360	BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3361	left_rec = &el->l_recs[index];
3362
3363	if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3364	    le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3365		/* we meet with a cross extent block merge. */
3366		ret = ocfs2_get_right_path(et, left_path, &right_path);
3367		if (ret) {
3368			mlog_errno(ret);
3369			return ret;
3370		}
3371
3372		right_el = path_leaf_el(right_path);
3373		next_free = le16_to_cpu(right_el->l_next_free_rec);
3374		BUG_ON(next_free <= 0);
3375		right_rec = &right_el->l_recs[0];
3376		if (ocfs2_is_empty_extent(right_rec)) {
3377			BUG_ON(next_free <= 1);
3378			right_rec = &right_el->l_recs[1];
3379		}
3380
3381		BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3382		       le16_to_cpu(left_rec->e_leaf_clusters) !=
3383		       le32_to_cpu(right_rec->e_cpos));
3384
3385		subtree_index = ocfs2_find_subtree_root(et, left_path,
3386							right_path);
3387
3388		ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3389						      handle->h_buffer_credits,
3390						      right_path);
3391		if (ret) {
3392			mlog_errno(ret);
3393			goto out;
3394		}
3395
3396		root_bh = left_path->p_node[subtree_index].bh;
3397		BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3398
3399		ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3400						   subtree_index);
3401		if (ret) {
3402			mlog_errno(ret);
3403			goto out;
3404		}
3405
3406		for (i = subtree_index + 1;
3407		     i < path_num_items(right_path); i++) {
3408			ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3409							   right_path, i);
3410			if (ret) {
3411				mlog_errno(ret);
3412				goto out;
3413			}
3414
3415			ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3416							   left_path, i);
3417			if (ret) {
3418				mlog_errno(ret);
3419				goto out;
3420			}
3421		}
3422
3423	} else {
3424		BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3425		right_rec = &el->l_recs[index + 1];
3426	}
3427
3428	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3429					   path_num_items(left_path) - 1);
3430	if (ret) {
3431		mlog_errno(ret);
3432		goto out;
3433	}
3434
3435	le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3436
3437	le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3438	le64_add_cpu(&right_rec->e_blkno,
3439		     -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3440					       split_clusters));
3441	le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3442
3443	ocfs2_cleanup_merge(el, index);
3444
3445	ocfs2_journal_dirty(handle, bh);
3446	if (right_path) {
3447		ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3448		ocfs2_complete_edge_insert(handle, left_path, right_path,
3449					   subtree_index);
3450	}
3451out:
3452	ocfs2_free_path(right_path);
3453	return ret;
3454}
3455
3456static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3457			       struct ocfs2_path *right_path,
3458			       struct ocfs2_path **ret_left_path)
3459{
3460	int ret;
3461	u32 left_cpos;
3462	struct ocfs2_path *left_path = NULL;
3463
3464	*ret_left_path = NULL;
3465
3466	/* This function shouldn't be called for non-trees. */
3467	BUG_ON(right_path->p_tree_depth == 0);
3468
3469	ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3470					    right_path, &left_cpos);
3471	if (ret) {
3472		mlog_errno(ret);
3473		goto out;
3474	}
3475
3476	/* This function shouldn't be called for the leftmost leaf. */
3477	BUG_ON(left_cpos == 0);
3478
3479	left_path = ocfs2_new_path_from_path(right_path);
3480	if (!left_path) {
3481		ret = -ENOMEM;
3482		mlog_errno(ret);
3483		goto out;
3484	}
3485
3486	ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3487	if (ret) {
3488		mlog_errno(ret);
3489		goto out;
3490	}
3491
3492	*ret_left_path = left_path;
3493out:
3494	if (ret)
3495		ocfs2_free_path(left_path);
3496	return ret;
3497}
3498
3499/*
3500 * Remove split_rec clusters from the record at index and merge them
3501 * onto the tail of the record "before" it.
3502 * For index > 0, the "before" means the extent rec at index - 1.
3503 *
3504 * For index == 0, the "before" means the last record of the previous
3505 * extent block. And there is also a situation that we may need to
3506 * remove the rightmost leaf extent block in the right_path and change
3507 * the right path to indicate the new rightmost path.
3508 */
3509static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3510				handle_t *handle,
3511				struct ocfs2_extent_tree *et,
3512				struct ocfs2_extent_rec *split_rec,
3513				struct ocfs2_cached_dealloc_ctxt *dealloc,
3514				int index)
3515{
3516	int ret, i, subtree_index = 0, has_empty_extent = 0;
3517	unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3518	struct ocfs2_extent_rec *left_rec;
3519	struct ocfs2_extent_rec *right_rec;
3520	struct ocfs2_extent_list *el = path_leaf_el(right_path);
3521	struct buffer_head *bh = path_leaf_bh(right_path);
3522	struct buffer_head *root_bh = NULL;
3523	struct ocfs2_path *left_path = NULL;
3524	struct ocfs2_extent_list *left_el;
3525
3526	BUG_ON(index < 0);
3527
3528	right_rec = &el->l_recs[index];
3529	if (index == 0) {
3530		/* we meet with a cross extent block merge. */
3531		ret = ocfs2_get_left_path(et, right_path, &left_path);
3532		if (ret) {
3533			mlog_errno(ret);
3534			return ret;
3535		}
3536
3537		left_el = path_leaf_el(left_path);
3538		BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3539		       le16_to_cpu(left_el->l_count));
3540
3541		left_rec = &left_el->l_recs[
3542				le16_to_cpu(left_el->l_next_free_rec) - 1];
3543		BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3544		       le16_to_cpu(left_rec->e_leaf_clusters) !=
3545		       le32_to_cpu(split_rec->e_cpos));
3546
3547		subtree_index = ocfs2_find_subtree_root(et, left_path,
3548							right_path);
3549
3550		ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3551						      handle->h_buffer_credits,
3552						      left_path);
3553		if (ret) {
3554			mlog_errno(ret);
3555			goto out;
3556		}
3557
3558		root_bh = left_path->p_node[subtree_index].bh;
3559		BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3560
3561		ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3562						   subtree_index);
3563		if (ret) {
3564			mlog_errno(ret);
3565			goto out;
3566		}
3567
3568		for (i = subtree_index + 1;
3569		     i < path_num_items(right_path); i++) {
3570			ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3571							   right_path, i);
3572			if (ret) {
3573				mlog_errno(ret);
3574				goto out;
3575			}
3576
3577			ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3578							   left_path, i);
3579			if (ret) {
3580				mlog_errno(ret);
3581				goto out;
3582			}
3583		}
3584	} else {
3585		left_rec = &el->l_recs[index - 1];
3586		if (ocfs2_is_empty_extent(&el->l_recs[0]))
3587			has_empty_extent = 1;
3588	}
3589
3590	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3591					   path_num_items(right_path) - 1);
3592	if (ret) {
3593		mlog_errno(ret);
3594		goto out;
3595	}
3596
3597	if (has_empty_extent && index == 1) {
3598		/*
3599		 * The easy case - we can just plop the record right in.
3600		 */
3601		*left_rec = *split_rec;
3602	} else
3603		le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3604
3605	le32_add_cpu(&right_rec->e_cpos, split_clusters);
3606	le64_add_cpu(&right_rec->e_blkno,
3607		     ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3608					      split_clusters));
3609	le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3610
3611	ocfs2_cleanup_merge(el, index);
3612
3613	ocfs2_journal_dirty(handle, bh);
3614	if (left_path) {
3615		ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3616
3617		/*
3618		 * In the situation that the right_rec is empty and the extent
3619		 * block is empty also,  ocfs2_complete_edge_insert can't handle
3620		 * it and we need to delete the right extent block.
3621		 */
3622		if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3623		    le16_to_cpu(el->l_next_free_rec) == 1) {
3624			/* extend credit for ocfs2_remove_rightmost_path */
3625			ret = ocfs2_extend_rotate_transaction(handle, 0,
3626					handle->h_buffer_credits,
3627					right_path);
3628			if (ret) {
3629				mlog_errno(ret);
3630				goto out;
3631			}
3632
3633			ret = ocfs2_remove_rightmost_path(handle, et,
3634							  right_path,
3635							  dealloc);
3636			if (ret) {
3637				mlog_errno(ret);
3638				goto out;
3639			}
3640
3641			/* Now the rightmost extent block has been deleted.
3642			 * So we use the new rightmost path.
3643			 */
3644			ocfs2_mv_path(right_path, left_path);
3645			left_path = NULL;
3646		} else
3647			ocfs2_complete_edge_insert(handle, left_path,
3648						   right_path, subtree_index);
3649	}
3650out:
3651	ocfs2_free_path(left_path);
3652	return ret;
3653}
3654
3655static int ocfs2_try_to_merge_extent(handle_t *handle,
3656				     struct ocfs2_extent_tree *et,
3657				     struct ocfs2_path *path,
3658				     int split_index,
3659				     struct ocfs2_extent_rec *split_rec,
3660				     struct ocfs2_cached_dealloc_ctxt *dealloc,
3661				     struct ocfs2_merge_ctxt *ctxt)
3662{
3663	int ret = 0;
3664	struct ocfs2_extent_list *el = path_leaf_el(path);
3665	struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3666
3667	BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3668
3669	if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3670		/* extend credit for ocfs2_remove_rightmost_path */
3671		ret = ocfs2_extend_rotate_transaction(handle, 0,
3672				handle->h_buffer_credits,
3673				path);
3674		if (ret) {
3675			mlog_errno(ret);
3676			goto out;
3677		}
3678		/*
3679		 * The merge code will need to create an empty
3680		 * extent to take the place of the newly
3681		 * emptied slot. Remove any pre-existing empty
3682		 * extents - having more than one in a leaf is
3683		 * illegal.
3684		 */
3685		ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3686		if (ret) {
3687			mlog_errno(ret);
3688			goto out;
3689		}
3690		split_index--;
3691		rec = &el->l_recs[split_index];
3692	}
3693
3694	if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3695		/*
3696		 * Left-right contig implies this.
3697		 */
3698		BUG_ON(!ctxt->c_split_covers_rec);
3699
3700		/*
3701		 * Since the leftright insert always covers the entire
3702		 * extent, this call will delete the insert record
3703		 * entirely, resulting in an empty extent record added to
3704		 * the extent block.
3705		 *
3706		 * Since the adding of an empty extent shifts
3707		 * everything back to the right, there's no need to
3708		 * update split_index here.
3709		 *
3710		 * When the split_index is zero, we need to merge it to the
3711		 * prevoius extent block. It is more efficient and easier
3712		 * if we do merge_right first and merge_left later.
3713		 */
3714		ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3715					    split_index);
3716		if (ret) {
3717			mlog_errno(ret);
3718			goto out;
3719		}
3720
3721		/*
3722		 * We can only get this from logic error above.
3723		 */
3724		BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3725
3726		/* extend credit for ocfs2_remove_rightmost_path */
3727		ret = ocfs2_extend_rotate_transaction(handle, 0,
3728					handle->h_buffer_credits,
3729					path);
3730		if (ret) {
3731			mlog_errno(ret);
3732			goto out;
3733		}
3734
3735		/* The merge left us with an empty extent, remove it. */
3736		ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3737		if (ret) {
3738			mlog_errno(ret);
3739			goto out;
3740		}
3741
3742		rec = &el->l_recs[split_index];
3743
3744		/*
3745		 * Note that we don't pass split_rec here on purpose -
3746		 * we've merged it into the rec already.
3747		 */
3748		ret = ocfs2_merge_rec_left(path, handle, et, rec,
3749					   dealloc, split_index);
3750
3751		if (ret) {
3752			mlog_errno(ret);
3753			goto out;
3754		}
3755
3756		/* extend credit for ocfs2_remove_rightmost_path */
3757		ret = ocfs2_extend_rotate_transaction(handle, 0,
3758				handle->h_buffer_credits,
3759				path);
3760		if (ret) {
3761			mlog_errno(ret);
3762			goto out;
3763		}
3764
3765		ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3766		/*
3767		 * Error from this last rotate is not critical, so
3768		 * print but don't bubble it up.
3769		 */
3770		if (ret)
3771			mlog_errno(ret);
3772		ret = 0;
3773	} else {
3774		/*
3775		 * Merge a record to the left or right.
3776		 *
3777		 * 'contig_type' is relative to the existing record,
3778		 * so for example, if we're "right contig", it's to
3779		 * the record on the left (hence the left merge).
3780		 */
3781		if (ctxt->c_contig_type == CONTIG_RIGHT) {
3782			ret = ocfs2_merge_rec_left(path, handle, et,
3783						   split_rec, dealloc,
3784						   split_index);
3785			if (ret) {
3786				mlog_errno(ret);
3787				goto out;
3788			}
3789		} else {
3790			ret = ocfs2_merge_rec_right(path, handle,
3791						    et, split_rec,
3792						    split_index);
3793			if (ret) {
3794				mlog_errno(ret);
3795				goto out;
3796			}
3797		}
3798
3799		if (ctxt->c_split_covers_rec) {
3800			/* extend credit for ocfs2_remove_rightmost_path */
3801			ret = ocfs2_extend_rotate_transaction(handle, 0,
3802					handle->h_buffer_credits,
3803					path);
3804			if (ret) {
3805				mlog_errno(ret);
3806				ret = 0;
3807				goto out;
3808			}
3809
3810			/*
3811			 * The merge may have left an empty extent in
3812			 * our leaf. Try to rotate it away.
3813			 */
3814			ret = ocfs2_rotate_tree_left(handle, et, path,
3815						     dealloc);
3816			if (ret)
3817				mlog_errno(ret);
3818			ret = 0;
3819		}
3820	}
3821
3822out:
3823	return ret;
3824}
3825
3826static void ocfs2_subtract_from_rec(struct super_block *sb,
3827				    enum ocfs2_split_type split,
3828				    struct ocfs2_extent_rec *rec,
3829				    struct ocfs2_extent_rec *split_rec)
3830{
3831	u64 len_blocks;
3832
3833	len_blocks = ocfs2_clusters_to_blocks(sb,
3834				le16_to_cpu(split_rec->e_leaf_clusters));
3835
3836	if (split == SPLIT_LEFT) {
3837		/*
3838		 * Region is on the left edge of the existing
3839		 * record.
3840		 */
3841		le32_add_cpu(&rec->e_cpos,
3842			     le16_to_cpu(split_rec->e_leaf_clusters));
3843		le64_add_cpu(&rec->e_blkno, len_blocks);
3844		le16_add_cpu(&rec->e_leaf_clusters,
3845			     -le16_to_cpu(split_rec->e_leaf_clusters));
3846	} else {
3847		/*
3848		 * Region is on the right edge of the existing
3849		 * record.
3850		 */
3851		le16_add_cpu(&rec->e_leaf_clusters,
3852			     -le16_to_cpu(split_rec->e_leaf_clusters));
3853	}
3854}
3855
3856/*
3857 * Do the final bits of extent record insertion at the target leaf
3858 * list. If this leaf is part of an allocation tree, it is assumed
3859 * that the tree above has been prepared.
3860 */
3861static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3862				 struct ocfs2_extent_rec *insert_rec,
3863				 struct ocfs2_extent_list *el,
3864				 struct ocfs2_insert_type *insert)
3865{
3866	int i = insert->ins_contig_index;
3867	unsigned int range;
3868	struct ocfs2_extent_rec *rec;
3869
3870	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3871
3872	if (insert->ins_split != SPLIT_NONE) {
3873		i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3874		BUG_ON(i == -1);
3875		rec = &el->l_recs[i];
3876		ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3877					insert->ins_split, rec,
3878					insert_rec);
3879		goto rotate;
3880	}
3881
3882	/*
3883	 * Contiguous insert - either left or right.
3884	 */
3885	if (insert->ins_contig != CONTIG_NONE) {
3886		rec = &el->l_recs[i];
3887		if (insert->ins_contig == CONTIG_LEFT) {
3888			rec->e_blkno = insert_rec->e_blkno;
3889			rec->e_cpos = insert_rec->e_cpos;
3890		}
3891		le16_add_cpu(&rec->e_leaf_clusters,
3892			     le16_to_cpu(insert_rec->e_leaf_clusters));
3893		return;
3894	}
3895
3896	/*
3897	 * Handle insert into an empty leaf.
3898	 */
3899	if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3900	    ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3901	     ocfs2_is_empty_extent(&el->l_recs[0]))) {
3902		el->l_recs[0] = *insert_rec;
3903		el->l_next_free_rec = cpu_to_le16(1);
3904		return;
3905	}
3906
3907	/*
3908	 * Appending insert.
3909	 */
3910	if (insert->ins_appending == APPEND_TAIL) {
3911		i = le16_to_cpu(el->l_next_free_rec) - 1;
3912		rec = &el->l_recs[i];
3913		range = le32_to_cpu(rec->e_cpos)
3914			+ le16_to_cpu(rec->e_leaf_clusters);
3915		BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3916
3917		mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3918				le16_to_cpu(el->l_count),
3919				"owner %llu, depth %u, count %u, next free %u, "
3920				"rec.cpos %u, rec.clusters %u, "
3921				"insert.cpos %u, insert.clusters %u\n",
3922				ocfs2_metadata_cache_owner(et->et_ci),
3923				le16_to_cpu(el->l_tree_depth),
3924				le16_to_cpu(el->l_count),
3925				le16_to_cpu(el->l_next_free_rec),
3926				le32_to_cpu(el->l_recs[i].e_cpos),
3927				le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3928				le32_to_cpu(insert_rec->e_cpos),
3929				le16_to_cpu(insert_rec->e_leaf_clusters));
3930		i++;
3931		el->l_recs[i] = *insert_rec;
3932		le16_add_cpu(&el->l_next_free_rec, 1);
3933		return;
3934	}
3935
3936rotate:
3937	/*
3938	 * Ok, we have to rotate.
3939	 *
3940	 * At this point, it is safe to assume that inserting into an
3941	 * empty leaf and appending to a leaf have both been handled
3942	 * above.
3943	 *
3944	 * This leaf needs to have space, either by the empty 1st
3945	 * extent record, or by virtue of an l_next_rec < l_count.
3946	 */
3947	ocfs2_rotate_leaf(el, insert_rec);
3948}
3949
3950static void ocfs2_adjust_rightmost_records(handle_t *handle,
3951					   struct ocfs2_extent_tree *et,
3952					   struct ocfs2_path *path,
3953					   struct ocfs2_extent_rec *insert_rec)
3954{
3955	int i, next_free;
3956	struct buffer_head *bh;
3957	struct ocfs2_extent_list *el;
3958	struct ocfs2_extent_rec *rec;
3959
3960	/*
3961	 * Update everything except the leaf block.
3962	 */
3963	for (i = 0; i < path->p_tree_depth; i++) {
3964		bh = path->p_node[i].bh;
3965		el = path->p_node[i].el;
3966
3967		next_free = le16_to_cpu(el->l_next_free_rec);
3968		if (next_free == 0) {
3969			ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3970				    "Owner %llu has a bad extent list\n",
3971				    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3972			return;
3973		}
3974
3975		rec = &el->l_recs[next_free - 1];
3976
3977		rec->e_int_clusters = insert_rec->e_cpos;
3978		le32_add_cpu(&rec->e_int_clusters,
3979			     le16_to_cpu(insert_rec->e_leaf_clusters));
3980		le32_add_cpu(&rec->e_int_clusters,
3981			     -le32_to_cpu(rec->e_cpos));
3982
3983		ocfs2_journal_dirty(handle, bh);
3984	}
3985}
3986
3987static int ocfs2_append_rec_to_path(handle_t *handle,
3988				    struct ocfs2_extent_tree *et,
3989				    struct ocfs2_extent_rec *insert_rec,
3990				    struct ocfs2_path *right_path,
3991				    struct ocfs2_path **ret_left_path)
3992{
3993	int ret, next_free;
3994	struct ocfs2_extent_list *el;
3995	struct ocfs2_path *left_path = NULL;
3996
3997	*ret_left_path = NULL;
3998
3999	/*
4000	 * This shouldn't happen for non-trees. The extent rec cluster
4001	 * count manipulation below only works for interior nodes.
4002	 */
4003	BUG_ON(right_path->p_tree_depth == 0);
4004
4005	/*
4006	 * If our appending insert is at the leftmost edge of a leaf,
4007	 * then we might need to update the rightmost records of the
4008	 * neighboring path.
4009	 */
4010	el = path_leaf_el(right_path);
4011	next_free = le16_to_cpu(el->l_next_free_rec);
4012	if (next_free == 0 ||
4013	    (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4014		u32 left_cpos;
4015
4016		ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4017						    right_path, &left_cpos);
4018		if (ret) {
4019			mlog_errno(ret);
4020			goto out;
4021		}
4022
4023		trace_ocfs2_append_rec_to_path(
4024			(unsigned long long)
4025			ocfs2_metadata_cache_owner(et->et_ci),
4026			le32_to_cpu(insert_rec->e_cpos),
4027			left_cpos);
4028
4029		/*
4030		 * No need to worry if the append is already in the
4031		 * leftmost leaf.
4032		 */
4033		if (left_cpos) {
4034			left_path = ocfs2_new_path_from_path(right_path);
4035			if (!left_path) {
4036				ret = -ENOMEM;
4037				mlog_errno(ret);
4038				goto out;
4039			}
4040
4041			ret = ocfs2_find_path(et->et_ci, left_path,
4042					      left_cpos);
4043			if (ret) {
4044				mlog_errno(ret);
4045				goto out;
4046			}
4047
4048			/*
4049			 * ocfs2_insert_path() will pass the left_path to the
4050			 * journal for us.
4051			 */
4052		}
4053	}
4054
4055	ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4056	if (ret) {
4057		mlog_errno(ret);
4058		goto out;
4059	}
4060
4061	ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4062
4063	*ret_left_path = left_path;
4064	ret = 0;
4065out:
4066	if (ret != 0)
4067		ocfs2_free_path(left_path);
4068
4069	return ret;
4070}
4071
4072static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4073			       struct ocfs2_path *left_path,
4074			       struct ocfs2_path *right_path,
4075			       struct ocfs2_extent_rec *split_rec,
4076			       enum ocfs2_split_type split)
4077{
4078	int index;
4079	u32 cpos = le32_to_cpu(split_rec->e_cpos);
4080	struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4081	struct ocfs2_extent_rec *rec, *tmprec;
4082
4083	right_el = path_leaf_el(right_path);
4084	if (left_path)
4085		left_el = path_leaf_el(left_path);
4086
4087	el = right_el;
4088	insert_el = right_el;
4089	index = ocfs2_search_extent_list(el, cpos);
4090	if (index != -1) {
4091		if (index == 0 && left_path) {
4092			BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4093
4094			/*
4095			 * This typically means that the record
4096			 * started in the left path but moved to the
4097			 * right as a result of rotation. We either
4098			 * move the existing record to the left, or we
4099			 * do the later insert there.
4100			 *
4101			 * In this case, the left path should always
4102			 * exist as the rotate code will have passed
4103			 * it back for a post-insert update.
4104			 */
4105
4106			if (split == SPLIT_LEFT) {
4107				/*
4108				 * It's a left split. Since we know
4109				 * that the rotate code gave us an
4110				 * empty extent in the left path, we
4111				 * can just do the insert there.
4112				 */
4113				insert_el = left_el;
4114			} else {
4115				/*
4116				 * Right split - we have to move the
4117				 * existing record over to the left
4118				 * leaf. The insert will be into the
4119				 * newly created empty extent in the
4120				 * right leaf.
4121				 */
4122				tmprec = &right_el->l_recs[index];
4123				ocfs2_rotate_leaf(left_el, tmprec);
4124				el = left_el;
4125
4126				memset(tmprec, 0, sizeof(*tmprec));
4127				index = ocfs2_search_extent_list(left_el, cpos);
4128				BUG_ON(index == -1);
4129			}
4130		}
4131	} else {
4132		BUG_ON(!left_path);
4133		BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4134		/*
4135		 * Left path is easy - we can just allow the insert to
4136		 * happen.
4137		 */
4138		el = left_el;
4139		insert_el = left_el;
4140		index = ocfs2_search_extent_list(el, cpos);
4141		BUG_ON(index == -1);
4142	}
4143
4144	rec = &el->l_recs[index];
4145	ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4146				split, rec, split_rec);
4147	ocfs2_rotate_leaf(insert_el, split_rec);
4148}
4149
4150/*
4151 * This function only does inserts on an allocation b-tree. For tree
4152 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4153 *
4154 * right_path is the path we want to do the actual insert
4155 * in. left_path should only be passed in if we need to update that
4156 * portion of the tree after an edge insert.
4157 */
4158static int ocfs2_insert_path(handle_t *handle,
4159			     struct ocfs2_extent_tree *et,
4160			     struct ocfs2_path *left_path,
4161			     struct ocfs2_path *right_path,
4162			     struct ocfs2_extent_rec *insert_rec,
4163			     struct ocfs2_insert_type *insert)
4164{
4165	int ret, subtree_index;
4166	struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4167
4168	if (left_path) {
4169		/*
4170		 * There's a chance that left_path got passed back to
4171		 * us without being accounted for in the
4172		 * journal. Extend our transaction here to be sure we
4173		 * can change those blocks.
4174		 */
4175		ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4176		if (ret < 0) {
4177			mlog_errno(ret);
4178			goto out;
4179		}
4180
4181		ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4182		if (ret < 0) {
4183			mlog_errno(ret);
4184			goto out;
4185		}
4186	}
4187
4188	/*
4189	 * Pass both paths to the journal. The majority of inserts
4190	 * will be touching all components anyway.
4191	 */
4192	ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4193	if (ret < 0) {
4194		mlog_errno(ret);
4195		goto out;
4196	}
4197
4198	if (insert->ins_split != SPLIT_NONE) {
4199		/*
4200		 * We could call ocfs2_insert_at_leaf() for some types
4201		 * of splits, but it's easier to just let one separate
4202		 * function sort it all out.
4203		 */
4204		ocfs2_split_record(et, left_path, right_path,
4205				   insert_rec, insert->ins_split);
4206
4207		/*
4208		 * Split might have modified either leaf and we don't
4209		 * have a guarantee that the later edge insert will
4210		 * dirty this for us.
4211		 */
4212		if (left_path)
4213			ocfs2_journal_dirty(handle,
4214					    path_leaf_bh(left_path));
4215	} else
4216		ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4217				     insert);
4218
4219	ocfs2_journal_dirty(handle, leaf_bh);
4220
4221	if (left_path) {
4222		/*
4223		 * The rotate code has indicated that we need to fix
4224		 * up portions of the tree after the insert.
4225		 *
4226		 * XXX: Should we extend the transaction here?
4227		 */
4228		subtree_index = ocfs2_find_subtree_root(et, left_path,
4229							right_path);
4230		ocfs2_complete_edge_insert(handle, left_path, right_path,
4231					   subtree_index);
4232	}
4233
4234	ret = 0;
4235out:
4236	return ret;
4237}
4238
4239static int ocfs2_do_insert_extent(handle_t *handle,
4240				  struct ocfs2_extent_tree *et,
4241				  struct ocfs2_extent_rec *insert_rec,
4242				  struct ocfs2_insert_type *type)
4243{
4244	int ret, rotate = 0;
4245	u32 cpos;
4246	struct ocfs2_path *right_path = NULL;
4247	struct ocfs2_path *left_path = NULL;
4248	struct ocfs2_extent_list *el;
4249
4250	el = et->et_root_el;
4251
4252	ret = ocfs2_et_root_journal_access(handle, et,
4253					   OCFS2_JOURNAL_ACCESS_WRITE);
4254	if (ret) {
4255		mlog_errno(ret);
4256		goto out;
4257	}
4258
4259	if (le16_to_cpu(el->l_tree_depth) == 0) {
4260		ocfs2_insert_at_leaf(et, insert_rec, el, type);
4261		goto out_update_clusters;
4262	}
4263
4264	right_path = ocfs2_new_path_from_et(et);
4265	if (!right_path) {
4266		ret = -ENOMEM;
4267		mlog_errno(ret);
4268		goto out;
4269	}
4270
4271	/*
4272	 * Determine the path to start with. Rotations need the
4273	 * rightmost path, everything else can go directly to the
4274	 * target leaf.
4275	 */
4276	cpos = le32_to_cpu(insert_rec->e_cpos);
4277	if (type->ins_appending == APPEND_NONE &&
4278	    type->ins_contig == CONTIG_NONE) {
4279		rotate = 1;
4280		cpos = UINT_MAX;
4281	}
4282
4283	ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4284	if (ret) {
4285		mlog_errno(ret);
4286		goto out;
4287	}
4288
4289	/*
4290	 * Rotations and appends need special treatment - they modify
4291	 * parts of the tree's above them.
4292	 *
4293	 * Both might pass back a path immediate to the left of the
4294	 * one being inserted to. This will be cause
4295	 * ocfs2_insert_path() to modify the rightmost records of
4296	 * left_path to account for an edge insert.
4297	 *
4298	 * XXX: When modifying this code, keep in mind that an insert
4299	 * can wind up skipping both of these two special cases...
4300	 */
4301	if (rotate) {
4302		ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4303					      le32_to_cpu(insert_rec->e_cpos),
4304					      right_path, &left_path);
4305		if (ret) {
4306			mlog_errno(ret);
4307			goto out;
4308		}
4309
4310		/*
4311		 * ocfs2_rotate_tree_right() might have extended the
4312		 * transaction without re-journaling our tree root.
4313		 */
4314		ret = ocfs2_et_root_journal_access(handle, et,
4315						   OCFS2_JOURNAL_ACCESS_WRITE);
4316		if (ret) {
4317			mlog_errno(ret);
4318			goto out;
4319		}
4320	} else if (type->ins_appending == APPEND_TAIL
4321		   && type->ins_contig != CONTIG_LEFT) {
4322		ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4323					       right_path, &left_path);
4324		if (ret) {
4325			mlog_errno(ret);
4326			goto out;
4327		}
4328	}
4329
4330	ret = ocfs2_insert_path(handle, et, left_path, right_path,
4331				insert_rec, type);
4332	if (ret) {
4333		mlog_errno(ret);
4334		goto out;
4335	}
4336
4337out_update_clusters:
4338	if (type->ins_split == SPLIT_NONE)
4339		ocfs2_et_update_clusters(et,
4340					 le16_to_cpu(insert_rec->e_leaf_clusters));
4341
4342	ocfs2_journal_dirty(handle, et->et_root_bh);
4343
4344out:
4345	ocfs2_free_path(left_path);
4346	ocfs2_free_path(right_path);
4347
4348	return ret;
4349}
4350
4351static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4352			       struct ocfs2_path *path,
4353			       struct ocfs2_extent_list *el, int index,
4354			       struct ocfs2_extent_rec *split_rec,
4355			       struct ocfs2_merge_ctxt *ctxt)
4356{
4357	int status = 0;
4358	enum ocfs2_contig_type ret = CONTIG_NONE;
4359	u32 left_cpos, right_cpos;
4360	struct ocfs2_extent_rec *rec = NULL;
4361	struct ocfs2_extent_list *new_el;
4362	struct ocfs2_path *left_path = NULL, *right_path = NULL;
4363	struct buffer_head *bh;
4364	struct ocfs2_extent_block *eb;
4365	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4366
4367	if (index > 0) {
4368		rec = &el->l_recs[index - 1];
4369	} else if (path->p_tree_depth > 0) {
4370		status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4371		if (status)
4372			goto exit;
4373
4374		if (left_cpos != 0) {
4375			left_path = ocfs2_new_path_from_path(path);
4376			if (!left_path) {
4377				status = -ENOMEM;
4378				mlog_errno(status);
4379				goto exit;
4380			}
4381
4382			status = ocfs2_find_path(et->et_ci, left_path,
4383						 left_cpos);
4384			if (status)
4385				goto free_left_path;
4386
4387			new_el = path_leaf_el(left_path);
4388
4389			if (le16_to_cpu(new_el->l_next_free_rec) !=
4390			    le16_to_cpu(new_el->l_count)) {
4391				bh = path_leaf_bh(left_path);
4392				eb = (struct ocfs2_extent_block *)bh->b_data;
4393				status = ocfs2_error(sb,
4394						"Extent block #%llu has an invalid l_next_free_rec of %d.  It should have matched the l_count of %d\n",
4395						(unsigned long long)le64_to_cpu(eb->h_blkno),
4396						le16_to_cpu(new_el->l_next_free_rec),
4397						le16_to_cpu(new_el->l_count));
4398				goto free_left_path;
4399			}
4400			rec = &new_el->l_recs[
4401				le16_to_cpu(new_el->l_next_free_rec) - 1];
4402		}
4403	}
4404
4405	/*
4406	 * We're careful to check for an empty extent record here -
4407	 * the merge code will know what to do if it sees one.
4408	 */
4409	if (rec) {
4410		if (index == 1 && ocfs2_is_empty_extent(rec)) {
4411			if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4412				ret = CONTIG_RIGHT;
4413		} else {
4414			ret = ocfs2_et_extent_contig(et, rec, split_rec);
4415		}
4416	}
4417
4418	rec = NULL;
4419	if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4420		rec = &el->l_recs[index + 1];
4421	else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4422		 path->p_tree_depth > 0) {
4423		status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4424		if (status)
4425			goto free_left_path;
4426
4427		if (right_cpos == 0)
4428			goto free_left_path;
4429
4430		right_path = ocfs2_new_path_from_path(path);
4431		if (!right_path) {
4432			status = -ENOMEM;
4433			mlog_errno(status);
4434			goto free_left_path;
4435		}
4436
4437		status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4438		if (status)
4439			goto free_right_path;
4440
4441		new_el = path_leaf_el(right_path);
4442		rec = &new_el->l_recs[0];
4443		if (ocfs2_is_empty_extent(rec)) {
4444			if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4445				bh = path_leaf_bh(right_path);
4446				eb = (struct ocfs2_extent_block *)bh->b_data;
4447				status = ocfs2_error(sb,
4448						"Extent block #%llu has an invalid l_next_free_rec of %d\n",
4449						(unsigned long long)le64_to_cpu(eb->h_blkno),
4450						le16_to_cpu(new_el->l_next_free_rec));
4451				goto free_right_path;
4452			}
4453			rec = &new_el->l_recs[1];
4454		}
4455	}
4456
4457	if (rec) {
4458		enum ocfs2_contig_type contig_type;
4459
4460		contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4461
4462		if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4463			ret = CONTIG_LEFTRIGHT;
4464		else if (ret == CONTIG_NONE)
4465			ret = contig_type;
4466	}
4467
4468free_right_path:
4469	ocfs2_free_path(right_path);
4470free_left_path:
4471	ocfs2_free_path(left_path);
4472exit:
4473	if (status == 0)
4474		ctxt->c_contig_type = ret;
4475
4476	return status;
4477}
4478
4479static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4480				     struct ocfs2_insert_type *insert,
4481				     struct ocfs2_extent_list *el,
4482				     struct ocfs2_extent_rec *insert_rec)
4483{
4484	int i;
4485	enum ocfs2_contig_type contig_type = CONTIG_NONE;
4486
4487	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4488
4489	for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4490		contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4491						     insert_rec);
4492		if (contig_type != CONTIG_NONE) {
4493			insert->ins_contig_index = i;
4494			break;
4495		}
4496	}
4497	insert->ins_contig = contig_type;
4498
4499	if (insert->ins_contig != CONTIG_NONE) {
4500		struct ocfs2_extent_rec *rec =
4501				&el->l_recs[insert->ins_contig_index];
4502		unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4503				   le16_to_cpu(insert_rec->e_leaf_clusters);
4504
4505		/*
4506		 * Caller might want us to limit the size of extents, don't
4507		 * calculate contiguousness if we might exceed that limit.
4508		 */
4509		if (et->et_max_leaf_clusters &&
4510		    (len > et->et_max_leaf_clusters))
4511			insert->ins_contig = CONTIG_NONE;
4512	}
4513}
4514
4515/*
4516 * This should only be called against the righmost leaf extent list.
4517 *
4518 * ocfs2_figure_appending_type() will figure out whether we'll have to
4519 * insert at the tail of the rightmost leaf.
4520 *
4521 * This should also work against the root extent list for tree's with 0
4522 * depth. If we consider the root extent list to be the rightmost leaf node
4523 * then the logic here makes sense.
4524 */
4525static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4526					struct ocfs2_extent_list *el,
4527					struct ocfs2_extent_rec *insert_rec)
4528{
4529	int i;
4530	u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4531	struct ocfs2_extent_rec *rec;
4532
4533	insert->ins_appending = APPEND_NONE;
4534
4535	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4536
4537	if (!el->l_next_free_rec)
4538		goto set_tail_append;
4539
4540	if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4541		/* Were all records empty? */
4542		if (le16_to_cpu(el->l_next_free_rec) == 1)
4543			goto set_tail_append;
4544	}
4545
4546	i = le16_to_cpu(el->l_next_free_rec) - 1;
4547	rec = &el->l_recs[i];
4548
4549	if (cpos >=
4550	    (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4551		goto set_tail_append;
4552
4553	return;
4554
4555set_tail_append:
4556	insert->ins_appending = APPEND_TAIL;
4557}
4558
4559/*
4560 * Helper function called at the beginning of an insert.
4561 *
4562 * This computes a few things that are commonly used in the process of
4563 * inserting into the btree:
4564 *   - Whether the new extent is contiguous with an existing one.
4565 *   - The current tree depth.
4566 *   - Whether the insert is an appending one.
4567 *   - The total # of free records in the tree.
4568 *
4569 * All of the information is stored on the ocfs2_insert_type
4570 * structure.
4571 */
4572static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4573				    struct buffer_head **last_eb_bh,
4574				    struct ocfs2_extent_rec *insert_rec,
4575				    int *free_records,
4576				    struct ocfs2_insert_type *insert)
4577{
4578	int ret;
4579	struct ocfs2_extent_block *eb;
4580	struct ocfs2_extent_list *el;
4581	struct ocfs2_path *path = NULL;
4582	struct buffer_head *bh = NULL;
4583
4584	insert->ins_split = SPLIT_NONE;
4585
4586	el = et->et_root_el;
4587	insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4588
4589	if (el->l_tree_depth) {
4590		/*
4591		 * If we have tree depth, we read in the
4592		 * rightmost extent block ahead of time as
4593		 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4594		 * may want it later.
4595		 */
4596		ret = ocfs2_read_extent_block(et->et_ci,
4597					      ocfs2_et_get_last_eb_blk(et),
4598					      &bh);
4599		if (ret) {
4600			mlog_errno(ret);
4601			goto out;
4602		}
4603		eb = (struct ocfs2_extent_block *) bh->b_data;
4604		el = &eb->h_list;
4605	}
4606
4607	/*
4608	 * Unless we have a contiguous insert, we'll need to know if
4609	 * there is room left in our allocation tree for another
4610	 * extent record.
4611	 *
4612	 * XXX: This test is simplistic, we can search for empty
4613	 * extent records too.
4614	 */
4615	*free_records = le16_to_cpu(el->l_count) -
4616		le16_to_cpu(el->l_next_free_rec);
4617
4618	if (!insert->ins_tree_depth) {
4619		ocfs2_figure_contig_type(et, insert, el, insert_rec);
4620		ocfs2_figure_appending_type(insert, el, insert_rec);
4621		return 0;
4622	}
4623
4624	path = ocfs2_new_path_from_et(et);
4625	if (!path) {
4626		ret = -ENOMEM;
4627		mlog_errno(ret);
4628		goto out;
4629	}
4630
4631	/*
4632	 * In the case that we're inserting past what the tree
4633	 * currently accounts for, ocfs2_find_path() will return for
4634	 * us the rightmost tree path. This is accounted for below in
4635	 * the appending code.
4636	 */
4637	ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4638	if (ret) {
4639		mlog_errno(ret);
4640		goto out;
4641	}
4642
4643	el = path_leaf_el(path);
4644
4645	/*
4646	 * Now that we have the path, there's two things we want to determine:
4647	 * 1) Contiguousness (also set contig_index if this is so)
4648	 *
4649	 * 2) Are we doing an append? We can trivially break this up
4650         *     into two types of appends: simple record append, or a
4651         *     rotate inside the tail leaf.
4652	 */
4653	ocfs2_figure_contig_type(et, insert, el, insert_rec);
4654
4655	/*
4656	 * The insert code isn't quite ready to deal with all cases of
4657	 * left contiguousness. Specifically, if it's an insert into
4658	 * the 1st record in a leaf, it will require the adjustment of
4659	 * cluster count on the last record of the path directly to it's
4660	 * left. For now, just catch that case and fool the layers
4661	 * above us. This works just fine for tree_depth == 0, which
4662	 * is why we allow that above.
4663	 */
4664	if (insert->ins_contig == CONTIG_LEFT &&
4665	    insert->ins_contig_index == 0)
4666		insert->ins_contig = CONTIG_NONE;
4667
4668	/*
4669	 * Ok, so we can simply compare against last_eb to figure out
4670	 * whether the path doesn't exist. This will only happen in
4671	 * the case that we're doing a tail append, so maybe we can
4672	 * take advantage of that information somehow.
4673	 */
4674	if (ocfs2_et_get_last_eb_blk(et) ==
4675	    path_leaf_bh(path)->b_blocknr) {
4676		/*
4677		 * Ok, ocfs2_find_path() returned us the rightmost
4678		 * tree path. This might be an appending insert. There are
4679		 * two cases:
4680		 *    1) We're doing a true append at the tail:
4681		 *	-This might even be off the end of the leaf
4682		 *    2) We're "appending" by rotating in the tail
4683		 */
4684		ocfs2_figure_appending_type(insert, el, insert_rec);
4685	}
4686
4687out:
4688	ocfs2_free_path(path);
4689
4690	if (ret == 0)
4691		*last_eb_bh = bh;
4692	else
4693		brelse(bh);
4694	return ret;
4695}
4696
4697/*
4698 * Insert an extent into a btree.
4699 *
4700 * The caller needs to update the owning btree's cluster count.
4701 */
4702int ocfs2_insert_extent(handle_t *handle,
4703			struct ocfs2_extent_tree *et,
4704			u32 cpos,
4705			u64 start_blk,
4706			u32 new_clusters,
4707			u8 flags,
4708			struct ocfs2_alloc_context *meta_ac)
4709{
4710	int status;
4711	int uninitialized_var(free_records);
4712	struct buffer_head *last_eb_bh = NULL;
4713	struct ocfs2_insert_type insert = {0, };
4714	struct ocfs2_extent_rec rec;
4715
4716	trace_ocfs2_insert_extent_start(
4717		(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4718		cpos, new_clusters);
4719
4720	memset(&rec, 0, sizeof(rec));
4721	rec.e_cpos = cpu_to_le32(cpos);
4722	rec.e_blkno = cpu_to_le64(start_blk);
4723	rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4724	rec.e_flags = flags;
4725	status = ocfs2_et_insert_check(et, &rec);
4726	if (status) {
4727		mlog_errno(status);
4728		goto bail;
4729	}
4730
4731	status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4732					  &free_records, &insert);
4733	if (status < 0) {
4734		mlog_errno(status);
4735		goto bail;
4736	}
4737
4738	trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4739				  insert.ins_contig_index, free_records,
4740				  insert.ins_tree_depth);
4741
4742	if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4743		status = ocfs2_grow_tree(handle, et,
4744					 &insert.ins_tree_depth, &last_eb_bh,
4745					 meta_ac);
4746		if (status) {
4747			mlog_errno(status);
4748			goto bail;
4749		}
4750	}
4751
4752	/* Finally, we can add clusters. This might rotate the tree for us. */
4753	status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4754	if (status < 0)
4755		mlog_errno(status);
4756	else
4757		ocfs2_et_extent_map_insert(et, &rec);
4758
4759bail:
4760	brelse(last_eb_bh);
4761
4762	return status;
4763}
4764
4765/*
4766 * Allcate and add clusters into the extent b-tree.
4767 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4768 * The extent b-tree's root is specified by et, and
4769 * it is not limited to the file storage. Any extent tree can use this
4770 * function if it implements the proper ocfs2_extent_tree.
4771 */
4772int ocfs2_add_clusters_in_btree(handle_t *handle,
4773				struct ocfs2_extent_tree *et,
4774				u32 *logical_offset,
4775				u32 clusters_to_add,
4776				int mark_unwritten,
4777				struct ocfs2_alloc_context *data_ac,
4778				struct ocfs2_alloc_context *meta_ac,
4779				enum ocfs2_alloc_restarted *reason_ret)
4780{
4781	int status = 0, err = 0;
4782	int need_free = 0;
4783	int free_extents;
4784	enum ocfs2_alloc_restarted reason = RESTART_NONE;
4785	u32 bit_off, num_bits;
4786	u64 block;
4787	u8 flags = 0;
4788	struct ocfs2_super *osb =
4789		OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4790
4791	BUG_ON(!clusters_to_add);
4792
4793	if (mark_unwritten)
4794		flags = OCFS2_EXT_UNWRITTEN;
4795
4796	free_extents = ocfs2_num_free_extents(et);
4797	if (free_extents < 0) {
4798		status = free_extents;
4799		mlog_errno(status);
4800		goto leave;
4801	}
4802
4803	/* there are two cases which could cause us to EAGAIN in the
4804	 * we-need-more-metadata case:
4805	 * 1) we haven't reserved *any*
4806	 * 2) we are so fragmented, we've needed to add metadata too
4807	 *    many times. */
4808	if (!free_extents && !meta_ac) {
4809		err = -1;
4810		status = -EAGAIN;
4811		reason = RESTART_META;
4812		goto leave;
4813	} else if ((!free_extents)
4814		   && (ocfs2_alloc_context_bits_left(meta_ac)
4815		       < ocfs2_extend_meta_needed(et->et_root_el))) {
4816		err = -2;
4817		status = -EAGAIN;
4818		reason = RESTART_META;
4819		goto leave;
4820	}
4821
4822	status = __ocfs2_claim_clusters(handle, data_ac, 1,
4823					clusters_to_add, &bit_off, &num_bits);
4824	if (status < 0) {
4825		if (status != -ENOSPC)
4826			mlog_errno(status);
4827		goto leave;
4828	}
4829
4830	BUG_ON(num_bits > clusters_to_add);
4831
4832	/* reserve our write early -- insert_extent may update the tree root */
4833	status = ocfs2_et_root_journal_access(handle, et,
4834					      OCFS2_JOURNAL_ACCESS_WRITE);
4835	if (status < 0) {
4836		mlog_errno(status);
4837		need_free = 1;
4838		goto bail;
4839	}
4840
4841	block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4842	trace_ocfs2_add_clusters_in_btree(
4843	     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4844	     bit_off, num_bits);
4845	status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4846				     num_bits, flags, meta_ac);
4847	if (status < 0) {
4848		mlog_errno(status);
4849		need_free = 1;
4850		goto bail;
4851	}
4852
4853	ocfs2_journal_dirty(handle, et->et_root_bh);
4854
4855	clusters_to_add -= num_bits;
4856	*logical_offset += num_bits;
4857
4858	if (clusters_to_add) {
4859		err = clusters_to_add;
4860		status = -EAGAIN;
4861		reason = RESTART_TRANS;
4862	}
4863
4864bail:
4865	if (need_free) {
4866		if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4867			ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4868					bit_off, num_bits);
4869		else
4870			ocfs2_free_clusters(handle,
4871					data_ac->ac_inode,
4872					data_ac->ac_bh,
4873					ocfs2_clusters_to_blocks(osb->sb, bit_off),
4874					num_bits);
4875	}
4876
4877leave:
4878	if (reason_ret)
4879		*reason_ret = reason;
4880	trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4881	return status;
4882}
4883
4884static void ocfs2_make_right_split_rec(struct super_block *sb,
4885				       struct ocfs2_extent_rec *split_rec,
4886				       u32 cpos,
4887				       struct ocfs2_extent_rec *rec)
4888{
4889	u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4890	u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4891
4892	memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4893
4894	split_rec->e_cpos = cpu_to_le32(cpos);
4895	split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4896
4897	split_rec->e_blkno = rec->e_blkno;
4898	le64_add_cpu(&split_rec->e_blkno,
4899		     ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4900
4901	split_rec->e_flags = rec->e_flags;
4902}
4903
4904static int ocfs2_split_and_insert(handle_t *handle,
4905				  struct ocfs2_extent_tree *et,
4906				  struct ocfs2_path *path,
4907				  struct buffer_head **last_eb_bh,
4908				  int split_index,
4909				  struct ocfs2_extent_rec *orig_split_rec,
4910				  struct ocfs2_alloc_context *meta_ac)
4911{
4912	int ret = 0, depth;
4913	unsigned int insert_range, rec_range, do_leftright = 0;
4914	struct ocfs2_extent_rec tmprec;
4915	struct ocfs2_extent_list *rightmost_el;
4916	struct ocfs2_extent_rec rec;
4917	struct ocfs2_extent_rec split_rec = *orig_split_rec;
4918	struct ocfs2_insert_type insert;
4919	struct ocfs2_extent_block *eb;
4920
4921leftright:
4922	/*
4923	 * Store a copy of the record on the stack - it might move
4924	 * around as the tree is manipulated below.
4925	 */
4926	rec = path_leaf_el(path)->l_recs[split_index];
4927
4928	rightmost_el = et->et_root_el;
4929
4930	depth = le16_to_cpu(rightmost_el->l_tree_depth);
4931	if (depth) {
4932		BUG_ON(!(*last_eb_bh));
4933		eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4934		rightmost_el = &eb->h_list;
4935	}
4936
4937	if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4938	    le16_to_cpu(rightmost_el->l_count)) {
4939		ret = ocfs2_grow_tree(handle, et,
4940				      &depth, last_eb_bh, meta_ac);
4941		if (ret) {
4942			mlog_errno(ret);
4943			goto out;
4944		}
4945	}
4946
4947	memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4948	insert.ins_appending = APPEND_NONE;
4949	insert.ins_contig = CONTIG_NONE;
4950	insert.ins_tree_depth = depth;
4951
4952	insert_range = le32_to_cpu(split_rec.e_cpos) +
4953		le16_to_cpu(split_rec.e_leaf_clusters);
4954	rec_range = le32_to_cpu(rec.e_cpos) +
4955		le16_to_cpu(rec.e_leaf_clusters);
4956
4957	if (split_rec.e_cpos == rec.e_cpos) {
4958		insert.ins_split = SPLIT_LEFT;
4959	} else if (insert_range == rec_range) {
4960		insert.ins_split = SPLIT_RIGHT;
4961	} else {
4962		/*
4963		 * Left/right split. We fake this as a right split
4964		 * first and then make a second pass as a left split.
4965		 */
4966		insert.ins_split = SPLIT_RIGHT;
4967
4968		ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4969					   &tmprec, insert_range, &rec);
4970
4971		split_rec = tmprec;
4972
4973		BUG_ON(do_leftright);
4974		do_leftright = 1;
4975	}
4976
4977	ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4978	if (ret) {
4979		mlog_errno(ret);
4980		goto out;
4981	}
4982
4983	if (do_leftright == 1) {
4984		u32 cpos;
4985		struct ocfs2_extent_list *el;
4986
4987		do_leftright++;
4988		split_rec = *orig_split_rec;
4989
4990		ocfs2_reinit_path(path, 1);
4991
4992		cpos = le32_to_cpu(split_rec.e_cpos);
4993		ret = ocfs2_find_path(et->et_ci, path, cpos);
4994		if (ret) {
4995			mlog_errno(ret);
4996			goto out;
4997		}
4998
4999		el = path_leaf_el(path);
5000		split_index = ocfs2_search_extent_list(el, cpos);
5001		if (split_index == -1) {
5002			ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5003				    "Owner %llu has an extent at cpos %u which can no longer be found\n",
5004				    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5005				    cpos);
5006			ret = -EROFS;
5007			goto out;
5008		}
5009		goto leftright;
5010	}
5011out:
5012
5013	return ret;
5014}
5015
5016static int ocfs2_replace_extent_rec(handle_t *handle,
5017				    struct ocfs2_extent_tree *et,
5018				    struct ocfs2_path *path,
5019				    struct ocfs2_extent_list *el,
5020				    int split_index,
5021				    struct ocfs2_extent_rec *split_rec)
5022{
5023	int ret;
5024
5025	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5026					   path_num_items(path) - 1);
5027	if (ret) {
5028		mlog_errno(ret);
5029		goto out;
5030	}
5031
5032	el->l_recs[split_index] = *split_rec;
5033
5034	ocfs2_journal_dirty(handle, path_leaf_bh(path));
5035out:
5036	return ret;
5037}
5038
5039/*
5040 * Split part or all of the extent record at split_index in the leaf
5041 * pointed to by path. Merge with the contiguous extent record if needed.
5042 *
5043 * Care is taken to handle contiguousness so as to not grow the tree.
5044 *
5045 * meta_ac is not strictly necessary - we only truly need it if growth
5046 * of the tree is required. All other cases will degrade into a less
5047 * optimal tree layout.
5048 *
5049 * last_eb_bh should be the rightmost leaf block for any extent
5050 * btree. Since a split may grow the tree or a merge might shrink it,
5051 * the caller cannot trust the contents of that buffer after this call.
5052 *
5053 * This code is optimized for readability - several passes might be
5054 * made over certain portions of the tree. All of those blocks will
5055 * have been brought into cache (and pinned via the journal), so the
5056 * extra overhead is not expressed in terms of disk reads.
5057 */
5058int ocfs2_split_extent(handle_t *handle,
5059		       struct ocfs2_extent_tree *et,
5060		       struct ocfs2_path *path,
5061		       int split_index,
5062		       struct ocfs2_extent_rec *split_rec,
5063		       struct ocfs2_alloc_context *meta_ac,
5064		       struct ocfs2_cached_dealloc_ctxt *dealloc)
5065{
5066	int ret = 0;
5067	struct ocfs2_extent_list *el = path_leaf_el(path);
5068	struct buffer_head *last_eb_bh = NULL;
5069	struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5070	struct ocfs2_merge_ctxt ctxt;
5071
5072	if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5073	    ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5074	     (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5075		ret = -EIO;
5076		mlog_errno(ret);
5077		goto out;
5078	}
5079
5080	ret = ocfs2_figure_merge_contig_type(et, path, el,
5081					     split_index,
5082					     split_rec,
5083					     &ctxt);
5084	if (ret) {
5085		mlog_errno(ret);
5086		goto out;
5087	}
5088
5089	/*
5090	 * The core merge / split code wants to know how much room is
5091	 * left in this allocation tree, so we pass the
5092	 * rightmost extent list.
5093	 */
5094	if (path->p_tree_depth) {
5095		ret = ocfs2_read_extent_block(et->et_ci,
5096					      ocfs2_et_get_last_eb_blk(et),
5097					      &last_eb_bh);
5098		if (ret) {
5099			mlog_errno(ret);
5100			goto out;
5101		}
5102	}
5103
5104	if (rec->e_cpos == split_rec->e_cpos &&
5105	    rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5106		ctxt.c_split_covers_rec = 1;
5107	else
5108		ctxt.c_split_covers_rec = 0;
5109
5110	ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5111
5112	trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5113				 ctxt.c_has_empty_extent,
5114				 ctxt.c_split_covers_rec);
5115
5116	if (ctxt.c_contig_type == CONTIG_NONE) {
5117		if (ctxt.c_split_covers_rec)
5118			ret = ocfs2_replace_extent_rec(handle, et, path, el,
5119						       split_index, split_rec);
5120		else
5121			ret = ocfs2_split_and_insert(handle, et, path,
5122						     &last_eb_bh, split_index,
5123						     split_rec, meta_ac);
5124		if (ret)
5125			mlog_errno(ret);
5126	} else {
5127		ret = ocfs2_try_to_merge_extent(handle, et, path,
5128						split_index, split_rec,
5129						dealloc, &ctxt);
5130		if (ret)
5131			mlog_errno(ret);
5132	}
5133
5134out:
5135	brelse(last_eb_bh);
5136	return ret;
5137}
5138
5139/*
5140 * Change the flags of the already-existing extent at cpos for len clusters.
5141 *
5142 * new_flags: the flags we want to set.
5143 * clear_flags: the flags we want to clear.
5144 * phys: the new physical offset we want this new extent starts from.
5145 *
5146 * If the existing extent is larger than the request, initiate a
5147 * split. An attempt will be made at merging with adjacent extents.
5148 *
5149 * The caller is responsible for passing down meta_ac if we'll need it.
5150 */
5151int ocfs2_change_extent_flag(handle_t *handle,
5152			     struct ocfs2_extent_tree *et,
5153			     u32 cpos, u32 len, u32 phys,
5154			     struct ocfs2_alloc_context *meta_ac,
5155			     struct ocfs2_cached_dealloc_ctxt *dealloc,
5156			     int new_flags, int clear_flags)
5157{
5158	int ret, index;
5159	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5160	u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5161	struct ocfs2_extent_rec split_rec;
5162	struct ocfs2_path *left_path = NULL;
5163	struct ocfs2_extent_list *el;
5164	struct ocfs2_extent_rec *rec;
5165
5166	left_path = ocfs2_new_path_from_et(et);
5167	if (!left_path) {
5168		ret = -ENOMEM;
5169		mlog_errno(ret);
5170		goto out;
5171	}
5172
5173	ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5174	if (ret) {
5175		mlog_errno(ret);
5176		goto out;
5177	}
5178	el = path_leaf_el(left_path);
5179
5180	index = ocfs2_search_extent_list(el, cpos);
5181	if (index == -1) {
5182		ocfs2_error(sb,
5183			    "Owner %llu has an extent at cpos %u which can no longer be found\n",
5184			    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5185			    cpos);
5186		ret = -EROFS;
5187		goto out;
5188	}
5189
5190	ret = -EIO;
5191	rec = &el->l_recs[index];
5192	if (new_flags && (rec->e_flags & new_flags)) {
5193		mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5194		     "extent that already had them\n",
5195		     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5196		     new_flags);
5197		goto out;
5198	}
5199
5200	if (clear_flags && !(rec->e_flags & clear_flags)) {
5201		mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5202		     "extent that didn't have them\n",
5203		     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5204		     clear_flags);
5205		goto out;
5206	}
5207
5208	memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5209	split_rec.e_cpos = cpu_to_le32(cpos);
5210	split_rec.e_leaf_clusters = cpu_to_le16(len);
5211	split_rec.e_blkno = cpu_to_le64(start_blkno);
5212	split_rec.e_flags = rec->e_flags;
5213	if (new_flags)
5214		split_rec.e_flags |= new_flags;
5215	if (clear_flags)
5216		split_rec.e_flags &= ~clear_flags;
5217
5218	ret = ocfs2_split_extent(handle, et, left_path,
5219				 index, &split_rec, meta_ac,
5220				 dealloc);
5221	if (ret)
5222		mlog_errno(ret);
5223
5224out:
5225	ocfs2_free_path(left_path);
5226	return ret;
5227
5228}
5229
5230/*
5231 * Mark the already-existing extent at cpos as written for len clusters.
5232 * This removes the unwritten extent flag.
5233 *
5234 * If the existing extent is larger than the request, initiate a
5235 * split. An attempt will be made at merging with adjacent extents.
5236 *
5237 * The caller is responsible for passing down meta_ac if we'll need it.
5238 */
5239int ocfs2_mark_extent_written(struct inode *inode,
5240			      struct ocfs2_extent_tree *et,
5241			      handle_t *handle, u32 cpos, u32 len, u32 phys,
5242			      struct ocfs2_alloc_context *meta_ac,
5243			      struct ocfs2_cached_dealloc_ctxt *dealloc)
5244{
5245	int ret;
5246
5247	trace_ocfs2_mark_extent_written(
5248		(unsigned long long)OCFS2_I(inode)->ip_blkno,
5249		cpos, len, phys);
5250
5251	if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5252		ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents that are being written to, but the feature bit is not set in the super block\n",
5253			    (unsigned long long)OCFS2_I(inode)->ip_blkno);
5254		ret = -EROFS;
5255		goto out;
5256	}
5257
5258	/*
5259	 * XXX: This should be fixed up so that we just re-insert the
5260	 * next extent records.
5261	 */
5262	ocfs2_et_extent_map_truncate(et, 0);
5263
5264	ret = ocfs2_change_extent_flag(handle, et, cpos,
5265				       len, phys, meta_ac, dealloc,
5266				       0, OCFS2_EXT_UNWRITTEN);
5267	if (ret)
5268		mlog_errno(ret);
5269
5270out:
5271	return ret;
5272}
5273
5274static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5275			    struct ocfs2_path *path,
5276			    int index, u32 new_range,
5277			    struct ocfs2_alloc_context *meta_ac)
5278{
5279	int ret, depth, credits;
5280	struct buffer_head *last_eb_bh = NULL;
5281	struct ocfs2_extent_block *eb;
5282	struct ocfs2_extent_list *rightmost_el, *el;
5283	struct ocfs2_extent_rec split_rec;
5284	struct ocfs2_extent_rec *rec;
5285	struct ocfs2_insert_type insert;
5286
5287	/*
5288	 * Setup the record to split before we grow the tree.
5289	 */
5290	el = path_leaf_el(path);
5291	rec = &el->l_recs[index];
5292	ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5293				   &split_rec, new_range, rec);
5294
5295	depth = path->p_tree_depth;
5296	if (depth > 0) {
5297		ret = ocfs2_read_extent_block(et->et_ci,
5298					      ocfs2_et_get_last_eb_blk(et),
5299					      &last_eb_bh);
5300		if (ret < 0) {
5301			mlog_errno(ret);
5302			goto out;
5303		}
5304
5305		eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5306		rightmost_el = &eb->h_list;
5307	} else
5308		rightmost_el = path_leaf_el(path);
5309
5310	credits = path->p_tree_depth +
5311		  ocfs2_extend_meta_needed(et->et_root_el);
5312	ret = ocfs2_extend_trans(handle, credits);
5313	if (ret) {
5314		mlog_errno(ret);
5315		goto out;
5316	}
5317
5318	if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5319	    le16_to_cpu(rightmost_el->l_count)) {
5320		ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5321				      meta_ac);
5322		if (ret) {
5323			mlog_errno(ret);
5324			goto out;
5325		}
5326	}
5327
5328	memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5329	insert.ins_appending = APPEND_NONE;
5330	insert.ins_contig = CONTIG_NONE;
5331	insert.ins_split = SPLIT_RIGHT;
5332	insert.ins_tree_depth = depth;
5333
5334	ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5335	if (ret)
5336		mlog_errno(ret);
5337
5338out:
5339	brelse(last_eb_bh);
5340	return ret;
5341}
5342
5343static int ocfs2_truncate_rec(handle_t *handle,
5344			      struct ocfs2_extent_tree *et,
5345			      struct ocfs2_path *path, int index,
5346			      struct ocfs2_cached_dealloc_ctxt *dealloc,
5347			      u32 cpos, u32 len)
5348{
5349	int ret;
5350	u32 left_cpos, rec_range, trunc_range;
5351	int is_rightmost_tree_rec = 0;
5352	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5353	struct ocfs2_path *left_path = NULL;
5354	struct ocfs2_extent_list *el = path_leaf_el(path);
5355	struct ocfs2_extent_rec *rec;
5356	struct ocfs2_extent_block *eb;
5357
5358	if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5359		/* extend credit for ocfs2_remove_rightmost_path */
5360		ret = ocfs2_extend_rotate_transaction(handle, 0,
5361				handle->h_buffer_credits,
5362				path);
5363		if (ret) {
5364			mlog_errno(ret);
5365			goto out;
5366		}
5367
5368		ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5369		if (ret) {
5370			mlog_errno(ret);
5371			goto out;
5372		}
5373
5374		index--;
5375	}
5376
5377	if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5378	    path->p_tree_depth) {
5379		/*
5380		 * Check whether this is the rightmost tree record. If
5381		 * we remove all of this record or part of its right
5382		 * edge then an update of the record lengths above it
5383		 * will be required.
5384		 */
5385		eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5386		if (eb->h_next_leaf_blk == 0)
5387			is_rightmost_tree_rec = 1;
5388	}
5389
5390	rec = &el->l_recs[index];
5391	if (index == 0 && path->p_tree_depth &&
5392	    le32_to_cpu(rec->e_cpos) == cpos) {
5393		/*
5394		 * Changing the leftmost offset (via partial or whole
5395		 * record truncate) of an interior (or rightmost) path
5396		 * means we have to update the subtree that is formed
5397		 * by this leaf and the one to it's left.
5398		 *
5399		 * There are two cases we can skip:
5400		 *   1) Path is the leftmost one in our btree.
5401		 *   2) The leaf is rightmost and will be empty after
5402		 *      we remove the extent record - the rotate code
5403		 *      knows how to update the newly formed edge.
5404		 */
5405
5406		ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5407		if (ret) {
5408			mlog_errno(ret);
5409			goto out;
5410		}
5411
5412		if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5413			left_path = ocfs2_new_path_from_path(path);
5414			if (!left_path) {
5415				ret = -ENOMEM;
5416				mlog_errno(ret);
5417				goto out;
5418			}
5419
5420			ret = ocfs2_find_path(et->et_ci, left_path,
5421					      left_cpos);
5422			if (ret) {
5423				mlog_errno(ret);
5424				goto out;
5425			}
5426		}
5427	}
5428
5429	ret = ocfs2_extend_rotate_transaction(handle, 0,
5430					      handle->h_buffer_credits,
5431					      path);
5432	if (ret) {
5433		mlog_errno(ret);
5434		goto out;
5435	}
5436
5437	ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5438	if (ret) {
5439		mlog_errno(ret);
5440		goto out;
5441	}
5442
5443	ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5444	if (ret) {
5445		mlog_errno(ret);
5446		goto out;
5447	}
5448
5449	rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5450	trunc_range = cpos + len;
5451
5452	if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5453		int next_free;
5454
5455		memset(rec, 0, sizeof(*rec));
5456		ocfs2_cleanup_merge(el, index);
5457
5458		next_free = le16_to_cpu(el->l_next_free_rec);
5459		if (is_rightmost_tree_rec && next_free > 1) {
5460			/*
5461			 * We skip the edge update if this path will
5462			 * be deleted by the rotate code.
5463			 */
5464			rec = &el->l_recs[next_free - 1];
5465			ocfs2_adjust_rightmost_records(handle, et, path,
5466						       rec);
5467		}
5468	} else if (le32_to_cpu(rec->e_cpos) == cpos) {
5469		/* Remove leftmost portion of the record. */
5470		le32_add_cpu(&rec->e_cpos, len);
5471		le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5472		le16_add_cpu(&rec->e_leaf_clusters, -len);
5473	} else if (rec_range == trunc_range) {
5474		/* Remove rightmost portion of the record */
5475		le16_add_cpu(&rec->e_leaf_clusters, -len);
5476		if (is_rightmost_tree_rec)
5477			ocfs2_adjust_rightmost_records(handle, et, path, rec);
5478	} else {
5479		/* Caller should have trapped this. */
5480		mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5481		     "(%u, %u)\n",
5482		     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5483		     le32_to_cpu(rec->e_cpos),
5484		     le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5485		BUG();
5486	}
5487
5488	if (left_path) {
5489		int subtree_index;
5490
5491		subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5492		ocfs2_complete_edge_insert(handle, left_path, path,
5493					   subtree_index);
5494	}
5495
5496	ocfs2_journal_dirty(handle, path_leaf_bh(path));
5497
5498	ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5499	if (ret)
5500		mlog_errno(ret);
5501
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	}
5636
5637out:
5638	ocfs2_free_path(path);
5639	return ret;
5640}
5641
5642/*
5643 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5644 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5645 * number to reserve some extra blocks, and it only handles meta
5646 * data allocations.
5647 *
5648 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5649 * and punching holes.
5650 */
5651static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5652					      struct ocfs2_extent_tree *et,
5653					      u32 extents_to_split,
5654					      struct ocfs2_alloc_context **ac,
5655					      int extra_blocks)
5656{
5657	int ret = 0, num_free_extents;
5658	unsigned int max_recs_needed = 2 * extents_to_split;
5659	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5660
5661	*ac = NULL;
5662
5663	num_free_extents = ocfs2_num_free_extents(et);
5664	if (num_free_extents < 0) {
5665		ret = num_free_extents;
5666		mlog_errno(ret);
5667		goto out;
5668	}
5669
5670	if (!num_free_extents ||
5671	    (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5672		extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5673
5674	if (extra_blocks) {
5675		ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5676		if (ret < 0) {
5677			if (ret != -ENOSPC)
5678				mlog_errno(ret);
5679		}
5680	}
5681
5682out:
5683	if (ret) {
5684		if (*ac) {
5685			ocfs2_free_alloc_context(*ac);
5686			*ac = NULL;
5687		}
5688	}
5689
5690	return ret;
5691}
5692
5693int ocfs2_remove_btree_range(struct inode *inode,
5694			     struct ocfs2_extent_tree *et,
5695			     u32 cpos, u32 phys_cpos, u32 len, int flags,
5696			     struct ocfs2_cached_dealloc_ctxt *dealloc,
5697			     u64 refcount_loc, bool refcount_tree_locked)
5698{
5699	int ret, credits = 0, extra_blocks = 0;
5700	u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5701	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5702	struct inode *tl_inode = osb->osb_tl_inode;
5703	handle_t *handle;
5704	struct ocfs2_alloc_context *meta_ac = NULL;
5705	struct ocfs2_refcount_tree *ref_tree = NULL;
5706
5707	if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5708		BUG_ON(!ocfs2_is_refcount_inode(inode));
5709
5710		if (!refcount_tree_locked) {
5711			ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5712						       &ref_tree, NULL);
5713			if (ret) {
5714				mlog_errno(ret);
5715				goto bail;
5716			}
5717		}
5718
5719		ret = ocfs2_prepare_refcount_change_for_del(inode,
5720							    refcount_loc,
5721							    phys_blkno,
5722							    len,
5723							    &credits,
5724							    &extra_blocks);
5725		if (ret < 0) {
5726			mlog_errno(ret);
5727			goto bail;
5728		}
5729	}
5730
5731	ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5732						 extra_blocks);
5733	if (ret) {
5734		mlog_errno(ret);
5735		goto bail;
5736	}
5737
5738	inode_lock(tl_inode);
5739
5740	if (ocfs2_truncate_log_needs_flush(osb)) {
5741		ret = __ocfs2_flush_truncate_log(osb);
5742		if (ret < 0) {
5743			mlog_errno(ret);
5744			goto out;
5745		}
5746	}
5747
5748	handle = ocfs2_start_trans(osb,
5749			ocfs2_remove_extent_credits(osb->sb) + credits);
5750	if (IS_ERR(handle)) {
5751		ret = PTR_ERR(handle);
5752		mlog_errno(ret);
5753		goto out;
5754	}
5755
5756	ret = ocfs2_et_root_journal_access(handle, et,
5757					   OCFS2_JOURNAL_ACCESS_WRITE);
5758	if (ret) {
5759		mlog_errno(ret);
5760		goto out_commit;
5761	}
5762
5763	dquot_free_space_nodirty(inode,
5764				  ocfs2_clusters_to_bytes(inode->i_sb, len));
5765
5766	ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5767	if (ret) {
5768		mlog_errno(ret);
5769		goto out_commit;
5770	}
5771
5772	ocfs2_et_update_clusters(et, -len);
5773	ocfs2_update_inode_fsync_trans(handle, inode, 1);
5774
5775	ocfs2_journal_dirty(handle, et->et_root_bh);
5776
5777	if (phys_blkno) {
5778		if (flags & OCFS2_EXT_REFCOUNTED)
5779			ret = ocfs2_decrease_refcount(inode, handle,
5780					ocfs2_blocks_to_clusters(osb->sb,
5781								 phys_blkno),
5782					len, meta_ac,
5783					dealloc, 1);
5784		else
5785			ret = ocfs2_truncate_log_append(osb, handle,
5786							phys_blkno, len);
5787		if (ret)
5788			mlog_errno(ret);
5789
5790	}
5791
5792out_commit:
5793	ocfs2_commit_trans(osb, handle);
5794out:
5795	inode_unlock(tl_inode);
5796bail:
5797	if (meta_ac)
5798		ocfs2_free_alloc_context(meta_ac);
5799
5800	if (ref_tree)
5801		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5802
5803	return ret;
5804}
5805
5806int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5807{
5808	struct buffer_head *tl_bh = osb->osb_tl_bh;
5809	struct ocfs2_dinode *di;
5810	struct ocfs2_truncate_log *tl;
5811
5812	di = (struct ocfs2_dinode *) tl_bh->b_data;
5813	tl = &di->id2.i_dealloc;
5814
5815	mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5816			"slot %d, invalid truncate log parameters: used = "
5817			"%u, count = %u\n", osb->slot_num,
5818			le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5819	return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5820}
5821
5822static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5823					   unsigned int new_start)
5824{
5825	unsigned int tail_index;
5826	unsigned int current_tail;
5827
5828	/* No records, nothing to coalesce */
5829	if (!le16_to_cpu(tl->tl_used))
5830		return 0;
5831
5832	tail_index = le16_to_cpu(tl->tl_used) - 1;
5833	current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5834	current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5835
5836	return current_tail == new_start;
5837}
5838
5839int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5840			      handle_t *handle,
5841			      u64 start_blk,
5842			      unsigned int num_clusters)
5843{
5844	int status, index;
5845	unsigned int start_cluster, tl_count;
5846	struct inode *tl_inode = osb->osb_tl_inode;
5847	struct buffer_head *tl_bh = osb->osb_tl_bh;
5848	struct ocfs2_dinode *di;
5849	struct ocfs2_truncate_log *tl;
5850
5851	BUG_ON(inode_trylock(tl_inode));
5852
5853	start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5854
5855	di = (struct ocfs2_dinode *) tl_bh->b_data;
5856
5857	/* tl_bh is loaded from ocfs2_truncate_log_init().  It's validated
5858	 * by the underlying call to ocfs2_read_inode_block(), so any
5859	 * corruption is a code bug */
5860	BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5861
5862	tl = &di->id2.i_dealloc;
5863	tl_count = le16_to_cpu(tl->tl_count);
5864	mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5865			tl_count == 0,
5866			"Truncate record count on #%llu invalid "
5867			"wanted %u, actual %u\n",
5868			(unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5869			ocfs2_truncate_recs_per_inode(osb->sb),
5870			le16_to_cpu(tl->tl_count));
5871
5872	/* Caller should have known to flush before calling us. */
5873	index = le16_to_cpu(tl->tl_used);
5874	if (index >= tl_count) {
5875		status = -ENOSPC;
5876		mlog_errno(status);
5877		goto bail;
5878	}
5879
5880	status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5881					 OCFS2_JOURNAL_ACCESS_WRITE);
5882	if (status < 0) {
5883		mlog_errno(status);
5884		goto bail;
5885	}
5886
5887	trace_ocfs2_truncate_log_append(
5888		(unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5889		start_cluster, num_clusters);
5890	if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5891		/*
5892		 * Move index back to the record we are coalescing with.
5893		 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5894		 */
5895		index--;
5896
5897		num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5898		trace_ocfs2_truncate_log_append(
5899			(unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5900			index, le32_to_cpu(tl->tl_recs[index].t_start),
5901			num_clusters);
5902	} else {
5903		tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5904		tl->tl_used = cpu_to_le16(index + 1);
5905	}
5906	tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5907
5908	ocfs2_journal_dirty(handle, tl_bh);
5909
5910	osb->truncated_clusters += num_clusters;
5911bail:
5912	return status;
5913}
5914
5915static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5916					 struct inode *data_alloc_inode,
5917					 struct buffer_head *data_alloc_bh)
5918{
5919	int status = 0;
5920	int i;
5921	unsigned int num_clusters;
5922	u64 start_blk;
5923	struct ocfs2_truncate_rec rec;
5924	struct ocfs2_dinode *di;
5925	struct ocfs2_truncate_log *tl;
5926	struct inode *tl_inode = osb->osb_tl_inode;
5927	struct buffer_head *tl_bh = osb->osb_tl_bh;
5928	handle_t *handle;
5929
5930	di = (struct ocfs2_dinode *) tl_bh->b_data;
5931	tl = &di->id2.i_dealloc;
5932	i = le16_to_cpu(tl->tl_used) - 1;
5933	while (i >= 0) {
5934		handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5935		if (IS_ERR(handle)) {
5936			status = PTR_ERR(handle);
5937			mlog_errno(status);
5938			goto bail;
5939		}
5940
5941		/* Caller has given us at least enough credits to
5942		 * update the truncate log dinode */
5943		status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5944						 OCFS2_JOURNAL_ACCESS_WRITE);
5945		if (status < 0) {
5946			mlog_errno(status);
5947			goto bail;
5948		}
5949
5950		tl->tl_used = cpu_to_le16(i);
5951
5952		ocfs2_journal_dirty(handle, tl_bh);
5953
5954		rec = tl->tl_recs[i];
5955		start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5956						    le32_to_cpu(rec.t_start));
5957		num_clusters = le32_to_cpu(rec.t_clusters);
5958
5959		/* if start_blk is not set, we ignore the record as
5960		 * invalid. */
5961		if (start_blk) {
5962			trace_ocfs2_replay_truncate_records(
5963				(unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5964				i, le32_to_cpu(rec.t_start), num_clusters);
5965
5966			status = ocfs2_free_clusters(handle, data_alloc_inode,
5967						     data_alloc_bh, start_blk,
5968						     num_clusters);
5969			if (status < 0) {
5970				mlog_errno(status);
5971				goto bail;
5972			}
5973		}
5974
5975		ocfs2_commit_trans(osb, handle);
5976		i--;
5977	}
5978
5979	osb->truncated_clusters = 0;
5980
5981bail:
5982	return status;
5983}
5984
5985/* Expects you to already be holding tl_inode->i_mutex */
5986int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5987{
5988	int status;
5989	unsigned int num_to_flush;
5990	struct inode *tl_inode = osb->osb_tl_inode;
5991	struct inode *data_alloc_inode = NULL;
5992	struct buffer_head *tl_bh = osb->osb_tl_bh;
5993	struct buffer_head *data_alloc_bh = NULL;
5994	struct ocfs2_dinode *di;
5995	struct ocfs2_truncate_log *tl;
5996	struct ocfs2_journal *journal = osb->journal;
5997
5998	BUG_ON(inode_trylock(tl_inode));
5999
6000	di = (struct ocfs2_dinode *) tl_bh->b_data;
6001
6002	/* tl_bh is loaded from ocfs2_truncate_log_init().  It's validated
6003	 * by the underlying call to ocfs2_read_inode_block(), so any
6004	 * corruption is a code bug */
6005	BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6006
6007	tl = &di->id2.i_dealloc;
6008	num_to_flush = le16_to_cpu(tl->tl_used);
6009	trace_ocfs2_flush_truncate_log(
6010		(unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6011		num_to_flush);
6012	if (!num_to_flush) {
6013		status = 0;
6014		goto out;
6015	}
6016
6017	/* Appending truncate log(TA) and and flushing truncate log(TF) are
6018	 * two separated transactions. They can be both committed but not
6019	 * checkpointed. If crash occurs then, both two transaction will be
6020	 * replayed with several already released to global bitmap clusters.
6021	 * Then truncate log will be replayed resulting in cluster double free.
6022	 */
6023	jbd2_journal_lock_updates(journal->j_journal);
6024	status = jbd2_journal_flush(journal->j_journal);
6025	jbd2_journal_unlock_updates(journal->j_journal);
6026	if (status < 0) {
6027		mlog_errno(status);
6028		goto out;
6029	}
6030
6031	data_alloc_inode = ocfs2_get_system_file_inode(osb,
6032						       GLOBAL_BITMAP_SYSTEM_INODE,
6033						       OCFS2_INVALID_SLOT);
6034	if (!data_alloc_inode) {
6035		status = -EINVAL;
6036		mlog(ML_ERROR, "Could not get bitmap inode!\n");
6037		goto out;
6038	}
6039
6040	inode_lock(data_alloc_inode);
6041
6042	status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6043	if (status < 0) {
6044		mlog_errno(status);
6045		goto out_mutex;
6046	}
6047
6048	status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6049					       data_alloc_bh);
6050	if (status < 0)
6051		mlog_errno(status);
6052
6053	brelse(data_alloc_bh);
6054	ocfs2_inode_unlock(data_alloc_inode, 1);
6055
6056out_mutex:
6057	inode_unlock(data_alloc_inode);
6058	iput(data_alloc_inode);
6059
6060out:
6061	return status;
6062}
6063
6064int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6065{
6066	int status;
6067	struct inode *tl_inode = osb->osb_tl_inode;
6068
6069	inode_lock(tl_inode);
6070	status = __ocfs2_flush_truncate_log(osb);
6071	inode_unlock(tl_inode);
6072
6073	return status;
6074}
6075
6076static void ocfs2_truncate_log_worker(struct work_struct *work)
6077{
6078	int status;
6079	struct ocfs2_super *osb =
6080		container_of(work, struct ocfs2_super,
6081			     osb_truncate_log_wq.work);
6082
6083	status = ocfs2_flush_truncate_log(osb);
6084	if (status < 0)
6085		mlog_errno(status);
6086	else
6087		ocfs2_init_steal_slots(osb);
6088}
6089
6090#define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6091void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6092				       int cancel)
6093{
6094	if (osb->osb_tl_inode &&
6095			atomic_read(&osb->osb_tl_disable) == 0) {
6096		/* We want to push off log flushes while truncates are
6097		 * still running. */
6098		if (cancel)
6099			cancel_delayed_work(&osb->osb_truncate_log_wq);
6100
6101		queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6102				   OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6103	}
6104}
6105
6106/*
6107 * Try to flush truncate logs if we can free enough clusters from it.
6108 * As for return value, "< 0" means error, "0" no space and "1" means
6109 * we have freed enough spaces and let the caller try to allocate again.
6110 */
6111int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6112					unsigned int needed)
6113{
6114	tid_t target;
6115	int ret = 0;
6116	unsigned int truncated_clusters;
6117
6118	inode_lock(osb->osb_tl_inode);
6119	truncated_clusters = osb->truncated_clusters;
6120	inode_unlock(osb->osb_tl_inode);
6121
6122	/*
6123	 * Check whether we can succeed in allocating if we free
6124	 * the truncate log.
6125	 */
6126	if (truncated_clusters < needed)
6127		goto out;
6128
6129	ret = ocfs2_flush_truncate_log(osb);
6130	if (ret) {
6131		mlog_errno(ret);
6132		goto out;
6133	}
6134
6135	if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6136		jbd2_log_wait_commit(osb->journal->j_journal, target);
6137		ret = 1;
6138	}
6139out:
6140	return ret;
6141}
6142
6143static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6144				       int slot_num,
6145				       struct inode **tl_inode,
6146				       struct buffer_head **tl_bh)
6147{
6148	int status;
6149	struct inode *inode = NULL;
6150	struct buffer_head *bh = NULL;
6151
6152	inode = ocfs2_get_system_file_inode(osb,
6153					   TRUNCATE_LOG_SYSTEM_INODE,
6154					   slot_num);
6155	if (!inode) {
6156		status = -EINVAL;
6157		mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6158		goto bail;
6159	}
6160
6161	status = ocfs2_read_inode_block(inode, &bh);
6162	if (status < 0) {
6163		iput(inode);
6164		mlog_errno(status);
6165		goto bail;
6166	}
6167
6168	*tl_inode = inode;
6169	*tl_bh    = bh;
6170bail:
6171	return status;
6172}
6173
6174/* called during the 1st stage of node recovery. we stamp a clean
6175 * truncate log and pass back a copy for processing later. if the
6176 * truncate log does not require processing, a *tl_copy is set to
6177 * NULL. */
6178int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6179				      int slot_num,
6180				      struct ocfs2_dinode **tl_copy)
6181{
6182	int status;
6183	struct inode *tl_inode = NULL;
6184	struct buffer_head *tl_bh = NULL;
6185	struct ocfs2_dinode *di;
6186	struct ocfs2_truncate_log *tl;
6187
6188	*tl_copy = NULL;
6189
6190	trace_ocfs2_begin_truncate_log_recovery(slot_num);
6191
6192	status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6193	if (status < 0) {
6194		mlog_errno(status);
6195		goto bail;
6196	}
6197
6198	di = (struct ocfs2_dinode *) tl_bh->b_data;
6199
6200	/* tl_bh is loaded from ocfs2_get_truncate_log_info().  It's
6201	 * validated by the underlying call to ocfs2_read_inode_block(),
6202	 * so any corruption is a code bug */
6203	BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6204
6205	tl = &di->id2.i_dealloc;
6206	if (le16_to_cpu(tl->tl_used)) {
6207		trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6208
6209		/*
6210		 * Assuming the write-out below goes well, this copy will be
6211		 * passed back to recovery for processing.
6212		 */
6213		*tl_copy = kmemdup(tl_bh->b_data, tl_bh->b_size, GFP_KERNEL);
6214		if (!(*tl_copy)) {
6215			status = -ENOMEM;
6216			mlog_errno(status);
6217			goto bail;
6218		}
6219
6220		/* All we need to do to clear the truncate log is set
6221		 * tl_used. */
6222		tl->tl_used = 0;
6223
6224		ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6225		status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6226		if (status < 0) {
6227			mlog_errno(status);
6228			goto bail;
6229		}
6230	}
6231
6232bail:
6233	iput(tl_inode);
6234	brelse(tl_bh);
6235
6236	if (status < 0) {
6237		kfree(*tl_copy);
6238		*tl_copy = NULL;
6239		mlog_errno(status);
6240	}
6241
6242	return status;
6243}
6244
6245int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6246					 struct ocfs2_dinode *tl_copy)
6247{
6248	int status = 0;
6249	int i;
6250	unsigned int clusters, num_recs, start_cluster;
6251	u64 start_blk;
6252	handle_t *handle;
6253	struct inode *tl_inode = osb->osb_tl_inode;
6254	struct ocfs2_truncate_log *tl;
6255
6256	if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6257		mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6258		return -EINVAL;
6259	}
6260
6261	tl = &tl_copy->id2.i_dealloc;
6262	num_recs = le16_to_cpu(tl->tl_used);
6263	trace_ocfs2_complete_truncate_log_recovery(
6264		(unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6265		num_recs);
6266
6267	inode_lock(tl_inode);
6268	for(i = 0; i < num_recs; i++) {
6269		if (ocfs2_truncate_log_needs_flush(osb)) {
6270			status = __ocfs2_flush_truncate_log(osb);
6271			if (status < 0) {
6272				mlog_errno(status);
6273				goto bail_up;
6274			}
6275		}
6276
6277		handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6278		if (IS_ERR(handle)) {
6279			status = PTR_ERR(handle);
6280			mlog_errno(status);
6281			goto bail_up;
6282		}
6283
6284		clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6285		start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6286		start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6287
6288		status = ocfs2_truncate_log_append(osb, handle,
6289						   start_blk, clusters);
6290		ocfs2_commit_trans(osb, handle);
6291		if (status < 0) {
6292			mlog_errno(status);
6293			goto bail_up;
6294		}
6295	}
6296
6297bail_up:
6298	inode_unlock(tl_inode);
6299
6300	return status;
6301}
6302
6303void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6304{
6305	int status;
6306	struct inode *tl_inode = osb->osb_tl_inode;
6307
6308	atomic_set(&osb->osb_tl_disable, 1);
6309
6310	if (tl_inode) {
6311		cancel_delayed_work(&osb->osb_truncate_log_wq);
6312		flush_workqueue(osb->ocfs2_wq);
6313
6314		status = ocfs2_flush_truncate_log(osb);
6315		if (status < 0)
6316			mlog_errno(status);
6317
6318		brelse(osb->osb_tl_bh);
6319		iput(osb->osb_tl_inode);
6320	}
6321}
6322
6323int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6324{
6325	int status;
6326	struct inode *tl_inode = NULL;
6327	struct buffer_head *tl_bh = NULL;
6328
6329	status = ocfs2_get_truncate_log_info(osb,
6330					     osb->slot_num,
6331					     &tl_inode,
6332					     &tl_bh);
6333	if (status < 0)
6334		mlog_errno(status);
6335
6336	/* ocfs2_truncate_log_shutdown keys on the existence of
6337	 * osb->osb_tl_inode so we don't set any of the osb variables
6338	 * until we're sure all is well. */
6339	INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6340			  ocfs2_truncate_log_worker);
6341	atomic_set(&osb->osb_tl_disable, 0);
6342	osb->osb_tl_bh    = tl_bh;
6343	osb->osb_tl_inode = tl_inode;
6344
6345	return status;
6346}
6347
6348/*
6349 * Delayed de-allocation of suballocator blocks.
6350 *
6351 * Some sets of block de-allocations might involve multiple suballocator inodes.
6352 *
6353 * The locking for this can get extremely complicated, especially when
6354 * the suballocator inodes to delete from aren't known until deep
6355 * within an unrelated codepath.
6356 *
6357 * ocfs2_extent_block structures are a good example of this - an inode
6358 * btree could have been grown by any number of nodes each allocating
6359 * out of their own suballoc inode.
6360 *
6361 * These structures allow the delay of block de-allocation until a
6362 * later time, when locking of multiple cluster inodes won't cause
6363 * deadlock.
6364 */
6365
6366/*
6367 * Describe a single bit freed from a suballocator.  For the block
6368 * suballocators, it represents one block.  For the global cluster
6369 * allocator, it represents some clusters and free_bit indicates
6370 * clusters number.
6371 */
6372struct ocfs2_cached_block_free {
6373	struct ocfs2_cached_block_free		*free_next;
6374	u64					free_bg;
6375	u64					free_blk;
6376	unsigned int				free_bit;
6377};
6378
6379struct ocfs2_per_slot_free_list {
6380	struct ocfs2_per_slot_free_list		*f_next_suballocator;
6381	int					f_inode_type;
6382	int					f_slot;
6383	struct ocfs2_cached_block_free		*f_first;
6384};
6385
6386static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6387				    int sysfile_type,
6388				    int slot,
6389				    struct ocfs2_cached_block_free *head)
6390{
6391	int ret;
6392	u64 bg_blkno;
6393	handle_t *handle;
6394	struct inode *inode;
6395	struct buffer_head *di_bh = NULL;
6396	struct ocfs2_cached_block_free *tmp;
6397
6398	inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6399	if (!inode) {
6400		ret = -EINVAL;
6401		mlog_errno(ret);
6402		goto out;
6403	}
6404
6405	inode_lock(inode);
6406
6407	ret = ocfs2_inode_lock(inode, &di_bh, 1);
6408	if (ret) {
6409		mlog_errno(ret);
6410		goto out_mutex;
6411	}
6412
6413	while (head) {
6414		if (head->free_bg)
6415			bg_blkno = head->free_bg;
6416		else
6417			bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6418							      head->free_bit);
6419		handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6420		if (IS_ERR(handle)) {
6421			ret = PTR_ERR(handle);
6422			mlog_errno(ret);
6423			goto out_unlock;
6424		}
6425
6426		trace_ocfs2_free_cached_blocks(
6427		     (unsigned long long)head->free_blk, head->free_bit);
6428
6429		ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6430					       head->free_bit, bg_blkno, 1);
6431		if (ret)
6432			mlog_errno(ret);
6433
6434		ocfs2_commit_trans(osb, handle);
6435
6436		tmp = head;
6437		head = head->free_next;
6438		kfree(tmp);
6439	}
6440
6441out_unlock:
6442	ocfs2_inode_unlock(inode, 1);
6443	brelse(di_bh);
6444out_mutex:
6445	inode_unlock(inode);
6446	iput(inode);
6447out:
6448	while(head) {
6449		/* Premature exit may have left some dangling items. */
6450		tmp = head;
6451		head = head->free_next;
6452		kfree(tmp);
6453	}
6454
6455	return ret;
6456}
6457
6458int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6459				u64 blkno, unsigned int bit)
6460{
6461	int ret = 0;
6462	struct ocfs2_cached_block_free *item;
6463
6464	item = kzalloc(sizeof(*item), GFP_NOFS);
6465	if (item == NULL) {
6466		ret = -ENOMEM;
6467		mlog_errno(ret);
6468		return ret;
6469	}
6470
6471	trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6472
6473	item->free_blk = blkno;
6474	item->free_bit = bit;
6475	item->free_next = ctxt->c_global_allocator;
6476
6477	ctxt->c_global_allocator = item;
6478	return ret;
6479}
6480
6481static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6482				      struct ocfs2_cached_block_free *head)
6483{
6484	struct ocfs2_cached_block_free *tmp;
6485	struct inode *tl_inode = osb->osb_tl_inode;
6486	handle_t *handle;
6487	int ret = 0;
6488
6489	inode_lock(tl_inode);
6490
6491	while (head) {
6492		if (ocfs2_truncate_log_needs_flush(osb)) {
6493			ret = __ocfs2_flush_truncate_log(osb);
6494			if (ret < 0) {
6495				mlog_errno(ret);
6496				break;
6497			}
6498		}
6499
6500		handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6501		if (IS_ERR(handle)) {
6502			ret = PTR_ERR(handle);
6503			mlog_errno(ret);
6504			break;
6505		}
6506
6507		ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6508						head->free_bit);
6509
6510		ocfs2_commit_trans(osb, handle);
6511		tmp = head;
6512		head = head->free_next;
6513		kfree(tmp);
6514
6515		if (ret < 0) {
6516			mlog_errno(ret);
6517			break;
6518		}
6519	}
6520
6521	inode_unlock(tl_inode);
6522
6523	while (head) {
6524		/* Premature exit may have left some dangling items. */
6525		tmp = head;
6526		head = head->free_next;
6527		kfree(tmp);
6528	}
6529
6530	return ret;
6531}
6532
6533int ocfs2_run_deallocs(struct ocfs2_super *osb,
6534		       struct ocfs2_cached_dealloc_ctxt *ctxt)
6535{
6536	int ret = 0, ret2;
6537	struct ocfs2_per_slot_free_list *fl;
6538
6539	if (!ctxt)
6540		return 0;
6541
6542	while (ctxt->c_first_suballocator) {
6543		fl = ctxt->c_first_suballocator;
6544
6545		if (fl->f_first) {
6546			trace_ocfs2_run_deallocs(fl->f_inode_type,
6547						 fl->f_slot);
6548			ret2 = ocfs2_free_cached_blocks(osb,
6549							fl->f_inode_type,
6550							fl->f_slot,
6551							fl->f_first);
6552			if (ret2)
6553				mlog_errno(ret2);
6554			if (!ret)
6555				ret = ret2;
6556		}
6557
6558		ctxt->c_first_suballocator = fl->f_next_suballocator;
6559		kfree(fl);
6560	}
6561
6562	if (ctxt->c_global_allocator) {
6563		ret2 = ocfs2_free_cached_clusters(osb,
6564						  ctxt->c_global_allocator);
6565		if (ret2)
6566			mlog_errno(ret2);
6567		if (!ret)
6568			ret = ret2;
6569
6570		ctxt->c_global_allocator = NULL;
6571	}
6572
6573	return ret;
6574}
6575
6576static struct ocfs2_per_slot_free_list *
6577ocfs2_find_per_slot_free_list(int type,
6578			      int slot,
6579			      struct ocfs2_cached_dealloc_ctxt *ctxt)
6580{
6581	struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6582
6583	while (fl) {
6584		if (fl->f_inode_type == type && fl->f_slot == slot)
6585			return fl;
6586
6587		fl = fl->f_next_suballocator;
6588	}
6589
6590	fl = kmalloc(sizeof(*fl), GFP_NOFS);
6591	if (fl) {
6592		fl->f_inode_type = type;
6593		fl->f_slot = slot;
6594		fl->f_first = NULL;
6595		fl->f_next_suballocator = ctxt->c_first_suballocator;
6596
6597		ctxt->c_first_suballocator = fl;
6598	}
6599	return fl;
6600}
6601
6602static struct ocfs2_per_slot_free_list *
6603ocfs2_find_preferred_free_list(int type,
6604			       int preferred_slot,
6605			       int *real_slot,
6606			       struct ocfs2_cached_dealloc_ctxt *ctxt)
6607{
6608	struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6609
6610	while (fl) {
6611		if (fl->f_inode_type == type && fl->f_slot == preferred_slot) {
6612			*real_slot = fl->f_slot;
6613			return fl;
6614		}
6615
6616		fl = fl->f_next_suballocator;
6617	}
6618
6619	/* If we can't find any free list matching preferred slot, just use
6620	 * the first one.
6621	 */
6622	fl = ctxt->c_first_suballocator;
6623	*real_slot = fl->f_slot;
6624
6625	return fl;
6626}
6627
6628/* Return Value 1 indicates empty */
6629static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et)
6630{
6631	struct ocfs2_per_slot_free_list *fl = NULL;
6632
6633	if (!et->et_dealloc)
6634		return 1;
6635
6636	fl = et->et_dealloc->c_first_suballocator;
6637	if (!fl)
6638		return 1;
6639
6640	if (!fl->f_first)
6641		return 1;
6642
6643	return 0;
6644}
6645
6646/* If extent was deleted from tree due to extent rotation and merging, and
6647 * no metadata is reserved ahead of time. Try to reuse some extents
6648 * just deleted. This is only used to reuse extent blocks.
6649 * It is supposed to find enough extent blocks in dealloc if our estimation
6650 * on metadata is accurate.
6651 */
6652static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
6653					struct ocfs2_extent_tree *et,
6654					struct buffer_head **new_eb_bh,
6655					int blk_wanted, int *blk_given)
6656{
6657	int i, status = 0, real_slot;
6658	struct ocfs2_cached_dealloc_ctxt *dealloc;
6659	struct ocfs2_per_slot_free_list *fl;
6660	struct ocfs2_cached_block_free *bf;
6661	struct ocfs2_extent_block *eb;
6662	struct ocfs2_super *osb =
6663		OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
6664
6665	*blk_given = 0;
6666
6667	/* If extent tree doesn't have a dealloc, this is not faulty. Just
6668	 * tell upper caller dealloc can't provide any block and it should
6669	 * ask for alloc to claim more space.
6670	 */
6671	dealloc = et->et_dealloc;
6672	if (!dealloc)
6673		goto bail;
6674
6675	for (i = 0; i < blk_wanted; i++) {
6676		/* Prefer to use local slot */
6677		fl = ocfs2_find_preferred_free_list(EXTENT_ALLOC_SYSTEM_INODE,
6678						    osb->slot_num, &real_slot,
6679						    dealloc);
6680		/* If no more block can be reused, we should claim more
6681		 * from alloc. Just return here normally.
6682		 */
6683		if (!fl) {
6684			status = 0;
6685			break;
6686		}
6687
6688		bf = fl->f_first;
6689		fl->f_first = bf->free_next;
6690
6691		new_eb_bh[i] = sb_getblk(osb->sb, bf->free_blk);
6692		if (new_eb_bh[i] == NULL) {
6693			status = -ENOMEM;
6694			mlog_errno(status);
6695			goto bail;
6696		}
6697
6698		mlog(0, "Reusing block(%llu) from "
6699		     "dealloc(local slot:%d, real slot:%d)\n",
6700		     bf->free_blk, osb->slot_num, real_slot);
6701
6702		ocfs2_set_new_buffer_uptodate(et->et_ci, new_eb_bh[i]);
6703
6704		status = ocfs2_journal_access_eb(handle, et->et_ci,
6705						 new_eb_bh[i],
6706						 OCFS2_JOURNAL_ACCESS_CREATE);
6707		if (status < 0) {
6708			mlog_errno(status);
6709			goto bail;
6710		}
6711
6712		memset(new_eb_bh[i]->b_data, 0, osb->sb->s_blocksize);
6713		eb = (struct ocfs2_extent_block *) new_eb_bh[i]->b_data;
6714
6715		/* We can't guarantee that buffer head is still cached, so
6716		 * polutlate the extent block again.
6717		 */
6718		strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
6719		eb->h_blkno = cpu_to_le64(bf->free_blk);
6720		eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
6721		eb->h_suballoc_slot = cpu_to_le16(real_slot);
6722		eb->h_suballoc_loc = cpu_to_le64(bf->free_bg);
6723		eb->h_suballoc_bit = cpu_to_le16(bf->free_bit);
6724		eb->h_list.l_count =
6725			cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
6726
6727		/* We'll also be dirtied by the caller, so
6728		 * this isn't absolutely necessary.
6729		 */
6730		ocfs2_journal_dirty(handle, new_eb_bh[i]);
6731
6732		if (!fl->f_first) {
6733			dealloc->c_first_suballocator = fl->f_next_suballocator;
6734			kfree(fl);
6735		}
6736		kfree(bf);
6737	}
6738
6739	*blk_given = i;
6740
6741bail:
6742	if (unlikely(status < 0)) {
6743		for (i = 0; i < blk_wanted; i++)
6744			brelse(new_eb_bh[i]);
6745	}
6746
6747	return status;
6748}
6749
6750int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6751			      int type, int slot, u64 suballoc,
6752			      u64 blkno, unsigned int bit)
6753{
6754	int ret;
6755	struct ocfs2_per_slot_free_list *fl;
6756	struct ocfs2_cached_block_free *item;
6757
6758	fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6759	if (fl == NULL) {
6760		ret = -ENOMEM;
6761		mlog_errno(ret);
6762		goto out;
6763	}
6764
6765	item = kzalloc(sizeof(*item), GFP_NOFS);
6766	if (item == NULL) {
6767		ret = -ENOMEM;
6768		mlog_errno(ret);
6769		goto out;
6770	}
6771
6772	trace_ocfs2_cache_block_dealloc(type, slot,
6773					(unsigned long long)suballoc,
6774					(unsigned long long)blkno, bit);
6775
6776	item->free_bg = suballoc;
6777	item->free_blk = blkno;
6778	item->free_bit = bit;
6779	item->free_next = fl->f_first;
6780
6781	fl->f_first = item;
6782
6783	ret = 0;
6784out:
6785	return ret;
6786}
6787
6788static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6789					 struct ocfs2_extent_block *eb)
6790{
6791	return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6792					 le16_to_cpu(eb->h_suballoc_slot),
6793					 le64_to_cpu(eb->h_suballoc_loc),
6794					 le64_to_cpu(eb->h_blkno),
6795					 le16_to_cpu(eb->h_suballoc_bit));
6796}
6797
6798static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6799{
6800	set_buffer_uptodate(bh);
6801	mark_buffer_dirty(bh);
6802	return 0;
6803}
6804
6805void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6806			      unsigned int from, unsigned int to,
6807			      struct page *page, int zero, u64 *phys)
6808{
6809	int ret, partial = 0;
6810	loff_t start_byte = ((loff_t)page->index << PAGE_SHIFT) + from;
6811	loff_t length = to - from;
6812
6813	ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6814	if (ret)
6815		mlog_errno(ret);
6816
6817	if (zero)
6818		zero_user_segment(page, from, to);
6819
6820	/*
6821	 * Need to set the buffers we zero'd into uptodate
6822	 * here if they aren't - ocfs2_map_page_blocks()
6823	 * might've skipped some
6824	 */
6825	ret = walk_page_buffers(handle, page_buffers(page),
6826				from, to, &partial,
6827				ocfs2_zero_func);
6828	if (ret < 0)
6829		mlog_errno(ret);
6830	else if (ocfs2_should_order_data(inode)) {
6831		ret = ocfs2_jbd2_inode_add_write(handle, inode,
6832						 start_byte, length);
6833		if (ret < 0)
6834			mlog_errno(ret);
6835	}
6836
6837	if (!partial)
6838		SetPageUptodate(page);
6839
6840	flush_dcache_page(page);
6841}
6842
6843static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6844				     loff_t end, struct page **pages,
6845				     int numpages, u64 phys, handle_t *handle)
6846{
6847	int i;
6848	struct page *page;
6849	unsigned int from, to = PAGE_SIZE;
6850	struct super_block *sb = inode->i_sb;
6851
6852	BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6853
6854	if (numpages == 0)
6855		goto out;
6856
6857	to = PAGE_SIZE;
6858	for(i = 0; i < numpages; i++) {
6859		page = pages[i];
6860
6861		from = start & (PAGE_SIZE - 1);
6862		if ((end >> PAGE_SHIFT) == page->index)
6863			to = end & (PAGE_SIZE - 1);
6864
6865		BUG_ON(from > PAGE_SIZE);
6866		BUG_ON(to > PAGE_SIZE);
6867
6868		ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6869					 &phys);
6870
6871		start = (page->index + 1) << PAGE_SHIFT;
6872	}
6873out:
6874	if (pages)
6875		ocfs2_unlock_and_free_pages(pages, numpages);
6876}
6877
6878int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6879		     struct page **pages, int *num)
6880{
6881	int numpages, ret = 0;
6882	struct address_space *mapping = inode->i_mapping;
6883	unsigned long index;
6884	loff_t last_page_bytes;
6885
6886	BUG_ON(start > end);
6887
6888	numpages = 0;
6889	last_page_bytes = PAGE_ALIGN(end);
6890	index = start >> PAGE_SHIFT;
6891	do {
6892		pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6893		if (!pages[numpages]) {
6894			ret = -ENOMEM;
6895			mlog_errno(ret);
6896			goto out;
6897		}
6898
6899		numpages++;
6900		index++;
6901	} while (index < (last_page_bytes >> PAGE_SHIFT));
6902
6903out:
6904	if (ret != 0) {
6905		if (pages)
6906			ocfs2_unlock_and_free_pages(pages, numpages);
6907		numpages = 0;
6908	}
6909
6910	*num = numpages;
6911
6912	return ret;
6913}
6914
6915static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6916				struct page **pages, int *num)
6917{
6918	struct super_block *sb = inode->i_sb;
6919
6920	BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6921	       (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6922
6923	return ocfs2_grab_pages(inode, start, end, pages, num);
6924}
6925
6926/*
6927 * Zero the area past i_size but still within an allocated
6928 * cluster. This avoids exposing nonzero data on subsequent file
6929 * extends.
6930 *
6931 * We need to call this before i_size is updated on the inode because
6932 * otherwise block_write_full_page() will skip writeout of pages past
6933 * i_size. The new_i_size parameter is passed for this reason.
6934 */
6935int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6936				  u64 range_start, u64 range_end)
6937{
6938	int ret = 0, numpages;
6939	struct page **pages = NULL;
6940	u64 phys;
6941	unsigned int ext_flags;
6942	struct super_block *sb = inode->i_sb;
6943
6944	/*
6945	 * File systems which don't support sparse files zero on every
6946	 * extend.
6947	 */
6948	if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6949		return 0;
6950
6951	pages = kcalloc(ocfs2_pages_per_cluster(sb),
6952			sizeof(struct page *), GFP_NOFS);
6953	if (pages == NULL) {
6954		ret = -ENOMEM;
6955		mlog_errno(ret);
6956		goto out;
6957	}
6958
6959	if (range_start == range_end)
6960		goto out;
6961
6962	ret = ocfs2_extent_map_get_blocks(inode,
6963					  range_start >> sb->s_blocksize_bits,
6964					  &phys, NULL, &ext_flags);
6965	if (ret) {
6966		mlog_errno(ret);
6967		goto out;
6968	}
6969
6970	/*
6971	 * Tail is a hole, or is marked unwritten. In either case, we
6972	 * can count on read and write to return/push zero's.
6973	 */
6974	if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6975		goto out;
6976
6977	ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6978				   &numpages);
6979	if (ret) {
6980		mlog_errno(ret);
6981		goto out;
6982	}
6983
6984	ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6985				 numpages, phys, handle);
6986
6987	/*
6988	 * Initiate writeout of the pages we zero'd here. We don't
6989	 * wait on them - the truncate_inode_pages() call later will
6990	 * do that for us.
6991	 */
6992	ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6993				       range_end - 1);
6994	if (ret)
6995		mlog_errno(ret);
6996
6997out:
6998	kfree(pages);
6999
7000	return ret;
7001}
7002
7003static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
7004					     struct ocfs2_dinode *di)
7005{
7006	unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
7007	unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
7008
7009	if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
7010		memset(&di->id2, 0, blocksize -
7011				    offsetof(struct ocfs2_dinode, id2) -
7012				    xattrsize);
7013	else
7014		memset(&di->id2, 0, blocksize -
7015				    offsetof(struct ocfs2_dinode, id2));
7016}
7017
7018void ocfs2_dinode_new_extent_list(struct inode *inode,
7019				  struct ocfs2_dinode *di)
7020{
7021	ocfs2_zero_dinode_id2_with_xattr(inode, di);
7022	di->id2.i_list.l_tree_depth = 0;
7023	di->id2.i_list.l_next_free_rec = 0;
7024	di->id2.i_list.l_count = cpu_to_le16(
7025		ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7026}
7027
7028void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7029{
7030	struct ocfs2_inode_info *oi = OCFS2_I(inode);
7031	struct ocfs2_inline_data *idata = &di->id2.i_data;
7032
7033	spin_lock(&oi->ip_lock);
7034	oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7035	di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7036	spin_unlock(&oi->ip_lock);
7037
7038	/*
7039	 * We clear the entire i_data structure here so that all
7040	 * fields can be properly initialized.
7041	 */
7042	ocfs2_zero_dinode_id2_with_xattr(inode, di);
7043
7044	idata->id_count = cpu_to_le16(
7045			ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7046}
7047
7048int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7049					 struct buffer_head *di_bh)
7050{
7051	int ret, i, has_data, num_pages = 0;
7052	int need_free = 0;
7053	u32 bit_off, num;
7054	handle_t *handle;
7055	u64 uninitialized_var(block);
7056	struct ocfs2_inode_info *oi = OCFS2_I(inode);
7057	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7058	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7059	struct ocfs2_alloc_context *data_ac = NULL;
7060	struct page **pages = NULL;
7061	loff_t end = osb->s_clustersize;
7062	struct ocfs2_extent_tree et;
7063	int did_quota = 0;
7064
7065	has_data = i_size_read(inode) ? 1 : 0;
7066
7067	if (has_data) {
7068		pages = kcalloc(ocfs2_pages_per_cluster(osb->sb),
7069				sizeof(struct page *), GFP_NOFS);
7070		if (pages == NULL) {
7071			ret = -ENOMEM;
7072			mlog_errno(ret);
7073			return ret;
7074		}
7075
7076		ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7077		if (ret) {
7078			mlog_errno(ret);
7079			goto free_pages;
7080		}
7081	}
7082
7083	handle = ocfs2_start_trans(osb,
7084				   ocfs2_inline_to_extents_credits(osb->sb));
7085	if (IS_ERR(handle)) {
7086		ret = PTR_ERR(handle);
7087		mlog_errno(ret);
7088		goto out;
7089	}
7090
7091	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7092				      OCFS2_JOURNAL_ACCESS_WRITE);
7093	if (ret) {
7094		mlog_errno(ret);
7095		goto out_commit;
7096	}
7097
7098	if (has_data) {
7099		unsigned int page_end;
 
7100		u64 phys;
7101
7102		ret = dquot_alloc_space_nodirty(inode,
7103				       ocfs2_clusters_to_bytes(osb->sb, 1));
7104		if (ret)
7105			goto out_commit;
7106		did_quota = 1;
7107
7108		data_ac->ac_resv = &oi->ip_la_data_resv;
7109
7110		ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
7111					   &num);
7112		if (ret) {
7113			mlog_errno(ret);
7114			goto out_commit;
7115		}
7116
7117		/*
7118		 * Save two copies, one for insert, and one that can
7119		 * be changed by ocfs2_map_and_dirty_page() below.
7120		 */
7121		block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7122
7123		/*
7124		 * Non sparse file systems zero on extend, so no need
7125		 * to do that now.
7126		 */
7127		if (!ocfs2_sparse_alloc(osb) &&
7128		    PAGE_SIZE < osb->s_clustersize)
7129			end = PAGE_SIZE;
7130
7131		ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages);
7132		if (ret) {
7133			mlog_errno(ret);
7134			need_free = 1;
7135			goto out_commit;
7136		}
7137
7138		/*
7139		 * This should populate the 1st page for us and mark
7140		 * it up to date.
7141		 */
7142		ret = ocfs2_read_inline_data(inode, pages[0], di_bh);
7143		if (ret) {
7144			mlog_errno(ret);
7145			need_free = 1;
7146			goto out_unlock;
7147		}
7148
7149		page_end = PAGE_SIZE;
7150		if (PAGE_SIZE > osb->s_clustersize)
7151			page_end = osb->s_clustersize;
7152
7153		for (i = 0; i < num_pages; i++)
7154			ocfs2_map_and_dirty_page(inode, handle, 0, page_end,
7155						 pages[i], i > 0, &phys);
7156	}
7157
7158	spin_lock(&oi->ip_lock);
7159	oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7160	di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7161	spin_unlock(&oi->ip_lock);
7162
7163	ocfs2_update_inode_fsync_trans(handle, inode, 1);
7164	ocfs2_dinode_new_extent_list(inode, di);
7165
7166	ocfs2_journal_dirty(handle, di_bh);
7167
7168	if (has_data) {
7169		/*
7170		 * An error at this point should be extremely rare. If
7171		 * this proves to be false, we could always re-build
7172		 * the in-inode data from our pages.
7173		 */
7174		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7175		ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7176		if (ret) {
7177			mlog_errno(ret);
7178			need_free = 1;
7179			goto out_unlock;
7180		}
7181
7182		inode->i_blocks = ocfs2_inode_sector_count(inode);
7183	}
7184
7185out_unlock:
7186	if (pages)
7187		ocfs2_unlock_and_free_pages(pages, num_pages);
7188
7189out_commit:
7190	if (ret < 0 && did_quota)
7191		dquot_free_space_nodirty(inode,
7192					  ocfs2_clusters_to_bytes(osb->sb, 1));
7193
7194	if (need_free) {
7195		if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7196			ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7197					bit_off, num);
7198		else
7199			ocfs2_free_clusters(handle,
7200					data_ac->ac_inode,
7201					data_ac->ac_bh,
7202					ocfs2_clusters_to_blocks(osb->sb, bit_off),
7203					num);
7204	}
7205
7206	ocfs2_commit_trans(osb, handle);
7207
7208out:
7209	if (data_ac)
7210		ocfs2_free_alloc_context(data_ac);
7211free_pages:
7212	kfree(pages);
7213	return ret;
7214}
7215
7216/*
7217 * It is expected, that by the time you call this function,
7218 * inode->i_size and fe->i_size have been adjusted.
7219 *
7220 * WARNING: This will kfree the truncate context
7221 */
7222int ocfs2_commit_truncate(struct ocfs2_super *osb,
7223			  struct inode *inode,
7224			  struct buffer_head *di_bh)
7225{
7226	int status = 0, i, flags = 0;
7227	u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7228	u64 blkno = 0;
7229	struct ocfs2_extent_list *el;
7230	struct ocfs2_extent_rec *rec;
7231	struct ocfs2_path *path = NULL;
7232	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7233	struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7234	u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7235	struct ocfs2_extent_tree et;
7236	struct ocfs2_cached_dealloc_ctxt dealloc;
7237	struct ocfs2_refcount_tree *ref_tree = NULL;
7238
7239	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7240	ocfs2_init_dealloc_ctxt(&dealloc);
7241
7242	new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7243						     i_size_read(inode));
7244
7245	path = ocfs2_new_path(di_bh, &di->id2.i_list,
7246			      ocfs2_journal_access_di);
7247	if (!path) {
7248		status = -ENOMEM;
7249		mlog_errno(status);
7250		goto bail;
7251	}
7252
7253	ocfs2_extent_map_trunc(inode, new_highest_cpos);
7254
7255start:
7256	/*
7257	 * Check that we still have allocation to delete.
7258	 */
7259	if (OCFS2_I(inode)->ip_clusters == 0) {
7260		status = 0;
7261		goto bail;
7262	}
7263
7264	/*
7265	 * Truncate always works against the rightmost tree branch.
7266	 */
7267	status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7268	if (status) {
7269		mlog_errno(status);
7270		goto bail;
7271	}
7272
7273	trace_ocfs2_commit_truncate(
7274		(unsigned long long)OCFS2_I(inode)->ip_blkno,
7275		new_highest_cpos,
7276		OCFS2_I(inode)->ip_clusters,
7277		path->p_tree_depth);
7278
7279	/*
7280	 * By now, el will point to the extent list on the bottom most
7281	 * portion of this tree. Only the tail record is considered in
7282	 * each pass.
7283	 *
7284	 * We handle the following cases, in order:
7285	 * - empty extent: delete the remaining branch
7286	 * - remove the entire record
7287	 * - remove a partial record
7288	 * - no record needs to be removed (truncate has completed)
7289	 */
7290	el = path_leaf_el(path);
7291	if (le16_to_cpu(el->l_next_free_rec) == 0) {
7292		ocfs2_error(inode->i_sb,
7293			    "Inode %llu has empty extent block at %llu\n",
7294			    (unsigned long long)OCFS2_I(inode)->ip_blkno,
7295			    (unsigned long long)path_leaf_bh(path)->b_blocknr);
7296		status = -EROFS;
7297		goto bail;
7298	}
7299
7300	i = le16_to_cpu(el->l_next_free_rec) - 1;
7301	rec = &el->l_recs[i];
7302	flags = rec->e_flags;
7303	range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7304
7305	if (i == 0 && ocfs2_is_empty_extent(rec)) {
7306		/*
7307		 * Lower levels depend on this never happening, but it's best
7308		 * to check it up here before changing the tree.
7309		*/
7310		if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7311			mlog(ML_ERROR, "Inode %lu has an empty "
7312				    "extent record, depth %u\n", inode->i_ino,
7313				    le16_to_cpu(root_el->l_tree_depth));
7314			status = ocfs2_remove_rightmost_empty_extent(osb,
7315					&et, path, &dealloc);
7316			if (status) {
7317				mlog_errno(status);
7318				goto bail;
7319			}
7320
7321			ocfs2_reinit_path(path, 1);
7322			goto start;
7323		} else {
7324			trunc_cpos = le32_to_cpu(rec->e_cpos);
7325			trunc_len = 0;
7326			blkno = 0;
7327		}
7328	} else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7329		/*
7330		 * Truncate entire record.
7331		 */
7332		trunc_cpos = le32_to_cpu(rec->e_cpos);
7333		trunc_len = ocfs2_rec_clusters(el, rec);
7334		blkno = le64_to_cpu(rec->e_blkno);
7335	} else if (range > new_highest_cpos) {
7336		/*
7337		 * Partial truncate. it also should be
7338		 * the last truncate we're doing.
7339		 */
7340		trunc_cpos = new_highest_cpos;
7341		trunc_len = range - new_highest_cpos;
7342		coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7343		blkno = le64_to_cpu(rec->e_blkno) +
7344				ocfs2_clusters_to_blocks(inode->i_sb, coff);
7345	} else {
7346		/*
7347		 * Truncate completed, leave happily.
7348		 */
7349		status = 0;
7350		goto bail;
7351	}
7352
7353	phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7354
7355	if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7356		status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7357				&ref_tree, NULL);
7358		if (status) {
7359			mlog_errno(status);
7360			goto bail;
7361		}
7362	}
7363
7364	status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7365					  phys_cpos, trunc_len, flags, &dealloc,
7366					  refcount_loc, true);
7367	if (status < 0) {
7368		mlog_errno(status);
7369		goto bail;
7370	}
7371
7372	ocfs2_reinit_path(path, 1);
7373
7374	/*
7375	 * The check above will catch the case where we've truncated
7376	 * away all allocation.
7377	 */
7378	goto start;
7379
7380bail:
7381	if (ref_tree)
7382		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7383
7384	ocfs2_schedule_truncate_log_flush(osb, 1);
7385
7386	ocfs2_run_deallocs(osb, &dealloc);
7387
7388	ocfs2_free_path(path);
7389
7390	return status;
7391}
7392
7393/*
7394 * 'start' is inclusive, 'end' is not.
7395 */
7396int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7397			  unsigned int start, unsigned int end, int trunc)
7398{
7399	int ret;
7400	unsigned int numbytes;
7401	handle_t *handle;
7402	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7403	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7404	struct ocfs2_inline_data *idata = &di->id2.i_data;
7405
 
 
 
 
7406	if (end > i_size_read(inode))
7407		end = i_size_read(inode);
7408
7409	BUG_ON(start > end);
7410
7411	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7412	    !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7413	    !ocfs2_supports_inline_data(osb)) {
7414		ocfs2_error(inode->i_sb,
7415			    "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7416			    (unsigned long long)OCFS2_I(inode)->ip_blkno,
7417			    le16_to_cpu(di->i_dyn_features),
7418			    OCFS2_I(inode)->ip_dyn_features,
7419			    osb->s_feature_incompat);
7420		ret = -EROFS;
7421		goto out;
7422	}
7423
7424	handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7425	if (IS_ERR(handle)) {
7426		ret = PTR_ERR(handle);
7427		mlog_errno(ret);
7428		goto out;
7429	}
7430
7431	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7432				      OCFS2_JOURNAL_ACCESS_WRITE);
7433	if (ret) {
7434		mlog_errno(ret);
7435		goto out_commit;
7436	}
7437
7438	numbytes = end - start;
7439	memset(idata->id_data + start, 0, numbytes);
7440
7441	/*
7442	 * No need to worry about the data page here - it's been
7443	 * truncated already and inline data doesn't need it for
7444	 * pushing zero's to disk, so we'll let readpage pick it up
7445	 * later.
7446	 */
7447	if (trunc) {
7448		i_size_write(inode, start);
7449		di->i_size = cpu_to_le64(start);
7450	}
7451
7452	inode->i_blocks = ocfs2_inode_sector_count(inode);
7453	inode->i_ctime = inode->i_mtime = current_time(inode);
7454
7455	di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7456	di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7457
7458	ocfs2_update_inode_fsync_trans(handle, inode, 1);
7459	ocfs2_journal_dirty(handle, di_bh);
7460
7461out_commit:
7462	ocfs2_commit_trans(osb, handle);
7463
7464out:
7465	return ret;
7466}
7467
7468static int ocfs2_trim_extent(struct super_block *sb,
7469			     struct ocfs2_group_desc *gd,
7470			     u64 group, u32 start, u32 count)
7471{
7472	u64 discard, bcount;
7473	struct ocfs2_super *osb = OCFS2_SB(sb);
7474
7475	bcount = ocfs2_clusters_to_blocks(sb, count);
7476	discard = ocfs2_clusters_to_blocks(sb, start);
7477
7478	/*
7479	 * For the first cluster group, the gd->bg_blkno is not at the start
7480	 * of the group, but at an offset from the start. If we add it while
7481	 * calculating discard for first group, we will wrongly start fstrim a
7482	 * few blocks after the desried start block and the range can cross
7483	 * over into the next cluster group. So, add it only if this is not
7484	 * the first cluster group.
7485	 */
7486	if (group != osb->first_cluster_group_blkno)
7487		discard += le64_to_cpu(gd->bg_blkno);
7488
7489	trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7490
7491	return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7492}
7493
7494static int ocfs2_trim_group(struct super_block *sb,
7495			    struct ocfs2_group_desc *gd, u64 group,
7496			    u32 start, u32 max, u32 minbits)
7497{
7498	int ret = 0, count = 0, next;
7499	void *bitmap = gd->bg_bitmap;
7500
7501	if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7502		return 0;
7503
7504	trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7505			       start, max, minbits);
7506
7507	while (start < max) {
7508		start = ocfs2_find_next_zero_bit(bitmap, max, start);
7509		if (start >= max)
7510			break;
7511		next = ocfs2_find_next_bit(bitmap, max, start);
7512
7513		if ((next - start) >= minbits) {
7514			ret = ocfs2_trim_extent(sb, gd, group,
7515						start, next - start);
7516			if (ret < 0) {
7517				mlog_errno(ret);
7518				break;
7519			}
7520			count += next - start;
7521		}
7522		start = next + 1;
7523
7524		if (fatal_signal_pending(current)) {
7525			count = -ERESTARTSYS;
7526			break;
7527		}
7528
7529		if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7530			break;
7531	}
7532
7533	if (ret < 0)
7534		count = ret;
7535
7536	return count;
7537}
7538
7539static
7540int ocfs2_trim_mainbm(struct super_block *sb, struct fstrim_range *range)
7541{
7542	struct ocfs2_super *osb = OCFS2_SB(sb);
7543	u64 start, len, trimmed = 0, first_group, last_group = 0, group = 0;
7544	int ret, cnt;
7545	u32 first_bit, last_bit, minlen;
7546	struct buffer_head *main_bm_bh = NULL;
7547	struct inode *main_bm_inode = NULL;
7548	struct buffer_head *gd_bh = NULL;
7549	struct ocfs2_dinode *main_bm;
7550	struct ocfs2_group_desc *gd = NULL;
7551
7552	start = range->start >> osb->s_clustersize_bits;
7553	len = range->len >> osb->s_clustersize_bits;
7554	minlen = range->minlen >> osb->s_clustersize_bits;
7555
7556	if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7557		return -EINVAL;
7558
7559	trace_ocfs2_trim_mainbm(start, len, minlen);
7560
7561next_group:
7562	main_bm_inode = ocfs2_get_system_file_inode(osb,
7563						    GLOBAL_BITMAP_SYSTEM_INODE,
7564						    OCFS2_INVALID_SLOT);
7565	if (!main_bm_inode) {
7566		ret = -EIO;
7567		mlog_errno(ret);
7568		goto out;
7569	}
7570
7571	inode_lock(main_bm_inode);
7572
7573	ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7574	if (ret < 0) {
7575		mlog_errno(ret);
7576		goto out_mutex;
7577	}
7578	main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7579
7580	/*
7581	 * Do some check before trim the first group.
7582	 */
7583	if (!group) {
7584		if (start >= le32_to_cpu(main_bm->i_clusters)) {
7585			ret = -EINVAL;
7586			goto out_unlock;
7587		}
7588
7589		if (start + len > le32_to_cpu(main_bm->i_clusters))
7590			len = le32_to_cpu(main_bm->i_clusters) - start;
7591
7592		/*
7593		 * Determine first and last group to examine based on
7594		 * start and len
7595		 */
7596		first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7597		if (first_group == osb->first_cluster_group_blkno)
7598			first_bit = start;
7599		else
7600			first_bit = start - ocfs2_blocks_to_clusters(sb,
7601								first_group);
7602		last_group = ocfs2_which_cluster_group(main_bm_inode,
7603						       start + len - 1);
7604		group = first_group;
7605	}
7606
7607	do {
7608		if (first_bit + len >= osb->bitmap_cpg)
7609			last_bit = osb->bitmap_cpg;
7610		else
7611			last_bit = first_bit + len;
7612
7613		ret = ocfs2_read_group_descriptor(main_bm_inode,
7614						  main_bm, group,
7615						  &gd_bh);
7616		if (ret < 0) {
7617			mlog_errno(ret);
7618			break;
7619		}
7620
7621		gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7622		cnt = ocfs2_trim_group(sb, gd, group,
7623				       first_bit, last_bit, minlen);
7624		brelse(gd_bh);
7625		gd_bh = NULL;
7626		if (cnt < 0) {
7627			ret = cnt;
7628			mlog_errno(ret);
7629			break;
7630		}
7631
7632		trimmed += cnt;
7633		len -= osb->bitmap_cpg - first_bit;
7634		first_bit = 0;
7635		if (group == osb->first_cluster_group_blkno)
7636			group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7637		else
7638			group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7639	} while (0);
7640
7641out_unlock:
7642	ocfs2_inode_unlock(main_bm_inode, 0);
7643	brelse(main_bm_bh);
7644	main_bm_bh = NULL;
7645out_mutex:
7646	inode_unlock(main_bm_inode);
7647	iput(main_bm_inode);
7648
7649	/*
7650	 * If all the groups trim are not done or failed, but we should release
7651	 * main_bm related locks for avoiding the current IO starve, then go to
7652	 * trim the next group
7653	 */
7654	if (ret >= 0 && group <= last_group)
 
7655		goto next_group;
 
7656out:
7657	range->len = trimmed * sb->s_blocksize;
7658	return ret;
7659}
7660
7661int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7662{
7663	int ret;
7664	struct ocfs2_super *osb = OCFS2_SB(sb);
7665	struct ocfs2_trim_fs_info info, *pinfo = NULL;
7666
7667	ocfs2_trim_fs_lock_res_init(osb);
7668
7669	trace_ocfs2_trim_fs(range->start, range->len, range->minlen);
7670
7671	ret = ocfs2_trim_fs_lock(osb, NULL, 1);
7672	if (ret < 0) {
7673		if (ret != -EAGAIN) {
7674			mlog_errno(ret);
7675			ocfs2_trim_fs_lock_res_uninit(osb);
7676			return ret;
7677		}
7678
7679		mlog(ML_NOTICE, "Wait for trim on device (%s) to "
7680		     "finish, which is running from another node.\n",
7681		     osb->dev_str);
7682		ret = ocfs2_trim_fs_lock(osb, &info, 0);
7683		if (ret < 0) {
7684			mlog_errno(ret);
7685			ocfs2_trim_fs_lock_res_uninit(osb);
7686			return ret;
7687		}
7688
7689		if (info.tf_valid && info.tf_success &&
7690		    info.tf_start == range->start &&
7691		    info.tf_len == range->len &&
7692		    info.tf_minlen == range->minlen) {
7693			/* Avoid sending duplicated trim to a shared device */
7694			mlog(ML_NOTICE, "The same trim on device (%s) was "
7695			     "just done from node (%u), return.\n",
7696			     osb->dev_str, info.tf_nodenum);
7697			range->len = info.tf_trimlen;
7698			goto out;
7699		}
7700	}
7701
7702	info.tf_nodenum = osb->node_num;
7703	info.tf_start = range->start;
7704	info.tf_len = range->len;
7705	info.tf_minlen = range->minlen;
7706
7707	ret = ocfs2_trim_mainbm(sb, range);
7708
7709	info.tf_trimlen = range->len;
7710	info.tf_success = (ret < 0 ? 0 : 1);
7711	pinfo = &info;
7712out:
7713	ocfs2_trim_fs_unlock(osb, pinfo);
7714	ocfs2_trim_fs_lock_res_uninit(osb);
7715	return ret;
7716}
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
 
 
   3 * alloc.c
   4 *
   5 * Extent allocs and frees
   6 *
   7 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   8 */
   9
  10#include <linux/fs.h>
  11#include <linux/types.h>
  12#include <linux/slab.h>
  13#include <linux/highmem.h>
  14#include <linux/swap.h>
  15#include <linux/quotaops.h>
  16#include <linux/blkdev.h>
  17#include <linux/sched/signal.h>
  18
  19#include <cluster/masklog.h>
  20
  21#include "ocfs2.h"
  22
  23#include "alloc.h"
  24#include "aops.h"
  25#include "blockcheck.h"
  26#include "dlmglue.h"
  27#include "extent_map.h"
  28#include "inode.h"
  29#include "journal.h"
  30#include "localalloc.h"
  31#include "suballoc.h"
  32#include "sysfile.h"
  33#include "file.h"
  34#include "super.h"
  35#include "uptodate.h"
  36#include "xattr.h"
  37#include "refcounttree.h"
  38#include "ocfs2_trace.h"
  39
  40#include "buffer_head_io.h"
  41
  42enum ocfs2_contig_type {
  43	CONTIG_NONE = 0,
  44	CONTIG_LEFT,
  45	CONTIG_RIGHT,
  46	CONTIG_LEFTRIGHT,
  47};
  48
  49static enum ocfs2_contig_type
  50	ocfs2_extent_rec_contig(struct super_block *sb,
  51				struct ocfs2_extent_rec *ext,
  52				struct ocfs2_extent_rec *insert_rec);
  53/*
  54 * Operations for a specific extent tree type.
  55 *
  56 * To implement an on-disk btree (extent tree) type in ocfs2, add
  57 * an ocfs2_extent_tree_operations structure and the matching
  58 * ocfs2_init_<thingy>_extent_tree() function.  That's pretty much it
  59 * for the allocation portion of the extent tree.
  60 */
  61struct ocfs2_extent_tree_operations {
  62	/*
  63	 * last_eb_blk is the block number of the right most leaf extent
  64	 * block.  Most on-disk structures containing an extent tree store
  65	 * this value for fast access.  The ->eo_set_last_eb_blk() and
  66	 * ->eo_get_last_eb_blk() operations access this value.  They are
  67	 *  both required.
  68	 */
  69	void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
  70				   u64 blkno);
  71	u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);
  72
  73	/*
  74	 * The on-disk structure usually keeps track of how many total
  75	 * clusters are stored in this extent tree.  This function updates
  76	 * that value.  new_clusters is the delta, and must be
  77	 * added to the total.  Required.
  78	 */
  79	void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
  80				   u32 new_clusters);
  81
  82	/*
  83	 * If this extent tree is supported by an extent map, insert
  84	 * a record into the map.
  85	 */
  86	void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
  87				     struct ocfs2_extent_rec *rec);
  88
  89	/*
  90	 * If this extent tree is supported by an extent map, truncate the
  91	 * map to clusters,
  92	 */
  93	void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
  94				       u32 clusters);
  95
  96	/*
  97	 * If ->eo_insert_check() exists, it is called before rec is
  98	 * inserted into the extent tree.  It is optional.
  99	 */
 100	int (*eo_insert_check)(struct ocfs2_extent_tree *et,
 101			       struct ocfs2_extent_rec *rec);
 102	int (*eo_sanity_check)(struct ocfs2_extent_tree *et);
 103
 104	/*
 105	 * --------------------------------------------------------------
 106	 * The remaining are internal to ocfs2_extent_tree and don't have
 107	 * accessor functions
 108	 */
 109
 110	/*
 111	 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
 112	 * It is required.
 113	 */
 114	void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);
 115
 116	/*
 117	 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
 118	 * it exists.  If it does not, et->et_max_leaf_clusters is set
 119	 * to 0 (unlimited).  Optional.
 120	 */
 121	void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);
 122
 123	/*
 124	 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
 125	 * are contiguous or not. Optional. Don't need to set it if use
 126	 * ocfs2_extent_rec as the tree leaf.
 127	 */
 128	enum ocfs2_contig_type
 129		(*eo_extent_contig)(struct ocfs2_extent_tree *et,
 130				    struct ocfs2_extent_rec *ext,
 131				    struct ocfs2_extent_rec *insert_rec);
 132};
 133
 134
 135/*
 136 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
 137 * in the methods.
 138 */
 139static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
 140static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
 141					 u64 blkno);
 142static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
 143					 u32 clusters);
 144static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
 145					   struct ocfs2_extent_rec *rec);
 146static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
 147					     u32 clusters);
 148static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
 149				     struct ocfs2_extent_rec *rec);
 150static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
 151static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);
 152
 153static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
 154					struct ocfs2_extent_tree *et,
 155					struct buffer_head **new_eb_bh,
 156					int blk_wanted, int *blk_given);
 157static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et);
 158
 159static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
 160	.eo_set_last_eb_blk	= ocfs2_dinode_set_last_eb_blk,
 161	.eo_get_last_eb_blk	= ocfs2_dinode_get_last_eb_blk,
 162	.eo_update_clusters	= ocfs2_dinode_update_clusters,
 163	.eo_extent_map_insert	= ocfs2_dinode_extent_map_insert,
 164	.eo_extent_map_truncate	= ocfs2_dinode_extent_map_truncate,
 165	.eo_insert_check	= ocfs2_dinode_insert_check,
 166	.eo_sanity_check	= ocfs2_dinode_sanity_check,
 167	.eo_fill_root_el	= ocfs2_dinode_fill_root_el,
 168};
 169
 170static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
 171					 u64 blkno)
 172{
 173	struct ocfs2_dinode *di = et->et_object;
 174
 175	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 176	di->i_last_eb_blk = cpu_to_le64(blkno);
 177}
 178
 179static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
 180{
 181	struct ocfs2_dinode *di = et->et_object;
 182
 183	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 184	return le64_to_cpu(di->i_last_eb_blk);
 185}
 186
 187static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
 188					 u32 clusters)
 189{
 190	struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
 191	struct ocfs2_dinode *di = et->et_object;
 192
 193	le32_add_cpu(&di->i_clusters, clusters);
 194	spin_lock(&oi->ip_lock);
 195	oi->ip_clusters = le32_to_cpu(di->i_clusters);
 196	spin_unlock(&oi->ip_lock);
 197}
 198
 199static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
 200					   struct ocfs2_extent_rec *rec)
 201{
 202	struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
 203
 204	ocfs2_extent_map_insert_rec(inode, rec);
 205}
 206
 207static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
 208					     u32 clusters)
 209{
 210	struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;
 211
 212	ocfs2_extent_map_trunc(inode, clusters);
 213}
 214
 215static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
 216				     struct ocfs2_extent_rec *rec)
 217{
 218	struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
 219	struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);
 220
 221	BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
 222	mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
 223			(oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
 224			"Device %s, asking for sparse allocation: inode %llu, "
 225			"cpos %u, clusters %u\n",
 226			osb->dev_str,
 227			(unsigned long long)oi->ip_blkno,
 228			rec->e_cpos, oi->ip_clusters);
 229
 230	return 0;
 231}
 232
 233static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
 234{
 235	struct ocfs2_dinode *di = et->et_object;
 236
 237	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
 238	BUG_ON(!OCFS2_IS_VALID_DINODE(di));
 239
 240	return 0;
 241}
 242
 243static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
 244{
 245	struct ocfs2_dinode *di = et->et_object;
 246
 247	et->et_root_el = &di->id2.i_list;
 248}
 249
 250
 251static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
 252{
 253	struct ocfs2_xattr_value_buf *vb = et->et_object;
 254
 255	et->et_root_el = &vb->vb_xv->xr_list;
 256}
 257
 258static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
 259					      u64 blkno)
 260{
 261	struct ocfs2_xattr_value_buf *vb = et->et_object;
 262
 263	vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
 264}
 265
 266static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
 267{
 268	struct ocfs2_xattr_value_buf *vb = et->et_object;
 269
 270	return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
 271}
 272
 273static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
 274					      u32 clusters)
 275{
 276	struct ocfs2_xattr_value_buf *vb = et->et_object;
 277
 278	le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
 279}
 280
 281static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
 282	.eo_set_last_eb_blk	= ocfs2_xattr_value_set_last_eb_blk,
 283	.eo_get_last_eb_blk	= ocfs2_xattr_value_get_last_eb_blk,
 284	.eo_update_clusters	= ocfs2_xattr_value_update_clusters,
 285	.eo_fill_root_el	= ocfs2_xattr_value_fill_root_el,
 286};
 287
 288static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
 289{
 290	struct ocfs2_xattr_block *xb = et->et_object;
 291
 292	et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
 293}
 294
 295static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
 296{
 297	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
 298	et->et_max_leaf_clusters =
 299		ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
 300}
 301
 302static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
 303					     u64 blkno)
 304{
 305	struct ocfs2_xattr_block *xb = et->et_object;
 306	struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
 307
 308	xt->xt_last_eb_blk = cpu_to_le64(blkno);
 309}
 310
 311static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
 312{
 313	struct ocfs2_xattr_block *xb = et->et_object;
 314	struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;
 315
 316	return le64_to_cpu(xt->xt_last_eb_blk);
 317}
 318
 319static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
 320					     u32 clusters)
 321{
 322	struct ocfs2_xattr_block *xb = et->et_object;
 323
 324	le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
 325}
 326
 327static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
 328	.eo_set_last_eb_blk	= ocfs2_xattr_tree_set_last_eb_blk,
 329	.eo_get_last_eb_blk	= ocfs2_xattr_tree_get_last_eb_blk,
 330	.eo_update_clusters	= ocfs2_xattr_tree_update_clusters,
 331	.eo_fill_root_el	= ocfs2_xattr_tree_fill_root_el,
 332	.eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
 333};
 334
 335static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
 336					  u64 blkno)
 337{
 338	struct ocfs2_dx_root_block *dx_root = et->et_object;
 339
 340	dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
 341}
 342
 343static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
 344{
 345	struct ocfs2_dx_root_block *dx_root = et->et_object;
 346
 347	return le64_to_cpu(dx_root->dr_last_eb_blk);
 348}
 349
 350static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
 351					  u32 clusters)
 352{
 353	struct ocfs2_dx_root_block *dx_root = et->et_object;
 354
 355	le32_add_cpu(&dx_root->dr_clusters, clusters);
 356}
 357
 358static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
 359{
 360	struct ocfs2_dx_root_block *dx_root = et->et_object;
 361
 362	BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));
 363
 364	return 0;
 365}
 366
 367static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
 368{
 369	struct ocfs2_dx_root_block *dx_root = et->et_object;
 370
 371	et->et_root_el = &dx_root->dr_list;
 372}
 373
 374static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
 375	.eo_set_last_eb_blk	= ocfs2_dx_root_set_last_eb_blk,
 376	.eo_get_last_eb_blk	= ocfs2_dx_root_get_last_eb_blk,
 377	.eo_update_clusters	= ocfs2_dx_root_update_clusters,
 378	.eo_sanity_check	= ocfs2_dx_root_sanity_check,
 379	.eo_fill_root_el	= ocfs2_dx_root_fill_root_el,
 380};
 381
 382static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
 383{
 384	struct ocfs2_refcount_block *rb = et->et_object;
 385
 386	et->et_root_el = &rb->rf_list;
 387}
 388
 389static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
 390						u64 blkno)
 391{
 392	struct ocfs2_refcount_block *rb = et->et_object;
 393
 394	rb->rf_last_eb_blk = cpu_to_le64(blkno);
 395}
 396
 397static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
 398{
 399	struct ocfs2_refcount_block *rb = et->et_object;
 400
 401	return le64_to_cpu(rb->rf_last_eb_blk);
 402}
 403
 404static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
 405						u32 clusters)
 406{
 407	struct ocfs2_refcount_block *rb = et->et_object;
 408
 409	le32_add_cpu(&rb->rf_clusters, clusters);
 410}
 411
 412static enum ocfs2_contig_type
 413ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
 414				  struct ocfs2_extent_rec *ext,
 415				  struct ocfs2_extent_rec *insert_rec)
 416{
 417	return CONTIG_NONE;
 418}
 419
 420static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
 421	.eo_set_last_eb_blk	= ocfs2_refcount_tree_set_last_eb_blk,
 422	.eo_get_last_eb_blk	= ocfs2_refcount_tree_get_last_eb_blk,
 423	.eo_update_clusters	= ocfs2_refcount_tree_update_clusters,
 424	.eo_fill_root_el	= ocfs2_refcount_tree_fill_root_el,
 425	.eo_extent_contig	= ocfs2_refcount_tree_extent_contig,
 426};
 427
 428static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
 429				     struct ocfs2_caching_info *ci,
 430				     struct buffer_head *bh,
 431				     ocfs2_journal_access_func access,
 432				     void *obj,
 433				     const struct ocfs2_extent_tree_operations *ops)
 434{
 435	et->et_ops = ops;
 436	et->et_root_bh = bh;
 437	et->et_ci = ci;
 438	et->et_root_journal_access = access;
 439	if (!obj)
 440		obj = (void *)bh->b_data;
 441	et->et_object = obj;
 442	et->et_dealloc = NULL;
 443
 444	et->et_ops->eo_fill_root_el(et);
 445	if (!et->et_ops->eo_fill_max_leaf_clusters)
 446		et->et_max_leaf_clusters = 0;
 447	else
 448		et->et_ops->eo_fill_max_leaf_clusters(et);
 449}
 450
 451void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
 452				   struct ocfs2_caching_info *ci,
 453				   struct buffer_head *bh)
 454{
 455	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
 456				 NULL, &ocfs2_dinode_et_ops);
 457}
 458
 459void ocfs2_init_xattr_tree_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_xb,
 464				 NULL, &ocfs2_xattr_tree_et_ops);
 465}
 466
 467void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
 468					struct ocfs2_caching_info *ci,
 469					struct ocfs2_xattr_value_buf *vb)
 470{
 471	__ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
 472				 &ocfs2_xattr_value_et_ops);
 473}
 474
 475void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
 476				    struct ocfs2_caching_info *ci,
 477				    struct buffer_head *bh)
 478{
 479	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
 480				 NULL, &ocfs2_dx_root_et_ops);
 481}
 482
 483void ocfs2_init_refcount_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_rb,
 488				 NULL, &ocfs2_refcount_tree_et_ops);
 489}
 490
 491static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
 492					    u64 new_last_eb_blk)
 493{
 494	et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
 495}
 496
 497static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
 498{
 499	return et->et_ops->eo_get_last_eb_blk(et);
 500}
 501
 502static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
 503					    u32 clusters)
 504{
 505	et->et_ops->eo_update_clusters(et, clusters);
 506}
 507
 508static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
 509					      struct ocfs2_extent_rec *rec)
 510{
 511	if (et->et_ops->eo_extent_map_insert)
 512		et->et_ops->eo_extent_map_insert(et, rec);
 513}
 514
 515static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
 516						u32 clusters)
 517{
 518	if (et->et_ops->eo_extent_map_truncate)
 519		et->et_ops->eo_extent_map_truncate(et, clusters);
 520}
 521
 522static inline int ocfs2_et_root_journal_access(handle_t *handle,
 523					       struct ocfs2_extent_tree *et,
 524					       int type)
 525{
 526	return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
 527					  type);
 528}
 529
 530static inline enum ocfs2_contig_type
 531	ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
 532			       struct ocfs2_extent_rec *rec,
 533			       struct ocfs2_extent_rec *insert_rec)
 534{
 535	if (et->et_ops->eo_extent_contig)
 536		return et->et_ops->eo_extent_contig(et, rec, insert_rec);
 537
 538	return ocfs2_extent_rec_contig(
 539				ocfs2_metadata_cache_get_super(et->et_ci),
 540				rec, insert_rec);
 541}
 542
 543static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
 544					struct ocfs2_extent_rec *rec)
 545{
 546	int ret = 0;
 547
 548	if (et->et_ops->eo_insert_check)
 549		ret = et->et_ops->eo_insert_check(et, rec);
 550	return ret;
 551}
 552
 553static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
 554{
 555	int ret = 0;
 556
 557	if (et->et_ops->eo_sanity_check)
 558		ret = et->et_ops->eo_sanity_check(et);
 559	return ret;
 560}
 561
 562static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
 563					 struct ocfs2_extent_block *eb);
 564static void ocfs2_adjust_rightmost_records(handle_t *handle,
 565					   struct ocfs2_extent_tree *et,
 566					   struct ocfs2_path *path,
 567					   struct ocfs2_extent_rec *insert_rec);
 568/*
 569 * Reset the actual path elements so that we can re-use the structure
 570 * to build another path. Generally, this involves freeing the buffer
 571 * heads.
 572 */
 573void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
 574{
 575	int i, start = 0, depth = 0;
 576	struct ocfs2_path_item *node;
 577
 578	if (keep_root)
 579		start = 1;
 580
 581	for(i = start; i < path_num_items(path); i++) {
 582		node = &path->p_node[i];
 583
 584		brelse(node->bh);
 585		node->bh = NULL;
 586		node->el = NULL;
 587	}
 588
 589	/*
 590	 * Tree depth may change during truncate, or insert. If we're
 591	 * keeping the root extent list, then make sure that our path
 592	 * structure reflects the proper depth.
 593	 */
 594	if (keep_root)
 595		depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
 596	else
 597		path_root_access(path) = NULL;
 598
 599	path->p_tree_depth = depth;
 600}
 601
 602void ocfs2_free_path(struct ocfs2_path *path)
 603{
 604	if (path) {
 605		ocfs2_reinit_path(path, 0);
 606		kfree(path);
 607	}
 608}
 609
 610/*
 611 * All the elements of src into dest. After this call, src could be freed
 612 * without affecting dest.
 613 *
 614 * Both paths should have the same root. Any non-root elements of dest
 615 * will be freed.
 616 */
 617static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
 618{
 619	int i;
 620
 621	BUG_ON(path_root_bh(dest) != path_root_bh(src));
 622	BUG_ON(path_root_el(dest) != path_root_el(src));
 623	BUG_ON(path_root_access(dest) != path_root_access(src));
 624
 625	ocfs2_reinit_path(dest, 1);
 626
 627	for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
 628		dest->p_node[i].bh = src->p_node[i].bh;
 629		dest->p_node[i].el = src->p_node[i].el;
 630
 631		if (dest->p_node[i].bh)
 632			get_bh(dest->p_node[i].bh);
 633	}
 634}
 635
 636/*
 637 * Make the *dest path the same as src and re-initialize src path to
 638 * have a root only.
 639 */
 640static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
 641{
 642	int i;
 643
 644	BUG_ON(path_root_bh(dest) != path_root_bh(src));
 645	BUG_ON(path_root_access(dest) != path_root_access(src));
 646
 647	for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
 648		brelse(dest->p_node[i].bh);
 649
 650		dest->p_node[i].bh = src->p_node[i].bh;
 651		dest->p_node[i].el = src->p_node[i].el;
 652
 653		src->p_node[i].bh = NULL;
 654		src->p_node[i].el = NULL;
 655	}
 656}
 657
 658/*
 659 * Insert an extent block at given index.
 660 *
 661 * This will not take an additional reference on eb_bh.
 662 */
 663static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
 664					struct buffer_head *eb_bh)
 665{
 666	struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;
 667
 668	/*
 669	 * Right now, no root bh is an extent block, so this helps
 670	 * catch code errors with dinode trees. The assertion can be
 671	 * safely removed if we ever need to insert extent block
 672	 * structures at the root.
 673	 */
 674	BUG_ON(index == 0);
 675
 676	path->p_node[index].bh = eb_bh;
 677	path->p_node[index].el = &eb->h_list;
 678}
 679
 680static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
 681					 struct ocfs2_extent_list *root_el,
 682					 ocfs2_journal_access_func access)
 683{
 684	struct ocfs2_path *path;
 685
 686	BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);
 687
 688	path = kzalloc(sizeof(*path), GFP_NOFS);
 689	if (path) {
 690		path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
 691		get_bh(root_bh);
 692		path_root_bh(path) = root_bh;
 693		path_root_el(path) = root_el;
 694		path_root_access(path) = access;
 695	}
 696
 697	return path;
 698}
 699
 700struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
 701{
 702	return ocfs2_new_path(path_root_bh(path), path_root_el(path),
 703			      path_root_access(path));
 704}
 705
 706struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
 707{
 708	return ocfs2_new_path(et->et_root_bh, et->et_root_el,
 709			      et->et_root_journal_access);
 710}
 711
 712/*
 713 * Journal the buffer at depth idx.  All idx>0 are extent_blocks,
 714 * otherwise it's the root_access function.
 715 *
 716 * I don't like the way this function's name looks next to
 717 * ocfs2_journal_access_path(), but I don't have a better one.
 718 */
 719int ocfs2_path_bh_journal_access(handle_t *handle,
 720				 struct ocfs2_caching_info *ci,
 721				 struct ocfs2_path *path,
 722				 int idx)
 723{
 724	ocfs2_journal_access_func access = path_root_access(path);
 725
 726	if (!access)
 727		access = ocfs2_journal_access;
 728
 729	if (idx)
 730		access = ocfs2_journal_access_eb;
 731
 732	return access(handle, ci, path->p_node[idx].bh,
 733		      OCFS2_JOURNAL_ACCESS_WRITE);
 734}
 735
 736/*
 737 * Convenience function to journal all components in a path.
 738 */
 739int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
 740			      handle_t *handle,
 741			      struct ocfs2_path *path)
 742{
 743	int i, ret = 0;
 744
 745	if (!path)
 746		goto out;
 747
 748	for(i = 0; i < path_num_items(path); i++) {
 749		ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
 750		if (ret < 0) {
 751			mlog_errno(ret);
 752			goto out;
 753		}
 754	}
 755
 756out:
 757	return ret;
 758}
 759
 760/*
 761 * Return the index of the extent record which contains cluster #v_cluster.
 762 * -1 is returned if it was not found.
 763 *
 764 * Should work fine on interior and exterior nodes.
 765 */
 766int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
 767{
 768	int ret = -1;
 769	int i;
 770	struct ocfs2_extent_rec *rec;
 771	u32 rec_end, rec_start, clusters;
 772
 773	for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
 774		rec = &el->l_recs[i];
 775
 776		rec_start = le32_to_cpu(rec->e_cpos);
 777		clusters = ocfs2_rec_clusters(el, rec);
 778
 779		rec_end = rec_start + clusters;
 780
 781		if (v_cluster >= rec_start && v_cluster < rec_end) {
 782			ret = i;
 783			break;
 784		}
 785	}
 786
 787	return ret;
 788}
 789
 790/*
 791 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
 792 * ocfs2_extent_rec_contig only work properly against leaf nodes!
 793 */
 794static int ocfs2_block_extent_contig(struct super_block *sb,
 795				     struct ocfs2_extent_rec *ext,
 796				     u64 blkno)
 797{
 798	u64 blk_end = le64_to_cpu(ext->e_blkno);
 799
 800	blk_end += ocfs2_clusters_to_blocks(sb,
 801				    le16_to_cpu(ext->e_leaf_clusters));
 802
 803	return blkno == blk_end;
 804}
 805
 806static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
 807				  struct ocfs2_extent_rec *right)
 808{
 809	u32 left_range;
 810
 811	left_range = le32_to_cpu(left->e_cpos) +
 812		le16_to_cpu(left->e_leaf_clusters);
 813
 814	return (left_range == le32_to_cpu(right->e_cpos));
 815}
 816
 817static enum ocfs2_contig_type
 818	ocfs2_extent_rec_contig(struct super_block *sb,
 819				struct ocfs2_extent_rec *ext,
 820				struct ocfs2_extent_rec *insert_rec)
 821{
 822	u64 blkno = le64_to_cpu(insert_rec->e_blkno);
 823
 824	/*
 825	 * Refuse to coalesce extent records with different flag
 826	 * fields - we don't want to mix unwritten extents with user
 827	 * data.
 828	 */
 829	if (ext->e_flags != insert_rec->e_flags)
 830		return CONTIG_NONE;
 831
 832	if (ocfs2_extents_adjacent(ext, insert_rec) &&
 833	    ocfs2_block_extent_contig(sb, ext, blkno))
 834			return CONTIG_RIGHT;
 835
 836	blkno = le64_to_cpu(ext->e_blkno);
 837	if (ocfs2_extents_adjacent(insert_rec, ext) &&
 838	    ocfs2_block_extent_contig(sb, insert_rec, blkno))
 839		return CONTIG_LEFT;
 840
 841	return CONTIG_NONE;
 842}
 843
 844/*
 845 * NOTE: We can have pretty much any combination of contiguousness and
 846 * appending.
 847 *
 848 * The usefulness of APPEND_TAIL is more in that it lets us know that
 849 * we'll have to update the path to that leaf.
 850 */
 851enum ocfs2_append_type {
 852	APPEND_NONE = 0,
 853	APPEND_TAIL,
 854};
 855
 856enum ocfs2_split_type {
 857	SPLIT_NONE = 0,
 858	SPLIT_LEFT,
 859	SPLIT_RIGHT,
 860};
 861
 862struct ocfs2_insert_type {
 863	enum ocfs2_split_type	ins_split;
 864	enum ocfs2_append_type	ins_appending;
 865	enum ocfs2_contig_type	ins_contig;
 866	int			ins_contig_index;
 867	int			ins_tree_depth;
 868};
 869
 870struct ocfs2_merge_ctxt {
 871	enum ocfs2_contig_type	c_contig_type;
 872	int			c_has_empty_extent;
 873	int			c_split_covers_rec;
 874};
 875
 876static int ocfs2_validate_extent_block(struct super_block *sb,
 877				       struct buffer_head *bh)
 878{
 879	int rc;
 880	struct ocfs2_extent_block *eb =
 881		(struct ocfs2_extent_block *)bh->b_data;
 882
 883	trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);
 884
 885	BUG_ON(!buffer_uptodate(bh));
 886
 887	/*
 888	 * If the ecc fails, we return the error but otherwise
 889	 * leave the filesystem running.  We know any error is
 890	 * local to this block.
 891	 */
 892	rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
 893	if (rc) {
 894		mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
 895		     (unsigned long long)bh->b_blocknr);
 896		return rc;
 897	}
 898
 899	/*
 900	 * Errors after here are fatal.
 901	 */
 902
 903	if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
 904		rc = ocfs2_error(sb,
 905				 "Extent block #%llu has bad signature %.*s\n",
 906				 (unsigned long long)bh->b_blocknr, 7,
 907				 eb->h_signature);
 908		goto bail;
 909	}
 910
 911	if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
 912		rc = ocfs2_error(sb,
 913				 "Extent block #%llu has an invalid h_blkno of %llu\n",
 914				 (unsigned long long)bh->b_blocknr,
 915				 (unsigned long long)le64_to_cpu(eb->h_blkno));
 916		goto bail;
 917	}
 918
 919	if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation)
 920		rc = ocfs2_error(sb,
 921				 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
 922				 (unsigned long long)bh->b_blocknr,
 923				 le32_to_cpu(eb->h_fs_generation));
 924bail:
 925	return rc;
 926}
 927
 928int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
 929			    struct buffer_head **bh)
 930{
 931	int rc;
 932	struct buffer_head *tmp = *bh;
 933
 934	rc = ocfs2_read_block(ci, eb_blkno, &tmp,
 935			      ocfs2_validate_extent_block);
 936
 937	/* If ocfs2_read_block() got us a new bh, pass it up. */
 938	if (!rc && !*bh)
 939		*bh = tmp;
 940
 941	return rc;
 942}
 943
 944
 945/*
 946 * How many free extents have we got before we need more meta data?
 947 */
 948int ocfs2_num_free_extents(struct ocfs2_extent_tree *et)
 949{
 950	int retval;
 951	struct ocfs2_extent_list *el = NULL;
 952	struct ocfs2_extent_block *eb;
 953	struct buffer_head *eb_bh = NULL;
 954	u64 last_eb_blk = 0;
 955
 956	el = et->et_root_el;
 957	last_eb_blk = ocfs2_et_get_last_eb_blk(et);
 958
 959	if (last_eb_blk) {
 960		retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
 961						 &eb_bh);
 962		if (retval < 0) {
 963			mlog_errno(retval);
 964			goto bail;
 965		}
 966		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 967		el = &eb->h_list;
 968	}
 969
 970	BUG_ON(el->l_tree_depth != 0);
 971
 972	retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
 973bail:
 974	brelse(eb_bh);
 975
 976	trace_ocfs2_num_free_extents(retval);
 977	return retval;
 978}
 979
 980/* expects array to already be allocated
 981 *
 982 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
 983 * l_count for you
 984 */
 985static int ocfs2_create_new_meta_bhs(handle_t *handle,
 986				     struct ocfs2_extent_tree *et,
 987				     int wanted,
 988				     struct ocfs2_alloc_context *meta_ac,
 989				     struct buffer_head *bhs[])
 990{
 991	int count, status, i;
 992	u16 suballoc_bit_start;
 993	u32 num_got;
 994	u64 suballoc_loc, first_blkno;
 995	struct ocfs2_super *osb =
 996		OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
 997	struct ocfs2_extent_block *eb;
 998
 999	count = 0;
1000	while (count < wanted) {
1001		status = ocfs2_claim_metadata(handle,
1002					      meta_ac,
1003					      wanted - count,
1004					      &suballoc_loc,
1005					      &suballoc_bit_start,
1006					      &num_got,
1007					      &first_blkno);
1008		if (status < 0) {
1009			mlog_errno(status);
1010			goto bail;
1011		}
1012
1013		for(i = count;  i < (num_got + count); i++) {
1014			bhs[i] = sb_getblk(osb->sb, first_blkno);
1015			if (bhs[i] == NULL) {
1016				status = -ENOMEM;
1017				mlog_errno(status);
1018				goto bail;
1019			}
1020			ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);
1021
1022			status = ocfs2_journal_access_eb(handle, et->et_ci,
1023							 bhs[i],
1024							 OCFS2_JOURNAL_ACCESS_CREATE);
1025			if (status < 0) {
1026				mlog_errno(status);
1027				goto bail;
1028			}
1029
1030			memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
1031			eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
1032			/* Ok, setup the minimal stuff here. */
1033			strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
1034			eb->h_blkno = cpu_to_le64(first_blkno);
1035			eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
1036			eb->h_suballoc_slot =
1037				cpu_to_le16(meta_ac->ac_alloc_slot);
1038			eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
1039			eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
1040			eb->h_list.l_count =
1041				cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
1042
1043			suballoc_bit_start++;
1044			first_blkno++;
1045
1046			/* We'll also be dirtied by the caller, so
1047			 * this isn't absolutely necessary. */
1048			ocfs2_journal_dirty(handle, bhs[i]);
1049		}
1050
1051		count += num_got;
1052	}
1053
1054	status = 0;
1055bail:
1056	if (status < 0) {
1057		for(i = 0; i < wanted; i++) {
1058			brelse(bhs[i]);
1059			bhs[i] = NULL;
1060		}
 
1061	}
1062	return status;
1063}
1064
1065/*
1066 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
1067 *
1068 * Returns the sum of the rightmost extent rec logical offset and
1069 * cluster count.
1070 *
1071 * ocfs2_add_branch() uses this to determine what logical cluster
1072 * value should be populated into the leftmost new branch records.
1073 *
1074 * ocfs2_shift_tree_depth() uses this to determine the # clusters
1075 * value for the new topmost tree record.
1076 */
1077static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list  *el)
1078{
1079	int i;
1080
1081	i = le16_to_cpu(el->l_next_free_rec) - 1;
1082
1083	return le32_to_cpu(el->l_recs[i].e_cpos) +
1084		ocfs2_rec_clusters(el, &el->l_recs[i]);
1085}
1086
1087/*
1088 * Change range of the branches in the right most path according to the leaf
1089 * extent block's rightmost record.
1090 */
1091static int ocfs2_adjust_rightmost_branch(handle_t *handle,
1092					 struct ocfs2_extent_tree *et)
1093{
1094	int status;
1095	struct ocfs2_path *path = NULL;
1096	struct ocfs2_extent_list *el;
1097	struct ocfs2_extent_rec *rec;
1098
1099	path = ocfs2_new_path_from_et(et);
1100	if (!path) {
1101		status = -ENOMEM;
1102		return status;
1103	}
1104
1105	status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
1106	if (status < 0) {
1107		mlog_errno(status);
1108		goto out;
1109	}
1110
1111	status = ocfs2_extend_trans(handle, path_num_items(path));
1112	if (status < 0) {
1113		mlog_errno(status);
1114		goto out;
1115	}
1116
1117	status = ocfs2_journal_access_path(et->et_ci, handle, path);
1118	if (status < 0) {
1119		mlog_errno(status);
1120		goto out;
1121	}
1122
1123	el = path_leaf_el(path);
1124	rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];
1125
1126	ocfs2_adjust_rightmost_records(handle, et, path, rec);
1127
1128out:
1129	ocfs2_free_path(path);
1130	return status;
1131}
1132
1133/*
1134 * Add an entire tree branch to our inode. eb_bh is the extent block
1135 * to start at, if we don't want to start the branch at the root
1136 * structure.
1137 *
1138 * last_eb_bh is required as we have to update it's next_leaf pointer
1139 * for the new last extent block.
1140 *
1141 * the new branch will be 'empty' in the sense that every block will
1142 * contain a single record with cluster count == 0.
1143 */
1144static int ocfs2_add_branch(handle_t *handle,
1145			    struct ocfs2_extent_tree *et,
1146			    struct buffer_head *eb_bh,
1147			    struct buffer_head **last_eb_bh,
1148			    struct ocfs2_alloc_context *meta_ac)
1149{
1150	int status, new_blocks, i, block_given = 0;
1151	u64 next_blkno, new_last_eb_blk;
1152	struct buffer_head *bh;
1153	struct buffer_head **new_eb_bhs = NULL;
1154	struct ocfs2_extent_block *eb;
1155	struct ocfs2_extent_list  *eb_el;
1156	struct ocfs2_extent_list  *el;
1157	u32 new_cpos, root_end;
1158
1159	BUG_ON(!last_eb_bh || !*last_eb_bh);
1160
1161	if (eb_bh) {
1162		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
1163		el = &eb->h_list;
1164	} else
1165		el = et->et_root_el;
1166
1167	/* we never add a branch to a leaf. */
1168	BUG_ON(!el->l_tree_depth);
1169
1170	new_blocks = le16_to_cpu(el->l_tree_depth);
1171
1172	eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
1173	new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
1174	root_end = ocfs2_sum_rightmost_rec(et->et_root_el);
1175
1176	/*
1177	 * If there is a gap before the root end and the real end
1178	 * of the righmost leaf block, we need to remove the gap
1179	 * between new_cpos and root_end first so that the tree
1180	 * is consistent after we add a new branch(it will start
1181	 * from new_cpos).
1182	 */
1183	if (root_end > new_cpos) {
1184		trace_ocfs2_adjust_rightmost_branch(
1185			(unsigned long long)
1186			ocfs2_metadata_cache_owner(et->et_ci),
1187			root_end, new_cpos);
1188
1189		status = ocfs2_adjust_rightmost_branch(handle, et);
1190		if (status) {
1191			mlog_errno(status);
1192			goto bail;
1193		}
1194	}
1195
1196	/* allocate the number of new eb blocks we need */
1197	new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
1198			     GFP_KERNEL);
1199	if (!new_eb_bhs) {
1200		status = -ENOMEM;
1201		mlog_errno(status);
1202		goto bail;
1203	}
1204
1205	/* Firstyly, try to reuse dealloc since we have already estimated how
1206	 * many extent blocks we may use.
1207	 */
1208	if (!ocfs2_is_dealloc_empty(et)) {
1209		status = ocfs2_reuse_blk_from_dealloc(handle, et,
1210						      new_eb_bhs, new_blocks,
1211						      &block_given);
1212		if (status < 0) {
1213			mlog_errno(status);
1214			goto bail;
1215		}
1216	}
1217
1218	BUG_ON(block_given > new_blocks);
1219
1220	if (block_given < new_blocks) {
1221		BUG_ON(!meta_ac);
1222		status = ocfs2_create_new_meta_bhs(handle, et,
1223						   new_blocks - block_given,
1224						   meta_ac,
1225						   &new_eb_bhs[block_given]);
1226		if (status < 0) {
1227			mlog_errno(status);
1228			goto bail;
1229		}
1230	}
1231
1232	/* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
1233	 * linked with the rest of the tree.
1234	 * conversly, new_eb_bhs[0] is the new bottommost leaf.
1235	 *
1236	 * when we leave the loop, new_last_eb_blk will point to the
1237	 * newest leaf, and next_blkno will point to the topmost extent
1238	 * block. */
1239	next_blkno = new_last_eb_blk = 0;
1240	for(i = 0; i < new_blocks; i++) {
1241		bh = new_eb_bhs[i];
1242		eb = (struct ocfs2_extent_block *) bh->b_data;
1243		/* ocfs2_create_new_meta_bhs() should create it right! */
1244		BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1245		eb_el = &eb->h_list;
1246
1247		status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
1248						 OCFS2_JOURNAL_ACCESS_CREATE);
1249		if (status < 0) {
1250			mlog_errno(status);
1251			goto bail;
1252		}
1253
1254		eb->h_next_leaf_blk = 0;
1255		eb_el->l_tree_depth = cpu_to_le16(i);
1256		eb_el->l_next_free_rec = cpu_to_le16(1);
1257		/*
1258		 * This actually counts as an empty extent as
1259		 * c_clusters == 0
1260		 */
1261		eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
1262		eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
1263		/*
1264		 * eb_el isn't always an interior node, but even leaf
1265		 * nodes want a zero'd flags and reserved field so
1266		 * this gets the whole 32 bits regardless of use.
1267		 */
1268		eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
1269		if (!eb_el->l_tree_depth)
1270			new_last_eb_blk = le64_to_cpu(eb->h_blkno);
1271
1272		ocfs2_journal_dirty(handle, bh);
1273		next_blkno = le64_to_cpu(eb->h_blkno);
1274	}
1275
1276	/* This is a bit hairy. We want to update up to three blocks
1277	 * here without leaving any of them in an inconsistent state
1278	 * in case of error. We don't have to worry about
1279	 * journal_dirty erroring as it won't unless we've aborted the
1280	 * handle (in which case we would never be here) so reserving
1281	 * the write with journal_access is all we need to do. */
1282	status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
1283					 OCFS2_JOURNAL_ACCESS_WRITE);
1284	if (status < 0) {
1285		mlog_errno(status);
1286		goto bail;
1287	}
1288	status = ocfs2_et_root_journal_access(handle, et,
1289					      OCFS2_JOURNAL_ACCESS_WRITE);
1290	if (status < 0) {
1291		mlog_errno(status);
1292		goto bail;
1293	}
1294	if (eb_bh) {
1295		status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
1296						 OCFS2_JOURNAL_ACCESS_WRITE);
1297		if (status < 0) {
1298			mlog_errno(status);
1299			goto bail;
1300		}
1301	}
1302
1303	/* Link the new branch into the rest of the tree (el will
1304	 * either be on the root_bh, or the extent block passed in. */
1305	i = le16_to_cpu(el->l_next_free_rec);
1306	el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
1307	el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
1308	el->l_recs[i].e_int_clusters = 0;
1309	le16_add_cpu(&el->l_next_free_rec, 1);
1310
1311	/* fe needs a new last extent block pointer, as does the
1312	 * next_leaf on the previously last-extent-block. */
1313	ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);
1314
1315	eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
1316	eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);
1317
1318	ocfs2_journal_dirty(handle, *last_eb_bh);
1319	ocfs2_journal_dirty(handle, et->et_root_bh);
1320	if (eb_bh)
1321		ocfs2_journal_dirty(handle, eb_bh);
1322
1323	/*
1324	 * Some callers want to track the rightmost leaf so pass it
1325	 * back here.
1326	 */
1327	brelse(*last_eb_bh);
1328	get_bh(new_eb_bhs[0]);
1329	*last_eb_bh = new_eb_bhs[0];
1330
1331	status = 0;
1332bail:
1333	if (new_eb_bhs) {
1334		for (i = 0; i < new_blocks; i++)
1335			brelse(new_eb_bhs[i]);
1336		kfree(new_eb_bhs);
1337	}
1338
1339	return status;
1340}
1341
1342/*
1343 * adds another level to the allocation tree.
1344 * returns back the new extent block so you can add a branch to it
1345 * after this call.
1346 */
1347static int ocfs2_shift_tree_depth(handle_t *handle,
1348				  struct ocfs2_extent_tree *et,
1349				  struct ocfs2_alloc_context *meta_ac,
1350				  struct buffer_head **ret_new_eb_bh)
1351{
1352	int status, i, block_given = 0;
1353	u32 new_clusters;
1354	struct buffer_head *new_eb_bh = NULL;
1355	struct ocfs2_extent_block *eb;
1356	struct ocfs2_extent_list  *root_el;
1357	struct ocfs2_extent_list  *eb_el;
1358
1359	if (!ocfs2_is_dealloc_empty(et)) {
1360		status = ocfs2_reuse_blk_from_dealloc(handle, et,
1361						      &new_eb_bh, 1,
1362						      &block_given);
1363	} else if (meta_ac) {
1364		status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
1365						   &new_eb_bh);
1366
1367	} else {
1368		BUG();
1369	}
1370
1371	if (status < 0) {
1372		mlog_errno(status);
1373		goto bail;
1374	}
1375
1376	eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
1377	/* ocfs2_create_new_meta_bhs() should create it right! */
1378	BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
1379
1380	eb_el = &eb->h_list;
1381	root_el = et->et_root_el;
1382
1383	status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
1384					 OCFS2_JOURNAL_ACCESS_CREATE);
1385	if (status < 0) {
1386		mlog_errno(status);
1387		goto bail;
1388	}
1389
1390	/* copy the root extent list data into the new extent block */
1391	eb_el->l_tree_depth = root_el->l_tree_depth;
1392	eb_el->l_next_free_rec = root_el->l_next_free_rec;
1393	for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1394		eb_el->l_recs[i] = root_el->l_recs[i];
1395
1396	ocfs2_journal_dirty(handle, new_eb_bh);
1397
1398	status = ocfs2_et_root_journal_access(handle, et,
1399					      OCFS2_JOURNAL_ACCESS_WRITE);
1400	if (status < 0) {
1401		mlog_errno(status);
1402		goto bail;
1403	}
1404
1405	new_clusters = ocfs2_sum_rightmost_rec(eb_el);
1406
1407	/* update root_bh now */
1408	le16_add_cpu(&root_el->l_tree_depth, 1);
1409	root_el->l_recs[0].e_cpos = 0;
1410	root_el->l_recs[0].e_blkno = eb->h_blkno;
1411	root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
1412	for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
1413		memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
1414	root_el->l_next_free_rec = cpu_to_le16(1);
1415
1416	/* If this is our 1st tree depth shift, then last_eb_blk
1417	 * becomes the allocated extent block */
1418	if (root_el->l_tree_depth == cpu_to_le16(1))
1419		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
1420
1421	ocfs2_journal_dirty(handle, et->et_root_bh);
1422
1423	*ret_new_eb_bh = new_eb_bh;
1424	new_eb_bh = NULL;
1425	status = 0;
1426bail:
1427	brelse(new_eb_bh);
1428
1429	return status;
1430}
1431
1432/*
1433 * Should only be called when there is no space left in any of the
1434 * leaf nodes. What we want to do is find the lowest tree depth
1435 * non-leaf extent block with room for new records. There are three
1436 * valid results of this search:
1437 *
1438 * 1) a lowest extent block is found, then we pass it back in
1439 *    *lowest_eb_bh and return '0'
1440 *
1441 * 2) the search fails to find anything, but the root_el has room. We
1442 *    pass NULL back in *lowest_eb_bh, but still return '0'
1443 *
1444 * 3) the search fails to find anything AND the root_el is full, in
1445 *    which case we return > 0
1446 *
1447 * return status < 0 indicates an error.
1448 */
1449static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
1450				    struct buffer_head **target_bh)
1451{
1452	int status = 0, i;
1453	u64 blkno;
1454	struct ocfs2_extent_block *eb;
1455	struct ocfs2_extent_list  *el;
1456	struct buffer_head *bh = NULL;
1457	struct buffer_head *lowest_bh = NULL;
1458
1459	*target_bh = NULL;
1460
1461	el = et->et_root_el;
1462
1463	while(le16_to_cpu(el->l_tree_depth) > 1) {
1464		if (le16_to_cpu(el->l_next_free_rec) == 0) {
1465			status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1466					"Owner %llu has empty extent list (next_free_rec == 0)\n",
1467					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
1468			goto bail;
1469		}
1470		i = le16_to_cpu(el->l_next_free_rec) - 1;
1471		blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1472		if (!blkno) {
1473			status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
1474					"Owner %llu has extent list where extent # %d has no physical block start\n",
1475					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
1476			goto bail;
1477		}
1478
1479		brelse(bh);
1480		bh = NULL;
1481
1482		status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
1483		if (status < 0) {
1484			mlog_errno(status);
1485			goto bail;
1486		}
1487
1488		eb = (struct ocfs2_extent_block *) bh->b_data;
1489		el = &eb->h_list;
1490
1491		if (le16_to_cpu(el->l_next_free_rec) <
1492		    le16_to_cpu(el->l_count)) {
1493			brelse(lowest_bh);
1494			lowest_bh = bh;
1495			get_bh(lowest_bh);
1496		}
1497	}
1498
1499	/* If we didn't find one and the fe doesn't have any room,
1500	 * then return '1' */
1501	el = et->et_root_el;
1502	if (!lowest_bh && (el->l_next_free_rec == el->l_count))
1503		status = 1;
1504
1505	*target_bh = lowest_bh;
1506bail:
1507	brelse(bh);
1508
1509	return status;
1510}
1511
1512/*
1513 * Grow a b-tree so that it has more records.
1514 *
1515 * We might shift the tree depth in which case existing paths should
1516 * be considered invalid.
1517 *
1518 * Tree depth after the grow is returned via *final_depth.
1519 *
1520 * *last_eb_bh will be updated by ocfs2_add_branch().
1521 */
1522static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
1523			   int *final_depth, struct buffer_head **last_eb_bh,
1524			   struct ocfs2_alloc_context *meta_ac)
1525{
1526	int ret, shift;
1527	struct ocfs2_extent_list *el = et->et_root_el;
1528	int depth = le16_to_cpu(el->l_tree_depth);
1529	struct buffer_head *bh = NULL;
1530
1531	BUG_ON(meta_ac == NULL && ocfs2_is_dealloc_empty(et));
1532
1533	shift = ocfs2_find_branch_target(et, &bh);
1534	if (shift < 0) {
1535		ret = shift;
1536		mlog_errno(ret);
1537		goto out;
1538	}
1539
1540	/* We traveled all the way to the bottom of the allocation tree
1541	 * and didn't find room for any more extents - we need to add
1542	 * another tree level */
1543	if (shift) {
1544		BUG_ON(bh);
1545		trace_ocfs2_grow_tree(
1546			(unsigned long long)
1547			ocfs2_metadata_cache_owner(et->et_ci),
1548			depth);
1549
1550		/* ocfs2_shift_tree_depth will return us a buffer with
1551		 * the new extent block (so we can pass that to
1552		 * ocfs2_add_branch). */
1553		ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
1554		if (ret < 0) {
1555			mlog_errno(ret);
1556			goto out;
1557		}
1558		depth++;
1559		if (depth == 1) {
1560			/*
1561			 * Special case: we have room now if we shifted from
1562			 * tree_depth 0, so no more work needs to be done.
1563			 *
1564			 * We won't be calling add_branch, so pass
1565			 * back *last_eb_bh as the new leaf. At depth
1566			 * zero, it should always be null so there's
1567			 * no reason to brelse.
1568			 */
1569			BUG_ON(*last_eb_bh);
1570			get_bh(bh);
1571			*last_eb_bh = bh;
1572			goto out;
1573		}
1574	}
1575
1576	/* call ocfs2_add_branch to add the final part of the tree with
1577	 * the new data. */
1578	ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
1579			       meta_ac);
1580	if (ret < 0)
1581		mlog_errno(ret);
1582
1583out:
1584	if (final_depth)
1585		*final_depth = depth;
1586	brelse(bh);
1587	return ret;
1588}
1589
1590/*
1591 * This function will discard the rightmost extent record.
1592 */
1593static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
1594{
1595	int next_free = le16_to_cpu(el->l_next_free_rec);
1596	int count = le16_to_cpu(el->l_count);
1597	unsigned int num_bytes;
1598
1599	BUG_ON(!next_free);
1600	/* This will cause us to go off the end of our extent list. */
1601	BUG_ON(next_free >= count);
1602
1603	num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;
1604
1605	memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
1606}
1607
1608static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
1609			      struct ocfs2_extent_rec *insert_rec)
1610{
1611	int i, insert_index, next_free, has_empty, num_bytes;
1612	u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
1613	struct ocfs2_extent_rec *rec;
1614
1615	next_free = le16_to_cpu(el->l_next_free_rec);
1616	has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);
1617
1618	BUG_ON(!next_free);
1619
1620	/* The tree code before us didn't allow enough room in the leaf. */
1621	BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);
1622
1623	/*
1624	 * The easiest way to approach this is to just remove the
1625	 * empty extent and temporarily decrement next_free.
1626	 */
1627	if (has_empty) {
1628		/*
1629		 * If next_free was 1 (only an empty extent), this
1630		 * loop won't execute, which is fine. We still want
1631		 * the decrement above to happen.
1632		 */
1633		for(i = 0; i < (next_free - 1); i++)
1634			el->l_recs[i] = el->l_recs[i+1];
1635
1636		next_free--;
1637	}
1638
1639	/*
1640	 * Figure out what the new record index should be.
1641	 */
1642	for(i = 0; i < next_free; i++) {
1643		rec = &el->l_recs[i];
1644
1645		if (insert_cpos < le32_to_cpu(rec->e_cpos))
1646			break;
1647	}
1648	insert_index = i;
1649
1650	trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
1651				has_empty, next_free,
1652				le16_to_cpu(el->l_count));
1653
1654	BUG_ON(insert_index < 0);
1655	BUG_ON(insert_index >= le16_to_cpu(el->l_count));
1656	BUG_ON(insert_index > next_free);
1657
1658	/*
1659	 * No need to memmove if we're just adding to the tail.
1660	 */
1661	if (insert_index != next_free) {
1662		BUG_ON(next_free >= le16_to_cpu(el->l_count));
1663
1664		num_bytes = next_free - insert_index;
1665		num_bytes *= sizeof(struct ocfs2_extent_rec);
1666		memmove(&el->l_recs[insert_index + 1],
1667			&el->l_recs[insert_index],
1668			num_bytes);
1669	}
1670
1671	/*
1672	 * Either we had an empty extent, and need to re-increment or
1673	 * there was no empty extent on a non full rightmost leaf node,
1674	 * in which case we still need to increment.
1675	 */
1676	next_free++;
1677	el->l_next_free_rec = cpu_to_le16(next_free);
1678	/*
1679	 * Make sure none of the math above just messed up our tree.
1680	 */
1681	BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));
1682
1683	el->l_recs[insert_index] = *insert_rec;
1684
1685}
1686
1687static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
1688{
1689	int size, num_recs = le16_to_cpu(el->l_next_free_rec);
1690
1691	BUG_ON(num_recs == 0);
1692
1693	if (ocfs2_is_empty_extent(&el->l_recs[0])) {
1694		num_recs--;
1695		size = num_recs * sizeof(struct ocfs2_extent_rec);
1696		memmove(&el->l_recs[0], &el->l_recs[1], size);
1697		memset(&el->l_recs[num_recs], 0,
1698		       sizeof(struct ocfs2_extent_rec));
1699		el->l_next_free_rec = cpu_to_le16(num_recs);
1700	}
1701}
1702
1703/*
1704 * Create an empty extent record .
1705 *
1706 * l_next_free_rec may be updated.
1707 *
1708 * If an empty extent already exists do nothing.
1709 */
1710static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
1711{
1712	int next_free = le16_to_cpu(el->l_next_free_rec);
1713
1714	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
1715
1716	if (next_free == 0)
1717		goto set_and_inc;
1718
1719	if (ocfs2_is_empty_extent(&el->l_recs[0]))
1720		return;
1721
1722	mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
1723			"Asked to create an empty extent in a full list:\n"
1724			"count = %u, tree depth = %u",
1725			le16_to_cpu(el->l_count),
1726			le16_to_cpu(el->l_tree_depth));
1727
1728	ocfs2_shift_records_right(el);
1729
1730set_and_inc:
1731	le16_add_cpu(&el->l_next_free_rec, 1);
1732	memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
1733}
1734
1735/*
1736 * For a rotation which involves two leaf nodes, the "root node" is
1737 * the lowest level tree node which contains a path to both leafs. This
1738 * resulting set of information can be used to form a complete "subtree"
1739 *
1740 * This function is passed two full paths from the dinode down to a
1741 * pair of adjacent leaves. It's task is to figure out which path
1742 * index contains the subtree root - this can be the root index itself
1743 * in a worst-case rotation.
1744 *
1745 * The array index of the subtree root is passed back.
1746 */
1747int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
1748			    struct ocfs2_path *left,
1749			    struct ocfs2_path *right)
1750{
1751	int i = 0;
1752
1753	/*
1754	 * Check that the caller passed in two paths from the same tree.
1755	 */
1756	BUG_ON(path_root_bh(left) != path_root_bh(right));
1757
1758	do {
1759		i++;
1760
1761		/*
1762		 * The caller didn't pass two adjacent paths.
1763		 */
1764		mlog_bug_on_msg(i > left->p_tree_depth,
1765				"Owner %llu, left depth %u, right depth %u\n"
1766				"left leaf blk %llu, right leaf blk %llu\n",
1767				(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
1768				left->p_tree_depth, right->p_tree_depth,
1769				(unsigned long long)path_leaf_bh(left)->b_blocknr,
1770				(unsigned long long)path_leaf_bh(right)->b_blocknr);
1771	} while (left->p_node[i].bh->b_blocknr ==
1772		 right->p_node[i].bh->b_blocknr);
1773
1774	return i - 1;
1775}
1776
1777typedef void (path_insert_t)(void *, struct buffer_head *);
1778
1779/*
1780 * Traverse a btree path in search of cpos, starting at root_el.
1781 *
1782 * This code can be called with a cpos larger than the tree, in which
1783 * case it will return the rightmost path.
1784 */
1785static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
1786			     struct ocfs2_extent_list *root_el, u32 cpos,
1787			     path_insert_t *func, void *data)
1788{
1789	int i, ret = 0;
1790	u32 range;
1791	u64 blkno;
1792	struct buffer_head *bh = NULL;
1793	struct ocfs2_extent_block *eb;
1794	struct ocfs2_extent_list *el;
1795	struct ocfs2_extent_rec *rec;
1796
1797	el = root_el;
1798	while (el->l_tree_depth) {
1799		if (le16_to_cpu(el->l_next_free_rec) == 0) {
1800			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1801				    "Owner %llu has empty extent list at depth %u\n",
1802				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
1803				    le16_to_cpu(el->l_tree_depth));
1804			ret = -EROFS;
1805			goto out;
1806
1807		}
1808
1809		for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
1810			rec = &el->l_recs[i];
1811
1812			/*
1813			 * In the case that cpos is off the allocation
1814			 * tree, this should just wind up returning the
1815			 * rightmost record.
1816			 */
1817			range = le32_to_cpu(rec->e_cpos) +
1818				ocfs2_rec_clusters(el, rec);
1819			if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
1820			    break;
1821		}
1822
1823		blkno = le64_to_cpu(el->l_recs[i].e_blkno);
1824		if (blkno == 0) {
1825			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1826				    "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
1827				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
1828				    le16_to_cpu(el->l_tree_depth), i);
1829			ret = -EROFS;
1830			goto out;
1831		}
1832
1833		brelse(bh);
1834		bh = NULL;
1835		ret = ocfs2_read_extent_block(ci, blkno, &bh);
1836		if (ret) {
1837			mlog_errno(ret);
1838			goto out;
1839		}
1840
1841		eb = (struct ocfs2_extent_block *) bh->b_data;
1842		el = &eb->h_list;
1843
1844		if (le16_to_cpu(el->l_next_free_rec) >
1845		    le16_to_cpu(el->l_count)) {
1846			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
1847				    "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
1848				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
1849				    (unsigned long long)bh->b_blocknr,
1850				    le16_to_cpu(el->l_next_free_rec),
1851				    le16_to_cpu(el->l_count));
1852			ret = -EROFS;
1853			goto out;
1854		}
1855
1856		if (func)
1857			func(data, bh);
1858	}
1859
1860out:
1861	/*
1862	 * Catch any trailing bh that the loop didn't handle.
1863	 */
1864	brelse(bh);
1865
1866	return ret;
1867}
1868
1869/*
1870 * Given an initialized path (that is, it has a valid root extent
1871 * list), this function will traverse the btree in search of the path
1872 * which would contain cpos.
1873 *
1874 * The path traveled is recorded in the path structure.
1875 *
1876 * Note that this will not do any comparisons on leaf node extent
1877 * records, so it will work fine in the case that we just added a tree
1878 * branch.
1879 */
1880struct find_path_data {
1881	int index;
1882	struct ocfs2_path *path;
1883};
1884static void find_path_ins(void *data, struct buffer_head *bh)
1885{
1886	struct find_path_data *fp = data;
1887
1888	get_bh(bh);
1889	ocfs2_path_insert_eb(fp->path, fp->index, bh);
1890	fp->index++;
1891}
1892int ocfs2_find_path(struct ocfs2_caching_info *ci,
1893		    struct ocfs2_path *path, u32 cpos)
1894{
1895	struct find_path_data data;
1896
1897	data.index = 1;
1898	data.path = path;
1899	return __ocfs2_find_path(ci, path_root_el(path), cpos,
1900				 find_path_ins, &data);
1901}
1902
1903static void find_leaf_ins(void *data, struct buffer_head *bh)
1904{
1905	struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
1906	struct ocfs2_extent_list *el = &eb->h_list;
1907	struct buffer_head **ret = data;
1908
1909	/* We want to retain only the leaf block. */
1910	if (le16_to_cpu(el->l_tree_depth) == 0) {
1911		get_bh(bh);
1912		*ret = bh;
1913	}
1914}
1915/*
1916 * Find the leaf block in the tree which would contain cpos. No
1917 * checking of the actual leaf is done.
1918 *
1919 * Some paths want to call this instead of allocating a path structure
1920 * and calling ocfs2_find_path().
1921 *
1922 * This function doesn't handle non btree extent lists.
1923 */
1924int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
1925		    struct ocfs2_extent_list *root_el, u32 cpos,
1926		    struct buffer_head **leaf_bh)
1927{
1928	int ret;
1929	struct buffer_head *bh = NULL;
1930
1931	ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
1932	if (ret) {
1933		mlog_errno(ret);
1934		goto out;
1935	}
1936
1937	*leaf_bh = bh;
1938out:
1939	return ret;
1940}
1941
1942/*
1943 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1944 *
1945 * Basically, we've moved stuff around at the bottom of the tree and
1946 * we need to fix up the extent records above the changes to reflect
1947 * the new changes.
1948 *
1949 * left_rec: the record on the left.
1950 * right_rec: the record to the right of left_rec
1951 * right_child_el: is the child list pointed to by right_rec
1952 *
1953 * By definition, this only works on interior nodes.
1954 */
1955static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
1956				  struct ocfs2_extent_rec *right_rec,
1957				  struct ocfs2_extent_list *right_child_el)
1958{
1959	u32 left_clusters, right_end;
1960
1961	/*
1962	 * Interior nodes never have holes. Their cpos is the cpos of
1963	 * the leftmost record in their child list. Their cluster
1964	 * count covers the full theoretical range of their child list
1965	 * - the range between their cpos and the cpos of the record
1966	 * immediately to their right.
1967	 */
1968	left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
1969	if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
1970		BUG_ON(right_child_el->l_tree_depth);
1971		BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
1972		left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
1973	}
1974	left_clusters -= le32_to_cpu(left_rec->e_cpos);
1975	left_rec->e_int_clusters = cpu_to_le32(left_clusters);
1976
1977	/*
1978	 * Calculate the rightmost cluster count boundary before
1979	 * moving cpos - we will need to adjust clusters after
1980	 * updating e_cpos to keep the same highest cluster count.
1981	 */
1982	right_end = le32_to_cpu(right_rec->e_cpos);
1983	right_end += le32_to_cpu(right_rec->e_int_clusters);
1984
1985	right_rec->e_cpos = left_rec->e_cpos;
1986	le32_add_cpu(&right_rec->e_cpos, left_clusters);
1987
1988	right_end -= le32_to_cpu(right_rec->e_cpos);
1989	right_rec->e_int_clusters = cpu_to_le32(right_end);
1990}
1991
1992/*
1993 * Adjust the adjacent root node records involved in a
1994 * rotation. left_el_blkno is passed in as a key so that we can easily
1995 * find it's index in the root list.
1996 */
1997static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
1998				      struct ocfs2_extent_list *left_el,
1999				      struct ocfs2_extent_list *right_el,
2000				      u64 left_el_blkno)
2001{
2002	int i;
2003
2004	BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
2005	       le16_to_cpu(left_el->l_tree_depth));
2006
2007	for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
2008		if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
2009			break;
2010	}
2011
2012	/*
2013	 * The path walking code should have never returned a root and
2014	 * two paths which are not adjacent.
2015	 */
2016	BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));
2017
2018	ocfs2_adjust_adjacent_records(&root_el->l_recs[i],
2019				      &root_el->l_recs[i + 1], right_el);
2020}
2021
2022/*
2023 * We've changed a leaf block (in right_path) and need to reflect that
2024 * change back up the subtree.
2025 *
2026 * This happens in multiple places:
2027 *   - When we've moved an extent record from the left path leaf to the right
2028 *     path leaf to make room for an empty extent in the left path leaf.
2029 *   - When our insert into the right path leaf is at the leftmost edge
2030 *     and requires an update of the path immediately to it's left. This
2031 *     can occur at the end of some types of rotation and appending inserts.
2032 *   - When we've adjusted the last extent record in the left path leaf and the
2033 *     1st extent record in the right path leaf during cross extent block merge.
2034 */
2035static void ocfs2_complete_edge_insert(handle_t *handle,
2036				       struct ocfs2_path *left_path,
2037				       struct ocfs2_path *right_path,
2038				       int subtree_index)
2039{
2040	int i, idx;
2041	struct ocfs2_extent_list *el, *left_el, *right_el;
2042	struct ocfs2_extent_rec *left_rec, *right_rec;
2043	struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2044
2045	/*
2046	 * Update the counts and position values within all the
2047	 * interior nodes to reflect the leaf rotation we just did.
2048	 *
2049	 * The root node is handled below the loop.
2050	 *
2051	 * We begin the loop with right_el and left_el pointing to the
2052	 * leaf lists and work our way up.
2053	 *
2054	 * NOTE: within this loop, left_el and right_el always refer
2055	 * to the *child* lists.
2056	 */
2057	left_el = path_leaf_el(left_path);
2058	right_el = path_leaf_el(right_path);
2059	for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
2060		trace_ocfs2_complete_edge_insert(i);
2061
2062		/*
2063		 * One nice property of knowing that all of these
2064		 * nodes are below the root is that we only deal with
2065		 * the leftmost right node record and the rightmost
2066		 * left node record.
2067		 */
2068		el = left_path->p_node[i].el;
2069		idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
2070		left_rec = &el->l_recs[idx];
2071
2072		el = right_path->p_node[i].el;
2073		right_rec = &el->l_recs[0];
2074
2075		ocfs2_adjust_adjacent_records(left_rec, right_rec, right_el);
2076
2077		ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
2078		ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
2079
2080		/*
2081		 * Setup our list pointers now so that the current
2082		 * parents become children in the next iteration.
2083		 */
2084		left_el = left_path->p_node[i].el;
2085		right_el = right_path->p_node[i].el;
2086	}
2087
2088	/*
2089	 * At the root node, adjust the two adjacent records which
2090	 * begin our path to the leaves.
2091	 */
2092
2093	el = left_path->p_node[subtree_index].el;
2094	left_el = left_path->p_node[subtree_index + 1].el;
2095	right_el = right_path->p_node[subtree_index + 1].el;
2096
2097	ocfs2_adjust_root_records(el, left_el, right_el,
2098				  left_path->p_node[subtree_index + 1].bh->b_blocknr);
2099
2100	root_bh = left_path->p_node[subtree_index].bh;
2101
2102	ocfs2_journal_dirty(handle, root_bh);
2103}
2104
2105static int ocfs2_rotate_subtree_right(handle_t *handle,
2106				      struct ocfs2_extent_tree *et,
2107				      struct ocfs2_path *left_path,
2108				      struct ocfs2_path *right_path,
2109				      int subtree_index)
2110{
2111	int ret, i;
2112	struct buffer_head *right_leaf_bh;
2113	struct buffer_head *left_leaf_bh = NULL;
2114	struct buffer_head *root_bh;
2115	struct ocfs2_extent_list *right_el, *left_el;
2116	struct ocfs2_extent_rec move_rec;
2117
2118	left_leaf_bh = path_leaf_bh(left_path);
2119	left_el = path_leaf_el(left_path);
2120
2121	if (left_el->l_next_free_rec != left_el->l_count) {
2122		ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
2123			    "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
2124			    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2125			    (unsigned long long)left_leaf_bh->b_blocknr,
2126			    le16_to_cpu(left_el->l_next_free_rec));
2127		return -EROFS;
2128	}
2129
2130	/*
2131	 * This extent block may already have an empty record, so we
2132	 * return early if so.
2133	 */
2134	if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
2135		return 0;
2136
2137	root_bh = left_path->p_node[subtree_index].bh;
2138	BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2139
2140	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2141					   subtree_index);
2142	if (ret) {
2143		mlog_errno(ret);
2144		goto out;
2145	}
2146
2147	for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2148		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2149						   right_path, i);
2150		if (ret) {
2151			mlog_errno(ret);
2152			goto out;
2153		}
2154
2155		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2156						   left_path, i);
2157		if (ret) {
2158			mlog_errno(ret);
2159			goto out;
2160		}
2161	}
2162
2163	right_leaf_bh = path_leaf_bh(right_path);
2164	right_el = path_leaf_el(right_path);
2165
2166	/* This is a code error, not a disk corruption. */
2167	mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
2168			"because rightmost leaf block %llu is empty\n",
2169			(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2170			(unsigned long long)right_leaf_bh->b_blocknr);
2171
2172	ocfs2_create_empty_extent(right_el);
2173
2174	ocfs2_journal_dirty(handle, right_leaf_bh);
2175
2176	/* Do the copy now. */
2177	i = le16_to_cpu(left_el->l_next_free_rec) - 1;
2178	move_rec = left_el->l_recs[i];
2179	right_el->l_recs[0] = move_rec;
2180
2181	/*
2182	 * Clear out the record we just copied and shift everything
2183	 * over, leaving an empty extent in the left leaf.
2184	 *
2185	 * We temporarily subtract from next_free_rec so that the
2186	 * shift will lose the tail record (which is now defunct).
2187	 */
2188	le16_add_cpu(&left_el->l_next_free_rec, -1);
2189	ocfs2_shift_records_right(left_el);
2190	memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2191	le16_add_cpu(&left_el->l_next_free_rec, 1);
2192
2193	ocfs2_journal_dirty(handle, left_leaf_bh);
2194
2195	ocfs2_complete_edge_insert(handle, left_path, right_path,
2196				   subtree_index);
2197
2198out:
2199	return ret;
2200}
2201
2202/*
2203 * Given a full path, determine what cpos value would return us a path
2204 * containing the leaf immediately to the left of the current one.
2205 *
2206 * Will return zero if the path passed in is already the leftmost path.
2207 */
2208int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
2209				  struct ocfs2_path *path, u32 *cpos)
2210{
2211	int i, j, ret = 0;
2212	u64 blkno;
2213	struct ocfs2_extent_list *el;
2214
2215	BUG_ON(path->p_tree_depth == 0);
2216
2217	*cpos = 0;
2218
2219	blkno = path_leaf_bh(path)->b_blocknr;
2220
2221	/* Start at the tree node just above the leaf and work our way up. */
2222	i = path->p_tree_depth - 1;
2223	while (i >= 0) {
2224		el = path->p_node[i].el;
2225
2226		/*
2227		 * Find the extent record just before the one in our
2228		 * path.
2229		 */
2230		for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2231			if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2232				if (j == 0) {
2233					if (i == 0) {
2234						/*
2235						 * We've determined that the
2236						 * path specified is already
2237						 * the leftmost one - return a
2238						 * cpos of zero.
2239						 */
2240						goto out;
2241					}
2242					/*
2243					 * The leftmost record points to our
2244					 * leaf - we need to travel up the
2245					 * tree one level.
2246					 */
2247					goto next_node;
2248				}
2249
2250				*cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
2251				*cpos = *cpos + ocfs2_rec_clusters(el,
2252							   &el->l_recs[j - 1]);
2253				*cpos = *cpos - 1;
2254				goto out;
2255			}
2256		}
2257
2258		/*
2259		 * If we got here, we never found a valid node where
2260		 * the tree indicated one should be.
2261		 */
2262		ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2263			    (unsigned long long)blkno);
2264		ret = -EROFS;
2265		goto out;
2266
2267next_node:
2268		blkno = path->p_node[i].bh->b_blocknr;
2269		i--;
2270	}
2271
2272out:
2273	return ret;
2274}
2275
2276/*
2277 * Extend the transaction by enough credits to complete the rotation,
2278 * and still leave at least the original number of credits allocated
2279 * to this transaction.
2280 */
2281static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
2282					   int op_credits,
2283					   struct ocfs2_path *path)
2284{
2285	int ret = 0;
2286	int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;
2287
2288	if (jbd2_handle_buffer_credits(handle) < credits)
2289		ret = ocfs2_extend_trans(handle,
2290				credits - jbd2_handle_buffer_credits(handle));
2291
2292	return ret;
2293}
2294
2295/*
2296 * Trap the case where we're inserting into the theoretical range past
2297 * the _actual_ left leaf range. Otherwise, we'll rotate a record
2298 * whose cpos is less than ours into the right leaf.
2299 *
2300 * It's only necessary to look at the rightmost record of the left
2301 * leaf because the logic that calls us should ensure that the
2302 * theoretical ranges in the path components above the leaves are
2303 * correct.
2304 */
2305static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
2306						 u32 insert_cpos)
2307{
2308	struct ocfs2_extent_list *left_el;
2309	struct ocfs2_extent_rec *rec;
2310	int next_free;
2311
2312	left_el = path_leaf_el(left_path);
2313	next_free = le16_to_cpu(left_el->l_next_free_rec);
2314	rec = &left_el->l_recs[next_free - 1];
2315
2316	if (insert_cpos > le32_to_cpu(rec->e_cpos))
2317		return 1;
2318	return 0;
2319}
2320
2321static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
2322{
2323	int next_free = le16_to_cpu(el->l_next_free_rec);
2324	unsigned int range;
2325	struct ocfs2_extent_rec *rec;
2326
2327	if (next_free == 0)
2328		return 0;
2329
2330	rec = &el->l_recs[0];
2331	if (ocfs2_is_empty_extent(rec)) {
2332		/* Empty list. */
2333		if (next_free == 1)
2334			return 0;
2335		rec = &el->l_recs[1];
2336	}
2337
2338	range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2339	if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
2340		return 1;
2341	return 0;
2342}
2343
2344/*
2345 * Rotate all the records in a btree right one record, starting at insert_cpos.
2346 *
2347 * The path to the rightmost leaf should be passed in.
2348 *
2349 * The array is assumed to be large enough to hold an entire path (tree depth).
2350 *
2351 * Upon successful return from this function:
2352 *
2353 * - The 'right_path' array will contain a path to the leaf block
2354 *   whose range contains e_cpos.
2355 * - That leaf block will have a single empty extent in list index 0.
2356 * - In the case that the rotation requires a post-insert update,
2357 *   *ret_left_path will contain a valid path which can be passed to
2358 *   ocfs2_insert_path().
2359 */
2360static int ocfs2_rotate_tree_right(handle_t *handle,
2361				   struct ocfs2_extent_tree *et,
2362				   enum ocfs2_split_type split,
2363				   u32 insert_cpos,
2364				   struct ocfs2_path *right_path,
2365				   struct ocfs2_path **ret_left_path)
2366{
2367	int ret, start, orig_credits = jbd2_handle_buffer_credits(handle);
2368	u32 cpos;
2369	struct ocfs2_path *left_path = NULL;
2370	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2371
2372	*ret_left_path = NULL;
2373
2374	left_path = ocfs2_new_path_from_path(right_path);
2375	if (!left_path) {
2376		ret = -ENOMEM;
2377		mlog_errno(ret);
2378		goto out;
2379	}
2380
2381	ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2382	if (ret) {
2383		mlog_errno(ret);
2384		goto out;
2385	}
2386
2387	trace_ocfs2_rotate_tree_right(
2388		(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2389		insert_cpos, cpos);
2390
2391	/*
2392	 * What we want to do here is:
2393	 *
2394	 * 1) Start with the rightmost path.
2395	 *
2396	 * 2) Determine a path to the leaf block directly to the left
2397	 *    of that leaf.
2398	 *
2399	 * 3) Determine the 'subtree root' - the lowest level tree node
2400	 *    which contains a path to both leaves.
2401	 *
2402	 * 4) Rotate the subtree.
2403	 *
2404	 * 5) Find the next subtree by considering the left path to be
2405	 *    the new right path.
2406	 *
2407	 * The check at the top of this while loop also accepts
2408	 * insert_cpos == cpos because cpos is only a _theoretical_
2409	 * value to get us the left path - insert_cpos might very well
2410	 * be filling that hole.
2411	 *
2412	 * Stop at a cpos of '0' because we either started at the
2413	 * leftmost branch (i.e., a tree with one branch and a
2414	 * rotation inside of it), or we've gone as far as we can in
2415	 * rotating subtrees.
2416	 */
2417	while (cpos && insert_cpos <= cpos) {
2418		trace_ocfs2_rotate_tree_right(
2419			(unsigned long long)
2420			ocfs2_metadata_cache_owner(et->et_ci),
2421			insert_cpos, cpos);
2422
2423		ret = ocfs2_find_path(et->et_ci, left_path, cpos);
2424		if (ret) {
2425			mlog_errno(ret);
2426			goto out;
2427		}
2428
2429		mlog_bug_on_msg(path_leaf_bh(left_path) ==
2430				path_leaf_bh(right_path),
2431				"Owner %llu: error during insert of %u "
2432				"(left path cpos %u) results in two identical "
2433				"paths ending at %llu\n",
2434				(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2435				insert_cpos, cpos,
2436				(unsigned long long)
2437				path_leaf_bh(left_path)->b_blocknr);
2438
2439		if (split == SPLIT_NONE &&
2440		    ocfs2_rotate_requires_path_adjustment(left_path,
2441							  insert_cpos)) {
2442
2443			/*
2444			 * We've rotated the tree as much as we
2445			 * should. The rest is up to
2446			 * ocfs2_insert_path() to complete, after the
2447			 * record insertion. We indicate this
2448			 * situation by returning the left path.
2449			 *
2450			 * The reason we don't adjust the records here
2451			 * before the record insert is that an error
2452			 * later might break the rule where a parent
2453			 * record e_cpos will reflect the actual
2454			 * e_cpos of the 1st nonempty record of the
2455			 * child list.
2456			 */
2457			*ret_left_path = left_path;
2458			goto out_ret_path;
2459		}
2460
2461		start = ocfs2_find_subtree_root(et, left_path, right_path);
2462
2463		trace_ocfs2_rotate_subtree(start,
2464			(unsigned long long)
2465			right_path->p_node[start].bh->b_blocknr,
2466			right_path->p_tree_depth);
2467
2468		ret = ocfs2_extend_rotate_transaction(handle, start,
2469						      orig_credits, right_path);
2470		if (ret) {
2471			mlog_errno(ret);
2472			goto out;
2473		}
2474
2475		ret = ocfs2_rotate_subtree_right(handle, et, left_path,
2476						 right_path, start);
2477		if (ret) {
2478			mlog_errno(ret);
2479			goto out;
2480		}
2481
2482		if (split != SPLIT_NONE &&
2483		    ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
2484						insert_cpos)) {
2485			/*
2486			 * A rotate moves the rightmost left leaf
2487			 * record over to the leftmost right leaf
2488			 * slot. If we're doing an extent split
2489			 * instead of a real insert, then we have to
2490			 * check that the extent to be split wasn't
2491			 * just moved over. If it was, then we can
2492			 * exit here, passing left_path back -
2493			 * ocfs2_split_extent() is smart enough to
2494			 * search both leaves.
2495			 */
2496			*ret_left_path = left_path;
2497			goto out_ret_path;
2498		}
2499
2500		/*
2501		 * There is no need to re-read the next right path
2502		 * as we know that it'll be our current left
2503		 * path. Optimize by copying values instead.
2504		 */
2505		ocfs2_mv_path(right_path, left_path);
2506
2507		ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
2508		if (ret) {
2509			mlog_errno(ret);
2510			goto out;
2511		}
2512	}
2513
2514out:
2515	ocfs2_free_path(left_path);
2516
2517out_ret_path:
2518	return ret;
2519}
2520
2521static int ocfs2_update_edge_lengths(handle_t *handle,
2522				     struct ocfs2_extent_tree *et,
2523				     struct ocfs2_path *path)
2524{
2525	int i, idx, ret;
2526	struct ocfs2_extent_rec *rec;
2527	struct ocfs2_extent_list *el;
2528	struct ocfs2_extent_block *eb;
2529	u32 range;
2530
2531	ret = ocfs2_journal_access_path(et->et_ci, handle, path);
2532	if (ret) {
2533		mlog_errno(ret);
2534		goto out;
2535	}
2536
2537	/* Path should always be rightmost. */
2538	eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
2539	BUG_ON(eb->h_next_leaf_blk != 0ULL);
2540
2541	el = &eb->h_list;
2542	BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
2543	idx = le16_to_cpu(el->l_next_free_rec) - 1;
2544	rec = &el->l_recs[idx];
2545	range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
2546
2547	for (i = 0; i < path->p_tree_depth; i++) {
2548		el = path->p_node[i].el;
2549		idx = le16_to_cpu(el->l_next_free_rec) - 1;
2550		rec = &el->l_recs[idx];
2551
2552		rec->e_int_clusters = cpu_to_le32(range);
2553		le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));
2554
2555		ocfs2_journal_dirty(handle, path->p_node[i].bh);
2556	}
2557out:
2558	return ret;
2559}
2560
2561static void ocfs2_unlink_path(handle_t *handle,
2562			      struct ocfs2_extent_tree *et,
2563			      struct ocfs2_cached_dealloc_ctxt *dealloc,
2564			      struct ocfs2_path *path, int unlink_start)
2565{
2566	int ret, i;
2567	struct ocfs2_extent_block *eb;
2568	struct ocfs2_extent_list *el;
2569	struct buffer_head *bh;
2570
2571	for(i = unlink_start; i < path_num_items(path); i++) {
2572		bh = path->p_node[i].bh;
2573
2574		eb = (struct ocfs2_extent_block *)bh->b_data;
2575		/*
2576		 * Not all nodes might have had their final count
2577		 * decremented by the caller - handle this here.
2578		 */
2579		el = &eb->h_list;
2580		if (le16_to_cpu(el->l_next_free_rec) > 1) {
2581			mlog(ML_ERROR,
2582			     "Inode %llu, attempted to remove extent block "
2583			     "%llu with %u records\n",
2584			     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
2585			     (unsigned long long)le64_to_cpu(eb->h_blkno),
2586			     le16_to_cpu(el->l_next_free_rec));
2587
2588			ocfs2_journal_dirty(handle, bh);
2589			ocfs2_remove_from_cache(et->et_ci, bh);
2590			continue;
2591		}
2592
2593		el->l_next_free_rec = 0;
2594		memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
2595
2596		ocfs2_journal_dirty(handle, bh);
2597
2598		ret = ocfs2_cache_extent_block_free(dealloc, eb);
2599		if (ret)
2600			mlog_errno(ret);
2601
2602		ocfs2_remove_from_cache(et->et_ci, bh);
2603	}
2604}
2605
2606static void ocfs2_unlink_subtree(handle_t *handle,
2607				 struct ocfs2_extent_tree *et,
2608				 struct ocfs2_path *left_path,
2609				 struct ocfs2_path *right_path,
2610				 int subtree_index,
2611				 struct ocfs2_cached_dealloc_ctxt *dealloc)
2612{
2613	int i;
2614	struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
2615	struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
2616	struct ocfs2_extent_block *eb;
2617
2618	eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;
2619
2620	for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
2621		if (root_el->l_recs[i].e_blkno == eb->h_blkno)
2622			break;
2623
2624	BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));
2625
2626	memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
2627	le16_add_cpu(&root_el->l_next_free_rec, -1);
2628
2629	eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2630	eb->h_next_leaf_blk = 0;
2631
2632	ocfs2_journal_dirty(handle, root_bh);
2633	ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2634
2635	ocfs2_unlink_path(handle, et, dealloc, right_path,
2636			  subtree_index + 1);
2637}
2638
2639static int ocfs2_rotate_subtree_left(handle_t *handle,
2640				     struct ocfs2_extent_tree *et,
2641				     struct ocfs2_path *left_path,
2642				     struct ocfs2_path *right_path,
2643				     int subtree_index,
2644				     struct ocfs2_cached_dealloc_ctxt *dealloc,
2645				     int *deleted)
2646{
2647	int ret, i, del_right_subtree = 0, right_has_empty = 0;
2648	struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
2649	struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
2650	struct ocfs2_extent_block *eb;
2651
2652	*deleted = 0;
2653
2654	right_leaf_el = path_leaf_el(right_path);
2655	left_leaf_el = path_leaf_el(left_path);
2656	root_bh = left_path->p_node[subtree_index].bh;
2657	BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
2658
2659	if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
2660		return 0;
2661
2662	eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
2663	if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
2664		/*
2665		 * It's legal for us to proceed if the right leaf is
2666		 * the rightmost one and it has an empty extent. There
2667		 * are two cases to handle - whether the leaf will be
2668		 * empty after removal or not. If the leaf isn't empty
2669		 * then just remove the empty extent up front. The
2670		 * next block will handle empty leaves by flagging
2671		 * them for unlink.
2672		 *
2673		 * Non rightmost leaves will throw -EAGAIN and the
2674		 * caller can manually move the subtree and retry.
2675		 */
2676
2677		if (eb->h_next_leaf_blk != 0ULL)
2678			return -EAGAIN;
2679
2680		if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
2681			ret = ocfs2_journal_access_eb(handle, et->et_ci,
2682						      path_leaf_bh(right_path),
2683						      OCFS2_JOURNAL_ACCESS_WRITE);
2684			if (ret) {
2685				mlog_errno(ret);
2686				goto out;
2687			}
2688
2689			ocfs2_remove_empty_extent(right_leaf_el);
2690		} else
2691			right_has_empty = 1;
2692	}
2693
2694	if (eb->h_next_leaf_blk == 0ULL &&
2695	    le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
2696		/*
2697		 * We have to update i_last_eb_blk during the meta
2698		 * data delete.
2699		 */
2700		ret = ocfs2_et_root_journal_access(handle, et,
2701						   OCFS2_JOURNAL_ACCESS_WRITE);
2702		if (ret) {
2703			mlog_errno(ret);
2704			goto out;
2705		}
2706
2707		del_right_subtree = 1;
2708	}
2709
2710	/*
2711	 * Getting here with an empty extent in the right path implies
2712	 * that it's the rightmost path and will be deleted.
2713	 */
2714	BUG_ON(right_has_empty && !del_right_subtree);
2715
2716	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
2717					   subtree_index);
2718	if (ret) {
2719		mlog_errno(ret);
2720		goto out;
2721	}
2722
2723	for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
2724		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2725						   right_path, i);
2726		if (ret) {
2727			mlog_errno(ret);
2728			goto out;
2729		}
2730
2731		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2732						   left_path, i);
2733		if (ret) {
2734			mlog_errno(ret);
2735			goto out;
2736		}
2737	}
2738
2739	if (!right_has_empty) {
2740		/*
2741		 * Only do this if we're moving a real
2742		 * record. Otherwise, the action is delayed until
2743		 * after removal of the right path in which case we
2744		 * can do a simple shift to remove the empty extent.
2745		 */
2746		ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
2747		memset(&right_leaf_el->l_recs[0], 0,
2748		       sizeof(struct ocfs2_extent_rec));
2749	}
2750	if (eb->h_next_leaf_blk == 0ULL) {
2751		/*
2752		 * Move recs over to get rid of empty extent, decrease
2753		 * next_free. This is allowed to remove the last
2754		 * extent in our leaf (setting l_next_free_rec to
2755		 * zero) - the delete code below won't care.
2756		 */
2757		ocfs2_remove_empty_extent(right_leaf_el);
2758	}
2759
2760	ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
2761	ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
2762
2763	if (del_right_subtree) {
2764		ocfs2_unlink_subtree(handle, et, left_path, right_path,
2765				     subtree_index, dealloc);
2766		ret = ocfs2_update_edge_lengths(handle, et, left_path);
2767		if (ret) {
2768			mlog_errno(ret);
2769			goto out;
2770		}
2771
2772		eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
2773		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
2774
2775		/*
2776		 * Removal of the extent in the left leaf was skipped
2777		 * above so we could delete the right path
2778		 * 1st.
2779		 */
2780		if (right_has_empty)
2781			ocfs2_remove_empty_extent(left_leaf_el);
2782
2783		ocfs2_journal_dirty(handle, et_root_bh);
2784
2785		*deleted = 1;
2786	} else
2787		ocfs2_complete_edge_insert(handle, left_path, right_path,
2788					   subtree_index);
2789
2790out:
2791	return ret;
2792}
2793
2794/*
2795 * Given a full path, determine what cpos value would return us a path
2796 * containing the leaf immediately to the right of the current one.
2797 *
2798 * Will return zero if the path passed in is already the rightmost path.
2799 *
2800 * This looks similar, but is subtly different to
2801 * ocfs2_find_cpos_for_left_leaf().
2802 */
2803int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
2804				   struct ocfs2_path *path, u32 *cpos)
2805{
2806	int i, j, ret = 0;
2807	u64 blkno;
2808	struct ocfs2_extent_list *el;
2809
2810	*cpos = 0;
2811
2812	if (path->p_tree_depth == 0)
2813		return 0;
2814
2815	blkno = path_leaf_bh(path)->b_blocknr;
2816
2817	/* Start at the tree node just above the leaf and work our way up. */
2818	i = path->p_tree_depth - 1;
2819	while (i >= 0) {
2820		int next_free;
2821
2822		el = path->p_node[i].el;
2823
2824		/*
2825		 * Find the extent record just after the one in our
2826		 * path.
2827		 */
2828		next_free = le16_to_cpu(el->l_next_free_rec);
2829		for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
2830			if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
2831				if (j == (next_free - 1)) {
2832					if (i == 0) {
2833						/*
2834						 * We've determined that the
2835						 * path specified is already
2836						 * the rightmost one - return a
2837						 * cpos of zero.
2838						 */
2839						goto out;
2840					}
2841					/*
2842					 * The rightmost record points to our
2843					 * leaf - we need to travel up the
2844					 * tree one level.
2845					 */
2846					goto next_node;
2847				}
2848
2849				*cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
2850				goto out;
2851			}
2852		}
2853
2854		/*
2855		 * If we got here, we never found a valid node where
2856		 * the tree indicated one should be.
2857		 */
2858		ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
2859			    (unsigned long long)blkno);
2860		ret = -EROFS;
2861		goto out;
2862
2863next_node:
2864		blkno = path->p_node[i].bh->b_blocknr;
2865		i--;
2866	}
2867
2868out:
2869	return ret;
2870}
2871
2872static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
2873					    struct ocfs2_extent_tree *et,
2874					    struct ocfs2_path *path)
2875{
2876	int ret;
2877	struct buffer_head *bh = path_leaf_bh(path);
2878	struct ocfs2_extent_list *el = path_leaf_el(path);
2879
2880	if (!ocfs2_is_empty_extent(&el->l_recs[0]))
2881		return 0;
2882
2883	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
2884					   path_num_items(path) - 1);
2885	if (ret) {
2886		mlog_errno(ret);
2887		goto out;
2888	}
2889
2890	ocfs2_remove_empty_extent(el);
2891	ocfs2_journal_dirty(handle, bh);
2892
2893out:
2894	return ret;
2895}
2896
2897static int __ocfs2_rotate_tree_left(handle_t *handle,
2898				    struct ocfs2_extent_tree *et,
2899				    int orig_credits,
2900				    struct ocfs2_path *path,
2901				    struct ocfs2_cached_dealloc_ctxt *dealloc,
2902				    struct ocfs2_path **empty_extent_path)
2903{
2904	int ret, subtree_root, deleted;
2905	u32 right_cpos;
2906	struct ocfs2_path *left_path = NULL;
2907	struct ocfs2_path *right_path = NULL;
2908	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
2909
2910	if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
2911		return 0;
2912
2913	*empty_extent_path = NULL;
2914
2915	ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
2916	if (ret) {
2917		mlog_errno(ret);
2918		goto out;
2919	}
2920
2921	left_path = ocfs2_new_path_from_path(path);
2922	if (!left_path) {
2923		ret = -ENOMEM;
2924		mlog_errno(ret);
2925		goto out;
2926	}
2927
2928	ocfs2_cp_path(left_path, path);
2929
2930	right_path = ocfs2_new_path_from_path(path);
2931	if (!right_path) {
2932		ret = -ENOMEM;
2933		mlog_errno(ret);
2934		goto out;
2935	}
2936
2937	while (right_cpos) {
2938		ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
2939		if (ret) {
2940			mlog_errno(ret);
2941			goto out;
2942		}
2943
2944		subtree_root = ocfs2_find_subtree_root(et, left_path,
2945						       right_path);
2946
2947		trace_ocfs2_rotate_subtree(subtree_root,
2948		     (unsigned long long)
2949		     right_path->p_node[subtree_root].bh->b_blocknr,
2950		     right_path->p_tree_depth);
2951
2952		ret = ocfs2_extend_rotate_transaction(handle, 0,
2953						      orig_credits, left_path);
2954		if (ret) {
2955			mlog_errno(ret);
2956			goto out;
2957		}
2958
2959		/*
2960		 * Caller might still want to make changes to the
2961		 * tree root, so re-add it to the journal here.
2962		 */
2963		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
2964						   left_path, 0);
2965		if (ret) {
2966			mlog_errno(ret);
2967			goto out;
2968		}
2969
2970		ret = ocfs2_rotate_subtree_left(handle, et, left_path,
2971						right_path, subtree_root,
2972						dealloc, &deleted);
2973		if (ret == -EAGAIN) {
2974			/*
2975			 * The rotation has to temporarily stop due to
2976			 * the right subtree having an empty
2977			 * extent. Pass it back to the caller for a
2978			 * fixup.
2979			 */
2980			*empty_extent_path = right_path;
2981			right_path = NULL;
2982			goto out;
2983		}
2984		if (ret) {
2985			mlog_errno(ret);
2986			goto out;
2987		}
2988
2989		/*
2990		 * The subtree rotate might have removed records on
2991		 * the rightmost edge. If so, then rotation is
2992		 * complete.
2993		 */
2994		if (deleted)
2995			break;
2996
2997		ocfs2_mv_path(left_path, right_path);
2998
2999		ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
3000						     &right_cpos);
3001		if (ret) {
3002			mlog_errno(ret);
3003			goto out;
3004		}
3005	}
3006
3007out:
3008	ocfs2_free_path(right_path);
3009	ocfs2_free_path(left_path);
3010
3011	return ret;
3012}
3013
3014static int ocfs2_remove_rightmost_path(handle_t *handle,
3015				struct ocfs2_extent_tree *et,
3016				struct ocfs2_path *path,
3017				struct ocfs2_cached_dealloc_ctxt *dealloc)
3018{
3019	int ret, subtree_index;
3020	u32 cpos;
3021	struct ocfs2_path *left_path = NULL;
3022	struct ocfs2_extent_block *eb;
3023	struct ocfs2_extent_list *el;
3024
3025	ret = ocfs2_et_sanity_check(et);
3026	if (ret)
3027		goto out;
3028
3029	ret = ocfs2_journal_access_path(et->et_ci, handle, path);
3030	if (ret) {
3031		mlog_errno(ret);
3032		goto out;
3033	}
3034
3035	ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3036					    path, &cpos);
3037	if (ret) {
3038		mlog_errno(ret);
3039		goto out;
3040	}
3041
3042	if (cpos) {
3043		/*
3044		 * We have a path to the left of this one - it needs
3045		 * an update too.
3046		 */
3047		left_path = ocfs2_new_path_from_path(path);
3048		if (!left_path) {
3049			ret = -ENOMEM;
3050			mlog_errno(ret);
3051			goto out;
3052		}
3053
3054		ret = ocfs2_find_path(et->et_ci, left_path, cpos);
3055		if (ret) {
3056			mlog_errno(ret);
3057			goto out;
3058		}
3059
3060		ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
3061		if (ret) {
3062			mlog_errno(ret);
3063			goto out;
3064		}
3065
3066		subtree_index = ocfs2_find_subtree_root(et, left_path, path);
3067
3068		ocfs2_unlink_subtree(handle, et, left_path, path,
3069				     subtree_index, dealloc);
3070		ret = ocfs2_update_edge_lengths(handle, et, left_path);
3071		if (ret) {
3072			mlog_errno(ret);
3073			goto out;
3074		}
3075
3076		eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
3077		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
3078	} else {
3079		/*
3080		 * 'path' is also the leftmost path which
3081		 * means it must be the only one. This gets
3082		 * handled differently because we want to
3083		 * revert the root back to having extents
3084		 * in-line.
3085		 */
3086		ocfs2_unlink_path(handle, et, dealloc, path, 1);
3087
3088		el = et->et_root_el;
3089		el->l_tree_depth = 0;
3090		el->l_next_free_rec = 0;
3091		memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3092
3093		ocfs2_et_set_last_eb_blk(et, 0);
3094	}
3095
3096	ocfs2_journal_dirty(handle, path_root_bh(path));
3097
3098out:
3099	ocfs2_free_path(left_path);
3100	return ret;
3101}
3102
3103static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
3104				struct ocfs2_extent_tree *et,
3105				struct ocfs2_path *path,
3106				struct ocfs2_cached_dealloc_ctxt *dealloc)
3107{
3108	handle_t *handle;
3109	int ret;
3110	int credits = path->p_tree_depth * 2 + 1;
3111
3112	handle = ocfs2_start_trans(osb, credits);
3113	if (IS_ERR(handle)) {
3114		ret = PTR_ERR(handle);
3115		mlog_errno(ret);
3116		return ret;
3117	}
3118
3119	ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
3120	if (ret)
3121		mlog_errno(ret);
3122
3123	ocfs2_commit_trans(osb, handle);
3124	return ret;
3125}
3126
3127/*
3128 * Left rotation of btree records.
3129 *
3130 * In many ways, this is (unsurprisingly) the opposite of right
3131 * rotation. We start at some non-rightmost path containing an empty
3132 * extent in the leaf block. The code works its way to the rightmost
3133 * path by rotating records to the left in every subtree.
3134 *
3135 * This is used by any code which reduces the number of extent records
3136 * in a leaf. After removal, an empty record should be placed in the
3137 * leftmost list position.
3138 *
3139 * This won't handle a length update of the rightmost path records if
3140 * the rightmost tree leaf record is removed so the caller is
3141 * responsible for detecting and correcting that.
3142 */
3143static int ocfs2_rotate_tree_left(handle_t *handle,
3144				  struct ocfs2_extent_tree *et,
3145				  struct ocfs2_path *path,
3146				  struct ocfs2_cached_dealloc_ctxt *dealloc)
3147{
3148	int ret, orig_credits = jbd2_handle_buffer_credits(handle);
3149	struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
3150	struct ocfs2_extent_block *eb;
3151	struct ocfs2_extent_list *el;
3152
3153	el = path_leaf_el(path);
3154	if (!ocfs2_is_empty_extent(&el->l_recs[0]))
3155		return 0;
3156
3157	if (path->p_tree_depth == 0) {
3158rightmost_no_delete:
3159		/*
3160		 * Inline extents. This is trivially handled, so do
3161		 * it up front.
3162		 */
3163		ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
3164		if (ret)
3165			mlog_errno(ret);
3166		goto out;
3167	}
3168
3169	/*
3170	 * Handle rightmost branch now. There's several cases:
3171	 *  1) simple rotation leaving records in there. That's trivial.
3172	 *  2) rotation requiring a branch delete - there's no more
3173	 *     records left. Two cases of this:
3174	 *     a) There are branches to the left.
3175	 *     b) This is also the leftmost (the only) branch.
3176	 *
3177	 *  1) is handled via ocfs2_rotate_rightmost_leaf_left()
3178	 *  2a) we need the left branch so that we can update it with the unlink
3179	 *  2b) we need to bring the root back to inline extents.
3180	 */
3181
3182	eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
3183	el = &eb->h_list;
3184	if (eb->h_next_leaf_blk == 0) {
3185		/*
3186		 * This gets a bit tricky if we're going to delete the
3187		 * rightmost path. Get the other cases out of the way
3188		 * 1st.
3189		 */
3190		if (le16_to_cpu(el->l_next_free_rec) > 1)
3191			goto rightmost_no_delete;
3192
3193		if (le16_to_cpu(el->l_next_free_rec) == 0) {
3194			ret = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3195					"Owner %llu has empty extent block at %llu\n",
3196					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
3197					(unsigned long long)le64_to_cpu(eb->h_blkno));
3198			goto out;
3199		}
3200
3201		/*
3202		 * XXX: The caller can not trust "path" any more after
3203		 * this as it will have been deleted. What do we do?
3204		 *
3205		 * In theory the rotate-for-merge code will never get
3206		 * here because it'll always ask for a rotate in a
3207		 * nonempty list.
3208		 */
3209
3210		ret = ocfs2_remove_rightmost_path(handle, et, path,
3211						  dealloc);
3212		if (ret)
3213			mlog_errno(ret);
3214		goto out;
3215	}
3216
3217	/*
3218	 * Now we can loop, remembering the path we get from -EAGAIN
3219	 * and restarting from there.
3220	 */
3221try_rotate:
3222	ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
3223				       dealloc, &restart_path);
3224	if (ret && ret != -EAGAIN) {
3225		mlog_errno(ret);
3226		goto out;
3227	}
3228
3229	while (ret == -EAGAIN) {
3230		tmp_path = restart_path;
3231		restart_path = NULL;
3232
3233		ret = __ocfs2_rotate_tree_left(handle, et, orig_credits,
3234					       tmp_path, dealloc,
3235					       &restart_path);
3236		if (ret && ret != -EAGAIN) {
3237			mlog_errno(ret);
3238			goto out;
3239		}
3240
3241		ocfs2_free_path(tmp_path);
3242		tmp_path = NULL;
3243
3244		if (ret == 0)
3245			goto try_rotate;
3246	}
3247
3248out:
3249	ocfs2_free_path(tmp_path);
3250	ocfs2_free_path(restart_path);
3251	return ret;
3252}
3253
3254static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el,
3255				int index)
3256{
3257	struct ocfs2_extent_rec *rec = &el->l_recs[index];
3258	unsigned int size;
3259
3260	if (rec->e_leaf_clusters == 0) {
3261		/*
3262		 * We consumed all of the merged-from record. An empty
3263		 * extent cannot exist anywhere but the 1st array
3264		 * position, so move things over if the merged-from
3265		 * record doesn't occupy that position.
3266		 *
3267		 * This creates a new empty extent so the caller
3268		 * should be smart enough to have removed any existing
3269		 * ones.
3270		 */
3271		if (index > 0) {
3272			BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
3273			size = index * sizeof(struct ocfs2_extent_rec);
3274			memmove(&el->l_recs[1], &el->l_recs[0], size);
3275		}
3276
3277		/*
3278		 * Always memset - the caller doesn't check whether it
3279		 * created an empty extent, so there could be junk in
3280		 * the other fields.
3281		 */
3282		memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
3283	}
3284}
3285
3286static int ocfs2_get_right_path(struct ocfs2_extent_tree *et,
3287				struct ocfs2_path *left_path,
3288				struct ocfs2_path **ret_right_path)
3289{
3290	int ret;
3291	u32 right_cpos;
3292	struct ocfs2_path *right_path = NULL;
3293	struct ocfs2_extent_list *left_el;
3294
3295	*ret_right_path = NULL;
3296
3297	/* This function shouldn't be called for non-trees. */
3298	BUG_ON(left_path->p_tree_depth == 0);
3299
3300	left_el = path_leaf_el(left_path);
3301	BUG_ON(left_el->l_next_free_rec != left_el->l_count);
3302
3303	ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3304					     left_path, &right_cpos);
3305	if (ret) {
3306		mlog_errno(ret);
3307		goto out;
3308	}
3309
3310	/* This function shouldn't be called for the rightmost leaf. */
3311	BUG_ON(right_cpos == 0);
3312
3313	right_path = ocfs2_new_path_from_path(left_path);
3314	if (!right_path) {
3315		ret = -ENOMEM;
3316		mlog_errno(ret);
3317		goto out;
3318	}
3319
3320	ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
3321	if (ret) {
3322		mlog_errno(ret);
3323		goto out;
3324	}
3325
3326	*ret_right_path = right_path;
3327out:
3328	if (ret)
3329		ocfs2_free_path(right_path);
3330	return ret;
3331}
3332
3333/*
3334 * Remove split_rec clusters from the record at index and merge them
3335 * onto the beginning of the record "next" to it.
3336 * For index < l_count - 1, the next means the extent rec at index + 1.
3337 * For index == l_count - 1, the "next" means the 1st extent rec of the
3338 * next extent block.
3339 */
3340static int ocfs2_merge_rec_right(struct ocfs2_path *left_path,
3341				 handle_t *handle,
3342				 struct ocfs2_extent_tree *et,
3343				 struct ocfs2_extent_rec *split_rec,
3344				 int index)
3345{
3346	int ret, next_free, i;
3347	unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3348	struct ocfs2_extent_rec *left_rec;
3349	struct ocfs2_extent_rec *right_rec;
3350	struct ocfs2_extent_list *right_el;
3351	struct ocfs2_path *right_path = NULL;
3352	int subtree_index = 0;
3353	struct ocfs2_extent_list *el = path_leaf_el(left_path);
3354	struct buffer_head *bh = path_leaf_bh(left_path);
3355	struct buffer_head *root_bh = NULL;
3356
3357	BUG_ON(index >= le16_to_cpu(el->l_next_free_rec));
3358	left_rec = &el->l_recs[index];
3359
3360	if (index == le16_to_cpu(el->l_next_free_rec) - 1 &&
3361	    le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) {
3362		/* we meet with a cross extent block merge. */
3363		ret = ocfs2_get_right_path(et, left_path, &right_path);
3364		if (ret) {
3365			mlog_errno(ret);
3366			return ret;
3367		}
3368
3369		right_el = path_leaf_el(right_path);
3370		next_free = le16_to_cpu(right_el->l_next_free_rec);
3371		BUG_ON(next_free <= 0);
3372		right_rec = &right_el->l_recs[0];
3373		if (ocfs2_is_empty_extent(right_rec)) {
3374			BUG_ON(next_free <= 1);
3375			right_rec = &right_el->l_recs[1];
3376		}
3377
3378		BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3379		       le16_to_cpu(left_rec->e_leaf_clusters) !=
3380		       le32_to_cpu(right_rec->e_cpos));
3381
3382		subtree_index = ocfs2_find_subtree_root(et, left_path,
3383							right_path);
3384
3385		ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3386					jbd2_handle_buffer_credits(handle),
3387					right_path);
3388		if (ret) {
3389			mlog_errno(ret);
3390			goto out;
3391		}
3392
3393		root_bh = left_path->p_node[subtree_index].bh;
3394		BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3395
3396		ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3397						   subtree_index);
3398		if (ret) {
3399			mlog_errno(ret);
3400			goto out;
3401		}
3402
3403		for (i = subtree_index + 1;
3404		     i < path_num_items(right_path); i++) {
3405			ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3406							   right_path, i);
3407			if (ret) {
3408				mlog_errno(ret);
3409				goto out;
3410			}
3411
3412			ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3413							   left_path, i);
3414			if (ret) {
3415				mlog_errno(ret);
3416				goto out;
3417			}
3418		}
3419
3420	} else {
3421		BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1);
3422		right_rec = &el->l_recs[index + 1];
3423	}
3424
3425	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path,
3426					   path_num_items(left_path) - 1);
3427	if (ret) {
3428		mlog_errno(ret);
3429		goto out;
3430	}
3431
3432	le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters);
3433
3434	le32_add_cpu(&right_rec->e_cpos, -split_clusters);
3435	le64_add_cpu(&right_rec->e_blkno,
3436		     -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3437					       split_clusters));
3438	le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters);
3439
3440	ocfs2_cleanup_merge(el, index);
3441
3442	ocfs2_journal_dirty(handle, bh);
3443	if (right_path) {
3444		ocfs2_journal_dirty(handle, path_leaf_bh(right_path));
3445		ocfs2_complete_edge_insert(handle, left_path, right_path,
3446					   subtree_index);
3447	}
3448out:
3449	ocfs2_free_path(right_path);
3450	return ret;
3451}
3452
3453static int ocfs2_get_left_path(struct ocfs2_extent_tree *et,
3454			       struct ocfs2_path *right_path,
3455			       struct ocfs2_path **ret_left_path)
3456{
3457	int ret;
3458	u32 left_cpos;
3459	struct ocfs2_path *left_path = NULL;
3460
3461	*ret_left_path = NULL;
3462
3463	/* This function shouldn't be called for non-trees. */
3464	BUG_ON(right_path->p_tree_depth == 0);
3465
3466	ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
3467					    right_path, &left_cpos);
3468	if (ret) {
3469		mlog_errno(ret);
3470		goto out;
3471	}
3472
3473	/* This function shouldn't be called for the leftmost leaf. */
3474	BUG_ON(left_cpos == 0);
3475
3476	left_path = ocfs2_new_path_from_path(right_path);
3477	if (!left_path) {
3478		ret = -ENOMEM;
3479		mlog_errno(ret);
3480		goto out;
3481	}
3482
3483	ret = ocfs2_find_path(et->et_ci, left_path, left_cpos);
3484	if (ret) {
3485		mlog_errno(ret);
3486		goto out;
3487	}
3488
3489	*ret_left_path = left_path;
3490out:
3491	if (ret)
3492		ocfs2_free_path(left_path);
3493	return ret;
3494}
3495
3496/*
3497 * Remove split_rec clusters from the record at index and merge them
3498 * onto the tail of the record "before" it.
3499 * For index > 0, the "before" means the extent rec at index - 1.
3500 *
3501 * For index == 0, the "before" means the last record of the previous
3502 * extent block. And there is also a situation that we may need to
3503 * remove the rightmost leaf extent block in the right_path and change
3504 * the right path to indicate the new rightmost path.
3505 */
3506static int ocfs2_merge_rec_left(struct ocfs2_path *right_path,
3507				handle_t *handle,
3508				struct ocfs2_extent_tree *et,
3509				struct ocfs2_extent_rec *split_rec,
3510				struct ocfs2_cached_dealloc_ctxt *dealloc,
3511				int index)
3512{
3513	int ret, i, subtree_index = 0, has_empty_extent = 0;
3514	unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters);
3515	struct ocfs2_extent_rec *left_rec;
3516	struct ocfs2_extent_rec *right_rec;
3517	struct ocfs2_extent_list *el = path_leaf_el(right_path);
3518	struct buffer_head *bh = path_leaf_bh(right_path);
3519	struct buffer_head *root_bh = NULL;
3520	struct ocfs2_path *left_path = NULL;
3521	struct ocfs2_extent_list *left_el;
3522
3523	BUG_ON(index < 0);
3524
3525	right_rec = &el->l_recs[index];
3526	if (index == 0) {
3527		/* we meet with a cross extent block merge. */
3528		ret = ocfs2_get_left_path(et, right_path, &left_path);
3529		if (ret) {
3530			mlog_errno(ret);
3531			return ret;
3532		}
3533
3534		left_el = path_leaf_el(left_path);
3535		BUG_ON(le16_to_cpu(left_el->l_next_free_rec) !=
3536		       le16_to_cpu(left_el->l_count));
3537
3538		left_rec = &left_el->l_recs[
3539				le16_to_cpu(left_el->l_next_free_rec) - 1];
3540		BUG_ON(le32_to_cpu(left_rec->e_cpos) +
3541		       le16_to_cpu(left_rec->e_leaf_clusters) !=
3542		       le32_to_cpu(split_rec->e_cpos));
3543
3544		subtree_index = ocfs2_find_subtree_root(et, left_path,
3545							right_path);
3546
3547		ret = ocfs2_extend_rotate_transaction(handle, subtree_index,
3548					jbd2_handle_buffer_credits(handle),
3549					left_path);
3550		if (ret) {
3551			mlog_errno(ret);
3552			goto out;
3553		}
3554
3555		root_bh = left_path->p_node[subtree_index].bh;
3556		BUG_ON(root_bh != right_path->p_node[subtree_index].bh);
3557
3558		ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3559						   subtree_index);
3560		if (ret) {
3561			mlog_errno(ret);
3562			goto out;
3563		}
3564
3565		for (i = subtree_index + 1;
3566		     i < path_num_items(right_path); i++) {
3567			ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3568							   right_path, i);
3569			if (ret) {
3570				mlog_errno(ret);
3571				goto out;
3572			}
3573
3574			ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
3575							   left_path, i);
3576			if (ret) {
3577				mlog_errno(ret);
3578				goto out;
3579			}
3580		}
3581	} else {
3582		left_rec = &el->l_recs[index - 1];
3583		if (ocfs2_is_empty_extent(&el->l_recs[0]))
3584			has_empty_extent = 1;
3585	}
3586
3587	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
3588					   path_num_items(right_path) - 1);
3589	if (ret) {
3590		mlog_errno(ret);
3591		goto out;
3592	}
3593
3594	if (has_empty_extent && index == 1) {
3595		/*
3596		 * The easy case - we can just plop the record right in.
3597		 */
3598		*left_rec = *split_rec;
3599	} else
3600		le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters);
3601
3602	le32_add_cpu(&right_rec->e_cpos, split_clusters);
3603	le64_add_cpu(&right_rec->e_blkno,
3604		     ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci),
3605					      split_clusters));
3606	le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters);
3607
3608	ocfs2_cleanup_merge(el, index);
3609
3610	ocfs2_journal_dirty(handle, bh);
3611	if (left_path) {
3612		ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
3613
3614		/*
3615		 * In the situation that the right_rec is empty and the extent
3616		 * block is empty also,  ocfs2_complete_edge_insert can't handle
3617		 * it and we need to delete the right extent block.
3618		 */
3619		if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 &&
3620		    le16_to_cpu(el->l_next_free_rec) == 1) {
3621			/* extend credit for ocfs2_remove_rightmost_path */
3622			ret = ocfs2_extend_rotate_transaction(handle, 0,
3623					jbd2_handle_buffer_credits(handle),
3624					right_path);
3625			if (ret) {
3626				mlog_errno(ret);
3627				goto out;
3628			}
3629
3630			ret = ocfs2_remove_rightmost_path(handle, et,
3631							  right_path,
3632							  dealloc);
3633			if (ret) {
3634				mlog_errno(ret);
3635				goto out;
3636			}
3637
3638			/* Now the rightmost extent block has been deleted.
3639			 * So we use the new rightmost path.
3640			 */
3641			ocfs2_mv_path(right_path, left_path);
3642			left_path = NULL;
3643		} else
3644			ocfs2_complete_edge_insert(handle, left_path,
3645						   right_path, subtree_index);
3646	}
3647out:
3648	ocfs2_free_path(left_path);
3649	return ret;
3650}
3651
3652static int ocfs2_try_to_merge_extent(handle_t *handle,
3653				     struct ocfs2_extent_tree *et,
3654				     struct ocfs2_path *path,
3655				     int split_index,
3656				     struct ocfs2_extent_rec *split_rec,
3657				     struct ocfs2_cached_dealloc_ctxt *dealloc,
3658				     struct ocfs2_merge_ctxt *ctxt)
3659{
3660	int ret = 0;
3661	struct ocfs2_extent_list *el = path_leaf_el(path);
3662	struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
3663
3664	BUG_ON(ctxt->c_contig_type == CONTIG_NONE);
3665
3666	if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) {
3667		/* extend credit for ocfs2_remove_rightmost_path */
3668		ret = ocfs2_extend_rotate_transaction(handle, 0,
3669				jbd2_handle_buffer_credits(handle),
3670				path);
3671		if (ret) {
3672			mlog_errno(ret);
3673			goto out;
3674		}
3675		/*
3676		 * The merge code will need to create an empty
3677		 * extent to take the place of the newly
3678		 * emptied slot. Remove any pre-existing empty
3679		 * extents - having more than one in a leaf is
3680		 * illegal.
3681		 */
3682		ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3683		if (ret) {
3684			mlog_errno(ret);
3685			goto out;
3686		}
3687		split_index--;
3688		rec = &el->l_recs[split_index];
3689	}
3690
3691	if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) {
3692		/*
3693		 * Left-right contig implies this.
3694		 */
3695		BUG_ON(!ctxt->c_split_covers_rec);
3696
3697		/*
3698		 * Since the leftright insert always covers the entire
3699		 * extent, this call will delete the insert record
3700		 * entirely, resulting in an empty extent record added to
3701		 * the extent block.
3702		 *
3703		 * Since the adding of an empty extent shifts
3704		 * everything back to the right, there's no need to
3705		 * update split_index here.
3706		 *
3707		 * When the split_index is zero, we need to merge it to the
3708		 * prevoius extent block. It is more efficient and easier
3709		 * if we do merge_right first and merge_left later.
3710		 */
3711		ret = ocfs2_merge_rec_right(path, handle, et, split_rec,
3712					    split_index);
3713		if (ret) {
3714			mlog_errno(ret);
3715			goto out;
3716		}
3717
3718		/*
3719		 * We can only get this from logic error above.
3720		 */
3721		BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0]));
3722
3723		/* extend credit for ocfs2_remove_rightmost_path */
3724		ret = ocfs2_extend_rotate_transaction(handle, 0,
3725					jbd2_handle_buffer_credits(handle),
3726					path);
3727		if (ret) {
3728			mlog_errno(ret);
3729			goto out;
3730		}
3731
3732		/* The merge left us with an empty extent, remove it. */
3733		ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3734		if (ret) {
3735			mlog_errno(ret);
3736			goto out;
3737		}
3738
3739		rec = &el->l_recs[split_index];
3740
3741		/*
3742		 * Note that we don't pass split_rec here on purpose -
3743		 * we've merged it into the rec already.
3744		 */
3745		ret = ocfs2_merge_rec_left(path, handle, et, rec,
3746					   dealloc, split_index);
3747
3748		if (ret) {
3749			mlog_errno(ret);
3750			goto out;
3751		}
3752
3753		/* extend credit for ocfs2_remove_rightmost_path */
3754		ret = ocfs2_extend_rotate_transaction(handle, 0,
3755				jbd2_handle_buffer_credits(handle),
3756				path);
3757		if (ret) {
3758			mlog_errno(ret);
3759			goto out;
3760		}
3761
3762		ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
3763		/*
3764		 * Error from this last rotate is not critical, so
3765		 * print but don't bubble it up.
3766		 */
3767		if (ret)
3768			mlog_errno(ret);
3769		ret = 0;
3770	} else {
3771		/*
3772		 * Merge a record to the left or right.
3773		 *
3774		 * 'contig_type' is relative to the existing record,
3775		 * so for example, if we're "right contig", it's to
3776		 * the record on the left (hence the left merge).
3777		 */
3778		if (ctxt->c_contig_type == CONTIG_RIGHT) {
3779			ret = ocfs2_merge_rec_left(path, handle, et,
3780						   split_rec, dealloc,
3781						   split_index);
3782			if (ret) {
3783				mlog_errno(ret);
3784				goto out;
3785			}
3786		} else {
3787			ret = ocfs2_merge_rec_right(path, handle,
3788						    et, split_rec,
3789						    split_index);
3790			if (ret) {
3791				mlog_errno(ret);
3792				goto out;
3793			}
3794		}
3795
3796		if (ctxt->c_split_covers_rec) {
3797			/* extend credit for ocfs2_remove_rightmost_path */
3798			ret = ocfs2_extend_rotate_transaction(handle, 0,
3799					jbd2_handle_buffer_credits(handle),
3800					path);
3801			if (ret) {
3802				mlog_errno(ret);
3803				ret = 0;
3804				goto out;
3805			}
3806
3807			/*
3808			 * The merge may have left an empty extent in
3809			 * our leaf. Try to rotate it away.
3810			 */
3811			ret = ocfs2_rotate_tree_left(handle, et, path,
3812						     dealloc);
3813			if (ret)
3814				mlog_errno(ret);
3815			ret = 0;
3816		}
3817	}
3818
3819out:
3820	return ret;
3821}
3822
3823static void ocfs2_subtract_from_rec(struct super_block *sb,
3824				    enum ocfs2_split_type split,
3825				    struct ocfs2_extent_rec *rec,
3826				    struct ocfs2_extent_rec *split_rec)
3827{
3828	u64 len_blocks;
3829
3830	len_blocks = ocfs2_clusters_to_blocks(sb,
3831				le16_to_cpu(split_rec->e_leaf_clusters));
3832
3833	if (split == SPLIT_LEFT) {
3834		/*
3835		 * Region is on the left edge of the existing
3836		 * record.
3837		 */
3838		le32_add_cpu(&rec->e_cpos,
3839			     le16_to_cpu(split_rec->e_leaf_clusters));
3840		le64_add_cpu(&rec->e_blkno, len_blocks);
3841		le16_add_cpu(&rec->e_leaf_clusters,
3842			     -le16_to_cpu(split_rec->e_leaf_clusters));
3843	} else {
3844		/*
3845		 * Region is on the right edge of the existing
3846		 * record.
3847		 */
3848		le16_add_cpu(&rec->e_leaf_clusters,
3849			     -le16_to_cpu(split_rec->e_leaf_clusters));
3850	}
3851}
3852
3853/*
3854 * Do the final bits of extent record insertion at the target leaf
3855 * list. If this leaf is part of an allocation tree, it is assumed
3856 * that the tree above has been prepared.
3857 */
3858static void ocfs2_insert_at_leaf(struct ocfs2_extent_tree *et,
3859				 struct ocfs2_extent_rec *insert_rec,
3860				 struct ocfs2_extent_list *el,
3861				 struct ocfs2_insert_type *insert)
3862{
3863	int i = insert->ins_contig_index;
3864	unsigned int range;
3865	struct ocfs2_extent_rec *rec;
3866
3867	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
3868
3869	if (insert->ins_split != SPLIT_NONE) {
3870		i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos));
3871		BUG_ON(i == -1);
3872		rec = &el->l_recs[i];
3873		ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
3874					insert->ins_split, rec,
3875					insert_rec);
3876		goto rotate;
3877	}
3878
3879	/*
3880	 * Contiguous insert - either left or right.
3881	 */
3882	if (insert->ins_contig != CONTIG_NONE) {
3883		rec = &el->l_recs[i];
3884		if (insert->ins_contig == CONTIG_LEFT) {
3885			rec->e_blkno = insert_rec->e_blkno;
3886			rec->e_cpos = insert_rec->e_cpos;
3887		}
3888		le16_add_cpu(&rec->e_leaf_clusters,
3889			     le16_to_cpu(insert_rec->e_leaf_clusters));
3890		return;
3891	}
3892
3893	/*
3894	 * Handle insert into an empty leaf.
3895	 */
3896	if (le16_to_cpu(el->l_next_free_rec) == 0 ||
3897	    ((le16_to_cpu(el->l_next_free_rec) == 1) &&
3898	     ocfs2_is_empty_extent(&el->l_recs[0]))) {
3899		el->l_recs[0] = *insert_rec;
3900		el->l_next_free_rec = cpu_to_le16(1);
3901		return;
3902	}
3903
3904	/*
3905	 * Appending insert.
3906	 */
3907	if (insert->ins_appending == APPEND_TAIL) {
3908		i = le16_to_cpu(el->l_next_free_rec) - 1;
3909		rec = &el->l_recs[i];
3910		range = le32_to_cpu(rec->e_cpos)
3911			+ le16_to_cpu(rec->e_leaf_clusters);
3912		BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
3913
3914		mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
3915				le16_to_cpu(el->l_count),
3916				"owner %llu, depth %u, count %u, next free %u, "
3917				"rec.cpos %u, rec.clusters %u, "
3918				"insert.cpos %u, insert.clusters %u\n",
3919				ocfs2_metadata_cache_owner(et->et_ci),
3920				le16_to_cpu(el->l_tree_depth),
3921				le16_to_cpu(el->l_count),
3922				le16_to_cpu(el->l_next_free_rec),
3923				le32_to_cpu(el->l_recs[i].e_cpos),
3924				le16_to_cpu(el->l_recs[i].e_leaf_clusters),
3925				le32_to_cpu(insert_rec->e_cpos),
3926				le16_to_cpu(insert_rec->e_leaf_clusters));
3927		i++;
3928		el->l_recs[i] = *insert_rec;
3929		le16_add_cpu(&el->l_next_free_rec, 1);
3930		return;
3931	}
3932
3933rotate:
3934	/*
3935	 * Ok, we have to rotate.
3936	 *
3937	 * At this point, it is safe to assume that inserting into an
3938	 * empty leaf and appending to a leaf have both been handled
3939	 * above.
3940	 *
3941	 * This leaf needs to have space, either by the empty 1st
3942	 * extent record, or by virtue of an l_next_free_rec < l_count.
3943	 */
3944	ocfs2_rotate_leaf(el, insert_rec);
3945}
3946
3947static void ocfs2_adjust_rightmost_records(handle_t *handle,
3948					   struct ocfs2_extent_tree *et,
3949					   struct ocfs2_path *path,
3950					   struct ocfs2_extent_rec *insert_rec)
3951{
3952	int i, next_free;
3953	struct buffer_head *bh;
3954	struct ocfs2_extent_list *el;
3955	struct ocfs2_extent_rec *rec;
3956
3957	/*
3958	 * Update everything except the leaf block.
3959	 */
3960	for (i = 0; i < path->p_tree_depth; i++) {
3961		bh = path->p_node[i].bh;
3962		el = path->p_node[i].el;
3963
3964		next_free = le16_to_cpu(el->l_next_free_rec);
3965		if (next_free == 0) {
3966			ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
3967				    "Owner %llu has a bad extent list\n",
3968				    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
3969			return;
3970		}
3971
3972		rec = &el->l_recs[next_free - 1];
3973
3974		rec->e_int_clusters = insert_rec->e_cpos;
3975		le32_add_cpu(&rec->e_int_clusters,
3976			     le16_to_cpu(insert_rec->e_leaf_clusters));
3977		le32_add_cpu(&rec->e_int_clusters,
3978			     -le32_to_cpu(rec->e_cpos));
3979
3980		ocfs2_journal_dirty(handle, bh);
3981	}
3982}
3983
3984static int ocfs2_append_rec_to_path(handle_t *handle,
3985				    struct ocfs2_extent_tree *et,
3986				    struct ocfs2_extent_rec *insert_rec,
3987				    struct ocfs2_path *right_path,
3988				    struct ocfs2_path **ret_left_path)
3989{
3990	int ret, next_free;
3991	struct ocfs2_extent_list *el;
3992	struct ocfs2_path *left_path = NULL;
3993
3994	*ret_left_path = NULL;
3995
3996	/*
3997	 * This shouldn't happen for non-trees. The extent rec cluster
3998	 * count manipulation below only works for interior nodes.
3999	 */
4000	BUG_ON(right_path->p_tree_depth == 0);
4001
4002	/*
4003	 * If our appending insert is at the leftmost edge of a leaf,
4004	 * then we might need to update the rightmost records of the
4005	 * neighboring path.
4006	 */
4007	el = path_leaf_el(right_path);
4008	next_free = le16_to_cpu(el->l_next_free_rec);
4009	if (next_free == 0 ||
4010	    (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
4011		u32 left_cpos;
4012
4013		ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
4014						    right_path, &left_cpos);
4015		if (ret) {
4016			mlog_errno(ret);
4017			goto out;
4018		}
4019
4020		trace_ocfs2_append_rec_to_path(
4021			(unsigned long long)
4022			ocfs2_metadata_cache_owner(et->et_ci),
4023			le32_to_cpu(insert_rec->e_cpos),
4024			left_cpos);
4025
4026		/*
4027		 * No need to worry if the append is already in the
4028		 * leftmost leaf.
4029		 */
4030		if (left_cpos) {
4031			left_path = ocfs2_new_path_from_path(right_path);
4032			if (!left_path) {
4033				ret = -ENOMEM;
4034				mlog_errno(ret);
4035				goto out;
4036			}
4037
4038			ret = ocfs2_find_path(et->et_ci, left_path,
4039					      left_cpos);
4040			if (ret) {
4041				mlog_errno(ret);
4042				goto out;
4043			}
4044
4045			/*
4046			 * ocfs2_insert_path() will pass the left_path to the
4047			 * journal for us.
4048			 */
4049		}
4050	}
4051
4052	ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4053	if (ret) {
4054		mlog_errno(ret);
4055		goto out;
4056	}
4057
4058	ocfs2_adjust_rightmost_records(handle, et, right_path, insert_rec);
4059
4060	*ret_left_path = left_path;
4061	ret = 0;
4062out:
4063	if (ret != 0)
4064		ocfs2_free_path(left_path);
4065
4066	return ret;
4067}
4068
4069static void ocfs2_split_record(struct ocfs2_extent_tree *et,
4070			       struct ocfs2_path *left_path,
4071			       struct ocfs2_path *right_path,
4072			       struct ocfs2_extent_rec *split_rec,
4073			       enum ocfs2_split_type split)
4074{
4075	int index;
4076	u32 cpos = le32_to_cpu(split_rec->e_cpos);
4077	struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el;
4078	struct ocfs2_extent_rec *rec, *tmprec;
4079
4080	right_el = path_leaf_el(right_path);
4081	if (left_path)
4082		left_el = path_leaf_el(left_path);
4083
4084	el = right_el;
4085	insert_el = right_el;
4086	index = ocfs2_search_extent_list(el, cpos);
4087	if (index != -1) {
4088		if (index == 0 && left_path) {
4089			BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0]));
4090
4091			/*
4092			 * This typically means that the record
4093			 * started in the left path but moved to the
4094			 * right as a result of rotation. We either
4095			 * move the existing record to the left, or we
4096			 * do the later insert there.
4097			 *
4098			 * In this case, the left path should always
4099			 * exist as the rotate code will have passed
4100			 * it back for a post-insert update.
4101			 */
4102
4103			if (split == SPLIT_LEFT) {
4104				/*
4105				 * It's a left split. Since we know
4106				 * that the rotate code gave us an
4107				 * empty extent in the left path, we
4108				 * can just do the insert there.
4109				 */
4110				insert_el = left_el;
4111			} else {
4112				/*
4113				 * Right split - we have to move the
4114				 * existing record over to the left
4115				 * leaf. The insert will be into the
4116				 * newly created empty extent in the
4117				 * right leaf.
4118				 */
4119				tmprec = &right_el->l_recs[index];
4120				ocfs2_rotate_leaf(left_el, tmprec);
4121				el = left_el;
4122
4123				memset(tmprec, 0, sizeof(*tmprec));
4124				index = ocfs2_search_extent_list(left_el, cpos);
4125				BUG_ON(index == -1);
4126			}
4127		}
4128	} else {
4129		BUG_ON(!left_path);
4130		BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0]));
4131		/*
4132		 * Left path is easy - we can just allow the insert to
4133		 * happen.
4134		 */
4135		el = left_el;
4136		insert_el = left_el;
4137		index = ocfs2_search_extent_list(el, cpos);
4138		BUG_ON(index == -1);
4139	}
4140
4141	rec = &el->l_recs[index];
4142	ocfs2_subtract_from_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4143				split, rec, split_rec);
4144	ocfs2_rotate_leaf(insert_el, split_rec);
4145}
4146
4147/*
4148 * This function only does inserts on an allocation b-tree. For tree
4149 * depth = 0, ocfs2_insert_at_leaf() is called directly.
4150 *
4151 * right_path is the path we want to do the actual insert
4152 * in. left_path should only be passed in if we need to update that
4153 * portion of the tree after an edge insert.
4154 */
4155static int ocfs2_insert_path(handle_t *handle,
4156			     struct ocfs2_extent_tree *et,
4157			     struct ocfs2_path *left_path,
4158			     struct ocfs2_path *right_path,
4159			     struct ocfs2_extent_rec *insert_rec,
4160			     struct ocfs2_insert_type *insert)
4161{
4162	int ret, subtree_index;
4163	struct buffer_head *leaf_bh = path_leaf_bh(right_path);
4164
4165	if (left_path) {
4166		/*
4167		 * There's a chance that left_path got passed back to
4168		 * us without being accounted for in the
4169		 * journal. Extend our transaction here to be sure we
4170		 * can change those blocks.
4171		 */
4172		ret = ocfs2_extend_trans(handle, left_path->p_tree_depth);
4173		if (ret < 0) {
4174			mlog_errno(ret);
4175			goto out;
4176		}
4177
4178		ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
4179		if (ret < 0) {
4180			mlog_errno(ret);
4181			goto out;
4182		}
4183	}
4184
4185	/*
4186	 * Pass both paths to the journal. The majority of inserts
4187	 * will be touching all components anyway.
4188	 */
4189	ret = ocfs2_journal_access_path(et->et_ci, handle, right_path);
4190	if (ret < 0) {
4191		mlog_errno(ret);
4192		goto out;
4193	}
4194
4195	if (insert->ins_split != SPLIT_NONE) {
4196		/*
4197		 * We could call ocfs2_insert_at_leaf() for some types
4198		 * of splits, but it's easier to just let one separate
4199		 * function sort it all out.
4200		 */
4201		ocfs2_split_record(et, left_path, right_path,
4202				   insert_rec, insert->ins_split);
4203
4204		/*
4205		 * Split might have modified either leaf and we don't
4206		 * have a guarantee that the later edge insert will
4207		 * dirty this for us.
4208		 */
4209		if (left_path)
4210			ocfs2_journal_dirty(handle,
4211					    path_leaf_bh(left_path));
4212	} else
4213		ocfs2_insert_at_leaf(et, insert_rec, path_leaf_el(right_path),
4214				     insert);
4215
4216	ocfs2_journal_dirty(handle, leaf_bh);
4217
4218	if (left_path) {
4219		/*
4220		 * The rotate code has indicated that we need to fix
4221		 * up portions of the tree after the insert.
4222		 *
4223		 * XXX: Should we extend the transaction here?
4224		 */
4225		subtree_index = ocfs2_find_subtree_root(et, left_path,
4226							right_path);
4227		ocfs2_complete_edge_insert(handle, left_path, right_path,
4228					   subtree_index);
4229	}
4230
4231	ret = 0;
4232out:
4233	return ret;
4234}
4235
4236static int ocfs2_do_insert_extent(handle_t *handle,
4237				  struct ocfs2_extent_tree *et,
4238				  struct ocfs2_extent_rec *insert_rec,
4239				  struct ocfs2_insert_type *type)
4240{
4241	int ret, rotate = 0;
4242	u32 cpos;
4243	struct ocfs2_path *right_path = NULL;
4244	struct ocfs2_path *left_path = NULL;
4245	struct ocfs2_extent_list *el;
4246
4247	el = et->et_root_el;
4248
4249	ret = ocfs2_et_root_journal_access(handle, et,
4250					   OCFS2_JOURNAL_ACCESS_WRITE);
4251	if (ret) {
4252		mlog_errno(ret);
4253		goto out;
4254	}
4255
4256	if (le16_to_cpu(el->l_tree_depth) == 0) {
4257		ocfs2_insert_at_leaf(et, insert_rec, el, type);
4258		goto out_update_clusters;
4259	}
4260
4261	right_path = ocfs2_new_path_from_et(et);
4262	if (!right_path) {
4263		ret = -ENOMEM;
4264		mlog_errno(ret);
4265		goto out;
4266	}
4267
4268	/*
4269	 * Determine the path to start with. Rotations need the
4270	 * rightmost path, everything else can go directly to the
4271	 * target leaf.
4272	 */
4273	cpos = le32_to_cpu(insert_rec->e_cpos);
4274	if (type->ins_appending == APPEND_NONE &&
4275	    type->ins_contig == CONTIG_NONE) {
4276		rotate = 1;
4277		cpos = UINT_MAX;
4278	}
4279
4280	ret = ocfs2_find_path(et->et_ci, right_path, cpos);
4281	if (ret) {
4282		mlog_errno(ret);
4283		goto out;
4284	}
4285
4286	/*
4287	 * Rotations and appends need special treatment - they modify
4288	 * parts of the tree's above them.
4289	 *
4290	 * Both might pass back a path immediate to the left of the
4291	 * one being inserted to. This will be cause
4292	 * ocfs2_insert_path() to modify the rightmost records of
4293	 * left_path to account for an edge insert.
4294	 *
4295	 * XXX: When modifying this code, keep in mind that an insert
4296	 * can wind up skipping both of these two special cases...
4297	 */
4298	if (rotate) {
4299		ret = ocfs2_rotate_tree_right(handle, et, type->ins_split,
4300					      le32_to_cpu(insert_rec->e_cpos),
4301					      right_path, &left_path);
4302		if (ret) {
4303			mlog_errno(ret);
4304			goto out;
4305		}
4306
4307		/*
4308		 * ocfs2_rotate_tree_right() might have extended the
4309		 * transaction without re-journaling our tree root.
4310		 */
4311		ret = ocfs2_et_root_journal_access(handle, et,
4312						   OCFS2_JOURNAL_ACCESS_WRITE);
4313		if (ret) {
4314			mlog_errno(ret);
4315			goto out;
4316		}
4317	} else if (type->ins_appending == APPEND_TAIL
4318		   && type->ins_contig != CONTIG_LEFT) {
4319		ret = ocfs2_append_rec_to_path(handle, et, insert_rec,
4320					       right_path, &left_path);
4321		if (ret) {
4322			mlog_errno(ret);
4323			goto out;
4324		}
4325	}
4326
4327	ret = ocfs2_insert_path(handle, et, left_path, right_path,
4328				insert_rec, type);
4329	if (ret) {
4330		mlog_errno(ret);
4331		goto out;
4332	}
4333
4334out_update_clusters:
4335	if (type->ins_split == SPLIT_NONE)
4336		ocfs2_et_update_clusters(et,
4337					 le16_to_cpu(insert_rec->e_leaf_clusters));
4338
4339	ocfs2_journal_dirty(handle, et->et_root_bh);
4340
4341out:
4342	ocfs2_free_path(left_path);
4343	ocfs2_free_path(right_path);
4344
4345	return ret;
4346}
4347
4348static int ocfs2_figure_merge_contig_type(struct ocfs2_extent_tree *et,
4349			       struct ocfs2_path *path,
4350			       struct ocfs2_extent_list *el, int index,
4351			       struct ocfs2_extent_rec *split_rec,
4352			       struct ocfs2_merge_ctxt *ctxt)
4353{
4354	int status = 0;
4355	enum ocfs2_contig_type ret = CONTIG_NONE;
4356	u32 left_cpos, right_cpos;
4357	struct ocfs2_extent_rec *rec = NULL;
4358	struct ocfs2_extent_list *new_el;
4359	struct ocfs2_path *left_path = NULL, *right_path = NULL;
4360	struct buffer_head *bh;
4361	struct ocfs2_extent_block *eb;
4362	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
4363
4364	if (index > 0) {
4365		rec = &el->l_recs[index - 1];
4366	} else if (path->p_tree_depth > 0) {
4367		status = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
4368		if (status)
4369			goto exit;
4370
4371		if (left_cpos != 0) {
4372			left_path = ocfs2_new_path_from_path(path);
4373			if (!left_path) {
4374				status = -ENOMEM;
4375				mlog_errno(status);
4376				goto exit;
4377			}
4378
4379			status = ocfs2_find_path(et->et_ci, left_path,
4380						 left_cpos);
4381			if (status)
4382				goto free_left_path;
4383
4384			new_el = path_leaf_el(left_path);
4385
4386			if (le16_to_cpu(new_el->l_next_free_rec) !=
4387			    le16_to_cpu(new_el->l_count)) {
4388				bh = path_leaf_bh(left_path);
4389				eb = (struct ocfs2_extent_block *)bh->b_data;
4390				status = ocfs2_error(sb,
4391						"Extent block #%llu has an invalid l_next_free_rec of %d.  It should have matched the l_count of %d\n",
4392						(unsigned long long)le64_to_cpu(eb->h_blkno),
4393						le16_to_cpu(new_el->l_next_free_rec),
4394						le16_to_cpu(new_el->l_count));
4395				goto free_left_path;
4396			}
4397			rec = &new_el->l_recs[
4398				le16_to_cpu(new_el->l_next_free_rec) - 1];
4399		}
4400	}
4401
4402	/*
4403	 * We're careful to check for an empty extent record here -
4404	 * the merge code will know what to do if it sees one.
4405	 */
4406	if (rec) {
4407		if (index == 1 && ocfs2_is_empty_extent(rec)) {
4408			if (split_rec->e_cpos == el->l_recs[index].e_cpos)
4409				ret = CONTIG_RIGHT;
4410		} else {
4411			ret = ocfs2_et_extent_contig(et, rec, split_rec);
4412		}
4413	}
4414
4415	rec = NULL;
4416	if (index < (le16_to_cpu(el->l_next_free_rec) - 1))
4417		rec = &el->l_recs[index + 1];
4418	else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) &&
4419		 path->p_tree_depth > 0) {
4420		status = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
4421		if (status)
4422			goto free_left_path;
4423
4424		if (right_cpos == 0)
4425			goto free_left_path;
4426
4427		right_path = ocfs2_new_path_from_path(path);
4428		if (!right_path) {
4429			status = -ENOMEM;
4430			mlog_errno(status);
4431			goto free_left_path;
4432		}
4433
4434		status = ocfs2_find_path(et->et_ci, right_path, right_cpos);
4435		if (status)
4436			goto free_right_path;
4437
4438		new_el = path_leaf_el(right_path);
4439		rec = &new_el->l_recs[0];
4440		if (ocfs2_is_empty_extent(rec)) {
4441			if (le16_to_cpu(new_el->l_next_free_rec) <= 1) {
4442				bh = path_leaf_bh(right_path);
4443				eb = (struct ocfs2_extent_block *)bh->b_data;
4444				status = ocfs2_error(sb,
4445						"Extent block #%llu has an invalid l_next_free_rec of %d\n",
4446						(unsigned long long)le64_to_cpu(eb->h_blkno),
4447						le16_to_cpu(new_el->l_next_free_rec));
4448				goto free_right_path;
4449			}
4450			rec = &new_el->l_recs[1];
4451		}
4452	}
4453
4454	if (rec) {
4455		enum ocfs2_contig_type contig_type;
4456
4457		contig_type = ocfs2_et_extent_contig(et, rec, split_rec);
4458
4459		if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT)
4460			ret = CONTIG_LEFTRIGHT;
4461		else if (ret == CONTIG_NONE)
4462			ret = contig_type;
4463	}
4464
4465free_right_path:
4466	ocfs2_free_path(right_path);
4467free_left_path:
4468	ocfs2_free_path(left_path);
4469exit:
4470	if (status == 0)
4471		ctxt->c_contig_type = ret;
4472
4473	return status;
4474}
4475
4476static void ocfs2_figure_contig_type(struct ocfs2_extent_tree *et,
4477				     struct ocfs2_insert_type *insert,
4478				     struct ocfs2_extent_list *el,
4479				     struct ocfs2_extent_rec *insert_rec)
4480{
4481	int i;
4482	enum ocfs2_contig_type contig_type = CONTIG_NONE;
4483
4484	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4485
4486	for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
4487		contig_type = ocfs2_et_extent_contig(et, &el->l_recs[i],
4488						     insert_rec);
4489		if (contig_type != CONTIG_NONE) {
4490			insert->ins_contig_index = i;
4491			break;
4492		}
4493	}
4494	insert->ins_contig = contig_type;
4495
4496	if (insert->ins_contig != CONTIG_NONE) {
4497		struct ocfs2_extent_rec *rec =
4498				&el->l_recs[insert->ins_contig_index];
4499		unsigned int len = le16_to_cpu(rec->e_leaf_clusters) +
4500				   le16_to_cpu(insert_rec->e_leaf_clusters);
4501
4502		/*
4503		 * Caller might want us to limit the size of extents, don't
4504		 * calculate contiguousness if we might exceed that limit.
4505		 */
4506		if (et->et_max_leaf_clusters &&
4507		    (len > et->et_max_leaf_clusters))
4508			insert->ins_contig = CONTIG_NONE;
4509	}
4510}
4511
4512/*
4513 * This should only be called against the righmost leaf extent list.
4514 *
4515 * ocfs2_figure_appending_type() will figure out whether we'll have to
4516 * insert at the tail of the rightmost leaf.
4517 *
4518 * This should also work against the root extent list for tree's with 0
4519 * depth. If we consider the root extent list to be the rightmost leaf node
4520 * then the logic here makes sense.
4521 */
4522static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
4523					struct ocfs2_extent_list *el,
4524					struct ocfs2_extent_rec *insert_rec)
4525{
4526	int i;
4527	u32 cpos = le32_to_cpu(insert_rec->e_cpos);
4528	struct ocfs2_extent_rec *rec;
4529
4530	insert->ins_appending = APPEND_NONE;
4531
4532	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);
4533
4534	if (!el->l_next_free_rec)
4535		goto set_tail_append;
4536
4537	if (ocfs2_is_empty_extent(&el->l_recs[0])) {
4538		/* Were all records empty? */
4539		if (le16_to_cpu(el->l_next_free_rec) == 1)
4540			goto set_tail_append;
4541	}
4542
4543	i = le16_to_cpu(el->l_next_free_rec) - 1;
4544	rec = &el->l_recs[i];
4545
4546	if (cpos >=
4547	    (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)))
4548		goto set_tail_append;
4549
4550	return;
4551
4552set_tail_append:
4553	insert->ins_appending = APPEND_TAIL;
4554}
4555
4556/*
4557 * Helper function called at the beginning of an insert.
4558 *
4559 * This computes a few things that are commonly used in the process of
4560 * inserting into the btree:
4561 *   - Whether the new extent is contiguous with an existing one.
4562 *   - The current tree depth.
4563 *   - Whether the insert is an appending one.
4564 *   - The total # of free records in the tree.
4565 *
4566 * All of the information is stored on the ocfs2_insert_type
4567 * structure.
4568 */
4569static int ocfs2_figure_insert_type(struct ocfs2_extent_tree *et,
4570				    struct buffer_head **last_eb_bh,
4571				    struct ocfs2_extent_rec *insert_rec,
4572				    int *free_records,
4573				    struct ocfs2_insert_type *insert)
4574{
4575	int ret;
4576	struct ocfs2_extent_block *eb;
4577	struct ocfs2_extent_list *el;
4578	struct ocfs2_path *path = NULL;
4579	struct buffer_head *bh = NULL;
4580
4581	insert->ins_split = SPLIT_NONE;
4582
4583	el = et->et_root_el;
4584	insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
4585
4586	if (el->l_tree_depth) {
4587		/*
4588		 * If we have tree depth, we read in the
4589		 * rightmost extent block ahead of time as
4590		 * ocfs2_figure_insert_type() and ocfs2_add_branch()
4591		 * may want it later.
4592		 */
4593		ret = ocfs2_read_extent_block(et->et_ci,
4594					      ocfs2_et_get_last_eb_blk(et),
4595					      &bh);
4596		if (ret) {
4597			mlog_errno(ret);
4598			goto out;
4599		}
4600		eb = (struct ocfs2_extent_block *) bh->b_data;
4601		el = &eb->h_list;
4602	}
4603
4604	/*
4605	 * Unless we have a contiguous insert, we'll need to know if
4606	 * there is room left in our allocation tree for another
4607	 * extent record.
4608	 *
4609	 * XXX: This test is simplistic, we can search for empty
4610	 * extent records too.
4611	 */
4612	*free_records = le16_to_cpu(el->l_count) -
4613		le16_to_cpu(el->l_next_free_rec);
4614
4615	if (!insert->ins_tree_depth) {
4616		ocfs2_figure_contig_type(et, insert, el, insert_rec);
4617		ocfs2_figure_appending_type(insert, el, insert_rec);
4618		return 0;
4619	}
4620
4621	path = ocfs2_new_path_from_et(et);
4622	if (!path) {
4623		ret = -ENOMEM;
4624		mlog_errno(ret);
4625		goto out;
4626	}
4627
4628	/*
4629	 * In the case that we're inserting past what the tree
4630	 * currently accounts for, ocfs2_find_path() will return for
4631	 * us the rightmost tree path. This is accounted for below in
4632	 * the appending code.
4633	 */
4634	ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos));
4635	if (ret) {
4636		mlog_errno(ret);
4637		goto out;
4638	}
4639
4640	el = path_leaf_el(path);
4641
4642	/*
4643	 * Now that we have the path, there's two things we want to determine:
4644	 * 1) Contiguousness (also set contig_index if this is so)
4645	 *
4646	 * 2) Are we doing an append? We can trivially break this up
4647         *     into two types of appends: simple record append, or a
4648         *     rotate inside the tail leaf.
4649	 */
4650	ocfs2_figure_contig_type(et, insert, el, insert_rec);
4651
4652	/*
4653	 * The insert code isn't quite ready to deal with all cases of
4654	 * left contiguousness. Specifically, if it's an insert into
4655	 * the 1st record in a leaf, it will require the adjustment of
4656	 * cluster count on the last record of the path directly to it's
4657	 * left. For now, just catch that case and fool the layers
4658	 * above us. This works just fine for tree_depth == 0, which
4659	 * is why we allow that above.
4660	 */
4661	if (insert->ins_contig == CONTIG_LEFT &&
4662	    insert->ins_contig_index == 0)
4663		insert->ins_contig = CONTIG_NONE;
4664
4665	/*
4666	 * Ok, so we can simply compare against last_eb to figure out
4667	 * whether the path doesn't exist. This will only happen in
4668	 * the case that we're doing a tail append, so maybe we can
4669	 * take advantage of that information somehow.
4670	 */
4671	if (ocfs2_et_get_last_eb_blk(et) ==
4672	    path_leaf_bh(path)->b_blocknr) {
4673		/*
4674		 * Ok, ocfs2_find_path() returned us the rightmost
4675		 * tree path. This might be an appending insert. There are
4676		 * two cases:
4677		 *    1) We're doing a true append at the tail:
4678		 *	-This might even be off the end of the leaf
4679		 *    2) We're "appending" by rotating in the tail
4680		 */
4681		ocfs2_figure_appending_type(insert, el, insert_rec);
4682	}
4683
4684out:
4685	ocfs2_free_path(path);
4686
4687	if (ret == 0)
4688		*last_eb_bh = bh;
4689	else
4690		brelse(bh);
4691	return ret;
4692}
4693
4694/*
4695 * Insert an extent into a btree.
4696 *
4697 * The caller needs to update the owning btree's cluster count.
4698 */
4699int ocfs2_insert_extent(handle_t *handle,
4700			struct ocfs2_extent_tree *et,
4701			u32 cpos,
4702			u64 start_blk,
4703			u32 new_clusters,
4704			u8 flags,
4705			struct ocfs2_alloc_context *meta_ac)
4706{
4707	int status;
4708	int free_records;
4709	struct buffer_head *last_eb_bh = NULL;
4710	struct ocfs2_insert_type insert = {0, };
4711	struct ocfs2_extent_rec rec;
4712
4713	trace_ocfs2_insert_extent_start(
4714		(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4715		cpos, new_clusters);
4716
4717	memset(&rec, 0, sizeof(rec));
4718	rec.e_cpos = cpu_to_le32(cpos);
4719	rec.e_blkno = cpu_to_le64(start_blk);
4720	rec.e_leaf_clusters = cpu_to_le16(new_clusters);
4721	rec.e_flags = flags;
4722	status = ocfs2_et_insert_check(et, &rec);
4723	if (status) {
4724		mlog_errno(status);
4725		goto bail;
4726	}
4727
4728	status = ocfs2_figure_insert_type(et, &last_eb_bh, &rec,
4729					  &free_records, &insert);
4730	if (status < 0) {
4731		mlog_errno(status);
4732		goto bail;
4733	}
4734
4735	trace_ocfs2_insert_extent(insert.ins_appending, insert.ins_contig,
4736				  insert.ins_contig_index, free_records,
4737				  insert.ins_tree_depth);
4738
4739	if (insert.ins_contig == CONTIG_NONE && free_records == 0) {
4740		status = ocfs2_grow_tree(handle, et,
4741					 &insert.ins_tree_depth, &last_eb_bh,
4742					 meta_ac);
4743		if (status) {
4744			mlog_errno(status);
4745			goto bail;
4746		}
4747	}
4748
4749	/* Finally, we can add clusters. This might rotate the tree for us. */
4750	status = ocfs2_do_insert_extent(handle, et, &rec, &insert);
4751	if (status < 0)
4752		mlog_errno(status);
4753	else
4754		ocfs2_et_extent_map_insert(et, &rec);
4755
4756bail:
4757	brelse(last_eb_bh);
4758
4759	return status;
4760}
4761
4762/*
4763 * Allcate and add clusters into the extent b-tree.
4764 * The new clusters(clusters_to_add) will be inserted at logical_offset.
4765 * The extent b-tree's root is specified by et, and
4766 * it is not limited to the file storage. Any extent tree can use this
4767 * function if it implements the proper ocfs2_extent_tree.
4768 */
4769int ocfs2_add_clusters_in_btree(handle_t *handle,
4770				struct ocfs2_extent_tree *et,
4771				u32 *logical_offset,
4772				u32 clusters_to_add,
4773				int mark_unwritten,
4774				struct ocfs2_alloc_context *data_ac,
4775				struct ocfs2_alloc_context *meta_ac,
4776				enum ocfs2_alloc_restarted *reason_ret)
4777{
4778	int status = 0, err = 0;
4779	int need_free = 0;
4780	int free_extents;
4781	enum ocfs2_alloc_restarted reason = RESTART_NONE;
4782	u32 bit_off, num_bits;
4783	u64 block;
4784	u8 flags = 0;
4785	struct ocfs2_super *osb =
4786		OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
4787
4788	BUG_ON(!clusters_to_add);
4789
4790	if (mark_unwritten)
4791		flags = OCFS2_EXT_UNWRITTEN;
4792
4793	free_extents = ocfs2_num_free_extents(et);
4794	if (free_extents < 0) {
4795		status = free_extents;
4796		mlog_errno(status);
4797		goto leave;
4798	}
4799
4800	/* there are two cases which could cause us to EAGAIN in the
4801	 * we-need-more-metadata case:
4802	 * 1) we haven't reserved *any*
4803	 * 2) we are so fragmented, we've needed to add metadata too
4804	 *    many times. */
4805	if (!free_extents && !meta_ac) {
4806		err = -1;
4807		status = -EAGAIN;
4808		reason = RESTART_META;
4809		goto leave;
4810	} else if ((!free_extents)
4811		   && (ocfs2_alloc_context_bits_left(meta_ac)
4812		       < ocfs2_extend_meta_needed(et->et_root_el))) {
4813		err = -2;
4814		status = -EAGAIN;
4815		reason = RESTART_META;
4816		goto leave;
4817	}
4818
4819	status = __ocfs2_claim_clusters(handle, data_ac, 1,
4820					clusters_to_add, &bit_off, &num_bits);
4821	if (status < 0) {
4822		if (status != -ENOSPC)
4823			mlog_errno(status);
4824		goto leave;
4825	}
4826
4827	BUG_ON(num_bits > clusters_to_add);
4828
4829	/* reserve our write early -- insert_extent may update the tree root */
4830	status = ocfs2_et_root_journal_access(handle, et,
4831					      OCFS2_JOURNAL_ACCESS_WRITE);
4832	if (status < 0) {
4833		mlog_errno(status);
4834		need_free = 1;
4835		goto bail;
4836	}
4837
4838	block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
4839	trace_ocfs2_add_clusters_in_btree(
4840	     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
4841	     bit_off, num_bits);
4842	status = ocfs2_insert_extent(handle, et, *logical_offset, block,
4843				     num_bits, flags, meta_ac);
4844	if (status < 0) {
4845		mlog_errno(status);
4846		need_free = 1;
4847		goto bail;
4848	}
4849
4850	ocfs2_journal_dirty(handle, et->et_root_bh);
4851
4852	clusters_to_add -= num_bits;
4853	*logical_offset += num_bits;
4854
4855	if (clusters_to_add) {
4856		err = clusters_to_add;
4857		status = -EAGAIN;
4858		reason = RESTART_TRANS;
4859	}
4860
4861bail:
4862	if (need_free) {
4863		if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
4864			ocfs2_free_local_alloc_bits(osb, handle, data_ac,
4865					bit_off, num_bits);
4866		else
4867			ocfs2_free_clusters(handle,
4868					data_ac->ac_inode,
4869					data_ac->ac_bh,
4870					ocfs2_clusters_to_blocks(osb->sb, bit_off),
4871					num_bits);
4872	}
4873
4874leave:
4875	if (reason_ret)
4876		*reason_ret = reason;
4877	trace_ocfs2_add_clusters_in_btree_ret(status, reason, err);
4878	return status;
4879}
4880
4881static void ocfs2_make_right_split_rec(struct super_block *sb,
4882				       struct ocfs2_extent_rec *split_rec,
4883				       u32 cpos,
4884				       struct ocfs2_extent_rec *rec)
4885{
4886	u32 rec_cpos = le32_to_cpu(rec->e_cpos);
4887	u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters);
4888
4889	memset(split_rec, 0, sizeof(struct ocfs2_extent_rec));
4890
4891	split_rec->e_cpos = cpu_to_le32(cpos);
4892	split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos);
4893
4894	split_rec->e_blkno = rec->e_blkno;
4895	le64_add_cpu(&split_rec->e_blkno,
4896		     ocfs2_clusters_to_blocks(sb, cpos - rec_cpos));
4897
4898	split_rec->e_flags = rec->e_flags;
4899}
4900
4901static int ocfs2_split_and_insert(handle_t *handle,
4902				  struct ocfs2_extent_tree *et,
4903				  struct ocfs2_path *path,
4904				  struct buffer_head **last_eb_bh,
4905				  int split_index,
4906				  struct ocfs2_extent_rec *orig_split_rec,
4907				  struct ocfs2_alloc_context *meta_ac)
4908{
4909	int ret = 0, depth;
4910	unsigned int insert_range, rec_range, do_leftright = 0;
4911	struct ocfs2_extent_rec tmprec;
4912	struct ocfs2_extent_list *rightmost_el;
4913	struct ocfs2_extent_rec rec;
4914	struct ocfs2_extent_rec split_rec = *orig_split_rec;
4915	struct ocfs2_insert_type insert;
4916	struct ocfs2_extent_block *eb;
4917
4918leftright:
4919	/*
4920	 * Store a copy of the record on the stack - it might move
4921	 * around as the tree is manipulated below.
4922	 */
4923	rec = path_leaf_el(path)->l_recs[split_index];
4924
4925	rightmost_el = et->et_root_el;
4926
4927	depth = le16_to_cpu(rightmost_el->l_tree_depth);
4928	if (depth) {
4929		BUG_ON(!(*last_eb_bh));
4930		eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
4931		rightmost_el = &eb->h_list;
4932	}
4933
4934	if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
4935	    le16_to_cpu(rightmost_el->l_count)) {
4936		ret = ocfs2_grow_tree(handle, et,
4937				      &depth, last_eb_bh, meta_ac);
4938		if (ret) {
4939			mlog_errno(ret);
4940			goto out;
4941		}
4942	}
4943
4944	memset(&insert, 0, sizeof(struct ocfs2_insert_type));
4945	insert.ins_appending = APPEND_NONE;
4946	insert.ins_contig = CONTIG_NONE;
4947	insert.ins_tree_depth = depth;
4948
4949	insert_range = le32_to_cpu(split_rec.e_cpos) +
4950		le16_to_cpu(split_rec.e_leaf_clusters);
4951	rec_range = le32_to_cpu(rec.e_cpos) +
4952		le16_to_cpu(rec.e_leaf_clusters);
4953
4954	if (split_rec.e_cpos == rec.e_cpos) {
4955		insert.ins_split = SPLIT_LEFT;
4956	} else if (insert_range == rec_range) {
4957		insert.ins_split = SPLIT_RIGHT;
4958	} else {
4959		/*
4960		 * Left/right split. We fake this as a right split
4961		 * first and then make a second pass as a left split.
4962		 */
4963		insert.ins_split = SPLIT_RIGHT;
4964
4965		ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
4966					   &tmprec, insert_range, &rec);
4967
4968		split_rec = tmprec;
4969
4970		BUG_ON(do_leftright);
4971		do_leftright = 1;
4972	}
4973
4974	ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
4975	if (ret) {
4976		mlog_errno(ret);
4977		goto out;
4978	}
4979
4980	if (do_leftright == 1) {
4981		u32 cpos;
4982		struct ocfs2_extent_list *el;
4983
4984		do_leftright++;
4985		split_rec = *orig_split_rec;
4986
4987		ocfs2_reinit_path(path, 1);
4988
4989		cpos = le32_to_cpu(split_rec.e_cpos);
4990		ret = ocfs2_find_path(et->et_ci, path, cpos);
4991		if (ret) {
4992			mlog_errno(ret);
4993			goto out;
4994		}
4995
4996		el = path_leaf_el(path);
4997		split_index = ocfs2_search_extent_list(el, cpos);
4998		if (split_index == -1) {
4999			ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5000				    "Owner %llu has an extent at cpos %u which can no longer be found\n",
5001				    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5002				    cpos);
5003			ret = -EROFS;
5004			goto out;
5005		}
5006		goto leftright;
5007	}
5008out:
5009
5010	return ret;
5011}
5012
5013static int ocfs2_replace_extent_rec(handle_t *handle,
5014				    struct ocfs2_extent_tree *et,
5015				    struct ocfs2_path *path,
5016				    struct ocfs2_extent_list *el,
5017				    int split_index,
5018				    struct ocfs2_extent_rec *split_rec)
5019{
5020	int ret;
5021
5022	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
5023					   path_num_items(path) - 1);
5024	if (ret) {
5025		mlog_errno(ret);
5026		goto out;
5027	}
5028
5029	el->l_recs[split_index] = *split_rec;
5030
5031	ocfs2_journal_dirty(handle, path_leaf_bh(path));
5032out:
5033	return ret;
5034}
5035
5036/*
5037 * Split part or all of the extent record at split_index in the leaf
5038 * pointed to by path. Merge with the contiguous extent record if needed.
5039 *
5040 * Care is taken to handle contiguousness so as to not grow the tree.
5041 *
5042 * meta_ac is not strictly necessary - we only truly need it if growth
5043 * of the tree is required. All other cases will degrade into a less
5044 * optimal tree layout.
5045 *
5046 * last_eb_bh should be the rightmost leaf block for any extent
5047 * btree. Since a split may grow the tree or a merge might shrink it,
5048 * the caller cannot trust the contents of that buffer after this call.
5049 *
5050 * This code is optimized for readability - several passes might be
5051 * made over certain portions of the tree. All of those blocks will
5052 * have been brought into cache (and pinned via the journal), so the
5053 * extra overhead is not expressed in terms of disk reads.
5054 */
5055int ocfs2_split_extent(handle_t *handle,
5056		       struct ocfs2_extent_tree *et,
5057		       struct ocfs2_path *path,
5058		       int split_index,
5059		       struct ocfs2_extent_rec *split_rec,
5060		       struct ocfs2_alloc_context *meta_ac,
5061		       struct ocfs2_cached_dealloc_ctxt *dealloc)
5062{
5063	int ret = 0;
5064	struct ocfs2_extent_list *el = path_leaf_el(path);
5065	struct buffer_head *last_eb_bh = NULL;
5066	struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
5067	struct ocfs2_merge_ctxt ctxt;
5068
5069	if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) ||
5070	    ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) <
5071	     (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) {
5072		ret = -EIO;
5073		mlog_errno(ret);
5074		goto out;
5075	}
5076
5077	ret = ocfs2_figure_merge_contig_type(et, path, el,
5078					     split_index,
5079					     split_rec,
5080					     &ctxt);
5081	if (ret) {
5082		mlog_errno(ret);
5083		goto out;
5084	}
5085
5086	/*
5087	 * The core merge / split code wants to know how much room is
5088	 * left in this allocation tree, so we pass the
5089	 * rightmost extent list.
5090	 */
5091	if (path->p_tree_depth) {
5092		ret = ocfs2_read_extent_block(et->et_ci,
5093					      ocfs2_et_get_last_eb_blk(et),
5094					      &last_eb_bh);
5095		if (ret) {
5096			mlog_errno(ret);
5097			goto out;
5098		}
5099	}
5100
5101	if (rec->e_cpos == split_rec->e_cpos &&
5102	    rec->e_leaf_clusters == split_rec->e_leaf_clusters)
5103		ctxt.c_split_covers_rec = 1;
5104	else
5105		ctxt.c_split_covers_rec = 0;
5106
5107	ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]);
5108
5109	trace_ocfs2_split_extent(split_index, ctxt.c_contig_type,
5110				 ctxt.c_has_empty_extent,
5111				 ctxt.c_split_covers_rec);
5112
5113	if (ctxt.c_contig_type == CONTIG_NONE) {
5114		if (ctxt.c_split_covers_rec)
5115			ret = ocfs2_replace_extent_rec(handle, et, path, el,
5116						       split_index, split_rec);
5117		else
5118			ret = ocfs2_split_and_insert(handle, et, path,
5119						     &last_eb_bh, split_index,
5120						     split_rec, meta_ac);
5121		if (ret)
5122			mlog_errno(ret);
5123	} else {
5124		ret = ocfs2_try_to_merge_extent(handle, et, path,
5125						split_index, split_rec,
5126						dealloc, &ctxt);
5127		if (ret)
5128			mlog_errno(ret);
5129	}
5130
5131out:
5132	brelse(last_eb_bh);
5133	return ret;
5134}
5135
5136/*
5137 * Change the flags of the already-existing extent at cpos for len clusters.
5138 *
5139 * new_flags: the flags we want to set.
5140 * clear_flags: the flags we want to clear.
5141 * phys: the new physical offset we want this new extent starts from.
5142 *
5143 * If the existing extent is larger than the request, initiate a
5144 * split. An attempt will be made at merging with adjacent extents.
5145 *
5146 * The caller is responsible for passing down meta_ac if we'll need it.
5147 */
5148int ocfs2_change_extent_flag(handle_t *handle,
5149			     struct ocfs2_extent_tree *et,
5150			     u32 cpos, u32 len, u32 phys,
5151			     struct ocfs2_alloc_context *meta_ac,
5152			     struct ocfs2_cached_dealloc_ctxt *dealloc,
5153			     int new_flags, int clear_flags)
5154{
5155	int ret, index;
5156	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5157	u64 start_blkno = ocfs2_clusters_to_blocks(sb, phys);
5158	struct ocfs2_extent_rec split_rec;
5159	struct ocfs2_path *left_path = NULL;
5160	struct ocfs2_extent_list *el;
5161	struct ocfs2_extent_rec *rec;
5162
5163	left_path = ocfs2_new_path_from_et(et);
5164	if (!left_path) {
5165		ret = -ENOMEM;
5166		mlog_errno(ret);
5167		goto out;
5168	}
5169
5170	ret = ocfs2_find_path(et->et_ci, left_path, cpos);
5171	if (ret) {
5172		mlog_errno(ret);
5173		goto out;
5174	}
5175	el = path_leaf_el(left_path);
5176
5177	index = ocfs2_search_extent_list(el, cpos);
5178	if (index == -1) {
5179		ocfs2_error(sb,
5180			    "Owner %llu has an extent at cpos %u which can no longer be found\n",
5181			    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5182			    cpos);
5183		ret = -EROFS;
5184		goto out;
5185	}
5186
5187	ret = -EIO;
5188	rec = &el->l_recs[index];
5189	if (new_flags && (rec->e_flags & new_flags)) {
5190		mlog(ML_ERROR, "Owner %llu tried to set %d flags on an "
5191		     "extent that already had them\n",
5192		     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5193		     new_flags);
5194		goto out;
5195	}
5196
5197	if (clear_flags && !(rec->e_flags & clear_flags)) {
5198		mlog(ML_ERROR, "Owner %llu tried to clear %d flags on an "
5199		     "extent that didn't have them\n",
5200		     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5201		     clear_flags);
5202		goto out;
5203	}
5204
5205	memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec));
5206	split_rec.e_cpos = cpu_to_le32(cpos);
5207	split_rec.e_leaf_clusters = cpu_to_le16(len);
5208	split_rec.e_blkno = cpu_to_le64(start_blkno);
5209	split_rec.e_flags = rec->e_flags;
5210	if (new_flags)
5211		split_rec.e_flags |= new_flags;
5212	if (clear_flags)
5213		split_rec.e_flags &= ~clear_flags;
5214
5215	ret = ocfs2_split_extent(handle, et, left_path,
5216				 index, &split_rec, meta_ac,
5217				 dealloc);
5218	if (ret)
5219		mlog_errno(ret);
5220
5221out:
5222	ocfs2_free_path(left_path);
5223	return ret;
5224
5225}
5226
5227/*
5228 * Mark the already-existing extent at cpos as written for len clusters.
5229 * This removes the unwritten extent flag.
5230 *
5231 * If the existing extent is larger than the request, initiate a
5232 * split. An attempt will be made at merging with adjacent extents.
5233 *
5234 * The caller is responsible for passing down meta_ac if we'll need it.
5235 */
5236int ocfs2_mark_extent_written(struct inode *inode,
5237			      struct ocfs2_extent_tree *et,
5238			      handle_t *handle, u32 cpos, u32 len, u32 phys,
5239			      struct ocfs2_alloc_context *meta_ac,
5240			      struct ocfs2_cached_dealloc_ctxt *dealloc)
5241{
5242	int ret;
5243
5244	trace_ocfs2_mark_extent_written(
5245		(unsigned long long)OCFS2_I(inode)->ip_blkno,
5246		cpos, len, phys);
5247
5248	if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) {
5249		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",
5250			    (unsigned long long)OCFS2_I(inode)->ip_blkno);
5251		ret = -EROFS;
5252		goto out;
5253	}
5254
5255	/*
5256	 * XXX: This should be fixed up so that we just re-insert the
5257	 * next extent records.
5258	 */
5259	ocfs2_et_extent_map_truncate(et, 0);
5260
5261	ret = ocfs2_change_extent_flag(handle, et, cpos,
5262				       len, phys, meta_ac, dealloc,
5263				       0, OCFS2_EXT_UNWRITTEN);
5264	if (ret)
5265		mlog_errno(ret);
5266
5267out:
5268	return ret;
5269}
5270
5271static int ocfs2_split_tree(handle_t *handle, struct ocfs2_extent_tree *et,
5272			    struct ocfs2_path *path,
5273			    int index, u32 new_range,
5274			    struct ocfs2_alloc_context *meta_ac)
5275{
5276	int ret, depth, credits;
5277	struct buffer_head *last_eb_bh = NULL;
5278	struct ocfs2_extent_block *eb;
5279	struct ocfs2_extent_list *rightmost_el, *el;
5280	struct ocfs2_extent_rec split_rec;
5281	struct ocfs2_extent_rec *rec;
5282	struct ocfs2_insert_type insert;
5283
5284	/*
5285	 * Setup the record to split before we grow the tree.
5286	 */
5287	el = path_leaf_el(path);
5288	rec = &el->l_recs[index];
5289	ocfs2_make_right_split_rec(ocfs2_metadata_cache_get_super(et->et_ci),
5290				   &split_rec, new_range, rec);
5291
5292	depth = path->p_tree_depth;
5293	if (depth > 0) {
5294		ret = ocfs2_read_extent_block(et->et_ci,
5295					      ocfs2_et_get_last_eb_blk(et),
5296					      &last_eb_bh);
5297		if (ret < 0) {
5298			mlog_errno(ret);
5299			goto out;
5300		}
5301
5302		eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
5303		rightmost_el = &eb->h_list;
5304	} else
5305		rightmost_el = path_leaf_el(path);
5306
5307	credits = path->p_tree_depth +
5308		  ocfs2_extend_meta_needed(et->et_root_el);
5309	ret = ocfs2_extend_trans(handle, credits);
5310	if (ret) {
5311		mlog_errno(ret);
5312		goto out;
5313	}
5314
5315	if (le16_to_cpu(rightmost_el->l_next_free_rec) ==
5316	    le16_to_cpu(rightmost_el->l_count)) {
5317		ret = ocfs2_grow_tree(handle, et, &depth, &last_eb_bh,
5318				      meta_ac);
5319		if (ret) {
5320			mlog_errno(ret);
5321			goto out;
5322		}
5323	}
5324
5325	memset(&insert, 0, sizeof(struct ocfs2_insert_type));
5326	insert.ins_appending = APPEND_NONE;
5327	insert.ins_contig = CONTIG_NONE;
5328	insert.ins_split = SPLIT_RIGHT;
5329	insert.ins_tree_depth = depth;
5330
5331	ret = ocfs2_do_insert_extent(handle, et, &split_rec, &insert);
5332	if (ret)
5333		mlog_errno(ret);
5334
5335out:
5336	brelse(last_eb_bh);
5337	return ret;
5338}
5339
5340static int ocfs2_truncate_rec(handle_t *handle,
5341			      struct ocfs2_extent_tree *et,
5342			      struct ocfs2_path *path, int index,
5343			      struct ocfs2_cached_dealloc_ctxt *dealloc,
5344			      u32 cpos, u32 len)
5345{
5346	int ret;
5347	u32 left_cpos, rec_range, trunc_range;
5348	int is_rightmost_tree_rec = 0;
5349	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
5350	struct ocfs2_path *left_path = NULL;
5351	struct ocfs2_extent_list *el = path_leaf_el(path);
5352	struct ocfs2_extent_rec *rec;
5353	struct ocfs2_extent_block *eb;
5354
5355	if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) {
5356		/* extend credit for ocfs2_remove_rightmost_path */
5357		ret = ocfs2_extend_rotate_transaction(handle, 0,
5358				jbd2_handle_buffer_credits(handle),
5359				path);
5360		if (ret) {
5361			mlog_errno(ret);
5362			goto out;
5363		}
5364
5365		ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5366		if (ret) {
5367			mlog_errno(ret);
5368			goto out;
5369		}
5370
5371		index--;
5372	}
5373
5374	if (index == (le16_to_cpu(el->l_next_free_rec) - 1) &&
5375	    path->p_tree_depth) {
5376		/*
5377		 * Check whether this is the rightmost tree record. If
5378		 * we remove all of this record or part of its right
5379		 * edge then an update of the record lengths above it
5380		 * will be required.
5381		 */
5382		eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
5383		if (eb->h_next_leaf_blk == 0)
5384			is_rightmost_tree_rec = 1;
5385	}
5386
5387	rec = &el->l_recs[index];
5388	if (index == 0 && path->p_tree_depth &&
5389	    le32_to_cpu(rec->e_cpos) == cpos) {
5390		/*
5391		 * Changing the leftmost offset (via partial or whole
5392		 * record truncate) of an interior (or rightmost) path
5393		 * means we have to update the subtree that is formed
5394		 * by this leaf and the one to it's left.
5395		 *
5396		 * There are two cases we can skip:
5397		 *   1) Path is the leftmost one in our btree.
5398		 *   2) The leaf is rightmost and will be empty after
5399		 *      we remove the extent record - the rotate code
5400		 *      knows how to update the newly formed edge.
5401		 */
5402
5403		ret = ocfs2_find_cpos_for_left_leaf(sb, path, &left_cpos);
5404		if (ret) {
5405			mlog_errno(ret);
5406			goto out;
5407		}
5408
5409		if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) {
5410			left_path = ocfs2_new_path_from_path(path);
5411			if (!left_path) {
5412				ret = -ENOMEM;
5413				mlog_errno(ret);
5414				goto out;
5415			}
5416
5417			ret = ocfs2_find_path(et->et_ci, left_path,
5418					      left_cpos);
5419			if (ret) {
5420				mlog_errno(ret);
5421				goto out;
5422			}
5423		}
5424	}
5425
5426	ret = ocfs2_extend_rotate_transaction(handle, 0,
5427					jbd2_handle_buffer_credits(handle),
5428					path);
5429	if (ret) {
5430		mlog_errno(ret);
5431		goto out;
5432	}
5433
5434	ret = ocfs2_journal_access_path(et->et_ci, handle, path);
5435	if (ret) {
5436		mlog_errno(ret);
5437		goto out;
5438	}
5439
5440	ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
5441	if (ret) {
5442		mlog_errno(ret);
5443		goto out;
5444	}
5445
5446	rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5447	trunc_range = cpos + len;
5448
5449	if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) {
5450		int next_free;
5451
5452		memset(rec, 0, sizeof(*rec));
5453		ocfs2_cleanup_merge(el, index);
5454
5455		next_free = le16_to_cpu(el->l_next_free_rec);
5456		if (is_rightmost_tree_rec && next_free > 1) {
5457			/*
5458			 * We skip the edge update if this path will
5459			 * be deleted by the rotate code.
5460			 */
5461			rec = &el->l_recs[next_free - 1];
5462			ocfs2_adjust_rightmost_records(handle, et, path,
5463						       rec);
5464		}
5465	} else if (le32_to_cpu(rec->e_cpos) == cpos) {
5466		/* Remove leftmost portion of the record. */
5467		le32_add_cpu(&rec->e_cpos, len);
5468		le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len));
5469		le16_add_cpu(&rec->e_leaf_clusters, -len);
5470	} else if (rec_range == trunc_range) {
5471		/* Remove rightmost portion of the record */
5472		le16_add_cpu(&rec->e_leaf_clusters, -len);
5473		if (is_rightmost_tree_rec)
5474			ocfs2_adjust_rightmost_records(handle, et, path, rec);
5475	} else {
5476		/* Caller should have trapped this. */
5477		mlog(ML_ERROR, "Owner %llu: Invalid record truncate: (%u, %u) "
5478		     "(%u, %u)\n",
5479		     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5480		     le32_to_cpu(rec->e_cpos),
5481		     le16_to_cpu(rec->e_leaf_clusters), cpos, len);
5482		BUG();
5483	}
5484
5485	if (left_path) {
5486		int subtree_index;
5487
5488		subtree_index = ocfs2_find_subtree_root(et, left_path, path);
5489		ocfs2_complete_edge_insert(handle, left_path, path,
5490					   subtree_index);
5491	}
5492
5493	ocfs2_journal_dirty(handle, path_leaf_bh(path));
5494
5495	ret = ocfs2_rotate_tree_left(handle, et, path, dealloc);
5496	if (ret)
5497		mlog_errno(ret);
5498
5499out:
5500	ocfs2_free_path(left_path);
5501	return ret;
5502}
5503
5504int ocfs2_remove_extent(handle_t *handle,
5505			struct ocfs2_extent_tree *et,
5506			u32 cpos, u32 len,
5507			struct ocfs2_alloc_context *meta_ac,
5508			struct ocfs2_cached_dealloc_ctxt *dealloc)
5509{
5510	int ret, index;
5511	u32 rec_range, trunc_range;
5512	struct ocfs2_extent_rec *rec;
5513	struct ocfs2_extent_list *el;
5514	struct ocfs2_path *path = NULL;
5515
5516	/*
5517	 * XXX: Why are we truncating to 0 instead of wherever this
5518	 * affects us?
5519	 */
5520	ocfs2_et_extent_map_truncate(et, 0);
5521
5522	path = ocfs2_new_path_from_et(et);
5523	if (!path) {
5524		ret = -ENOMEM;
5525		mlog_errno(ret);
5526		goto out;
5527	}
5528
5529	ret = ocfs2_find_path(et->et_ci, path, cpos);
5530	if (ret) {
5531		mlog_errno(ret);
5532		goto out;
5533	}
5534
5535	el = path_leaf_el(path);
5536	index = ocfs2_search_extent_list(el, cpos);
5537	if (index == -1) {
5538		ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5539			    "Owner %llu has an extent at cpos %u which can no longer be found\n",
5540			    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5541			    cpos);
5542		ret = -EROFS;
5543		goto out;
5544	}
5545
5546	/*
5547	 * We have 3 cases of extent removal:
5548	 *   1) Range covers the entire extent rec
5549	 *   2) Range begins or ends on one edge of the extent rec
5550	 *   3) Range is in the middle of the extent rec (no shared edges)
5551	 *
5552	 * For case 1 we remove the extent rec and left rotate to
5553	 * fill the hole.
5554	 *
5555	 * For case 2 we just shrink the existing extent rec, with a
5556	 * tree update if the shrinking edge is also the edge of an
5557	 * extent block.
5558	 *
5559	 * For case 3 we do a right split to turn the extent rec into
5560	 * something case 2 can handle.
5561	 */
5562	rec = &el->l_recs[index];
5563	rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
5564	trunc_range = cpos + len;
5565
5566	BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range);
5567
5568	trace_ocfs2_remove_extent(
5569		(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5570		cpos, len, index, le32_to_cpu(rec->e_cpos),
5571		ocfs2_rec_clusters(el, rec));
5572
5573	if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) {
5574		ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5575					 cpos, len);
5576		if (ret) {
5577			mlog_errno(ret);
5578			goto out;
5579		}
5580	} else {
5581		ret = ocfs2_split_tree(handle, et, path, index,
5582				       trunc_range, meta_ac);
5583		if (ret) {
5584			mlog_errno(ret);
5585			goto out;
5586		}
5587
5588		/*
5589		 * The split could have manipulated the tree enough to
5590		 * move the record location, so we have to look for it again.
5591		 */
5592		ocfs2_reinit_path(path, 1);
5593
5594		ret = ocfs2_find_path(et->et_ci, path, cpos);
5595		if (ret) {
5596			mlog_errno(ret);
5597			goto out;
5598		}
5599
5600		el = path_leaf_el(path);
5601		index = ocfs2_search_extent_list(el, cpos);
5602		if (index == -1) {
5603			ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5604				    "Owner %llu: split at cpos %u lost record\n",
5605				    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5606				    cpos);
5607			ret = -EROFS;
5608			goto out;
5609		}
5610
5611		/*
5612		 * Double check our values here. If anything is fishy,
5613		 * it's easier to catch it at the top level.
5614		 */
5615		rec = &el->l_recs[index];
5616		rec_range = le32_to_cpu(rec->e_cpos) +
5617			ocfs2_rec_clusters(el, rec);
5618		if (rec_range != trunc_range) {
5619			ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
5620				    "Owner %llu: error after split at cpos %u trunc len %u, existing record is (%u,%u)\n",
5621				    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
5622				    cpos, len, le32_to_cpu(rec->e_cpos),
5623				    ocfs2_rec_clusters(el, rec));
5624			ret = -EROFS;
5625			goto out;
5626		}
5627
5628		ret = ocfs2_truncate_rec(handle, et, path, index, dealloc,
5629					 cpos, len);
5630		if (ret)
5631			mlog_errno(ret);
5632	}
5633
5634out:
5635	ocfs2_free_path(path);
5636	return ret;
5637}
5638
5639/*
5640 * ocfs2_reserve_blocks_for_rec_trunc() would look basically the
5641 * same as ocfs2_lock_alloctors(), except for it accepts a blocks
5642 * number to reserve some extra blocks, and it only handles meta
5643 * data allocations.
5644 *
5645 * Currently, only ocfs2_remove_btree_range() uses it for truncating
5646 * and punching holes.
5647 */
5648static int ocfs2_reserve_blocks_for_rec_trunc(struct inode *inode,
5649					      struct ocfs2_extent_tree *et,
5650					      u32 extents_to_split,
5651					      struct ocfs2_alloc_context **ac,
5652					      int extra_blocks)
5653{
5654	int ret = 0, num_free_extents;
5655	unsigned int max_recs_needed = 2 * extents_to_split;
5656	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5657
5658	*ac = NULL;
5659
5660	num_free_extents = ocfs2_num_free_extents(et);
5661	if (num_free_extents < 0) {
5662		ret = num_free_extents;
5663		mlog_errno(ret);
5664		goto out;
5665	}
5666
5667	if (!num_free_extents ||
5668	    (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
5669		extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
5670
5671	if (extra_blocks) {
5672		ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, ac);
5673		if (ret < 0) {
5674			if (ret != -ENOSPC)
5675				mlog_errno(ret);
5676		}
5677	}
5678
5679out:
5680	if (ret) {
5681		if (*ac) {
5682			ocfs2_free_alloc_context(*ac);
5683			*ac = NULL;
5684		}
5685	}
5686
5687	return ret;
5688}
5689
5690int ocfs2_remove_btree_range(struct inode *inode,
5691			     struct ocfs2_extent_tree *et,
5692			     u32 cpos, u32 phys_cpos, u32 len, int flags,
5693			     struct ocfs2_cached_dealloc_ctxt *dealloc,
5694			     u64 refcount_loc, bool refcount_tree_locked)
5695{
5696	int ret, credits = 0, extra_blocks = 0;
5697	u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
5698	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
5699	struct inode *tl_inode = osb->osb_tl_inode;
5700	handle_t *handle;
5701	struct ocfs2_alloc_context *meta_ac = NULL;
5702	struct ocfs2_refcount_tree *ref_tree = NULL;
5703
5704	if ((flags & OCFS2_EXT_REFCOUNTED) && len) {
5705		BUG_ON(!ocfs2_is_refcount_inode(inode));
5706
5707		if (!refcount_tree_locked) {
5708			ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
5709						       &ref_tree, NULL);
5710			if (ret) {
5711				mlog_errno(ret);
5712				goto bail;
5713			}
5714		}
5715
5716		ret = ocfs2_prepare_refcount_change_for_del(inode,
5717							    refcount_loc,
5718							    phys_blkno,
5719							    len,
5720							    &credits,
5721							    &extra_blocks);
5722		if (ret < 0) {
5723			mlog_errno(ret);
5724			goto bail;
5725		}
5726	}
5727
5728	ret = ocfs2_reserve_blocks_for_rec_trunc(inode, et, 1, &meta_ac,
5729						 extra_blocks);
5730	if (ret) {
5731		mlog_errno(ret);
5732		goto bail;
5733	}
5734
5735	inode_lock(tl_inode);
5736
5737	if (ocfs2_truncate_log_needs_flush(osb)) {
5738		ret = __ocfs2_flush_truncate_log(osb);
5739		if (ret < 0) {
5740			mlog_errno(ret);
5741			goto out;
5742		}
5743	}
5744
5745	handle = ocfs2_start_trans(osb,
5746			ocfs2_remove_extent_credits(osb->sb) + credits);
5747	if (IS_ERR(handle)) {
5748		ret = PTR_ERR(handle);
5749		mlog_errno(ret);
5750		goto out;
5751	}
5752
5753	ret = ocfs2_et_root_journal_access(handle, et,
5754					   OCFS2_JOURNAL_ACCESS_WRITE);
5755	if (ret) {
5756		mlog_errno(ret);
5757		goto out_commit;
5758	}
5759
5760	dquot_free_space_nodirty(inode,
5761				  ocfs2_clusters_to_bytes(inode->i_sb, len));
5762
5763	ret = ocfs2_remove_extent(handle, et, cpos, len, meta_ac, dealloc);
5764	if (ret) {
5765		mlog_errno(ret);
5766		goto out_commit;
5767	}
5768
5769	ocfs2_et_update_clusters(et, -len);
5770	ocfs2_update_inode_fsync_trans(handle, inode, 1);
5771
5772	ocfs2_journal_dirty(handle, et->et_root_bh);
5773
5774	if (phys_blkno) {
5775		if (flags & OCFS2_EXT_REFCOUNTED)
5776			ret = ocfs2_decrease_refcount(inode, handle,
5777					ocfs2_blocks_to_clusters(osb->sb,
5778								 phys_blkno),
5779					len, meta_ac,
5780					dealloc, 1);
5781		else
5782			ret = ocfs2_truncate_log_append(osb, handle,
5783							phys_blkno, len);
5784		if (ret)
5785			mlog_errno(ret);
5786
5787	}
5788
5789out_commit:
5790	ocfs2_commit_trans(osb, handle);
5791out:
5792	inode_unlock(tl_inode);
5793bail:
5794	if (meta_ac)
5795		ocfs2_free_alloc_context(meta_ac);
5796
5797	if (ref_tree)
5798		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
5799
5800	return ret;
5801}
5802
5803int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb)
5804{
5805	struct buffer_head *tl_bh = osb->osb_tl_bh;
5806	struct ocfs2_dinode *di;
5807	struct ocfs2_truncate_log *tl;
5808
5809	di = (struct ocfs2_dinode *) tl_bh->b_data;
5810	tl = &di->id2.i_dealloc;
5811
5812	mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count),
5813			"slot %d, invalid truncate log parameters: used = "
5814			"%u, count = %u\n", osb->slot_num,
5815			le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count));
5816	return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count);
5817}
5818
5819static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl,
5820					   unsigned int new_start)
5821{
5822	unsigned int tail_index;
5823	unsigned int current_tail;
5824
5825	/* No records, nothing to coalesce */
5826	if (!le16_to_cpu(tl->tl_used))
5827		return 0;
5828
5829	tail_index = le16_to_cpu(tl->tl_used) - 1;
5830	current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start);
5831	current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters);
5832
5833	return current_tail == new_start;
5834}
5835
5836int ocfs2_truncate_log_append(struct ocfs2_super *osb,
5837			      handle_t *handle,
5838			      u64 start_blk,
5839			      unsigned int num_clusters)
5840{
5841	int status, index;
5842	unsigned int start_cluster, tl_count;
5843	struct inode *tl_inode = osb->osb_tl_inode;
5844	struct buffer_head *tl_bh = osb->osb_tl_bh;
5845	struct ocfs2_dinode *di;
5846	struct ocfs2_truncate_log *tl;
5847
5848	BUG_ON(inode_trylock(tl_inode));
5849
5850	start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk);
5851
5852	di = (struct ocfs2_dinode *) tl_bh->b_data;
5853
5854	/* tl_bh is loaded from ocfs2_truncate_log_init().  It's validated
5855	 * by the underlying call to ocfs2_read_inode_block(), so any
5856	 * corruption is a code bug */
5857	BUG_ON(!OCFS2_IS_VALID_DINODE(di));
5858
5859	tl = &di->id2.i_dealloc;
5860	tl_count = le16_to_cpu(tl->tl_count);
5861	mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) ||
5862			tl_count == 0,
5863			"Truncate record count on #%llu invalid "
5864			"wanted %u, actual %u\n",
5865			(unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5866			ocfs2_truncate_recs_per_inode(osb->sb),
5867			le16_to_cpu(tl->tl_count));
5868
5869	/* Caller should have known to flush before calling us. */
5870	index = le16_to_cpu(tl->tl_used);
5871	if (index >= tl_count) {
5872		status = -ENOSPC;
5873		mlog_errno(status);
5874		goto bail;
5875	}
5876
5877	status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5878					 OCFS2_JOURNAL_ACCESS_WRITE);
5879	if (status < 0) {
5880		mlog_errno(status);
5881		goto bail;
5882	}
5883
5884	trace_ocfs2_truncate_log_append(
5885		(unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index,
5886		start_cluster, num_clusters);
5887	if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) {
5888		/*
5889		 * Move index back to the record we are coalescing with.
5890		 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
5891		 */
5892		index--;
5893
5894		num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters);
5895		trace_ocfs2_truncate_log_append(
5896			(unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5897			index, le32_to_cpu(tl->tl_recs[index].t_start),
5898			num_clusters);
5899	} else {
5900		tl->tl_recs[index].t_start = cpu_to_le32(start_cluster);
5901		tl->tl_used = cpu_to_le16(index + 1);
5902	}
5903	tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters);
5904
5905	ocfs2_journal_dirty(handle, tl_bh);
5906
5907	osb->truncated_clusters += num_clusters;
5908bail:
5909	return status;
5910}
5911
5912static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
5913					 struct inode *data_alloc_inode,
5914					 struct buffer_head *data_alloc_bh)
5915{
5916	int status = 0;
5917	int i;
5918	unsigned int num_clusters;
5919	u64 start_blk;
5920	struct ocfs2_truncate_rec rec;
5921	struct ocfs2_dinode *di;
5922	struct ocfs2_truncate_log *tl;
5923	struct inode *tl_inode = osb->osb_tl_inode;
5924	struct buffer_head *tl_bh = osb->osb_tl_bh;
5925	handle_t *handle;
5926
5927	di = (struct ocfs2_dinode *) tl_bh->b_data;
5928	tl = &di->id2.i_dealloc;
5929	i = le16_to_cpu(tl->tl_used) - 1;
5930	while (i >= 0) {
5931		handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
5932		if (IS_ERR(handle)) {
5933			status = PTR_ERR(handle);
5934			mlog_errno(status);
5935			goto bail;
5936		}
5937
5938		/* Caller has given us at least enough credits to
5939		 * update the truncate log dinode */
5940		status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
5941						 OCFS2_JOURNAL_ACCESS_WRITE);
5942		if (status < 0) {
5943			mlog_errno(status);
5944			goto bail;
5945		}
5946
5947		tl->tl_used = cpu_to_le16(i);
5948
5949		ocfs2_journal_dirty(handle, tl_bh);
5950
5951		rec = tl->tl_recs[i];
5952		start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb,
5953						    le32_to_cpu(rec.t_start));
5954		num_clusters = le32_to_cpu(rec.t_clusters);
5955
5956		/* if start_blk is not set, we ignore the record as
5957		 * invalid. */
5958		if (start_blk) {
5959			trace_ocfs2_replay_truncate_records(
5960				(unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
5961				i, le32_to_cpu(rec.t_start), num_clusters);
5962
5963			status = ocfs2_free_clusters(handle, data_alloc_inode,
5964						     data_alloc_bh, start_blk,
5965						     num_clusters);
5966			if (status < 0) {
5967				mlog_errno(status);
5968				goto bail;
5969			}
5970		}
5971
5972		ocfs2_commit_trans(osb, handle);
5973		i--;
5974	}
5975
5976	osb->truncated_clusters = 0;
5977
5978bail:
5979	return status;
5980}
5981
5982/* Expects you to already be holding tl_inode->i_mutex */
5983int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
5984{
5985	int status;
5986	unsigned int num_to_flush;
5987	struct inode *tl_inode = osb->osb_tl_inode;
5988	struct inode *data_alloc_inode = NULL;
5989	struct buffer_head *tl_bh = osb->osb_tl_bh;
5990	struct buffer_head *data_alloc_bh = NULL;
5991	struct ocfs2_dinode *di;
5992	struct ocfs2_truncate_log *tl;
5993	struct ocfs2_journal *journal = osb->journal;
5994
5995	BUG_ON(inode_trylock(tl_inode));
5996
5997	di = (struct ocfs2_dinode *) tl_bh->b_data;
5998
5999	/* tl_bh is loaded from ocfs2_truncate_log_init().  It's validated
6000	 * by the underlying call to ocfs2_read_inode_block(), so any
6001	 * corruption is a code bug */
6002	BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6003
6004	tl = &di->id2.i_dealloc;
6005	num_to_flush = le16_to_cpu(tl->tl_used);
6006	trace_ocfs2_flush_truncate_log(
6007		(unsigned long long)OCFS2_I(tl_inode)->ip_blkno,
6008		num_to_flush);
6009	if (!num_to_flush) {
6010		status = 0;
6011		goto out;
6012	}
6013
6014	/* Appending truncate log(TA) and flushing truncate log(TF) are
6015	 * two separated transactions. They can be both committed but not
6016	 * checkpointed. If crash occurs then, both two transaction will be
6017	 * replayed with several already released to global bitmap clusters.
6018	 * Then truncate log will be replayed resulting in cluster double free.
6019	 */
6020	jbd2_journal_lock_updates(journal->j_journal);
6021	status = jbd2_journal_flush(journal->j_journal, 0);
6022	jbd2_journal_unlock_updates(journal->j_journal);
6023	if (status < 0) {
6024		mlog_errno(status);
6025		goto out;
6026	}
6027
6028	data_alloc_inode = ocfs2_get_system_file_inode(osb,
6029						       GLOBAL_BITMAP_SYSTEM_INODE,
6030						       OCFS2_INVALID_SLOT);
6031	if (!data_alloc_inode) {
6032		status = -EINVAL;
6033		mlog(ML_ERROR, "Could not get bitmap inode!\n");
6034		goto out;
6035	}
6036
6037	inode_lock(data_alloc_inode);
6038
6039	status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1);
6040	if (status < 0) {
6041		mlog_errno(status);
6042		goto out_mutex;
6043	}
6044
6045	status = ocfs2_replay_truncate_records(osb, data_alloc_inode,
6046					       data_alloc_bh);
6047	if (status < 0)
6048		mlog_errno(status);
6049
6050	brelse(data_alloc_bh);
6051	ocfs2_inode_unlock(data_alloc_inode, 1);
6052
6053out_mutex:
6054	inode_unlock(data_alloc_inode);
6055	iput(data_alloc_inode);
6056
6057out:
6058	return status;
6059}
6060
6061int ocfs2_flush_truncate_log(struct ocfs2_super *osb)
6062{
6063	int status;
6064	struct inode *tl_inode = osb->osb_tl_inode;
6065
6066	inode_lock(tl_inode);
6067	status = __ocfs2_flush_truncate_log(osb);
6068	inode_unlock(tl_inode);
6069
6070	return status;
6071}
6072
6073static void ocfs2_truncate_log_worker(struct work_struct *work)
6074{
6075	int status;
6076	struct ocfs2_super *osb =
6077		container_of(work, struct ocfs2_super,
6078			     osb_truncate_log_wq.work);
6079
6080	status = ocfs2_flush_truncate_log(osb);
6081	if (status < 0)
6082		mlog_errno(status);
6083	else
6084		ocfs2_init_steal_slots(osb);
6085}
6086
6087#define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
6088void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb,
6089				       int cancel)
6090{
6091	if (osb->osb_tl_inode &&
6092			atomic_read(&osb->osb_tl_disable) == 0) {
6093		/* We want to push off log flushes while truncates are
6094		 * still running. */
6095		if (cancel)
6096			cancel_delayed_work(&osb->osb_truncate_log_wq);
6097
6098		queue_delayed_work(osb->ocfs2_wq, &osb->osb_truncate_log_wq,
6099				   OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
6100	}
6101}
6102
6103/*
6104 * Try to flush truncate logs if we can free enough clusters from it.
6105 * As for return value, "< 0" means error, "0" no space and "1" means
6106 * we have freed enough spaces and let the caller try to allocate again.
6107 */
6108int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
6109					unsigned int needed)
6110{
6111	tid_t target;
6112	int ret = 0;
6113	unsigned int truncated_clusters;
6114
6115	inode_lock(osb->osb_tl_inode);
6116	truncated_clusters = osb->truncated_clusters;
6117	inode_unlock(osb->osb_tl_inode);
6118
6119	/*
6120	 * Check whether we can succeed in allocating if we free
6121	 * the truncate log.
6122	 */
6123	if (truncated_clusters < needed)
6124		goto out;
6125
6126	ret = ocfs2_flush_truncate_log(osb);
6127	if (ret) {
6128		mlog_errno(ret);
6129		goto out;
6130	}
6131
6132	if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
6133		jbd2_log_wait_commit(osb->journal->j_journal, target);
6134		ret = 1;
6135	}
6136out:
6137	return ret;
6138}
6139
6140static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb,
6141				       int slot_num,
6142				       struct inode **tl_inode,
6143				       struct buffer_head **tl_bh)
6144{
6145	int status;
6146	struct inode *inode = NULL;
6147	struct buffer_head *bh = NULL;
6148
6149	inode = ocfs2_get_system_file_inode(osb,
6150					   TRUNCATE_LOG_SYSTEM_INODE,
6151					   slot_num);
6152	if (!inode) {
6153		status = -EINVAL;
6154		mlog(ML_ERROR, "Could not get load truncate log inode!\n");
6155		goto bail;
6156	}
6157
6158	status = ocfs2_read_inode_block(inode, &bh);
6159	if (status < 0) {
6160		iput(inode);
6161		mlog_errno(status);
6162		goto bail;
6163	}
6164
6165	*tl_inode = inode;
6166	*tl_bh    = bh;
6167bail:
6168	return status;
6169}
6170
6171/* called during the 1st stage of node recovery. we stamp a clean
6172 * truncate log and pass back a copy for processing later. if the
6173 * truncate log does not require processing, a *tl_copy is set to
6174 * NULL. */
6175int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb,
6176				      int slot_num,
6177				      struct ocfs2_dinode **tl_copy)
6178{
6179	int status;
6180	struct inode *tl_inode = NULL;
6181	struct buffer_head *tl_bh = NULL;
6182	struct ocfs2_dinode *di;
6183	struct ocfs2_truncate_log *tl;
6184
6185	*tl_copy = NULL;
6186
6187	trace_ocfs2_begin_truncate_log_recovery(slot_num);
6188
6189	status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh);
6190	if (status < 0) {
6191		mlog_errno(status);
6192		goto bail;
6193	}
6194
6195	di = (struct ocfs2_dinode *) tl_bh->b_data;
6196
6197	/* tl_bh is loaded from ocfs2_get_truncate_log_info().  It's
6198	 * validated by the underlying call to ocfs2_read_inode_block(),
6199	 * so any corruption is a code bug */
6200	BUG_ON(!OCFS2_IS_VALID_DINODE(di));
6201
6202	tl = &di->id2.i_dealloc;
6203	if (le16_to_cpu(tl->tl_used)) {
6204		trace_ocfs2_truncate_log_recovery_num(le16_to_cpu(tl->tl_used));
6205
6206		/*
6207		 * Assuming the write-out below goes well, this copy will be
6208		 * passed back to recovery for processing.
6209		 */
6210		*tl_copy = kmemdup(tl_bh->b_data, tl_bh->b_size, GFP_KERNEL);
6211		if (!(*tl_copy)) {
6212			status = -ENOMEM;
6213			mlog_errno(status);
6214			goto bail;
6215		}
6216
6217		/* All we need to do to clear the truncate log is set
6218		 * tl_used. */
6219		tl->tl_used = 0;
6220
6221		ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check);
6222		status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode));
6223		if (status < 0) {
6224			mlog_errno(status);
6225			goto bail;
6226		}
6227	}
6228
6229bail:
6230	iput(tl_inode);
6231	brelse(tl_bh);
6232
6233	if (status < 0) {
6234		kfree(*tl_copy);
6235		*tl_copy = NULL;
6236		mlog_errno(status);
6237	}
6238
6239	return status;
6240}
6241
6242int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb,
6243					 struct ocfs2_dinode *tl_copy)
6244{
6245	int status = 0;
6246	int i;
6247	unsigned int clusters, num_recs, start_cluster;
6248	u64 start_blk;
6249	handle_t *handle;
6250	struct inode *tl_inode = osb->osb_tl_inode;
6251	struct ocfs2_truncate_log *tl;
6252
6253	if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) {
6254		mlog(ML_ERROR, "Asked to recover my own truncate log!\n");
6255		return -EINVAL;
6256	}
6257
6258	tl = &tl_copy->id2.i_dealloc;
6259	num_recs = le16_to_cpu(tl->tl_used);
6260	trace_ocfs2_complete_truncate_log_recovery(
6261		(unsigned long long)le64_to_cpu(tl_copy->i_blkno),
6262		num_recs);
6263
6264	inode_lock(tl_inode);
6265	for(i = 0; i < num_recs; i++) {
6266		if (ocfs2_truncate_log_needs_flush(osb)) {
6267			status = __ocfs2_flush_truncate_log(osb);
6268			if (status < 0) {
6269				mlog_errno(status);
6270				goto bail_up;
6271			}
6272		}
6273
6274		handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6275		if (IS_ERR(handle)) {
6276			status = PTR_ERR(handle);
6277			mlog_errno(status);
6278			goto bail_up;
6279		}
6280
6281		clusters = le32_to_cpu(tl->tl_recs[i].t_clusters);
6282		start_cluster = le32_to_cpu(tl->tl_recs[i].t_start);
6283		start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster);
6284
6285		status = ocfs2_truncate_log_append(osb, handle,
6286						   start_blk, clusters);
6287		ocfs2_commit_trans(osb, handle);
6288		if (status < 0) {
6289			mlog_errno(status);
6290			goto bail_up;
6291		}
6292	}
6293
6294bail_up:
6295	inode_unlock(tl_inode);
6296
6297	return status;
6298}
6299
6300void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb)
6301{
6302	int status;
6303	struct inode *tl_inode = osb->osb_tl_inode;
6304
6305	atomic_set(&osb->osb_tl_disable, 1);
6306
6307	if (tl_inode) {
6308		cancel_delayed_work(&osb->osb_truncate_log_wq);
6309		flush_workqueue(osb->ocfs2_wq);
6310
6311		status = ocfs2_flush_truncate_log(osb);
6312		if (status < 0)
6313			mlog_errno(status);
6314
6315		brelse(osb->osb_tl_bh);
6316		iput(osb->osb_tl_inode);
6317	}
6318}
6319
6320int ocfs2_truncate_log_init(struct ocfs2_super *osb)
6321{
6322	int status;
6323	struct inode *tl_inode = NULL;
6324	struct buffer_head *tl_bh = NULL;
6325
6326	status = ocfs2_get_truncate_log_info(osb,
6327					     osb->slot_num,
6328					     &tl_inode,
6329					     &tl_bh);
6330	if (status < 0)
6331		mlog_errno(status);
6332
6333	/* ocfs2_truncate_log_shutdown keys on the existence of
6334	 * osb->osb_tl_inode so we don't set any of the osb variables
6335	 * until we're sure all is well. */
6336	INIT_DELAYED_WORK(&osb->osb_truncate_log_wq,
6337			  ocfs2_truncate_log_worker);
6338	atomic_set(&osb->osb_tl_disable, 0);
6339	osb->osb_tl_bh    = tl_bh;
6340	osb->osb_tl_inode = tl_inode;
6341
6342	return status;
6343}
6344
6345/*
6346 * Delayed de-allocation of suballocator blocks.
6347 *
6348 * Some sets of block de-allocations might involve multiple suballocator inodes.
6349 *
6350 * The locking for this can get extremely complicated, especially when
6351 * the suballocator inodes to delete from aren't known until deep
6352 * within an unrelated codepath.
6353 *
6354 * ocfs2_extent_block structures are a good example of this - an inode
6355 * btree could have been grown by any number of nodes each allocating
6356 * out of their own suballoc inode.
6357 *
6358 * These structures allow the delay of block de-allocation until a
6359 * later time, when locking of multiple cluster inodes won't cause
6360 * deadlock.
6361 */
6362
6363/*
6364 * Describe a single bit freed from a suballocator.  For the block
6365 * suballocators, it represents one block.  For the global cluster
6366 * allocator, it represents some clusters and free_bit indicates
6367 * clusters number.
6368 */
6369struct ocfs2_cached_block_free {
6370	struct ocfs2_cached_block_free		*free_next;
6371	u64					free_bg;
6372	u64					free_blk;
6373	unsigned int				free_bit;
6374};
6375
6376struct ocfs2_per_slot_free_list {
6377	struct ocfs2_per_slot_free_list		*f_next_suballocator;
6378	int					f_inode_type;
6379	int					f_slot;
6380	struct ocfs2_cached_block_free		*f_first;
6381};
6382
6383static int ocfs2_free_cached_blocks(struct ocfs2_super *osb,
6384				    int sysfile_type,
6385				    int slot,
6386				    struct ocfs2_cached_block_free *head)
6387{
6388	int ret;
6389	u64 bg_blkno;
6390	handle_t *handle;
6391	struct inode *inode;
6392	struct buffer_head *di_bh = NULL;
6393	struct ocfs2_cached_block_free *tmp;
6394
6395	inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot);
6396	if (!inode) {
6397		ret = -EINVAL;
6398		mlog_errno(ret);
6399		goto out;
6400	}
6401
6402	inode_lock(inode);
6403
6404	ret = ocfs2_inode_lock(inode, &di_bh, 1);
6405	if (ret) {
6406		mlog_errno(ret);
6407		goto out_mutex;
6408	}
6409
6410	while (head) {
6411		if (head->free_bg)
6412			bg_blkno = head->free_bg;
6413		else
6414			bg_blkno = ocfs2_which_suballoc_group(head->free_blk,
6415							      head->free_bit);
6416		handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE);
6417		if (IS_ERR(handle)) {
6418			ret = PTR_ERR(handle);
6419			mlog_errno(ret);
6420			goto out_unlock;
6421		}
6422
6423		trace_ocfs2_free_cached_blocks(
6424		     (unsigned long long)head->free_blk, head->free_bit);
6425
6426		ret = ocfs2_free_suballoc_bits(handle, inode, di_bh,
6427					       head->free_bit, bg_blkno, 1);
6428		if (ret)
6429			mlog_errno(ret);
6430
6431		ocfs2_commit_trans(osb, handle);
6432
6433		tmp = head;
6434		head = head->free_next;
6435		kfree(tmp);
6436	}
6437
6438out_unlock:
6439	ocfs2_inode_unlock(inode, 1);
6440	brelse(di_bh);
6441out_mutex:
6442	inode_unlock(inode);
6443	iput(inode);
6444out:
6445	while(head) {
6446		/* Premature exit may have left some dangling items. */
6447		tmp = head;
6448		head = head->free_next;
6449		kfree(tmp);
6450	}
6451
6452	return ret;
6453}
6454
6455int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6456				u64 blkno, unsigned int bit)
6457{
6458	int ret = 0;
6459	struct ocfs2_cached_block_free *item;
6460
6461	item = kzalloc(sizeof(*item), GFP_NOFS);
6462	if (item == NULL) {
6463		ret = -ENOMEM;
6464		mlog_errno(ret);
6465		return ret;
6466	}
6467
6468	trace_ocfs2_cache_cluster_dealloc((unsigned long long)blkno, bit);
6469
6470	item->free_blk = blkno;
6471	item->free_bit = bit;
6472	item->free_next = ctxt->c_global_allocator;
6473
6474	ctxt->c_global_allocator = item;
6475	return ret;
6476}
6477
6478static int ocfs2_free_cached_clusters(struct ocfs2_super *osb,
6479				      struct ocfs2_cached_block_free *head)
6480{
6481	struct ocfs2_cached_block_free *tmp;
6482	struct inode *tl_inode = osb->osb_tl_inode;
6483	handle_t *handle;
6484	int ret = 0;
6485
6486	inode_lock(tl_inode);
6487
6488	while (head) {
6489		if (ocfs2_truncate_log_needs_flush(osb)) {
6490			ret = __ocfs2_flush_truncate_log(osb);
6491			if (ret < 0) {
6492				mlog_errno(ret);
6493				break;
6494			}
6495		}
6496
6497		handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE);
6498		if (IS_ERR(handle)) {
6499			ret = PTR_ERR(handle);
6500			mlog_errno(ret);
6501			break;
6502		}
6503
6504		ret = ocfs2_truncate_log_append(osb, handle, head->free_blk,
6505						head->free_bit);
6506
6507		ocfs2_commit_trans(osb, handle);
6508		tmp = head;
6509		head = head->free_next;
6510		kfree(tmp);
6511
6512		if (ret < 0) {
6513			mlog_errno(ret);
6514			break;
6515		}
6516	}
6517
6518	inode_unlock(tl_inode);
6519
6520	while (head) {
6521		/* Premature exit may have left some dangling items. */
6522		tmp = head;
6523		head = head->free_next;
6524		kfree(tmp);
6525	}
6526
6527	return ret;
6528}
6529
6530int ocfs2_run_deallocs(struct ocfs2_super *osb,
6531		       struct ocfs2_cached_dealloc_ctxt *ctxt)
6532{
6533	int ret = 0, ret2;
6534	struct ocfs2_per_slot_free_list *fl;
6535
6536	if (!ctxt)
6537		return 0;
6538
6539	while (ctxt->c_first_suballocator) {
6540		fl = ctxt->c_first_suballocator;
6541
6542		if (fl->f_first) {
6543			trace_ocfs2_run_deallocs(fl->f_inode_type,
6544						 fl->f_slot);
6545			ret2 = ocfs2_free_cached_blocks(osb,
6546							fl->f_inode_type,
6547							fl->f_slot,
6548							fl->f_first);
6549			if (ret2)
6550				mlog_errno(ret2);
6551			if (!ret)
6552				ret = ret2;
6553		}
6554
6555		ctxt->c_first_suballocator = fl->f_next_suballocator;
6556		kfree(fl);
6557	}
6558
6559	if (ctxt->c_global_allocator) {
6560		ret2 = ocfs2_free_cached_clusters(osb,
6561						  ctxt->c_global_allocator);
6562		if (ret2)
6563			mlog_errno(ret2);
6564		if (!ret)
6565			ret = ret2;
6566
6567		ctxt->c_global_allocator = NULL;
6568	}
6569
6570	return ret;
6571}
6572
6573static struct ocfs2_per_slot_free_list *
6574ocfs2_find_per_slot_free_list(int type,
6575			      int slot,
6576			      struct ocfs2_cached_dealloc_ctxt *ctxt)
6577{
6578	struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6579
6580	while (fl) {
6581		if (fl->f_inode_type == type && fl->f_slot == slot)
6582			return fl;
6583
6584		fl = fl->f_next_suballocator;
6585	}
6586
6587	fl = kmalloc(sizeof(*fl), GFP_NOFS);
6588	if (fl) {
6589		fl->f_inode_type = type;
6590		fl->f_slot = slot;
6591		fl->f_first = NULL;
6592		fl->f_next_suballocator = ctxt->c_first_suballocator;
6593
6594		ctxt->c_first_suballocator = fl;
6595	}
6596	return fl;
6597}
6598
6599static struct ocfs2_per_slot_free_list *
6600ocfs2_find_preferred_free_list(int type,
6601			       int preferred_slot,
6602			       int *real_slot,
6603			       struct ocfs2_cached_dealloc_ctxt *ctxt)
6604{
6605	struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator;
6606
6607	while (fl) {
6608		if (fl->f_inode_type == type && fl->f_slot == preferred_slot) {
6609			*real_slot = fl->f_slot;
6610			return fl;
6611		}
6612
6613		fl = fl->f_next_suballocator;
6614	}
6615
6616	/* If we can't find any free list matching preferred slot, just use
6617	 * the first one.
6618	 */
6619	fl = ctxt->c_first_suballocator;
6620	*real_slot = fl->f_slot;
6621
6622	return fl;
6623}
6624
6625/* Return Value 1 indicates empty */
6626static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et)
6627{
6628	struct ocfs2_per_slot_free_list *fl = NULL;
6629
6630	if (!et->et_dealloc)
6631		return 1;
6632
6633	fl = et->et_dealloc->c_first_suballocator;
6634	if (!fl)
6635		return 1;
6636
6637	if (!fl->f_first)
6638		return 1;
6639
6640	return 0;
6641}
6642
6643/* If extent was deleted from tree due to extent rotation and merging, and
6644 * no metadata is reserved ahead of time. Try to reuse some extents
6645 * just deleted. This is only used to reuse extent blocks.
6646 * It is supposed to find enough extent blocks in dealloc if our estimation
6647 * on metadata is accurate.
6648 */
6649static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
6650					struct ocfs2_extent_tree *et,
6651					struct buffer_head **new_eb_bh,
6652					int blk_wanted, int *blk_given)
6653{
6654	int i, status = 0, real_slot;
6655	struct ocfs2_cached_dealloc_ctxt *dealloc;
6656	struct ocfs2_per_slot_free_list *fl;
6657	struct ocfs2_cached_block_free *bf;
6658	struct ocfs2_extent_block *eb;
6659	struct ocfs2_super *osb =
6660		OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
6661
6662	*blk_given = 0;
6663
6664	/* If extent tree doesn't have a dealloc, this is not faulty. Just
6665	 * tell upper caller dealloc can't provide any block and it should
6666	 * ask for alloc to claim more space.
6667	 */
6668	dealloc = et->et_dealloc;
6669	if (!dealloc)
6670		goto bail;
6671
6672	for (i = 0; i < blk_wanted; i++) {
6673		/* Prefer to use local slot */
6674		fl = ocfs2_find_preferred_free_list(EXTENT_ALLOC_SYSTEM_INODE,
6675						    osb->slot_num, &real_slot,
6676						    dealloc);
6677		/* If no more block can be reused, we should claim more
6678		 * from alloc. Just return here normally.
6679		 */
6680		if (!fl) {
6681			status = 0;
6682			break;
6683		}
6684
6685		bf = fl->f_first;
6686		fl->f_first = bf->free_next;
6687
6688		new_eb_bh[i] = sb_getblk(osb->sb, bf->free_blk);
6689		if (new_eb_bh[i] == NULL) {
6690			status = -ENOMEM;
6691			mlog_errno(status);
6692			goto bail;
6693		}
6694
6695		mlog(0, "Reusing block(%llu) from "
6696		     "dealloc(local slot:%d, real slot:%d)\n",
6697		     bf->free_blk, osb->slot_num, real_slot);
6698
6699		ocfs2_set_new_buffer_uptodate(et->et_ci, new_eb_bh[i]);
6700
6701		status = ocfs2_journal_access_eb(handle, et->et_ci,
6702						 new_eb_bh[i],
6703						 OCFS2_JOURNAL_ACCESS_CREATE);
6704		if (status < 0) {
6705			mlog_errno(status);
6706			goto bail;
6707		}
6708
6709		memset(new_eb_bh[i]->b_data, 0, osb->sb->s_blocksize);
6710		eb = (struct ocfs2_extent_block *) new_eb_bh[i]->b_data;
6711
6712		/* We can't guarantee that buffer head is still cached, so
6713		 * polutlate the extent block again.
6714		 */
6715		strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
6716		eb->h_blkno = cpu_to_le64(bf->free_blk);
6717		eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
6718		eb->h_suballoc_slot = cpu_to_le16(real_slot);
6719		eb->h_suballoc_loc = cpu_to_le64(bf->free_bg);
6720		eb->h_suballoc_bit = cpu_to_le16(bf->free_bit);
6721		eb->h_list.l_count =
6722			cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));
6723
6724		/* We'll also be dirtied by the caller, so
6725		 * this isn't absolutely necessary.
6726		 */
6727		ocfs2_journal_dirty(handle, new_eb_bh[i]);
6728
6729		if (!fl->f_first) {
6730			dealloc->c_first_suballocator = fl->f_next_suballocator;
6731			kfree(fl);
6732		}
6733		kfree(bf);
6734	}
6735
6736	*blk_given = i;
6737
6738bail:
6739	if (unlikely(status < 0)) {
6740		for (i = 0; i < blk_wanted; i++)
6741			brelse(new_eb_bh[i]);
6742	}
6743
6744	return status;
6745}
6746
6747int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt,
6748			      int type, int slot, u64 suballoc,
6749			      u64 blkno, unsigned int bit)
6750{
6751	int ret;
6752	struct ocfs2_per_slot_free_list *fl;
6753	struct ocfs2_cached_block_free *item;
6754
6755	fl = ocfs2_find_per_slot_free_list(type, slot, ctxt);
6756	if (fl == NULL) {
6757		ret = -ENOMEM;
6758		mlog_errno(ret);
6759		goto out;
6760	}
6761
6762	item = kzalloc(sizeof(*item), GFP_NOFS);
6763	if (item == NULL) {
6764		ret = -ENOMEM;
6765		mlog_errno(ret);
6766		goto out;
6767	}
6768
6769	trace_ocfs2_cache_block_dealloc(type, slot,
6770					(unsigned long long)suballoc,
6771					(unsigned long long)blkno, bit);
6772
6773	item->free_bg = suballoc;
6774	item->free_blk = blkno;
6775	item->free_bit = bit;
6776	item->free_next = fl->f_first;
6777
6778	fl->f_first = item;
6779
6780	ret = 0;
6781out:
6782	return ret;
6783}
6784
6785static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
6786					 struct ocfs2_extent_block *eb)
6787{
6788	return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE,
6789					 le16_to_cpu(eb->h_suballoc_slot),
6790					 le64_to_cpu(eb->h_suballoc_loc),
6791					 le64_to_cpu(eb->h_blkno),
6792					 le16_to_cpu(eb->h_suballoc_bit));
6793}
6794
6795static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh)
6796{
6797	set_buffer_uptodate(bh);
6798	mark_buffer_dirty(bh);
6799	return 0;
6800}
6801
6802void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle,
6803			      unsigned int from, unsigned int to,
6804			      struct page *page, int zero, u64 *phys)
6805{
6806	int ret, partial = 0;
6807	loff_t start_byte = ((loff_t)page->index << PAGE_SHIFT) + from;
6808	loff_t length = to - from;
6809
6810	ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0);
6811	if (ret)
6812		mlog_errno(ret);
6813
6814	if (zero)
6815		zero_user_segment(page, from, to);
6816
6817	/*
6818	 * Need to set the buffers we zero'd into uptodate
6819	 * here if they aren't - ocfs2_map_page_blocks()
6820	 * might've skipped some
6821	 */
6822	ret = walk_page_buffers(handle, page_buffers(page),
6823				from, to, &partial,
6824				ocfs2_zero_func);
6825	if (ret < 0)
6826		mlog_errno(ret);
6827	else if (ocfs2_should_order_data(inode)) {
6828		ret = ocfs2_jbd2_inode_add_write(handle, inode,
6829						 start_byte, length);
6830		if (ret < 0)
6831			mlog_errno(ret);
6832	}
6833
6834	if (!partial)
6835		SetPageUptodate(page);
6836
6837	flush_dcache_page(page);
6838}
6839
6840static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start,
6841				     loff_t end, struct page **pages,
6842				     int numpages, u64 phys, handle_t *handle)
6843{
6844	int i;
6845	struct page *page;
6846	unsigned int from, to = PAGE_SIZE;
6847	struct super_block *sb = inode->i_sb;
6848
6849	BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb)));
6850
6851	if (numpages == 0)
6852		goto out;
6853
6854	to = PAGE_SIZE;
6855	for(i = 0; i < numpages; i++) {
6856		page = pages[i];
6857
6858		from = start & (PAGE_SIZE - 1);
6859		if ((end >> PAGE_SHIFT) == page->index)
6860			to = end & (PAGE_SIZE - 1);
6861
6862		BUG_ON(from > PAGE_SIZE);
6863		BUG_ON(to > PAGE_SIZE);
6864
6865		ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1,
6866					 &phys);
6867
6868		start = (page->index + 1) << PAGE_SHIFT;
6869	}
6870out:
6871	if (pages)
6872		ocfs2_unlock_and_free_pages(pages, numpages);
6873}
6874
6875int ocfs2_grab_pages(struct inode *inode, loff_t start, loff_t end,
6876		     struct page **pages, int *num)
6877{
6878	int numpages, ret = 0;
6879	struct address_space *mapping = inode->i_mapping;
6880	unsigned long index;
6881	loff_t last_page_bytes;
6882
6883	BUG_ON(start > end);
6884
6885	numpages = 0;
6886	last_page_bytes = PAGE_ALIGN(end);
6887	index = start >> PAGE_SHIFT;
6888	do {
6889		pages[numpages] = find_or_create_page(mapping, index, GFP_NOFS);
6890		if (!pages[numpages]) {
6891			ret = -ENOMEM;
6892			mlog_errno(ret);
6893			goto out;
6894		}
6895
6896		numpages++;
6897		index++;
6898	} while (index < (last_page_bytes >> PAGE_SHIFT));
6899
6900out:
6901	if (ret != 0) {
6902		if (pages)
6903			ocfs2_unlock_and_free_pages(pages, numpages);
6904		numpages = 0;
6905	}
6906
6907	*num = numpages;
6908
6909	return ret;
6910}
6911
6912static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end,
6913				struct page **pages, int *num)
6914{
6915	struct super_block *sb = inode->i_sb;
6916
6917	BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits !=
6918	       (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits);
6919
6920	return ocfs2_grab_pages(inode, start, end, pages, num);
6921}
6922
6923/*
6924 * Zero the area past i_size but still within an allocated
6925 * cluster. This avoids exposing nonzero data on subsequent file
6926 * extends.
6927 *
6928 * We need to call this before i_size is updated on the inode because
6929 * otherwise block_write_full_page() will skip writeout of pages past
6930 * i_size. The new_i_size parameter is passed for this reason.
6931 */
6932int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle,
6933				  u64 range_start, u64 range_end)
6934{
6935	int ret = 0, numpages;
6936	struct page **pages = NULL;
6937	u64 phys;
6938	unsigned int ext_flags;
6939	struct super_block *sb = inode->i_sb;
6940
6941	/*
6942	 * File systems which don't support sparse files zero on every
6943	 * extend.
6944	 */
6945	if (!ocfs2_sparse_alloc(OCFS2_SB(sb)))
6946		return 0;
6947
6948	pages = kcalloc(ocfs2_pages_per_cluster(sb),
6949			sizeof(struct page *), GFP_NOFS);
6950	if (pages == NULL) {
6951		ret = -ENOMEM;
6952		mlog_errno(ret);
6953		goto out;
6954	}
6955
6956	if (range_start == range_end)
6957		goto out;
6958
6959	ret = ocfs2_extent_map_get_blocks(inode,
6960					  range_start >> sb->s_blocksize_bits,
6961					  &phys, NULL, &ext_flags);
6962	if (ret) {
6963		mlog_errno(ret);
6964		goto out;
6965	}
6966
6967	/*
6968	 * Tail is a hole, or is marked unwritten. In either case, we
6969	 * can count on read and write to return/push zero's.
6970	 */
6971	if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN)
6972		goto out;
6973
6974	ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages,
6975				   &numpages);
6976	if (ret) {
6977		mlog_errno(ret);
6978		goto out;
6979	}
6980
6981	ocfs2_zero_cluster_pages(inode, range_start, range_end, pages,
6982				 numpages, phys, handle);
6983
6984	/*
6985	 * Initiate writeout of the pages we zero'd here. We don't
6986	 * wait on them - the truncate_inode_pages() call later will
6987	 * do that for us.
6988	 */
6989	ret = filemap_fdatawrite_range(inode->i_mapping, range_start,
6990				       range_end - 1);
6991	if (ret)
6992		mlog_errno(ret);
6993
6994out:
6995	kfree(pages);
6996
6997	return ret;
6998}
6999
7000static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode,
7001					     struct ocfs2_dinode *di)
7002{
7003	unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits;
7004	unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size);
7005
7006	if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL)
7007		memset(&di->id2, 0, blocksize -
7008				    offsetof(struct ocfs2_dinode, id2) -
7009				    xattrsize);
7010	else
7011		memset(&di->id2, 0, blocksize -
7012				    offsetof(struct ocfs2_dinode, id2));
7013}
7014
7015void ocfs2_dinode_new_extent_list(struct inode *inode,
7016				  struct ocfs2_dinode *di)
7017{
7018	ocfs2_zero_dinode_id2_with_xattr(inode, di);
7019	di->id2.i_list.l_tree_depth = 0;
7020	di->id2.i_list.l_next_free_rec = 0;
7021	di->id2.i_list.l_count = cpu_to_le16(
7022		ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di));
7023}
7024
7025void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di)
7026{
7027	struct ocfs2_inode_info *oi = OCFS2_I(inode);
7028	struct ocfs2_inline_data *idata = &di->id2.i_data;
7029
7030	spin_lock(&oi->ip_lock);
7031	oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL;
7032	di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7033	spin_unlock(&oi->ip_lock);
7034
7035	/*
7036	 * We clear the entire i_data structure here so that all
7037	 * fields can be properly initialized.
7038	 */
7039	ocfs2_zero_dinode_id2_with_xattr(inode, di);
7040
7041	idata->id_count = cpu_to_le16(
7042			ocfs2_max_inline_data_with_xattr(inode->i_sb, di));
7043}
7044
7045int ocfs2_convert_inline_data_to_extents(struct inode *inode,
7046					 struct buffer_head *di_bh)
7047{
7048	int ret, has_data, num_pages = 0;
7049	int need_free = 0;
7050	u32 bit_off, num;
7051	handle_t *handle;
7052	u64 block;
7053	struct ocfs2_inode_info *oi = OCFS2_I(inode);
7054	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7055	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7056	struct ocfs2_alloc_context *data_ac = NULL;
7057	struct page *page = NULL;
 
7058	struct ocfs2_extent_tree et;
7059	int did_quota = 0;
7060
7061	has_data = i_size_read(inode) ? 1 : 0;
7062
7063	if (has_data) {
 
 
 
 
 
 
 
 
7064		ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
7065		if (ret) {
7066			mlog_errno(ret);
7067			goto out;
7068		}
7069	}
7070
7071	handle = ocfs2_start_trans(osb,
7072				   ocfs2_inline_to_extents_credits(osb->sb));
7073	if (IS_ERR(handle)) {
7074		ret = PTR_ERR(handle);
7075		mlog_errno(ret);
7076		goto out;
7077	}
7078
7079	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7080				      OCFS2_JOURNAL_ACCESS_WRITE);
7081	if (ret) {
7082		mlog_errno(ret);
7083		goto out_commit;
7084	}
7085
7086	if (has_data) {
7087		unsigned int page_end = min_t(unsigned, PAGE_SIZE,
7088							osb->s_clustersize);
7089		u64 phys;
7090
7091		ret = dquot_alloc_space_nodirty(inode,
7092				       ocfs2_clusters_to_bytes(osb->sb, 1));
7093		if (ret)
7094			goto out_commit;
7095		did_quota = 1;
7096
7097		data_ac->ac_resv = &oi->ip_la_data_resv;
7098
7099		ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
7100					   &num);
7101		if (ret) {
7102			mlog_errno(ret);
7103			goto out_commit;
7104		}
7105
7106		/*
7107		 * Save two copies, one for insert, and one that can
7108		 * be changed by ocfs2_map_and_dirty_page() below.
7109		 */
7110		block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off);
7111
7112		ret = ocfs2_grab_eof_pages(inode, 0, page_end, &page,
7113					   &num_pages);
 
 
 
 
 
 
 
7114		if (ret) {
7115			mlog_errno(ret);
7116			need_free = 1;
7117			goto out_commit;
7118		}
7119
7120		/*
7121		 * This should populate the 1st page for us and mark
7122		 * it up to date.
7123		 */
7124		ret = ocfs2_read_inline_data(inode, page, di_bh);
7125		if (ret) {
7126			mlog_errno(ret);
7127			need_free = 1;
7128			goto out_unlock;
7129		}
7130
7131		ocfs2_map_and_dirty_page(inode, handle, 0, page_end, page, 0,
7132					 &phys);
 
 
 
 
 
7133	}
7134
7135	spin_lock(&oi->ip_lock);
7136	oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
7137	di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
7138	spin_unlock(&oi->ip_lock);
7139
7140	ocfs2_update_inode_fsync_trans(handle, inode, 1);
7141	ocfs2_dinode_new_extent_list(inode, di);
7142
7143	ocfs2_journal_dirty(handle, di_bh);
7144
7145	if (has_data) {
7146		/*
7147		 * An error at this point should be extremely rare. If
7148		 * this proves to be false, we could always re-build
7149		 * the in-inode data from our pages.
7150		 */
7151		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7152		ret = ocfs2_insert_extent(handle, &et, 0, block, 1, 0, NULL);
7153		if (ret) {
7154			mlog_errno(ret);
7155			need_free = 1;
7156			goto out_unlock;
7157		}
7158
7159		inode->i_blocks = ocfs2_inode_sector_count(inode);
7160	}
7161
7162out_unlock:
7163	if (page)
7164		ocfs2_unlock_and_free_pages(&page, num_pages);
7165
7166out_commit:
7167	if (ret < 0 && did_quota)
7168		dquot_free_space_nodirty(inode,
7169					  ocfs2_clusters_to_bytes(osb->sb, 1));
7170
7171	if (need_free) {
7172		if (data_ac->ac_which == OCFS2_AC_USE_LOCAL)
7173			ocfs2_free_local_alloc_bits(osb, handle, data_ac,
7174					bit_off, num);
7175		else
7176			ocfs2_free_clusters(handle,
7177					data_ac->ac_inode,
7178					data_ac->ac_bh,
7179					ocfs2_clusters_to_blocks(osb->sb, bit_off),
7180					num);
7181	}
7182
7183	ocfs2_commit_trans(osb, handle);
7184
7185out:
7186	if (data_ac)
7187		ocfs2_free_alloc_context(data_ac);
 
 
7188	return ret;
7189}
7190
7191/*
7192 * It is expected, that by the time you call this function,
7193 * inode->i_size and fe->i_size have been adjusted.
7194 *
7195 * WARNING: This will kfree the truncate context
7196 */
7197int ocfs2_commit_truncate(struct ocfs2_super *osb,
7198			  struct inode *inode,
7199			  struct buffer_head *di_bh)
7200{
7201	int status = 0, i, flags = 0;
7202	u32 new_highest_cpos, range, trunc_cpos, trunc_len, phys_cpos, coff;
7203	u64 blkno = 0;
7204	struct ocfs2_extent_list *el;
7205	struct ocfs2_extent_rec *rec;
7206	struct ocfs2_path *path = NULL;
7207	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7208	struct ocfs2_extent_list *root_el = &(di->id2.i_list);
7209	u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
7210	struct ocfs2_extent_tree et;
7211	struct ocfs2_cached_dealloc_ctxt dealloc;
7212	struct ocfs2_refcount_tree *ref_tree = NULL;
7213
7214	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
7215	ocfs2_init_dealloc_ctxt(&dealloc);
7216
7217	new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
7218						     i_size_read(inode));
7219
7220	path = ocfs2_new_path(di_bh, &di->id2.i_list,
7221			      ocfs2_journal_access_di);
7222	if (!path) {
7223		status = -ENOMEM;
7224		mlog_errno(status);
7225		goto bail;
7226	}
7227
7228	ocfs2_extent_map_trunc(inode, new_highest_cpos);
7229
7230start:
7231	/*
7232	 * Check that we still have allocation to delete.
7233	 */
7234	if (OCFS2_I(inode)->ip_clusters == 0) {
7235		status = 0;
7236		goto bail;
7237	}
7238
7239	/*
7240	 * Truncate always works against the rightmost tree branch.
7241	 */
7242	status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX);
7243	if (status) {
7244		mlog_errno(status);
7245		goto bail;
7246	}
7247
7248	trace_ocfs2_commit_truncate(
7249		(unsigned long long)OCFS2_I(inode)->ip_blkno,
7250		new_highest_cpos,
7251		OCFS2_I(inode)->ip_clusters,
7252		path->p_tree_depth);
7253
7254	/*
7255	 * By now, el will point to the extent list on the bottom most
7256	 * portion of this tree. Only the tail record is considered in
7257	 * each pass.
7258	 *
7259	 * We handle the following cases, in order:
7260	 * - empty extent: delete the remaining branch
7261	 * - remove the entire record
7262	 * - remove a partial record
7263	 * - no record needs to be removed (truncate has completed)
7264	 */
7265	el = path_leaf_el(path);
7266	if (le16_to_cpu(el->l_next_free_rec) == 0) {
7267		ocfs2_error(inode->i_sb,
7268			    "Inode %llu has empty extent block at %llu\n",
7269			    (unsigned long long)OCFS2_I(inode)->ip_blkno,
7270			    (unsigned long long)path_leaf_bh(path)->b_blocknr);
7271		status = -EROFS;
7272		goto bail;
7273	}
7274
7275	i = le16_to_cpu(el->l_next_free_rec) - 1;
7276	rec = &el->l_recs[i];
7277	flags = rec->e_flags;
7278	range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
7279
7280	if (i == 0 && ocfs2_is_empty_extent(rec)) {
7281		/*
7282		 * Lower levels depend on this never happening, but it's best
7283		 * to check it up here before changing the tree.
7284		*/
7285		if (root_el->l_tree_depth && rec->e_int_clusters == 0) {
7286			mlog(ML_ERROR, "Inode %lu has an empty "
7287				    "extent record, depth %u\n", inode->i_ino,
7288				    le16_to_cpu(root_el->l_tree_depth));
7289			status = ocfs2_remove_rightmost_empty_extent(osb,
7290					&et, path, &dealloc);
7291			if (status) {
7292				mlog_errno(status);
7293				goto bail;
7294			}
7295
7296			ocfs2_reinit_path(path, 1);
7297			goto start;
7298		} else {
7299			trunc_cpos = le32_to_cpu(rec->e_cpos);
7300			trunc_len = 0;
7301			blkno = 0;
7302		}
7303	} else if (le32_to_cpu(rec->e_cpos) >= new_highest_cpos) {
7304		/*
7305		 * Truncate entire record.
7306		 */
7307		trunc_cpos = le32_to_cpu(rec->e_cpos);
7308		trunc_len = ocfs2_rec_clusters(el, rec);
7309		blkno = le64_to_cpu(rec->e_blkno);
7310	} else if (range > new_highest_cpos) {
7311		/*
7312		 * Partial truncate. it also should be
7313		 * the last truncate we're doing.
7314		 */
7315		trunc_cpos = new_highest_cpos;
7316		trunc_len = range - new_highest_cpos;
7317		coff = new_highest_cpos - le32_to_cpu(rec->e_cpos);
7318		blkno = le64_to_cpu(rec->e_blkno) +
7319				ocfs2_clusters_to_blocks(inode->i_sb, coff);
7320	} else {
7321		/*
7322		 * Truncate completed, leave happily.
7323		 */
7324		status = 0;
7325		goto bail;
7326	}
7327
7328	phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
7329
7330	if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
7331		status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
7332				&ref_tree, NULL);
7333		if (status) {
7334			mlog_errno(status);
7335			goto bail;
7336		}
7337	}
7338
7339	status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
7340					  phys_cpos, trunc_len, flags, &dealloc,
7341					  refcount_loc, true);
7342	if (status < 0) {
7343		mlog_errno(status);
7344		goto bail;
7345	}
7346
7347	ocfs2_reinit_path(path, 1);
7348
7349	/*
7350	 * The check above will catch the case where we've truncated
7351	 * away all allocation.
7352	 */
7353	goto start;
7354
7355bail:
7356	if (ref_tree)
7357		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
7358
7359	ocfs2_schedule_truncate_log_flush(osb, 1);
7360
7361	ocfs2_run_deallocs(osb, &dealloc);
7362
7363	ocfs2_free_path(path);
7364
7365	return status;
7366}
7367
7368/*
7369 * 'start' is inclusive, 'end' is not.
7370 */
7371int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh,
7372			  unsigned int start, unsigned int end, int trunc)
7373{
7374	int ret;
7375	unsigned int numbytes;
7376	handle_t *handle;
7377	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
7378	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
7379	struct ocfs2_inline_data *idata = &di->id2.i_data;
7380
7381	/* No need to punch hole beyond i_size. */
7382	if (start >= i_size_read(inode))
7383		return 0;
7384
7385	if (end > i_size_read(inode))
7386		end = i_size_read(inode);
7387
7388	BUG_ON(start > end);
7389
7390	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) ||
7391	    !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) ||
7392	    !ocfs2_supports_inline_data(osb)) {
7393		ocfs2_error(inode->i_sb,
7394			    "Inline data flags for inode %llu don't agree! Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n",
7395			    (unsigned long long)OCFS2_I(inode)->ip_blkno,
7396			    le16_to_cpu(di->i_dyn_features),
7397			    OCFS2_I(inode)->ip_dyn_features,
7398			    osb->s_feature_incompat);
7399		ret = -EROFS;
7400		goto out;
7401	}
7402
7403	handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
7404	if (IS_ERR(handle)) {
7405		ret = PTR_ERR(handle);
7406		mlog_errno(ret);
7407		goto out;
7408	}
7409
7410	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
7411				      OCFS2_JOURNAL_ACCESS_WRITE);
7412	if (ret) {
7413		mlog_errno(ret);
7414		goto out_commit;
7415	}
7416
7417	numbytes = end - start;
7418	memset(idata->id_data + start, 0, numbytes);
7419
7420	/*
7421	 * No need to worry about the data page here - it's been
7422	 * truncated already and inline data doesn't need it for
7423	 * pushing zero's to disk, so we'll let readpage pick it up
7424	 * later.
7425	 */
7426	if (trunc) {
7427		i_size_write(inode, start);
7428		di->i_size = cpu_to_le64(start);
7429	}
7430
7431	inode->i_blocks = ocfs2_inode_sector_count(inode);
7432	inode->i_ctime = inode->i_mtime = current_time(inode);
7433
7434	di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
7435	di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
7436
7437	ocfs2_update_inode_fsync_trans(handle, inode, 1);
7438	ocfs2_journal_dirty(handle, di_bh);
7439
7440out_commit:
7441	ocfs2_commit_trans(osb, handle);
7442
7443out:
7444	return ret;
7445}
7446
7447static int ocfs2_trim_extent(struct super_block *sb,
7448			     struct ocfs2_group_desc *gd,
7449			     u64 group, u32 start, u32 count)
7450{
7451	u64 discard, bcount;
7452	struct ocfs2_super *osb = OCFS2_SB(sb);
7453
7454	bcount = ocfs2_clusters_to_blocks(sb, count);
7455	discard = ocfs2_clusters_to_blocks(sb, start);
7456
7457	/*
7458	 * For the first cluster group, the gd->bg_blkno is not at the start
7459	 * of the group, but at an offset from the start. If we add it while
7460	 * calculating discard for first group, we will wrongly start fstrim a
7461	 * few blocks after the desried start block and the range can cross
7462	 * over into the next cluster group. So, add it only if this is not
7463	 * the first cluster group.
7464	 */
7465	if (group != osb->first_cluster_group_blkno)
7466		discard += le64_to_cpu(gd->bg_blkno);
7467
7468	trace_ocfs2_trim_extent(sb, (unsigned long long)discard, bcount);
7469
7470	return sb_issue_discard(sb, discard, bcount, GFP_NOFS, 0);
7471}
7472
7473static int ocfs2_trim_group(struct super_block *sb,
7474			    struct ocfs2_group_desc *gd, u64 group,
7475			    u32 start, u32 max, u32 minbits)
7476{
7477	int ret = 0, count = 0, next;
7478	void *bitmap = gd->bg_bitmap;
7479
7480	if (le16_to_cpu(gd->bg_free_bits_count) < minbits)
7481		return 0;
7482
7483	trace_ocfs2_trim_group((unsigned long long)le64_to_cpu(gd->bg_blkno),
7484			       start, max, minbits);
7485
7486	while (start < max) {
7487		start = ocfs2_find_next_zero_bit(bitmap, max, start);
7488		if (start >= max)
7489			break;
7490		next = ocfs2_find_next_bit(bitmap, max, start);
7491
7492		if ((next - start) >= minbits) {
7493			ret = ocfs2_trim_extent(sb, gd, group,
7494						start, next - start);
7495			if (ret < 0) {
7496				mlog_errno(ret);
7497				break;
7498			}
7499			count += next - start;
7500		}
7501		start = next + 1;
7502
7503		if (fatal_signal_pending(current)) {
7504			count = -ERESTARTSYS;
7505			break;
7506		}
7507
7508		if ((le16_to_cpu(gd->bg_free_bits_count) - count) < minbits)
7509			break;
7510	}
7511
7512	if (ret < 0)
7513		count = ret;
7514
7515	return count;
7516}
7517
7518static
7519int ocfs2_trim_mainbm(struct super_block *sb, struct fstrim_range *range)
7520{
7521	struct ocfs2_super *osb = OCFS2_SB(sb);
7522	u64 start, len, trimmed = 0, first_group, last_group = 0, group = 0;
7523	int ret, cnt;
7524	u32 first_bit, last_bit, minlen;
7525	struct buffer_head *main_bm_bh = NULL;
7526	struct inode *main_bm_inode = NULL;
7527	struct buffer_head *gd_bh = NULL;
7528	struct ocfs2_dinode *main_bm;
7529	struct ocfs2_group_desc *gd = NULL;
7530
7531	start = range->start >> osb->s_clustersize_bits;
7532	len = range->len >> osb->s_clustersize_bits;
7533	minlen = range->minlen >> osb->s_clustersize_bits;
7534
7535	if (minlen >= osb->bitmap_cpg || range->len < sb->s_blocksize)
7536		return -EINVAL;
7537
7538	trace_ocfs2_trim_mainbm(start, len, minlen);
7539
7540next_group:
7541	main_bm_inode = ocfs2_get_system_file_inode(osb,
7542						    GLOBAL_BITMAP_SYSTEM_INODE,
7543						    OCFS2_INVALID_SLOT);
7544	if (!main_bm_inode) {
7545		ret = -EIO;
7546		mlog_errno(ret);
7547		goto out;
7548	}
7549
7550	inode_lock(main_bm_inode);
7551
7552	ret = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 0);
7553	if (ret < 0) {
7554		mlog_errno(ret);
7555		goto out_mutex;
7556	}
7557	main_bm = (struct ocfs2_dinode *)main_bm_bh->b_data;
7558
7559	/*
7560	 * Do some check before trim the first group.
7561	 */
7562	if (!group) {
7563		if (start >= le32_to_cpu(main_bm->i_clusters)) {
7564			ret = -EINVAL;
7565			goto out_unlock;
7566		}
7567
7568		if (start + len > le32_to_cpu(main_bm->i_clusters))
7569			len = le32_to_cpu(main_bm->i_clusters) - start;
7570
7571		/*
7572		 * Determine first and last group to examine based on
7573		 * start and len
7574		 */
7575		first_group = ocfs2_which_cluster_group(main_bm_inode, start);
7576		if (first_group == osb->first_cluster_group_blkno)
7577			first_bit = start;
7578		else
7579			first_bit = start - ocfs2_blocks_to_clusters(sb,
7580								first_group);
7581		last_group = ocfs2_which_cluster_group(main_bm_inode,
7582						       start + len - 1);
7583		group = first_group;
7584	}
7585
7586	do {
7587		if (first_bit + len >= osb->bitmap_cpg)
7588			last_bit = osb->bitmap_cpg;
7589		else
7590			last_bit = first_bit + len;
7591
7592		ret = ocfs2_read_group_descriptor(main_bm_inode,
7593						  main_bm, group,
7594						  &gd_bh);
7595		if (ret < 0) {
7596			mlog_errno(ret);
7597			break;
7598		}
7599
7600		gd = (struct ocfs2_group_desc *)gd_bh->b_data;
7601		cnt = ocfs2_trim_group(sb, gd, group,
7602				       first_bit, last_bit, minlen);
7603		brelse(gd_bh);
7604		gd_bh = NULL;
7605		if (cnt < 0) {
7606			ret = cnt;
7607			mlog_errno(ret);
7608			break;
7609		}
7610
7611		trimmed += cnt;
7612		len -= osb->bitmap_cpg - first_bit;
7613		first_bit = 0;
7614		if (group == osb->first_cluster_group_blkno)
7615			group = ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7616		else
7617			group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg);
7618	} while (0);
7619
7620out_unlock:
7621	ocfs2_inode_unlock(main_bm_inode, 0);
7622	brelse(main_bm_bh);
7623	main_bm_bh = NULL;
7624out_mutex:
7625	inode_unlock(main_bm_inode);
7626	iput(main_bm_inode);
7627
7628	/*
7629	 * If all the groups trim are not done or failed, but we should release
7630	 * main_bm related locks for avoiding the current IO starve, then go to
7631	 * trim the next group
7632	 */
7633	if (ret >= 0 && group <= last_group) {
7634		cond_resched();
7635		goto next_group;
7636	}
7637out:
7638	range->len = trimmed * sb->s_blocksize;
7639	return ret;
7640}
7641
7642int ocfs2_trim_fs(struct super_block *sb, struct fstrim_range *range)
7643{
7644	int ret;
7645	struct ocfs2_super *osb = OCFS2_SB(sb);
7646	struct ocfs2_trim_fs_info info, *pinfo = NULL;
7647
7648	ocfs2_trim_fs_lock_res_init(osb);
7649
7650	trace_ocfs2_trim_fs(range->start, range->len, range->minlen);
7651
7652	ret = ocfs2_trim_fs_lock(osb, NULL, 1);
7653	if (ret < 0) {
7654		if (ret != -EAGAIN) {
7655			mlog_errno(ret);
7656			ocfs2_trim_fs_lock_res_uninit(osb);
7657			return ret;
7658		}
7659
7660		mlog(ML_NOTICE, "Wait for trim on device (%s) to "
7661		     "finish, which is running from another node.\n",
7662		     osb->dev_str);
7663		ret = ocfs2_trim_fs_lock(osb, &info, 0);
7664		if (ret < 0) {
7665			mlog_errno(ret);
7666			ocfs2_trim_fs_lock_res_uninit(osb);
7667			return ret;
7668		}
7669
7670		if (info.tf_valid && info.tf_success &&
7671		    info.tf_start == range->start &&
7672		    info.tf_len == range->len &&
7673		    info.tf_minlen == range->minlen) {
7674			/* Avoid sending duplicated trim to a shared device */
7675			mlog(ML_NOTICE, "The same trim on device (%s) was "
7676			     "just done from node (%u), return.\n",
7677			     osb->dev_str, info.tf_nodenum);
7678			range->len = info.tf_trimlen;
7679			goto out;
7680		}
7681	}
7682
7683	info.tf_nodenum = osb->node_num;
7684	info.tf_start = range->start;
7685	info.tf_len = range->len;
7686	info.tf_minlen = range->minlen;
7687
7688	ret = ocfs2_trim_mainbm(sb, range);
7689
7690	info.tf_trimlen = range->len;
7691	info.tf_success = (ret < 0 ? 0 : 1);
7692	pinfo = &info;
7693out:
7694	ocfs2_trim_fs_unlock(osb, pinfo);
7695	ocfs2_trim_fs_lock_res_uninit(osb);
7696	return ret;
7697}