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