xfs_buf_item_recover.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include "xfs.h"
   7#include "xfs_fs.h"
   8#include "xfs_shared.h"
   9#include "xfs_format.h"
  10#include "xfs_log_format.h"
  11#include "xfs_trans_resv.h"
  12#include "xfs_bit.h"
  13#include "xfs_mount.h"
  14#include "xfs_trans.h"
  15#include "xfs_buf_item.h"
  16#include "xfs_trans_priv.h"
  17#include "xfs_trace.h"
  18#include "xfs_log.h"
  19#include "xfs_log_priv.h"
  20#include "xfs_log_recover.h"
  21#include "xfs_error.h"
  22#include "xfs_inode.h"
  23#include "xfs_dir2.h"
  24#include "xfs_quota.h"
  25#include "xfs_alloc.h"
  26#include "xfs_ag.h"
  27#include "xfs_sb.h"
  28#include "xfs_rtgroup.h"
  29#include "xfs_rtbitmap.h"
  30
  31/*
  32 * This is the number of entries in the l_buf_cancel_table used during
  33 * recovery.
  34 */
  35#define	XLOG_BC_TABLE_SIZE	64
  36
  37#define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
  38	((log)->l_buf_cancel_table + ((uint64_t)blkno % XLOG_BC_TABLE_SIZE))
  39
  40/*
  41 * This structure is used during recovery to record the buf log items which
  42 * have been canceled and should not be replayed.
  43 */
  44struct xfs_buf_cancel {
  45	xfs_daddr_t		bc_blkno;
  46	uint			bc_len;
  47	int			bc_refcount;
  48	struct list_head	bc_list;
  49};
  50
  51static struct xfs_buf_cancel *
  52xlog_find_buffer_cancelled(
  53	struct xlog		*log,
  54	xfs_daddr_t		blkno,
  55	uint			len)
  56{
  57	struct list_head	*bucket;
  58	struct xfs_buf_cancel	*bcp;
  59
  60	if (!log->l_buf_cancel_table)
  61		return NULL;
  62
  63	bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
  64	list_for_each_entry(bcp, bucket, bc_list) {
  65		if (bcp->bc_blkno == blkno && bcp->bc_len == len)
  66			return bcp;
  67	}
  68
  69	return NULL;
  70}
  71
  72static bool
  73xlog_add_buffer_cancelled(
  74	struct xlog		*log,
  75	xfs_daddr_t		blkno,
  76	uint			len)
  77{
  78	struct xfs_buf_cancel	*bcp;
  79
  80	/*
  81	 * If we find an existing cancel record, this indicates that the buffer
  82	 * was cancelled multiple times.  To ensure that during pass 2 we keep
  83	 * the record in the table until we reach its last occurrence in the
  84	 * log, a reference count is kept to tell how many times we expect to
  85	 * see this record during the second pass.
  86	 */
  87	bcp = xlog_find_buffer_cancelled(log, blkno, len);
  88	if (bcp) {
  89		bcp->bc_refcount++;
  90		return false;
  91	}
  92
  93	bcp = kmalloc(sizeof(struct xfs_buf_cancel), GFP_KERNEL | __GFP_NOFAIL);
  94	bcp->bc_blkno = blkno;
  95	bcp->bc_len = len;
  96	bcp->bc_refcount = 1;
  97	list_add_tail(&bcp->bc_list, XLOG_BUF_CANCEL_BUCKET(log, blkno));
  98	return true;
  99}
 100
 101/*
 102 * Check if there is and entry for blkno, len in the buffer cancel record table.
 103 */
 104bool
 105xlog_is_buffer_cancelled(
 106	struct xlog		*log,
 107	xfs_daddr_t		blkno,
 108	uint			len)
 109{
 110	return xlog_find_buffer_cancelled(log, blkno, len) != NULL;
 111}
 112
 113/*
 114 * Check if there is and entry for blkno, len in the buffer cancel record table,
 115 * and decremented the reference count on it if there is one.
 116 *
 117 * Remove the cancel record once the refcount hits zero, so that if the same
 118 * buffer is re-used again after its last cancellation we actually replay the
 119 * changes made at that point.
 120 */
 121static bool
 122xlog_put_buffer_cancelled(
 123	struct xlog		*log,
 124	xfs_daddr_t		blkno,
 125	uint			len)
 126{
 127	struct xfs_buf_cancel	*bcp;
 128
 129	bcp = xlog_find_buffer_cancelled(log, blkno, len);
 130	if (!bcp) {
 131		ASSERT(0);
 132		return false;
 133	}
 134
 135	if (--bcp->bc_refcount == 0) {
 136		list_del(&bcp->bc_list);
 137		kfree(bcp);
 138	}
 139	return true;
 140}
 141
 142/* log buffer item recovery */
 143
 144/*
 145 * Sort buffer items for log recovery.  Most buffer items should end up on the
 146 * buffer list and are recovered first, with the following exceptions:
 147 *
 148 * 1. XFS_BLF_CANCEL buffers must be processed last because some log items
 149 *    might depend on the incor ecancellation record, and replaying a cancelled
 150 *    buffer item can remove the incore record.
 151 *
 152 * 2. XFS_BLF_INODE_BUF buffers are handled after most regular items so that
 153 *    we replay di_next_unlinked only after flushing the inode 'free' state
 154 *    to the inode buffer.
 155 *
 156 * See xlog_recover_reorder_trans for more details.
 157 */
 158STATIC enum xlog_recover_reorder
 159xlog_recover_buf_reorder(
 160	struct xlog_recover_item	*item)
 161{
 162	struct xfs_buf_log_format	*buf_f = item->ri_buf[0].iov_base;
 163
 164	if (buf_f->blf_flags & XFS_BLF_CANCEL)
 165		return XLOG_REORDER_CANCEL_LIST;
 166	if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
 167		return XLOG_REORDER_INODE_BUFFER_LIST;
 168	return XLOG_REORDER_BUFFER_LIST;
 169}
 170
 171STATIC void
 172xlog_recover_buf_ra_pass2(
 173	struct xlog                     *log,
 174	struct xlog_recover_item        *item)
 175{
 176	struct xfs_buf_log_format	*buf_f = item->ri_buf[0].iov_base;
 177
 178	xlog_buf_readahead(log, buf_f->blf_blkno, buf_f->blf_len, NULL);
 179}
 180
 181/*
 182 * Build up the table of buf cancel records so that we don't replay cancelled
 183 * data in the second pass.
 184 */
 185static int
 186xlog_recover_buf_commit_pass1(
 187	struct xlog			*log,
 188	struct xlog_recover_item	*item)
 189{
 190	struct xfs_buf_log_format	*bf = item->ri_buf[0].iov_base;
 191
 192	if (!xfs_buf_log_check_iovec(&item->ri_buf[0])) {
 193		xfs_err(log->l_mp, "bad buffer log item size (%zd)",
 194				item->ri_buf[0].iov_len);
 195		return -EFSCORRUPTED;
 196	}
 197
 198	if (!(bf->blf_flags & XFS_BLF_CANCEL))
 199		trace_xfs_log_recover_buf_not_cancel(log, bf);
 200	else if (xlog_add_buffer_cancelled(log, bf->blf_blkno, bf->blf_len))
 201		trace_xfs_log_recover_buf_cancel_add(log, bf);
 202	else
 203		trace_xfs_log_recover_buf_cancel_ref_inc(log, bf);
 204	return 0;
 205}
 206
 207/*
 208 * Validate the recovered buffer is of the correct type and attach the
 209 * appropriate buffer operations to them for writeback. Magic numbers are in a
 210 * few places:
 211 *	the first 16 bits of the buffer (inode buffer, dquot buffer),
 212 *	the first 32 bits of the buffer (most blocks),
 213 *	inside a struct xfs_da_blkinfo at the start of the buffer.
 214 */
 215static void
 216xlog_recover_validate_buf_type(
 217	struct xfs_mount		*mp,
 218	struct xfs_buf			*bp,
 219	struct xfs_buf_log_format	*buf_f,
 220	xfs_lsn_t			current_lsn)
 221{
 222	struct xfs_da_blkinfo		*info = bp->b_addr;
 223	uint32_t			magic32;
 224	uint16_t			magic16;
 225	uint16_t			magicda;
 226	char				*warnmsg = NULL;
 227
 228	/*
 229	 * We can only do post recovery validation on items on CRC enabled
 230	 * fielsystems as we need to know when the buffer was written to be able
 231	 * to determine if we should have replayed the item. If we replay old
 232	 * metadata over a newer buffer, then it will enter a temporarily
 233	 * inconsistent state resulting in verification failures. Hence for now
 234	 * just avoid the verification stage for non-crc filesystems
 235	 */
 236	if (!xfs_has_crc(mp))
 237		return;
 238
 239	magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
 240	magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
 241	magicda = be16_to_cpu(info->magic);
 242	switch (xfs_blft_from_flags(buf_f)) {
 243	case XFS_BLFT_BTREE_BUF:
 244		switch (magic32) {
 245		case XFS_ABTB_CRC_MAGIC:
 246		case XFS_ABTB_MAGIC:
 247			bp->b_ops = &xfs_bnobt_buf_ops;
 248			break;
 249		case XFS_ABTC_CRC_MAGIC:
 250		case XFS_ABTC_MAGIC:
 251			bp->b_ops = &xfs_cntbt_buf_ops;
 252			break;
 253		case XFS_IBT_CRC_MAGIC:
 254		case XFS_IBT_MAGIC:
 255			bp->b_ops = &xfs_inobt_buf_ops;
 256			break;
 257		case XFS_FIBT_CRC_MAGIC:
 258		case XFS_FIBT_MAGIC:
 259			bp->b_ops = &xfs_finobt_buf_ops;
 260			break;
 261		case XFS_BMAP_CRC_MAGIC:
 262		case XFS_BMAP_MAGIC:
 263			bp->b_ops = &xfs_bmbt_buf_ops;
 264			break;
 265		case XFS_RTRMAP_CRC_MAGIC:
 266			bp->b_ops = &xfs_rtrmapbt_buf_ops;
 267			break;
 268		case XFS_RMAP_CRC_MAGIC:
 269			bp->b_ops = &xfs_rmapbt_buf_ops;
 270			break;
 271		case XFS_REFC_CRC_MAGIC:
 272			bp->b_ops = &xfs_refcountbt_buf_ops;
 273			break;
 274		case XFS_RTREFC_CRC_MAGIC:
 275			bp->b_ops = &xfs_rtrefcountbt_buf_ops;
 276			break;
 277		default:
 278			warnmsg = "Bad btree block magic!";
 279			break;
 280		}
 281		break;
 282	case XFS_BLFT_AGF_BUF:
 283		if (magic32 != XFS_AGF_MAGIC) {
 284			warnmsg = "Bad AGF block magic!";
 285			break;
 286		}
 287		bp->b_ops = &xfs_agf_buf_ops;
 288		break;
 289	case XFS_BLFT_AGFL_BUF:
 290		if (magic32 != XFS_AGFL_MAGIC) {
 291			warnmsg = "Bad AGFL block magic!";
 292			break;
 293		}
 294		bp->b_ops = &xfs_agfl_buf_ops;
 295		break;
 296	case XFS_BLFT_AGI_BUF:
 297		if (magic32 != XFS_AGI_MAGIC) {
 298			warnmsg = "Bad AGI block magic!";
 299			break;
 300		}
 301		bp->b_ops = &xfs_agi_buf_ops;
 302		break;
 303	case XFS_BLFT_UDQUOT_BUF:
 304	case XFS_BLFT_PDQUOT_BUF:
 305	case XFS_BLFT_GDQUOT_BUF:
 306#ifdef CONFIG_XFS_QUOTA
 307		if (magic16 != XFS_DQUOT_MAGIC) {
 308			warnmsg = "Bad DQUOT block magic!";
 309			break;
 310		}
 311		bp->b_ops = &xfs_dquot_buf_ops;
 312#else
 313		xfs_alert(mp,
 314	"Trying to recover dquots without QUOTA support built in!");
 315		ASSERT(0);
 316#endif
 317		break;
 318	case XFS_BLFT_DINO_BUF:
 319		if (magic16 != XFS_DINODE_MAGIC) {
 320			warnmsg = "Bad INODE block magic!";
 321			break;
 322		}
 323		bp->b_ops = &xfs_inode_buf_ops;
 324		break;
 325	case XFS_BLFT_SYMLINK_BUF:
 326		if (magic32 != XFS_SYMLINK_MAGIC) {
 327			warnmsg = "Bad symlink block magic!";
 328			break;
 329		}
 330		bp->b_ops = &xfs_symlink_buf_ops;
 331		break;
 332	case XFS_BLFT_DIR_BLOCK_BUF:
 333		if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
 334		    magic32 != XFS_DIR3_BLOCK_MAGIC) {
 335			warnmsg = "Bad dir block magic!";
 336			break;
 337		}
 338		bp->b_ops = &xfs_dir3_block_buf_ops;
 339		break;
 340	case XFS_BLFT_DIR_DATA_BUF:
 341		if (magic32 != XFS_DIR2_DATA_MAGIC &&
 342		    magic32 != XFS_DIR3_DATA_MAGIC) {
 343			warnmsg = "Bad dir data magic!";
 344			break;
 345		}
 346		bp->b_ops = &xfs_dir3_data_buf_ops;
 347		break;
 348	case XFS_BLFT_DIR_FREE_BUF:
 349		if (magic32 != XFS_DIR2_FREE_MAGIC &&
 350		    magic32 != XFS_DIR3_FREE_MAGIC) {
 351			warnmsg = "Bad dir3 free magic!";
 352			break;
 353		}
 354		bp->b_ops = &xfs_dir3_free_buf_ops;
 355		break;
 356	case XFS_BLFT_DIR_LEAF1_BUF:
 357		if (magicda != XFS_DIR2_LEAF1_MAGIC &&
 358		    magicda != XFS_DIR3_LEAF1_MAGIC) {
 359			warnmsg = "Bad dir leaf1 magic!";
 360			break;
 361		}
 362		bp->b_ops = &xfs_dir3_leaf1_buf_ops;
 363		break;
 364	case XFS_BLFT_DIR_LEAFN_BUF:
 365		if (magicda != XFS_DIR2_LEAFN_MAGIC &&
 366		    magicda != XFS_DIR3_LEAFN_MAGIC) {
 367			warnmsg = "Bad dir leafn magic!";
 368			break;
 369		}
 370		bp->b_ops = &xfs_dir3_leafn_buf_ops;
 371		break;
 372	case XFS_BLFT_DA_NODE_BUF:
 373		if (magicda != XFS_DA_NODE_MAGIC &&
 374		    magicda != XFS_DA3_NODE_MAGIC) {
 375			warnmsg = "Bad da node magic!";
 376			break;
 377		}
 378		bp->b_ops = &xfs_da3_node_buf_ops;
 379		break;
 380	case XFS_BLFT_ATTR_LEAF_BUF:
 381		if (magicda != XFS_ATTR_LEAF_MAGIC &&
 382		    magicda != XFS_ATTR3_LEAF_MAGIC) {
 383			warnmsg = "Bad attr leaf magic!";
 384			break;
 385		}
 386		bp->b_ops = &xfs_attr3_leaf_buf_ops;
 387		break;
 388	case XFS_BLFT_ATTR_RMT_BUF:
 389		if (magic32 != XFS_ATTR3_RMT_MAGIC) {
 390			warnmsg = "Bad attr remote magic!";
 391			break;
 392		}
 393		bp->b_ops = &xfs_attr3_rmt_buf_ops;
 394		break;
 395	case XFS_BLFT_SB_BUF:
 396		if (magic32 != XFS_SB_MAGIC) {
 397			warnmsg = "Bad SB block magic!";
 398			break;
 399		}
 400		bp->b_ops = &xfs_sb_buf_ops;
 401		break;
 402#ifdef CONFIG_XFS_RT
 403	case XFS_BLFT_RTBITMAP_BUF:
 404		if (xfs_has_rtgroups(mp) && magic32 != XFS_RTBITMAP_MAGIC) {
 405			warnmsg = "Bad rtbitmap magic!";
 406			break;
 407		}
 408		bp->b_ops = xfs_rtblock_ops(mp, XFS_RTGI_BITMAP);
 409		break;
 410	case XFS_BLFT_RTSUMMARY_BUF:
 411		if (xfs_has_rtgroups(mp) && magic32 != XFS_RTSUMMARY_MAGIC) {
 412			warnmsg = "Bad rtsummary magic!";
 413			break;
 414		}
 415		bp->b_ops = xfs_rtblock_ops(mp, XFS_RTGI_SUMMARY);
 416		break;
 417#endif /* CONFIG_XFS_RT */
 418	default:
 419		xfs_warn(mp, "Unknown buffer type %d!",
 420			 xfs_blft_from_flags(buf_f));
 421		break;
 422	}
 423
 424	/*
 425	 * Nothing else to do in the case of a NULL current LSN as this means
 426	 * the buffer is more recent than the change in the log and will be
 427	 * skipped.
 428	 */
 429	if (current_lsn == NULLCOMMITLSN)
 430		return;
 431
 432	if (warnmsg) {
 433		xfs_warn(mp, warnmsg);
 434		ASSERT(0);
 435	}
 436
 437	/*
 438	 * We must update the metadata LSN of the buffer as it is written out to
 439	 * ensure that older transactions never replay over this one and corrupt
 440	 * the buffer. This can occur if log recovery is interrupted at some
 441	 * point after the current transaction completes, at which point a
 442	 * subsequent mount starts recovery from the beginning.
 443	 *
 444	 * Write verifiers update the metadata LSN from log items attached to
 445	 * the buffer. Therefore, initialize a bli purely to carry the LSN to
 446	 * the verifier.
 447	 */
 448	if (bp->b_ops) {
 449		struct xfs_buf_log_item	*bip;
 450
 451		bp->b_flags |= _XBF_LOGRECOVERY;
 452		xfs_buf_item_init(bp, mp);
 453		bip = bp->b_log_item;
 454		bip->bli_item.li_lsn = current_lsn;
 455	}
 456}
 457
 458/*
 459 * Perform a 'normal' buffer recovery.  Each logged region of the
 460 * buffer should be copied over the corresponding region in the
 461 * given buffer.  The bitmap in the buf log format structure indicates
 462 * where to place the logged data.
 463 */
 464STATIC void
 465xlog_recover_do_reg_buffer(
 466	struct xfs_mount		*mp,
 467	struct xlog_recover_item	*item,
 468	struct xfs_buf			*bp,
 469	struct xfs_buf_log_format	*buf_f,
 470	xfs_lsn_t			current_lsn)
 471{
 472	int			i;
 473	int			bit;
 474	int			nbits;
 475	xfs_failaddr_t		fa;
 476	const size_t		size_disk_dquot = sizeof(struct xfs_disk_dquot);
 477
 478	trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
 479
 480	bit = 0;
 481	i = 1;  /* 0 is the buf format structure */
 482	while (1) {
 483		bit = xfs_next_bit(buf_f->blf_data_map,
 484				   buf_f->blf_map_size, bit);
 485		if (bit == -1)
 486			break;
 487		nbits = xfs_contig_bits(buf_f->blf_data_map,
 488					buf_f->blf_map_size, bit);
 489		ASSERT(nbits > 0);
 490		ASSERT(item->ri_buf[i].iov_base != NULL);
 491		ASSERT(item->ri_buf[i].iov_len % XFS_BLF_CHUNK == 0);
 492		ASSERT(BBTOB(bp->b_length) >=
 493		       ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
 494
 495		/*
 496		 * The dirty regions logged in the buffer, even though
 497		 * contiguous, may span multiple chunks. This is because the
 498		 * dirty region may span a physical page boundary in a buffer
 499		 * and hence be split into two separate vectors for writing into
 500		 * the log. Hence we need to trim nbits back to the length of
 501		 * the current region being copied out of the log.
 502		 */
 503		if (item->ri_buf[i].iov_len < (nbits << XFS_BLF_SHIFT))
 504			nbits = item->ri_buf[i].iov_len >> XFS_BLF_SHIFT;
 505
 506		/*
 507		 * Do a sanity check if this is a dquot buffer. Just checking
 508		 * the first dquot in the buffer should do. XXXThis is
 509		 * probably a good thing to do for other buf types also.
 510		 */
 511		fa = NULL;
 512		if (buf_f->blf_flags &
 513		   (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
 514			if (item->ri_buf[i].iov_base == NULL) {
 515				xfs_alert(mp,
 516					"XFS: NULL dquot in %s.", __func__);
 517				goto next;
 518			}
 519			if (item->ri_buf[i].iov_len < size_disk_dquot) {
 520				xfs_alert(mp,
 521					"XFS: dquot too small (%zd) in %s.",
 522					item->ri_buf[i].iov_len, __func__);
 523				goto next;
 524			}
 525			fa = xfs_dquot_verify(mp, item->ri_buf[i].iov_base, -1);
 526			if (fa) {
 527				xfs_alert(mp,
 528	"dquot corrupt at %pS trying to replay into block 0x%llx",
 529					fa, xfs_buf_daddr(bp));
 530				goto next;
 531			}
 532		}
 533
 534		memcpy(xfs_buf_offset(bp,
 535			(uint)bit << XFS_BLF_SHIFT),	/* dest */
 536			item->ri_buf[i].iov_base,		/* source */
 537			nbits<<XFS_BLF_SHIFT);		/* length */
 538 next:
 539		i++;
 540		bit += nbits;
 541	}
 542
 543	/* Shouldn't be any more regions */
 544	ASSERT(i == item->ri_total);
 545
 546	xlog_recover_validate_buf_type(mp, bp, buf_f, current_lsn);
 547}
 548
 549/*
 550 * Perform a dquot buffer recovery.
 551 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
 552 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
 553 * Else, treat it as a regular buffer and do recovery.
 554 *
 555 * Return false if the buffer was tossed and true if we recovered the buffer to
 556 * indicate to the caller if the buffer needs writing.
 557 */
 558STATIC bool
 559xlog_recover_do_dquot_buffer(
 560	struct xfs_mount		*mp,
 561	struct xlog			*log,
 562	struct xlog_recover_item	*item,
 563	struct xfs_buf			*bp,
 564	struct xfs_buf_log_format	*buf_f)
 565{
 566	uint			type;
 567
 568	trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
 569
 570	/*
 571	 * Filesystems are required to send in quota flags at mount time.
 572	 */
 573	if (!mp->m_qflags)
 574		return false;
 575
 576	type = 0;
 577	if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
 578		type |= XFS_DQTYPE_USER;
 579	if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
 580		type |= XFS_DQTYPE_PROJ;
 581	if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
 582		type |= XFS_DQTYPE_GROUP;
 583	/*
 584	 * This type of quotas was turned off, so ignore this buffer
 585	 */
 586	if (log->l_quotaoffs_flag & type)
 587		return false;
 588
 589	xlog_recover_do_reg_buffer(mp, item, bp, buf_f, NULLCOMMITLSN);
 590	return true;
 591}
 592
 593/*
 594 * Perform recovery for a buffer full of inodes.  In these buffers, the only
 595 * data which should be recovered is that which corresponds to the
 596 * di_next_unlinked pointers in the on disk inode structures.  The rest of the
 597 * data for the inodes is always logged through the inodes themselves rather
 598 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
 599 *
 600 * The only time when buffers full of inodes are fully recovered is when the
 601 * buffer is full of newly allocated inodes.  In this case the buffer will
 602 * not be marked as an inode buffer and so will be sent to
 603 * xlog_recover_do_reg_buffer() below during recovery.
 604 */
 605STATIC int
 606xlog_recover_do_inode_buffer(
 607	struct xfs_mount		*mp,
 608	struct xlog_recover_item	*item,
 609	struct xfs_buf			*bp,
 610	struct xfs_buf_log_format	*buf_f)
 611{
 612	int				i;
 613	int				item_index = 0;
 614	int				bit = 0;
 615	int				nbits = 0;
 616	int				reg_buf_offset = 0;
 617	int				reg_buf_bytes = 0;
 618	int				next_unlinked_offset;
 619	int				inodes_per_buf;
 620	xfs_agino_t			*logged_nextp;
 621	xfs_agino_t			*buffer_nextp;
 622
 623	trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
 624
 625	/*
 626	 * Post recovery validation only works properly on CRC enabled
 627	 * filesystems.
 628	 */
 629	if (xfs_has_crc(mp))
 630		bp->b_ops = &xfs_inode_buf_ops;
 631
 632	inodes_per_buf = BBTOB(bp->b_length) >> mp->m_sb.sb_inodelog;
 633	for (i = 0; i < inodes_per_buf; i++) {
 634		next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
 635			offsetof(struct xfs_dinode, di_next_unlinked);
 636
 637		while (next_unlinked_offset >=
 638		       (reg_buf_offset + reg_buf_bytes)) {
 639			/*
 640			 * The next di_next_unlinked field is beyond
 641			 * the current logged region.  Find the next
 642			 * logged region that contains or is beyond
 643			 * the current di_next_unlinked field.
 644			 */
 645			bit += nbits;
 646			bit = xfs_next_bit(buf_f->blf_data_map,
 647					   buf_f->blf_map_size, bit);
 648
 649			/*
 650			 * If there are no more logged regions in the
 651			 * buffer, then we're done.
 652			 */
 653			if (bit == -1)
 654				return 0;
 655
 656			nbits = xfs_contig_bits(buf_f->blf_data_map,
 657						buf_f->blf_map_size, bit);
 658			ASSERT(nbits > 0);
 659			reg_buf_offset = bit << XFS_BLF_SHIFT;
 660			reg_buf_bytes = nbits << XFS_BLF_SHIFT;
 661			item_index++;
 662		}
 663
 664		/*
 665		 * If the current logged region starts after the current
 666		 * di_next_unlinked field, then move on to the next
 667		 * di_next_unlinked field.
 668		 */
 669		if (next_unlinked_offset < reg_buf_offset)
 670			continue;
 671
 672		ASSERT(item->ri_buf[item_index].iov_base != NULL);
 673		ASSERT((item->ri_buf[item_index].iov_len % XFS_BLF_CHUNK) == 0);
 674		ASSERT((reg_buf_offset + reg_buf_bytes) <= BBTOB(bp->b_length));
 675
 676		/*
 677		 * The current logged region contains a copy of the
 678		 * current di_next_unlinked field.  Extract its value
 679		 * and copy it to the buffer copy.
 680		 */
 681		logged_nextp = item->ri_buf[item_index].iov_base +
 682				next_unlinked_offset - reg_buf_offset;
 683		if (XFS_IS_CORRUPT(mp, *logged_nextp == 0)) {
 684			xfs_alert(mp,
 685		"Bad inode buffer log record (ptr = "PTR_FMT", bp = "PTR_FMT"). "
 686		"Trying to replay bad (0) inode di_next_unlinked field.",
 687				item, bp);
 688			return -EFSCORRUPTED;
 689		}
 690
 691		buffer_nextp = xfs_buf_offset(bp, next_unlinked_offset);
 692		*buffer_nextp = *logged_nextp;
 693
 694		/*
 695		 * If necessary, recalculate the CRC in the on-disk inode. We
 696		 * have to leave the inode in a consistent state for whoever
 697		 * reads it next....
 698		 */
 699		xfs_dinode_calc_crc(mp,
 700				xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
 701
 702	}
 703
 704	return 0;
 705}
 706
 707/*
 708 * Update the in-memory superblock and perag structures from the primary SB
 709 * buffer.
 710 *
 711 * This is required because transactions running after growfs may require the
 712 * updated values to be set in a previous fully commit transaction.
 713 */
 714static int
 715xlog_recover_do_primary_sb_buffer(
 716	struct xfs_mount		*mp,
 717	struct xlog_recover_item	*item,
 718	struct xfs_buf			*bp,
 719	struct xfs_buf_log_format	*buf_f,
 720	xfs_lsn_t			current_lsn)
 721{
 722	struct xfs_dsb			*dsb = bp->b_addr;
 723	xfs_agnumber_t			orig_agcount = mp->m_sb.sb_agcount;
 724	xfs_rgnumber_t			orig_rgcount = mp->m_sb.sb_rgcount;
 725	int				error;
 726
 727	xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn);
 728
 729	if (orig_agcount == 0) {
 730		xfs_alert(mp, "Trying to grow file system without AGs");
 731		return -EFSCORRUPTED;
 732	}
 733
 734	/*
 735	 * Update the in-core super block from the freshly recovered on-disk one.
 736	 */
 737	xfs_sb_from_disk(&mp->m_sb, dsb);
 738
 739	/*
 740	 * Grow can change the device size.  Mirror that into the buftarg.
 741	 */
 742	mp->m_ddev_targp->bt_nr_sectors =
 743		XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
 744	if (mp->m_rtdev_targp && mp->m_rtdev_targp != mp->m_ddev_targp) {
 745		mp->m_rtdev_targp->bt_nr_sectors =
 746			XFS_FSB_TO_BB(mp, mp->m_sb.sb_rblocks);
 747	}
 748
 749	if (mp->m_sb.sb_agcount < orig_agcount) {
 750		xfs_alert(mp, "Shrinking AG count in log recovery not supported");
 751		return -EFSCORRUPTED;
 752	}
 753	if (mp->m_sb.sb_rgcount < orig_rgcount) {
 754		xfs_warn(mp,
 755 "Shrinking rtgroup count in log recovery not supported");
 756		return -EFSCORRUPTED;
 757	}
 758
 759	/*
 760	 * If the last AG was grown or shrunk, we also need to update the
 761	 * length in the in-core perag structure and values depending on it.
 762	 */
 763	error = xfs_update_last_ag_size(mp, orig_agcount);
 764	if (error)
 765		return error;
 766
 767	/*
 768	 * If the last rtgroup was grown or shrunk, we also need to update the
 769	 * length in the in-core rtgroup structure and values depending on it.
 770	 * Ignore this on any filesystem with zero rtgroups.
 771	 */
 772	if (orig_rgcount > 0) {
 773		error = xfs_update_last_rtgroup_size(mp, orig_rgcount);
 774		if (error)
 775			return error;
 776	}
 777
 778	/*
 779	 * Initialize the new perags, and also update various block and inode
 780	 * allocator setting based off the number of AGs or total blocks.
 781	 * Because of the latter this also needs to happen if the agcount did
 782	 * not change.
 783	 */
 784	error = xfs_initialize_perag(mp, orig_agcount, mp->m_sb.sb_agcount,
 785			mp->m_sb.sb_dblocks, &mp->m_maxagi);
 786	if (error) {
 787		xfs_warn(mp, "Failed recovery per-ag init: %d", error);
 788		return error;
 789	}
 790	mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
 791
 792	error = xfs_initialize_rtgroups(mp, orig_rgcount, mp->m_sb.sb_rgcount,
 793			mp->m_sb.sb_rextents);
 794	if (error) {
 795		xfs_warn(mp, "Failed recovery rtgroup init: %d", error);
 796		return error;
 797	}
 798	return 0;
 799}
 800
 801/*
 802 * V5 filesystems know the age of the buffer on disk being recovered. We can
 803 * have newer objects on disk than we are replaying, and so for these cases we
 804 * don't want to replay the current change as that will make the buffer contents
 805 * temporarily invalid on disk.
 806 *
 807 * The magic number might not match the buffer type we are going to recover
 808 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags.  Hence
 809 * extract the LSN of the existing object in the buffer based on it's current
 810 * magic number.  If we don't recognise the magic number in the buffer, then
 811 * return a LSN of -1 so that the caller knows it was an unrecognised block and
 812 * so can recover the buffer.
 813 *
 814 * Note: we cannot rely solely on magic number matches to determine that the
 815 * buffer has a valid LSN - we also need to verify that it belongs to this
 816 * filesystem, so we need to extract the object's LSN and compare it to that
 817 * which we read from the superblock. If the UUIDs don't match, then we've got a
 818 * stale metadata block from an old filesystem instance that we need to recover
 819 * over the top of.
 820 */
 821static xfs_lsn_t
 822xlog_recover_get_buf_lsn(
 823	struct xfs_mount	*mp,
 824	struct xfs_buf		*bp,
 825	struct xfs_buf_log_format *buf_f)
 826{
 827	uint32_t		magic32;
 828	uint16_t		magic16;
 829	uint16_t		magicda;
 830	void			*blk = bp->b_addr;
 831	uuid_t			*uuid;
 832	xfs_lsn_t		lsn = -1;
 833	uint16_t		blft;
 834
 835	/* v4 filesystems always recover immediately */
 836	if (!xfs_has_crc(mp))
 837		goto recover_immediately;
 838
 839	/*
 840	 * realtime bitmap and summary file blocks do not have magic numbers or
 841	 * UUIDs, so we must recover them immediately.
 842	 */
 843	blft = xfs_blft_from_flags(buf_f);
 844	if (!xfs_has_rtgroups(mp) && (blft == XFS_BLFT_RTBITMAP_BUF ||
 845				      blft == XFS_BLFT_RTSUMMARY_BUF))
 846		goto recover_immediately;
 847
 848	magic32 = be32_to_cpu(*(__be32 *)blk);
 849	switch (magic32) {
 850	case XFS_RTSUMMARY_MAGIC:
 851	case XFS_RTBITMAP_MAGIC: {
 852		struct xfs_rtbuf_blkinfo	*hdr = blk;
 853
 854		lsn = be64_to_cpu(hdr->rt_lsn);
 855		uuid = &hdr->rt_uuid;
 856		break;
 857	}
 858	case XFS_ABTB_CRC_MAGIC:
 859	case XFS_ABTC_CRC_MAGIC:
 860	case XFS_ABTB_MAGIC:
 861	case XFS_ABTC_MAGIC:
 862	case XFS_RMAP_CRC_MAGIC:
 863	case XFS_REFC_CRC_MAGIC:
 864	case XFS_FIBT_CRC_MAGIC:
 865	case XFS_FIBT_MAGIC:
 866	case XFS_IBT_CRC_MAGIC:
 867	case XFS_IBT_MAGIC: {
 868		struct xfs_btree_block *btb = blk;
 869
 870		lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
 871		uuid = &btb->bb_u.s.bb_uuid;
 872		break;
 873	}
 874	case XFS_RTRMAP_CRC_MAGIC:
 875	case XFS_RTREFC_CRC_MAGIC:
 876	case XFS_BMAP_CRC_MAGIC:
 877	case XFS_BMAP_MAGIC: {
 878		struct xfs_btree_block *btb = blk;
 879
 880		lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
 881		uuid = &btb->bb_u.l.bb_uuid;
 882		break;
 883	}
 884	case XFS_AGF_MAGIC:
 885		lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
 886		uuid = &((struct xfs_agf *)blk)->agf_uuid;
 887		break;
 888	case XFS_AGFL_MAGIC:
 889		lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
 890		uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
 891		break;
 892	case XFS_AGI_MAGIC:
 893		lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
 894		uuid = &((struct xfs_agi *)blk)->agi_uuid;
 895		break;
 896	case XFS_SYMLINK_MAGIC:
 897		lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
 898		uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
 899		break;
 900	case XFS_DIR3_BLOCK_MAGIC:
 901	case XFS_DIR3_DATA_MAGIC:
 902	case XFS_DIR3_FREE_MAGIC:
 903		lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
 904		uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
 905		break;
 906	case XFS_ATTR3_RMT_MAGIC:
 907		/*
 908		 * Remote attr blocks are written synchronously, rather than
 909		 * being logged. That means they do not contain a valid LSN
 910		 * (i.e. transactionally ordered) in them, and hence any time we
 911		 * see a buffer to replay over the top of a remote attribute
 912		 * block we should simply do so.
 913		 */
 914		goto recover_immediately;
 915	case XFS_SB_MAGIC:
 916		/*
 917		 * superblock uuids are magic. We may or may not have a
 918		 * sb_meta_uuid on disk, but it will be set in the in-core
 919		 * superblock. We set the uuid pointer for verification
 920		 * according to the superblock feature mask to ensure we check
 921		 * the relevant UUID in the superblock.
 922		 */
 923		lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
 924		if (xfs_has_metauuid(mp))
 925			uuid = &((struct xfs_dsb *)blk)->sb_meta_uuid;
 926		else
 927			uuid = &((struct xfs_dsb *)blk)->sb_uuid;
 928		break;
 929	default:
 930		break;
 931	}
 932
 933	if (lsn != (xfs_lsn_t)-1) {
 934		if (!uuid_equal(&mp->m_sb.sb_meta_uuid, uuid))
 935			goto recover_immediately;
 936		return lsn;
 937	}
 938
 939	magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
 940	switch (magicda) {
 941	case XFS_DIR3_LEAF1_MAGIC:
 942	case XFS_DIR3_LEAFN_MAGIC:
 943	case XFS_ATTR3_LEAF_MAGIC:
 944	case XFS_DA3_NODE_MAGIC:
 945		lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
 946		uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
 947		break;
 948	default:
 949		break;
 950	}
 951
 952	if (lsn != (xfs_lsn_t)-1) {
 953		if (!uuid_equal(&mp->m_sb.sb_meta_uuid, uuid))
 954			goto recover_immediately;
 955		return lsn;
 956	}
 957
 958	/*
 959	 * We do individual object checks on dquot and inode buffers as they
 960	 * have their own individual LSN records. Also, we could have a stale
 961	 * buffer here, so we have to at least recognise these buffer types.
 962	 *
 963	 * A notd complexity here is inode unlinked list processing - it logs
 964	 * the inode directly in the buffer, but we don't know which inodes have
 965	 * been modified, and there is no global buffer LSN. Hence we need to
 966	 * recover all inode buffer types immediately. This problem will be
 967	 * fixed by logical logging of the unlinked list modifications.
 968	 */
 969	magic16 = be16_to_cpu(*(__be16 *)blk);
 970	switch (magic16) {
 971	case XFS_DQUOT_MAGIC:
 972	case XFS_DINODE_MAGIC:
 973		goto recover_immediately;
 974	default:
 975		break;
 976	}
 977
 978	/* unknown buffer contents, recover immediately */
 979
 980recover_immediately:
 981	return (xfs_lsn_t)-1;
 982
 983}
 984
 985/*
 986 * This routine replays a modification made to a buffer at runtime.
 987 * There are actually two types of buffer, regular and inode, which
 988 * are handled differently.  Inode buffers are handled differently
 989 * in that we only recover a specific set of data from them, namely
 990 * the inode di_next_unlinked fields.  This is because all other inode
 991 * data is actually logged via inode records and any data we replay
 992 * here which overlaps that may be stale.
 993 *
 994 * When meta-data buffers are freed at run time we log a buffer item
 995 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
 996 * of the buffer in the log should not be replayed at recovery time.
 997 * This is so that if the blocks covered by the buffer are reused for
 998 * file data before we crash we don't end up replaying old, freed
 999 * meta-data into a user's file.
1000 *
1001 * To handle the cancellation of buffer log items, we make two passes
1002 * over the log during recovery.  During the first we build a table of
1003 * those buffers which have been cancelled, and during the second we
1004 * only replay those buffers which do not have corresponding cancel
1005 * records in the table.  See xlog_recover_buf_pass[1,2] above
1006 * for more details on the implementation of the table of cancel records.
1007 */
1008STATIC int
1009xlog_recover_buf_commit_pass2(
1010	struct xlog			*log,
1011	struct list_head		*buffer_list,
1012	struct xlog_recover_item	*item,
1013	xfs_lsn_t			current_lsn)
1014{
1015	struct xfs_buf_log_format	*buf_f = item->ri_buf[0].iov_base;
1016	struct xfs_mount		*mp = log->l_mp;
1017	struct xfs_buf			*bp;
1018	int				error;
1019	xfs_lsn_t			lsn;
1020
1021	/*
1022	 * In this pass we only want to recover all the buffers which have
1023	 * not been cancelled and are not cancellation buffers themselves.
1024	 */
1025	if (buf_f->blf_flags & XFS_BLF_CANCEL) {
1026		if (xlog_put_buffer_cancelled(log, buf_f->blf_blkno,
1027				buf_f->blf_len))
1028			goto cancelled;
1029	} else {
1030
1031		if (xlog_is_buffer_cancelled(log, buf_f->blf_blkno,
1032				buf_f->blf_len))
1033			goto cancelled;
1034	}
1035
1036	trace_xfs_log_recover_buf_recover(log, buf_f);
1037	error = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
1038			  0, &bp, NULL);
1039	if (error)
1040		return error;
1041
1042	/*
1043	 * Recover the buffer only if we get an LSN from it and it's less than
1044	 * the lsn of the transaction we are replaying.
1045	 *
1046	 * Note that we have to be extremely careful of readahead here.
1047	 * Readahead does not attach verfiers to the buffers so if we don't
1048	 * actually do any replay after readahead because of the LSN we found
1049	 * in the buffer if more recent than that current transaction then we
1050	 * need to attach the verifier directly. Failure to do so can lead to
1051	 * future recovery actions (e.g. EFI and unlinked list recovery) can
1052	 * operate on the buffers and they won't get the verifier attached. This
1053	 * can lead to blocks on disk having the correct content but a stale
1054	 * CRC.
1055	 *
1056	 * It is safe to assume these clean buffers are currently up to date.
1057	 * If the buffer is dirtied by a later transaction being replayed, then
1058	 * the verifier will be reset to match whatever recover turns that
1059	 * buffer into.
1060	 */
1061	lsn = xlog_recover_get_buf_lsn(mp, bp, buf_f);
1062	if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
1063		trace_xfs_log_recover_buf_skip(log, buf_f);
1064		xlog_recover_validate_buf_type(mp, bp, buf_f, NULLCOMMITLSN);
1065
1066		/*
1067		 * We're skipping replay of this buffer log item due to the log
1068		 * item LSN being behind the ondisk buffer.  Verify the buffer
1069		 * contents since we aren't going to run the write verifier.
1070		 */
1071		if (bp->b_ops) {
1072			bp->b_ops->verify_read(bp);
1073			error = bp->b_error;
1074		}
1075		goto out_release;
1076	}
1077
1078	if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
1079		error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
1080		if (error)
1081			goto out_release;
1082	} else if (buf_f->blf_flags &
1083		  (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
1084		bool	dirty;
1085
1086		dirty = xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
1087		if (!dirty)
1088			goto out_release;
1089	} else if ((xfs_blft_from_flags(buf_f) & XFS_BLFT_SB_BUF) &&
1090			xfs_buf_daddr(bp) == 0) {
1091		error = xlog_recover_do_primary_sb_buffer(mp, item, bp, buf_f,
1092				current_lsn);
1093		if (error)
1094			goto out_writebuf;
1095
1096		/* Update the rt superblock if we have one. */
1097		if (xfs_has_rtsb(mp) && mp->m_rtsb_bp) {
1098			struct xfs_buf	*rtsb_bp = mp->m_rtsb_bp;
1099
1100			xfs_buf_lock(rtsb_bp);
1101			xfs_buf_hold(rtsb_bp);
1102			xfs_update_rtsb(rtsb_bp, bp);
1103			rtsb_bp->b_flags |= _XBF_LOGRECOVERY;
1104			xfs_buf_delwri_queue(rtsb_bp, buffer_list);
1105			xfs_buf_relse(rtsb_bp);
1106		}
1107	} else {
1108		xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn);
1109	}
1110
1111	/*
1112	 * Buffer held by buf log item during 'normal' buffer recovery must
1113	 * be committed through buffer I/O submission path to ensure proper
1114	 * release. When error occurs during sb buffer recovery, log shutdown
1115	 * will be done before submitting buffer list so that buffers can be
1116	 * released correctly through ioend failure path.
1117	 */
1118out_writebuf:
1119
1120	/*
1121	 * Perform delayed write on the buffer.  Asynchronous writes will be
1122	 * slower when taking into account all the buffers to be flushed.
1123	 *
1124	 * Also make sure that only inode buffers with good sizes stay in
1125	 * the buffer cache.  The kernel moves inodes in buffers of 1 block
1126	 * or inode_cluster_size bytes, whichever is bigger.  The inode
1127	 * buffers in the log can be a different size if the log was generated
1128	 * by an older kernel using unclustered inode buffers or a newer kernel
1129	 * running with a different inode cluster size.  Regardless, if
1130	 * the inode buffer size isn't max(blocksize, inode_cluster_size)
1131	 * for *our* value of inode_cluster_size, then we need to keep
1132	 * the buffer out of the buffer cache so that the buffer won't
1133	 * overlap with future reads of those inodes.
1134	 */
1135	if (XFS_DINODE_MAGIC ==
1136	    be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
1137	    (BBTOB(bp->b_length) != M_IGEO(log->l_mp)->inode_cluster_size)) {
1138		xfs_buf_stale(bp);
1139		error = xfs_bwrite(bp);
1140	} else {
1141		ASSERT(bp->b_mount == mp);
1142		bp->b_flags |= _XBF_LOGRECOVERY;
1143		xfs_buf_delwri_queue(bp, buffer_list);
1144	}
1145
1146out_release:
1147	xfs_buf_relse(bp);
1148	return error;
1149cancelled:
1150	trace_xfs_log_recover_buf_cancel(log, buf_f);
1151	return 0;
1152}
1153
1154const struct xlog_recover_item_ops xlog_buf_item_ops = {
1155	.item_type		= XFS_LI_BUF,
1156	.reorder		= xlog_recover_buf_reorder,
1157	.ra_pass2		= xlog_recover_buf_ra_pass2,
1158	.commit_pass1		= xlog_recover_buf_commit_pass1,
1159	.commit_pass2		= xlog_recover_buf_commit_pass2,
1160};
1161
1162#ifdef DEBUG
1163void
1164xlog_check_buf_cancel_table(
1165	struct xlog	*log)
1166{
1167	int		i;
1168
1169	for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
1170		ASSERT(list_empty(&log->l_buf_cancel_table[i]));
1171}
1172#endif
1173
1174int
1175xlog_alloc_buf_cancel_table(
1176	struct xlog	*log)
1177{
1178	void		*p;
1179	int		i;
1180
1181	ASSERT(log->l_buf_cancel_table == NULL);
1182
1183	p = kmalloc_array(XLOG_BC_TABLE_SIZE, sizeof(struct list_head),
1184			  GFP_KERNEL);
1185	if (!p)
1186		return -ENOMEM;
1187
1188	log->l_buf_cancel_table = p;
1189	for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
1190		INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
1191
1192	return 0;
1193}
1194
1195void
1196xlog_free_buf_cancel_table(
1197	struct xlog	*log)
1198{
1199	int		i;
1200
1201	if (!log->l_buf_cancel_table)
1202		return;
1203
1204	for (i = 0; i < XLOG_BC_TABLE_SIZE; i++) {
1205		struct xfs_buf_cancel	*bc;
1206
1207		while ((bc = list_first_entry_or_null(
1208				&log->l_buf_cancel_table[i],
1209				struct xfs_buf_cancel, bc_list))) {
1210			list_del(&bc->bc_list);
1211			kfree(bc);
1212		}
1213	}
1214
1215	kfree(log->l_buf_cancel_table);
1216	log->l_buf_cancel_table = NULL;
1217}