PageRenderTime 237ms CodeModel.GetById 7ms app.highlight 192ms RepoModel.GetById 1ms app.codeStats 2ms

/fs/xfs/xfs_log_recover.c

https://bitbucket.org/emiliolopez/linux
C | 5829 lines | 3640 code | 561 blank | 1628 comment | 673 complexity | 175fc7830e2a6de277b7ad8431160061 MD5 | raw file
   1/*
   2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3 * All Rights Reserved.
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it would be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write the Free Software Foundation,
  16 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18#include "xfs.h"
  19#include "xfs_fs.h"
  20#include "xfs_shared.h"
  21#include "xfs_format.h"
  22#include "xfs_log_format.h"
  23#include "xfs_trans_resv.h"
  24#include "xfs_bit.h"
  25#include "xfs_sb.h"
  26#include "xfs_mount.h"
  27#include "xfs_da_format.h"
  28#include "xfs_da_btree.h"
  29#include "xfs_inode.h"
  30#include "xfs_trans.h"
  31#include "xfs_log.h"
  32#include "xfs_log_priv.h"
  33#include "xfs_log_recover.h"
  34#include "xfs_inode_item.h"
  35#include "xfs_extfree_item.h"
  36#include "xfs_trans_priv.h"
  37#include "xfs_alloc.h"
  38#include "xfs_ialloc.h"
  39#include "xfs_quota.h"
  40#include "xfs_cksum.h"
  41#include "xfs_trace.h"
  42#include "xfs_icache.h"
  43#include "xfs_bmap_btree.h"
  44#include "xfs_error.h"
  45#include "xfs_dir2.h"
  46#include "xfs_rmap_item.h"
  47#include "xfs_buf_item.h"
  48#include "xfs_refcount_item.h"
  49#include "xfs_bmap_item.h"
  50
  51#define BLK_AVG(blk1, blk2)	((blk1+blk2) >> 1)
  52
  53STATIC int
  54xlog_find_zeroed(
  55	struct xlog	*,
  56	xfs_daddr_t	*);
  57STATIC int
  58xlog_clear_stale_blocks(
  59	struct xlog	*,
  60	xfs_lsn_t);
  61#if defined(DEBUG)
  62STATIC void
  63xlog_recover_check_summary(
  64	struct xlog *);
  65#else
  66#define	xlog_recover_check_summary(log)
  67#endif
  68STATIC int
  69xlog_do_recovery_pass(
  70        struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
  71
  72/*
  73 * This structure is used during recovery to record the buf log items which
  74 * have been canceled and should not be replayed.
  75 */
  76struct xfs_buf_cancel {
  77	xfs_daddr_t		bc_blkno;
  78	uint			bc_len;
  79	int			bc_refcount;
  80	struct list_head	bc_list;
  81};
  82
  83/*
  84 * Sector aligned buffer routines for buffer create/read/write/access
  85 */
  86
  87/*
  88 * Verify the given count of basic blocks is valid number of blocks
  89 * to specify for an operation involving the given XFS log buffer.
  90 * Returns nonzero if the count is valid, 0 otherwise.
  91 */
  92
  93static inline int
  94xlog_buf_bbcount_valid(
  95	struct xlog	*log,
  96	int		bbcount)
  97{
  98	return bbcount > 0 && bbcount <= log->l_logBBsize;
  99}
 100
 101/*
 102 * Allocate a buffer to hold log data.  The buffer needs to be able
 103 * to map to a range of nbblks basic blocks at any valid (basic
 104 * block) offset within the log.
 105 */
 106STATIC xfs_buf_t *
 107xlog_get_bp(
 108	struct xlog	*log,
 109	int		nbblks)
 110{
 111	struct xfs_buf	*bp;
 112
 113	if (!xlog_buf_bbcount_valid(log, nbblks)) {
 114		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 115			nbblks);
 116		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 117		return NULL;
 118	}
 119
 120	/*
 121	 * We do log I/O in units of log sectors (a power-of-2
 122	 * multiple of the basic block size), so we round up the
 123	 * requested size to accommodate the basic blocks required
 124	 * for complete log sectors.
 125	 *
 126	 * In addition, the buffer may be used for a non-sector-
 127	 * aligned block offset, in which case an I/O of the
 128	 * requested size could extend beyond the end of the
 129	 * buffer.  If the requested size is only 1 basic block it
 130	 * will never straddle a sector boundary, so this won't be
 131	 * an issue.  Nor will this be a problem if the log I/O is
 132	 * done in basic blocks (sector size 1).  But otherwise we
 133	 * extend the buffer by one extra log sector to ensure
 134	 * there's space to accommodate this possibility.
 135	 */
 136	if (nbblks > 1 && log->l_sectBBsize > 1)
 137		nbblks += log->l_sectBBsize;
 138	nbblks = round_up(nbblks, log->l_sectBBsize);
 139
 140	bp = xfs_buf_get_uncached(log->l_mp->m_logdev_targp, nbblks, 0);
 141	if (bp)
 142		xfs_buf_unlock(bp);
 143	return bp;
 144}
 145
 146STATIC void
 147xlog_put_bp(
 148	xfs_buf_t	*bp)
 149{
 150	xfs_buf_free(bp);
 151}
 152
 153/*
 154 * Return the address of the start of the given block number's data
 155 * in a log buffer.  The buffer covers a log sector-aligned region.
 156 */
 157STATIC char *
 158xlog_align(
 159	struct xlog	*log,
 160	xfs_daddr_t	blk_no,
 161	int		nbblks,
 162	struct xfs_buf	*bp)
 163{
 164	xfs_daddr_t	offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
 165
 166	ASSERT(offset + nbblks <= bp->b_length);
 167	return bp->b_addr + BBTOB(offset);
 168}
 169
 170
 171/*
 172 * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
 173 */
 174STATIC int
 175xlog_bread_noalign(
 176	struct xlog	*log,
 177	xfs_daddr_t	blk_no,
 178	int		nbblks,
 179	struct xfs_buf	*bp)
 180{
 181	int		error;
 182
 183	if (!xlog_buf_bbcount_valid(log, nbblks)) {
 184		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 185			nbblks);
 186		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 187		return -EFSCORRUPTED;
 188	}
 189
 190	blk_no = round_down(blk_no, log->l_sectBBsize);
 191	nbblks = round_up(nbblks, log->l_sectBBsize);
 192
 193	ASSERT(nbblks > 0);
 194	ASSERT(nbblks <= bp->b_length);
 195
 196	XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
 197	bp->b_flags |= XBF_READ;
 198	bp->b_io_length = nbblks;
 199	bp->b_error = 0;
 200
 201	error = xfs_buf_submit_wait(bp);
 202	if (error && !XFS_FORCED_SHUTDOWN(log->l_mp))
 203		xfs_buf_ioerror_alert(bp, __func__);
 204	return error;
 205}
 206
 207STATIC int
 208xlog_bread(
 209	struct xlog	*log,
 210	xfs_daddr_t	blk_no,
 211	int		nbblks,
 212	struct xfs_buf	*bp,
 213	char		**offset)
 214{
 215	int		error;
 216
 217	error = xlog_bread_noalign(log, blk_no, nbblks, bp);
 218	if (error)
 219		return error;
 220
 221	*offset = xlog_align(log, blk_no, nbblks, bp);
 222	return 0;
 223}
 224
 225/*
 226 * Read at an offset into the buffer. Returns with the buffer in it's original
 227 * state regardless of the result of the read.
 228 */
 229STATIC int
 230xlog_bread_offset(
 231	struct xlog	*log,
 232	xfs_daddr_t	blk_no,		/* block to read from */
 233	int		nbblks,		/* blocks to read */
 234	struct xfs_buf	*bp,
 235	char		*offset)
 236{
 237	char		*orig_offset = bp->b_addr;
 238	int		orig_len = BBTOB(bp->b_length);
 239	int		error, error2;
 240
 241	error = xfs_buf_associate_memory(bp, offset, BBTOB(nbblks));
 242	if (error)
 243		return error;
 244
 245	error = xlog_bread_noalign(log, blk_no, nbblks, bp);
 246
 247	/* must reset buffer pointer even on error */
 248	error2 = xfs_buf_associate_memory(bp, orig_offset, orig_len);
 249	if (error)
 250		return error;
 251	return error2;
 252}
 253
 254/*
 255 * Write out the buffer at the given block for the given number of blocks.
 256 * The buffer is kept locked across the write and is returned locked.
 257 * This can only be used for synchronous log writes.
 258 */
 259STATIC int
 260xlog_bwrite(
 261	struct xlog	*log,
 262	xfs_daddr_t	blk_no,
 263	int		nbblks,
 264	struct xfs_buf	*bp)
 265{
 266	int		error;
 267
 268	if (!xlog_buf_bbcount_valid(log, nbblks)) {
 269		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 270			nbblks);
 271		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 272		return -EFSCORRUPTED;
 273	}
 274
 275	blk_no = round_down(blk_no, log->l_sectBBsize);
 276	nbblks = round_up(nbblks, log->l_sectBBsize);
 277
 278	ASSERT(nbblks > 0);
 279	ASSERT(nbblks <= bp->b_length);
 280
 281	XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
 282	xfs_buf_hold(bp);
 283	xfs_buf_lock(bp);
 284	bp->b_io_length = nbblks;
 285	bp->b_error = 0;
 286
 287	error = xfs_bwrite(bp);
 288	if (error)
 289		xfs_buf_ioerror_alert(bp, __func__);
 290	xfs_buf_relse(bp);
 291	return error;
 292}
 293
 294#ifdef DEBUG
 295/*
 296 * dump debug superblock and log record information
 297 */
 298STATIC void
 299xlog_header_check_dump(
 300	xfs_mount_t		*mp,
 301	xlog_rec_header_t	*head)
 302{
 303	xfs_debug(mp, "%s:  SB : uuid = %pU, fmt = %d",
 304		__func__, &mp->m_sb.sb_uuid, XLOG_FMT);
 305	xfs_debug(mp, "    log : uuid = %pU, fmt = %d",
 306		&head->h_fs_uuid, be32_to_cpu(head->h_fmt));
 307}
 308#else
 309#define xlog_header_check_dump(mp, head)
 310#endif
 311
 312/*
 313 * check log record header for recovery
 314 */
 315STATIC int
 316xlog_header_check_recover(
 317	xfs_mount_t		*mp,
 318	xlog_rec_header_t	*head)
 319{
 320	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 321
 322	/*
 323	 * IRIX doesn't write the h_fmt field and leaves it zeroed
 324	 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
 325	 * a dirty log created in IRIX.
 326	 */
 327	if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
 328		xfs_warn(mp,
 329	"dirty log written in incompatible format - can't recover");
 330		xlog_header_check_dump(mp, head);
 331		XFS_ERROR_REPORT("xlog_header_check_recover(1)",
 332				 XFS_ERRLEVEL_HIGH, mp);
 333		return -EFSCORRUPTED;
 334	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 335		xfs_warn(mp,
 336	"dirty log entry has mismatched uuid - can't recover");
 337		xlog_header_check_dump(mp, head);
 338		XFS_ERROR_REPORT("xlog_header_check_recover(2)",
 339				 XFS_ERRLEVEL_HIGH, mp);
 340		return -EFSCORRUPTED;
 341	}
 342	return 0;
 343}
 344
 345/*
 346 * read the head block of the log and check the header
 347 */
 348STATIC int
 349xlog_header_check_mount(
 350	xfs_mount_t		*mp,
 351	xlog_rec_header_t	*head)
 352{
 353	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 354
 355	if (uuid_is_null(&head->h_fs_uuid)) {
 356		/*
 357		 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
 358		 * h_fs_uuid is null, we assume this log was last mounted
 359		 * by IRIX and continue.
 360		 */
 361		xfs_warn(mp, "null uuid in log - IRIX style log");
 362	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 363		xfs_warn(mp, "log has mismatched uuid - can't recover");
 364		xlog_header_check_dump(mp, head);
 365		XFS_ERROR_REPORT("xlog_header_check_mount",
 366				 XFS_ERRLEVEL_HIGH, mp);
 367		return -EFSCORRUPTED;
 368	}
 369	return 0;
 370}
 371
 372STATIC void
 373xlog_recover_iodone(
 374	struct xfs_buf	*bp)
 375{
 376	if (bp->b_error) {
 377		/*
 378		 * We're not going to bother about retrying
 379		 * this during recovery. One strike!
 380		 */
 381		if (!XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
 382			xfs_buf_ioerror_alert(bp, __func__);
 383			xfs_force_shutdown(bp->b_target->bt_mount,
 384						SHUTDOWN_META_IO_ERROR);
 385		}
 386	}
 387
 388	/*
 389	 * On v5 supers, a bli could be attached to update the metadata LSN.
 390	 * Clean it up.
 391	 */
 392	if (bp->b_fspriv)
 393		xfs_buf_item_relse(bp);
 394	ASSERT(bp->b_fspriv == NULL);
 395
 396	bp->b_iodone = NULL;
 397	xfs_buf_ioend(bp);
 398}
 399
 400/*
 401 * This routine finds (to an approximation) the first block in the physical
 402 * log which contains the given cycle.  It uses a binary search algorithm.
 403 * Note that the algorithm can not be perfect because the disk will not
 404 * necessarily be perfect.
 405 */
 406STATIC int
 407xlog_find_cycle_start(
 408	struct xlog	*log,
 409	struct xfs_buf	*bp,
 410	xfs_daddr_t	first_blk,
 411	xfs_daddr_t	*last_blk,
 412	uint		cycle)
 413{
 414	char		*offset;
 415	xfs_daddr_t	mid_blk;
 416	xfs_daddr_t	end_blk;
 417	uint		mid_cycle;
 418	int		error;
 419
 420	end_blk = *last_blk;
 421	mid_blk = BLK_AVG(first_blk, end_blk);
 422	while (mid_blk != first_blk && mid_blk != end_blk) {
 423		error = xlog_bread(log, mid_blk, 1, bp, &offset);
 424		if (error)
 425			return error;
 426		mid_cycle = xlog_get_cycle(offset);
 427		if (mid_cycle == cycle)
 428			end_blk = mid_blk;   /* last_half_cycle == mid_cycle */
 429		else
 430			first_blk = mid_blk; /* first_half_cycle == mid_cycle */
 431		mid_blk = BLK_AVG(first_blk, end_blk);
 432	}
 433	ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
 434	       (mid_blk == end_blk && mid_blk-1 == first_blk));
 435
 436	*last_blk = end_blk;
 437
 438	return 0;
 439}
 440
 441/*
 442 * Check that a range of blocks does not contain stop_on_cycle_no.
 443 * Fill in *new_blk with the block offset where such a block is
 444 * found, or with -1 (an invalid block number) if there is no such
 445 * block in the range.  The scan needs to occur from front to back
 446 * and the pointer into the region must be updated since a later
 447 * routine will need to perform another test.
 448 */
 449STATIC int
 450xlog_find_verify_cycle(
 451	struct xlog	*log,
 452	xfs_daddr_t	start_blk,
 453	int		nbblks,
 454	uint		stop_on_cycle_no,
 455	xfs_daddr_t	*new_blk)
 456{
 457	xfs_daddr_t	i, j;
 458	uint		cycle;
 459	xfs_buf_t	*bp;
 460	xfs_daddr_t	bufblks;
 461	char		*buf = NULL;
 462	int		error = 0;
 463
 464	/*
 465	 * Greedily allocate a buffer big enough to handle the full
 466	 * range of basic blocks we'll be examining.  If that fails,
 467	 * try a smaller size.  We need to be able to read at least
 468	 * a log sector, or we're out of luck.
 469	 */
 470	bufblks = 1 << ffs(nbblks);
 471	while (bufblks > log->l_logBBsize)
 472		bufblks >>= 1;
 473	while (!(bp = xlog_get_bp(log, bufblks))) {
 474		bufblks >>= 1;
 475		if (bufblks < log->l_sectBBsize)
 476			return -ENOMEM;
 477	}
 478
 479	for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
 480		int	bcount;
 481
 482		bcount = min(bufblks, (start_blk + nbblks - i));
 483
 484		error = xlog_bread(log, i, bcount, bp, &buf);
 485		if (error)
 486			goto out;
 487
 488		for (j = 0; j < bcount; j++) {
 489			cycle = xlog_get_cycle(buf);
 490			if (cycle == stop_on_cycle_no) {
 491				*new_blk = i+j;
 492				goto out;
 493			}
 494
 495			buf += BBSIZE;
 496		}
 497	}
 498
 499	*new_blk = -1;
 500
 501out:
 502	xlog_put_bp(bp);
 503	return error;
 504}
 505
 506/*
 507 * Potentially backup over partial log record write.
 508 *
 509 * In the typical case, last_blk is the number of the block directly after
 510 * a good log record.  Therefore, we subtract one to get the block number
 511 * of the last block in the given buffer.  extra_bblks contains the number
 512 * of blocks we would have read on a previous read.  This happens when the
 513 * last log record is split over the end of the physical log.
 514 *
 515 * extra_bblks is the number of blocks potentially verified on a previous
 516 * call to this routine.
 517 */
 518STATIC int
 519xlog_find_verify_log_record(
 520	struct xlog		*log,
 521	xfs_daddr_t		start_blk,
 522	xfs_daddr_t		*last_blk,
 523	int			extra_bblks)
 524{
 525	xfs_daddr_t		i;
 526	xfs_buf_t		*bp;
 527	char			*offset = NULL;
 528	xlog_rec_header_t	*head = NULL;
 529	int			error = 0;
 530	int			smallmem = 0;
 531	int			num_blks = *last_blk - start_blk;
 532	int			xhdrs;
 533
 534	ASSERT(start_blk != 0 || *last_blk != start_blk);
 535
 536	if (!(bp = xlog_get_bp(log, num_blks))) {
 537		if (!(bp = xlog_get_bp(log, 1)))
 538			return -ENOMEM;
 539		smallmem = 1;
 540	} else {
 541		error = xlog_bread(log, start_blk, num_blks, bp, &offset);
 542		if (error)
 543			goto out;
 544		offset += ((num_blks - 1) << BBSHIFT);
 545	}
 546
 547	for (i = (*last_blk) - 1; i >= 0; i--) {
 548		if (i < start_blk) {
 549			/* valid log record not found */
 550			xfs_warn(log->l_mp,
 551		"Log inconsistent (didn't find previous header)");
 552			ASSERT(0);
 553			error = -EIO;
 554			goto out;
 555		}
 556
 557		if (smallmem) {
 558			error = xlog_bread(log, i, 1, bp, &offset);
 559			if (error)
 560				goto out;
 561		}
 562
 563		head = (xlog_rec_header_t *)offset;
 564
 565		if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
 566			break;
 567
 568		if (!smallmem)
 569			offset -= BBSIZE;
 570	}
 571
 572	/*
 573	 * We hit the beginning of the physical log & still no header.  Return
 574	 * to caller.  If caller can handle a return of -1, then this routine
 575	 * will be called again for the end of the physical log.
 576	 */
 577	if (i == -1) {
 578		error = 1;
 579		goto out;
 580	}
 581
 582	/*
 583	 * We have the final block of the good log (the first block
 584	 * of the log record _before_ the head. So we check the uuid.
 585	 */
 586	if ((error = xlog_header_check_mount(log->l_mp, head)))
 587		goto out;
 588
 589	/*
 590	 * We may have found a log record header before we expected one.
 591	 * last_blk will be the 1st block # with a given cycle #.  We may end
 592	 * up reading an entire log record.  In this case, we don't want to
 593	 * reset last_blk.  Only when last_blk points in the middle of a log
 594	 * record do we update last_blk.
 595	 */
 596	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
 597		uint	h_size = be32_to_cpu(head->h_size);
 598
 599		xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
 600		if (h_size % XLOG_HEADER_CYCLE_SIZE)
 601			xhdrs++;
 602	} else {
 603		xhdrs = 1;
 604	}
 605
 606	if (*last_blk - i + extra_bblks !=
 607	    BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
 608		*last_blk = i;
 609
 610out:
 611	xlog_put_bp(bp);
 612	return error;
 613}
 614
 615/*
 616 * Head is defined to be the point of the log where the next log write
 617 * could go.  This means that incomplete LR writes at the end are
 618 * eliminated when calculating the head.  We aren't guaranteed that previous
 619 * LR have complete transactions.  We only know that a cycle number of
 620 * current cycle number -1 won't be present in the log if we start writing
 621 * from our current block number.
 622 *
 623 * last_blk contains the block number of the first block with a given
 624 * cycle number.
 625 *
 626 * Return: zero if normal, non-zero if error.
 627 */
 628STATIC int
 629xlog_find_head(
 630	struct xlog	*log,
 631	xfs_daddr_t	*return_head_blk)
 632{
 633	xfs_buf_t	*bp;
 634	char		*offset;
 635	xfs_daddr_t	new_blk, first_blk, start_blk, last_blk, head_blk;
 636	int		num_scan_bblks;
 637	uint		first_half_cycle, last_half_cycle;
 638	uint		stop_on_cycle;
 639	int		error, log_bbnum = log->l_logBBsize;
 640
 641	/* Is the end of the log device zeroed? */
 642	error = xlog_find_zeroed(log, &first_blk);
 643	if (error < 0) {
 644		xfs_warn(log->l_mp, "empty log check failed");
 645		return error;
 646	}
 647	if (error == 1) {
 648		*return_head_blk = first_blk;
 649
 650		/* Is the whole lot zeroed? */
 651		if (!first_blk) {
 652			/* Linux XFS shouldn't generate totally zeroed logs -
 653			 * mkfs etc write a dummy unmount record to a fresh
 654			 * log so we can store the uuid in there
 655			 */
 656			xfs_warn(log->l_mp, "totally zeroed log");
 657		}
 658
 659		return 0;
 660	}
 661
 662	first_blk = 0;			/* get cycle # of 1st block */
 663	bp = xlog_get_bp(log, 1);
 664	if (!bp)
 665		return -ENOMEM;
 666
 667	error = xlog_bread(log, 0, 1, bp, &offset);
 668	if (error)
 669		goto bp_err;
 670
 671	first_half_cycle = xlog_get_cycle(offset);
 672
 673	last_blk = head_blk = log_bbnum - 1;	/* get cycle # of last block */
 674	error = xlog_bread(log, last_blk, 1, bp, &offset);
 675	if (error)
 676		goto bp_err;
 677
 678	last_half_cycle = xlog_get_cycle(offset);
 679	ASSERT(last_half_cycle != 0);
 680
 681	/*
 682	 * If the 1st half cycle number is equal to the last half cycle number,
 683	 * then the entire log is stamped with the same cycle number.  In this
 684	 * case, head_blk can't be set to zero (which makes sense).  The below
 685	 * math doesn't work out properly with head_blk equal to zero.  Instead,
 686	 * we set it to log_bbnum which is an invalid block number, but this
 687	 * value makes the math correct.  If head_blk doesn't changed through
 688	 * all the tests below, *head_blk is set to zero at the very end rather
 689	 * than log_bbnum.  In a sense, log_bbnum and zero are the same block
 690	 * in a circular file.
 691	 */
 692	if (first_half_cycle == last_half_cycle) {
 693		/*
 694		 * In this case we believe that the entire log should have
 695		 * cycle number last_half_cycle.  We need to scan backwards
 696		 * from the end verifying that there are no holes still
 697		 * containing last_half_cycle - 1.  If we find such a hole,
 698		 * then the start of that hole will be the new head.  The
 699		 * simple case looks like
 700		 *        x | x ... | x - 1 | x
 701		 * Another case that fits this picture would be
 702		 *        x | x + 1 | x ... | x
 703		 * In this case the head really is somewhere at the end of the
 704		 * log, as one of the latest writes at the beginning was
 705		 * incomplete.
 706		 * One more case is
 707		 *        x | x + 1 | x ... | x - 1 | x
 708		 * This is really the combination of the above two cases, and
 709		 * the head has to end up at the start of the x-1 hole at the
 710		 * end of the log.
 711		 *
 712		 * In the 256k log case, we will read from the beginning to the
 713		 * end of the log and search for cycle numbers equal to x-1.
 714		 * We don't worry about the x+1 blocks that we encounter,
 715		 * because we know that they cannot be the head since the log
 716		 * started with x.
 717		 */
 718		head_blk = log_bbnum;
 719		stop_on_cycle = last_half_cycle - 1;
 720	} else {
 721		/*
 722		 * In this case we want to find the first block with cycle
 723		 * number matching last_half_cycle.  We expect the log to be
 724		 * some variation on
 725		 *        x + 1 ... | x ... | x
 726		 * The first block with cycle number x (last_half_cycle) will
 727		 * be where the new head belongs.  First we do a binary search
 728		 * for the first occurrence of last_half_cycle.  The binary
 729		 * search may not be totally accurate, so then we scan back
 730		 * from there looking for occurrences of last_half_cycle before
 731		 * us.  If that backwards scan wraps around the beginning of
 732		 * the log, then we look for occurrences of last_half_cycle - 1
 733		 * at the end of the log.  The cases we're looking for look
 734		 * like
 735		 *                               v binary search stopped here
 736		 *        x + 1 ... | x | x + 1 | x ... | x
 737		 *                   ^ but we want to locate this spot
 738		 * or
 739		 *        <---------> less than scan distance
 740		 *        x + 1 ... | x ... | x - 1 | x
 741		 *                           ^ we want to locate this spot
 742		 */
 743		stop_on_cycle = last_half_cycle;
 744		if ((error = xlog_find_cycle_start(log, bp, first_blk,
 745						&head_blk, last_half_cycle)))
 746			goto bp_err;
 747	}
 748
 749	/*
 750	 * Now validate the answer.  Scan back some number of maximum possible
 751	 * blocks and make sure each one has the expected cycle number.  The
 752	 * maximum is determined by the total possible amount of buffering
 753	 * in the in-core log.  The following number can be made tighter if
 754	 * we actually look at the block size of the filesystem.
 755	 */
 756	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
 757	if (head_blk >= num_scan_bblks) {
 758		/*
 759		 * We are guaranteed that the entire check can be performed
 760		 * in one buffer.
 761		 */
 762		start_blk = head_blk - num_scan_bblks;
 763		if ((error = xlog_find_verify_cycle(log,
 764						start_blk, num_scan_bblks,
 765						stop_on_cycle, &new_blk)))
 766			goto bp_err;
 767		if (new_blk != -1)
 768			head_blk = new_blk;
 769	} else {		/* need to read 2 parts of log */
 770		/*
 771		 * We are going to scan backwards in the log in two parts.
 772		 * First we scan the physical end of the log.  In this part
 773		 * of the log, we are looking for blocks with cycle number
 774		 * last_half_cycle - 1.
 775		 * If we find one, then we know that the log starts there, as
 776		 * we've found a hole that didn't get written in going around
 777		 * the end of the physical log.  The simple case for this is
 778		 *        x + 1 ... | x ... | x - 1 | x
 779		 *        <---------> less than scan distance
 780		 * If all of the blocks at the end of the log have cycle number
 781		 * last_half_cycle, then we check the blocks at the start of
 782		 * the log looking for occurrences of last_half_cycle.  If we
 783		 * find one, then our current estimate for the location of the
 784		 * first occurrence of last_half_cycle is wrong and we move
 785		 * back to the hole we've found.  This case looks like
 786		 *        x + 1 ... | x | x + 1 | x ...
 787		 *                               ^ binary search stopped here
 788		 * Another case we need to handle that only occurs in 256k
 789		 * logs is
 790		 *        x + 1 ... | x ... | x+1 | x ...
 791		 *                   ^ binary search stops here
 792		 * In a 256k log, the scan at the end of the log will see the
 793		 * x + 1 blocks.  We need to skip past those since that is
 794		 * certainly not the head of the log.  By searching for
 795		 * last_half_cycle-1 we accomplish that.
 796		 */
 797		ASSERT(head_blk <= INT_MAX &&
 798			(xfs_daddr_t) num_scan_bblks >= head_blk);
 799		start_blk = log_bbnum - (num_scan_bblks - head_blk);
 800		if ((error = xlog_find_verify_cycle(log, start_blk,
 801					num_scan_bblks - (int)head_blk,
 802					(stop_on_cycle - 1), &new_blk)))
 803			goto bp_err;
 804		if (new_blk != -1) {
 805			head_blk = new_blk;
 806			goto validate_head;
 807		}
 808
 809		/*
 810		 * Scan beginning of log now.  The last part of the physical
 811		 * log is good.  This scan needs to verify that it doesn't find
 812		 * the last_half_cycle.
 813		 */
 814		start_blk = 0;
 815		ASSERT(head_blk <= INT_MAX);
 816		if ((error = xlog_find_verify_cycle(log,
 817					start_blk, (int)head_blk,
 818					stop_on_cycle, &new_blk)))
 819			goto bp_err;
 820		if (new_blk != -1)
 821			head_blk = new_blk;
 822	}
 823
 824validate_head:
 825	/*
 826	 * Now we need to make sure head_blk is not pointing to a block in
 827	 * the middle of a log record.
 828	 */
 829	num_scan_bblks = XLOG_REC_SHIFT(log);
 830	if (head_blk >= num_scan_bblks) {
 831		start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
 832
 833		/* start ptr at last block ptr before head_blk */
 834		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
 835		if (error == 1)
 836			error = -EIO;
 837		if (error)
 838			goto bp_err;
 839	} else {
 840		start_blk = 0;
 841		ASSERT(head_blk <= INT_MAX);
 842		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
 843		if (error < 0)
 844			goto bp_err;
 845		if (error == 1) {
 846			/* We hit the beginning of the log during our search */
 847			start_blk = log_bbnum - (num_scan_bblks - head_blk);
 848			new_blk = log_bbnum;
 849			ASSERT(start_blk <= INT_MAX &&
 850				(xfs_daddr_t) log_bbnum-start_blk >= 0);
 851			ASSERT(head_blk <= INT_MAX);
 852			error = xlog_find_verify_log_record(log, start_blk,
 853							&new_blk, (int)head_blk);
 854			if (error == 1)
 855				error = -EIO;
 856			if (error)
 857				goto bp_err;
 858			if (new_blk != log_bbnum)
 859				head_blk = new_blk;
 860		} else if (error)
 861			goto bp_err;
 862	}
 863
 864	xlog_put_bp(bp);
 865	if (head_blk == log_bbnum)
 866		*return_head_blk = 0;
 867	else
 868		*return_head_blk = head_blk;
 869	/*
 870	 * When returning here, we have a good block number.  Bad block
 871	 * means that during a previous crash, we didn't have a clean break
 872	 * from cycle number N to cycle number N-1.  In this case, we need
 873	 * to find the first block with cycle number N-1.
 874	 */
 875	return 0;
 876
 877 bp_err:
 878	xlog_put_bp(bp);
 879
 880	if (error)
 881		xfs_warn(log->l_mp, "failed to find log head");
 882	return error;
 883}
 884
 885/*
 886 * Seek backwards in the log for log record headers.
 887 *
 888 * Given a starting log block, walk backwards until we find the provided number
 889 * of records or hit the provided tail block. The return value is the number of
 890 * records encountered or a negative error code. The log block and buffer
 891 * pointer of the last record seen are returned in rblk and rhead respectively.
 892 */
 893STATIC int
 894xlog_rseek_logrec_hdr(
 895	struct xlog		*log,
 896	xfs_daddr_t		head_blk,
 897	xfs_daddr_t		tail_blk,
 898	int			count,
 899	struct xfs_buf		*bp,
 900	xfs_daddr_t		*rblk,
 901	struct xlog_rec_header	**rhead,
 902	bool			*wrapped)
 903{
 904	int			i;
 905	int			error;
 906	int			found = 0;
 907	char			*offset = NULL;
 908	xfs_daddr_t		end_blk;
 909
 910	*wrapped = false;
 911
 912	/*
 913	 * Walk backwards from the head block until we hit the tail or the first
 914	 * block in the log.
 915	 */
 916	end_blk = head_blk > tail_blk ? tail_blk : 0;
 917	for (i = (int) head_blk - 1; i >= end_blk; i--) {
 918		error = xlog_bread(log, i, 1, bp, &offset);
 919		if (error)
 920			goto out_error;
 921
 922		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 923			*rblk = i;
 924			*rhead = (struct xlog_rec_header *) offset;
 925			if (++found == count)
 926				break;
 927		}
 928	}
 929
 930	/*
 931	 * If we haven't hit the tail block or the log record header count,
 932	 * start looking again from the end of the physical log. Note that
 933	 * callers can pass head == tail if the tail is not yet known.
 934	 */
 935	if (tail_blk >= head_blk && found != count) {
 936		for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
 937			error = xlog_bread(log, i, 1, bp, &offset);
 938			if (error)
 939				goto out_error;
 940
 941			if (*(__be32 *)offset ==
 942			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 943				*wrapped = true;
 944				*rblk = i;
 945				*rhead = (struct xlog_rec_header *) offset;
 946				if (++found == count)
 947					break;
 948			}
 949		}
 950	}
 951
 952	return found;
 953
 954out_error:
 955	return error;
 956}
 957
 958/*
 959 * Seek forward in the log for log record headers.
 960 *
 961 * Given head and tail blocks, walk forward from the tail block until we find
 962 * the provided number of records or hit the head block. The return value is the
 963 * number of records encountered or a negative error code. The log block and
 964 * buffer pointer of the last record seen are returned in rblk and rhead
 965 * respectively.
 966 */
 967STATIC int
 968xlog_seek_logrec_hdr(
 969	struct xlog		*log,
 970	xfs_daddr_t		head_blk,
 971	xfs_daddr_t		tail_blk,
 972	int			count,
 973	struct xfs_buf		*bp,
 974	xfs_daddr_t		*rblk,
 975	struct xlog_rec_header	**rhead,
 976	bool			*wrapped)
 977{
 978	int			i;
 979	int			error;
 980	int			found = 0;
 981	char			*offset = NULL;
 982	xfs_daddr_t		end_blk;
 983
 984	*wrapped = false;
 985
 986	/*
 987	 * Walk forward from the tail block until we hit the head or the last
 988	 * block in the log.
 989	 */
 990	end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
 991	for (i = (int) tail_blk; i <= end_blk; i++) {
 992		error = xlog_bread(log, i, 1, bp, &offset);
 993		if (error)
 994			goto out_error;
 995
 996		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 997			*rblk = i;
 998			*rhead = (struct xlog_rec_header *) offset;
 999			if (++found == count)
1000				break;
1001		}
1002	}
1003
1004	/*
1005	 * If we haven't hit the head block or the log record header count,
1006	 * start looking again from the start of the physical log.
1007	 */
1008	if (tail_blk > head_blk && found != count) {
1009		for (i = 0; i < (int) head_blk; i++) {
1010			error = xlog_bread(log, i, 1, bp, &offset);
1011			if (error)
1012				goto out_error;
1013
1014			if (*(__be32 *)offset ==
1015			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
1016				*wrapped = true;
1017				*rblk = i;
1018				*rhead = (struct xlog_rec_header *) offset;
1019				if (++found == count)
1020					break;
1021			}
1022		}
1023	}
1024
1025	return found;
1026
1027out_error:
1028	return error;
1029}
1030
1031/*
1032 * Check the log tail for torn writes. This is required when torn writes are
1033 * detected at the head and the head had to be walked back to a previous record.
1034 * The tail of the previous record must now be verified to ensure the torn
1035 * writes didn't corrupt the previous tail.
1036 *
1037 * Return an error if CRC verification fails as recovery cannot proceed.
1038 */
1039STATIC int
1040xlog_verify_tail(
1041	struct xlog		*log,
1042	xfs_daddr_t		head_blk,
1043	xfs_daddr_t		tail_blk)
1044{
1045	struct xlog_rec_header	*thead;
1046	struct xfs_buf		*bp;
1047	xfs_daddr_t		first_bad;
1048	int			count;
1049	int			error = 0;
1050	bool			wrapped;
1051	xfs_daddr_t		tmp_head;
1052
1053	bp = xlog_get_bp(log, 1);
1054	if (!bp)
1055		return -ENOMEM;
1056
1057	/*
1058	 * Seek XLOG_MAX_ICLOGS + 1 records past the current tail record to get
1059	 * a temporary head block that points after the last possible
1060	 * concurrently written record of the tail.
1061	 */
1062	count = xlog_seek_logrec_hdr(log, head_blk, tail_blk,
1063				     XLOG_MAX_ICLOGS + 1, bp, &tmp_head, &thead,
1064				     &wrapped);
1065	if (count < 0) {
1066		error = count;
1067		goto out;
1068	}
1069
1070	/*
1071	 * If the call above didn't find XLOG_MAX_ICLOGS + 1 records, we ran
1072	 * into the actual log head. tmp_head points to the start of the record
1073	 * so update it to the actual head block.
1074	 */
1075	if (count < XLOG_MAX_ICLOGS + 1)
1076		tmp_head = head_blk;
1077
1078	/*
1079	 * We now have a tail and temporary head block that covers at least
1080	 * XLOG_MAX_ICLOGS records from the tail. We need to verify that these
1081	 * records were completely written. Run a CRC verification pass from
1082	 * tail to head and return the result.
1083	 */
1084	error = xlog_do_recovery_pass(log, tmp_head, tail_blk,
1085				      XLOG_RECOVER_CRCPASS, &first_bad);
1086
1087out:
1088	xlog_put_bp(bp);
1089	return error;
1090}
1091
1092/*
1093 * Detect and trim torn writes from the head of the log.
1094 *
1095 * Storage without sector atomicity guarantees can result in torn writes in the
1096 * log in the event of a crash. Our only means to detect this scenario is via
1097 * CRC verification. While we can't always be certain that CRC verification
1098 * failure is due to a torn write vs. an unrelated corruption, we do know that
1099 * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1100 * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1101 * the log and treat failures in this range as torn writes as a matter of
1102 * policy. In the event of CRC failure, the head is walked back to the last good
1103 * record in the log and the tail is updated from that record and verified.
1104 */
1105STATIC int
1106xlog_verify_head(
1107	struct xlog		*log,
1108	xfs_daddr_t		*head_blk,	/* in/out: unverified head */
1109	xfs_daddr_t		*tail_blk,	/* out: tail block */
1110	struct xfs_buf		*bp,
1111	xfs_daddr_t		*rhead_blk,	/* start blk of last record */
1112	struct xlog_rec_header	**rhead,	/* ptr to last record */
1113	bool			*wrapped)	/* last rec. wraps phys. log */
1114{
1115	struct xlog_rec_header	*tmp_rhead;
1116	struct xfs_buf		*tmp_bp;
1117	xfs_daddr_t		first_bad;
1118	xfs_daddr_t		tmp_rhead_blk;
1119	int			found;
1120	int			error;
1121	bool			tmp_wrapped;
1122
1123	/*
1124	 * Check the head of the log for torn writes. Search backwards from the
1125	 * head until we hit the tail or the maximum number of log record I/Os
1126	 * that could have been in flight at one time. Use a temporary buffer so
1127	 * we don't trash the rhead/bp pointers from the caller.
1128	 */
1129	tmp_bp = xlog_get_bp(log, 1);
1130	if (!tmp_bp)
1131		return -ENOMEM;
1132	error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1133				      XLOG_MAX_ICLOGS, tmp_bp, &tmp_rhead_blk,
1134				      &tmp_rhead, &tmp_wrapped);
1135	xlog_put_bp(tmp_bp);
1136	if (error < 0)
1137		return error;
1138
1139	/*
1140	 * Now run a CRC verification pass over the records starting at the
1141	 * block found above to the current head. If a CRC failure occurs, the
1142	 * log block of the first bad record is saved in first_bad.
1143	 */
1144	error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1145				      XLOG_RECOVER_CRCPASS, &first_bad);
1146	if (error == -EFSBADCRC) {
1147		/*
1148		 * We've hit a potential torn write. Reset the error and warn
1149		 * about it.
1150		 */
1151		error = 0;
1152		xfs_warn(log->l_mp,
1153"Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1154			 first_bad, *head_blk);
1155
1156		/*
1157		 * Get the header block and buffer pointer for the last good
1158		 * record before the bad record.
1159		 *
1160		 * Note that xlog_find_tail() clears the blocks at the new head
1161		 * (i.e., the records with invalid CRC) if the cycle number
1162		 * matches the the current cycle.
1163		 */
1164		found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1, bp,
1165					      rhead_blk, rhead, wrapped);
1166		if (found < 0)
1167			return found;
1168		if (found == 0)		/* XXX: right thing to do here? */
1169			return -EIO;
1170
1171		/*
1172		 * Reset the head block to the starting block of the first bad
1173		 * log record and set the tail block based on the last good
1174		 * record.
1175		 *
1176		 * Bail out if the updated head/tail match as this indicates
1177		 * possible corruption outside of the acceptable
1178		 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1179		 */
1180		*head_blk = first_bad;
1181		*tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1182		if (*head_blk == *tail_blk) {
1183			ASSERT(0);
1184			return 0;
1185		}
1186
1187		/*
1188		 * Now verify the tail based on the updated head. This is
1189		 * required because the torn writes trimmed from the head could
1190		 * have been written over the tail of a previous record. Return
1191		 * any errors since recovery cannot proceed if the tail is
1192		 * corrupt.
1193		 *
1194		 * XXX: This leaves a gap in truly robust protection from torn
1195		 * writes in the log. If the head is behind the tail, the tail
1196		 * pushes forward to create some space and then a crash occurs
1197		 * causing the writes into the previous record's tail region to
1198		 * tear, log recovery isn't able to recover.
1199		 *
1200		 * How likely is this to occur? If possible, can we do something
1201		 * more intelligent here? Is it safe to push the tail forward if
1202		 * we can determine that the tail is within the range of the
1203		 * torn write (e.g., the kernel can only overwrite the tail if
1204		 * it has actually been pushed forward)? Alternatively, could we
1205		 * somehow prevent this condition at runtime?
1206		 */
1207		error = xlog_verify_tail(log, *head_blk, *tail_blk);
1208	}
1209
1210	return error;
1211}
1212
1213/*
1214 * Check whether the head of the log points to an unmount record. In other
1215 * words, determine whether the log is clean. If so, update the in-core state
1216 * appropriately.
1217 */
1218static int
1219xlog_check_unmount_rec(
1220	struct xlog		*log,
1221	xfs_daddr_t		*head_blk,
1222	xfs_daddr_t		*tail_blk,
1223	struct xlog_rec_header	*rhead,
1224	xfs_daddr_t		rhead_blk,
1225	struct xfs_buf		*bp,
1226	bool			*clean)
1227{
1228	struct xlog_op_header	*op_head;
1229	xfs_daddr_t		umount_data_blk;
1230	xfs_daddr_t		after_umount_blk;
1231	int			hblks;
1232	int			error;
1233	char			*offset;
1234
1235	*clean = false;
1236
1237	/*
1238	 * Look for unmount record. If we find it, then we know there was a
1239	 * clean unmount. Since 'i' could be the last block in the physical
1240	 * log, we convert to a log block before comparing to the head_blk.
1241	 *
1242	 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1243	 * below. We won't want to clear the unmount record if there is one, so
1244	 * we pass the lsn of the unmount record rather than the block after it.
1245	 */
1246	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1247		int	h_size = be32_to_cpu(rhead->h_size);
1248		int	h_version = be32_to_cpu(rhead->h_version);
1249
1250		if ((h_version & XLOG_VERSION_2) &&
1251		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1252			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1253			if (h_size % XLOG_HEADER_CYCLE_SIZE)
1254				hblks++;
1255		} else {
1256			hblks = 1;
1257		}
1258	} else {
1259		hblks = 1;
1260	}
1261	after_umount_blk = rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len));
1262	after_umount_blk = do_mod(after_umount_blk, log->l_logBBsize);
1263	if (*head_blk == after_umount_blk &&
1264	    be32_to_cpu(rhead->h_num_logops) == 1) {
1265		umount_data_blk = rhead_blk + hblks;
1266		umount_data_blk = do_mod(umount_data_blk, log->l_logBBsize);
1267		error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
1268		if (error)
1269			return error;
1270
1271		op_head = (struct xlog_op_header *)offset;
1272		if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1273			/*
1274			 * Set tail and last sync so that newly written log
1275			 * records will point recovery to after the current
1276			 * unmount record.
1277			 */
1278			xlog_assign_atomic_lsn(&log->l_tail_lsn,
1279					log->l_curr_cycle, after_umount_blk);
1280			xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1281					log->l_curr_cycle, after_umount_blk);
1282			*tail_blk = after_umount_blk;
1283
1284			*clean = true;
1285		}
1286	}
1287
1288	return 0;
1289}
1290
1291static void
1292xlog_set_state(
1293	struct xlog		*log,
1294	xfs_daddr_t		head_blk,
1295	struct xlog_rec_header	*rhead,
1296	xfs_daddr_t		rhead_blk,
1297	bool			bump_cycle)
1298{
1299	/*
1300	 * Reset log values according to the state of the log when we
1301	 * crashed.  In the case where head_blk == 0, we bump curr_cycle
1302	 * one because the next write starts a new cycle rather than
1303	 * continuing the cycle of the last good log record.  At this
1304	 * point we have guaranteed that all partial log records have been
1305	 * accounted for.  Therefore, we know that the last good log record
1306	 * written was complete and ended exactly on the end boundary
1307	 * of the physical log.
1308	 */
1309	log->l_prev_block = rhead_blk;
1310	log->l_curr_block = (int)head_blk;
1311	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1312	if (bump_cycle)
1313		log->l_curr_cycle++;
1314	atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1315	atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1316	xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1317					BBTOB(log->l_curr_block));
1318	xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1319					BBTOB(log->l_curr_block));
1320}
1321
1322/*
1323 * Find the sync block number or the tail of the log.
1324 *
1325 * This will be the block number of the last record to have its
1326 * associated buffers synced to disk.  Every log record header has
1327 * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
1328 * to get a sync block number.  The only concern is to figure out which
1329 * log record header to believe.
1330 *
1331 * The following algorithm uses the log record header with the largest
1332 * lsn.  The entire log record does not need to be valid.  We only care
1333 * that the header is valid.
1334 *
1335 * We could speed up search by using current head_blk buffer, but it is not
1336 * available.
1337 */
1338STATIC int
1339xlog_find_tail(
1340	struct xlog		*log,
1341	xfs_daddr_t		*head_blk,
1342	xfs_daddr_t		*tail_blk)
1343{
1344	xlog_rec_header_t	*rhead;
1345	char			*offset = NULL;
1346	xfs_buf_t		*bp;
1347	int			error;
1348	xfs_daddr_t		rhead_blk;
1349	xfs_lsn_t		tail_lsn;
1350	bool			wrapped = false;
1351	bool			clean = false;
1352
1353	/*
1354	 * Find previous log record
1355	 */
1356	if ((error = xlog_find_head(log, head_blk)))
1357		return error;
1358	ASSERT(*head_blk < INT_MAX);
1359
1360	bp = xlog_get_bp(log, 1);
1361	if (!bp)
1362		return -ENOMEM;
1363	if (*head_blk == 0) {				/* special case */
1364		error = xlog_bread(log, 0, 1, bp, &offset);
1365		if (error)
1366			goto done;
1367
1368		if (xlog_get_cycle(offset) == 0) {
1369			*tail_blk = 0;
1370			/* leave all other log inited values alone */
1371			goto done;
1372		}
1373	}
1374
1375	/*
1376	 * Search backwards through the log looking for the log record header
1377	 * block. This wraps all the way back around to the head so something is
1378	 * seriously wrong if we can't find it.
1379	 */
1380	error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, bp,
1381				      &rhead_blk, &rhead, &wrapped);
1382	if (error < 0)
1383		return error;
1384	if (!error) {
1385		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1386		return -EIO;
1387	}
1388	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1389
1390	/*
1391	 * Set the log state based on the current head record.
1392	 */
1393	xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1394	tail_lsn = atomic64_read(&log->l_tail_lsn);
1395
1396	/*
1397	 * Look for an unmount record at the head of the log. This sets the log
1398	 * state to determine whether recovery is necessary.
1399	 */
1400	error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1401				       rhead_blk, bp, &clean);
1402	if (error)
1403		goto done;
1404
1405	/*
1406	 * Verify the log head if the log is not clean (e.g., we have anything
1407	 * but an unmount record at the head). This uses CRC verification to
1408	 * detect and trim torn writes. If discovered, CRC failures are
1409	 * considered torn writes and the log head is trimmed accordingly.
1410	 *
1411	 * Note that we can only run CRC verification when the log is dirty
1412	 * because there's no guarantee that the log data behind an unmount
1413	 * record is compatible with the current architecture.
1414	 */
1415	if (!clean) {
1416		xfs_daddr_t	orig_head = *head_blk;
1417
1418		error = xlog_verify_head(log, head_blk, tail_blk, bp,
1419					 &rhead_blk, &rhead, &wrapped);
1420		if (error)
1421			goto done;
1422
1423		/* update in-core state again if the head changed */
1424		if (*head_blk != orig_head) {
1425			xlog_set_state(log, *head_blk, rhead, rhead_blk,
1426				       wrapped);
1427			tail_lsn = atomic64_read(&log->l_tail_lsn);
1428			error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1429						       rhead, rhead_blk, bp,
1430						       &clean);
1431			if (error)
1432				goto done;
1433		}
1434	}
1435
1436	/*
1437	 * Note that the unmount was clean. If the unmount was not clean, we
1438	 * need to know this to rebuild the superblock counters from the perag
1439	 * headers if we have a filesystem using non-persistent counters.
1440	 */
1441	if (clean)
1442		log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1443
1444	/*
1445	 * Make sure that there are no blocks in front of the head
1446	 * with the same cycle number as the head.  This can happen
1447	 * because we allow multiple outstanding log writes concurrently,
1448	 * and the later writes might make it out before earlier ones.
1449	 *
1450	 * We use the lsn from before modifying it so that we'll never
1451	 * overwrite the unmount record after a clean unmount.
1452	 *
1453	 * Do this only if we are going to recover the filesystem
1454	 *
1455	 * NOTE: This used to say "if (!readonly)"
1456	 * However on Linux, we can & do recover a read-only filesystem.
1457	 * We only skip recovery if NORECOVERY is specified on mount,
1458	 * in which case we would not be here.
1459	 *
1460	 * But... if the -device- itself is readonly, just skip this.
1461	 * We can't recover this device anyway, so it won't matter.
1462	 */
1463	if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
1464		error = xlog_clear_stale_blocks(log, tail_lsn);
1465
1466done:
1467	xlog_put_bp(bp);
1468
1469	if (error)
1470		xfs_warn(log->l_mp, "failed to locate log tail");
1471	return error;
1472}
1473
1474/*
1475 * Is the log zeroed at all?
1476 *
1477 * The last binary search should be changed to perform an X block read
1478 * once X becomes small enough.  You can then search linearly through
1479 * the X blocks.  This will cut down on the number of reads we need to do.
1480 *
1481 * If the log is partially zeroed, this routine will pass back the blkno
1482 * of the first block with cycle number 0.  It won't have a complete LR
1483 * preceding it.
1484 *
1485 * Return:
1486 *	0  => the log is completely written to
1487 *	1 => use *blk_no as the first block of the log
1488 *	<0 => error has occurred
1489 */
1490STATIC int
1491xlog_find_zeroed(
1492	struct xlog	*log,
1493	xfs_daddr_t	*blk_no)
1494{
1495	xfs_buf_t	*bp;
1496	char		*offset;
1497	uint	        first_cycle, last_cycle;
1498	xfs_daddr_t	new_blk, last_blk, start_blk;
1499	xfs_daddr_t     num_scan_bblks;
1500	int	        error, log_bbnum = log->l_logBBsize;
1501
1502	*blk_no = 0;
1503
1504	/* check totally zeroed log */
1505	bp = xlog_get_bp(log, 1);
1506	if (!bp)
1507		return -ENOMEM;
1508	error = xlog_bread(log, 0, 1, bp, &offset);
1509	if (error)
1510		goto bp_err;
1511
1512	first_cycle = xlog_get_cycle(offset);
1513	if (first_cycle == 0) {		/* completely zeroed log */
1514		*blk_no = 0;
1515		xlog_put_bp(bp);
1516		return 1;
1517	}
1518
1519	/* check partially zeroed log */
1520	error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1521	if (error)
1522		goto bp_err;
1523
1524	last_cycle = xlog_get_cycle(offset);
1525	if (last_cycle != 0) {		/* log completely written to */
1526		xlog_put_bp(bp);
1527		return 0;
1528	} else if (first_cycle != 1) {
1529		/*
1530		 * If the cycle of the last block is zero, the cycle of
1531		 * the first block must be 1. If it's not, maybe we're
1532		 * not looking at a log... Bail out.
1533		 */
1534		xfs_warn(log->l_mp,
1535			"Log inconsistent or not a log (last==0, first!=1)");
1536		error = -EINVAL;
1537		goto bp_err;
1538	}
1539
1540	/* we have a partially zeroed log */
1541	last_blk = log_bbnum-1;
1542	if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1543		goto bp_err;
1544
1545	/*
1546	 * Validate the answer.  Because there is no way to guarantee that
1547	 * the entire log is made up of log records which are the same size,
1548	 * we scan over the defined maximum blocks.  At this point, the maximum
1549	 * is not chosen to mean anything special.   XXXmiken
1550	 */
1551	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1552	ASSERT(num_scan_bblks <= INT_MAX);
1553
1554	if (last_blk < num_scan_bblks)
1555		num_scan_bblks = last_blk;
1556	start_blk = last_blk - num_scan_bblks;
1557
1558	/*
1559	 * We search for any instances of cycle number 0 that occur before
1560	 * our current estimate of the head.  What we're trying to detect is
1561	 *        1 ... | 0 | 1 | 0...
1562	 *                       ^ binary search ends here
1563	 */
1564	if ((error = xlog_find_verify_cycle(log, start_blk,
1565					 (int)num_scan_bblks, 0, &new_blk)))
1566		goto bp_err;
1567	if (new_blk != -1)
1568		last_blk = new_blk;
1569
1570	/*
1571	 * Potentially backup over partial log record write.  We don't need
1572	 * to search the end of the log because we know it is zero.
1573	 */
1574	error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1575	if (error == 1)
1576		error = -EIO;
1577	if (error)
1578		goto bp_err;
1579
1580	*blk_no = last_blk;
1581bp_err:
1582	xlog_put_bp(bp);
1583	if (error)
1584		return error;
1585	return 1;
1586}
1587
1588/*
1589 * These are simple subroutines used by xlog_clear_stale_blocks() below
1590 * to initialize a buffer full of empty log record headers and write
1591 * them into the log.
1592 */
1593STATIC void
1594xlog_add_record(
1595	struct xlog		*log,
1596	char			*buf,
1597	int			cycle,
1598	int			block,
1599	int			tail_cycle,
1600	int			tail_block)
1601{
1602	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1603
1604	memset(buf, 0, BBSIZE);
1605	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1606	recp->h_cycle = cpu_to_be32(cycle);
1607	recp->h_version = cpu_to_be32(
1608			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1609	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1610	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1611	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1612	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1613}
1614
1615STATIC int
1616xlog_write_log_records(
1617	struct xlog	*log,
1618	int		cycle,
1619	int		start_block,
1620	int		blocks,
1621	int		tail_cycle,
1622	int		tail_block)
1623{
1624	char		*offset;
1625	xfs_buf_t	*bp;
1626	int		balign, ealign;
1627	int		sectbb = log->l_sectBBsize;
1628	int		end_block = start_block + blocks;
1629	int		bufblks;
1630	int		error = 0;
1631	int		i, j = 0;
1632
1633	/*
1634	 * Greedily allocate a buffer big enough to handle the full
1635	 * range of basic blocks to be written.  If that fails, try
1636	 * a smaller size.  We need to be able to write at least a
1637	 * log sector, or we're out of luck.
1638	 */
1639	bufblks = 1 << ffs(blocks);
1640	while (bufblks > log->l_logBBsize)
1641		bufblks >>= 1;
1642	while (!(bp = xlog_get_bp(log, bufblks))) {
1643		bufblks >>= 1;
1644		if (bufblks < sectbb)
1645			return -ENOMEM;
1646	}
1647
1648	/* We may need to do a read at the start to fill in part of
1649	 * the buffer in the starting sector not covered by the first
1650	 * write below.
1651	 */
1652	balign = round_down(start_block, sectbb);
1653	if (balign != start_block) {
1654		error = xlog_bread_noalign(log, start_block, 1, bp);
1655		if (error)
1656			goto out_put_bp;
1657
1658		j = start_block - balign;
1659	}
1660
1661	for (i = start_block; i < end_block; i += bufblks) {
1662		int		bcount, endcount;
1663
1664		bcount = min(bufblks, end_block - start_block);
1665		endcount = bcount - j;
1666
1667		/* We may need to do a read at the end to fill in part of
1668		 * the buffer in the final sector not covered by the write.
1669		 * If this is the same sector as the above read, skip it.
1670		 */
1671		ealign = round_down(end_block, sectbb);
1672		if (j == 0 && (start_block + endcount > ealign)) {
1673			offset = bp->b_addr + BBTOB(ealign - start_block);
1674			error = xlog_bread_offset(log, ealign, sectbb,
1675							bp, offset);
1676			if (error)
1677				break;
1678
1679		}
1680
1681		offset = xlog_align(log, start_block, endcount, bp);
1682		for (; j < endcount; j++) {
1683			xlog_add_record(log, offset, cycle, i+j,
1684					tail_cycle, tail_block);
1685			offset += BBSIZE;
1686		}
1687		error = xlog_bwrite(log, start_block, endcount, bp);
1688		if (error)
1689			break;
1690		start_block += endcount;
1691		j = 0;
1692	}
1693
1694 out_put_bp:
1695	xlog_put_bp(bp);
1696	return error;
1697}
1698
1699/*
1700 * This routine is called to blow away any incomplete log writes out
1701 * in front of the log head.  We do this so that we won't become confused
1702 * if we come up, write only a little bit more, and then crash again.
1703 * If we leave the partial log records out there, this situation could
1704 * cause us to think those partial writes are valid blocks since they
1705 * have the current cycle number.  We get rid of them by overwriting them
1706 * with empty log records with the old cycle number rather than the
1707 * current one.
1708 *
1709 * The tail lsn is passed in rather than taken from
1710 * the log so that we will not write over the unmount record after a
1711 * clean unmount in a 512 block log.  Doing so would leave the log without
1712 * any valid log records in it until a new one was written.  If we crashed
1713 * during that time we would not be able to recover.
1714 */
1715STATIC int
1716xlog_clear_stale_blocks(
1717	struct xlog	*log,
1718	xfs_lsn_t	tail_lsn)
1719{
1720	int		tail_cycle, head_cycle;
1721	int		tail_block, head_block;
1722	int		tail_distance, max_distance;
1723	int		distance;
1724	int		error;
1725
1726	tail_cycle = CYCLE_LSN(tail_lsn);
1727	tail_block = BLOCK_LSN(tail_lsn);
1728	head_cycle = log->l_curr_cycle;
1729	head_block = log->l_curr_block;
1730
1731	/*
1732	 * Figure out the distance between the new head of the log
1733	 * and the tail.  We want to write over any blocks beyond the
1734	 * head that we may have written just before the crash, but
1735	 * we don't want to overwrite the tail of the log.
1736	 */
1737	if (head_cycle == tail_cycle) {
1738		/*
1739		 * The tail is behind the head in the physical log,
1740		 * so the distance from the head to the tail is the
1741		 * distance from the head to the end of the log plus
1742		 * the distance from the beginning of the log to the
1743		 * tail.
1744		 */
1745		if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1746			XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1747					 XFS_ERRLEVEL_LOW, log->l_mp);
1748			return -EFSCORRUPTED;
1749		}
1750		tail_distance = tail_block + (log->l_logBBsize - head_block);
1751	} else {
1752		/*
1753		 * The head is behind the tail in the physical log,
1754		 * so the distance from the head to the tail is just
1755		 * the tail block minus the head block.
1756		 */
1757		if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1758			XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1759					 XFS_ERRLEVEL_LOW, log->l_mp);
1760			return -EFSCORRUPTED;
1761		}
1762		tail_distance = tail_block - head_block;
1763	}
1764
1765	/*
1766	 * If the head is right up against the tail, we can't clear
1767	 * anything.
1768	 */
1769	if (tail_distance <= 0) {
1770		ASSERT(tail_distance == 0);
1771		return 0;
1772	}
1773
1774	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1775	/*
1776	 * Take the smaller of the maximum amount of outstanding I/O
1777	 * we could have and the distance to the tail to clear out.
1778	 * We take the smaller so that we don't overwrite the tail and
1779	 * we don't waste all day writing from the head to the tail
1780	 * for no reason.
1781	 */
1782	max_distance = MIN(max_distance, tail_distance);
1783
1784	if ((head_block + max_distance) <= log->l_logBBsize) {
1785		/*
1786		 * We can stomp all the blocks we need to without
1787		 * wrapping around the end of the log.  Just do it
1788		 * in a single write.  Use the cycle number of the
1789		 * current cycle minus one so that the log will look like:
1790		 *     n ... | n - 1 ...
1791		 */
1792		error = xlog_write_log_records(log, (head_cycle - 1),
1793				head_block, max_distance, tail_cycle,
1794				tail_block);
1795		if (error)
1796			return error;
1797	} else {
1798		/*
1799		 * We need to wrap around the end of the physical log in
1800		 * order to clear all the blocks.  Do it in two separate
1801		 * I/Os.  The first write should be from the head to the
1802		 * end of the physical log, and it should use the current
1803		 * cycle number minus one just like above.
1804		 */
1805		distance = log->l_logBBsize - head_block;
1806		error = xlog_write_log_records(log, (head_cycle - 1),
1807				head_block, distance, tail_cycle,
1808				tail_block);
1809
1810		if (error)
1811			return error;
1812
1813		/*
1814		 * Now write the blocks at the start of the physical log.
1815		 * This writes the remainder of the blocks we want to clear.
1816		 * It uses the current cycle number since we're now on the
1817		 * same cycle as the head so that we get:
1818		 *    n ... n ... | n - 1 ...
1819		 *    ^^^^^ blocks we're writing
1820		 */
1821		distance = max_distance - (log->l_logBBsize - head_block);
1822		error = xlog_write_log_records(log, head_cycle, 0, distance,
1823				tail_cycle, tail_block);
1824		if (error)
1825			return error;
1826	}
1827
1828	return 0;
1829}
1830
1831/******************************************************************************
1832 *
1833 *		Log recover routines
1834 *
1835 ******************************************************************************
1836 */
1837
1838/*
1839 * Sort the log items in the transaction.
1840 *
1841 * The ordering constraints are defined by the inode allocation and unlink
1842 * behaviour. The rules are:
1843 *
1844 *	1. Every item is only logged once in a given transaction. Hence it
1845 *	   represents the last logged state of the item. Hence ordering is
1846 *	   dependent on the order in which operations need to be performed so
1847 *	   required initial conditions are always met.
1848 *
1849 *	2. Cancelled buffers are recorded in pass 1 in a separate table and
1850 *	   there's nothing to replay from them so we can simply cull them
1851 *	   from the transaction. However, we can't do that until after we've
1852 *	   replayed all the other items because they may be dependent on the
1853 *	   cancelled buffer and replaying the cancelled buffer can remove it
1854 *	   form the cancelled buffer table. Hence they have tobe done last.
1855 *
1856 *	3. Inode allocation buffers must be replayed before inode items that
1857 *	   read the buffer and replay changes into it. For filesystems using the
1858 *	   ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1859 *	   treated the same as inode allocation buffers as they create and
1860 *	   initialise the buffers directly.
1861 *
1862 *	4. Inode unlink buffers must be replayed after inode items are replayed.
1863 *	   This ensures that inodes are completely flushed to the inode buffer
1864 *	   in a "free" state before we remove the unlinked inode list pointer.
1865 *
1866 * Hence the ordering needs to be inode allocation buffers first, inode items
1867 * second, inode unlink buffers third and cancelled buffers last.
1868 *
1869 * But there's a problem with that - we can't tell an inode allocation buffer
1870 * apart from a regular buffer, so we can't separate them. We can, however,
1871 * tell an inode unlink buffer from the others, and so we can separate them out
1872 * from all the other buffers and move them to last.
1873 *
1874 * Hence, 4 lists, in order from head to tail:
1875 *	- buffer_list for all buffers except cancelled/inode unlink buffers
1876 *	- item_list for all non-buffer items
1877 *	- inode_buffer_list for inode unlink buffers
1878 *	- cancel_list for the cancelled buffers
1879 *
1880 * Note that we add objects to the tail of the lists so that first-to-last
1881 * ordering is preserved within the lists. Adding objects to the head of the
1882 * list means when we traverse from the head we walk them in last-to-first
1883 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1884 * but for all other items there may be specific ordering that we need to
1885 * preserve.
1886 */
1887STATIC int
1888xlog_recover_reorder_trans(
1889	struct xlog		*log,
1890	struct xlog_recover	*trans,
1891	int			pass)
1892{
1893	xlog_recover_item_t	*item, *n;
1894	int			error = 0;
1895	LIST_HEAD(sort_list);
1896	LIST_HEAD(cancel_list);
1897	LIST_HEAD(buffer_list);
1898	LIST_HEAD(inode_buffer_list);
1899	LIST_HEAD(inode_list);
1900
1901	list_splice_init(&trans->r_itemq, &sort_list);
1902	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1903		xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
1904
1905		switch (ITEM_TYPE(item)) {
1906		case XFS_LI_ICREATE:
1907			list_move_tail(&item->ri_list, &buffer_list);
1908			break;
1909		case XFS_LI_BUF:
1910			if (buf_f->blf_flags & XFS_BLF_CANCEL) {
1911				trace_xfs_log_recover_item_reorder_head(log,
1912							trans, item, pass);
1913				list_move(&item->ri_list, &cancel_list);
1914				break;
1915			}
1916			if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
1917				list_move(&item->ri_list, &inode_buffer_list);
1918				break;
1919			}
1920			list_move_tail(&item->ri_list, &buffer_list);
1921			break;
1922		case XFS_LI_INODE:
1923		case XFS_LI_DQUOT:
1924		case XFS_LI_QUOTAOFF:
1925		case XFS_LI_EFD:
1926		case XFS_LI_EFI:
1927		case XFS_LI_RUI:
1928		case XFS_LI_RUD:
1929		case XFS_LI_CUI:
1930		case XFS_LI_CUD:
1931		case XFS_LI_BUI:
1932		case XFS_LI_BUD:
1933			trace_xfs_log_recover_item_reorder_tail(log,
1934							trans, item, pass);
1935			list_move_tail(&item->ri_list, &inode_list);
1936			break;
1937		default:
1938			xfs_warn(log->l_mp,
1939				"%s: unrecognized type of log operation",
1940				__func__);
1941			ASSERT(0);
1942			/*
1943			 * return the remaining items back to the transaction
1944			 * item list so they can be freed in caller.
1945			 */
1946			if (!list_empty(&sort_list))
1947				list_splice_init(&sort_list, &trans->r_itemq);
1948			error = -EIO;
1949			goto out;
1950		}
1951	}
1952out:
1953	ASSERT(list_empty(&sort_list));
1954	if (!list_empty(&buffer_list))
1955		list_splice(&buffer_list, &trans->r_itemq);
1956	if (!list_empty(&inode_list))
1957		list_splice_tail(&inode_list, &trans->r_itemq);
1958	if (!list_empty(&inode_buffer_list))
1959		list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1960	if (!list_empty(&cancel_list))
1961		list_splice_tail(&cancel_list, &trans->r_itemq);
1962	return error;
1963}
1964
1965/*
1966 * Build up the table of buf cancel records so that we don't replay
1967 * cancelled data in the second pass.  For buffer records that are
1968 * not cancel records, there is nothing to do here so we just return.
1969 *
1970 * If we get a cancel record which is already in the table, this indicates
1971 * that the buffer was cancelled multiple times.  In order to ensure
1972 * that during pass 2 we keep the record in the table until we reach its
1973 * last occurrence in the log, we keep a reference count in the cancel
1974 * record in the table to tell us how many times we expect to see this
1975 * record during the second pass.
1976 */
1977STATIC int
1978xlog_recover_buffer_pass1(
1979	struct xlog			*log,
1980	struct xlog_recover_item	*item)
1981{
1982	xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
1983	struct list_head	*bucket;
1984	struct xfs_buf_cancel	*bcp;
1985
1986	/*
1987	 * If this isn't a cancel buffer item, then just return.
1988	 */
1989	if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1990		trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1991		return 0;
1992	}
1993
1994	/*
1995	 * Insert an xfs_buf_cancel record into the hash table of them.
1996	 * If there is already an identical record, bump its reference count.
1997	 */
1998	bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1999	list_for_each_entry(bcp, bucket, bc_list) {
2000		if (bcp->bc_blkno == buf_f->blf_blkno &&
2001		    bcp->bc_len == buf_f->blf_len) {
2002			bcp->bc_refcount++;
2003			trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
2004			return 0;
2005		}
2006	}
2007
2008	bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
2009	bcp->bc_blkno = buf_f->blf_blkno;
2010	bcp->bc_len = buf_f->blf_len;
2011	bcp->bc_refcount = 1;
2012	list_add_tail(&bcp->bc_list, bucket);
2013
2014	trace_xfs_log_recover_buf_cancel_add(log, buf_f);
2015	return 0;
2016}
2017
2018/*
2019 * Check to see whether the buffer being recovered has a corresponding
2020 * entry in the buffer cancel record table. If it is, return the cancel
2021 * buffer structure to the caller.
2022 */
2023STATIC struct xfs_buf_cancel *
2024xlog_peek_buffer_cancelled(
2025	struct xlog		*log,
2026	xfs_daddr_t		blkno,
2027	uint			len,
2028	unsigned short			flags)
2029{
2030	struct list_head	*bucket;
2031	struct xfs_buf_cancel	*bcp;
2032
2033	if (!log->l_buf_cancel_table) {
2034		/* empty table means no cancelled buffers in the log */
2035		ASSERT(!(flags & XFS_BLF_CANCEL));
2036		return NULL;
2037	}
2038
2039	bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
2040	list_for_each_entry(bcp, bucket, bc_list) {
2041		if (bcp->bc_blkno == blkno && bcp->bc_len == len)
2042			return bcp;
2043	}
2044
2045	/*
2046	 * We didn't find a corresponding entry in the table, so return 0 so
2047	 * that the buffer is NOT cancelled.
2048	 */
2049	ASSERT(!(flags & XFS_BLF_CANCEL));
2050	return NULL;
2051}
2052
2053/*
2054 * If the buffer is being cancelled then return 1 so that it will be cancelled,
2055 * otherwise return 0.  If the buffer is actually a buffer cancel item
2056 * (XFS_BLF_CANCEL is set), then decrement the refcount on the entry in the
2057 * table and remove it from the table if this is the last reference.
2058 *
2059 * We remove the cancel record from the table when we encounter its last
2060 * occurrence in the log so that if the same buffer is re-used again after its
2061 * last cancellation we actually replay the changes made at that point.
2062 */
2063STATIC int
2064xlog_check_buffer_cancelled(
2065	struct xlog		*log,
2066	xfs_daddr_t		blkno,
2067	uint			len,
2068	unsigned short			flags)
2069{
2070	struct xfs_buf_cancel	*bcp;
2071
2072	bcp = xlog_peek_buffer_cancelled(log, blkno, len, flags);
2073	if (!bcp)
2074		return 0;
2075
2076	/*
2077	 * We've go a match, so return 1 so that the recovery of this buffer
2078	 * is cancelled.  If this buffer is actually a buffer cancel log
2079	 * item, then decrement the refcount on the one in the table and
2080	 * remove it if this is the last reference.
2081	 */
2082	if (flags & XFS_BLF_CANCEL) {
2083		if (--bcp->bc_refcount == 0) {
2084			list_del(&bcp->bc_list);
2085			kmem_free(bcp);
2086		}
2087	}
2088	return 1;
2089}
2090
2091/*
2092 * Perform recovery for a buffer full of inodes.  In these buffers, the only
2093 * data which should be recovered is that which corresponds to the
2094 * di_next_unlinked pointers in the on disk inode structures.  The rest of the
2095 * data for the inodes is always logged through the inodes themselves rather
2096 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
2097 *
2098 * The only time when buffers full of inodes are fully recovered is when the
2099 * buffer is full of newly allocated inodes.  In this case the buffer will
2100 * not be marked as an inode buffer and so will be sent to
2101 * xlog_recover_do_reg_buffer() below during recovery.
2102 */
2103STATIC int
2104xlog_recover_do_inode_buffer(
2105	struct xfs_mount	*mp,
2106	xlog_recover_item_t	*item,
2107	struct xfs_buf		*bp,
2108	xfs_buf_log_format_t	*buf_f)
2109{
2110	int			i;
2111	int			item_index = 0;
2112	int			bit = 0;
2113	int			nbits = 0;
2114	int			reg_buf_offset = 0;
2115	int			reg_buf_bytes = 0;
2116	int			next_unlinked_offset;
2117	int			inodes_per_buf;
2118	xfs_agino_t		*logged_nextp;
2119	xfs_agino_t		*buffer_nextp;
2120
2121	trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
2122
2123	/*
2124	 * Post recovery validation only works properly on CRC enabled
2125	 * filesystems.
2126	 */
2127	if (xfs_sb_version_hascrc(&mp->m_sb))
2128		bp->b_ops = &xfs_inode_buf_ops;
2129
2130	inodes_per_buf = BBTOB(bp->b_io_length) >> mp->m_sb.sb_inodelog;
2131	for (i = 0; i < inodes_per_buf; i++) {
2132		next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
2133			offsetof(xfs_dinode_t, di_next_unlinked);
2134
2135		while (next_unlinked_offset >=
2136		       (reg_buf_offset + reg_buf_bytes)) {
2137			/*
2138			 * The next di_next_unlinked field is beyond
2139			 * the current logged region.  Find the next
2140			 * logged region that contains or is beyond
2141			 * the current di_next_unlinked field.
2142			 */
2143			bit += nbits;
2144			bit = xfs_next_bit(buf_f->blf_data_map,
2145					   buf_f->blf_map_size, bit);
2146
2147			/*
2148			 * If there are no more logged regions in the
2149			 * buffer, then we're done.
2150			 */
2151			if (bit == -1)
2152				return 0;
2153
2154			nbits = xfs_contig_bits(buf_f->blf_data_map,
2155						buf_f->blf_map_size, bit);
2156			ASSERT(nbits > 0);
2157			reg_buf_offset = bit << XFS_BLF_SHIFT;
2158			reg_buf_bytes = nbits << XFS_BLF_SHIFT;
2159			item_index++;
2160		}
2161
2162		/*
2163		 * If the current logged region starts after the current
2164		 * di_next_unlinked field, then move on to the next
2165		 * di_next_unlinked field.
2166		 */
2167		if (next_unlinked_offset < reg_buf_offset)
2168			continue;
2169
2170		ASSERT(item->ri_buf[item_index].i_addr != NULL);
2171		ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
2172		ASSERT((reg_buf_offset + reg_buf_bytes) <=
2173							BBTOB(bp->b_io_length));
2174
2175		/*
2176		 * The current logged region contains a copy of the
2177		 * current di_next_unlinked field.  Extract its value
2178		 * and copy it to the buffer copy.
2179		 */
2180		logged_nextp = item->ri_buf[item_index].i_addr +
2181				next_unlinked_offset - reg_buf_offset;
2182		if (unlikely(*logged_nextp == 0)) {
2183			xfs_alert(mp,
2184		"Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
2185		"Trying to replay bad (0) inode di_next_unlinked field.",
2186				item, bp);
2187			XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
2188					 XFS_ERRLEVEL_LOW, mp);
2189			return -EFSCORRUPTED;
2190		}
2191
2192		buffer_nextp = xfs_buf_offset(bp, next_unlinked_offset);
2193		*buffer_nextp = *logged_nextp;
2194
2195		/*
2196		 * If necessary, recalculate the CRC in the on-disk inode. We
2197		 * have to leave the inode in a consistent state for whoever
2198		 * reads it next....
2199		 */
2200		xfs_dinode_calc_crc(mp,
2201				xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
2202
2203	}
2204
2205	return 0;
2206}
2207
2208/*
2209 * V5 filesystems know the age of the buffer on disk being recovered. We can
2210 * have newer objects on disk than we are replaying, and so for these cases we
2211 * don't want to replay the current change as that will make the buffer contents
2212 * temporarily invalid on disk.
2213 *
2214 * The magic number might not match the buffer type we are going to recover
2215 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags.  Hence
2216 * extract the LSN of the existing object in the buffer based on it's current
2217 * magic number.  If we don't recognise the magic number in the buffer, then
2218 * return a LSN of -1 so that the caller knows it was an unrecognised block and
2219 * so can recover the buffer.
2220 *
2221 * Note: we cannot rely solely on magic number matches to determine that the
2222 * buffer has a valid LSN - we also need to verify that it belongs to this
2223 * filesystem, so we need to extract the object's LSN and compare it to that
2224 * which we read from the superblock. If the UUIDs don't match, then we've got a
2225 * stale metadata block from an old filesystem instance that we need to recover
2226 * over the top of.
2227 */
2228static xfs_lsn_t
2229xlog_recover_get_buf_lsn(
2230	struct xfs_mount	*mp,
2231	struct xfs_buf		*bp)
2232{
2233	uint32_t		magic32;
2234	uint16_t		magic16;
2235	uint16_t		magicda;
2236	void			*blk = bp->b_addr;
2237	uuid_t			*uuid;
2238	xfs_lsn_t		lsn = -1;
2239
2240	/* v4 filesystems always recover immediately */
2241	if (!xfs_sb_version_hascrc(&mp->m_sb))
2242		goto recover_immediately;
2243
2244	magic32 = be32_to_cpu(*(__be32 *)blk);
2245	switch (magic32) {
2246	case XFS_ABTB_CRC_MAGIC:
2247	case XFS_ABTC_CRC_MAGIC:
2248	case XFS_ABTB_MAGIC:
2249	case XFS_ABTC_MAGIC:
2250	case XFS_RMAP_CRC_MAGIC:
2251	case XFS_REFC_CRC_MAGIC:
2252	case XFS_IBT_CRC_MAGIC:
2253	case XFS_IBT_MAGIC: {
2254		struct xfs_btree_block *btb = blk;
2255
2256		lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
2257		uuid = &btb->bb_u.s.bb_uuid;
2258		break;
2259	}
2260	case XFS_BMAP_CRC_MAGIC:
2261	case XFS_BMAP_MAGIC: {
2262		struct xfs_btree_block *btb = blk;
2263
2264		lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
2265		uuid = &btb->bb_u.l.bb_uuid;
2266		break;
2267	}
2268	case XFS_AGF_MAGIC:
2269		lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
2270		uuid = &((struct xfs_agf *)blk)->agf_uuid;
2271		break;
2272	case XFS_AGFL_MAGIC:
2273		lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
2274		uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
2275		break;
2276	case XFS_AGI_MAGIC:
2277		lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
2278		uuid = &((struct xfs_agi *)blk)->agi_uuid;
2279		break;
2280	case XFS_SYMLINK_MAGIC:
2281		lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
2282		uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
2283		break;
2284	case XFS_DIR3_BLOCK_MAGIC:
2285	case XFS_DIR3_DATA_MAGIC:
2286	case XFS_DIR3_FREE_MAGIC:
2287		lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
2288		uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
2289		break;
2290	case XFS_ATTR3_RMT_MAGIC:
2291		/*
2292		 * Remote attr blocks are written synchronously, rather than
2293		 * being logged. That means they do not contain a valid LSN
2294		 * (i.e. transactionally ordered) in them, and hence any time we
2295		 * see a buffer to replay over the top of a remote attribute
2296		 * block we should simply do so.
2297		 */
2298		goto recover_immediately;
2299	case XFS_SB_MAGIC:
2300		/*
2301		 * superblock uuids are magic. We may or may not have a
2302		 * sb_meta_uuid on disk, but it will be set in the in-core
2303		 * superblock. We set the uuid pointer for verification
2304		 * according to the superblock feature mask to ensure we check
2305		 * the relevant UUID in the superblock.
2306		 */
2307		lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
2308		if (xfs_sb_version_hasmetauuid(&mp->m_sb))
2309			uuid = &((struct xfs_dsb *)blk)->sb_meta_uuid;
2310		else
2311			uuid = &((struct xfs_dsb *)blk)->sb_uuid;
2312		break;
2313	default:
2314		break;
2315	}
2316
2317	if (lsn != (xfs_lsn_t)-1) {
2318		if (!uuid_equal(&mp->m_sb.sb_meta_uuid, uuid))
2319			goto recover_immediately;
2320		return lsn;
2321	}
2322
2323	magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
2324	switch (magicda) {
2325	case XFS_DIR3_LEAF1_MAGIC:
2326	case XFS_DIR3_LEAFN_MAGIC:
2327	case XFS_DA3_NODE_MAGIC:
2328		lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
2329		uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
2330		break;
2331	default:
2332		break;
2333	}
2334
2335	if (lsn != (xfs_lsn_t)-1) {
2336		if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
2337			goto recover_immediately;
2338		return lsn;
2339	}
2340
2341	/*
2342	 * We do individual object checks on dquot and inode buffers as they
2343	 * have their own individual LSN records. Also, we could have a stale
2344	 * buffer here, so we have to at least recognise these buffer types.
2345	 *
2346	 * A notd complexity here is inode unlinked list processing - it logs
2347	 * the inode directly in the buffer, but we don't know which inodes have
2348	 * been modified, and there is no global buffer LSN. Hence we need to
2349	 * recover all inode buffer types immediately. This problem will be
2350	 * fixed by logical logging of the unlinked list modifications.
2351	 */
2352	magic16 = be16_to_cpu(*(__be16 *)blk);
2353	switch (magic16) {
2354	case XFS_DQUOT_MAGIC:
2355	case XFS_DINODE_MAGIC:
2356		goto recover_immediately;
2357	default:
2358		break;
2359	}
2360
2361	/* unknown buffer contents, recover immediately */
2362
2363recover_immediately:
2364	return (xfs_lsn_t)-1;
2365
2366}
2367
2368/*
2369 * Validate the recovered buffer is of the correct type and attach the
2370 * appropriate buffer operations to them for writeback. Magic numbers are in a
2371 * few places:
2372 *	the first 16 bits of the buffer (inode buffer, dquot buffer),
2373 *	the first 32 bits of the buffer (most blocks),
2374 *	inside a struct xfs_da_blkinfo at the start of the buffer.
2375 */
2376static void
2377xlog_recover_validate_buf_type(
2378	struct xfs_mount	*mp,
2379	struct xfs_buf		*bp,
2380	xfs_buf_log_format_t	*buf_f,
2381	xfs_lsn_t		current_lsn)
2382{
2383	struct xfs_da_blkinfo	*info = bp->b_addr;
2384	uint32_t		magic32;
2385	uint16_t		magic16;
2386	uint16_t		magicda;
2387	char			*warnmsg = NULL;
2388
2389	/*
2390	 * We can only do post recovery validation on items on CRC enabled
2391	 * fielsystems as we need to know when the buffer was written to be able
2392	 * to determine if we should have replayed the item. If we replay old
2393	 * metadata over a newer buffer, then it will enter a temporarily
2394	 * inconsistent state resulting in verification failures. Hence for now
2395	 * just avoid the verification stage for non-crc filesystems
2396	 */
2397	if (!xfs_sb_version_hascrc(&mp->m_sb))
2398		return;
2399
2400	magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
2401	magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
2402	magicda = be16_to_cpu(info->magic);
2403	switch (xfs_blft_from_flags(buf_f)) {
2404	case XFS_BLFT_BTREE_BUF:
2405		switch (magic32) {
2406		case XFS_ABTB_CRC_MAGIC:
2407		case XFS_ABTC_CRC_MAGIC:
2408		case XFS_ABTB_MAGIC:
2409		case XFS_ABTC_MAGIC:
2410			bp->b_ops = &xfs_allocbt_buf_ops;
2411			break;
2412		case XFS_IBT_CRC_MAGIC:
2413		case XFS_FIBT_CRC_MAGIC:
2414		case XFS_IBT_MAGIC:
2415		case XFS_FIBT_MAGIC:
2416			bp->b_ops = &xfs_inobt_buf_ops;
2417			break;
2418		case XFS_BMAP_CRC_MAGIC:
2419		case XFS_BMAP_MAGIC:
2420			bp->b_ops = &xfs_bmbt_buf_ops;
2421			break;
2422		case XFS_RMAP_CRC_MAGIC:
2423			bp->b_ops = &xfs_rmapbt_buf_ops;
2424			break;
2425		case XFS_REFC_CRC_MAGIC:
2426			bp->b_ops = &xfs_refcountbt_buf_ops;
2427			break;
2428		default:
2429			warnmsg = "Bad btree block magic!";
2430			break;
2431		}
2432		break;
2433	case XFS_BLFT_AGF_BUF:
2434		if (magic32 != XFS_AGF_MAGIC) {
2435			warnmsg = "Bad AGF block magic!";
2436			break;
2437		}
2438		bp->b_ops = &xfs_agf_buf_ops;
2439		break;
2440	case XFS_BLFT_AGFL_BUF:
2441		if (magic32 != XFS_AGFL_MAGIC) {
2442			warnmsg = "Bad AGFL block magic!";
2443			break;
2444		}
2445		bp->b_ops = &xfs_agfl_buf_ops;
2446		break;
2447	case XFS_BLFT_AGI_BUF:
2448		if (magic32 != XFS_AGI_MAGIC) {
2449			warnmsg = "Bad AGI block magic!";
2450			break;
2451		}
2452		bp->b_ops = &xfs_agi_buf_ops;
2453		break;
2454	case XFS_BLFT_UDQUOT_BUF:
2455	case XFS_BLFT_PDQUOT_BUF:
2456	case XFS_BLFT_GDQUOT_BUF:
2457#ifdef CONFIG_XFS_QUOTA
2458		if (magic16 != XFS_DQUOT_MAGIC) {
2459			warnmsg = "Bad DQUOT block magic!";
2460			break;
2461		}
2462		bp->b_ops = &xfs_dquot_buf_ops;
2463#else
2464		xfs_alert(mp,
2465	"Trying to recover dquots without QUOTA support built in!");
2466		ASSERT(0);
2467#endif
2468		break;
2469	case XFS_BLFT_DINO_BUF:
2470		if (magic16 != XFS_DINODE_MAGIC) {
2471			warnmsg = "Bad INODE block magic!";
2472			break;
2473		}
2474		bp->b_ops = &xfs_inode_buf_ops;
2475		break;
2476	case XFS_BLFT_SYMLINK_BUF:
2477		if (magic32 != XFS_SYMLINK_MAGIC) {
2478			warnmsg = "Bad symlink block magic!";
2479			break;
2480		}
2481		bp->b_ops = &xfs_symlink_buf_ops;
2482		break;
2483	case XFS_BLFT_DIR_BLOCK_BUF:
2484		if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
2485		    magic32 != XFS_DIR3_BLOCK_MAGIC) {
2486			warnmsg = "Bad dir block magic!";
2487			break;
2488		}
2489		bp->b_ops = &xfs_dir3_block_buf_ops;
2490		break;
2491	case XFS_BLFT_DIR_DATA_BUF:
2492		if (magic32 != XFS_DIR2_DATA_MAGIC &&
2493		    magic32 != XFS_DIR3_DATA_MAGIC) {
2494			warnmsg = "Bad dir data magic!";
2495			break;
2496		}
2497		bp->b_ops = &xfs_dir3_data_buf_ops;
2498		break;
2499	case XFS_BLFT_DIR_FREE_BUF:
2500		if (magic32 != XFS_DIR2_FREE_MAGIC &&
2501		    magic32 != XFS_DIR3_FREE_MAGIC) {
2502			warnmsg = "Bad dir3 free magic!";
2503			break;
2504		}
2505		bp->b_ops = &xfs_dir3_free_buf_ops;
2506		break;
2507	case XFS_BLFT_DIR_LEAF1_BUF:
2508		if (magicda != XFS_DIR2_LEAF1_MAGIC &&
2509		    magicda != XFS_DIR3_LEAF1_MAGIC) {
2510			warnmsg = "Bad dir leaf1 magic!";
2511			break;
2512		}
2513		bp->b_ops = &xfs_dir3_leaf1_buf_ops;
2514		break;
2515	case XFS_BLFT_DIR_LEAFN_BUF:
2516		if (magicda != XFS_DIR2_LEAFN_MAGIC &&
2517		    magicda != XFS_DIR3_LEAFN_MAGIC) {
2518			warnmsg = "Bad dir leafn magic!";
2519			break;
2520		}
2521		bp->b_ops = &xfs_dir3_leafn_buf_ops;
2522		break;
2523	case XFS_BLFT_DA_NODE_BUF:
2524		if (magicda != XFS_DA_NODE_MAGIC &&
2525		    magicda != XFS_DA3_NODE_MAGIC) {
2526			warnmsg = "Bad da node magic!";
2527			break;
2528		}
2529		bp->b_ops = &xfs_da3_node_buf_ops;
2530		break;
2531	case XFS_BLFT_ATTR_LEAF_BUF:
2532		if (magicda != XFS_ATTR_LEAF_MAGIC &&
2533		    magicda != XFS_ATTR3_LEAF_MAGIC) {
2534			warnmsg = "Bad attr leaf magic!";
2535			break;
2536		}
2537		bp->b_ops = &xfs_attr3_leaf_buf_ops;
2538		break;
2539	case XFS_BLFT_ATTR_RMT_BUF:
2540		if (magic32 != XFS_ATTR3_RMT_MAGIC) {
2541			warnmsg = "Bad attr remote magic!";
2542			break;
2543		}
2544		bp->b_ops = &xfs_attr3_rmt_buf_ops;
2545		break;
2546	case XFS_BLFT_SB_BUF:
2547		if (magic32 != XFS_SB_MAGIC) {
2548			warnmsg = "Bad SB block magic!";
2549			break;
2550		}
2551		bp->b_ops = &xfs_sb_buf_ops;
2552		break;
2553#ifdef CONFIG_XFS_RT
2554	case XFS_BLFT_RTBITMAP_BUF:
2555	case XFS_BLFT_RTSUMMARY_BUF:
2556		/* no magic numbers for verification of RT buffers */
2557		bp->b_ops = &xfs_rtbuf_ops;
2558		break;
2559#endif /* CONFIG_XFS_RT */
2560	default:
2561		xfs_warn(mp, "Unknown buffer type %d!",
2562			 xfs_blft_from_flags(buf_f));
2563		break;
2564	}
2565
2566	/*
2567	 * Nothing else to do in the case of a NULL current LSN as this means
2568	 * the buffer is more recent than the change in the log and will be
2569	 * skipped.
2570	 */
2571	if (current_lsn == NULLCOMMITLSN)
2572		return;
2573
2574	if (warnmsg) {
2575		xfs_warn(mp, warnmsg);
2576		ASSERT(0);
2577	}
2578
2579	/*
2580	 * We must update the metadata LSN of the buffer as it is written out to
2581	 * ensure that older transactions never replay over this one and corrupt
2582	 * the buffer. This can occur if log recovery is interrupted at some
2583	 * point after the current transaction completes, at which point a
2584	 * subsequent mount starts recovery from the beginning.
2585	 *
2586	 * Write verifiers update the metadata LSN from log items attached to
2587	 * the buffer. Therefore, initialize a bli purely to carry the LSN to
2588	 * the verifier. We'll clean it up in our ->iodone() callback.
2589	 */
2590	if (bp->b_ops) {
2591		struct xfs_buf_log_item	*bip;
2592
2593		ASSERT(!bp->b_iodone || bp->b_iodone == xlog_recover_iodone);
2594		bp->b_iodone = xlog_recover_iodone;
2595		xfs_buf_item_init(bp, mp);
2596		bip = bp->b_fspriv;
2597		bip->bli_item.li_lsn = current_lsn;
2598	}
2599}
2600
2601/*
2602 * Perform a 'normal' buffer recovery.  Each logged region of the
2603 * buffer should be copied over the corresponding region in the
2604 * given buffer.  The bitmap in the buf log format structure indicates
2605 * where to place the logged data.
2606 */
2607STATIC void
2608xlog_recover_do_reg_buffer(
2609	struct xfs_mount	*mp,
2610	xlog_recover_item_t	*item,
2611	struct xfs_buf		*bp,
2612	xfs_buf_log_format_t	*buf_f,
2613	xfs_lsn_t		current_lsn)
2614{
2615	int			i;
2616	int			bit;
2617	int			nbits;
2618	int                     error;
2619
2620	trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
2621
2622	bit = 0;
2623	i = 1;  /* 0 is the buf format structure */
2624	while (1) {
2625		bit = xfs_next_bit(buf_f->blf_data_map,
2626				   buf_f->blf_map_size, bit);
2627		if (bit == -1)
2628			break;
2629		nbits = xfs_contig_bits(buf_f->blf_data_map,
2630					buf_f->blf_map_size, bit);
2631		ASSERT(nbits > 0);
2632		ASSERT(item->ri_buf[i].i_addr != NULL);
2633		ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
2634		ASSERT(BBTOB(bp->b_io_length) >=
2635		       ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
2636
2637		/*
2638		 * The dirty regions logged in the buffer, even though
2639		 * contiguous, may span multiple chunks. This is because the
2640		 * dirty region may span a physical page boundary in a buffer
2641		 * and hence be split into two separate vectors for writing into
2642		 * the log. Hence we need to trim nbits back to the length of
2643		 * the current region being copied out of the log.
2644		 */
2645		if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
2646			nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
2647
2648		/*
2649		 * Do a sanity check if this is a dquot buffer. Just checking
2650		 * the first dquot in the buffer should do. XXXThis is
2651		 * probably a good thing to do for other buf types also.
2652		 */
2653		error = 0;
2654		if (buf_f->blf_flags &
2655		   (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2656			if (item->ri_buf[i].i_addr == NULL) {
2657				xfs_alert(mp,
2658					"XFS: NULL dquot in %s.", __func__);
2659				goto next;
2660			}
2661			if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
2662				xfs_alert(mp,
2663					"XFS: dquot too small (%d) in %s.",
2664					item->ri_buf[i].i_len, __func__);
2665				goto next;
2666			}
2667			error = xfs_dqcheck(mp, item->ri_buf[i].i_addr,
2668					       -1, 0, XFS_QMOPT_DOWARN,
2669					       "dquot_buf_recover");
2670			if (error)
2671				goto next;
2672		}
2673
2674		memcpy(xfs_buf_offset(bp,
2675			(uint)bit << XFS_BLF_SHIFT),	/* dest */
2676			item->ri_buf[i].i_addr,		/* source */
2677			nbits<<XFS_BLF_SHIFT);		/* length */
2678 next:
2679		i++;
2680		bit += nbits;
2681	}
2682
2683	/* Shouldn't be any more regions */
2684	ASSERT(i == item->ri_total);
2685
2686	xlog_recover_validate_buf_type(mp, bp, buf_f, current_lsn);
2687}
2688
2689/*
2690 * Perform a dquot buffer recovery.
2691 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
2692 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2693 * Else, treat it as a regular buffer and do recovery.
2694 *
2695 * Return false if the buffer was tossed and true if we recovered the buffer to
2696 * indicate to the caller if the buffer needs writing.
2697 */
2698STATIC bool
2699xlog_recover_do_dquot_buffer(
2700	struct xfs_mount		*mp,
2701	struct xlog			*log,
2702	struct xlog_recover_item	*item,
2703	struct xfs_buf			*bp,
2704	struct xfs_buf_log_format	*buf_f)
2705{
2706	uint			type;
2707
2708	trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2709
2710	/*
2711	 * Filesystems are required to send in quota flags at mount time.
2712	 */
2713	if (!mp->m_qflags)
2714		return false;
2715
2716	type = 0;
2717	if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2718		type |= XFS_DQ_USER;
2719	if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2720		type |= XFS_DQ_PROJ;
2721	if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2722		type |= XFS_DQ_GROUP;
2723	/*
2724	 * This type of quotas was turned off, so ignore this buffer
2725	 */
2726	if (log->l_quotaoffs_flag & type)
2727		return false;
2728
2729	xlog_recover_do_reg_buffer(mp, item, bp, buf_f, NULLCOMMITLSN);
2730	return true;
2731}
2732
2733/*
2734 * This routine replays a modification made to a buffer at runtime.
2735 * There are actually two types of buffer, regular and inode, which
2736 * are handled differently.  Inode buffers are handled differently
2737 * in that we only recover a specific set of data from them, namely
2738 * the inode di_next_unlinked fields.  This is because all other inode
2739 * data is actually logged via inode records and any data we replay
2740 * here which overlaps that may be stale.
2741 *
2742 * When meta-data buffers are freed at run time we log a buffer item
2743 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2744 * of the buffer in the log should not be replayed at recovery time.
2745 * This is so that if the blocks covered by the buffer are reused for
2746 * file data before we crash we don't end up replaying old, freed
2747 * meta-data into a user's file.
2748 *
2749 * To handle the cancellation of buffer log items, we make two passes
2750 * over the log during recovery.  During the first we build a table of
2751 * those buffers which have been cancelled, and during the second we
2752 * only replay those buffers which do not have corresponding cancel
2753 * records in the table.  See xlog_recover_buffer_pass[1,2] above
2754 * for more details on the implementation of the table of cancel records.
2755 */
2756STATIC int
2757xlog_recover_buffer_pass2(
2758	struct xlog			*log,
2759	struct list_head		*buffer_list,
2760	struct xlog_recover_item	*item,
2761	xfs_lsn_t			current_lsn)
2762{
2763	xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
2764	xfs_mount_t		*mp = log->l_mp;
2765	xfs_buf_t		*bp;
2766	int			error;
2767	uint			buf_flags;
2768	xfs_lsn_t		lsn;
2769
2770	/*
2771	 * In this pass we only want to recover all the buffers which have
2772	 * not been cancelled and are not cancellation buffers themselves.
2773	 */
2774	if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2775			buf_f->blf_len, buf_f->blf_flags)) {
2776		trace_xfs_log_recover_buf_cancel(log, buf_f);
2777		return 0;
2778	}
2779
2780	trace_xfs_log_recover_buf_recover(log, buf_f);
2781
2782	buf_flags = 0;
2783	if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
2784		buf_flags |= XBF_UNMAPPED;
2785
2786	bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2787			  buf_flags, NULL);
2788	if (!bp)
2789		return -ENOMEM;
2790	error = bp->b_error;
2791	if (error) {
2792		xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#1)");
2793		goto out_release;
2794	}
2795
2796	/*
2797	 * Recover the buffer only if we get an LSN from it and it's less than
2798	 * the lsn of the transaction we are replaying.
2799	 *
2800	 * Note that we have to be extremely careful of readahead here.
2801	 * Readahead does not attach verfiers to the buffers so if we don't
2802	 * actually do any replay after readahead because of the LSN we found
2803	 * in the buffer if more recent than that current transaction then we
2804	 * need to attach the verifier directly. Failure to do so can lead to
2805	 * future recovery actions (e.g. EFI and unlinked list recovery) can
2806	 * operate on the buffers and they won't get the verifier attached. This
2807	 * can lead to blocks on disk having the correct content but a stale
2808	 * CRC.
2809	 *
2810	 * It is safe to assume these clean buffers are currently up to date.
2811	 * If the buffer is dirtied by a later transaction being replayed, then
2812	 * the verifier will be reset to match whatever recover turns that
2813	 * buffer into.
2814	 */
2815	lsn = xlog_recover_get_buf_lsn(mp, bp);
2816	if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
2817		trace_xfs_log_recover_buf_skip(log, buf_f);
2818		xlog_recover_validate_buf_type(mp, bp, buf_f, NULLCOMMITLSN);
2819		goto out_release;
2820	}
2821
2822	if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2823		error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2824		if (error)
2825			goto out_release;
2826	} else if (buf_f->blf_flags &
2827		  (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2828		bool	dirty;
2829
2830		dirty = xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2831		if (!dirty)
2832			goto out_release;
2833	} else {
2834		xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn);
2835	}
2836
2837	/*
2838	 * Perform delayed write on the buffer.  Asynchronous writes will be
2839	 * slower when taking into account all the buffers to be flushed.
2840	 *
2841	 * Also make sure that only inode buffers with good sizes stay in
2842	 * the buffer cache.  The kernel moves inodes in buffers of 1 block
2843	 * or mp->m_inode_cluster_size bytes, whichever is bigger.  The inode
2844	 * buffers in the log can be a different size if the log was generated
2845	 * by an older kernel using unclustered inode buffers or a newer kernel
2846	 * running with a different inode cluster size.  Regardless, if the
2847	 * the inode buffer size isn't MAX(blocksize, mp->m_inode_cluster_size)
2848	 * for *our* value of mp->m_inode_cluster_size, then we need to keep
2849	 * the buffer out of the buffer cache so that the buffer won't
2850	 * overlap with future reads of those inodes.
2851	 */
2852	if (XFS_DINODE_MAGIC ==
2853	    be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2854	    (BBTOB(bp->b_io_length) != MAX(log->l_mp->m_sb.sb_blocksize,
2855			(uint32_t)log->l_mp->m_inode_cluster_size))) {
2856		xfs_buf_stale(bp);
2857		error = xfs_bwrite(bp);
2858	} else {
2859		ASSERT(bp->b_target->bt_mount == mp);
2860		bp->b_iodone = xlog_recover_iodone;
2861		xfs_buf_delwri_queue(bp, buffer_list);
2862	}
2863
2864out_release:
2865	xfs_buf_relse(bp);
2866	return error;
2867}
2868
2869/*
2870 * Inode fork owner changes
2871 *
2872 * If we have been told that we have to reparent the inode fork, it's because an
2873 * extent swap operation on a CRC enabled filesystem has been done and we are
2874 * replaying it. We need to walk the BMBT of the appropriate fork and change the
2875 * owners of it.
2876 *
2877 * The complexity here is that we don't have an inode context to work with, so
2878 * after we've replayed the inode we need to instantiate one.  This is where the
2879 * fun begins.
2880 *
2881 * We are in the middle of log recovery, so we can't run transactions. That
2882 * means we cannot use cache coherent inode instantiation via xfs_iget(), as
2883 * that will result in the corresponding iput() running the inode through
2884 * xfs_inactive(). If we've just replayed an inode core that changes the link
2885 * count to zero (i.e. it's been unlinked), then xfs_inactive() will run
2886 * transactions (bad!).
2887 *
2888 * So, to avoid this, we instantiate an inode directly from the inode core we've
2889 * just recovered. We have the buffer still locked, and all we really need to
2890 * instantiate is the inode core and the forks being modified. We can do this
2891 * manually, then run the inode btree owner change, and then tear down the
2892 * xfs_inode without having to run any transactions at all.
2893 *
2894 * Also, because we don't have a transaction context available here but need to
2895 * gather all the buffers we modify for writeback so we pass the buffer_list
2896 * instead for the operation to use.
2897 */
2898
2899STATIC int
2900xfs_recover_inode_owner_change(
2901	struct xfs_mount	*mp,
2902	struct xfs_dinode	*dip,
2903	struct xfs_inode_log_format *in_f,
2904	struct list_head	*buffer_list)
2905{
2906	struct xfs_inode	*ip;
2907	int			error;
2908
2909	ASSERT(in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER));
2910
2911	ip = xfs_inode_alloc(mp, in_f->ilf_ino);
2912	if (!ip)
2913		return -ENOMEM;
2914
2915	/* instantiate the inode */
2916	xfs_inode_from_disk(ip, dip);
2917	ASSERT(ip->i_d.di_version >= 3);
2918
2919	error = xfs_iformat_fork(ip, dip);
2920	if (error)
2921		goto out_free_ip;
2922
2923
2924	if (in_f->ilf_fields & XFS_ILOG_DOWNER) {
2925		ASSERT(in_f->ilf_fields & XFS_ILOG_DBROOT);
2926		error = xfs_bmbt_change_owner(NULL, ip, XFS_DATA_FORK,
2927					      ip->i_ino, buffer_list);
2928		if (error)
2929			goto out_free_ip;
2930	}
2931
2932	if (in_f->ilf_fields & XFS_ILOG_AOWNER) {
2933		ASSERT(in_f->ilf_fields & XFS_ILOG_ABROOT);
2934		error = xfs_bmbt_change_owner(NULL, ip, XFS_ATTR_FORK,
2935					      ip->i_ino, buffer_list);
2936		if (error)
2937			goto out_free_ip;
2938	}
2939
2940out_free_ip:
2941	xfs_inode_free(ip);
2942	return error;
2943}
2944
2945STATIC int
2946xlog_recover_inode_pass2(
2947	struct xlog			*log,
2948	struct list_head		*buffer_list,
2949	struct xlog_recover_item	*item,
2950	xfs_lsn_t			current_lsn)
2951{
2952	xfs_inode_log_format_t	*in_f;
2953	xfs_mount_t		*mp = log->l_mp;
2954	xfs_buf_t		*bp;
2955	xfs_dinode_t		*dip;
2956	int			len;
2957	char			*src;
2958	char			*dest;
2959	int			error;
2960	int			attr_index;
2961	uint			fields;
2962	struct xfs_log_dinode	*ldip;
2963	uint			isize;
2964	int			need_free = 0;
2965
2966	if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2967		in_f = item->ri_buf[0].i_addr;
2968	} else {
2969		in_f = kmem_alloc(sizeof(xfs_inode_log_format_t), KM_SLEEP);
2970		need_free = 1;
2971		error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2972		if (error)
2973			goto error;
2974	}
2975
2976	/*
2977	 * Inode buffers can be freed, look out for it,
2978	 * and do not replay the inode.
2979	 */
2980	if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2981					in_f->ilf_len, 0)) {
2982		error = 0;
2983		trace_xfs_log_recover_inode_cancel(log, in_f);
2984		goto error;
2985	}
2986	trace_xfs_log_recover_inode_recover(log, in_f);
2987
2988	bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len, 0,
2989			  &xfs_inode_buf_ops);
2990	if (!bp) {
2991		error = -ENOMEM;
2992		goto error;
2993	}
2994	error = bp->b_error;
2995	if (error) {
2996		xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#2)");
2997		goto out_release;
2998	}
2999	ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
3000	dip = xfs_buf_offset(bp, in_f->ilf_boffset);
3001
3002	/*
3003	 * Make sure the place we're flushing out to really looks
3004	 * like an inode!
3005	 */
3006	if (unlikely(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC))) {
3007		xfs_alert(mp,
3008	"%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
3009			__func__, dip, bp, in_f->ilf_ino);
3010		XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
3011				 XFS_ERRLEVEL_LOW, mp);
3012		error = -EFSCORRUPTED;
3013		goto out_release;
3014	}
3015	ldip = item->ri_buf[1].i_addr;
3016	if (unlikely(ldip->di_magic != XFS_DINODE_MAGIC)) {
3017		xfs_alert(mp,
3018			"%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
3019			__func__, item, in_f->ilf_ino);
3020		XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
3021				 XFS_ERRLEVEL_LOW, mp);
3022		error = -EFSCORRUPTED;
3023		goto out_release;
3024	}
3025
3026	/*
3027	 * If the inode has an LSN in it, recover the inode only if it's less
3028	 * than the lsn of the transaction we are replaying. Note: we still
3029	 * need to replay an owner change even though the inode is more recent
3030	 * than the transaction as there is no guarantee that all the btree
3031	 * blocks are more recent than this transaction, too.
3032	 */
3033	if (dip->di_version >= 3) {
3034		xfs_lsn_t	lsn = be64_to_cpu(dip->di_lsn);
3035
3036		if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3037			trace_xfs_log_recover_inode_skip(log, in_f);
3038			error = 0;
3039			goto out_owner_change;
3040		}
3041	}
3042
3043	/*
3044	 * di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
3045	 * are transactional and if ordering is necessary we can determine that
3046	 * more accurately by the LSN field in the V3 inode core. Don't trust
3047	 * the inode versions we might be changing them here - use the
3048	 * superblock flag to determine whether we need to look at di_flushiter
3049	 * to skip replay when the on disk inode is newer than the log one
3050	 */
3051	if (!xfs_sb_version_hascrc(&mp->m_sb) &&
3052	    ldip->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
3053		/*
3054		 * Deal with the wrap case, DI_MAX_FLUSH is less
3055		 * than smaller numbers
3056		 */
3057		if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
3058		    ldip->di_flushiter < (DI_MAX_FLUSH >> 1)) {
3059			/* do nothing */
3060		} else {
3061			trace_xfs_log_recover_inode_skip(log, in_f);
3062			error = 0;
3063			goto out_release;
3064		}
3065	}
3066
3067	/* Take the opportunity to reset the flush iteration count */
3068	ldip->di_flushiter = 0;
3069
3070	if (unlikely(S_ISREG(ldip->di_mode))) {
3071		if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3072		    (ldip->di_format != XFS_DINODE_FMT_BTREE)) {
3073			XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
3074					 XFS_ERRLEVEL_LOW, mp, ldip);
3075			xfs_alert(mp,
3076		"%s: Bad regular inode log record, rec ptr 0x%p, "
3077		"ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
3078				__func__, item, dip, bp, in_f->ilf_ino);
3079			error = -EFSCORRUPTED;
3080			goto out_release;
3081		}
3082	} else if (unlikely(S_ISDIR(ldip->di_mode))) {
3083		if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3084		    (ldip->di_format != XFS_DINODE_FMT_BTREE) &&
3085		    (ldip->di_format != XFS_DINODE_FMT_LOCAL)) {
3086			XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
3087					     XFS_ERRLEVEL_LOW, mp, ldip);
3088			xfs_alert(mp,
3089		"%s: Bad dir inode log record, rec ptr 0x%p, "
3090		"ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
3091				__func__, item, dip, bp, in_f->ilf_ino);
3092			error = -EFSCORRUPTED;
3093			goto out_release;
3094		}
3095	}
3096	if (unlikely(ldip->di_nextents + ldip->di_anextents > ldip->di_nblocks)){
3097		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
3098				     XFS_ERRLEVEL_LOW, mp, ldip);
3099		xfs_alert(mp,
3100	"%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
3101	"dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
3102			__func__, item, dip, bp, in_f->ilf_ino,
3103			ldip->di_nextents + ldip->di_anextents,
3104			ldip->di_nblocks);
3105		error = -EFSCORRUPTED;
3106		goto out_release;
3107	}
3108	if (unlikely(ldip->di_forkoff > mp->m_sb.sb_inodesize)) {
3109		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
3110				     XFS_ERRLEVEL_LOW, mp, ldip);
3111		xfs_alert(mp,
3112	"%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
3113	"dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__,
3114			item, dip, bp, in_f->ilf_ino, ldip->di_forkoff);
3115		error = -EFSCORRUPTED;
3116		goto out_release;
3117	}
3118	isize = xfs_log_dinode_size(ldip->di_version);
3119	if (unlikely(item->ri_buf[1].i_len > isize)) {
3120		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
3121				     XFS_ERRLEVEL_LOW, mp, ldip);
3122		xfs_alert(mp,
3123			"%s: Bad inode log record length %d, rec ptr 0x%p",
3124			__func__, item->ri_buf[1].i_len, item);
3125		error = -EFSCORRUPTED;
3126		goto out_release;
3127	}
3128
3129	/* recover the log dinode inode into the on disk inode */
3130	xfs_log_dinode_to_disk(ldip, dip);
3131
3132	/* the rest is in on-disk format */
3133	if (item->ri_buf[1].i_len > isize) {
3134		memcpy((char *)dip + isize,
3135			item->ri_buf[1].i_addr + isize,
3136			item->ri_buf[1].i_len - isize);
3137	}
3138
3139	fields = in_f->ilf_fields;
3140	switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
3141	case XFS_ILOG_DEV:
3142		xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
3143		break;
3144	case XFS_ILOG_UUID:
3145		memcpy(XFS_DFORK_DPTR(dip),
3146		       &in_f->ilf_u.ilfu_uuid,
3147		       sizeof(uuid_t));
3148		break;
3149	}
3150
3151	if (in_f->ilf_size == 2)
3152		goto out_owner_change;
3153	len = item->ri_buf[2].i_len;
3154	src = item->ri_buf[2].i_addr;
3155	ASSERT(in_f->ilf_size <= 4);
3156	ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
3157	ASSERT(!(fields & XFS_ILOG_DFORK) ||
3158	       (len == in_f->ilf_dsize));
3159
3160	switch (fields & XFS_ILOG_DFORK) {
3161	case XFS_ILOG_DDATA:
3162	case XFS_ILOG_DEXT:
3163		memcpy(XFS_DFORK_DPTR(dip), src, len);
3164		break;
3165
3166	case XFS_ILOG_DBROOT:
3167		xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
3168				 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
3169				 XFS_DFORK_DSIZE(dip, mp));
3170		break;
3171
3172	default:
3173		/*
3174		 * There are no data fork flags set.
3175		 */
3176		ASSERT((fields & XFS_ILOG_DFORK) == 0);
3177		break;
3178	}
3179
3180	/*
3181	 * If we logged any attribute data, recover it.  There may or
3182	 * may not have been any other non-core data logged in this
3183	 * transaction.
3184	 */
3185	if (in_f->ilf_fields & XFS_ILOG_AFORK) {
3186		if (in_f->ilf_fields & XFS_ILOG_DFORK) {
3187			attr_index = 3;
3188		} else {
3189			attr_index = 2;
3190		}
3191		len = item->ri_buf[attr_index].i_len;
3192		src = item->ri_buf[attr_index].i_addr;
3193		ASSERT(len == in_f->ilf_asize);
3194
3195		switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
3196		case XFS_ILOG_ADATA:
3197		case XFS_ILOG_AEXT:
3198			dest = XFS_DFORK_APTR(dip);
3199			ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
3200			memcpy(dest, src, len);
3201			break;
3202
3203		case XFS_ILOG_ABROOT:
3204			dest = XFS_DFORK_APTR(dip);
3205			xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
3206					 len, (xfs_bmdr_block_t*)dest,
3207					 XFS_DFORK_ASIZE(dip, mp));
3208			break;
3209
3210		default:
3211			xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
3212			ASSERT(0);
3213			error = -EIO;
3214			goto out_release;
3215		}
3216	}
3217
3218out_owner_change:
3219	if (in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER))
3220		error = xfs_recover_inode_owner_change(mp, dip, in_f,
3221						       buffer_list);
3222	/* re-generate the checksum. */
3223	xfs_dinode_calc_crc(log->l_mp, dip);
3224
3225	ASSERT(bp->b_target->bt_mount == mp);
3226	bp->b_iodone = xlog_recover_iodone;
3227	xfs_buf_delwri_queue(bp, buffer_list);
3228
3229out_release:
3230	xfs_buf_relse(bp);
3231error:
3232	if (need_free)
3233		kmem_free(in_f);
3234	return error;
3235}
3236
3237/*
3238 * Recover QUOTAOFF records. We simply make a note of it in the xlog
3239 * structure, so that we know not to do any dquot item or dquot buffer recovery,
3240 * of that type.
3241 */
3242STATIC int
3243xlog_recover_quotaoff_pass1(
3244	struct xlog			*log,
3245	struct xlog_recover_item	*item)
3246{
3247	xfs_qoff_logformat_t	*qoff_f = item->ri_buf[0].i_addr;
3248	ASSERT(qoff_f);
3249
3250	/*
3251	 * The logitem format's flag tells us if this was user quotaoff,
3252	 * group/project quotaoff or both.
3253	 */
3254	if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
3255		log->l_quotaoffs_flag |= XFS_DQ_USER;
3256	if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
3257		log->l_quotaoffs_flag |= XFS_DQ_PROJ;
3258	if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
3259		log->l_quotaoffs_flag |= XFS_DQ_GROUP;
3260
3261	return 0;
3262}
3263
3264/*
3265 * Recover a dquot record
3266 */
3267STATIC int
3268xlog_recover_dquot_pass2(
3269	struct xlog			*log,
3270	struct list_head		*buffer_list,
3271	struct xlog_recover_item	*item,
3272	xfs_lsn_t			current_lsn)
3273{
3274	xfs_mount_t		*mp = log->l_mp;
3275	xfs_buf_t		*bp;
3276	struct xfs_disk_dquot	*ddq, *recddq;
3277	int			error;
3278	xfs_dq_logformat_t	*dq_f;
3279	uint			type;
3280
3281
3282	/*
3283	 * Filesystems are required to send in quota flags at mount time.
3284	 */
3285	if (mp->m_qflags == 0)
3286		return 0;
3287
3288	recddq = item->ri_buf[1].i_addr;
3289	if (recddq == NULL) {
3290		xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
3291		return -EIO;
3292	}
3293	if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
3294		xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
3295			item->ri_buf[1].i_len, __func__);
3296		return -EIO;
3297	}
3298
3299	/*
3300	 * This type of quotas was turned off, so ignore this record.
3301	 */
3302	type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3303	ASSERT(type);
3304	if (log->l_quotaoffs_flag & type)
3305		return 0;
3306
3307	/*
3308	 * At this point we know that quota was _not_ turned off.
3309	 * Since the mount flags are not indicating to us otherwise, this
3310	 * must mean that quota is on, and the dquot needs to be replayed.
3311	 * Remember that we may not have fully recovered the superblock yet,
3312	 * so we can't do the usual trick of looking at the SB quota bits.
3313	 *
3314	 * The other possibility, of course, is that the quota subsystem was
3315	 * removed since the last mount - ENOSYS.
3316	 */
3317	dq_f = item->ri_buf[0].i_addr;
3318	ASSERT(dq_f);
3319	error = xfs_dqcheck(mp, recddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
3320			   "xlog_recover_dquot_pass2 (log copy)");
3321	if (error)
3322		return -EIO;
3323	ASSERT(dq_f->qlf_len == 1);
3324
3325	/*
3326	 * At this point we are assuming that the dquots have been allocated
3327	 * and hence the buffer has valid dquots stamped in it. It should,
3328	 * therefore, pass verifier validation. If the dquot is bad, then the
3329	 * we'll return an error here, so we don't need to specifically check
3330	 * the dquot in the buffer after the verifier has run.
3331	 */
3332	error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dq_f->qlf_blkno,
3333				   XFS_FSB_TO_BB(mp, dq_f->qlf_len), 0, &bp,
3334				   &xfs_dquot_buf_ops);
3335	if (error)
3336		return error;
3337
3338	ASSERT(bp);
3339	ddq = xfs_buf_offset(bp, dq_f->qlf_boffset);
3340
3341	/*
3342	 * If the dquot has an LSN in it, recover the dquot only if it's less
3343	 * than the lsn of the transaction we are replaying.
3344	 */
3345	if (xfs_sb_version_hascrc(&mp->m_sb)) {
3346		struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddq;
3347		xfs_lsn_t	lsn = be64_to_cpu(dqb->dd_lsn);
3348
3349		if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3350			goto out_release;
3351		}
3352	}
3353
3354	memcpy(ddq, recddq, item->ri_buf[1].i_len);
3355	if (xfs_sb_version_hascrc(&mp->m_sb)) {
3356		xfs_update_cksum((char *)ddq, sizeof(struct xfs_dqblk),
3357				 XFS_DQUOT_CRC_OFF);
3358	}
3359
3360	ASSERT(dq_f->qlf_size == 2);
3361	ASSERT(bp->b_target->bt_mount == mp);
3362	bp->b_iodone = xlog_recover_iodone;
3363	xfs_buf_delwri_queue(bp, buffer_list);
3364
3365out_release:
3366	xfs_buf_relse(bp);
3367	return 0;
3368}
3369
3370/*
3371 * This routine is called to create an in-core extent free intent
3372 * item from the efi format structure which was logged on disk.
3373 * It allocates an in-core efi, copies the extents from the format
3374 * structure into it, and adds the efi to the AIL with the given
3375 * LSN.
3376 */
3377STATIC int
3378xlog_recover_efi_pass2(
3379	struct xlog			*log,
3380	struct xlog_recover_item	*item,
3381	xfs_lsn_t			lsn)
3382{
3383	int				error;
3384	struct xfs_mount		*mp = log->l_mp;
3385	struct xfs_efi_log_item		*efip;
3386	struct xfs_efi_log_format	*efi_formatp;
3387
3388	efi_formatp = item->ri_buf[0].i_addr;
3389
3390	efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
3391	error = xfs_efi_copy_format(&item->ri_buf[0], &efip->efi_format);
3392	if (error) {
3393		xfs_efi_item_free(efip);
3394		return error;
3395	}
3396	atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
3397
3398	spin_lock(&log->l_ailp->xa_lock);
3399	/*
3400	 * The EFI has two references. One for the EFD and one for EFI to ensure
3401	 * it makes it into the AIL. Insert the EFI into the AIL directly and
3402	 * drop the EFI reference. Note that xfs_trans_ail_update() drops the
3403	 * AIL lock.
3404	 */
3405	xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
3406	xfs_efi_release(efip);
3407	return 0;
3408}
3409
3410
3411/*
3412 * This routine is called when an EFD format structure is found in a committed
3413 * transaction in the log. Its purpose is to cancel the corresponding EFI if it
3414 * was still in the log. To do this it searches the AIL for the EFI with an id
3415 * equal to that in the EFD format structure. If we find it we drop the EFD
3416 * reference, which removes the EFI from the AIL and frees it.
3417 */
3418STATIC int
3419xlog_recover_efd_pass2(
3420	struct xlog			*log,
3421	struct xlog_recover_item	*item)
3422{
3423	xfs_efd_log_format_t	*efd_formatp;
3424	xfs_efi_log_item_t	*efip = NULL;
3425	xfs_log_item_t		*lip;
3426	uint64_t		efi_id;
3427	struct xfs_ail_cursor	cur;
3428	struct xfs_ail		*ailp = log->l_ailp;
3429
3430	efd_formatp = item->ri_buf[0].i_addr;
3431	ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
3432		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
3433	       (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
3434		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
3435	efi_id = efd_formatp->efd_efi_id;
3436
3437	/*
3438	 * Search for the EFI with the id in the EFD format structure in the
3439	 * AIL.
3440	 */
3441	spin_lock(&ailp->xa_lock);
3442	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3443	while (lip != NULL) {
3444		if (lip->li_type == XFS_LI_EFI) {
3445			efip = (xfs_efi_log_item_t *)lip;
3446			if (efip->efi_format.efi_id == efi_id) {
3447				/*
3448				 * Drop the EFD reference to the EFI. This
3449				 * removes the EFI from the AIL and frees it.
3450				 */
3451				spin_unlock(&ailp->xa_lock);
3452				xfs_efi_release(efip);
3453				spin_lock(&ailp->xa_lock);
3454				break;
3455			}
3456		}
3457		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3458	}
3459
3460	xfs_trans_ail_cursor_done(&cur);
3461	spin_unlock(&ailp->xa_lock);
3462
3463	return 0;
3464}
3465
3466/*
3467 * This routine is called to create an in-core extent rmap update
3468 * item from the rui format structure which was logged on disk.
3469 * It allocates an in-core rui, copies the extents from the format
3470 * structure into it, and adds the rui to the AIL with the given
3471 * LSN.
3472 */
3473STATIC int
3474xlog_recover_rui_pass2(
3475	struct xlog			*log,
3476	struct xlog_recover_item	*item,
3477	xfs_lsn_t			lsn)
3478{
3479	int				error;
3480	struct xfs_mount		*mp = log->l_mp;
3481	struct xfs_rui_log_item		*ruip;
3482	struct xfs_rui_log_format	*rui_formatp;
3483
3484	rui_formatp = item->ri_buf[0].i_addr;
3485
3486	ruip = xfs_rui_init(mp, rui_formatp->rui_nextents);
3487	error = xfs_rui_copy_format(&item->ri_buf[0], &ruip->rui_format);
3488	if (error) {
3489		xfs_rui_item_free(ruip);
3490		return error;
3491	}
3492	atomic_set(&ruip->rui_next_extent, rui_formatp->rui_nextents);
3493
3494	spin_lock(&log->l_ailp->xa_lock);
3495	/*
3496	 * The RUI has two references. One for the RUD and one for RUI to ensure
3497	 * it makes it into the AIL. Insert the RUI into the AIL directly and
3498	 * drop the RUI reference. Note that xfs_trans_ail_update() drops the
3499	 * AIL lock.
3500	 */
3501	xfs_trans_ail_update(log->l_ailp, &ruip->rui_item, lsn);
3502	xfs_rui_release(ruip);
3503	return 0;
3504}
3505
3506
3507/*
3508 * This routine is called when an RUD format structure is found in a committed
3509 * transaction in the log. Its purpose is to cancel the corresponding RUI if it
3510 * was still in the log. To do this it searches the AIL for the RUI with an id
3511 * equal to that in the RUD format structure. If we find it we drop the RUD
3512 * reference, which removes the RUI from the AIL and frees it.
3513 */
3514STATIC int
3515xlog_recover_rud_pass2(
3516	struct xlog			*log,
3517	struct xlog_recover_item	*item)
3518{
3519	struct xfs_rud_log_format	*rud_formatp;
3520	struct xfs_rui_log_item		*ruip = NULL;
3521	struct xfs_log_item		*lip;
3522	uint64_t			rui_id;
3523	struct xfs_ail_cursor		cur;
3524	struct xfs_ail			*ailp = log->l_ailp;
3525
3526	rud_formatp = item->ri_buf[0].i_addr;
3527	ASSERT(item->ri_buf[0].i_len == sizeof(struct xfs_rud_log_format));
3528	rui_id = rud_formatp->rud_rui_id;
3529
3530	/*
3531	 * Search for the RUI with the id in the RUD format structure in the
3532	 * AIL.
3533	 */
3534	spin_lock(&ailp->xa_lock);
3535	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3536	while (lip != NULL) {
3537		if (lip->li_type == XFS_LI_RUI) {
3538			ruip = (struct xfs_rui_log_item *)lip;
3539			if (ruip->rui_format.rui_id == rui_id) {
3540				/*
3541				 * Drop the RUD reference to the RUI. This
3542				 * removes the RUI from the AIL and frees it.
3543				 */
3544				spin_unlock(&ailp->xa_lock);
3545				xfs_rui_release(ruip);
3546				spin_lock(&ailp->xa_lock);
3547				break;
3548			}
3549		}
3550		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3551	}
3552
3553	xfs_trans_ail_cursor_done(&cur);
3554	spin_unlock(&ailp->xa_lock);
3555
3556	return 0;
3557}
3558
3559/*
3560 * Copy an CUI format buffer from the given buf, and into the destination
3561 * CUI format structure.  The CUI/CUD items were designed not to need any
3562 * special alignment handling.
3563 */
3564static int
3565xfs_cui_copy_format(
3566	struct xfs_log_iovec		*buf,
3567	struct xfs_cui_log_format	*dst_cui_fmt)
3568{
3569	struct xfs_cui_log_format	*src_cui_fmt;
3570	uint				len;
3571
3572	src_cui_fmt = buf->i_addr;
3573	len = xfs_cui_log_format_sizeof(src_cui_fmt->cui_nextents);
3574
3575	if (buf->i_len == len) {
3576		memcpy(dst_cui_fmt, src_cui_fmt, len);
3577		return 0;
3578	}
3579	return -EFSCORRUPTED;
3580}
3581
3582/*
3583 * This routine is called to create an in-core extent refcount update
3584 * item from the cui format structure which was logged on disk.
3585 * It allocates an in-core cui, copies the extents from the format
3586 * structure into it, and adds the cui to the AIL with the given
3587 * LSN.
3588 */
3589STATIC int
3590xlog_recover_cui_pass2(
3591	struct xlog			*log,
3592	struct xlog_recover_item	*item,
3593	xfs_lsn_t			lsn)
3594{
3595	int				error;
3596	struct xfs_mount		*mp = log->l_mp;
3597	struct xfs_cui_log_item		*cuip;
3598	struct xfs_cui_log_format	*cui_formatp;
3599
3600	cui_formatp = item->ri_buf[0].i_addr;
3601
3602	cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
3603	error = xfs_cui_copy_format(&item->ri_buf[0], &cuip->cui_format);
3604	if (error) {
3605		xfs_cui_item_free(cuip);
3606		return error;
3607	}
3608	atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
3609
3610	spin_lock(&log->l_ailp->xa_lock);
3611	/*
3612	 * The CUI has two references. One for the CUD and one for CUI to ensure
3613	 * it makes it into the AIL. Insert the CUI into the AIL directly and
3614	 * drop the CUI reference. Note that xfs_trans_ail_update() drops the
3615	 * AIL lock.
3616	 */
3617	xfs_trans_ail_update(log->l_ailp, &cuip->cui_item, lsn);
3618	xfs_cui_release(cuip);
3619	return 0;
3620}
3621
3622
3623/*
3624 * This routine is called when an CUD format structure is found in a committed
3625 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
3626 * was still in the log. To do this it searches the AIL for the CUI with an id
3627 * equal to that in the CUD format structure. If we find it we drop the CUD
3628 * reference, which removes the CUI from the AIL and frees it.
3629 */
3630STATIC int
3631xlog_recover_cud_pass2(
3632	struct xlog			*log,
3633	struct xlog_recover_item	*item)
3634{
3635	struct xfs_cud_log_format	*cud_formatp;
3636	struct xfs_cui_log_item		*cuip = NULL;
3637	struct xfs_log_item		*lip;
3638	uint64_t			cui_id;
3639	struct xfs_ail_cursor		cur;
3640	struct xfs_ail			*ailp = log->l_ailp;
3641
3642	cud_formatp = item->ri_buf[0].i_addr;
3643	if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format))
3644		return -EFSCORRUPTED;
3645	cui_id = cud_formatp->cud_cui_id;
3646
3647	/*
3648	 * Search for the CUI with the id in the CUD format structure in the
3649	 * AIL.
3650	 */
3651	spin_lock(&ailp->xa_lock);
3652	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3653	while (lip != NULL) {
3654		if (lip->li_type == XFS_LI_CUI) {
3655			cuip = (struct xfs_cui_log_item *)lip;
3656			if (cuip->cui_format.cui_id == cui_id) {
3657				/*
3658				 * Drop the CUD reference to the CUI. This
3659				 * removes the CUI from the AIL and frees it.
3660				 */
3661				spin_unlock(&ailp->xa_lock);
3662				xfs_cui_release(cuip);
3663				spin_lock(&ailp->xa_lock);
3664				break;
3665			}
3666		}
3667		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3668	}
3669
3670	xfs_trans_ail_cursor_done(&cur);
3671	spin_unlock(&ailp->xa_lock);
3672
3673	return 0;
3674}
3675
3676/*
3677 * Copy an BUI format buffer from the given buf, and into the destination
3678 * BUI format structure.  The BUI/BUD items were designed not to need any
3679 * special alignment handling.
3680 */
3681static int
3682xfs_bui_copy_format(
3683	struct xfs_log_iovec		*buf,
3684	struct xfs_bui_log_format	*dst_bui_fmt)
3685{
3686	struct xfs_bui_log_format	*src_bui_fmt;
3687	uint				len;
3688
3689	src_bui_fmt = buf->i_addr;
3690	len = xfs_bui_log_format_sizeof(src_bui_fmt->bui_nextents);
3691
3692	if (buf->i_len == len) {
3693		memcpy(dst_bui_fmt, src_bui_fmt, len);
3694		return 0;
3695	}
3696	return -EFSCORRUPTED;
3697}
3698
3699/*
3700 * This routine is called to create an in-core extent bmap update
3701 * item from the bui format structure which was logged on disk.
3702 * It allocates an in-core bui, copies the extents from the format
3703 * structure into it, and adds the bui to the AIL with the given
3704 * LSN.
3705 */
3706STATIC int
3707xlog_recover_bui_pass2(
3708	struct xlog			*log,
3709	struct xlog_recover_item	*item,
3710	xfs_lsn_t			lsn)
3711{
3712	int				error;
3713	struct xfs_mount		*mp = log->l_mp;
3714	struct xfs_bui_log_item		*buip;
3715	struct xfs_bui_log_format	*bui_formatp;
3716
3717	bui_formatp = item->ri_buf[0].i_addr;
3718
3719	if (bui_formatp->bui_nextents != XFS_BUI_MAX_FAST_EXTENTS)
3720		return -EFSCORRUPTED;
3721	buip = xfs_bui_init(mp);
3722	error = xfs_bui_copy_format(&item->ri_buf[0], &buip->bui_format);
3723	if (error) {
3724		xfs_bui_item_free(buip);
3725		return error;
3726	}
3727	atomic_set(&buip->bui_next_extent, bui_formatp->bui_nextents);
3728
3729	spin_lock(&log->l_ailp->xa_lock);
3730	/*
3731	 * The RUI has two references. One for the RUD and one for RUI to ensure
3732	 * it makes it into the AIL. Insert the RUI into the AIL directly and
3733	 * drop the RUI reference. Note that xfs_trans_ail_update() drops the
3734	 * AIL lock.
3735	 */
3736	xfs_trans_ail_update(log->l_ailp, &buip->bui_item, lsn);
3737	xfs_bui_release(buip);
3738	return 0;
3739}
3740
3741
3742/*
3743 * This routine is called when an BUD format structure is found in a committed
3744 * transaction in the log. Its purpose is to cancel the corresponding BUI if it
3745 * was still in the log. To do this it searches the AIL for the BUI with an id
3746 * equal to that in the BUD format structure. If we find it we drop the BUD
3747 * reference, which removes the BUI from the AIL and frees it.
3748 */
3749STATIC int
3750xlog_recover_bud_pass2(
3751	struct xlog			*log,
3752	struct xlog_recover_item	*item)
3753{
3754	struct xfs_bud_log_format	*bud_formatp;
3755	struct xfs_bui_log_item		*buip = NULL;
3756	struct xfs_log_item		*lip;
3757	uint64_t			bui_id;
3758	struct xfs_ail_cursor		cur;
3759	struct xfs_ail			*ailp = log->l_ailp;
3760
3761	bud_formatp = item->ri_buf[0].i_addr;
3762	if (item->ri_buf[0].i_len != sizeof(struct xfs_bud_log_format))
3763		return -EFSCORRUPTED;
3764	bui_id = bud_formatp->bud_bui_id;
3765
3766	/*
3767	 * Search for the BUI with the id in the BUD format structure in the
3768	 * AIL.
3769	 */
3770	spin_lock(&ailp->xa_lock);
3771	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3772	while (lip != NULL) {
3773		if (lip->li_type == XFS_LI_BUI) {
3774			buip = (struct xfs_bui_log_item *)lip;
3775			if (buip->bui_format.bui_id == bui_id) {
3776				/*
3777				 * Drop the BUD reference to the BUI. This
3778				 * removes the BUI from the AIL and frees it.
3779				 */
3780				spin_unlock(&ailp->xa_lock);
3781				xfs_bui_release(buip);
3782				spin_lock(&ailp->xa_lock);
3783				break;
3784			}
3785		}
3786		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3787	}
3788
3789	xfs_trans_ail_cursor_done(&cur);
3790	spin_unlock(&ailp->xa_lock);
3791
3792	return 0;
3793}
3794
3795/*
3796 * This routine is called when an inode create format structure is found in a
3797 * committed transaction in the log.  It's purpose is to initialise the inodes
3798 * being allocated on disk. This requires us to get inode cluster buffers that
3799 * match the range to be initialised, stamped with inode templates and written
3800 * by delayed write so that subsequent modifications will hit the cached buffer
3801 * and only need writing out at the end of recovery.
3802 */
3803STATIC int
3804xlog_recover_do_icreate_pass2(
3805	struct xlog		*log,
3806	struct list_head	*buffer_list,
3807	xlog_recover_item_t	*item)
3808{
3809	struct xfs_mount	*mp = log->l_mp;
3810	struct xfs_icreate_log	*icl;
3811	xfs_agnumber_t		agno;
3812	xfs_agblock_t		agbno;
3813	unsigned int		count;
3814	unsigned int		isize;
3815	xfs_agblock_t		length;
3816	int			blks_per_cluster;
3817	int			bb_per_cluster;
3818	int			cancel_count;
3819	int			nbufs;
3820	int			i;
3821
3822	icl = (struct xfs_icreate_log *)item->ri_buf[0].i_addr;
3823	if (icl->icl_type != XFS_LI_ICREATE) {
3824		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad type");
3825		return -EINVAL;
3826	}
3827
3828	if (icl->icl_size != 1) {
3829		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad icl size");
3830		return -EINVAL;
3831	}
3832
3833	agno = be32_to_cpu(icl->icl_ag);
3834	if (agno >= mp->m_sb.sb_agcount) {
3835		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agno");
3836		return -EINVAL;
3837	}
3838	agbno = be32_to_cpu(icl->icl_agbno);
3839	if (!agbno || agbno == NULLAGBLOCK || agbno >= mp->m_sb.sb_agblocks) {
3840		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agbno");
3841		return -EINVAL;
3842	}
3843	isize = be32_to_cpu(icl->icl_isize);
3844	if (isize != mp->m_sb.sb_inodesize) {
3845		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad isize");
3846		return -EINVAL;
3847	}
3848	count = be32_to_cpu(icl->icl_count);
3849	if (!count) {
3850		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count");
3851		return -EINVAL;
3852	}
3853	length = be32_to_cpu(icl->icl_length);
3854	if (!length || length >= mp->m_sb.sb_agblocks) {
3855		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad length");
3856		return -EINVAL;
3857	}
3858
3859	/*
3860	 * The inode chunk is either full or sparse and we only support
3861	 * m_ialloc_min_blks sized sparse allocations at this time.
3862	 */
3863	if (length != mp->m_ialloc_blks &&
3864	    length != mp->m_ialloc_min_blks) {
3865		xfs_warn(log->l_mp,
3866			 "%s: unsupported chunk length", __FUNCTION__);
3867		return -EINVAL;
3868	}
3869
3870	/* verify inode count is consistent with extent length */
3871	if ((count >> mp->m_sb.sb_inopblog) != length) {
3872		xfs_warn(log->l_mp,
3873			 "%s: inconsistent inode count and chunk length",
3874			 __FUNCTION__);
3875		return -EINVAL;
3876	}
3877
3878	/*
3879	 * The icreate transaction can cover multiple cluster buffers and these
3880	 * buffers could have been freed and reused. Check the individual
3881	 * buffers for cancellation so we don't overwrite anything written after
3882	 * a cancellation.
3883	 */
3884	blks_per_cluster = xfs_icluster_size_fsb(mp);
3885	bb_per_cluster = XFS_FSB_TO_BB(mp, blks_per_cluster);
3886	nbufs = length / blks_per_cluster;
3887	for (i = 0, cancel_count = 0; i < nbufs; i++) {
3888		xfs_daddr_t	daddr;
3889
3890		daddr = XFS_AGB_TO_DADDR(mp, agno,
3891					 agbno + i * blks_per_cluster);
3892		if (xlog_check_buffer_cancelled(log, daddr, bb_per_cluster, 0))
3893			cancel_count++;
3894	}
3895
3896	/*
3897	 * We currently only use icreate for a single allocation at a time. This
3898	 * means we should expect either all or none of the buffers to be
3899	 * cancelled. Be conservative and skip replay if at least one buffer is
3900	 * cancelled, but warn the user that something is awry if the buffers
3901	 * are not consistent.
3902	 *
3903	 * XXX: This must be refined to only skip cancelled clusters once we use
3904	 * icreate for multiple chunk allocations.
3905	 */
3906	ASSERT(!cancel_count || cancel_count == nbufs);
3907	if (cancel_count) {
3908		if (cancel_count != nbufs)
3909			xfs_warn(mp,
3910	"WARNING: partial inode chunk cancellation, skipped icreate.");
3911		trace_xfs_log_recover_icreate_cancel(log, icl);
3912		return 0;
3913	}
3914
3915	trace_xfs_log_recover_icreate_recover(log, icl);
3916	return xfs_ialloc_inode_init(mp, NULL, buffer_list, count, agno, agbno,
3917				     length, be32_to_cpu(icl->icl_gen));
3918}
3919
3920STATIC void
3921xlog_recover_buffer_ra_pass2(
3922	struct xlog                     *log,
3923	struct xlog_recover_item        *item)
3924{
3925	struct xfs_buf_log_format	*buf_f = item->ri_buf[0].i_addr;
3926	struct xfs_mount		*mp = log->l_mp;
3927
3928	if (xlog_peek_buffer_cancelled(log, buf_f->blf_blkno,
3929			buf_f->blf_len, buf_f->blf_flags)) {
3930		return;
3931	}
3932
3933	xfs_buf_readahead(mp->m_ddev_targp, buf_f->blf_blkno,
3934				buf_f->blf_len, NULL);
3935}
3936
3937STATIC void
3938xlog_recover_inode_ra_pass2(
3939	struct xlog                     *log,
3940	struct xlog_recover_item        *item)
3941{
3942	struct xfs_inode_log_format	ilf_buf;
3943	struct xfs_inode_log_format	*ilfp;
3944	struct xfs_mount		*mp = log->l_mp;
3945	int			error;
3946
3947	if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
3948		ilfp = item->ri_buf[0].i_addr;
3949	} else {
3950		ilfp = &ilf_buf;
3951		memset(ilfp, 0, sizeof(*ilfp));
3952		error = xfs_inode_item_format_convert(&item->ri_buf[0], ilfp);
3953		if (error)
3954			return;
3955	}
3956
3957	if (xlog_peek_buffer_cancelled(log, ilfp->ilf_blkno, ilfp->ilf_len, 0))
3958		return;
3959
3960	xfs_buf_readahead(mp->m_ddev_targp, ilfp->ilf_blkno,
3961				ilfp->ilf_len, &xfs_inode_buf_ra_ops);
3962}
3963
3964STATIC void
3965xlog_recover_dquot_ra_pass2(
3966	struct xlog			*log,
3967	struct xlog_recover_item	*item)
3968{
3969	struct xfs_mount	*mp = log->l_mp;
3970	struct xfs_disk_dquot	*recddq;
3971	struct xfs_dq_logformat	*dq_f;
3972	uint			type;
3973	int			len;
3974
3975
3976	if (mp->m_qflags == 0)
3977		return;
3978
3979	recddq = item->ri_buf[1].i_addr;
3980	if (recddq == NULL)
3981		return;
3982	if (item->ri_buf[1].i_len < sizeof(struct xfs_disk_dquot))
3983		return;
3984
3985	type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3986	ASSERT(type);
3987	if (log->l_quotaoffs_flag & type)
3988		return;
3989
3990	dq_f = item->ri_buf[0].i_addr;
3991	ASSERT(dq_f);
3992	ASSERT(dq_f->qlf_len == 1);
3993
3994	len = XFS_FSB_TO_BB(mp, dq_f->qlf_len);
3995	if (xlog_peek_buffer_cancelled(log, dq_f->qlf_blkno, len, 0))
3996		return;
3997
3998	xfs_buf_readahead(mp->m_ddev_targp, dq_f->qlf_blkno, len,
3999			  &xfs_dquot_buf_ra_ops);
4000}
4001
4002STATIC void
4003xlog_recover_ra_pass2(
4004	struct xlog			*log,
4005	struct xlog_recover_item	*item)
4006{
4007	switch (ITEM_TYPE(item)) {
4008	case XFS_LI_BUF:
4009		xlog_recover_buffer_ra_pass2(log, item);
4010		break;
4011	case XFS_LI_INODE:
4012		xlog_recover_inode_ra_pass2(log, item);
4013		break;
4014	case XFS_LI_DQUOT:
4015		xlog_recover_dquot_ra_pass2(log, item);
4016		break;
4017	case XFS_LI_EFI:
4018	case XFS_LI_EFD:
4019	case XFS_LI_QUOTAOFF:
4020	case XFS_LI_RUI:
4021	case XFS_LI_RUD:
4022	case XFS_LI_CUI:
4023	case XFS_LI_CUD:
4024	case XFS_LI_BUI:
4025	case XFS_LI_BUD:
4026	default:
4027		break;
4028	}
4029}
4030
4031STATIC int
4032xlog_recover_commit_pass1(
4033	struct xlog			*log,
4034	struct xlog_recover		*trans,
4035	struct xlog_recover_item	*item)
4036{
4037	trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
4038
4039	switch (ITEM_TYPE(item)) {
4040	case XFS_LI_BUF:
4041		return xlog_recover_buffer_pass1(log, item);
4042	case XFS_LI_QUOTAOFF:
4043		return xlog_recover_quotaoff_pass1(log, item);
4044	case XFS_LI_INODE:
4045	case XFS_LI_EFI:
4046	case XFS_LI_EFD:
4047	case XFS_LI_DQUOT:
4048	case XFS_LI_ICREATE:
4049	case XFS_LI_RUI:
4050	case XFS_LI_RUD:
4051	case XFS_LI_CUI:
4052	case XFS_LI_CUD:
4053	case XFS_LI_BUI:
4054	case XFS_LI_BUD:
4055		/* nothing to do in pass 1 */
4056		return 0;
4057	default:
4058		xfs_warn(log->l_mp, "%s: invalid item type (%d)",
4059			__func__, ITEM_TYPE(item));
4060		ASSERT(0);
4061		return -EIO;
4062	}
4063}
4064
4065STATIC int
4066xlog_recover_commit_pass2(
4067	struct xlog			*log,
4068	struct xlog_recover		*trans,
4069	struct list_head		*buffer_list,
4070	struct xlog_recover_item	*item)
4071{
4072	trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
4073
4074	switch (ITEM_TYPE(item)) {
4075	case XFS_LI_BUF:
4076		return xlog_recover_buffer_pass2(log, buffer_list, item,
4077						 trans->r_lsn);
4078	case XFS_LI_INODE:
4079		return xlog_recover_inode_pass2(log, buffer_list, item,
4080						 trans->r_lsn);
4081	case XFS_LI_EFI:
4082		return xlog_recover_efi_pass2(log, item, trans->r_lsn);
4083	case XFS_LI_EFD:
4084		return xlog_recover_efd_pass2(log, item);
4085	case XFS_LI_RUI:
4086		return xlog_recover_rui_pass2(log, item, trans->r_lsn);
4087	case XFS_LI_RUD:
4088		return xlog_recover_rud_pass2(log, item);
4089	case XFS_LI_CUI:
4090		return xlog_recover_cui_pass2(log, item, trans->r_lsn);
4091	case XFS_LI_CUD:
4092		return xlog_recover_cud_pass2(log, item);
4093	case XFS_LI_BUI:
4094		return xlog_recover_bui_pass2(log, item, trans->r_lsn);
4095	case XFS_LI_BUD:
4096		return xlog_recover_bud_pass2(log, item);
4097	case XFS_LI_DQUOT:
4098		return xlog_recover_dquot_pass2(log, buffer_list, item,
4099						trans->r_lsn);
4100	case XFS_LI_ICREATE:
4101		return xlog_recover_do_icreate_pass2(log, buffer_list, item);
4102	case XFS_LI_QUOTAOFF:
4103		/* nothing to do in pass2 */
4104		return 0;
4105	default:
4106		xfs_warn(log->l_mp, "%s: invalid item type (%d)",
4107			__func__, ITEM_TYPE(item));
4108		ASSERT(0);
4109		return -EIO;
4110	}
4111}
4112
4113STATIC int
4114xlog_recover_items_pass2(
4115	struct xlog                     *log,
4116	struct xlog_recover             *trans,
4117	struct list_head                *buffer_list,
4118	struct list_head                *item_list)
4119{
4120	struct xlog_recover_item	*item;
4121	int				error = 0;
4122
4123	list_for_each_entry(item, item_list, ri_list) {
4124		error = xlog_recover_commit_pass2(log, trans,
4125					  buffer_list, item);
4126		if (error)
4127			return error;
4128	}
4129
4130	return error;
4131}
4132
4133/*
4134 * Perform the transaction.
4135 *
4136 * If the transaction modifies a buffer or inode, do it now.  Otherwise,
4137 * EFIs and EFDs get queued up by adding entries into the AIL for them.
4138 */
4139STATIC int
4140xlog_recover_commit_trans(
4141	struct xlog		*log,
4142	struct xlog_recover	*trans,
4143	int			pass,
4144	struct list_head	*buffer_list)
4145{
4146	int				error = 0;
4147	int				items_queued = 0;
4148	struct xlog_recover_item	*item;
4149	struct xlog_recover_item	*next;
4150	LIST_HEAD			(ra_list);
4151	LIST_HEAD			(done_list);
4152
4153	#define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
4154
4155	hlist_del_init(&trans->r_list);
4156
4157	error = xlog_recover_reorder_trans(log, trans, pass);
4158	if (error)
4159		return error;
4160
4161	list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
4162		switch (pass) {
4163		case XLOG_RECOVER_PASS1:
4164			error = xlog_recover_commit_pass1(log, trans, item);
4165			break;
4166		case XLOG_RECOVER_PASS2:
4167			xlog_recover_ra_pass2(log, item);
4168			list_move_tail(&item->ri_list, &ra_list);
4169			items_queued++;
4170			if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
4171				error = xlog_recover_items_pass2(log, trans,
4172						buffer_list, &ra_list);
4173				list_splice_tail_init(&ra_list, &done_list);
4174				items_queued = 0;
4175			}
4176
4177			break;
4178		default:
4179			ASSERT(0);
4180		}
4181
4182		if (error)
4183			goto out;
4184	}
4185
4186out:
4187	if (!list_empty(&ra_list)) {
4188		if (!error)
4189			error = xlog_recover_items_pass2(log, trans,
4190					buffer_list, &ra_list);
4191		list_splice_tail_init(&ra_list, &done_list);
4192	}
4193
4194	if (!list_empty(&done_list))
4195		list_splice_init(&done_list, &trans->r_itemq);
4196
4197	return error;
4198}
4199
4200STATIC void
4201xlog_recover_add_item(
4202	struct list_head	*head)
4203{
4204	xlog_recover_item_t	*item;
4205
4206	item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
4207	INIT_LIST_HEAD(&item->ri_list);
4208	list_add_tail(&item->ri_list, head);
4209}
4210
4211STATIC int
4212xlog_recover_add_to_cont_trans(
4213	struct xlog		*log,
4214	struct xlog_recover	*trans,
4215	char			*dp,
4216	int			len)
4217{
4218	xlog_recover_item_t	*item;
4219	char			*ptr, *old_ptr;
4220	int			old_len;
4221
4222	/*
4223	 * If the transaction is empty, the header was split across this and the
4224	 * previous record. Copy the rest of the header.
4225	 */
4226	if (list_empty(&trans->r_itemq)) {
4227		ASSERT(len <= sizeof(struct xfs_trans_header));
4228		if (len > sizeof(struct xfs_trans_header)) {
4229			xfs_warn(log->l_mp, "%s: bad header length", __func__);
4230			return -EIO;
4231		}
4232
4233		xlog_recover_add_item(&trans->r_itemq);
4234		ptr = (char *)&trans->r_theader +
4235				sizeof(struct xfs_trans_header) - len;
4236		memcpy(ptr, dp, len);
4237		return 0;
4238	}
4239
4240	/* take the tail entry */
4241	item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
4242
4243	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
4244	old_len = item->ri_buf[item->ri_cnt-1].i_len;
4245
4246	ptr = kmem_realloc(old_ptr, len + old_len, KM_SLEEP);
4247	memcpy(&ptr[old_len], dp, len);
4248	item->ri_buf[item->ri_cnt-1].i_len += len;
4249	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
4250	trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
4251	return 0;
4252}
4253
4254/*
4255 * The next region to add is the start of a new region.  It could be
4256 * a whole region or it could be the first part of a new region.  Because
4257 * of this, the assumption here is that the type and size fields of all
4258 * format structures fit into the first 32 bits of the structure.
4259 *
4260 * This works because all regions must be 32 bit aligned.  Therefore, we
4261 * either have both fields or we have neither field.  In the case we have
4262 * neither field, the data part of the region is zero length.  We only have
4263 * a log_op_header and can throw away the header since a new one will appear
4264 * later.  If we have at least 4 bytes, then we can determine how many regions
4265 * will appear in the current log item.
4266 */
4267STATIC int
4268xlog_recover_add_to_trans(
4269	struct xlog		*log,
4270	struct xlog_recover	*trans,
4271	char			*dp,
4272	int			len)
4273{
4274	xfs_inode_log_format_t	*in_f;			/* any will do */
4275	xlog_recover_item_t	*item;
4276	char			*ptr;
4277
4278	if (!len)
4279		return 0;
4280	if (list_empty(&trans->r_itemq)) {
4281		/* we need to catch log corruptions here */
4282		if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
4283			xfs_warn(log->l_mp, "%s: bad header magic number",
4284				__func__);
4285			ASSERT(0);
4286			return -EIO;
4287		}
4288
4289		if (len > sizeof(struct xfs_trans_header)) {
4290			xfs_warn(log->l_mp, "%s: bad header length", __func__);
4291			ASSERT(0);
4292			return -EIO;
4293		}
4294
4295		/*
4296		 * The transaction header can be arbitrarily split across op
4297		 * records. If we don't have the whole thing here, copy what we
4298		 * do have and handle the rest in the next record.
4299		 */
4300		if (len == sizeof(struct xfs_trans_header))
4301			xlog_recover_add_item(&trans->r_itemq);
4302		memcpy(&trans->r_theader, dp, len);
4303		return 0;
4304	}
4305
4306	ptr = kmem_alloc(len, KM_SLEEP);
4307	memcpy(ptr, dp, len);
4308	in_f = (xfs_inode_log_format_t *)ptr;
4309
4310	/* take the tail entry */
4311	item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
4312	if (item->ri_total != 0 &&
4313	     item->ri_total == item->ri_cnt) {
4314		/* tail item is in use, get a new one */
4315		xlog_recover_add_item(&trans->r_itemq);
4316		item = list_entry(trans->r_itemq.prev,
4317					xlog_recover_item_t, ri_list);
4318	}
4319
4320	if (item->ri_total == 0) {		/* first region to be added */
4321		if (in_f->ilf_size == 0 ||
4322		    in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
4323			xfs_warn(log->l_mp,
4324		"bad number of regions (%d) in inode log format",
4325				  in_f->ilf_size);
4326			ASSERT(0);
4327			kmem_free(ptr);
4328			return -EIO;
4329		}
4330
4331		item->ri_total = in_f->ilf_size;
4332		item->ri_buf =
4333			kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
4334				    KM_SLEEP);
4335	}
4336	ASSERT(item->ri_total > item->ri_cnt);
4337	/* Description region is ri_buf[0] */
4338	item->ri_buf[item->ri_cnt].i_addr = ptr;
4339	item->ri_buf[item->ri_cnt].i_len  = len;
4340	item->ri_cnt++;
4341	trace_xfs_log_recover_item_add(log, trans, item, 0);
4342	return 0;
4343}
4344
4345/*
4346 * Free up any resources allocated by the transaction
4347 *
4348 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
4349 */
4350STATIC void
4351xlog_recover_free_trans(
4352	struct xlog_recover	*trans)
4353{
4354	xlog_recover_item_t	*item, *n;
4355	int			i;
4356
4357	hlist_del_init(&trans->r_list);
4358
4359	list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
4360		/* Free the regions in the item. */
4361		list_del(&item->ri_list);
4362		for (i = 0; i < item->ri_cnt; i++)
4363			kmem_free(item->ri_buf[i].i_addr);
4364		/* Free the item itself */
4365		kmem_free(item->ri_buf);
4366		kmem_free(item);
4367	}
4368	/* Free the transaction recover structure */
4369	kmem_free(trans);
4370}
4371
4372/*
4373 * On error or completion, trans is freed.
4374 */
4375STATIC int
4376xlog_recovery_process_trans(
4377	struct xlog		*log,
4378	struct xlog_recover	*trans,
4379	char			*dp,
4380	unsigned int		len,
4381	unsigned int		flags,
4382	int			pass,
4383	struct list_head	*buffer_list)
4384{
4385	int			error = 0;
4386	bool			freeit = false;
4387
4388	/* mask off ophdr transaction container flags */
4389	flags &= ~XLOG_END_TRANS;
4390	if (flags & XLOG_WAS_CONT_TRANS)
4391		flags &= ~XLOG_CONTINUE_TRANS;
4392
4393	/*
4394	 * Callees must not free the trans structure. We'll decide if we need to
4395	 * free it or not based on the operation being done and it's result.
4396	 */
4397	switch (flags) {
4398	/* expected flag values */
4399	case 0:
4400	case XLOG_CONTINUE_TRANS:
4401		error = xlog_recover_add_to_trans(log, trans, dp, len);
4402		break;
4403	case XLOG_WAS_CONT_TRANS:
4404		error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
4405		break;
4406	case XLOG_COMMIT_TRANS:
4407		error = xlog_recover_commit_trans(log, trans, pass,
4408						  buffer_list);
4409		/* success or fail, we are now done with this transaction. */
4410		freeit = true;
4411		break;
4412
4413	/* unexpected flag values */
4414	case XLOG_UNMOUNT_TRANS:
4415		/* just skip trans */
4416		xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
4417		freeit = true;
4418		break;
4419	case XLOG_START_TRANS:
4420	default:
4421		xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
4422		ASSERT(0);
4423		error = -EIO;
4424		break;
4425	}
4426	if (error || freeit)
4427		xlog_recover_free_trans(trans);
4428	return error;
4429}
4430
4431/*
4432 * Lookup the transaction recovery structure associated with the ID in the
4433 * current ophdr. If the transaction doesn't exist and the start flag is set in
4434 * the ophdr, then allocate a new transaction for future ID matches to find.
4435 * Either way, return what we found during the lookup - an existing transaction
4436 * or nothing.
4437 */
4438STATIC struct xlog_recover *
4439xlog_recover_ophdr_to_trans(
4440	struct hlist_head	rhash[],
4441	struct xlog_rec_header	*rhead,
4442	struct xlog_op_header	*ohead)
4443{
4444	struct xlog_recover	*trans;
4445	xlog_tid_t		tid;
4446	struct hlist_head	*rhp;
4447
4448	tid = be32_to_cpu(ohead->oh_tid);
4449	rhp = &rhash[XLOG_RHASH(tid)];
4450	hlist_for_each_entry(trans, rhp, r_list) {
4451		if (trans->r_log_tid == tid)
4452			return trans;
4453	}
4454
4455	/*
4456	 * skip over non-start transaction headers - we could be
4457	 * processing slack space before the next transaction starts
4458	 */
4459	if (!(ohead->oh_flags & XLOG_START_TRANS))
4460		return NULL;
4461
4462	ASSERT(be32_to_cpu(ohead->oh_len) == 0);
4463
4464	/*
4465	 * This is a new transaction so allocate a new recovery container to
4466	 * hold the recovery ops that will follow.
4467	 */
4468	trans = kmem_zalloc(sizeof(struct xlog_recover), KM_SLEEP);
4469	trans->r_log_tid = tid;
4470	trans->r_lsn = be64_to_cpu(rhead->h_lsn);
4471	INIT_LIST_HEAD(&trans->r_itemq);
4472	INIT_HLIST_NODE(&trans->r_list);
4473	hlist_add_head(&trans->r_list, rhp);
4474
4475	/*
4476	 * Nothing more to do for this ophdr. Items to be added to this new
4477	 * transaction will be in subsequent ophdr containers.
4478	 */
4479	return NULL;
4480}
4481
4482STATIC int
4483xlog_recover_process_ophdr(
4484	struct xlog		*log,
4485	struct hlist_head	rhash[],
4486	struct xlog_rec_header	*rhead,
4487	struct xlog_op_header	*ohead,
4488	char			*dp,
4489	char			*end,
4490	int			pass,
4491	struct list_head	*buffer_list)
4492{
4493	struct xlog_recover	*trans;
4494	unsigned int		len;
4495	int			error;
4496
4497	/* Do we understand who wrote this op? */
4498	if (ohead->oh_clientid != XFS_TRANSACTION &&
4499	    ohead->oh_clientid != XFS_LOG) {
4500		xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
4501			__func__, ohead->oh_clientid);
4502		ASSERT(0);
4503		return -EIO;
4504	}
4505
4506	/*
4507	 * Check the ophdr contains all the data it is supposed to contain.
4508	 */
4509	len = be32_to_cpu(ohead->oh_len);
4510	if (dp + len > end) {
4511		xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
4512		WARN_ON(1);
4513		return -EIO;
4514	}
4515
4516	trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
4517	if (!trans) {
4518		/* nothing to do, so skip over this ophdr */
4519		return 0;
4520	}
4521
4522	/*
4523	 * The recovered buffer queue is drained only once we know that all
4524	 * recovery items for the current LSN have been processed. This is
4525	 * required because:
4526	 *
4527	 * - Buffer write submission updates the metadata LSN of the buffer.
4528	 * - Log recovery skips items with a metadata LSN >= the current LSN of
4529	 *   the recovery item.
4530	 * - Separate recovery items against the same metadata buffer can share
4531	 *   a current LSN. I.e., consider that the LSN of a recovery item is
4532	 *   defined as the starting LSN of the first record in which its
4533	 *   transaction appears, that a record can hold multiple transactions,
4534	 *   and/or that a transaction can span multiple records.
4535	 *
4536	 * In other words, we are allowed to submit a buffer from log recovery
4537	 * once per current LSN. Otherwise, we may incorrectly skip recovery
4538	 * items and cause corruption.
4539	 *
4540	 * We don't know up front whether buffers are updated multiple times per
4541	 * LSN. Therefore, track the current LSN of each commit log record as it
4542	 * is processed and drain the queue when it changes. Use commit records
4543	 * because they are ordered correctly by the logging code.
4544	 */
4545	if (log->l_recovery_lsn != trans->r_lsn &&
4546	    ohead->oh_flags & XLOG_COMMIT_TRANS) {
4547		error = xfs_buf_delwri_submit(buffer_list);
4548		if (error)
4549			return error;
4550		log->l_recovery_lsn = trans->r_lsn;
4551	}
4552
4553	return xlog_recovery_process_trans(log, trans, dp, len,
4554					   ohead->oh_flags, pass, buffer_list);
4555}
4556
4557/*
4558 * There are two valid states of the r_state field.  0 indicates that the
4559 * transaction structure is in a normal state.  We have either seen the
4560 * start of the transaction or the last operation we added was not a partial
4561 * operation.  If the last operation we added to the transaction was a
4562 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
4563 *
4564 * NOTE: skip LRs with 0 data length.
4565 */
4566STATIC int
4567xlog_recover_process_data(
4568	struct xlog		*log,
4569	struct hlist_head	rhash[],
4570	struct xlog_rec_header	*rhead,
4571	char			*dp,
4572	int			pass,
4573	struct list_head	*buffer_list)
4574{
4575	struct xlog_op_header	*ohead;
4576	char			*end;
4577	int			num_logops;
4578	int			error;
4579
4580	end = dp + be32_to_cpu(rhead->h_len);
4581	num_logops = be32_to_cpu(rhead->h_num_logops);
4582
4583	/* check the log format matches our own - else we can't recover */
4584	if (xlog_header_check_recover(log->l_mp, rhead))
4585		return -EIO;
4586
4587	trace_xfs_log_recover_record(log, rhead, pass);
4588	while ((dp < end) && num_logops) {
4589
4590		ohead = (struct xlog_op_header *)dp;
4591		dp += sizeof(*ohead);
4592		ASSERT(dp <= end);
4593
4594		/* errors will abort recovery */
4595		error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
4596						   dp, end, pass, buffer_list);
4597		if (error)
4598			return error;
4599
4600		dp += be32_to_cpu(ohead->oh_len);
4601		num_logops--;
4602	}
4603	return 0;
4604}
4605
4606/* Recover the EFI if necessary. */
4607STATIC int
4608xlog_recover_process_efi(
4609	struct xfs_mount		*mp,
4610	struct xfs_ail			*ailp,
4611	struct xfs_log_item		*lip)
4612{
4613	struct xfs_efi_log_item		*efip;
4614	int				error;
4615
4616	/*
4617	 * Skip EFIs that we've already processed.
4618	 */
4619	efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4620	if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags))
4621		return 0;
4622
4623	spin_unlock(&ailp->xa_lock);
4624	error = xfs_efi_recover(mp, efip);
4625	spin_lock(&ailp->xa_lock);
4626
4627	return error;
4628}
4629
4630/* Release the EFI since we're cancelling everything. */
4631STATIC void
4632xlog_recover_cancel_efi(
4633	struct xfs_mount		*mp,
4634	struct xfs_ail			*ailp,
4635	struct xfs_log_item		*lip)
4636{
4637	struct xfs_efi_log_item		*efip;
4638
4639	efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4640
4641	spin_unlock(&ailp->xa_lock);
4642	xfs_efi_release(efip);
4643	spin_lock(&ailp->xa_lock);
4644}
4645
4646/* Recover the RUI if necessary. */
4647STATIC int
4648xlog_recover_process_rui(
4649	struct xfs_mount		*mp,
4650	struct xfs_ail			*ailp,
4651	struct xfs_log_item		*lip)
4652{
4653	struct xfs_rui_log_item		*ruip;
4654	int				error;
4655
4656	/*
4657	 * Skip RUIs that we've already processed.
4658	 */
4659	ruip = container_of(lip, struct xfs_rui_log_item, rui_item);
4660	if (test_bit(XFS_RUI_RECOVERED, &ruip->rui_flags))
4661		return 0;
4662
4663	spin_unlock(&ailp->xa_lock);
4664	error = xfs_rui_recover(mp, ruip);
4665	spin_lock(&ailp->xa_lock);
4666
4667	return error;
4668}
4669
4670/* Release the RUI since we're cancelling everything. */
4671STATIC void
4672xlog_recover_cancel_rui(
4673	struct xfs_mount		*mp,
4674	struct xfs_ail			*ailp,
4675	struct xfs_log_item		*lip)
4676{
4677	struct xfs_rui_log_item		*ruip;
4678
4679	ruip = container_of(lip, struct xfs_rui_log_item, rui_item);
4680
4681	spin_unlock(&ailp->xa_lock);
4682	xfs_rui_release(ruip);
4683	spin_lock(&ailp->xa_lock);
4684}
4685
4686/* Recover the CUI if necessary. */
4687STATIC int
4688xlog_recover_process_cui(
4689	struct xfs_mount		*mp,
4690	struct xfs_ail			*ailp,
4691	struct xfs_log_item		*lip)
4692{
4693	struct xfs_cui_log_item		*cuip;
4694	int				error;
4695
4696	/*
4697	 * Skip CUIs that we've already processed.
4698	 */
4699	cuip = container_of(lip, struct xfs_cui_log_item, cui_item);
4700	if (test_bit(XFS_CUI_RECOVERED, &cuip->cui_flags))
4701		return 0;
4702
4703	spin_unlock(&ailp->xa_lock);
4704	error = xfs_cui_recover(mp, cuip);
4705	spin_lock(&ailp->xa_lock);
4706
4707	return error;
4708}
4709
4710/* Release the CUI since we're cancelling everything. */
4711STATIC void
4712xlog_recover_cancel_cui(
4713	struct xfs_mount		*mp,
4714	struct xfs_ail			*ailp,
4715	struct xfs_log_item		*lip)
4716{
4717	struct xfs_cui_log_item		*cuip;
4718
4719	cuip = container_of(lip, struct xfs_cui_log_item, cui_item);
4720
4721	spin_unlock(&ailp->xa_lock);
4722	xfs_cui_release(cuip);
4723	spin_lock(&ailp->xa_lock);
4724}
4725
4726/* Recover the BUI if necessary. */
4727STATIC int
4728xlog_recover_process_bui(
4729	struct xfs_mount		*mp,
4730	struct xfs_ail			*ailp,
4731	struct xfs_log_item		*lip)
4732{
4733	struct xfs_bui_log_item		*buip;
4734	int				error;
4735
4736	/*
4737	 * Skip BUIs that we've already processed.
4738	 */
4739	buip = container_of(lip, struct xfs_bui_log_item, bui_item);
4740	if (test_bit(XFS_BUI_RECOVERED, &buip->bui_flags))
4741		return 0;
4742
4743	spin_unlock(&ailp->xa_lock);
4744	error = xfs_bui_recover(mp, buip);
4745	spin_lock(&ailp->xa_lock);
4746
4747	return error;
4748}
4749
4750/* Release the BUI since we're cancelling everything. */
4751STATIC void
4752xlog_recover_cancel_bui(
4753	struct xfs_mount		*mp,
4754	struct xfs_ail			*ailp,
4755	struct xfs_log_item		*lip)
4756{
4757	struct xfs_bui_log_item		*buip;
4758
4759	buip = container_of(lip, struct xfs_bui_log_item, bui_item);
4760
4761	spin_unlock(&ailp->xa_lock);
4762	xfs_bui_release(buip);
4763	spin_lock(&ailp->xa_lock);
4764}
4765
4766/* Is this log item a deferred action intent? */
4767static inline bool xlog_item_is_intent(struct xfs_log_item *lip)
4768{
4769	switch (lip->li_type) {
4770	case XFS_LI_EFI:
4771	case XFS_LI_RUI:
4772	case XFS_LI_CUI:
4773	case XFS_LI_BUI:
4774		return true;
4775	default:
4776		return false;
4777	}
4778}
4779
4780/*
4781 * When this is called, all of the log intent items which did not have
4782 * corresponding log done items should be in the AIL.  What we do now
4783 * is update the data structures associated with each one.
4784 *
4785 * Since we process the log intent items in normal transactions, they
4786 * will be removed at some point after the commit.  This prevents us
4787 * from just walking down the list processing each one.  We'll use a
4788 * flag in the intent item to skip those that we've already processed
4789 * and use the AIL iteration mechanism's generation count to try to
4790 * speed this up at least a bit.
4791 *
4792 * When we start, we know that the intents are the only things in the
4793 * AIL.  As we process them, however, other items are added to the
4794 * AIL.
4795 */
4796STATIC int
4797xlog_recover_process_intents(
4798	struct xlog		*log)
4799{
4800	struct xfs_log_item	*lip;
4801	int			error = 0;
4802	struct xfs_ail_cursor	cur;
4803	struct xfs_ail		*ailp;
4804	xfs_lsn_t		last_lsn;
4805
4806	ailp = log->l_ailp;
4807	spin_lock(&ailp->xa_lock);
4808	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4809	last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
4810	while (lip != NULL) {
4811		/*
4812		 * We're done when we see something other than an intent.
4813		 * There should be no intents left in the AIL now.
4814		 */
4815		if (!xlog_item_is_intent(lip)) {
4816#ifdef DEBUG
4817			for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4818				ASSERT(!xlog_item_is_intent(lip));
4819#endif
4820			break;
4821		}
4822
4823		/*
4824		 * We should never see a redo item with a LSN higher than
4825		 * the last transaction we found in the log at the start
4826		 * of recovery.
4827		 */
4828		ASSERT(XFS_LSN_CMP(last_lsn, lip->li_lsn) >= 0);
4829
4830		switch (lip->li_type) {
4831		case XFS_LI_EFI:
4832			error = xlog_recover_process_efi(log->l_mp, ailp, lip);
4833			break;
4834		case XFS_LI_RUI:
4835			error = xlog_recover_process_rui(log->l_mp, ailp, lip);
4836			break;
4837		case XFS_LI_CUI:
4838			error = xlog_recover_process_cui(log->l_mp, ailp, lip);
4839			break;
4840		case XFS_LI_BUI:
4841			error = xlog_recover_process_bui(log->l_mp, ailp, lip);
4842			break;
4843		}
4844		if (error)
4845			goto out;
4846		lip = xfs_trans_ail_cursor_next(ailp, &cur);
4847	}
4848out:
4849	xfs_trans_ail_cursor_done(&cur);
4850	spin_unlock(&ailp->xa_lock);
4851	return error;
4852}
4853
4854/*
4855 * A cancel occurs when the mount has failed and we're bailing out.
4856 * Release all pending log intent items so they don't pin the AIL.
4857 */
4858STATIC int
4859xlog_recover_cancel_intents(
4860	struct xlog		*log)
4861{
4862	struct xfs_log_item	*lip;
4863	int			error = 0;
4864	struct xfs_ail_cursor	cur;
4865	struct xfs_ail		*ailp;
4866
4867	ailp = log->l_ailp;
4868	spin_lock(&ailp->xa_lock);
4869	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4870	while (lip != NULL) {
4871		/*
4872		 * We're done when we see something other than an intent.
4873		 * There should be no intents left in the AIL now.
4874		 */
4875		if (!xlog_item_is_intent(lip)) {
4876#ifdef DEBUG
4877			for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4878				ASSERT(!xlog_item_is_intent(lip));
4879#endif
4880			break;
4881		}
4882
4883		switch (lip->li_type) {
4884		case XFS_LI_EFI:
4885			xlog_recover_cancel_efi(log->l_mp, ailp, lip);
4886			break;
4887		case XFS_LI_RUI:
4888			xlog_recover_cancel_rui(log->l_mp, ailp, lip);
4889			break;
4890		case XFS_LI_CUI:
4891			xlog_recover_cancel_cui(log->l_mp, ailp, lip);
4892			break;
4893		case XFS_LI_BUI:
4894			xlog_recover_cancel_bui(log->l_mp, ailp, lip);
4895			break;
4896		}
4897
4898		lip = xfs_trans_ail_cursor_next(ailp, &cur);
4899	}
4900
4901	xfs_trans_ail_cursor_done(&cur);
4902	spin_unlock(&ailp->xa_lock);
4903	return error;
4904}
4905
4906/*
4907 * This routine performs a transaction to null out a bad inode pointer
4908 * in an agi unlinked inode hash bucket.
4909 */
4910STATIC void
4911xlog_recover_clear_agi_bucket(
4912	xfs_mount_t	*mp,
4913	xfs_agnumber_t	agno,
4914	int		bucket)
4915{
4916	xfs_trans_t	*tp;
4917	xfs_agi_t	*agi;
4918	xfs_buf_t	*agibp;
4919	int		offset;
4920	int		error;
4921
4922	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
4923	if (error)
4924		goto out_error;
4925
4926	error = xfs_read_agi(mp, tp, agno, &agibp);
4927	if (error)
4928		goto out_abort;
4929
4930	agi = XFS_BUF_TO_AGI(agibp);
4931	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
4932	offset = offsetof(xfs_agi_t, agi_unlinked) +
4933		 (sizeof(xfs_agino_t) * bucket);
4934	xfs_trans_log_buf(tp, agibp, offset,
4935			  (offset + sizeof(xfs_agino_t) - 1));
4936
4937	error = xfs_trans_commit(tp);
4938	if (error)
4939		goto out_error;
4940	return;
4941
4942out_abort:
4943	xfs_trans_cancel(tp);
4944out_error:
4945	xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
4946	return;
4947}
4948
4949STATIC xfs_agino_t
4950xlog_recover_process_one_iunlink(
4951	struct xfs_mount		*mp,
4952	xfs_agnumber_t			agno,
4953	xfs_agino_t			agino,
4954	int				bucket)
4955{
4956	struct xfs_buf			*ibp;
4957	struct xfs_dinode		*dip;
4958	struct xfs_inode		*ip;
4959	xfs_ino_t			ino;
4960	int				error;
4961
4962	ino = XFS_AGINO_TO_INO(mp, agno, agino);
4963	error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
4964	if (error)
4965		goto fail;
4966
4967	/*
4968	 * Get the on disk inode to find the next inode in the bucket.
4969	 */
4970	error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0, 0);
4971	if (error)
4972		goto fail_iput;
4973
4974	xfs_iflags_clear(ip, XFS_IRECOVERY);
4975	ASSERT(VFS_I(ip)->i_nlink == 0);
4976	ASSERT(VFS_I(ip)->i_mode != 0);
4977
4978	/* setup for the next pass */
4979	agino = be32_to_cpu(dip->di_next_unlinked);
4980	xfs_buf_relse(ibp);
4981
4982	/*
4983	 * Prevent any DMAPI event from being sent when the reference on
4984	 * the inode is dropped.
4985	 */
4986	ip->i_d.di_dmevmask = 0;
4987
4988	IRELE(ip);
4989	return agino;
4990
4991 fail_iput:
4992	IRELE(ip);
4993 fail:
4994	/*
4995	 * We can't read in the inode this bucket points to, or this inode
4996	 * is messed up.  Just ditch this bucket of inodes.  We will lose
4997	 * some inodes and space, but at least we won't hang.
4998	 *
4999	 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
5000	 * clear the inode pointer in the bucket.
5001	 */
5002	xlog_recover_clear_agi_bucket(mp, agno, bucket);
5003	return NULLAGINO;
5004}
5005
5006/*
5007 * xlog_iunlink_recover
5008 *
5009 * This is called during recovery to process any inodes which
5010 * we unlinked but not freed when the system crashed.  These
5011 * inodes will be on the lists in the AGI blocks.  What we do
5012 * here is scan all the AGIs and fully truncate and free any
5013 * inodes found on the lists.  Each inode is removed from the
5014 * lists when it has been fully truncated and is freed.  The
5015 * freeing of the inode and its removal from the list must be
5016 * atomic.
5017 */
5018STATIC void
5019xlog_recover_process_iunlinks(
5020	struct xlog	*log)
5021{
5022	xfs_mount_t	*mp;
5023	xfs_agnumber_t	agno;
5024	xfs_agi_t	*agi;
5025	xfs_buf_t	*agibp;
5026	xfs_agino_t	agino;
5027	int		bucket;
5028	int		error;
5029	uint		mp_dmevmask;
5030
5031	mp = log->l_mp;
5032
5033	/*
5034	 * Prevent any DMAPI event from being sent while in this function.
5035	 */
5036	mp_dmevmask = mp->m_dmevmask;
5037	mp->m_dmevmask = 0;
5038
5039	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5040		/*
5041		 * Find the agi for this ag.
5042		 */
5043		error = xfs_read_agi(mp, NULL, agno, &agibp);
5044		if (error) {
5045			/*
5046			 * AGI is b0rked. Don't process it.
5047			 *
5048			 * We should probably mark the filesystem as corrupt
5049			 * after we've recovered all the ag's we can....
5050			 */
5051			continue;
5052		}
5053		/*
5054		 * Unlock the buffer so that it can be acquired in the normal
5055		 * course of the transaction to truncate and free each inode.
5056		 * Because we are not racing with anyone else here for the AGI
5057		 * buffer, we don't even need to hold it locked to read the
5058		 * initial unlinked bucket entries out of the buffer. We keep
5059		 * buffer reference though, so that it stays pinned in memory
5060		 * while we need the buffer.
5061		 */
5062		agi = XFS_BUF_TO_AGI(agibp);
5063		xfs_buf_unlock(agibp);
5064
5065		for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
5066			agino = be32_to_cpu(agi->agi_unlinked[bucket]);
5067			while (agino != NULLAGINO) {
5068				agino = xlog_recover_process_one_iunlink(mp,
5069							agno, agino, bucket);
5070			}
5071		}
5072		xfs_buf_rele(agibp);
5073	}
5074
5075	mp->m_dmevmask = mp_dmevmask;
5076}
5077
5078STATIC int
5079xlog_unpack_data(
5080	struct xlog_rec_header	*rhead,
5081	char			*dp,
5082	struct xlog		*log)
5083{
5084	int			i, j, k;
5085
5086	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
5087		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
5088		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
5089		dp += BBSIZE;
5090	}
5091
5092	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
5093		xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
5094		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
5095			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
5096			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
5097			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
5098			dp += BBSIZE;
5099		}
5100	}
5101
5102	return 0;
5103}
5104
5105/*
5106 * CRC check, unpack and process a log record.
5107 */
5108STATIC int
5109xlog_recover_process(
5110	struct xlog		*log,
5111	struct hlist_head	rhash[],
5112	struct xlog_rec_header	*rhead,
5113	char			*dp,
5114	int			pass,
5115	struct list_head	*buffer_list)
5116{
5117	int			error;
5118	__le32			old_crc = rhead->h_crc;
5119	__le32			crc;
5120
5121
5122	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
5123
5124	/*
5125	 * Nothing else to do if this is a CRC verification pass. Just return
5126	 * if this a record with a non-zero crc. Unfortunately, mkfs always
5127	 * sets old_crc to 0 so we must consider this valid even on v5 supers.
5128	 * Otherwise, return EFSBADCRC on failure so the callers up the stack
5129	 * know precisely what failed.
5130	 */
5131	if (pass == XLOG_RECOVER_CRCPASS) {
5132		if (old_crc && crc != old_crc)
5133			return -EFSBADCRC;
5134		return 0;
5135	}
5136
5137	/*
5138	 * We're in the normal recovery path. Issue a warning if and only if the
5139	 * CRC in the header is non-zero. This is an advisory warning and the
5140	 * zero CRC check prevents warnings from being emitted when upgrading
5141	 * the kernel from one that does not add CRCs by default.
5142	 */
5143	if (crc != old_crc) {
5144		if (old_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
5145			xfs_alert(log->l_mp,
5146		"log record CRC mismatch: found 0x%x, expected 0x%x.",
5147					le32_to_cpu(old_crc),
5148					le32_to_cpu(crc));
5149			xfs_hex_dump(dp, 32);
5150		}
5151
5152		/*
5153		 * If the filesystem is CRC enabled, this mismatch becomes a
5154		 * fatal log corruption failure.
5155		 */
5156		if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
5157			return -EFSCORRUPTED;
5158	}
5159
5160	error = xlog_unpack_data(rhead, dp, log);
5161	if (error)
5162		return error;
5163
5164	return xlog_recover_process_data(log, rhash, rhead, dp, pass,
5165					 buffer_list);
5166}
5167
5168STATIC int
5169xlog_valid_rec_header(
5170	struct xlog		*log,
5171	struct xlog_rec_header	*rhead,
5172	xfs_daddr_t		blkno)
5173{
5174	int			hlen;
5175
5176	if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
5177		XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
5178				XFS_ERRLEVEL_LOW, log->l_mp);
5179		return -EFSCORRUPTED;
5180	}
5181	if (unlikely(
5182	    (!rhead->h_version ||
5183	    (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
5184		xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
5185			__func__, be32_to_cpu(rhead->h_version));
5186		return -EIO;
5187	}
5188
5189	/* LR body must have data or it wouldn't have been written */
5190	hlen = be32_to_cpu(rhead->h_len);
5191	if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
5192		XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
5193				XFS_ERRLEVEL_LOW, log->l_mp);
5194		return -EFSCORRUPTED;
5195	}
5196	if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
5197		XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
5198				XFS_ERRLEVEL_LOW, log->l_mp);
5199		return -EFSCORRUPTED;
5200	}
5201	return 0;
5202}
5203
5204/*
5205 * Read the log from tail to head and process the log records found.
5206 * Handle the two cases where the tail and head are in the same cycle
5207 * and where the active portion of the log wraps around the end of
5208 * the physical log separately.  The pass parameter is passed through
5209 * to the routines called to process the data and is not looked at
5210 * here.
5211 */
5212STATIC int
5213xlog_do_recovery_pass(
5214	struct xlog		*log,
5215	xfs_daddr_t		head_blk,
5216	xfs_daddr_t		tail_blk,
5217	int			pass,
5218	xfs_daddr_t		*first_bad)	/* out: first bad log rec */
5219{
5220	xlog_rec_header_t	*rhead;
5221	xfs_daddr_t		blk_no;
5222	xfs_daddr_t		rhead_blk;
5223	char			*offset;
5224	xfs_buf_t		*hbp, *dbp;
5225	int			error = 0, h_size, h_len;
5226	int			error2 = 0;
5227	int			bblks, split_bblks;
5228	int			hblks, split_hblks, wrapped_hblks;
5229	int			i;
5230	struct hlist_head	rhash[XLOG_RHASH_SIZE];
5231	LIST_HEAD		(buffer_list);
5232
5233	ASSERT(head_blk != tail_blk);
5234	rhead_blk = 0;
5235
5236	for (i = 0; i < XLOG_RHASH_SIZE; i++)
5237		INIT_HLIST_HEAD(&rhash[i]);
5238
5239	/*
5240	 * Read the header of the tail block and get the iclog buffer size from
5241	 * h_size.  Use this to tell how many sectors make up the log header.
5242	 */
5243	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
5244		/*
5245		 * When using variable length iclogs, read first sector of
5246		 * iclog header and extract the header size from it.  Get a
5247		 * new hbp that is the correct size.
5248		 */
5249		hbp = xlog_get_bp(log, 1);
5250		if (!hbp)
5251			return -ENOMEM;
5252
5253		error = xlog_bread(log, tail_blk, 1, hbp, &offset);
5254		if (error)
5255			goto bread_err1;
5256
5257		rhead = (xlog_rec_header_t *)offset;
5258		error = xlog_valid_rec_header(log, rhead, tail_blk);
5259		if (error)
5260			goto bread_err1;
5261
5262		/*
5263		 * xfsprogs has a bug where record length is based on lsunit but
5264		 * h_size (iclog size) is hardcoded to 32k. Now that we
5265		 * unconditionally CRC verify the unmount record, this means the
5266		 * log buffer can be too small for the record and cause an
5267		 * overrun.
5268		 *
5269		 * Detect this condition here. Use lsunit for the buffer size as
5270		 * long as this looks like the mkfs case. Otherwise, return an
5271		 * error to avoid a buffer overrun.
5272		 */
5273		h_size = be32_to_cpu(rhead->h_size);
5274		h_len = be32_to_cpu(rhead->h_len);
5275		if (h_len > h_size) {
5276			if (h_len <= log->l_mp->m_logbsize &&
5277			    be32_to_cpu(rhead->h_num_logops) == 1) {
5278				xfs_warn(log->l_mp,
5279		"invalid iclog size (%d bytes), using lsunit (%d bytes)",
5280					 h_size, log->l_mp->m_logbsize);
5281				h_size = log->l_mp->m_logbsize;
5282			} else
5283				return -EFSCORRUPTED;
5284		}
5285
5286		if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
5287		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
5288			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
5289			if (h_size % XLOG_HEADER_CYCLE_SIZE)
5290				hblks++;
5291			xlog_put_bp(hbp);
5292			hbp = xlog_get_bp(log, hblks);
5293		} else {
5294			hblks = 1;
5295		}
5296	} else {
5297		ASSERT(log->l_sectBBsize == 1);
5298		hblks = 1;
5299		hbp = xlog_get_bp(log, 1);
5300		h_size = XLOG_BIG_RECORD_BSIZE;
5301	}
5302
5303	if (!hbp)
5304		return -ENOMEM;
5305	dbp = xlog_get_bp(log, BTOBB(h_size));
5306	if (!dbp) {
5307		xlog_put_bp(hbp);
5308		return -ENOMEM;
5309	}
5310
5311	memset(rhash, 0, sizeof(rhash));
5312	blk_no = rhead_blk = tail_blk;
5313	if (tail_blk > head_blk) {
5314		/*
5315		 * Perform recovery around the end of the physical log.
5316		 * When the head is not on the same cycle number as the tail,
5317		 * we can't do a sequential recovery.
5318		 */
5319		while (blk_no < log->l_logBBsize) {
5320			/*
5321			 * Check for header wrapping around physical end-of-log
5322			 */
5323			offset = hbp->b_addr;
5324			split_hblks = 0;
5325			wrapped_hblks = 0;
5326			if (blk_no + hblks <= log->l_logBBsize) {
5327				/* Read header in one read */
5328				error = xlog_bread(log, blk_no, hblks, hbp,
5329						   &offset);
5330				if (error)
5331					goto bread_err2;
5332			} else {
5333				/* This LR is split across physical log end */
5334				if (blk_no != log->l_logBBsize) {
5335					/* some data before physical log end */
5336					ASSERT(blk_no <= INT_MAX);
5337					split_hblks = log->l_logBBsize - (int)blk_no;
5338					ASSERT(split_hblks > 0);
5339					error = xlog_bread(log, blk_no,
5340							   split_hblks, hbp,
5341							   &offset);
5342					if (error)
5343						goto bread_err2;
5344				}
5345
5346				/*
5347				 * Note: this black magic still works with
5348				 * large sector sizes (non-512) only because:
5349				 * - we increased the buffer size originally
5350				 *   by 1 sector giving us enough extra space
5351				 *   for the second read;
5352				 * - the log start is guaranteed to be sector
5353				 *   aligned;
5354				 * - we read the log end (LR header start)
5355				 *   _first_, then the log start (LR header end)
5356				 *   - order is important.
5357				 */
5358				wrapped_hblks = hblks - split_hblks;
5359				error = xlog_bread_offset(log, 0,
5360						wrapped_hblks, hbp,
5361						offset + BBTOB(split_hblks));
5362				if (error)
5363					goto bread_err2;
5364			}
5365			rhead = (xlog_rec_header_t *)offset;
5366			error = xlog_valid_rec_header(log, rhead,
5367						split_hblks ? blk_no : 0);
5368			if (error)
5369				goto bread_err2;
5370
5371			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
5372			blk_no += hblks;
5373
5374			/* Read in data for log record */
5375			if (blk_no + bblks <= log->l_logBBsize) {
5376				error = xlog_bread(log, blk_no, bblks, dbp,
5377						   &offset);
5378				if (error)
5379					goto bread_err2;
5380			} else {
5381				/* This log record is split across the
5382				 * physical end of log */
5383				offset = dbp->b_addr;
5384				split_bblks = 0;
5385				if (blk_no != log->l_logBBsize) {
5386					/* some data is before the physical
5387					 * end of log */
5388					ASSERT(!wrapped_hblks);
5389					ASSERT(blk_no <= INT_MAX);
5390					split_bblks =
5391						log->l_logBBsize - (int)blk_no;
5392					ASSERT(split_bblks > 0);
5393					error = xlog_bread(log, blk_no,
5394							split_bblks, dbp,
5395							&offset);
5396					if (error)
5397						goto bread_err2;
5398				}
5399
5400				/*
5401				 * Note: this black magic still works with
5402				 * large sector sizes (non-512) only because:
5403				 * - we increased the buffer size originally
5404				 *   by 1 sector giving us enough extra space
5405				 *   for the second read;
5406				 * - the log start is guaranteed to be sector
5407				 *   aligned;
5408				 * - we read the log end (LR header start)
5409				 *   _first_, then the log start (LR header end)
5410				 *   - order is important.
5411				 */
5412				error = xlog_bread_offset(log, 0,
5413						bblks - split_bblks, dbp,
5414						offset + BBTOB(split_bblks));
5415				if (error)
5416					goto bread_err2;
5417			}
5418
5419			error = xlog_recover_process(log, rhash, rhead, offset,
5420						     pass, &buffer_list);
5421			if (error)
5422				goto bread_err2;
5423
5424			blk_no += bblks;
5425			rhead_blk = blk_no;
5426		}
5427
5428		ASSERT(blk_no >= log->l_logBBsize);
5429		blk_no -= log->l_logBBsize;
5430		rhead_blk = blk_no;
5431	}
5432
5433	/* read first part of physical log */
5434	while (blk_no < head_blk) {
5435		error = xlog_bread(log, blk_no, hblks, hbp, &offset);
5436		if (error)
5437			goto bread_err2;
5438
5439		rhead = (xlog_rec_header_t *)offset;
5440		error = xlog_valid_rec_header(log, rhead, blk_no);
5441		if (error)
5442			goto bread_err2;
5443
5444		/* blocks in data section */
5445		bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
5446		error = xlog_bread(log, blk_no+hblks, bblks, dbp,
5447				   &offset);
5448		if (error)
5449			goto bread_err2;
5450
5451		error = xlog_recover_process(log, rhash, rhead, offset, pass,
5452					     &buffer_list);
5453		if (error)
5454			goto bread_err2;
5455
5456		blk_no += bblks + hblks;
5457		rhead_blk = blk_no;
5458	}
5459
5460 bread_err2:
5461	xlog_put_bp(dbp);
5462 bread_err1:
5463	xlog_put_bp(hbp);
5464
5465	/*
5466	 * Submit buffers that have been added from the last record processed,
5467	 * regardless of error status.
5468	 */
5469	if (!list_empty(&buffer_list))
5470		error2 = xfs_buf_delwri_submit(&buffer_list);
5471
5472	if (error && first_bad)
5473		*first_bad = rhead_blk;
5474
5475	/*
5476	 * Transactions are freed at commit time but transactions without commit
5477	 * records on disk are never committed. Free any that may be left in the
5478	 * hash table.
5479	 */
5480	for (i = 0; i < XLOG_RHASH_SIZE; i++) {
5481		struct hlist_node	*tmp;
5482		struct xlog_recover	*trans;
5483
5484		hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
5485			xlog_recover_free_trans(trans);
5486	}
5487
5488	return error ? error : error2;
5489}
5490
5491/*
5492 * Do the recovery of the log.  We actually do this in two phases.
5493 * The two passes are necessary in order to implement the function
5494 * of cancelling a record written into the log.  The first pass
5495 * determines those things which have been cancelled, and the
5496 * second pass replays log items normally except for those which
5497 * have been cancelled.  The handling of the replay and cancellations
5498 * takes place in the log item type specific routines.
5499 *
5500 * The table of items which have cancel records in the log is allocated
5501 * and freed at this level, since only here do we know when all of
5502 * the log recovery has been completed.
5503 */
5504STATIC int
5505xlog_do_log_recovery(
5506	struct xlog	*log,
5507	xfs_daddr_t	head_blk,
5508	xfs_daddr_t	tail_blk)
5509{
5510	int		error, i;
5511
5512	ASSERT(head_blk != tail_blk);
5513
5514	/*
5515	 * First do a pass to find all of the cancelled buf log items.
5516	 * Store them in the buf_cancel_table for use in the second pass.
5517	 */
5518	log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
5519						 sizeof(struct list_head),
5520						 KM_SLEEP);
5521	for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
5522		INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
5523
5524	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
5525				      XLOG_RECOVER_PASS1, NULL);
5526	if (error != 0) {
5527		kmem_free(log->l_buf_cancel_table);
5528		log->l_buf_cancel_table = NULL;
5529		return error;
5530	}
5531	/*
5532	 * Then do a second pass to actually recover the items in the log.
5533	 * When it is complete free the table of buf cancel items.
5534	 */
5535	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
5536				      XLOG_RECOVER_PASS2, NULL);
5537#ifdef DEBUG
5538	if (!error) {
5539		int	i;
5540
5541		for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
5542			ASSERT(list_empty(&log->l_buf_cancel_table[i]));
5543	}
5544#endif	/* DEBUG */
5545
5546	kmem_free(log->l_buf_cancel_table);
5547	log->l_buf_cancel_table = NULL;
5548
5549	return error;
5550}
5551
5552/*
5553 * Do the actual recovery
5554 */
5555STATIC int
5556xlog_do_recover(
5557	struct xlog	*log,
5558	xfs_daddr_t	head_blk,
5559	xfs_daddr_t	tail_blk)
5560{
5561	struct xfs_mount *mp = log->l_mp;
5562	int		error;
5563	xfs_buf_t	*bp;
5564	xfs_sb_t	*sbp;
5565
5566	/*
5567	 * First replay the images in the log.
5568	 */
5569	error = xlog_do_log_recovery(log, head_blk, tail_blk);
5570	if (error)
5571		return error;
5572
5573	/*
5574	 * If IO errors happened during recovery, bail out.
5575	 */
5576	if (XFS_FORCED_SHUTDOWN(mp)) {
5577		return -EIO;
5578	}
5579
5580	/*
5581	 * We now update the tail_lsn since much of the recovery has completed
5582	 * and there may be space available to use.  If there were no extent
5583	 * or iunlinks, we can free up the entire log and set the tail_lsn to
5584	 * be the last_sync_lsn.  This was set in xlog_find_tail to be the
5585	 * lsn of the last known good LR on disk.  If there are extent frees
5586	 * or iunlinks they will have some entries in the AIL; so we look at
5587	 * the AIL to determine how to set the tail_lsn.
5588	 */
5589	xlog_assign_tail_lsn(mp);
5590
5591	/*
5592	 * Now that we've finished replaying all buffer and inode
5593	 * updates, re-read in the superblock and reverify it.
5594	 */
5595	bp = xfs_getsb(mp, 0);
5596	bp->b_flags &= ~(XBF_DONE | XBF_ASYNC);
5597	ASSERT(!(bp->b_flags & XBF_WRITE));
5598	bp->b_flags |= XBF_READ;
5599	bp->b_ops = &xfs_sb_buf_ops;
5600
5601	error = xfs_buf_submit_wait(bp);
5602	if (error) {
5603		if (!XFS_FORCED_SHUTDOWN(mp)) {
5604			xfs_buf_ioerror_alert(bp, __func__);
5605			ASSERT(0);
5606		}
5607		xfs_buf_relse(bp);
5608		return error;
5609	}
5610
5611	/* Convert superblock from on-disk format */
5612	sbp = &mp->m_sb;
5613	xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
5614	xfs_buf_relse(bp);
5615
5616	/* re-initialise in-core superblock and geometry structures */
5617	xfs_reinit_percpu_counters(mp);
5618	error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
5619	if (error) {
5620		xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
5621		return error;
5622	}
5623	mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
5624
5625	xlog_recover_check_summary(log);
5626
5627	/* Normal transactions can now occur */
5628	log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
5629	return 0;
5630}
5631
5632/*
5633 * Perform recovery and re-initialize some log variables in xlog_find_tail.
5634 *
5635 * Return error or zero.
5636 */
5637int
5638xlog_recover(
5639	struct xlog	*log)
5640{
5641	xfs_daddr_t	head_blk, tail_blk;
5642	int		error;
5643
5644	/* find the tail of the log */
5645	error = xlog_find_tail(log, &head_blk, &tail_blk);
5646	if (error)
5647		return error;
5648
5649	/*
5650	 * The superblock was read before the log was available and thus the LSN
5651	 * could not be verified. Check the superblock LSN against the current
5652	 * LSN now that it's known.
5653	 */
5654	if (xfs_sb_version_hascrc(&log->l_mp->m_sb) &&
5655	    !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
5656		return -EINVAL;
5657
5658	if (tail_blk != head_blk) {
5659		/* There used to be a comment here:
5660		 *
5661		 * disallow recovery on read-only mounts.  note -- mount
5662		 * checks for ENOSPC and turns it into an intelligent
5663		 * error message.
5664		 * ...but this is no longer true.  Now, unless you specify
5665		 * NORECOVERY (in which case this function would never be
5666		 * called), we just go ahead and recover.  We do this all
5667		 * under the vfs layer, so we can get away with it unless
5668		 * the device itself is read-only, in which case we fail.
5669		 */
5670		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
5671			return error;
5672		}
5673
5674		/*
5675		 * Version 5 superblock log feature mask validation. We know the
5676		 * log is dirty so check if there are any unknown log features
5677		 * in what we need to recover. If there are unknown features
5678		 * (e.g. unsupported transactions, then simply reject the
5679		 * attempt at recovery before touching anything.
5680		 */
5681		if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
5682		    xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
5683					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
5684			xfs_warn(log->l_mp,
5685"Superblock has unknown incompatible log features (0x%x) enabled.",
5686				(log->l_mp->m_sb.sb_features_log_incompat &
5687					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
5688			xfs_warn(log->l_mp,
5689"The log can not be fully and/or safely recovered by this kernel.");
5690			xfs_warn(log->l_mp,
5691"Please recover the log on a kernel that supports the unknown features.");
5692			return -EINVAL;
5693		}
5694
5695		/*
5696		 * Delay log recovery if the debug hook is set. This is debug
5697		 * instrumention to coordinate simulation of I/O failures with
5698		 * log recovery.
5699		 */
5700		if (xfs_globals.log_recovery_delay) {
5701			xfs_notice(log->l_mp,
5702				"Delaying log recovery for %d seconds.",
5703				xfs_globals.log_recovery_delay);
5704			msleep(xfs_globals.log_recovery_delay * 1000);
5705		}
5706
5707		xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
5708				log->l_mp->m_logname ? log->l_mp->m_logname
5709						     : "internal");
5710
5711		error = xlog_do_recover(log, head_blk, tail_blk);
5712		log->l_flags |= XLOG_RECOVERY_NEEDED;
5713	}
5714	return error;
5715}
5716
5717/*
5718 * In the first part of recovery we replay inodes and buffers and build
5719 * up the list of extent free items which need to be processed.  Here
5720 * we process the extent free items and clean up the on disk unlinked
5721 * inode lists.  This is separated from the first part of recovery so
5722 * that the root and real-time bitmap inodes can be read in from disk in
5723 * between the two stages.  This is necessary so that we can free space
5724 * in the real-time portion of the file system.
5725 */
5726int
5727xlog_recover_finish(
5728	struct xlog	*log)
5729{
5730	/*
5731	 * Now we're ready to do the transactions needed for the
5732	 * rest of recovery.  Start with completing all the extent
5733	 * free intent records and then process the unlinked inode
5734	 * lists.  At this point, we essentially run in normal mode
5735	 * except that we're still performing recovery actions
5736	 * rather than accepting new requests.
5737	 */
5738	if (log->l_flags & XLOG_RECOVERY_NEEDED) {
5739		int	error;
5740		error = xlog_recover_process_intents(log);
5741		if (error) {
5742			xfs_alert(log->l_mp, "Failed to recover intents");
5743			return error;
5744		}
5745
5746		/*
5747		 * Sync the log to get all the intents out of the AIL.
5748		 * This isn't absolutely necessary, but it helps in
5749		 * case the unlink transactions would have problems
5750		 * pushing the intents out of the way.
5751		 */
5752		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
5753
5754		xlog_recover_process_iunlinks(log);
5755
5756		xlog_recover_check_summary(log);
5757
5758		xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
5759				log->l_mp->m_logname ? log->l_mp->m_logname
5760						     : "internal");
5761		log->l_flags &= ~XLOG_RECOVERY_NEEDED;
5762	} else {
5763		xfs_info(log->l_mp, "Ending clean mount");
5764	}
5765	return 0;
5766}
5767
5768int
5769xlog_recover_cancel(
5770	struct xlog	*log)
5771{
5772	int		error = 0;
5773
5774	if (log->l_flags & XLOG_RECOVERY_NEEDED)
5775		error = xlog_recover_cancel_intents(log);
5776
5777	return error;
5778}
5779
5780#if defined(DEBUG)
5781/*
5782 * Read all of the agf and agi counters and check that they
5783 * are consistent with the superblock counters.
5784 */
5785void
5786xlog_recover_check_summary(
5787	struct xlog	*log)
5788{
5789	xfs_mount_t	*mp;
5790	xfs_agf_t	*agfp;
5791	xfs_buf_t	*agfbp;
5792	xfs_buf_t	*agibp;
5793	xfs_agnumber_t	agno;
5794	uint64_t	freeblks;
5795	uint64_t	itotal;
5796	uint64_t	ifree;
5797	int		error;
5798
5799	mp = log->l_mp;
5800
5801	freeblks = 0LL;
5802	itotal = 0LL;
5803	ifree = 0LL;
5804	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5805		error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
5806		if (error) {
5807			xfs_alert(mp, "%s agf read failed agno %d error %d",
5808						__func__, agno, error);
5809		} else {
5810			agfp = XFS_BUF_TO_AGF(agfbp);
5811			freeblks += be32_to_cpu(agfp->agf_freeblks) +
5812				    be32_to_cpu(agfp->agf_flcount);
5813			xfs_buf_relse(agfbp);
5814		}
5815
5816		error = xfs_read_agi(mp, NULL, agno, &agibp);
5817		if (error) {
5818			xfs_alert(mp, "%s agi read failed agno %d error %d",
5819						__func__, agno, error);
5820		} else {
5821			struct xfs_agi	*agi = XFS_BUF_TO_AGI(agibp);
5822
5823			itotal += be32_to_cpu(agi->agi_count);
5824			ifree += be32_to_cpu(agi->agi_freecount);
5825			xfs_buf_relse(agibp);
5826		}
5827	}
5828}
5829#endif /* DEBUG */