Linux Audio

Check our new training course

Linux BSP upgrade and security maintenance

Need help to get security updates for your Linux BSP?
Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include "xfs.h"
   7#include "xfs_fs.h"
   8#include "xfs_shared.h"
   9#include "xfs_format.h"
  10#include "xfs_log_format.h"
  11#include "xfs_trans_resv.h"
  12#include "xfs_bit.h"
  13#include "xfs_sb.h"
  14#include "xfs_mount.h"
  15#include "xfs_defer.h"
  16#include "xfs_inode.h"
  17#include "xfs_trans.h"
  18#include "xfs_log.h"
  19#include "xfs_log_priv.h"
  20#include "xfs_log_recover.h"
 
 
  21#include "xfs_trans_priv.h"
  22#include "xfs_alloc.h"
  23#include "xfs_ialloc.h"
 
  24#include "xfs_trace.h"
  25#include "xfs_icache.h"
 
  26#include "xfs_error.h"
 
 
  27#include "xfs_buf_item.h"
  28#include "xfs_ag.h"
  29#include "xfs_quota.h"
  30#include "xfs_reflink.h"
  31
  32#define BLK_AVG(blk1, blk2)	((blk1+blk2) >> 1)
  33
  34STATIC int
  35xlog_find_zeroed(
  36	struct xlog	*,
  37	xfs_daddr_t	*);
  38STATIC int
  39xlog_clear_stale_blocks(
  40	struct xlog	*,
  41	xfs_lsn_t);
 
 
 
 
 
 
 
  42STATIC int
  43xlog_do_recovery_pass(
  44        struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
  45
  46/*
 
 
 
 
 
 
 
 
 
 
 
  47 * Sector aligned buffer routines for buffer create/read/write/access
  48 */
  49
  50/*
  51 * Verify the log-relative block number and length in basic blocks are valid for
  52 * an operation involving the given XFS log buffer. Returns true if the fields
  53 * are valid, false otherwise.
  54 */
  55static inline bool
  56xlog_verify_bno(
  57	struct xlog	*log,
  58	xfs_daddr_t	blk_no,
  59	int		bbcount)
  60{
  61	if (blk_no < 0 || blk_no >= log->l_logBBsize)
  62		return false;
  63	if (bbcount <= 0 || (blk_no + bbcount) > log->l_logBBsize)
  64		return false;
  65	return true;
  66}
  67
  68/*
  69 * Allocate a buffer to hold log data.  The buffer needs to be able to map to
  70 * a range of nbblks basic blocks at any valid offset within the log.
  71 */
  72static char *
  73xlog_alloc_buffer(
  74	struct xlog	*log,
  75	int		nbblks)
  76{
 
 
  77	/*
  78	 * Pass log block 0 since we don't have an addr yet, buffer will be
  79	 * verified on read.
  80	 */
  81	if (XFS_IS_CORRUPT(log->l_mp, !xlog_verify_bno(log, 0, nbblks))) {
  82		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
  83			nbblks);
 
  84		return NULL;
  85	}
  86
  87	/*
  88	 * We do log I/O in units of log sectors (a power-of-2 multiple of the
  89	 * basic block size), so we round up the requested size to accommodate
  90	 * the basic blocks required for complete log sectors.
  91	 *
  92	 * In addition, the buffer may be used for a non-sector-aligned block
  93	 * offset, in which case an I/O of the requested size could extend
  94	 * beyond the end of the buffer.  If the requested size is only 1 basic
  95	 * block it will never straddle a sector boundary, so this won't be an
  96	 * issue.  Nor will this be a problem if the log I/O is done in basic
  97	 * blocks (sector size 1).  But otherwise we extend the buffer by one
  98	 * extra log sector to ensure there's space to accommodate this
  99	 * possibility.
 100	 */
 101	if (nbblks > 1 && log->l_sectBBsize > 1)
 102		nbblks += log->l_sectBBsize;
 103	nbblks = round_up(nbblks, log->l_sectBBsize);
 104	return kvzalloc(BBTOB(nbblks), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
 105}
 106
 107/*
 108 * Return the address of the start of the given block number's data
 109 * in a log buffer.  The buffer covers a log sector-aligned region.
 110 */
 111static inline unsigned int
 112xlog_align(
 113	struct xlog	*log,
 114	xfs_daddr_t	blk_no)
 115{
 116	return BBTOB(blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1));
 117}
 118
 119static int
 120xlog_do_io(
 121	struct xlog		*log,
 122	xfs_daddr_t		blk_no,
 123	unsigned int		nbblks,
 124	char			*data,
 125	enum req_op		op)
 126{
 127	int			error;
 128
 129	if (XFS_IS_CORRUPT(log->l_mp, !xlog_verify_bno(log, blk_no, nbblks))) {
 130		xfs_warn(log->l_mp,
 131			 "Invalid log block/length (0x%llx, 0x%x) for buffer",
 132			 blk_no, nbblks);
 
 133		return -EFSCORRUPTED;
 134	}
 135
 136	blk_no = round_down(blk_no, log->l_sectBBsize);
 137	nbblks = round_up(nbblks, log->l_sectBBsize);
 138	ASSERT(nbblks > 0);
 139
 140	error = xfs_rw_bdev(log->l_targ->bt_bdev, log->l_logBBstart + blk_no,
 141			BBTOB(nbblks), data, op);
 142	if (error && !xlog_is_shutdown(log)) {
 143		xfs_alert(log->l_mp,
 144			  "log recovery %s I/O error at daddr 0x%llx len %d error %d",
 145			  op == REQ_OP_WRITE ? "write" : "read",
 146			  blk_no, nbblks, error);
 147	}
 148	return error;
 149}
 150
 151STATIC int
 152xlog_bread_noalign(
 153	struct xlog	*log,
 154	xfs_daddr_t	blk_no,
 155	int		nbblks,
 156	char		*data)
 157{
 158	return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
 159}
 160
 161STATIC int
 162xlog_bread(
 163	struct xlog	*log,
 164	xfs_daddr_t	blk_no,
 165	int		nbblks,
 166	char		*data,
 167	char		**offset)
 168{
 169	int		error;
 170
 171	error = xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
 172	if (!error)
 173		*offset = data + xlog_align(log, blk_no);
 174	return error;
 175}
 176
 177STATIC int
 178xlog_bwrite(
 179	struct xlog	*log,
 180	xfs_daddr_t	blk_no,
 181	int		nbblks,
 182	char		*data)
 183{
 184	return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_WRITE);
 185}
 186
 187#ifdef DEBUG
 188/*
 189 * dump debug superblock and log record information
 190 */
 191STATIC void
 192xlog_header_check_dump(
 193	xfs_mount_t		*mp,
 194	xlog_rec_header_t	*head)
 195{
 196	xfs_debug(mp, "%s:  SB : uuid = %pU, fmt = %d",
 197		__func__, &mp->m_sb.sb_uuid, XLOG_FMT);
 198	xfs_debug(mp, "    log : uuid = %pU, fmt = %d",
 199		&head->h_fs_uuid, be32_to_cpu(head->h_fmt));
 200}
 201#else
 202#define xlog_header_check_dump(mp, head)
 203#endif
 204
 205/*
 206 * check log record header for recovery
 207 */
 208STATIC int
 209xlog_header_check_recover(
 210	xfs_mount_t		*mp,
 211	xlog_rec_header_t	*head)
 212{
 213	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 214
 215	/*
 216	 * IRIX doesn't write the h_fmt field and leaves it zeroed
 217	 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
 218	 * a dirty log created in IRIX.
 219	 */
 220	if (XFS_IS_CORRUPT(mp, head->h_fmt != cpu_to_be32(XLOG_FMT))) {
 221		xfs_warn(mp,
 222	"dirty log written in incompatible format - can't recover");
 223		xlog_header_check_dump(mp, head);
 
 
 224		return -EFSCORRUPTED;
 225	}
 226	if (XFS_IS_CORRUPT(mp, !uuid_equal(&mp->m_sb.sb_uuid,
 227					   &head->h_fs_uuid))) {
 228		xfs_warn(mp,
 229	"dirty log entry has mismatched uuid - can't recover");
 230		xlog_header_check_dump(mp, head);
 
 
 231		return -EFSCORRUPTED;
 232	}
 233	return 0;
 234}
 235
 236/*
 237 * read the head block of the log and check the header
 238 */
 239STATIC int
 240xlog_header_check_mount(
 241	xfs_mount_t		*mp,
 242	xlog_rec_header_t	*head)
 243{
 244	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 245
 246	if (uuid_is_null(&head->h_fs_uuid)) {
 247		/*
 248		 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
 249		 * h_fs_uuid is null, we assume this log was last mounted
 250		 * by IRIX and continue.
 251		 */
 252		xfs_warn(mp, "null uuid in log - IRIX style log");
 253	} else if (XFS_IS_CORRUPT(mp, !uuid_equal(&mp->m_sb.sb_uuid,
 254						  &head->h_fs_uuid))) {
 255		xfs_warn(mp, "log has mismatched uuid - can't recover");
 256		xlog_header_check_dump(mp, head);
 
 
 257		return -EFSCORRUPTED;
 258	}
 259	return 0;
 260}
 261
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 262/*
 263 * This routine finds (to an approximation) the first block in the physical
 264 * log which contains the given cycle.  It uses a binary search algorithm.
 265 * Note that the algorithm can not be perfect because the disk will not
 266 * necessarily be perfect.
 267 */
 268STATIC int
 269xlog_find_cycle_start(
 270	struct xlog	*log,
 271	char		*buffer,
 272	xfs_daddr_t	first_blk,
 273	xfs_daddr_t	*last_blk,
 274	uint		cycle)
 275{
 276	char		*offset;
 277	xfs_daddr_t	mid_blk;
 278	xfs_daddr_t	end_blk;
 279	uint		mid_cycle;
 280	int		error;
 281
 282	end_blk = *last_blk;
 283	mid_blk = BLK_AVG(first_blk, end_blk);
 284	while (mid_blk != first_blk && mid_blk != end_blk) {
 285		error = xlog_bread(log, mid_blk, 1, buffer, &offset);
 286		if (error)
 287			return error;
 288		mid_cycle = xlog_get_cycle(offset);
 289		if (mid_cycle == cycle)
 290			end_blk = mid_blk;   /* last_half_cycle == mid_cycle */
 291		else
 292			first_blk = mid_blk; /* first_half_cycle == mid_cycle */
 293		mid_blk = BLK_AVG(first_blk, end_blk);
 294	}
 295	ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
 296	       (mid_blk == end_blk && mid_blk-1 == first_blk));
 297
 298	*last_blk = end_blk;
 299
 300	return 0;
 301}
 302
 303/*
 304 * Check that a range of blocks does not contain stop_on_cycle_no.
 305 * Fill in *new_blk with the block offset where such a block is
 306 * found, or with -1 (an invalid block number) if there is no such
 307 * block in the range.  The scan needs to occur from front to back
 308 * and the pointer into the region must be updated since a later
 309 * routine will need to perform another test.
 310 */
 311STATIC int
 312xlog_find_verify_cycle(
 313	struct xlog	*log,
 314	xfs_daddr_t	start_blk,
 315	int		nbblks,
 316	uint		stop_on_cycle_no,
 317	xfs_daddr_t	*new_blk)
 318{
 319	xfs_daddr_t	i, j;
 320	uint		cycle;
 321	char		*buffer;
 322	xfs_daddr_t	bufblks;
 323	char		*buf = NULL;
 324	int		error = 0;
 325
 326	/*
 327	 * Greedily allocate a buffer big enough to handle the full
 328	 * range of basic blocks we'll be examining.  If that fails,
 329	 * try a smaller size.  We need to be able to read at least
 330	 * a log sector, or we're out of luck.
 331	 */
 332	bufblks = roundup_pow_of_two(nbblks);
 333	while (bufblks > log->l_logBBsize)
 334		bufblks >>= 1;
 335	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
 336		bufblks >>= 1;
 337		if (bufblks < log->l_sectBBsize)
 338			return -ENOMEM;
 339	}
 340
 341	for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
 342		int	bcount;
 343
 344		bcount = min(bufblks, (start_blk + nbblks - i));
 345
 346		error = xlog_bread(log, i, bcount, buffer, &buf);
 347		if (error)
 348			goto out;
 349
 350		for (j = 0; j < bcount; j++) {
 351			cycle = xlog_get_cycle(buf);
 352			if (cycle == stop_on_cycle_no) {
 353				*new_blk = i+j;
 354				goto out;
 355			}
 356
 357			buf += BBSIZE;
 358		}
 359	}
 360
 361	*new_blk = -1;
 362
 363out:
 364	kmem_free(buffer);
 365	return error;
 366}
 367
 368static inline int
 369xlog_logrec_hblks(struct xlog *log, struct xlog_rec_header *rh)
 370{
 371	if (xfs_has_logv2(log->l_mp)) {
 372		int	h_size = be32_to_cpu(rh->h_size);
 373
 374		if ((be32_to_cpu(rh->h_version) & XLOG_VERSION_2) &&
 375		    h_size > XLOG_HEADER_CYCLE_SIZE)
 376			return DIV_ROUND_UP(h_size, XLOG_HEADER_CYCLE_SIZE);
 377	}
 378	return 1;
 379}
 380
 381/*
 382 * Potentially backup over partial log record write.
 383 *
 384 * In the typical case, last_blk is the number of the block directly after
 385 * a good log record.  Therefore, we subtract one to get the block number
 386 * of the last block in the given buffer.  extra_bblks contains the number
 387 * of blocks we would have read on a previous read.  This happens when the
 388 * last log record is split over the end of the physical log.
 389 *
 390 * extra_bblks is the number of blocks potentially verified on a previous
 391 * call to this routine.
 392 */
 393STATIC int
 394xlog_find_verify_log_record(
 395	struct xlog		*log,
 396	xfs_daddr_t		start_blk,
 397	xfs_daddr_t		*last_blk,
 398	int			extra_bblks)
 399{
 400	xfs_daddr_t		i;
 401	char			*buffer;
 402	char			*offset = NULL;
 403	xlog_rec_header_t	*head = NULL;
 404	int			error = 0;
 405	int			smallmem = 0;
 406	int			num_blks = *last_blk - start_blk;
 407	int			xhdrs;
 408
 409	ASSERT(start_blk != 0 || *last_blk != start_blk);
 410
 411	buffer = xlog_alloc_buffer(log, num_blks);
 412	if (!buffer) {
 413		buffer = xlog_alloc_buffer(log, 1);
 414		if (!buffer)
 415			return -ENOMEM;
 416		smallmem = 1;
 417	} else {
 418		error = xlog_bread(log, start_blk, num_blks, buffer, &offset);
 419		if (error)
 420			goto out;
 421		offset += ((num_blks - 1) << BBSHIFT);
 422	}
 423
 424	for (i = (*last_blk) - 1; i >= 0; i--) {
 425		if (i < start_blk) {
 426			/* valid log record not found */
 427			xfs_warn(log->l_mp,
 428		"Log inconsistent (didn't find previous header)");
 429			ASSERT(0);
 430			error = -EFSCORRUPTED;
 431			goto out;
 432		}
 433
 434		if (smallmem) {
 435			error = xlog_bread(log, i, 1, buffer, &offset);
 436			if (error)
 437				goto out;
 438		}
 439
 440		head = (xlog_rec_header_t *)offset;
 441
 442		if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
 443			break;
 444
 445		if (!smallmem)
 446			offset -= BBSIZE;
 447	}
 448
 449	/*
 450	 * We hit the beginning of the physical log & still no header.  Return
 451	 * to caller.  If caller can handle a return of -1, then this routine
 452	 * will be called again for the end of the physical log.
 453	 */
 454	if (i == -1) {
 455		error = 1;
 456		goto out;
 457	}
 458
 459	/*
 460	 * We have the final block of the good log (the first block
 461	 * of the log record _before_ the head. So we check the uuid.
 462	 */
 463	if ((error = xlog_header_check_mount(log->l_mp, head)))
 464		goto out;
 465
 466	/*
 467	 * We may have found a log record header before we expected one.
 468	 * last_blk will be the 1st block # with a given cycle #.  We may end
 469	 * up reading an entire log record.  In this case, we don't want to
 470	 * reset last_blk.  Only when last_blk points in the middle of a log
 471	 * record do we update last_blk.
 472	 */
 473	xhdrs = xlog_logrec_hblks(log, head);
 
 
 
 
 
 
 
 
 474
 475	if (*last_blk - i + extra_bblks !=
 476	    BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
 477		*last_blk = i;
 478
 479out:
 480	kmem_free(buffer);
 481	return error;
 482}
 483
 484/*
 485 * Head is defined to be the point of the log where the next log write
 486 * could go.  This means that incomplete LR writes at the end are
 487 * eliminated when calculating the head.  We aren't guaranteed that previous
 488 * LR have complete transactions.  We only know that a cycle number of
 489 * current cycle number -1 won't be present in the log if we start writing
 490 * from our current block number.
 491 *
 492 * last_blk contains the block number of the first block with a given
 493 * cycle number.
 494 *
 495 * Return: zero if normal, non-zero if error.
 496 */
 497STATIC int
 498xlog_find_head(
 499	struct xlog	*log,
 500	xfs_daddr_t	*return_head_blk)
 501{
 502	char		*buffer;
 503	char		*offset;
 504	xfs_daddr_t	new_blk, first_blk, start_blk, last_blk, head_blk;
 505	int		num_scan_bblks;
 506	uint		first_half_cycle, last_half_cycle;
 507	uint		stop_on_cycle;
 508	int		error, log_bbnum = log->l_logBBsize;
 509
 510	/* Is the end of the log device zeroed? */
 511	error = xlog_find_zeroed(log, &first_blk);
 512	if (error < 0) {
 513		xfs_warn(log->l_mp, "empty log check failed");
 514		return error;
 515	}
 516	if (error == 1) {
 517		*return_head_blk = first_blk;
 518
 519		/* Is the whole lot zeroed? */
 520		if (!first_blk) {
 521			/* Linux XFS shouldn't generate totally zeroed logs -
 522			 * mkfs etc write a dummy unmount record to a fresh
 523			 * log so we can store the uuid in there
 524			 */
 525			xfs_warn(log->l_mp, "totally zeroed log");
 526		}
 527
 528		return 0;
 529	}
 530
 531	first_blk = 0;			/* get cycle # of 1st block */
 532	buffer = xlog_alloc_buffer(log, 1);
 533	if (!buffer)
 534		return -ENOMEM;
 535
 536	error = xlog_bread(log, 0, 1, buffer, &offset);
 537	if (error)
 538		goto out_free_buffer;
 539
 540	first_half_cycle = xlog_get_cycle(offset);
 541
 542	last_blk = head_blk = log_bbnum - 1;	/* get cycle # of last block */
 543	error = xlog_bread(log, last_blk, 1, buffer, &offset);
 544	if (error)
 545		goto out_free_buffer;
 546
 547	last_half_cycle = xlog_get_cycle(offset);
 548	ASSERT(last_half_cycle != 0);
 549
 550	/*
 551	 * If the 1st half cycle number is equal to the last half cycle number,
 552	 * then the entire log is stamped with the same cycle number.  In this
 553	 * case, head_blk can't be set to zero (which makes sense).  The below
 554	 * math doesn't work out properly with head_blk equal to zero.  Instead,
 555	 * we set it to log_bbnum which is an invalid block number, but this
 556	 * value makes the math correct.  If head_blk doesn't changed through
 557	 * all the tests below, *head_blk is set to zero at the very end rather
 558	 * than log_bbnum.  In a sense, log_bbnum and zero are the same block
 559	 * in a circular file.
 560	 */
 561	if (first_half_cycle == last_half_cycle) {
 562		/*
 563		 * In this case we believe that the entire log should have
 564		 * cycle number last_half_cycle.  We need to scan backwards
 565		 * from the end verifying that there are no holes still
 566		 * containing last_half_cycle - 1.  If we find such a hole,
 567		 * then the start of that hole will be the new head.  The
 568		 * simple case looks like
 569		 *        x | x ... | x - 1 | x
 570		 * Another case that fits this picture would be
 571		 *        x | x + 1 | x ... | x
 572		 * In this case the head really is somewhere at the end of the
 573		 * log, as one of the latest writes at the beginning was
 574		 * incomplete.
 575		 * One more case is
 576		 *        x | x + 1 | x ... | x - 1 | x
 577		 * This is really the combination of the above two cases, and
 578		 * the head has to end up at the start of the x-1 hole at the
 579		 * end of the log.
 580		 *
 581		 * In the 256k log case, we will read from the beginning to the
 582		 * end of the log and search for cycle numbers equal to x-1.
 583		 * We don't worry about the x+1 blocks that we encounter,
 584		 * because we know that they cannot be the head since the log
 585		 * started with x.
 586		 */
 587		head_blk = log_bbnum;
 588		stop_on_cycle = last_half_cycle - 1;
 589	} else {
 590		/*
 591		 * In this case we want to find the first block with cycle
 592		 * number matching last_half_cycle.  We expect the log to be
 593		 * some variation on
 594		 *        x + 1 ... | x ... | x
 595		 * The first block with cycle number x (last_half_cycle) will
 596		 * be where the new head belongs.  First we do a binary search
 597		 * for the first occurrence of last_half_cycle.  The binary
 598		 * search may not be totally accurate, so then we scan back
 599		 * from there looking for occurrences of last_half_cycle before
 600		 * us.  If that backwards scan wraps around the beginning of
 601		 * the log, then we look for occurrences of last_half_cycle - 1
 602		 * at the end of the log.  The cases we're looking for look
 603		 * like
 604		 *                               v binary search stopped here
 605		 *        x + 1 ... | x | x + 1 | x ... | x
 606		 *                   ^ but we want to locate this spot
 607		 * or
 608		 *        <---------> less than scan distance
 609		 *        x + 1 ... | x ... | x - 1 | x
 610		 *                           ^ we want to locate this spot
 611		 */
 612		stop_on_cycle = last_half_cycle;
 613		error = xlog_find_cycle_start(log, buffer, first_blk, &head_blk,
 614				last_half_cycle);
 615		if (error)
 616			goto out_free_buffer;
 617	}
 618
 619	/*
 620	 * Now validate the answer.  Scan back some number of maximum possible
 621	 * blocks and make sure each one has the expected cycle number.  The
 622	 * maximum is determined by the total possible amount of buffering
 623	 * in the in-core log.  The following number can be made tighter if
 624	 * we actually look at the block size of the filesystem.
 625	 */
 626	num_scan_bblks = min_t(int, log_bbnum, XLOG_TOTAL_REC_SHIFT(log));
 627	if (head_blk >= num_scan_bblks) {
 628		/*
 629		 * We are guaranteed that the entire check can be performed
 630		 * in one buffer.
 631		 */
 632		start_blk = head_blk - num_scan_bblks;
 633		if ((error = xlog_find_verify_cycle(log,
 634						start_blk, num_scan_bblks,
 635						stop_on_cycle, &new_blk)))
 636			goto out_free_buffer;
 637		if (new_blk != -1)
 638			head_blk = new_blk;
 639	} else {		/* need to read 2 parts of log */
 640		/*
 641		 * We are going to scan backwards in the log in two parts.
 642		 * First we scan the physical end of the log.  In this part
 643		 * of the log, we are looking for blocks with cycle number
 644		 * last_half_cycle - 1.
 645		 * If we find one, then we know that the log starts there, as
 646		 * we've found a hole that didn't get written in going around
 647		 * the end of the physical log.  The simple case for this is
 648		 *        x + 1 ... | x ... | x - 1 | x
 649		 *        <---------> less than scan distance
 650		 * If all of the blocks at the end of the log have cycle number
 651		 * last_half_cycle, then we check the blocks at the start of
 652		 * the log looking for occurrences of last_half_cycle.  If we
 653		 * find one, then our current estimate for the location of the
 654		 * first occurrence of last_half_cycle is wrong and we move
 655		 * back to the hole we've found.  This case looks like
 656		 *        x + 1 ... | x | x + 1 | x ...
 657		 *                               ^ binary search stopped here
 658		 * Another case we need to handle that only occurs in 256k
 659		 * logs is
 660		 *        x + 1 ... | x ... | x+1 | x ...
 661		 *                   ^ binary search stops here
 662		 * In a 256k log, the scan at the end of the log will see the
 663		 * x + 1 blocks.  We need to skip past those since that is
 664		 * certainly not the head of the log.  By searching for
 665		 * last_half_cycle-1 we accomplish that.
 666		 */
 667		ASSERT(head_blk <= INT_MAX &&
 668			(xfs_daddr_t) num_scan_bblks >= head_blk);
 669		start_blk = log_bbnum - (num_scan_bblks - head_blk);
 670		if ((error = xlog_find_verify_cycle(log, start_blk,
 671					num_scan_bblks - (int)head_blk,
 672					(stop_on_cycle - 1), &new_blk)))
 673			goto out_free_buffer;
 674		if (new_blk != -1) {
 675			head_blk = new_blk;
 676			goto validate_head;
 677		}
 678
 679		/*
 680		 * Scan beginning of log now.  The last part of the physical
 681		 * log is good.  This scan needs to verify that it doesn't find
 682		 * the last_half_cycle.
 683		 */
 684		start_blk = 0;
 685		ASSERT(head_blk <= INT_MAX);
 686		if ((error = xlog_find_verify_cycle(log,
 687					start_blk, (int)head_blk,
 688					stop_on_cycle, &new_blk)))
 689			goto out_free_buffer;
 690		if (new_blk != -1)
 691			head_blk = new_blk;
 692	}
 693
 694validate_head:
 695	/*
 696	 * Now we need to make sure head_blk is not pointing to a block in
 697	 * the middle of a log record.
 698	 */
 699	num_scan_bblks = XLOG_REC_SHIFT(log);
 700	if (head_blk >= num_scan_bblks) {
 701		start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
 702
 703		/* start ptr at last block ptr before head_blk */
 704		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
 705		if (error == 1)
 706			error = -EIO;
 707		if (error)
 708			goto out_free_buffer;
 709	} else {
 710		start_blk = 0;
 711		ASSERT(head_blk <= INT_MAX);
 712		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
 713		if (error < 0)
 714			goto out_free_buffer;
 715		if (error == 1) {
 716			/* We hit the beginning of the log during our search */
 717			start_blk = log_bbnum - (num_scan_bblks - head_blk);
 718			new_blk = log_bbnum;
 719			ASSERT(start_blk <= INT_MAX &&
 720				(xfs_daddr_t) log_bbnum-start_blk >= 0);
 721			ASSERT(head_blk <= INT_MAX);
 722			error = xlog_find_verify_log_record(log, start_blk,
 723							&new_blk, (int)head_blk);
 724			if (error == 1)
 725				error = -EIO;
 726			if (error)
 727				goto out_free_buffer;
 728			if (new_blk != log_bbnum)
 729				head_blk = new_blk;
 730		} else if (error)
 731			goto out_free_buffer;
 732	}
 733
 734	kmem_free(buffer);
 735	if (head_blk == log_bbnum)
 736		*return_head_blk = 0;
 737	else
 738		*return_head_blk = head_blk;
 739	/*
 740	 * When returning here, we have a good block number.  Bad block
 741	 * means that during a previous crash, we didn't have a clean break
 742	 * from cycle number N to cycle number N-1.  In this case, we need
 743	 * to find the first block with cycle number N-1.
 744	 */
 745	return 0;
 746
 747out_free_buffer:
 748	kmem_free(buffer);
 749	if (error)
 750		xfs_warn(log->l_mp, "failed to find log head");
 751	return error;
 752}
 753
 754/*
 755 * Seek backwards in the log for log record headers.
 756 *
 757 * Given a starting log block, walk backwards until we find the provided number
 758 * of records or hit the provided tail block. The return value is the number of
 759 * records encountered or a negative error code. The log block and buffer
 760 * pointer of the last record seen are returned in rblk and rhead respectively.
 761 */
 762STATIC int
 763xlog_rseek_logrec_hdr(
 764	struct xlog		*log,
 765	xfs_daddr_t		head_blk,
 766	xfs_daddr_t		tail_blk,
 767	int			count,
 768	char			*buffer,
 769	xfs_daddr_t		*rblk,
 770	struct xlog_rec_header	**rhead,
 771	bool			*wrapped)
 772{
 773	int			i;
 774	int			error;
 775	int			found = 0;
 776	char			*offset = NULL;
 777	xfs_daddr_t		end_blk;
 778
 779	*wrapped = false;
 780
 781	/*
 782	 * Walk backwards from the head block until we hit the tail or the first
 783	 * block in the log.
 784	 */
 785	end_blk = head_blk > tail_blk ? tail_blk : 0;
 786	for (i = (int) head_blk - 1; i >= end_blk; i--) {
 787		error = xlog_bread(log, i, 1, buffer, &offset);
 788		if (error)
 789			goto out_error;
 790
 791		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 792			*rblk = i;
 793			*rhead = (struct xlog_rec_header *) offset;
 794			if (++found == count)
 795				break;
 796		}
 797	}
 798
 799	/*
 800	 * If we haven't hit the tail block or the log record header count,
 801	 * start looking again from the end of the physical log. Note that
 802	 * callers can pass head == tail if the tail is not yet known.
 803	 */
 804	if (tail_blk >= head_blk && found != count) {
 805		for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
 806			error = xlog_bread(log, i, 1, buffer, &offset);
 807			if (error)
 808				goto out_error;
 809
 810			if (*(__be32 *)offset ==
 811			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 812				*wrapped = true;
 813				*rblk = i;
 814				*rhead = (struct xlog_rec_header *) offset;
 815				if (++found == count)
 816					break;
 817			}
 818		}
 819	}
 820
 821	return found;
 822
 823out_error:
 824	return error;
 825}
 826
 827/*
 828 * Seek forward in the log for log record headers.
 829 *
 830 * Given head and tail blocks, walk forward from the tail block until we find
 831 * the provided number of records or hit the head block. The return value is the
 832 * number of records encountered or a negative error code. The log block and
 833 * buffer pointer of the last record seen are returned in rblk and rhead
 834 * respectively.
 835 */
 836STATIC int
 837xlog_seek_logrec_hdr(
 838	struct xlog		*log,
 839	xfs_daddr_t		head_blk,
 840	xfs_daddr_t		tail_blk,
 841	int			count,
 842	char			*buffer,
 843	xfs_daddr_t		*rblk,
 844	struct xlog_rec_header	**rhead,
 845	bool			*wrapped)
 846{
 847	int			i;
 848	int			error;
 849	int			found = 0;
 850	char			*offset = NULL;
 851	xfs_daddr_t		end_blk;
 852
 853	*wrapped = false;
 854
 855	/*
 856	 * Walk forward from the tail block until we hit the head or the last
 857	 * block in the log.
 858	 */
 859	end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
 860	for (i = (int) tail_blk; i <= end_blk; i++) {
 861		error = xlog_bread(log, i, 1, buffer, &offset);
 862		if (error)
 863			goto out_error;
 864
 865		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 866			*rblk = i;
 867			*rhead = (struct xlog_rec_header *) offset;
 868			if (++found == count)
 869				break;
 870		}
 871	}
 872
 873	/*
 874	 * If we haven't hit the head block or the log record header count,
 875	 * start looking again from the start of the physical log.
 876	 */
 877	if (tail_blk > head_blk && found != count) {
 878		for (i = 0; i < (int) head_blk; i++) {
 879			error = xlog_bread(log, i, 1, buffer, &offset);
 880			if (error)
 881				goto out_error;
 882
 883			if (*(__be32 *)offset ==
 884			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 885				*wrapped = true;
 886				*rblk = i;
 887				*rhead = (struct xlog_rec_header *) offset;
 888				if (++found == count)
 889					break;
 890			}
 891		}
 892	}
 893
 894	return found;
 895
 896out_error:
 897	return error;
 898}
 899
 900/*
 901 * Calculate distance from head to tail (i.e., unused space in the log).
 902 */
 903static inline int
 904xlog_tail_distance(
 905	struct xlog	*log,
 906	xfs_daddr_t	head_blk,
 907	xfs_daddr_t	tail_blk)
 908{
 909	if (head_blk < tail_blk)
 910		return tail_blk - head_blk;
 911
 912	return tail_blk + (log->l_logBBsize - head_blk);
 913}
 914
 915/*
 916 * Verify the log tail. This is particularly important when torn or incomplete
 917 * writes have been detected near the front of the log and the head has been
 918 * walked back accordingly.
 919 *
 920 * We also have to handle the case where the tail was pinned and the head
 921 * blocked behind the tail right before a crash. If the tail had been pushed
 922 * immediately prior to the crash and the subsequent checkpoint was only
 923 * partially written, it's possible it overwrote the last referenced tail in the
 924 * log with garbage. This is not a coherency problem because the tail must have
 925 * been pushed before it can be overwritten, but appears as log corruption to
 926 * recovery because we have no way to know the tail was updated if the
 927 * subsequent checkpoint didn't write successfully.
 928 *
 929 * Therefore, CRC check the log from tail to head. If a failure occurs and the
 930 * offending record is within max iclog bufs from the head, walk the tail
 931 * forward and retry until a valid tail is found or corruption is detected out
 932 * of the range of a possible overwrite.
 933 */
 934STATIC int
 935xlog_verify_tail(
 936	struct xlog		*log,
 937	xfs_daddr_t		head_blk,
 938	xfs_daddr_t		*tail_blk,
 939	int			hsize)
 940{
 941	struct xlog_rec_header	*thead;
 942	char			*buffer;
 943	xfs_daddr_t		first_bad;
 944	int			error = 0;
 945	bool			wrapped;
 946	xfs_daddr_t		tmp_tail;
 947	xfs_daddr_t		orig_tail = *tail_blk;
 948
 949	buffer = xlog_alloc_buffer(log, 1);
 950	if (!buffer)
 951		return -ENOMEM;
 952
 953	/*
 954	 * Make sure the tail points to a record (returns positive count on
 955	 * success).
 956	 */
 957	error = xlog_seek_logrec_hdr(log, head_blk, *tail_blk, 1, buffer,
 958			&tmp_tail, &thead, &wrapped);
 959	if (error < 0)
 960		goto out;
 961	if (*tail_blk != tmp_tail)
 962		*tail_blk = tmp_tail;
 963
 964	/*
 965	 * Run a CRC check from the tail to the head. We can't just check
 966	 * MAX_ICLOGS records past the tail because the tail may point to stale
 967	 * blocks cleared during the search for the head/tail. These blocks are
 968	 * overwritten with zero-length records and thus record count is not a
 969	 * reliable indicator of the iclog state before a crash.
 970	 */
 971	first_bad = 0;
 972	error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
 973				      XLOG_RECOVER_CRCPASS, &first_bad);
 974	while ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
 975		int	tail_distance;
 976
 977		/*
 978		 * Is corruption within range of the head? If so, retry from
 979		 * the next record. Otherwise return an error.
 980		 */
 981		tail_distance = xlog_tail_distance(log, head_blk, first_bad);
 982		if (tail_distance > BTOBB(XLOG_MAX_ICLOGS * hsize))
 983			break;
 984
 985		/* skip to the next record; returns positive count on success */
 986		error = xlog_seek_logrec_hdr(log, head_blk, first_bad, 2,
 987				buffer, &tmp_tail, &thead, &wrapped);
 988		if (error < 0)
 989			goto out;
 990
 991		*tail_blk = tmp_tail;
 992		first_bad = 0;
 993		error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
 994					      XLOG_RECOVER_CRCPASS, &first_bad);
 995	}
 996
 997	if (!error && *tail_blk != orig_tail)
 998		xfs_warn(log->l_mp,
 999		"Tail block (0x%llx) overwrite detected. Updated to 0x%llx",
1000			 orig_tail, *tail_blk);
1001out:
1002	kmem_free(buffer);
1003	return error;
1004}
1005
1006/*
1007 * Detect and trim torn writes from the head of the log.
1008 *
1009 * Storage without sector atomicity guarantees can result in torn writes in the
1010 * log in the event of a crash. Our only means to detect this scenario is via
1011 * CRC verification. While we can't always be certain that CRC verification
1012 * failure is due to a torn write vs. an unrelated corruption, we do know that
1013 * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1014 * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1015 * the log and treat failures in this range as torn writes as a matter of
1016 * policy. In the event of CRC failure, the head is walked back to the last good
1017 * record in the log and the tail is updated from that record and verified.
1018 */
1019STATIC int
1020xlog_verify_head(
1021	struct xlog		*log,
1022	xfs_daddr_t		*head_blk,	/* in/out: unverified head */
1023	xfs_daddr_t		*tail_blk,	/* out: tail block */
1024	char			*buffer,
1025	xfs_daddr_t		*rhead_blk,	/* start blk of last record */
1026	struct xlog_rec_header	**rhead,	/* ptr to last record */
1027	bool			*wrapped)	/* last rec. wraps phys. log */
1028{
1029	struct xlog_rec_header	*tmp_rhead;
1030	char			*tmp_buffer;
1031	xfs_daddr_t		first_bad;
1032	xfs_daddr_t		tmp_rhead_blk;
1033	int			found;
1034	int			error;
1035	bool			tmp_wrapped;
1036
1037	/*
1038	 * Check the head of the log for torn writes. Search backwards from the
1039	 * head until we hit the tail or the maximum number of log record I/Os
1040	 * that could have been in flight at one time. Use a temporary buffer so
1041	 * we don't trash the rhead/buffer pointers from the caller.
1042	 */
1043	tmp_buffer = xlog_alloc_buffer(log, 1);
1044	if (!tmp_buffer)
1045		return -ENOMEM;
1046	error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1047				      XLOG_MAX_ICLOGS, tmp_buffer,
1048				      &tmp_rhead_blk, &tmp_rhead, &tmp_wrapped);
1049	kmem_free(tmp_buffer);
1050	if (error < 0)
1051		return error;
1052
1053	/*
1054	 * Now run a CRC verification pass over the records starting at the
1055	 * block found above to the current head. If a CRC failure occurs, the
1056	 * log block of the first bad record is saved in first_bad.
1057	 */
1058	error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1059				      XLOG_RECOVER_CRCPASS, &first_bad);
1060	if ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1061		/*
1062		 * We've hit a potential torn write. Reset the error and warn
1063		 * about it.
1064		 */
1065		error = 0;
1066		xfs_warn(log->l_mp,
1067"Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1068			 first_bad, *head_blk);
1069
1070		/*
1071		 * Get the header block and buffer pointer for the last good
1072		 * record before the bad record.
1073		 *
1074		 * Note that xlog_find_tail() clears the blocks at the new head
1075		 * (i.e., the records with invalid CRC) if the cycle number
1076		 * matches the current cycle.
1077		 */
1078		found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1,
1079				buffer, rhead_blk, rhead, wrapped);
1080		if (found < 0)
1081			return found;
1082		if (found == 0)		/* XXX: right thing to do here? */
1083			return -EIO;
1084
1085		/*
1086		 * Reset the head block to the starting block of the first bad
1087		 * log record and set the tail block based on the last good
1088		 * record.
1089		 *
1090		 * Bail out if the updated head/tail match as this indicates
1091		 * possible corruption outside of the acceptable
1092		 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1093		 */
1094		*head_blk = first_bad;
1095		*tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1096		if (*head_blk == *tail_blk) {
1097			ASSERT(0);
1098			return 0;
1099		}
1100	}
1101	if (error)
1102		return error;
1103
1104	return xlog_verify_tail(log, *head_blk, tail_blk,
1105				be32_to_cpu((*rhead)->h_size));
1106}
1107
1108/*
1109 * We need to make sure we handle log wrapping properly, so we can't use the
1110 * calculated logbno directly. Make sure it wraps to the correct bno inside the
1111 * log.
1112 *
1113 * The log is limited to 32 bit sizes, so we use the appropriate modulus
1114 * operation here and cast it back to a 64 bit daddr on return.
1115 */
1116static inline xfs_daddr_t
1117xlog_wrap_logbno(
1118	struct xlog		*log,
1119	xfs_daddr_t		bno)
1120{
1121	int			mod;
1122
1123	div_s64_rem(bno, log->l_logBBsize, &mod);
1124	return mod;
1125}
1126
1127/*
1128 * Check whether the head of the log points to an unmount record. In other
1129 * words, determine whether the log is clean. If so, update the in-core state
1130 * appropriately.
1131 */
1132static int
1133xlog_check_unmount_rec(
1134	struct xlog		*log,
1135	xfs_daddr_t		*head_blk,
1136	xfs_daddr_t		*tail_blk,
1137	struct xlog_rec_header	*rhead,
1138	xfs_daddr_t		rhead_blk,
1139	char			*buffer,
1140	bool			*clean)
1141{
1142	struct xlog_op_header	*op_head;
1143	xfs_daddr_t		umount_data_blk;
1144	xfs_daddr_t		after_umount_blk;
1145	int			hblks;
1146	int			error;
1147	char			*offset;
1148
1149	*clean = false;
1150
1151	/*
1152	 * Look for unmount record. If we find it, then we know there was a
1153	 * clean unmount. Since 'i' could be the last block in the physical
1154	 * log, we convert to a log block before comparing to the head_blk.
1155	 *
1156	 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1157	 * below. We won't want to clear the unmount record if there is one, so
1158	 * we pass the lsn of the unmount record rather than the block after it.
1159	 */
1160	hblks = xlog_logrec_hblks(log, rhead);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1161	after_umount_blk = xlog_wrap_logbno(log,
1162			rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len)));
1163
1164	if (*head_blk == after_umount_blk &&
1165	    be32_to_cpu(rhead->h_num_logops) == 1) {
1166		umount_data_blk = xlog_wrap_logbno(log, rhead_blk + hblks);
1167		error = xlog_bread(log, umount_data_blk, 1, buffer, &offset);
1168		if (error)
1169			return error;
1170
1171		op_head = (struct xlog_op_header *)offset;
1172		if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1173			/*
1174			 * Set tail and last sync so that newly written log
1175			 * records will point recovery to after the current
1176			 * unmount record.
1177			 */
1178			xlog_assign_atomic_lsn(&log->l_tail_lsn,
1179					log->l_curr_cycle, after_umount_blk);
1180			xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1181					log->l_curr_cycle, after_umount_blk);
1182			*tail_blk = after_umount_blk;
1183
1184			*clean = true;
1185		}
1186	}
1187
1188	return 0;
1189}
1190
1191static void
1192xlog_set_state(
1193	struct xlog		*log,
1194	xfs_daddr_t		head_blk,
1195	struct xlog_rec_header	*rhead,
1196	xfs_daddr_t		rhead_blk,
1197	bool			bump_cycle)
1198{
1199	/*
1200	 * Reset log values according to the state of the log when we
1201	 * crashed.  In the case where head_blk == 0, we bump curr_cycle
1202	 * one because the next write starts a new cycle rather than
1203	 * continuing the cycle of the last good log record.  At this
1204	 * point we have guaranteed that all partial log records have been
1205	 * accounted for.  Therefore, we know that the last good log record
1206	 * written was complete and ended exactly on the end boundary
1207	 * of the physical log.
1208	 */
1209	log->l_prev_block = rhead_blk;
1210	log->l_curr_block = (int)head_blk;
1211	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1212	if (bump_cycle)
1213		log->l_curr_cycle++;
1214	atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1215	atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1216	xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1217					BBTOB(log->l_curr_block));
1218	xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1219					BBTOB(log->l_curr_block));
1220}
1221
1222/*
1223 * Find the sync block number or the tail of the log.
1224 *
1225 * This will be the block number of the last record to have its
1226 * associated buffers synced to disk.  Every log record header has
1227 * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
1228 * to get a sync block number.  The only concern is to figure out which
1229 * log record header to believe.
1230 *
1231 * The following algorithm uses the log record header with the largest
1232 * lsn.  The entire log record does not need to be valid.  We only care
1233 * that the header is valid.
1234 *
1235 * We could speed up search by using current head_blk buffer, but it is not
1236 * available.
1237 */
1238STATIC int
1239xlog_find_tail(
1240	struct xlog		*log,
1241	xfs_daddr_t		*head_blk,
1242	xfs_daddr_t		*tail_blk)
1243{
1244	xlog_rec_header_t	*rhead;
1245	char			*offset = NULL;
1246	char			*buffer;
1247	int			error;
1248	xfs_daddr_t		rhead_blk;
1249	xfs_lsn_t		tail_lsn;
1250	bool			wrapped = false;
1251	bool			clean = false;
1252
1253	/*
1254	 * Find previous log record
1255	 */
1256	if ((error = xlog_find_head(log, head_blk)))
1257		return error;
1258	ASSERT(*head_blk < INT_MAX);
1259
1260	buffer = xlog_alloc_buffer(log, 1);
1261	if (!buffer)
1262		return -ENOMEM;
1263	if (*head_blk == 0) {				/* special case */
1264		error = xlog_bread(log, 0, 1, buffer, &offset);
1265		if (error)
1266			goto done;
1267
1268		if (xlog_get_cycle(offset) == 0) {
1269			*tail_blk = 0;
1270			/* leave all other log inited values alone */
1271			goto done;
1272		}
1273	}
1274
1275	/*
1276	 * Search backwards through the log looking for the log record header
1277	 * block. This wraps all the way back around to the head so something is
1278	 * seriously wrong if we can't find it.
1279	 */
1280	error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, buffer,
1281				      &rhead_blk, &rhead, &wrapped);
1282	if (error < 0)
1283		goto done;
1284	if (!error) {
1285		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1286		error = -EFSCORRUPTED;
1287		goto done;
1288	}
1289	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1290
1291	/*
1292	 * Set the log state based on the current head record.
1293	 */
1294	xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1295	tail_lsn = atomic64_read(&log->l_tail_lsn);
1296
1297	/*
1298	 * Look for an unmount record at the head of the log. This sets the log
1299	 * state to determine whether recovery is necessary.
1300	 */
1301	error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1302				       rhead_blk, buffer, &clean);
1303	if (error)
1304		goto done;
1305
1306	/*
1307	 * Verify the log head if the log is not clean (e.g., we have anything
1308	 * but an unmount record at the head). This uses CRC verification to
1309	 * detect and trim torn writes. If discovered, CRC failures are
1310	 * considered torn writes and the log head is trimmed accordingly.
1311	 *
1312	 * Note that we can only run CRC verification when the log is dirty
1313	 * because there's no guarantee that the log data behind an unmount
1314	 * record is compatible with the current architecture.
1315	 */
1316	if (!clean) {
1317		xfs_daddr_t	orig_head = *head_blk;
1318
1319		error = xlog_verify_head(log, head_blk, tail_blk, buffer,
1320					 &rhead_blk, &rhead, &wrapped);
1321		if (error)
1322			goto done;
1323
1324		/* update in-core state again if the head changed */
1325		if (*head_blk != orig_head) {
1326			xlog_set_state(log, *head_blk, rhead, rhead_blk,
1327				       wrapped);
1328			tail_lsn = atomic64_read(&log->l_tail_lsn);
1329			error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1330						       rhead, rhead_blk, buffer,
1331						       &clean);
1332			if (error)
1333				goto done;
1334		}
1335	}
1336
1337	/*
1338	 * Note that the unmount was clean. If the unmount was not clean, we
1339	 * need to know this to rebuild the superblock counters from the perag
1340	 * headers if we have a filesystem using non-persistent counters.
1341	 */
1342	if (clean)
1343		set_bit(XFS_OPSTATE_CLEAN, &log->l_mp->m_opstate);
1344
1345	/*
1346	 * Make sure that there are no blocks in front of the head
1347	 * with the same cycle number as the head.  This can happen
1348	 * because we allow multiple outstanding log writes concurrently,
1349	 * and the later writes might make it out before earlier ones.
1350	 *
1351	 * We use the lsn from before modifying it so that we'll never
1352	 * overwrite the unmount record after a clean unmount.
1353	 *
1354	 * Do this only if we are going to recover the filesystem
1355	 *
1356	 * NOTE: This used to say "if (!readonly)"
1357	 * However on Linux, we can & do recover a read-only filesystem.
1358	 * We only skip recovery if NORECOVERY is specified on mount,
1359	 * in which case we would not be here.
1360	 *
1361	 * But... if the -device- itself is readonly, just skip this.
1362	 * We can't recover this device anyway, so it won't matter.
1363	 */
1364	if (!xfs_readonly_buftarg(log->l_targ))
1365		error = xlog_clear_stale_blocks(log, tail_lsn);
1366
1367done:
1368	kmem_free(buffer);
1369
1370	if (error)
1371		xfs_warn(log->l_mp, "failed to locate log tail");
1372	return error;
1373}
1374
1375/*
1376 * Is the log zeroed at all?
1377 *
1378 * The last binary search should be changed to perform an X block read
1379 * once X becomes small enough.  You can then search linearly through
1380 * the X blocks.  This will cut down on the number of reads we need to do.
1381 *
1382 * If the log is partially zeroed, this routine will pass back the blkno
1383 * of the first block with cycle number 0.  It won't have a complete LR
1384 * preceding it.
1385 *
1386 * Return:
1387 *	0  => the log is completely written to
1388 *	1 => use *blk_no as the first block of the log
1389 *	<0 => error has occurred
1390 */
1391STATIC int
1392xlog_find_zeroed(
1393	struct xlog	*log,
1394	xfs_daddr_t	*blk_no)
1395{
1396	char		*buffer;
1397	char		*offset;
1398	uint	        first_cycle, last_cycle;
1399	xfs_daddr_t	new_blk, last_blk, start_blk;
1400	xfs_daddr_t     num_scan_bblks;
1401	int	        error, log_bbnum = log->l_logBBsize;
1402
1403	*blk_no = 0;
1404
1405	/* check totally zeroed log */
1406	buffer = xlog_alloc_buffer(log, 1);
1407	if (!buffer)
1408		return -ENOMEM;
1409	error = xlog_bread(log, 0, 1, buffer, &offset);
1410	if (error)
1411		goto out_free_buffer;
1412
1413	first_cycle = xlog_get_cycle(offset);
1414	if (first_cycle == 0) {		/* completely zeroed log */
1415		*blk_no = 0;
1416		kmem_free(buffer);
1417		return 1;
1418	}
1419
1420	/* check partially zeroed log */
1421	error = xlog_bread(log, log_bbnum-1, 1, buffer, &offset);
1422	if (error)
1423		goto out_free_buffer;
1424
1425	last_cycle = xlog_get_cycle(offset);
1426	if (last_cycle != 0) {		/* log completely written to */
1427		kmem_free(buffer);
1428		return 0;
1429	}
1430
1431	/* we have a partially zeroed log */
1432	last_blk = log_bbnum-1;
1433	error = xlog_find_cycle_start(log, buffer, 0, &last_blk, 0);
1434	if (error)
1435		goto out_free_buffer;
1436
1437	/*
1438	 * Validate the answer.  Because there is no way to guarantee that
1439	 * the entire log is made up of log records which are the same size,
1440	 * we scan over the defined maximum blocks.  At this point, the maximum
1441	 * is not chosen to mean anything special.   XXXmiken
1442	 */
1443	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1444	ASSERT(num_scan_bblks <= INT_MAX);
1445
1446	if (last_blk < num_scan_bblks)
1447		num_scan_bblks = last_blk;
1448	start_blk = last_blk - num_scan_bblks;
1449
1450	/*
1451	 * We search for any instances of cycle number 0 that occur before
1452	 * our current estimate of the head.  What we're trying to detect is
1453	 *        1 ... | 0 | 1 | 0...
1454	 *                       ^ binary search ends here
1455	 */
1456	if ((error = xlog_find_verify_cycle(log, start_blk,
1457					 (int)num_scan_bblks, 0, &new_blk)))
1458		goto out_free_buffer;
1459	if (new_blk != -1)
1460		last_blk = new_blk;
1461
1462	/*
1463	 * Potentially backup over partial log record write.  We don't need
1464	 * to search the end of the log because we know it is zero.
1465	 */
1466	error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1467	if (error == 1)
1468		error = -EIO;
1469	if (error)
1470		goto out_free_buffer;
1471
1472	*blk_no = last_blk;
1473out_free_buffer:
1474	kmem_free(buffer);
1475	if (error)
1476		return error;
1477	return 1;
1478}
1479
1480/*
1481 * These are simple subroutines used by xlog_clear_stale_blocks() below
1482 * to initialize a buffer full of empty log record headers and write
1483 * them into the log.
1484 */
1485STATIC void
1486xlog_add_record(
1487	struct xlog		*log,
1488	char			*buf,
1489	int			cycle,
1490	int			block,
1491	int			tail_cycle,
1492	int			tail_block)
1493{
1494	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1495
1496	memset(buf, 0, BBSIZE);
1497	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1498	recp->h_cycle = cpu_to_be32(cycle);
1499	recp->h_version = cpu_to_be32(
1500			xfs_has_logv2(log->l_mp) ? 2 : 1);
1501	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1502	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1503	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1504	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1505}
1506
1507STATIC int
1508xlog_write_log_records(
1509	struct xlog	*log,
1510	int		cycle,
1511	int		start_block,
1512	int		blocks,
1513	int		tail_cycle,
1514	int		tail_block)
1515{
1516	char		*offset;
1517	char		*buffer;
1518	int		balign, ealign;
1519	int		sectbb = log->l_sectBBsize;
1520	int		end_block = start_block + blocks;
1521	int		bufblks;
1522	int		error = 0;
1523	int		i, j = 0;
1524
1525	/*
1526	 * Greedily allocate a buffer big enough to handle the full
1527	 * range of basic blocks to be written.  If that fails, try
1528	 * a smaller size.  We need to be able to write at least a
1529	 * log sector, or we're out of luck.
1530	 */
1531	bufblks = roundup_pow_of_two(blocks);
1532	while (bufblks > log->l_logBBsize)
1533		bufblks >>= 1;
1534	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
1535		bufblks >>= 1;
1536		if (bufblks < sectbb)
1537			return -ENOMEM;
1538	}
1539
1540	/* We may need to do a read at the start to fill in part of
1541	 * the buffer in the starting sector not covered by the first
1542	 * write below.
1543	 */
1544	balign = round_down(start_block, sectbb);
1545	if (balign != start_block) {
1546		error = xlog_bread_noalign(log, start_block, 1, buffer);
1547		if (error)
1548			goto out_free_buffer;
1549
1550		j = start_block - balign;
1551	}
1552
1553	for (i = start_block; i < end_block; i += bufblks) {
1554		int		bcount, endcount;
1555
1556		bcount = min(bufblks, end_block - start_block);
1557		endcount = bcount - j;
1558
1559		/* We may need to do a read at the end to fill in part of
1560		 * the buffer in the final sector not covered by the write.
1561		 * If this is the same sector as the above read, skip it.
1562		 */
1563		ealign = round_down(end_block, sectbb);
1564		if (j == 0 && (start_block + endcount > ealign)) {
1565			error = xlog_bread_noalign(log, ealign, sectbb,
1566					buffer + BBTOB(ealign - start_block));
1567			if (error)
1568				break;
1569
1570		}
1571
1572		offset = buffer + xlog_align(log, start_block);
1573		for (; j < endcount; j++) {
1574			xlog_add_record(log, offset, cycle, i+j,
1575					tail_cycle, tail_block);
1576			offset += BBSIZE;
1577		}
1578		error = xlog_bwrite(log, start_block, endcount, buffer);
1579		if (error)
1580			break;
1581		start_block += endcount;
1582		j = 0;
1583	}
1584
1585out_free_buffer:
1586	kmem_free(buffer);
1587	return error;
1588}
1589
1590/*
1591 * This routine is called to blow away any incomplete log writes out
1592 * in front of the log head.  We do this so that we won't become confused
1593 * if we come up, write only a little bit more, and then crash again.
1594 * If we leave the partial log records out there, this situation could
1595 * cause us to think those partial writes are valid blocks since they
1596 * have the current cycle number.  We get rid of them by overwriting them
1597 * with empty log records with the old cycle number rather than the
1598 * current one.
1599 *
1600 * The tail lsn is passed in rather than taken from
1601 * the log so that we will not write over the unmount record after a
1602 * clean unmount in a 512 block log.  Doing so would leave the log without
1603 * any valid log records in it until a new one was written.  If we crashed
1604 * during that time we would not be able to recover.
1605 */
1606STATIC int
1607xlog_clear_stale_blocks(
1608	struct xlog	*log,
1609	xfs_lsn_t	tail_lsn)
1610{
1611	int		tail_cycle, head_cycle;
1612	int		tail_block, head_block;
1613	int		tail_distance, max_distance;
1614	int		distance;
1615	int		error;
1616
1617	tail_cycle = CYCLE_LSN(tail_lsn);
1618	tail_block = BLOCK_LSN(tail_lsn);
1619	head_cycle = log->l_curr_cycle;
1620	head_block = log->l_curr_block;
1621
1622	/*
1623	 * Figure out the distance between the new head of the log
1624	 * and the tail.  We want to write over any blocks beyond the
1625	 * head that we may have written just before the crash, but
1626	 * we don't want to overwrite the tail of the log.
1627	 */
1628	if (head_cycle == tail_cycle) {
1629		/*
1630		 * The tail is behind the head in the physical log,
1631		 * so the distance from the head to the tail is the
1632		 * distance from the head to the end of the log plus
1633		 * the distance from the beginning of the log to the
1634		 * tail.
1635		 */
1636		if (XFS_IS_CORRUPT(log->l_mp,
1637				   head_block < tail_block ||
1638				   head_block >= log->l_logBBsize))
1639			return -EFSCORRUPTED;
 
1640		tail_distance = tail_block + (log->l_logBBsize - head_block);
1641	} else {
1642		/*
1643		 * The head is behind the tail in the physical log,
1644		 * so the distance from the head to the tail is just
1645		 * the tail block minus the head block.
1646		 */
1647		if (XFS_IS_CORRUPT(log->l_mp,
1648				   head_block >= tail_block ||
1649				   head_cycle != tail_cycle + 1))
1650			return -EFSCORRUPTED;
 
1651		tail_distance = tail_block - head_block;
1652	}
1653
1654	/*
1655	 * If the head is right up against the tail, we can't clear
1656	 * anything.
1657	 */
1658	if (tail_distance <= 0) {
1659		ASSERT(tail_distance == 0);
1660		return 0;
1661	}
1662
1663	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1664	/*
1665	 * Take the smaller of the maximum amount of outstanding I/O
1666	 * we could have and the distance to the tail to clear out.
1667	 * We take the smaller so that we don't overwrite the tail and
1668	 * we don't waste all day writing from the head to the tail
1669	 * for no reason.
1670	 */
1671	max_distance = min(max_distance, tail_distance);
1672
1673	if ((head_block + max_distance) <= log->l_logBBsize) {
1674		/*
1675		 * We can stomp all the blocks we need to without
1676		 * wrapping around the end of the log.  Just do it
1677		 * in a single write.  Use the cycle number of the
1678		 * current cycle minus one so that the log will look like:
1679		 *     n ... | n - 1 ...
1680		 */
1681		error = xlog_write_log_records(log, (head_cycle - 1),
1682				head_block, max_distance, tail_cycle,
1683				tail_block);
1684		if (error)
1685			return error;
1686	} else {
1687		/*
1688		 * We need to wrap around the end of the physical log in
1689		 * order to clear all the blocks.  Do it in two separate
1690		 * I/Os.  The first write should be from the head to the
1691		 * end of the physical log, and it should use the current
1692		 * cycle number minus one just like above.
1693		 */
1694		distance = log->l_logBBsize - head_block;
1695		error = xlog_write_log_records(log, (head_cycle - 1),
1696				head_block, distance, tail_cycle,
1697				tail_block);
1698
1699		if (error)
1700			return error;
1701
1702		/*
1703		 * Now write the blocks at the start of the physical log.
1704		 * This writes the remainder of the blocks we want to clear.
1705		 * It uses the current cycle number since we're now on the
1706		 * same cycle as the head so that we get:
1707		 *    n ... n ... | n - 1 ...
1708		 *    ^^^^^ blocks we're writing
1709		 */
1710		distance = max_distance - (log->l_logBBsize - head_block);
1711		error = xlog_write_log_records(log, head_cycle, 0, distance,
1712				tail_cycle, tail_block);
1713		if (error)
1714			return error;
1715	}
1716
1717	return 0;
1718}
1719
1720/*
1721 * Release the recovered intent item in the AIL that matches the given intent
1722 * type and intent id.
1723 */
1724void
1725xlog_recover_release_intent(
1726	struct xlog			*log,
1727	unsigned short			intent_type,
1728	uint64_t			intent_id)
1729{
1730	struct xfs_defer_pending	*dfp, *n;
1731
1732	list_for_each_entry_safe(dfp, n, &log->r_dfops, dfp_list) {
1733		struct xfs_log_item	*lip = dfp->dfp_intent;
1734
1735		if (lip->li_type != intent_type)
1736			continue;
1737		if (!lip->li_ops->iop_match(lip, intent_id))
1738			continue;
1739
1740		ASSERT(xlog_item_is_intent(lip));
1741
1742		xfs_defer_cancel_recovery(log->l_mp, dfp);
1743	}
1744}
1745
1746int
1747xlog_recover_iget(
1748	struct xfs_mount	*mp,
1749	xfs_ino_t		ino,
1750	struct xfs_inode	**ipp)
1751{
1752	int			error;
1753
1754	error = xfs_iget(mp, NULL, ino, 0, 0, ipp);
1755	if (error)
1756		return error;
1757
1758	error = xfs_qm_dqattach(*ipp);
1759	if (error) {
1760		xfs_irele(*ipp);
1761		return error;
1762	}
1763
1764	if (VFS_I(*ipp)->i_nlink == 0)
1765		xfs_iflags_set(*ipp, XFS_IRECOVERY);
1766
1767	return 0;
1768}
1769
1770/******************************************************************************
1771 *
1772 *		Log recover routines
1773 *
1774 ******************************************************************************
1775 */
1776static const struct xlog_recover_item_ops *xlog_recover_item_ops[] = {
1777	&xlog_buf_item_ops,
1778	&xlog_inode_item_ops,
1779	&xlog_dquot_item_ops,
1780	&xlog_quotaoff_item_ops,
1781	&xlog_icreate_item_ops,
1782	&xlog_efi_item_ops,
1783	&xlog_efd_item_ops,
1784	&xlog_rui_item_ops,
1785	&xlog_rud_item_ops,
1786	&xlog_cui_item_ops,
1787	&xlog_cud_item_ops,
1788	&xlog_bui_item_ops,
1789	&xlog_bud_item_ops,
1790	&xlog_attri_item_ops,
1791	&xlog_attrd_item_ops,
1792};
1793
1794static const struct xlog_recover_item_ops *
1795xlog_find_item_ops(
1796	struct xlog_recover_item		*item)
1797{
1798	unsigned int				i;
1799
1800	for (i = 0; i < ARRAY_SIZE(xlog_recover_item_ops); i++)
1801		if (ITEM_TYPE(item) == xlog_recover_item_ops[i]->item_type)
1802			return xlog_recover_item_ops[i];
1803
1804	return NULL;
1805}
1806
1807/*
1808 * Sort the log items in the transaction.
1809 *
1810 * The ordering constraints are defined by the inode allocation and unlink
1811 * behaviour. The rules are:
1812 *
1813 *	1. Every item is only logged once in a given transaction. Hence it
1814 *	   represents the last logged state of the item. Hence ordering is
1815 *	   dependent on the order in which operations need to be performed so
1816 *	   required initial conditions are always met.
1817 *
1818 *	2. Cancelled buffers are recorded in pass 1 in a separate table and
1819 *	   there's nothing to replay from them so we can simply cull them
1820 *	   from the transaction. However, we can't do that until after we've
1821 *	   replayed all the other items because they may be dependent on the
1822 *	   cancelled buffer and replaying the cancelled buffer can remove it
1823 *	   form the cancelled buffer table. Hence they have tobe done last.
1824 *
1825 *	3. Inode allocation buffers must be replayed before inode items that
1826 *	   read the buffer and replay changes into it. For filesystems using the
1827 *	   ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1828 *	   treated the same as inode allocation buffers as they create and
1829 *	   initialise the buffers directly.
1830 *
1831 *	4. Inode unlink buffers must be replayed after inode items are replayed.
1832 *	   This ensures that inodes are completely flushed to the inode buffer
1833 *	   in a "free" state before we remove the unlinked inode list pointer.
1834 *
1835 * Hence the ordering needs to be inode allocation buffers first, inode items
1836 * second, inode unlink buffers third and cancelled buffers last.
1837 *
1838 * But there's a problem with that - we can't tell an inode allocation buffer
1839 * apart from a regular buffer, so we can't separate them. We can, however,
1840 * tell an inode unlink buffer from the others, and so we can separate them out
1841 * from all the other buffers and move them to last.
1842 *
1843 * Hence, 4 lists, in order from head to tail:
1844 *	- buffer_list for all buffers except cancelled/inode unlink buffers
1845 *	- item_list for all non-buffer items
1846 *	- inode_buffer_list for inode unlink buffers
1847 *	- cancel_list for the cancelled buffers
1848 *
1849 * Note that we add objects to the tail of the lists so that first-to-last
1850 * ordering is preserved within the lists. Adding objects to the head of the
1851 * list means when we traverse from the head we walk them in last-to-first
1852 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1853 * but for all other items there may be specific ordering that we need to
1854 * preserve.
1855 */
1856STATIC int
1857xlog_recover_reorder_trans(
1858	struct xlog		*log,
1859	struct xlog_recover	*trans,
1860	int			pass)
1861{
1862	struct xlog_recover_item *item, *n;
1863	int			error = 0;
1864	LIST_HEAD(sort_list);
1865	LIST_HEAD(cancel_list);
1866	LIST_HEAD(buffer_list);
1867	LIST_HEAD(inode_buffer_list);
1868	LIST_HEAD(item_list);
1869
1870	list_splice_init(&trans->r_itemq, &sort_list);
1871	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1872		enum xlog_recover_reorder	fate = XLOG_REORDER_ITEM_LIST;
1873
1874		item->ri_ops = xlog_find_item_ops(item);
1875		if (!item->ri_ops) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1876			xfs_warn(log->l_mp,
1877				"%s: unrecognized type of log operation (%d)",
1878				__func__, ITEM_TYPE(item));
1879			ASSERT(0);
1880			/*
1881			 * return the remaining items back to the transaction
1882			 * item list so they can be freed in caller.
1883			 */
1884			if (!list_empty(&sort_list))
1885				list_splice_init(&sort_list, &trans->r_itemq);
1886			error = -EFSCORRUPTED;
1887			break;
1888		}
1889
1890		if (item->ri_ops->reorder)
1891			fate = item->ri_ops->reorder(item);
1892
1893		switch (fate) {
1894		case XLOG_REORDER_BUFFER_LIST:
1895			list_move_tail(&item->ri_list, &buffer_list);
1896			break;
1897		case XLOG_REORDER_CANCEL_LIST:
1898			trace_xfs_log_recover_item_reorder_head(log,
1899					trans, item, pass);
1900			list_move(&item->ri_list, &cancel_list);
1901			break;
1902		case XLOG_REORDER_INODE_BUFFER_LIST:
1903			list_move(&item->ri_list, &inode_buffer_list);
1904			break;
1905		case XLOG_REORDER_ITEM_LIST:
1906			trace_xfs_log_recover_item_reorder_tail(log,
1907							trans, item, pass);
1908			list_move_tail(&item->ri_list, &item_list);
1909			break;
1910		}
1911	}
1912
1913	ASSERT(list_empty(&sort_list));
1914	if (!list_empty(&buffer_list))
1915		list_splice(&buffer_list, &trans->r_itemq);
1916	if (!list_empty(&item_list))
1917		list_splice_tail(&item_list, &trans->r_itemq);
1918	if (!list_empty(&inode_buffer_list))
1919		list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1920	if (!list_empty(&cancel_list))
1921		list_splice_tail(&cancel_list, &trans->r_itemq);
1922	return error;
1923}
1924
1925void
1926xlog_buf_readahead(
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1927	struct xlog		*log,
1928	xfs_daddr_t		blkno,
1929	uint			len,
1930	const struct xfs_buf_ops *ops)
1931{
1932	if (!xlog_is_buffer_cancelled(log, blkno, len))
1933		xfs_buf_readahead(log->l_mp->m_ddev_targp, blkno, len, ops);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1934}
1935
1936/*
1937 * Create a deferred work structure for resuming and tracking the progress of a
1938 * log intent item that was found during recovery.
 
 
 
 
 
 
 
 
1939 */
1940void
1941xlog_recover_intent_item(
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1942	struct xlog			*log,
1943	struct xfs_log_item		*lip,
1944	xfs_lsn_t			lsn,
1945	const struct xfs_defer_op_type	*ops)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1946{
1947	ASSERT(xlog_item_is_intent(lip));
 
 
 
 
 
 
 
 
 
 
 
1948
1949	xfs_defer_start_recovery(lip, &log->r_dfops, ops);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1950
1951	/*
1952	 * Insert the intent into the AIL directly and drop one reference so
1953	 * that finishing or canceling the work will drop the other.
1954	 */
1955	xfs_trans_ail_insert(log->l_ailp, lip, lsn);
1956	lip->li_ops->iop_unpin(lip, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1957}
1958
1959STATIC int
1960xlog_recover_items_pass2(
1961	struct xlog                     *log,
1962	struct xlog_recover             *trans,
1963	struct list_head                *buffer_list,
1964	struct list_head                *item_list)
1965{
1966	struct xlog_recover_item	*item;
1967	int				error = 0;
1968
1969	list_for_each_entry(item, item_list, ri_list) {
1970		trace_xfs_log_recover_item_recover(log, trans, item,
1971				XLOG_RECOVER_PASS2);
1972
1973		if (item->ri_ops->commit_pass2)
1974			error = item->ri_ops->commit_pass2(log, buffer_list,
1975					item, trans->r_lsn);
1976		if (error)
1977			return error;
1978	}
1979
1980	return error;
1981}
1982
1983/*
1984 * Perform the transaction.
1985 *
1986 * If the transaction modifies a buffer or inode, do it now.  Otherwise,
1987 * EFIs and EFDs get queued up by adding entries into the AIL for them.
1988 */
1989STATIC int
1990xlog_recover_commit_trans(
1991	struct xlog		*log,
1992	struct xlog_recover	*trans,
1993	int			pass,
1994	struct list_head	*buffer_list)
1995{
1996	int				error = 0;
1997	int				items_queued = 0;
1998	struct xlog_recover_item	*item;
1999	struct xlog_recover_item	*next;
2000	LIST_HEAD			(ra_list);
2001	LIST_HEAD			(done_list);
2002
2003	#define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
2004
2005	hlist_del_init(&trans->r_list);
2006
2007	error = xlog_recover_reorder_trans(log, trans, pass);
2008	if (error)
2009		return error;
2010
2011	list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
2012		trace_xfs_log_recover_item_recover(log, trans, item, pass);
2013
2014		switch (pass) {
2015		case XLOG_RECOVER_PASS1:
2016			if (item->ri_ops->commit_pass1)
2017				error = item->ri_ops->commit_pass1(log, item);
2018			break;
2019		case XLOG_RECOVER_PASS2:
2020			if (item->ri_ops->ra_pass2)
2021				item->ri_ops->ra_pass2(log, item);
2022			list_move_tail(&item->ri_list, &ra_list);
2023			items_queued++;
2024			if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
2025				error = xlog_recover_items_pass2(log, trans,
2026						buffer_list, &ra_list);
2027				list_splice_tail_init(&ra_list, &done_list);
2028				items_queued = 0;
2029			}
2030
2031			break;
2032		default:
2033			ASSERT(0);
2034		}
2035
2036		if (error)
2037			goto out;
2038	}
2039
2040out:
2041	if (!list_empty(&ra_list)) {
2042		if (!error)
2043			error = xlog_recover_items_pass2(log, trans,
2044					buffer_list, &ra_list);
2045		list_splice_tail_init(&ra_list, &done_list);
2046	}
2047
2048	if (!list_empty(&done_list))
2049		list_splice_init(&done_list, &trans->r_itemq);
2050
2051	return error;
2052}
2053
2054STATIC void
2055xlog_recover_add_item(
2056	struct list_head	*head)
2057{
2058	struct xlog_recover_item *item;
2059
2060	item = kmem_zalloc(sizeof(struct xlog_recover_item), 0);
2061	INIT_LIST_HEAD(&item->ri_list);
2062	list_add_tail(&item->ri_list, head);
2063}
2064
2065STATIC int
2066xlog_recover_add_to_cont_trans(
2067	struct xlog		*log,
2068	struct xlog_recover	*trans,
2069	char			*dp,
2070	int			len)
2071{
2072	struct xlog_recover_item *item;
2073	char			*ptr, *old_ptr;
2074	int			old_len;
2075
2076	/*
2077	 * If the transaction is empty, the header was split across this and the
2078	 * previous record. Copy the rest of the header.
2079	 */
2080	if (list_empty(&trans->r_itemq)) {
2081		ASSERT(len <= sizeof(struct xfs_trans_header));
2082		if (len > sizeof(struct xfs_trans_header)) {
2083			xfs_warn(log->l_mp, "%s: bad header length", __func__);
2084			return -EFSCORRUPTED;
2085		}
2086
2087		xlog_recover_add_item(&trans->r_itemq);
2088		ptr = (char *)&trans->r_theader +
2089				sizeof(struct xfs_trans_header) - len;
2090		memcpy(ptr, dp, len);
2091		return 0;
2092	}
2093
2094	/* take the tail entry */
2095	item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2096			  ri_list);
2097
2098	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
2099	old_len = item->ri_buf[item->ri_cnt-1].i_len;
2100
2101	ptr = kvrealloc(old_ptr, old_len, len + old_len, GFP_KERNEL);
2102	if (!ptr)
2103		return -ENOMEM;
2104	memcpy(&ptr[old_len], dp, len);
2105	item->ri_buf[item->ri_cnt-1].i_len += len;
2106	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
2107	trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
2108	return 0;
2109}
2110
2111/*
2112 * The next region to add is the start of a new region.  It could be
2113 * a whole region or it could be the first part of a new region.  Because
2114 * of this, the assumption here is that the type and size fields of all
2115 * format structures fit into the first 32 bits of the structure.
2116 *
2117 * This works because all regions must be 32 bit aligned.  Therefore, we
2118 * either have both fields or we have neither field.  In the case we have
2119 * neither field, the data part of the region is zero length.  We only have
2120 * a log_op_header and can throw away the header since a new one will appear
2121 * later.  If we have at least 4 bytes, then we can determine how many regions
2122 * will appear in the current log item.
2123 */
2124STATIC int
2125xlog_recover_add_to_trans(
2126	struct xlog		*log,
2127	struct xlog_recover	*trans,
2128	char			*dp,
2129	int			len)
2130{
2131	struct xfs_inode_log_format	*in_f;			/* any will do */
2132	struct xlog_recover_item *item;
2133	char			*ptr;
2134
2135	if (!len)
2136		return 0;
2137	if (list_empty(&trans->r_itemq)) {
2138		/* we need to catch log corruptions here */
2139		if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
2140			xfs_warn(log->l_mp, "%s: bad header magic number",
2141				__func__);
2142			ASSERT(0);
2143			return -EFSCORRUPTED;
2144		}
2145
2146		if (len > sizeof(struct xfs_trans_header)) {
2147			xfs_warn(log->l_mp, "%s: bad header length", __func__);
2148			ASSERT(0);
2149			return -EFSCORRUPTED;
2150		}
2151
2152		/*
2153		 * The transaction header can be arbitrarily split across op
2154		 * records. If we don't have the whole thing here, copy what we
2155		 * do have and handle the rest in the next record.
2156		 */
2157		if (len == sizeof(struct xfs_trans_header))
2158			xlog_recover_add_item(&trans->r_itemq);
2159		memcpy(&trans->r_theader, dp, len);
2160		return 0;
2161	}
2162
2163	ptr = kmem_alloc(len, 0);
2164	memcpy(ptr, dp, len);
2165	in_f = (struct xfs_inode_log_format *)ptr;
2166
2167	/* take the tail entry */
2168	item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2169			  ri_list);
2170	if (item->ri_total != 0 &&
2171	     item->ri_total == item->ri_cnt) {
2172		/* tail item is in use, get a new one */
2173		xlog_recover_add_item(&trans->r_itemq);
2174		item = list_entry(trans->r_itemq.prev,
2175					struct xlog_recover_item, ri_list);
2176	}
2177
2178	if (item->ri_total == 0) {		/* first region to be added */
2179		if (in_f->ilf_size == 0 ||
2180		    in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
2181			xfs_warn(log->l_mp,
2182		"bad number of regions (%d) in inode log format",
2183				  in_f->ilf_size);
2184			ASSERT(0);
2185			kmem_free(ptr);
2186			return -EFSCORRUPTED;
2187		}
2188
2189		item->ri_total = in_f->ilf_size;
2190		item->ri_buf =
2191			kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
2192				    0);
2193	}
2194
2195	if (item->ri_total <= item->ri_cnt) {
2196		xfs_warn(log->l_mp,
2197	"log item region count (%d) overflowed size (%d)",
2198				item->ri_cnt, item->ri_total);
2199		ASSERT(0);
2200		kmem_free(ptr);
2201		return -EFSCORRUPTED;
2202	}
2203
2204	/* Description region is ri_buf[0] */
2205	item->ri_buf[item->ri_cnt].i_addr = ptr;
2206	item->ri_buf[item->ri_cnt].i_len  = len;
2207	item->ri_cnt++;
2208	trace_xfs_log_recover_item_add(log, trans, item, 0);
2209	return 0;
2210}
2211
2212/*
2213 * Free up any resources allocated by the transaction
2214 *
2215 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2216 */
2217STATIC void
2218xlog_recover_free_trans(
2219	struct xlog_recover	*trans)
2220{
2221	struct xlog_recover_item *item, *n;
2222	int			i;
2223
2224	hlist_del_init(&trans->r_list);
2225
2226	list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2227		/* Free the regions in the item. */
2228		list_del(&item->ri_list);
2229		for (i = 0; i < item->ri_cnt; i++)
2230			kmem_free(item->ri_buf[i].i_addr);
2231		/* Free the item itself */
2232		kmem_free(item->ri_buf);
2233		kmem_free(item);
2234	}
2235	/* Free the transaction recover structure */
2236	kmem_free(trans);
2237}
2238
2239/*
2240 * On error or completion, trans is freed.
2241 */
2242STATIC int
2243xlog_recovery_process_trans(
2244	struct xlog		*log,
2245	struct xlog_recover	*trans,
2246	char			*dp,
2247	unsigned int		len,
2248	unsigned int		flags,
2249	int			pass,
2250	struct list_head	*buffer_list)
2251{
2252	int			error = 0;
2253	bool			freeit = false;
2254
2255	/* mask off ophdr transaction container flags */
2256	flags &= ~XLOG_END_TRANS;
2257	if (flags & XLOG_WAS_CONT_TRANS)
2258		flags &= ~XLOG_CONTINUE_TRANS;
2259
2260	/*
2261	 * Callees must not free the trans structure. We'll decide if we need to
2262	 * free it or not based on the operation being done and it's result.
2263	 */
2264	switch (flags) {
2265	/* expected flag values */
2266	case 0:
2267	case XLOG_CONTINUE_TRANS:
2268		error = xlog_recover_add_to_trans(log, trans, dp, len);
2269		break;
2270	case XLOG_WAS_CONT_TRANS:
2271		error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
2272		break;
2273	case XLOG_COMMIT_TRANS:
2274		error = xlog_recover_commit_trans(log, trans, pass,
2275						  buffer_list);
2276		/* success or fail, we are now done with this transaction. */
2277		freeit = true;
2278		break;
2279
2280	/* unexpected flag values */
2281	case XLOG_UNMOUNT_TRANS:
2282		/* just skip trans */
2283		xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
2284		freeit = true;
2285		break;
2286	case XLOG_START_TRANS:
2287	default:
2288		xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
2289		ASSERT(0);
2290		error = -EFSCORRUPTED;
2291		break;
2292	}
2293	if (error || freeit)
2294		xlog_recover_free_trans(trans);
2295	return error;
2296}
2297
2298/*
2299 * Lookup the transaction recovery structure associated with the ID in the
2300 * current ophdr. If the transaction doesn't exist and the start flag is set in
2301 * the ophdr, then allocate a new transaction for future ID matches to find.
2302 * Either way, return what we found during the lookup - an existing transaction
2303 * or nothing.
2304 */
2305STATIC struct xlog_recover *
2306xlog_recover_ophdr_to_trans(
2307	struct hlist_head	rhash[],
2308	struct xlog_rec_header	*rhead,
2309	struct xlog_op_header	*ohead)
2310{
2311	struct xlog_recover	*trans;
2312	xlog_tid_t		tid;
2313	struct hlist_head	*rhp;
2314
2315	tid = be32_to_cpu(ohead->oh_tid);
2316	rhp = &rhash[XLOG_RHASH(tid)];
2317	hlist_for_each_entry(trans, rhp, r_list) {
2318		if (trans->r_log_tid == tid)
2319			return trans;
2320	}
2321
2322	/*
2323	 * skip over non-start transaction headers - we could be
2324	 * processing slack space before the next transaction starts
2325	 */
2326	if (!(ohead->oh_flags & XLOG_START_TRANS))
2327		return NULL;
2328
2329	ASSERT(be32_to_cpu(ohead->oh_len) == 0);
2330
2331	/*
2332	 * This is a new transaction so allocate a new recovery container to
2333	 * hold the recovery ops that will follow.
2334	 */
2335	trans = kmem_zalloc(sizeof(struct xlog_recover), 0);
2336	trans->r_log_tid = tid;
2337	trans->r_lsn = be64_to_cpu(rhead->h_lsn);
2338	INIT_LIST_HEAD(&trans->r_itemq);
2339	INIT_HLIST_NODE(&trans->r_list);
2340	hlist_add_head(&trans->r_list, rhp);
2341
2342	/*
2343	 * Nothing more to do for this ophdr. Items to be added to this new
2344	 * transaction will be in subsequent ophdr containers.
2345	 */
2346	return NULL;
2347}
2348
2349STATIC int
2350xlog_recover_process_ophdr(
2351	struct xlog		*log,
2352	struct hlist_head	rhash[],
2353	struct xlog_rec_header	*rhead,
2354	struct xlog_op_header	*ohead,
2355	char			*dp,
2356	char			*end,
2357	int			pass,
2358	struct list_head	*buffer_list)
2359{
2360	struct xlog_recover	*trans;
2361	unsigned int		len;
2362	int			error;
2363
2364	/* Do we understand who wrote this op? */
2365	if (ohead->oh_clientid != XFS_TRANSACTION &&
2366	    ohead->oh_clientid != XFS_LOG) {
2367		xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
2368			__func__, ohead->oh_clientid);
2369		ASSERT(0);
2370		return -EFSCORRUPTED;
2371	}
2372
2373	/*
2374	 * Check the ophdr contains all the data it is supposed to contain.
2375	 */
2376	len = be32_to_cpu(ohead->oh_len);
2377	if (dp + len > end) {
2378		xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
2379		WARN_ON(1);
2380		return -EFSCORRUPTED;
2381	}
2382
2383	trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
2384	if (!trans) {
2385		/* nothing to do, so skip over this ophdr */
2386		return 0;
2387	}
2388
2389	/*
2390	 * The recovered buffer queue is drained only once we know that all
2391	 * recovery items for the current LSN have been processed. This is
2392	 * required because:
2393	 *
2394	 * - Buffer write submission updates the metadata LSN of the buffer.
2395	 * - Log recovery skips items with a metadata LSN >= the current LSN of
2396	 *   the recovery item.
2397	 * - Separate recovery items against the same metadata buffer can share
2398	 *   a current LSN. I.e., consider that the LSN of a recovery item is
2399	 *   defined as the starting LSN of the first record in which its
2400	 *   transaction appears, that a record can hold multiple transactions,
2401	 *   and/or that a transaction can span multiple records.
2402	 *
2403	 * In other words, we are allowed to submit a buffer from log recovery
2404	 * once per current LSN. Otherwise, we may incorrectly skip recovery
2405	 * items and cause corruption.
2406	 *
2407	 * We don't know up front whether buffers are updated multiple times per
2408	 * LSN. Therefore, track the current LSN of each commit log record as it
2409	 * is processed and drain the queue when it changes. Use commit records
2410	 * because they are ordered correctly by the logging code.
2411	 */
2412	if (log->l_recovery_lsn != trans->r_lsn &&
2413	    ohead->oh_flags & XLOG_COMMIT_TRANS) {
2414		error = xfs_buf_delwri_submit(buffer_list);
2415		if (error)
2416			return error;
2417		log->l_recovery_lsn = trans->r_lsn;
2418	}
2419
2420	return xlog_recovery_process_trans(log, trans, dp, len,
2421					   ohead->oh_flags, pass, buffer_list);
2422}
2423
2424/*
2425 * There are two valid states of the r_state field.  0 indicates that the
2426 * transaction structure is in a normal state.  We have either seen the
2427 * start of the transaction or the last operation we added was not a partial
2428 * operation.  If the last operation we added to the transaction was a
2429 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2430 *
2431 * NOTE: skip LRs with 0 data length.
2432 */
2433STATIC int
2434xlog_recover_process_data(
2435	struct xlog		*log,
2436	struct hlist_head	rhash[],
2437	struct xlog_rec_header	*rhead,
2438	char			*dp,
2439	int			pass,
2440	struct list_head	*buffer_list)
2441{
2442	struct xlog_op_header	*ohead;
2443	char			*end;
2444	int			num_logops;
2445	int			error;
2446
2447	end = dp + be32_to_cpu(rhead->h_len);
2448	num_logops = be32_to_cpu(rhead->h_num_logops);
2449
2450	/* check the log format matches our own - else we can't recover */
2451	if (xlog_header_check_recover(log->l_mp, rhead))
2452		return -EIO;
2453
2454	trace_xfs_log_recover_record(log, rhead, pass);
2455	while ((dp < end) && num_logops) {
2456
2457		ohead = (struct xlog_op_header *)dp;
2458		dp += sizeof(*ohead);
2459		ASSERT(dp <= end);
2460
2461		/* errors will abort recovery */
2462		error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
2463						   dp, end, pass, buffer_list);
2464		if (error)
2465			return error;
2466
2467		dp += be32_to_cpu(ohead->oh_len);
2468		num_logops--;
2469	}
2470	return 0;
2471}
2472
2473/* Take all the collected deferred ops and finish them in order. */
2474static int
2475xlog_finish_defer_ops(
2476	struct xfs_mount	*mp,
2477	struct list_head	*capture_list)
 
2478{
2479	struct xfs_defer_capture *dfc, *next;
2480	struct xfs_trans	*tp;
2481	int			error = 0;
2482
2483	list_for_each_entry_safe(dfc, next, capture_list, dfc_list) {
2484		struct xfs_trans_res	resv;
2485		struct xfs_defer_resources dres;
2486
2487		/*
2488		 * Create a new transaction reservation from the captured
2489		 * information.  Set logcount to 1 to force the new transaction
2490		 * to regrant every roll so that we can make forward progress
2491		 * in recovery no matter how full the log might be.
2492		 */
2493		resv.tr_logres = dfc->dfc_logres;
2494		resv.tr_logcount = 1;
2495		resv.tr_logflags = XFS_TRANS_PERM_LOG_RES;
2496
2497		error = xfs_trans_alloc(mp, &resv, dfc->dfc_blkres,
2498				dfc->dfc_rtxres, XFS_TRANS_RESERVE, &tp);
2499		if (error) {
2500			xlog_force_shutdown(mp->m_log, SHUTDOWN_LOG_IO_ERROR);
2501			return error;
2502		}
2503
2504		/*
2505		 * Transfer to this new transaction all the dfops we captured
2506		 * from recovering a single intent item.
2507		 */
2508		list_del_init(&dfc->dfc_list);
2509		xfs_defer_ops_continue(dfc, tp, &dres);
2510		error = xfs_trans_commit(tp);
2511		xfs_defer_resources_rele(&dres);
2512		if (error)
2513			return error;
2514	}
2515
2516	ASSERT(list_empty(capture_list));
2517	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
2518}
2519
2520/* Release all the captured defer ops and capture structures in this list. */
2521static void
2522xlog_abort_defer_ops(
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2523	struct xfs_mount		*mp,
2524	struct list_head		*capture_list)
 
2525{
2526	struct xfs_defer_capture	*dfc;
2527	struct xfs_defer_capture	*next;
 
2528
2529	list_for_each_entry_safe(dfc, next, capture_list, dfc_list) {
2530		list_del_init(&dfc->dfc_list);
2531		xfs_defer_ops_capture_abort(mp, dfc);
 
 
 
 
 
 
 
 
 
 
 
 
 
2532	}
2533}
2534
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2535/*
2536 * When this is called, all of the log intent items which did not have
2537 * corresponding log done items should be in the AIL.  What we do now is update
2538 * the data structures associated with each one.
 
 
 
 
 
 
 
2539 *
2540 * Since we process the log intent items in normal transactions, they will be
2541 * removed at some point after the commit.  This prevents us from just walking
2542 * down the list processing each one.  We'll use a flag in the intent item to
2543 * skip those that we've already processed and use the AIL iteration mechanism's
2544 * generation count to try to speed this up at least a bit.
2545 *
2546 * When we start, we know that the intents are the only things in the AIL. As we
2547 * process them, however, other items are added to the AIL. Hence we know we
2548 * have started recovery on all the pending intents when we find an non-intent
2549 * item in the AIL.
2550 */
2551STATIC int
2552xlog_recover_process_intents(
2553	struct xlog			*log)
2554{
2555	LIST_HEAD(capture_list);
2556	struct xfs_defer_pending	*dfp, *n;
2557	int				error = 0;
 
 
2558#if defined(DEBUG) || defined(XFS_WARN)
2559	xfs_lsn_t			last_lsn;
 
2560
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2561	last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
2562#endif
2563
2564	list_for_each_entry_safe(dfp, n, &log->r_dfops, dfp_list) {
2565		ASSERT(xlog_item_is_intent(dfp->dfp_intent));
 
 
 
 
 
 
 
 
 
2566
2567		/*
2568		 * We should never see a redo item with a LSN higher than
2569		 * the last transaction we found in the log at the start
2570		 * of recovery.
2571		 */
2572		ASSERT(XFS_LSN_CMP(last_lsn, dfp->dfp_intent->li_lsn) >= 0);
2573
2574		/*
2575		 * NOTE: If your intent processing routine can create more
2576		 * deferred ops, you /must/ attach them to the capture list in
2577		 * the recover routine or else those subsequent intents will be
2578		 * replayed in the wrong order!
2579		 *
2580		 * The recovery function can free the log item, so we must not
2581		 * access dfp->dfp_intent after it returns.  It must dispose of
2582		 * @dfp if it returns 0.
2583		 */
2584		error = xfs_defer_finish_recovery(log->l_mp, dfp,
2585				&capture_list);
2586		if (error)
 
 
 
 
 
 
 
 
 
2587			break;
 
 
 
 
2588	}
2589	if (error)
2590		goto err;
2591
2592	error = xlog_finish_defer_ops(log->l_mp, &capture_list);
2593	if (error)
2594		goto err;
2595
2596	return 0;
2597err:
2598	xlog_abort_defer_ops(log->l_mp, &capture_list);
2599	return error;
2600}
2601
2602/*
2603 * A cancel occurs when the mount has failed and we're bailing out.  Release all
2604 * pending log intent items that we haven't started recovery on so they don't
2605 * pin the AIL.
2606 */
2607STATIC void
2608xlog_recover_cancel_intents(
2609	struct xlog			*log)
2610{
2611	struct xfs_defer_pending	*dfp, *n;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2612
2613	list_for_each_entry_safe(dfp, n, &log->r_dfops, dfp_list) {
2614		ASSERT(xlog_item_is_intent(dfp->dfp_intent));
 
 
 
 
 
 
 
 
 
 
 
 
2615
2616		xfs_defer_cancel_recovery(log->l_mp, dfp);
2617	}
2618}
2619
2620/*
2621 * Transfer ownership of the recovered pending work to the recovery transaction
2622 * and try to finish the work.  If there is more work to be done, the dfp will
2623 * remain attached to the transaction.  If not, the dfp is freed.
2624 */
2625int
2626xlog_recover_finish_intent(
2627	struct xfs_trans		*tp,
2628	struct xfs_defer_pending	*dfp)
2629{
2630	int				error;
2631
2632	list_move(&dfp->dfp_list, &tp->t_dfops);
2633	error = xfs_defer_finish_one(tp, dfp);
2634	if (error == -EAGAIN)
2635		return 0;
2636	return error;
2637}
2638
2639/*
2640 * This routine performs a transaction to null out a bad inode pointer
2641 * in an agi unlinked inode hash bucket.
2642 */
2643STATIC void
2644xlog_recover_clear_agi_bucket(
2645	struct xfs_perag	*pag,
2646	int			bucket)
2647{
2648	struct xfs_mount	*mp = pag->pag_mount;
2649	struct xfs_trans	*tp;
2650	struct xfs_agi		*agi;
2651	struct xfs_buf		*agibp;
2652	int			offset;
2653	int			error;
2654
2655	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
2656	if (error)
2657		goto out_error;
2658
2659	error = xfs_read_agi(pag, tp, &agibp);
2660	if (error)
2661		goto out_abort;
2662
2663	agi = agibp->b_addr;
2664	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
2665	offset = offsetof(xfs_agi_t, agi_unlinked) +
2666		 (sizeof(xfs_agino_t) * bucket);
2667	xfs_trans_log_buf(tp, agibp, offset,
2668			  (offset + sizeof(xfs_agino_t) - 1));
2669
2670	error = xfs_trans_commit(tp);
2671	if (error)
2672		goto out_error;
2673	return;
2674
2675out_abort:
2676	xfs_trans_cancel(tp);
2677out_error:
2678	xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__,
2679			pag->pag_agno);
2680	return;
2681}
2682
2683static int
2684xlog_recover_iunlink_bucket(
2685	struct xfs_perag	*pag,
2686	struct xfs_agi		*agi,
2687	int			bucket)
2688{
2689	struct xfs_mount	*mp = pag->pag_mount;
2690	struct xfs_inode	*prev_ip = NULL;
2691	struct xfs_inode	*ip;
2692	xfs_agino_t		prev_agino, agino;
2693	int			error = 0;
 
2694
2695	agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2696	while (agino != NULLAGINO) {
2697		error = xfs_iget(mp, NULL,
2698				XFS_AGINO_TO_INO(mp, pag->pag_agno, agino),
2699				0, 0, &ip);
2700		if (error)
2701			break;
2702
2703		ASSERT(VFS_I(ip)->i_nlink == 0);
2704		ASSERT(VFS_I(ip)->i_mode != 0);
2705		xfs_iflags_clear(ip, XFS_IRECOVERY);
2706		agino = ip->i_next_unlinked;
2707
2708		if (prev_ip) {
2709			ip->i_prev_unlinked = prev_agino;
2710			xfs_irele(prev_ip);
2711
2712			/*
2713			 * Ensure the inode is removed from the unlinked list
2714			 * before we continue so that it won't race with
2715			 * building the in-memory list here. This could be
2716			 * serialised with the agibp lock, but that just
2717			 * serialises via lockstepping and it's much simpler
2718			 * just to flush the inodegc queue and wait for it to
2719			 * complete.
2720			 */
2721			error = xfs_inodegc_flush(mp);
2722			if (error)
2723				break;
2724		}
2725
2726		prev_agino = agino;
2727		prev_ip = ip;
2728	}
2729
2730	if (prev_ip) {
2731		int	error2;
 
 
 
2732
2733		ip->i_prev_unlinked = prev_agino;
2734		xfs_irele(prev_ip);
2735
2736		error2 = xfs_inodegc_flush(mp);
2737		if (error2 && !error)
2738			return error2;
2739	}
2740	return error;
 
 
 
 
 
 
 
 
2741}
2742
2743/*
2744 * Recover AGI unlinked lists
2745 *
2746 * This is called during recovery to process any inodes which we unlinked but
2747 * not freed when the system crashed.  These inodes will be on the lists in the
2748 * AGI blocks. What we do here is scan all the AGIs and fully truncate and free
2749 * any inodes found on the lists. Each inode is removed from the lists when it
2750 * has been fully truncated and is freed. The freeing of the inode and its
2751 * removal from the list must be atomic.
2752 *
2753 * If everything we touch in the agi processing loop is already in memory, this
2754 * loop can hold the cpu for a long time. It runs without lock contention,
2755 * memory allocation contention, the need wait for IO, etc, and so will run
2756 * until we either run out of inodes to process, run low on memory or we run out
2757 * of log space.
2758 *
2759 * This behaviour is bad for latency on single CPU and non-preemptible kernels,
2760 * and can prevent other filesystem work (such as CIL pushes) from running. This
2761 * can lead to deadlocks if the recovery process runs out of log reservation
2762 * space. Hence we need to yield the CPU when there is other kernel work
2763 * scheduled on this CPU to ensure other scheduled work can run without undue
2764 * latency.
2765 */
2766static void
2767xlog_recover_iunlink_ag(
2768	struct xfs_perag	*pag)
2769{
2770	struct xfs_agi		*agi;
2771	struct xfs_buf		*agibp;
2772	int			bucket;
2773	int			error;
 
 
 
2774
2775	error = xfs_read_agi(pag, NULL, &agibp);
2776	if (error) {
 
2777		/*
2778		 * AGI is b0rked. Don't process it.
2779		 *
2780		 * We should probably mark the filesystem as corrupt after we've
2781		 * recovered all the ag's we can....
2782		 */
2783		return;
2784	}
2785
2786	/*
2787	 * Unlock the buffer so that it can be acquired in the normal course of
2788	 * the transaction to truncate and free each inode.  Because we are not
2789	 * racing with anyone else here for the AGI buffer, we don't even need
2790	 * to hold it locked to read the initial unlinked bucket entries out of
2791	 * the buffer. We keep buffer reference though, so that it stays pinned
2792	 * in memory while we need the buffer.
2793	 */
2794	agi = agibp->b_addr;
2795	xfs_buf_unlock(agibp);
2796
2797	for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
2798		error = xlog_recover_iunlink_bucket(pag, agi, bucket);
2799		if (error) {
2800			/*
2801			 * Bucket is unrecoverable, so only a repair scan can
2802			 * free the remaining unlinked inodes. Just empty the
2803			 * bucket and remaining inodes on it unreferenced and
2804			 * unfreeable.
2805			 */
2806			xlog_recover_clear_agi_bucket(pag, bucket);
2807		}
2808	}
2809
2810	xfs_buf_rele(agibp);
2811}
2812
2813static void
2814xlog_recover_process_iunlinks(
2815	struct xlog	*log)
2816{
2817	struct xfs_perag	*pag;
2818	xfs_agnumber_t		agno;
2819
2820	for_each_perag(log->l_mp, agno, pag)
2821		xlog_recover_iunlink_ag(pag);
 
 
 
 
 
 
 
 
2822}
2823
2824STATIC void
2825xlog_unpack_data(
2826	struct xlog_rec_header	*rhead,
2827	char			*dp,
2828	struct xlog		*log)
2829{
2830	int			i, j, k;
2831
2832	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
2833		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
2834		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
2835		dp += BBSIZE;
2836	}
2837
2838	if (xfs_has_logv2(log->l_mp)) {
2839		xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
2840		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
2841			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2842			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2843			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
2844			dp += BBSIZE;
2845		}
2846	}
2847}
2848
2849/*
2850 * CRC check, unpack and process a log record.
2851 */
2852STATIC int
2853xlog_recover_process(
2854	struct xlog		*log,
2855	struct hlist_head	rhash[],
2856	struct xlog_rec_header	*rhead,
2857	char			*dp,
2858	int			pass,
2859	struct list_head	*buffer_list)
2860{
2861	__le32			old_crc = rhead->h_crc;
2862	__le32			crc;
2863
2864	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
2865
2866	/*
2867	 * Nothing else to do if this is a CRC verification pass. Just return
2868	 * if this a record with a non-zero crc. Unfortunately, mkfs always
2869	 * sets old_crc to 0 so we must consider this valid even on v5 supers.
2870	 * Otherwise, return EFSBADCRC on failure so the callers up the stack
2871	 * know precisely what failed.
2872	 */
2873	if (pass == XLOG_RECOVER_CRCPASS) {
2874		if (old_crc && crc != old_crc)
2875			return -EFSBADCRC;
2876		return 0;
2877	}
2878
2879	/*
2880	 * We're in the normal recovery path. Issue a warning if and only if the
2881	 * CRC in the header is non-zero. This is an advisory warning and the
2882	 * zero CRC check prevents warnings from being emitted when upgrading
2883	 * the kernel from one that does not add CRCs by default.
2884	 */
2885	if (crc != old_crc) {
2886		if (old_crc || xfs_has_crc(log->l_mp)) {
2887			xfs_alert(log->l_mp,
2888		"log record CRC mismatch: found 0x%x, expected 0x%x.",
2889					le32_to_cpu(old_crc),
2890					le32_to_cpu(crc));
2891			xfs_hex_dump(dp, 32);
2892		}
2893
2894		/*
2895		 * If the filesystem is CRC enabled, this mismatch becomes a
2896		 * fatal log corruption failure.
2897		 */
2898		if (xfs_has_crc(log->l_mp)) {
2899			XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, log->l_mp);
2900			return -EFSCORRUPTED;
2901		}
2902	}
2903
2904	xlog_unpack_data(rhead, dp, log);
2905
2906	return xlog_recover_process_data(log, rhash, rhead, dp, pass,
2907					 buffer_list);
2908}
2909
2910STATIC int
2911xlog_valid_rec_header(
2912	struct xlog		*log,
2913	struct xlog_rec_header	*rhead,
2914	xfs_daddr_t		blkno,
2915	int			bufsize)
2916{
2917	int			hlen;
2918
2919	if (XFS_IS_CORRUPT(log->l_mp,
2920			   rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM)))
 
2921		return -EFSCORRUPTED;
2922	if (XFS_IS_CORRUPT(log->l_mp,
2923			   (!rhead->h_version ||
2924			   (be32_to_cpu(rhead->h_version) &
2925			    (~XLOG_VERSION_OKBITS))))) {
2926		xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
2927			__func__, be32_to_cpu(rhead->h_version));
2928		return -EFSCORRUPTED;
2929	}
2930
2931	/*
2932	 * LR body must have data (or it wouldn't have been written)
2933	 * and h_len must not be greater than LR buffer size.
2934	 */
2935	hlen = be32_to_cpu(rhead->h_len);
2936	if (XFS_IS_CORRUPT(log->l_mp, hlen <= 0 || hlen > bufsize))
 
 
2937		return -EFSCORRUPTED;
2938
2939	if (XFS_IS_CORRUPT(log->l_mp,
2940			   blkno > log->l_logBBsize || blkno > INT_MAX))
 
2941		return -EFSCORRUPTED;
 
2942	return 0;
2943}
2944
2945/*
2946 * Read the log from tail to head and process the log records found.
2947 * Handle the two cases where the tail and head are in the same cycle
2948 * and where the active portion of the log wraps around the end of
2949 * the physical log separately.  The pass parameter is passed through
2950 * to the routines called to process the data and is not looked at
2951 * here.
2952 */
2953STATIC int
2954xlog_do_recovery_pass(
2955	struct xlog		*log,
2956	xfs_daddr_t		head_blk,
2957	xfs_daddr_t		tail_blk,
2958	int			pass,
2959	xfs_daddr_t		*first_bad)	/* out: first bad log rec */
2960{
2961	xlog_rec_header_t	*rhead;
2962	xfs_daddr_t		blk_no, rblk_no;
2963	xfs_daddr_t		rhead_blk;
2964	char			*offset;
2965	char			*hbp, *dbp;
2966	int			error = 0, h_size, h_len;
2967	int			error2 = 0;
2968	int			bblks, split_bblks;
2969	int			hblks, split_hblks, wrapped_hblks;
2970	int			i;
2971	struct hlist_head	rhash[XLOG_RHASH_SIZE];
2972	LIST_HEAD		(buffer_list);
2973
2974	ASSERT(head_blk != tail_blk);
2975	blk_no = rhead_blk = tail_blk;
2976
2977	for (i = 0; i < XLOG_RHASH_SIZE; i++)
2978		INIT_HLIST_HEAD(&rhash[i]);
2979
2980	/*
2981	 * Read the header of the tail block and get the iclog buffer size from
2982	 * h_size.  Use this to tell how many sectors make up the log header.
2983	 */
2984	if (xfs_has_logv2(log->l_mp)) {
2985		/*
2986		 * When using variable length iclogs, read first sector of
2987		 * iclog header and extract the header size from it.  Get a
2988		 * new hbp that is the correct size.
2989		 */
2990		hbp = xlog_alloc_buffer(log, 1);
2991		if (!hbp)
2992			return -ENOMEM;
2993
2994		error = xlog_bread(log, tail_blk, 1, hbp, &offset);
2995		if (error)
2996			goto bread_err1;
2997
2998		rhead = (xlog_rec_header_t *)offset;
 
 
 
2999
3000		/*
3001		 * xfsprogs has a bug where record length is based on lsunit but
3002		 * h_size (iclog size) is hardcoded to 32k. Now that we
3003		 * unconditionally CRC verify the unmount record, this means the
3004		 * log buffer can be too small for the record and cause an
3005		 * overrun.
3006		 *
3007		 * Detect this condition here. Use lsunit for the buffer size as
3008		 * long as this looks like the mkfs case. Otherwise, return an
3009		 * error to avoid a buffer overrun.
3010		 */
3011		h_size = be32_to_cpu(rhead->h_size);
3012		h_len = be32_to_cpu(rhead->h_len);
3013		if (h_len > h_size && h_len <= log->l_mp->m_logbsize &&
3014		    rhead->h_num_logops == cpu_to_be32(1)) {
3015			xfs_warn(log->l_mp,
 
3016		"invalid iclog size (%d bytes), using lsunit (%d bytes)",
3017				 h_size, log->l_mp->m_logbsize);
3018			h_size = log->l_mp->m_logbsize;
 
 
3019		}
3020
3021		error = xlog_valid_rec_header(log, rhead, tail_blk, h_size);
3022		if (error)
3023			goto bread_err1;
3024
3025		hblks = xlog_logrec_hblks(log, rhead);
3026		if (hblks != 1) {
3027			kmem_free(hbp);
3028			hbp = xlog_alloc_buffer(log, hblks);
 
 
3029		}
3030	} else {
3031		ASSERT(log->l_sectBBsize == 1);
3032		hblks = 1;
3033		hbp = xlog_alloc_buffer(log, 1);
3034		h_size = XLOG_BIG_RECORD_BSIZE;
3035	}
3036
3037	if (!hbp)
3038		return -ENOMEM;
3039	dbp = xlog_alloc_buffer(log, BTOBB(h_size));
3040	if (!dbp) {
3041		kmem_free(hbp);
3042		return -ENOMEM;
3043	}
3044
3045	memset(rhash, 0, sizeof(rhash));
3046	if (tail_blk > head_blk) {
3047		/*
3048		 * Perform recovery around the end of the physical log.
3049		 * When the head is not on the same cycle number as the tail,
3050		 * we can't do a sequential recovery.
3051		 */
3052		while (blk_no < log->l_logBBsize) {
3053			/*
3054			 * Check for header wrapping around physical end-of-log
3055			 */
3056			offset = hbp;
3057			split_hblks = 0;
3058			wrapped_hblks = 0;
3059			if (blk_no + hblks <= log->l_logBBsize) {
3060				/* Read header in one read */
3061				error = xlog_bread(log, blk_no, hblks, hbp,
3062						   &offset);
3063				if (error)
3064					goto bread_err2;
3065			} else {
3066				/* This LR is split across physical log end */
3067				if (blk_no != log->l_logBBsize) {
3068					/* some data before physical log end */
3069					ASSERT(blk_no <= INT_MAX);
3070					split_hblks = log->l_logBBsize - (int)blk_no;
3071					ASSERT(split_hblks > 0);
3072					error = xlog_bread(log, blk_no,
3073							   split_hblks, hbp,
3074							   &offset);
3075					if (error)
3076						goto bread_err2;
3077				}
3078
3079				/*
3080				 * Note: this black magic still works with
3081				 * large sector sizes (non-512) only because:
3082				 * - we increased the buffer size originally
3083				 *   by 1 sector giving us enough extra space
3084				 *   for the second read;
3085				 * - the log start is guaranteed to be sector
3086				 *   aligned;
3087				 * - we read the log end (LR header start)
3088				 *   _first_, then the log start (LR header end)
3089				 *   - order is important.
3090				 */
3091				wrapped_hblks = hblks - split_hblks;
3092				error = xlog_bread_noalign(log, 0,
3093						wrapped_hblks,
3094						offset + BBTOB(split_hblks));
3095				if (error)
3096					goto bread_err2;
3097			}
3098			rhead = (xlog_rec_header_t *)offset;
3099			error = xlog_valid_rec_header(log, rhead,
3100					split_hblks ? blk_no : 0, h_size);
3101			if (error)
3102				goto bread_err2;
3103
3104			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3105			blk_no += hblks;
3106
3107			/*
3108			 * Read the log record data in multiple reads if it
3109			 * wraps around the end of the log. Note that if the
3110			 * header already wrapped, blk_no could point past the
3111			 * end of the log. The record data is contiguous in
3112			 * that case.
3113			 */
3114			if (blk_no + bblks <= log->l_logBBsize ||
3115			    blk_no >= log->l_logBBsize) {
3116				rblk_no = xlog_wrap_logbno(log, blk_no);
3117				error = xlog_bread(log, rblk_no, bblks, dbp,
3118						   &offset);
3119				if (error)
3120					goto bread_err2;
3121			} else {
3122				/* This log record is split across the
3123				 * physical end of log */
3124				offset = dbp;
3125				split_bblks = 0;
3126				if (blk_no != log->l_logBBsize) {
3127					/* some data is before the physical
3128					 * end of log */
3129					ASSERT(!wrapped_hblks);
3130					ASSERT(blk_no <= INT_MAX);
3131					split_bblks =
3132						log->l_logBBsize - (int)blk_no;
3133					ASSERT(split_bblks > 0);
3134					error = xlog_bread(log, blk_no,
3135							split_bblks, dbp,
3136							&offset);
3137					if (error)
3138						goto bread_err2;
3139				}
3140
3141				/*
3142				 * Note: this black magic still works with
3143				 * large sector sizes (non-512) only because:
3144				 * - we increased the buffer size originally
3145				 *   by 1 sector giving us enough extra space
3146				 *   for the second read;
3147				 * - the log start is guaranteed to be sector
3148				 *   aligned;
3149				 * - we read the log end (LR header start)
3150				 *   _first_, then the log start (LR header end)
3151				 *   - order is important.
3152				 */
3153				error = xlog_bread_noalign(log, 0,
3154						bblks - split_bblks,
3155						offset + BBTOB(split_bblks));
3156				if (error)
3157					goto bread_err2;
3158			}
3159
3160			error = xlog_recover_process(log, rhash, rhead, offset,
3161						     pass, &buffer_list);
3162			if (error)
3163				goto bread_err2;
3164
3165			blk_no += bblks;
3166			rhead_blk = blk_no;
3167		}
3168
3169		ASSERT(blk_no >= log->l_logBBsize);
3170		blk_no -= log->l_logBBsize;
3171		rhead_blk = blk_no;
3172	}
3173
3174	/* read first part of physical log */
3175	while (blk_no < head_blk) {
3176		error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3177		if (error)
3178			goto bread_err2;
3179
3180		rhead = (xlog_rec_header_t *)offset;
3181		error = xlog_valid_rec_header(log, rhead, blk_no, h_size);
3182		if (error)
3183			goto bread_err2;
3184
3185		/* blocks in data section */
3186		bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3187		error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3188				   &offset);
3189		if (error)
3190			goto bread_err2;
3191
3192		error = xlog_recover_process(log, rhash, rhead, offset, pass,
3193					     &buffer_list);
3194		if (error)
3195			goto bread_err2;
3196
3197		blk_no += bblks + hblks;
3198		rhead_blk = blk_no;
3199	}
3200
3201 bread_err2:
3202	kmem_free(dbp);
3203 bread_err1:
3204	kmem_free(hbp);
3205
3206	/*
3207	 * Submit buffers that have been added from the last record processed,
3208	 * regardless of error status.
3209	 */
3210	if (!list_empty(&buffer_list))
3211		error2 = xfs_buf_delwri_submit(&buffer_list);
3212
3213	if (error && first_bad)
3214		*first_bad = rhead_blk;
3215
3216	/*
3217	 * Transactions are freed at commit time but transactions without commit
3218	 * records on disk are never committed. Free any that may be left in the
3219	 * hash table.
3220	 */
3221	for (i = 0; i < XLOG_RHASH_SIZE; i++) {
3222		struct hlist_node	*tmp;
3223		struct xlog_recover	*trans;
3224
3225		hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
3226			xlog_recover_free_trans(trans);
3227	}
3228
3229	return error ? error : error2;
3230}
3231
3232/*
3233 * Do the recovery of the log.  We actually do this in two phases.
3234 * The two passes are necessary in order to implement the function
3235 * of cancelling a record written into the log.  The first pass
3236 * determines those things which have been cancelled, and the
3237 * second pass replays log items normally except for those which
3238 * have been cancelled.  The handling of the replay and cancellations
3239 * takes place in the log item type specific routines.
3240 *
3241 * The table of items which have cancel records in the log is allocated
3242 * and freed at this level, since only here do we know when all of
3243 * the log recovery has been completed.
3244 */
3245STATIC int
3246xlog_do_log_recovery(
3247	struct xlog	*log,
3248	xfs_daddr_t	head_blk,
3249	xfs_daddr_t	tail_blk)
3250{
3251	int		error;
3252
3253	ASSERT(head_blk != tail_blk);
3254
3255	/*
3256	 * First do a pass to find all of the cancelled buf log items.
3257	 * Store them in the buf_cancel_table for use in the second pass.
3258	 */
3259	error = xlog_alloc_buf_cancel_table(log);
3260	if (error)
3261		return error;
 
 
3262
3263	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3264				      XLOG_RECOVER_PASS1, NULL);
3265	if (error != 0)
3266		goto out_cancel;
3267
 
 
3268	/*
3269	 * Then do a second pass to actually recover the items in the log.
3270	 * When it is complete free the table of buf cancel items.
3271	 */
3272	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3273				      XLOG_RECOVER_PASS2, NULL);
3274	if (!error)
3275		xlog_check_buf_cancel_table(log);
3276out_cancel:
3277	xlog_free_buf_cancel_table(log);
 
 
 
 
 
 
 
 
3278	return error;
3279}
3280
3281/*
3282 * Do the actual recovery
3283 */
3284STATIC int
3285xlog_do_recover(
3286	struct xlog		*log,
3287	xfs_daddr_t		head_blk,
3288	xfs_daddr_t		tail_blk)
3289{
3290	struct xfs_mount	*mp = log->l_mp;
3291	struct xfs_buf		*bp = mp->m_sb_bp;
3292	struct xfs_sb		*sbp = &mp->m_sb;
3293	int			error;
3294
3295	trace_xfs_log_recover(log, head_blk, tail_blk);
3296
3297	/*
3298	 * First replay the images in the log.
3299	 */
3300	error = xlog_do_log_recovery(log, head_blk, tail_blk);
3301	if (error)
3302		return error;
3303
3304	if (xlog_is_shutdown(log))
 
 
 
3305		return -EIO;
 
3306
3307	/*
3308	 * We now update the tail_lsn since much of the recovery has completed
3309	 * and there may be space available to use.  If there were no extent
3310	 * or iunlinks, we can free up the entire log and set the tail_lsn to
3311	 * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3312	 * lsn of the last known good LR on disk.  If there are extent frees
3313	 * or iunlinks they will have some entries in the AIL; so we look at
3314	 * the AIL to determine how to set the tail_lsn.
3315	 */
3316	xlog_assign_tail_lsn(mp);
3317
3318	/*
3319	 * Now that we've finished replaying all buffer and inode updates,
3320	 * re-read the superblock and reverify it.
3321	 */
3322	xfs_buf_lock(bp);
3323	xfs_buf_hold(bp);
3324	error = _xfs_buf_read(bp, XBF_READ);
 
 
 
 
3325	if (error) {
3326		if (!xlog_is_shutdown(log)) {
3327			xfs_buf_ioerror_alert(bp, __this_address);
3328			ASSERT(0);
3329		}
3330		xfs_buf_relse(bp);
3331		return error;
3332	}
3333
3334	/* Convert superblock from on-disk format */
3335	xfs_sb_from_disk(sbp, bp->b_addr);
 
3336	xfs_buf_relse(bp);
3337
3338	/* re-initialise in-core superblock and geometry structures */
3339	mp->m_features |= xfs_sb_version_to_features(sbp);
3340	xfs_reinit_percpu_counters(mp);
3341	error = xfs_initialize_perag(mp, sbp->sb_agcount, sbp->sb_dblocks,
3342			&mp->m_maxagi);
3343	if (error) {
3344		xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
3345		return error;
3346	}
3347	mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
3348
 
 
3349	/* Normal transactions can now occur */
3350	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
3351	return 0;
3352}
3353
3354/*
3355 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3356 *
3357 * Return error or zero.
3358 */
3359int
3360xlog_recover(
3361	struct xlog	*log)
3362{
3363	xfs_daddr_t	head_blk, tail_blk;
3364	int		error;
3365
3366	/* find the tail of the log */
3367	error = xlog_find_tail(log, &head_blk, &tail_blk);
3368	if (error)
3369		return error;
3370
3371	/*
3372	 * The superblock was read before the log was available and thus the LSN
3373	 * could not be verified. Check the superblock LSN against the current
3374	 * LSN now that it's known.
3375	 */
3376	if (xfs_has_crc(log->l_mp) &&
3377	    !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
3378		return -EINVAL;
3379
3380	if (tail_blk != head_blk) {
3381		/* There used to be a comment here:
3382		 *
3383		 * disallow recovery on read-only mounts.  note -- mount
3384		 * checks for ENOSPC and turns it into an intelligent
3385		 * error message.
3386		 * ...but this is no longer true.  Now, unless you specify
3387		 * NORECOVERY (in which case this function would never be
3388		 * called), we just go ahead and recover.  We do this all
3389		 * under the vfs layer, so we can get away with it unless
3390		 * the device itself is read-only, in which case we fail.
3391		 */
3392		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3393			return error;
3394		}
3395
3396		/*
3397		 * Version 5 superblock log feature mask validation. We know the
3398		 * log is dirty so check if there are any unknown log features
3399		 * in what we need to recover. If there are unknown features
3400		 * (e.g. unsupported transactions, then simply reject the
3401		 * attempt at recovery before touching anything.
3402		 */
3403		if (xfs_sb_is_v5(&log->l_mp->m_sb) &&
3404		    xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
3405					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
3406			xfs_warn(log->l_mp,
3407"Superblock has unknown incompatible log features (0x%x) enabled.",
3408				(log->l_mp->m_sb.sb_features_log_incompat &
3409					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
3410			xfs_warn(log->l_mp,
3411"The log can not be fully and/or safely recovered by this kernel.");
3412			xfs_warn(log->l_mp,
3413"Please recover the log on a kernel that supports the unknown features.");
3414			return -EINVAL;
3415		}
3416
3417		/*
3418		 * Delay log recovery if the debug hook is set. This is debug
3419		 * instrumentation to coordinate simulation of I/O failures with
3420		 * log recovery.
3421		 */
3422		if (xfs_globals.log_recovery_delay) {
3423			xfs_notice(log->l_mp,
3424				"Delaying log recovery for %d seconds.",
3425				xfs_globals.log_recovery_delay);
3426			msleep(xfs_globals.log_recovery_delay * 1000);
3427		}
3428
3429		xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
3430				log->l_mp->m_logname ? log->l_mp->m_logname
3431						     : "internal");
3432
3433		error = xlog_do_recover(log, head_blk, tail_blk);
3434		set_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
3435	}
3436	return error;
3437}
3438
3439/*
3440 * In the first part of recovery we replay inodes and buffers and build up the
3441 * list of intents which need to be processed. Here we process the intents and
3442 * clean up the on disk unlinked inode lists. This is separated from the first
3443 * part of recovery so that the root and real-time bitmap inodes can be read in
3444 * from disk in between the two stages.  This is necessary so that we can free
3445 * space in the real-time portion of the file system.
 
3446 */
3447int
3448xlog_recover_finish(
3449	struct xlog	*log)
3450{
3451	int	error;
3452
3453	error = xlog_recover_process_intents(log);
3454	if (error) {
3455		/*
3456		 * Cancel all the unprocessed intent items now so that we don't
3457		 * leave them pinned in the AIL.  This can cause the AIL to
3458		 * livelock on the pinned item if anyone tries to push the AIL
3459		 * (inode reclaim does this) before we get around to
3460		 * xfs_log_mount_cancel.
3461		 */
3462		xlog_recover_cancel_intents(log);
3463		xfs_alert(log->l_mp, "Failed to recover intents");
3464		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
3465		return error;
3466	}
3467
3468	/*
3469	 * Sync the log to get all the intents out of the AIL.  This isn't
3470	 * absolutely necessary, but it helps in case the unlink transactions
3471	 * would have problems pushing the intents out of the way.
3472	 */
3473	xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3474
3475	/*
3476	 * Now that we've recovered the log and all the intents, we can clear
3477	 * the log incompat feature bits in the superblock because there's no
3478	 * longer anything to protect.  We rely on the AIL push to write out the
3479	 * updated superblock after everything else.
3480	 */
3481	if (xfs_clear_incompat_log_features(log->l_mp)) {
3482		error = xfs_sync_sb(log->l_mp, false);
3483		if (error < 0) {
3484			xfs_alert(log->l_mp,
3485	"Failed to clear log incompat features on recovery");
 
 
3486			return error;
3487		}
3488	}
3489
3490	xlog_recover_process_iunlinks(log);
3491
3492	/*
3493	 * Recover any CoW staging blocks that are still referenced by the
3494	 * ondisk refcount metadata.  During mount there cannot be any live
3495	 * staging extents as we have not permitted any user modifications.
3496	 * Therefore, it is safe to free them all right now, even on a
3497	 * read-only mount.
3498	 */
3499	error = xfs_reflink_recover_cow(log->l_mp);
3500	if (error) {
3501		xfs_alert(log->l_mp,
3502	"Failed to recover leftover CoW staging extents, err %d.",
3503				error);
3504		/*
3505		 * If we get an error here, make sure the log is shut down
3506		 * but return zero so that any log items committed since the
3507		 * end of intents processing can be pushed through the CIL
3508		 * and AIL.
3509		 */
3510		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
3511	}
 
3512
 
 
 
 
 
 
 
 
 
3513	return 0;
3514}
3515
3516void
3517xlog_recover_cancel(
3518	struct xlog	*log)
3519{
3520	if (xlog_recovery_needed(log))
3521		xlog_recover_cancel_intents(log);
3522}
3523
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include "xfs.h"
   7#include "xfs_fs.h"
   8#include "xfs_shared.h"
   9#include "xfs_format.h"
  10#include "xfs_log_format.h"
  11#include "xfs_trans_resv.h"
  12#include "xfs_bit.h"
  13#include "xfs_sb.h"
  14#include "xfs_mount.h"
  15#include "xfs_defer.h"
  16#include "xfs_inode.h"
  17#include "xfs_trans.h"
  18#include "xfs_log.h"
  19#include "xfs_log_priv.h"
  20#include "xfs_log_recover.h"
  21#include "xfs_inode_item.h"
  22#include "xfs_extfree_item.h"
  23#include "xfs_trans_priv.h"
  24#include "xfs_alloc.h"
  25#include "xfs_ialloc.h"
  26#include "xfs_quota.h"
  27#include "xfs_trace.h"
  28#include "xfs_icache.h"
  29#include "xfs_bmap_btree.h"
  30#include "xfs_error.h"
  31#include "xfs_dir2.h"
  32#include "xfs_rmap_item.h"
  33#include "xfs_buf_item.h"
  34#include "xfs_refcount_item.h"
  35#include "xfs_bmap_item.h"
 
  36
  37#define BLK_AVG(blk1, blk2)	((blk1+blk2) >> 1)
  38
  39STATIC int
  40xlog_find_zeroed(
  41	struct xlog	*,
  42	xfs_daddr_t	*);
  43STATIC int
  44xlog_clear_stale_blocks(
  45	struct xlog	*,
  46	xfs_lsn_t);
  47#if defined(DEBUG)
  48STATIC void
  49xlog_recover_check_summary(
  50	struct xlog *);
  51#else
  52#define	xlog_recover_check_summary(log)
  53#endif
  54STATIC int
  55xlog_do_recovery_pass(
  56        struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
  57
  58/*
  59 * This structure is used during recovery to record the buf log items which
  60 * have been canceled and should not be replayed.
  61 */
  62struct xfs_buf_cancel {
  63	xfs_daddr_t		bc_blkno;
  64	uint			bc_len;
  65	int			bc_refcount;
  66	struct list_head	bc_list;
  67};
  68
  69/*
  70 * Sector aligned buffer routines for buffer create/read/write/access
  71 */
  72
  73/*
  74 * Verify the log-relative block number and length in basic blocks are valid for
  75 * an operation involving the given XFS log buffer. Returns true if the fields
  76 * are valid, false otherwise.
  77 */
  78static inline bool
  79xlog_verify_bno(
  80	struct xlog	*log,
  81	xfs_daddr_t	blk_no,
  82	int		bbcount)
  83{
  84	if (blk_no < 0 || blk_no >= log->l_logBBsize)
  85		return false;
  86	if (bbcount <= 0 || (blk_no + bbcount) > log->l_logBBsize)
  87		return false;
  88	return true;
  89}
  90
  91/*
  92 * Allocate a buffer to hold log data.  The buffer needs to be able to map to
  93 * a range of nbblks basic blocks at any valid offset within the log.
  94 */
  95static char *
  96xlog_alloc_buffer(
  97	struct xlog	*log,
  98	int		nbblks)
  99{
 100	int align_mask = xfs_buftarg_dma_alignment(log->l_targ);
 101
 102	/*
 103	 * Pass log block 0 since we don't have an addr yet, buffer will be
 104	 * verified on read.
 105	 */
 106	if (!xlog_verify_bno(log, 0, nbblks)) {
 107		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 108			nbblks);
 109		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 110		return NULL;
 111	}
 112
 113	/*
 114	 * We do log I/O in units of log sectors (a power-of-2 multiple of the
 115	 * basic block size), so we round up the requested size to accommodate
 116	 * the basic blocks required for complete log sectors.
 117	 *
 118	 * In addition, the buffer may be used for a non-sector-aligned block
 119	 * offset, in which case an I/O of the requested size could extend
 120	 * beyond the end of the buffer.  If the requested size is only 1 basic
 121	 * block it will never straddle a sector boundary, so this won't be an
 122	 * issue.  Nor will this be a problem if the log I/O is done in basic
 123	 * blocks (sector size 1).  But otherwise we extend the buffer by one
 124	 * extra log sector to ensure there's space to accommodate this
 125	 * possibility.
 126	 */
 127	if (nbblks > 1 && log->l_sectBBsize > 1)
 128		nbblks += log->l_sectBBsize;
 129	nbblks = round_up(nbblks, log->l_sectBBsize);
 130	return kmem_alloc_io(BBTOB(nbblks), align_mask, KM_MAYFAIL | KM_ZERO);
 131}
 132
 133/*
 134 * Return the address of the start of the given block number's data
 135 * in a log buffer.  The buffer covers a log sector-aligned region.
 136 */
 137static inline unsigned int
 138xlog_align(
 139	struct xlog	*log,
 140	xfs_daddr_t	blk_no)
 141{
 142	return BBTOB(blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1));
 143}
 144
 145static int
 146xlog_do_io(
 147	struct xlog		*log,
 148	xfs_daddr_t		blk_no,
 149	unsigned int		nbblks,
 150	char			*data,
 151	unsigned int		op)
 152{
 153	int			error;
 154
 155	if (!xlog_verify_bno(log, blk_no, nbblks)) {
 156		xfs_warn(log->l_mp,
 157			 "Invalid log block/length (0x%llx, 0x%x) for buffer",
 158			 blk_no, nbblks);
 159		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 160		return -EFSCORRUPTED;
 161	}
 162
 163	blk_no = round_down(blk_no, log->l_sectBBsize);
 164	nbblks = round_up(nbblks, log->l_sectBBsize);
 165	ASSERT(nbblks > 0);
 166
 167	error = xfs_rw_bdev(log->l_targ->bt_bdev, log->l_logBBstart + blk_no,
 168			BBTOB(nbblks), data, op);
 169	if (error && !XFS_FORCED_SHUTDOWN(log->l_mp)) {
 170		xfs_alert(log->l_mp,
 171			  "log recovery %s I/O error at daddr 0x%llx len %d error %d",
 172			  op == REQ_OP_WRITE ? "write" : "read",
 173			  blk_no, nbblks, error);
 174	}
 175	return error;
 176}
 177
 178STATIC int
 179xlog_bread_noalign(
 180	struct xlog	*log,
 181	xfs_daddr_t	blk_no,
 182	int		nbblks,
 183	char		*data)
 184{
 185	return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
 186}
 187
 188STATIC int
 189xlog_bread(
 190	struct xlog	*log,
 191	xfs_daddr_t	blk_no,
 192	int		nbblks,
 193	char		*data,
 194	char		**offset)
 195{
 196	int		error;
 197
 198	error = xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
 199	if (!error)
 200		*offset = data + xlog_align(log, blk_no);
 201	return error;
 202}
 203
 204STATIC int
 205xlog_bwrite(
 206	struct xlog	*log,
 207	xfs_daddr_t	blk_no,
 208	int		nbblks,
 209	char		*data)
 210{
 211	return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_WRITE);
 212}
 213
 214#ifdef DEBUG
 215/*
 216 * dump debug superblock and log record information
 217 */
 218STATIC void
 219xlog_header_check_dump(
 220	xfs_mount_t		*mp,
 221	xlog_rec_header_t	*head)
 222{
 223	xfs_debug(mp, "%s:  SB : uuid = %pU, fmt = %d",
 224		__func__, &mp->m_sb.sb_uuid, XLOG_FMT);
 225	xfs_debug(mp, "    log : uuid = %pU, fmt = %d",
 226		&head->h_fs_uuid, be32_to_cpu(head->h_fmt));
 227}
 228#else
 229#define xlog_header_check_dump(mp, head)
 230#endif
 231
 232/*
 233 * check log record header for recovery
 234 */
 235STATIC int
 236xlog_header_check_recover(
 237	xfs_mount_t		*mp,
 238	xlog_rec_header_t	*head)
 239{
 240	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 241
 242	/*
 243	 * IRIX doesn't write the h_fmt field and leaves it zeroed
 244	 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
 245	 * a dirty log created in IRIX.
 246	 */
 247	if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
 248		xfs_warn(mp,
 249	"dirty log written in incompatible format - can't recover");
 250		xlog_header_check_dump(mp, head);
 251		XFS_ERROR_REPORT("xlog_header_check_recover(1)",
 252				 XFS_ERRLEVEL_HIGH, mp);
 253		return -EFSCORRUPTED;
 254	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 
 
 255		xfs_warn(mp,
 256	"dirty log entry has mismatched uuid - can't recover");
 257		xlog_header_check_dump(mp, head);
 258		XFS_ERROR_REPORT("xlog_header_check_recover(2)",
 259				 XFS_ERRLEVEL_HIGH, mp);
 260		return -EFSCORRUPTED;
 261	}
 262	return 0;
 263}
 264
 265/*
 266 * read the head block of the log and check the header
 267 */
 268STATIC int
 269xlog_header_check_mount(
 270	xfs_mount_t		*mp,
 271	xlog_rec_header_t	*head)
 272{
 273	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 274
 275	if (uuid_is_null(&head->h_fs_uuid)) {
 276		/*
 277		 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
 278		 * h_fs_uuid is null, we assume this log was last mounted
 279		 * by IRIX and continue.
 280		 */
 281		xfs_warn(mp, "null uuid in log - IRIX style log");
 282	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 
 283		xfs_warn(mp, "log has mismatched uuid - can't recover");
 284		xlog_header_check_dump(mp, head);
 285		XFS_ERROR_REPORT("xlog_header_check_mount",
 286				 XFS_ERRLEVEL_HIGH, mp);
 287		return -EFSCORRUPTED;
 288	}
 289	return 0;
 290}
 291
 292STATIC void
 293xlog_recover_iodone(
 294	struct xfs_buf	*bp)
 295{
 296	if (bp->b_error) {
 297		/*
 298		 * We're not going to bother about retrying
 299		 * this during recovery. One strike!
 300		 */
 301		if (!XFS_FORCED_SHUTDOWN(bp->b_mount)) {
 302			xfs_buf_ioerror_alert(bp, __func__);
 303			xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
 304		}
 305	}
 306
 307	/*
 308	 * On v5 supers, a bli could be attached to update the metadata LSN.
 309	 * Clean it up.
 310	 */
 311	if (bp->b_log_item)
 312		xfs_buf_item_relse(bp);
 313	ASSERT(bp->b_log_item == NULL);
 314
 315	bp->b_iodone = NULL;
 316	xfs_buf_ioend(bp);
 317}
 318
 319/*
 320 * This routine finds (to an approximation) the first block in the physical
 321 * log which contains the given cycle.  It uses a binary search algorithm.
 322 * Note that the algorithm can not be perfect because the disk will not
 323 * necessarily be perfect.
 324 */
 325STATIC int
 326xlog_find_cycle_start(
 327	struct xlog	*log,
 328	char		*buffer,
 329	xfs_daddr_t	first_blk,
 330	xfs_daddr_t	*last_blk,
 331	uint		cycle)
 332{
 333	char		*offset;
 334	xfs_daddr_t	mid_blk;
 335	xfs_daddr_t	end_blk;
 336	uint		mid_cycle;
 337	int		error;
 338
 339	end_blk = *last_blk;
 340	mid_blk = BLK_AVG(first_blk, end_blk);
 341	while (mid_blk != first_blk && mid_blk != end_blk) {
 342		error = xlog_bread(log, mid_blk, 1, buffer, &offset);
 343		if (error)
 344			return error;
 345		mid_cycle = xlog_get_cycle(offset);
 346		if (mid_cycle == cycle)
 347			end_blk = mid_blk;   /* last_half_cycle == mid_cycle */
 348		else
 349			first_blk = mid_blk; /* first_half_cycle == mid_cycle */
 350		mid_blk = BLK_AVG(first_blk, end_blk);
 351	}
 352	ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
 353	       (mid_blk == end_blk && mid_blk-1 == first_blk));
 354
 355	*last_blk = end_blk;
 356
 357	return 0;
 358}
 359
 360/*
 361 * Check that a range of blocks does not contain stop_on_cycle_no.
 362 * Fill in *new_blk with the block offset where such a block is
 363 * found, or with -1 (an invalid block number) if there is no such
 364 * block in the range.  The scan needs to occur from front to back
 365 * and the pointer into the region must be updated since a later
 366 * routine will need to perform another test.
 367 */
 368STATIC int
 369xlog_find_verify_cycle(
 370	struct xlog	*log,
 371	xfs_daddr_t	start_blk,
 372	int		nbblks,
 373	uint		stop_on_cycle_no,
 374	xfs_daddr_t	*new_blk)
 375{
 376	xfs_daddr_t	i, j;
 377	uint		cycle;
 378	char		*buffer;
 379	xfs_daddr_t	bufblks;
 380	char		*buf = NULL;
 381	int		error = 0;
 382
 383	/*
 384	 * Greedily allocate a buffer big enough to handle the full
 385	 * range of basic blocks we'll be examining.  If that fails,
 386	 * try a smaller size.  We need to be able to read at least
 387	 * a log sector, or we're out of luck.
 388	 */
 389	bufblks = 1 << ffs(nbblks);
 390	while (bufblks > log->l_logBBsize)
 391		bufblks >>= 1;
 392	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
 393		bufblks >>= 1;
 394		if (bufblks < log->l_sectBBsize)
 395			return -ENOMEM;
 396	}
 397
 398	for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
 399		int	bcount;
 400
 401		bcount = min(bufblks, (start_blk + nbblks - i));
 402
 403		error = xlog_bread(log, i, bcount, buffer, &buf);
 404		if (error)
 405			goto out;
 406
 407		for (j = 0; j < bcount; j++) {
 408			cycle = xlog_get_cycle(buf);
 409			if (cycle == stop_on_cycle_no) {
 410				*new_blk = i+j;
 411				goto out;
 412			}
 413
 414			buf += BBSIZE;
 415		}
 416	}
 417
 418	*new_blk = -1;
 419
 420out:
 421	kmem_free(buffer);
 422	return error;
 423}
 424
 
 
 
 
 
 
 
 
 
 
 
 
 
 425/*
 426 * Potentially backup over partial log record write.
 427 *
 428 * In the typical case, last_blk is the number of the block directly after
 429 * a good log record.  Therefore, we subtract one to get the block number
 430 * of the last block in the given buffer.  extra_bblks contains the number
 431 * of blocks we would have read on a previous read.  This happens when the
 432 * last log record is split over the end of the physical log.
 433 *
 434 * extra_bblks is the number of blocks potentially verified on a previous
 435 * call to this routine.
 436 */
 437STATIC int
 438xlog_find_verify_log_record(
 439	struct xlog		*log,
 440	xfs_daddr_t		start_blk,
 441	xfs_daddr_t		*last_blk,
 442	int			extra_bblks)
 443{
 444	xfs_daddr_t		i;
 445	char			*buffer;
 446	char			*offset = NULL;
 447	xlog_rec_header_t	*head = NULL;
 448	int			error = 0;
 449	int			smallmem = 0;
 450	int			num_blks = *last_blk - start_blk;
 451	int			xhdrs;
 452
 453	ASSERT(start_blk != 0 || *last_blk != start_blk);
 454
 455	buffer = xlog_alloc_buffer(log, num_blks);
 456	if (!buffer) {
 457		buffer = xlog_alloc_buffer(log, 1);
 458		if (!buffer)
 459			return -ENOMEM;
 460		smallmem = 1;
 461	} else {
 462		error = xlog_bread(log, start_blk, num_blks, buffer, &offset);
 463		if (error)
 464			goto out;
 465		offset += ((num_blks - 1) << BBSHIFT);
 466	}
 467
 468	for (i = (*last_blk) - 1; i >= 0; i--) {
 469		if (i < start_blk) {
 470			/* valid log record not found */
 471			xfs_warn(log->l_mp,
 472		"Log inconsistent (didn't find previous header)");
 473			ASSERT(0);
 474			error = -EIO;
 475			goto out;
 476		}
 477
 478		if (smallmem) {
 479			error = xlog_bread(log, i, 1, buffer, &offset);
 480			if (error)
 481				goto out;
 482		}
 483
 484		head = (xlog_rec_header_t *)offset;
 485
 486		if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
 487			break;
 488
 489		if (!smallmem)
 490			offset -= BBSIZE;
 491	}
 492
 493	/*
 494	 * We hit the beginning of the physical log & still no header.  Return
 495	 * to caller.  If caller can handle a return of -1, then this routine
 496	 * will be called again for the end of the physical log.
 497	 */
 498	if (i == -1) {
 499		error = 1;
 500		goto out;
 501	}
 502
 503	/*
 504	 * We have the final block of the good log (the first block
 505	 * of the log record _before_ the head. So we check the uuid.
 506	 */
 507	if ((error = xlog_header_check_mount(log->l_mp, head)))
 508		goto out;
 509
 510	/*
 511	 * We may have found a log record header before we expected one.
 512	 * last_blk will be the 1st block # with a given cycle #.  We may end
 513	 * up reading an entire log record.  In this case, we don't want to
 514	 * reset last_blk.  Only when last_blk points in the middle of a log
 515	 * record do we update last_blk.
 516	 */
 517	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
 518		uint	h_size = be32_to_cpu(head->h_size);
 519
 520		xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
 521		if (h_size % XLOG_HEADER_CYCLE_SIZE)
 522			xhdrs++;
 523	} else {
 524		xhdrs = 1;
 525	}
 526
 527	if (*last_blk - i + extra_bblks !=
 528	    BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
 529		*last_blk = i;
 530
 531out:
 532	kmem_free(buffer);
 533	return error;
 534}
 535
 536/*
 537 * Head is defined to be the point of the log where the next log write
 538 * could go.  This means that incomplete LR writes at the end are
 539 * eliminated when calculating the head.  We aren't guaranteed that previous
 540 * LR have complete transactions.  We only know that a cycle number of
 541 * current cycle number -1 won't be present in the log if we start writing
 542 * from our current block number.
 543 *
 544 * last_blk contains the block number of the first block with a given
 545 * cycle number.
 546 *
 547 * Return: zero if normal, non-zero if error.
 548 */
 549STATIC int
 550xlog_find_head(
 551	struct xlog	*log,
 552	xfs_daddr_t	*return_head_blk)
 553{
 554	char		*buffer;
 555	char		*offset;
 556	xfs_daddr_t	new_blk, first_blk, start_blk, last_blk, head_blk;
 557	int		num_scan_bblks;
 558	uint		first_half_cycle, last_half_cycle;
 559	uint		stop_on_cycle;
 560	int		error, log_bbnum = log->l_logBBsize;
 561
 562	/* Is the end of the log device zeroed? */
 563	error = xlog_find_zeroed(log, &first_blk);
 564	if (error < 0) {
 565		xfs_warn(log->l_mp, "empty log check failed");
 566		return error;
 567	}
 568	if (error == 1) {
 569		*return_head_blk = first_blk;
 570
 571		/* Is the whole lot zeroed? */
 572		if (!first_blk) {
 573			/* Linux XFS shouldn't generate totally zeroed logs -
 574			 * mkfs etc write a dummy unmount record to a fresh
 575			 * log so we can store the uuid in there
 576			 */
 577			xfs_warn(log->l_mp, "totally zeroed log");
 578		}
 579
 580		return 0;
 581	}
 582
 583	first_blk = 0;			/* get cycle # of 1st block */
 584	buffer = xlog_alloc_buffer(log, 1);
 585	if (!buffer)
 586		return -ENOMEM;
 587
 588	error = xlog_bread(log, 0, 1, buffer, &offset);
 589	if (error)
 590		goto out_free_buffer;
 591
 592	first_half_cycle = xlog_get_cycle(offset);
 593
 594	last_blk = head_blk = log_bbnum - 1;	/* get cycle # of last block */
 595	error = xlog_bread(log, last_blk, 1, buffer, &offset);
 596	if (error)
 597		goto out_free_buffer;
 598
 599	last_half_cycle = xlog_get_cycle(offset);
 600	ASSERT(last_half_cycle != 0);
 601
 602	/*
 603	 * If the 1st half cycle number is equal to the last half cycle number,
 604	 * then the entire log is stamped with the same cycle number.  In this
 605	 * case, head_blk can't be set to zero (which makes sense).  The below
 606	 * math doesn't work out properly with head_blk equal to zero.  Instead,
 607	 * we set it to log_bbnum which is an invalid block number, but this
 608	 * value makes the math correct.  If head_blk doesn't changed through
 609	 * all the tests below, *head_blk is set to zero at the very end rather
 610	 * than log_bbnum.  In a sense, log_bbnum and zero are the same block
 611	 * in a circular file.
 612	 */
 613	if (first_half_cycle == last_half_cycle) {
 614		/*
 615		 * In this case we believe that the entire log should have
 616		 * cycle number last_half_cycle.  We need to scan backwards
 617		 * from the end verifying that there are no holes still
 618		 * containing last_half_cycle - 1.  If we find such a hole,
 619		 * then the start of that hole will be the new head.  The
 620		 * simple case looks like
 621		 *        x | x ... | x - 1 | x
 622		 * Another case that fits this picture would be
 623		 *        x | x + 1 | x ... | x
 624		 * In this case the head really is somewhere at the end of the
 625		 * log, as one of the latest writes at the beginning was
 626		 * incomplete.
 627		 * One more case is
 628		 *        x | x + 1 | x ... | x - 1 | x
 629		 * This is really the combination of the above two cases, and
 630		 * the head has to end up at the start of the x-1 hole at the
 631		 * end of the log.
 632		 *
 633		 * In the 256k log case, we will read from the beginning to the
 634		 * end of the log and search for cycle numbers equal to x-1.
 635		 * We don't worry about the x+1 blocks that we encounter,
 636		 * because we know that they cannot be the head since the log
 637		 * started with x.
 638		 */
 639		head_blk = log_bbnum;
 640		stop_on_cycle = last_half_cycle - 1;
 641	} else {
 642		/*
 643		 * In this case we want to find the first block with cycle
 644		 * number matching last_half_cycle.  We expect the log to be
 645		 * some variation on
 646		 *        x + 1 ... | x ... | x
 647		 * The first block with cycle number x (last_half_cycle) will
 648		 * be where the new head belongs.  First we do a binary search
 649		 * for the first occurrence of last_half_cycle.  The binary
 650		 * search may not be totally accurate, so then we scan back
 651		 * from there looking for occurrences of last_half_cycle before
 652		 * us.  If that backwards scan wraps around the beginning of
 653		 * the log, then we look for occurrences of last_half_cycle - 1
 654		 * at the end of the log.  The cases we're looking for look
 655		 * like
 656		 *                               v binary search stopped here
 657		 *        x + 1 ... | x | x + 1 | x ... | x
 658		 *                   ^ but we want to locate this spot
 659		 * or
 660		 *        <---------> less than scan distance
 661		 *        x + 1 ... | x ... | x - 1 | x
 662		 *                           ^ we want to locate this spot
 663		 */
 664		stop_on_cycle = last_half_cycle;
 665		error = xlog_find_cycle_start(log, buffer, first_blk, &head_blk,
 666				last_half_cycle);
 667		if (error)
 668			goto out_free_buffer;
 669	}
 670
 671	/*
 672	 * Now validate the answer.  Scan back some number of maximum possible
 673	 * blocks and make sure each one has the expected cycle number.  The
 674	 * maximum is determined by the total possible amount of buffering
 675	 * in the in-core log.  The following number can be made tighter if
 676	 * we actually look at the block size of the filesystem.
 677	 */
 678	num_scan_bblks = min_t(int, log_bbnum, XLOG_TOTAL_REC_SHIFT(log));
 679	if (head_blk >= num_scan_bblks) {
 680		/*
 681		 * We are guaranteed that the entire check can be performed
 682		 * in one buffer.
 683		 */
 684		start_blk = head_blk - num_scan_bblks;
 685		if ((error = xlog_find_verify_cycle(log,
 686						start_blk, num_scan_bblks,
 687						stop_on_cycle, &new_blk)))
 688			goto out_free_buffer;
 689		if (new_blk != -1)
 690			head_blk = new_blk;
 691	} else {		/* need to read 2 parts of log */
 692		/*
 693		 * We are going to scan backwards in the log in two parts.
 694		 * First we scan the physical end of the log.  In this part
 695		 * of the log, we are looking for blocks with cycle number
 696		 * last_half_cycle - 1.
 697		 * If we find one, then we know that the log starts there, as
 698		 * we've found a hole that didn't get written in going around
 699		 * the end of the physical log.  The simple case for this is
 700		 *        x + 1 ... | x ... | x - 1 | x
 701		 *        <---------> less than scan distance
 702		 * If all of the blocks at the end of the log have cycle number
 703		 * last_half_cycle, then we check the blocks at the start of
 704		 * the log looking for occurrences of last_half_cycle.  If we
 705		 * find one, then our current estimate for the location of the
 706		 * first occurrence of last_half_cycle is wrong and we move
 707		 * back to the hole we've found.  This case looks like
 708		 *        x + 1 ... | x | x + 1 | x ...
 709		 *                               ^ binary search stopped here
 710		 * Another case we need to handle that only occurs in 256k
 711		 * logs is
 712		 *        x + 1 ... | x ... | x+1 | x ...
 713		 *                   ^ binary search stops here
 714		 * In a 256k log, the scan at the end of the log will see the
 715		 * x + 1 blocks.  We need to skip past those since that is
 716		 * certainly not the head of the log.  By searching for
 717		 * last_half_cycle-1 we accomplish that.
 718		 */
 719		ASSERT(head_blk <= INT_MAX &&
 720			(xfs_daddr_t) num_scan_bblks >= head_blk);
 721		start_blk = log_bbnum - (num_scan_bblks - head_blk);
 722		if ((error = xlog_find_verify_cycle(log, start_blk,
 723					num_scan_bblks - (int)head_blk,
 724					(stop_on_cycle - 1), &new_blk)))
 725			goto out_free_buffer;
 726		if (new_blk != -1) {
 727			head_blk = new_blk;
 728			goto validate_head;
 729		}
 730
 731		/*
 732		 * Scan beginning of log now.  The last part of the physical
 733		 * log is good.  This scan needs to verify that it doesn't find
 734		 * the last_half_cycle.
 735		 */
 736		start_blk = 0;
 737		ASSERT(head_blk <= INT_MAX);
 738		if ((error = xlog_find_verify_cycle(log,
 739					start_blk, (int)head_blk,
 740					stop_on_cycle, &new_blk)))
 741			goto out_free_buffer;
 742		if (new_blk != -1)
 743			head_blk = new_blk;
 744	}
 745
 746validate_head:
 747	/*
 748	 * Now we need to make sure head_blk is not pointing to a block in
 749	 * the middle of a log record.
 750	 */
 751	num_scan_bblks = XLOG_REC_SHIFT(log);
 752	if (head_blk >= num_scan_bblks) {
 753		start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
 754
 755		/* start ptr at last block ptr before head_blk */
 756		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
 757		if (error == 1)
 758			error = -EIO;
 759		if (error)
 760			goto out_free_buffer;
 761	} else {
 762		start_blk = 0;
 763		ASSERT(head_blk <= INT_MAX);
 764		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
 765		if (error < 0)
 766			goto out_free_buffer;
 767		if (error == 1) {
 768			/* We hit the beginning of the log during our search */
 769			start_blk = log_bbnum - (num_scan_bblks - head_blk);
 770			new_blk = log_bbnum;
 771			ASSERT(start_blk <= INT_MAX &&
 772				(xfs_daddr_t) log_bbnum-start_blk >= 0);
 773			ASSERT(head_blk <= INT_MAX);
 774			error = xlog_find_verify_log_record(log, start_blk,
 775							&new_blk, (int)head_blk);
 776			if (error == 1)
 777				error = -EIO;
 778			if (error)
 779				goto out_free_buffer;
 780			if (new_blk != log_bbnum)
 781				head_blk = new_blk;
 782		} else if (error)
 783			goto out_free_buffer;
 784	}
 785
 786	kmem_free(buffer);
 787	if (head_blk == log_bbnum)
 788		*return_head_blk = 0;
 789	else
 790		*return_head_blk = head_blk;
 791	/*
 792	 * When returning here, we have a good block number.  Bad block
 793	 * means that during a previous crash, we didn't have a clean break
 794	 * from cycle number N to cycle number N-1.  In this case, we need
 795	 * to find the first block with cycle number N-1.
 796	 */
 797	return 0;
 798
 799out_free_buffer:
 800	kmem_free(buffer);
 801	if (error)
 802		xfs_warn(log->l_mp, "failed to find log head");
 803	return error;
 804}
 805
 806/*
 807 * Seek backwards in the log for log record headers.
 808 *
 809 * Given a starting log block, walk backwards until we find the provided number
 810 * of records or hit the provided tail block. The return value is the number of
 811 * records encountered or a negative error code. The log block and buffer
 812 * pointer of the last record seen are returned in rblk and rhead respectively.
 813 */
 814STATIC int
 815xlog_rseek_logrec_hdr(
 816	struct xlog		*log,
 817	xfs_daddr_t		head_blk,
 818	xfs_daddr_t		tail_blk,
 819	int			count,
 820	char			*buffer,
 821	xfs_daddr_t		*rblk,
 822	struct xlog_rec_header	**rhead,
 823	bool			*wrapped)
 824{
 825	int			i;
 826	int			error;
 827	int			found = 0;
 828	char			*offset = NULL;
 829	xfs_daddr_t		end_blk;
 830
 831	*wrapped = false;
 832
 833	/*
 834	 * Walk backwards from the head block until we hit the tail or the first
 835	 * block in the log.
 836	 */
 837	end_blk = head_blk > tail_blk ? tail_blk : 0;
 838	for (i = (int) head_blk - 1; i >= end_blk; i--) {
 839		error = xlog_bread(log, i, 1, buffer, &offset);
 840		if (error)
 841			goto out_error;
 842
 843		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 844			*rblk = i;
 845			*rhead = (struct xlog_rec_header *) offset;
 846			if (++found == count)
 847				break;
 848		}
 849	}
 850
 851	/*
 852	 * If we haven't hit the tail block or the log record header count,
 853	 * start looking again from the end of the physical log. Note that
 854	 * callers can pass head == tail if the tail is not yet known.
 855	 */
 856	if (tail_blk >= head_blk && found != count) {
 857		for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
 858			error = xlog_bread(log, i, 1, buffer, &offset);
 859			if (error)
 860				goto out_error;
 861
 862			if (*(__be32 *)offset ==
 863			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 864				*wrapped = true;
 865				*rblk = i;
 866				*rhead = (struct xlog_rec_header *) offset;
 867				if (++found == count)
 868					break;
 869			}
 870		}
 871	}
 872
 873	return found;
 874
 875out_error:
 876	return error;
 877}
 878
 879/*
 880 * Seek forward in the log for log record headers.
 881 *
 882 * Given head and tail blocks, walk forward from the tail block until we find
 883 * the provided number of records or hit the head block. The return value is the
 884 * number of records encountered or a negative error code. The log block and
 885 * buffer pointer of the last record seen are returned in rblk and rhead
 886 * respectively.
 887 */
 888STATIC int
 889xlog_seek_logrec_hdr(
 890	struct xlog		*log,
 891	xfs_daddr_t		head_blk,
 892	xfs_daddr_t		tail_blk,
 893	int			count,
 894	char			*buffer,
 895	xfs_daddr_t		*rblk,
 896	struct xlog_rec_header	**rhead,
 897	bool			*wrapped)
 898{
 899	int			i;
 900	int			error;
 901	int			found = 0;
 902	char			*offset = NULL;
 903	xfs_daddr_t		end_blk;
 904
 905	*wrapped = false;
 906
 907	/*
 908	 * Walk forward from the tail block until we hit the head or the last
 909	 * block in the log.
 910	 */
 911	end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
 912	for (i = (int) tail_blk; i <= end_blk; i++) {
 913		error = xlog_bread(log, i, 1, buffer, &offset);
 914		if (error)
 915			goto out_error;
 916
 917		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 918			*rblk = i;
 919			*rhead = (struct xlog_rec_header *) offset;
 920			if (++found == count)
 921				break;
 922		}
 923	}
 924
 925	/*
 926	 * If we haven't hit the head block or the log record header count,
 927	 * start looking again from the start of the physical log.
 928	 */
 929	if (tail_blk > head_blk && found != count) {
 930		for (i = 0; i < (int) head_blk; i++) {
 931			error = xlog_bread(log, i, 1, buffer, &offset);
 932			if (error)
 933				goto out_error;
 934
 935			if (*(__be32 *)offset ==
 936			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 937				*wrapped = true;
 938				*rblk = i;
 939				*rhead = (struct xlog_rec_header *) offset;
 940				if (++found == count)
 941					break;
 942			}
 943		}
 944	}
 945
 946	return found;
 947
 948out_error:
 949	return error;
 950}
 951
 952/*
 953 * Calculate distance from head to tail (i.e., unused space in the log).
 954 */
 955static inline int
 956xlog_tail_distance(
 957	struct xlog	*log,
 958	xfs_daddr_t	head_blk,
 959	xfs_daddr_t	tail_blk)
 960{
 961	if (head_blk < tail_blk)
 962		return tail_blk - head_blk;
 963
 964	return tail_blk + (log->l_logBBsize - head_blk);
 965}
 966
 967/*
 968 * Verify the log tail. This is particularly important when torn or incomplete
 969 * writes have been detected near the front of the log and the head has been
 970 * walked back accordingly.
 971 *
 972 * We also have to handle the case where the tail was pinned and the head
 973 * blocked behind the tail right before a crash. If the tail had been pushed
 974 * immediately prior to the crash and the subsequent checkpoint was only
 975 * partially written, it's possible it overwrote the last referenced tail in the
 976 * log with garbage. This is not a coherency problem because the tail must have
 977 * been pushed before it can be overwritten, but appears as log corruption to
 978 * recovery because we have no way to know the tail was updated if the
 979 * subsequent checkpoint didn't write successfully.
 980 *
 981 * Therefore, CRC check the log from tail to head. If a failure occurs and the
 982 * offending record is within max iclog bufs from the head, walk the tail
 983 * forward and retry until a valid tail is found or corruption is detected out
 984 * of the range of a possible overwrite.
 985 */
 986STATIC int
 987xlog_verify_tail(
 988	struct xlog		*log,
 989	xfs_daddr_t		head_blk,
 990	xfs_daddr_t		*tail_blk,
 991	int			hsize)
 992{
 993	struct xlog_rec_header	*thead;
 994	char			*buffer;
 995	xfs_daddr_t		first_bad;
 996	int			error = 0;
 997	bool			wrapped;
 998	xfs_daddr_t		tmp_tail;
 999	xfs_daddr_t		orig_tail = *tail_blk;
1000
1001	buffer = xlog_alloc_buffer(log, 1);
1002	if (!buffer)
1003		return -ENOMEM;
1004
1005	/*
1006	 * Make sure the tail points to a record (returns positive count on
1007	 * success).
1008	 */
1009	error = xlog_seek_logrec_hdr(log, head_blk, *tail_blk, 1, buffer,
1010			&tmp_tail, &thead, &wrapped);
1011	if (error < 0)
1012		goto out;
1013	if (*tail_blk != tmp_tail)
1014		*tail_blk = tmp_tail;
1015
1016	/*
1017	 * Run a CRC check from the tail to the head. We can't just check
1018	 * MAX_ICLOGS records past the tail because the tail may point to stale
1019	 * blocks cleared during the search for the head/tail. These blocks are
1020	 * overwritten with zero-length records and thus record count is not a
1021	 * reliable indicator of the iclog state before a crash.
1022	 */
1023	first_bad = 0;
1024	error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1025				      XLOG_RECOVER_CRCPASS, &first_bad);
1026	while ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1027		int	tail_distance;
1028
1029		/*
1030		 * Is corruption within range of the head? If so, retry from
1031		 * the next record. Otherwise return an error.
1032		 */
1033		tail_distance = xlog_tail_distance(log, head_blk, first_bad);
1034		if (tail_distance > BTOBB(XLOG_MAX_ICLOGS * hsize))
1035			break;
1036
1037		/* skip to the next record; returns positive count on success */
1038		error = xlog_seek_logrec_hdr(log, head_blk, first_bad, 2,
1039				buffer, &tmp_tail, &thead, &wrapped);
1040		if (error < 0)
1041			goto out;
1042
1043		*tail_blk = tmp_tail;
1044		first_bad = 0;
1045		error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1046					      XLOG_RECOVER_CRCPASS, &first_bad);
1047	}
1048
1049	if (!error && *tail_blk != orig_tail)
1050		xfs_warn(log->l_mp,
1051		"Tail block (0x%llx) overwrite detected. Updated to 0x%llx",
1052			 orig_tail, *tail_blk);
1053out:
1054	kmem_free(buffer);
1055	return error;
1056}
1057
1058/*
1059 * Detect and trim torn writes from the head of the log.
1060 *
1061 * Storage without sector atomicity guarantees can result in torn writes in the
1062 * log in the event of a crash. Our only means to detect this scenario is via
1063 * CRC verification. While we can't always be certain that CRC verification
1064 * failure is due to a torn write vs. an unrelated corruption, we do know that
1065 * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1066 * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1067 * the log and treat failures in this range as torn writes as a matter of
1068 * policy. In the event of CRC failure, the head is walked back to the last good
1069 * record in the log and the tail is updated from that record and verified.
1070 */
1071STATIC int
1072xlog_verify_head(
1073	struct xlog		*log,
1074	xfs_daddr_t		*head_blk,	/* in/out: unverified head */
1075	xfs_daddr_t		*tail_blk,	/* out: tail block */
1076	char			*buffer,
1077	xfs_daddr_t		*rhead_blk,	/* start blk of last record */
1078	struct xlog_rec_header	**rhead,	/* ptr to last record */
1079	bool			*wrapped)	/* last rec. wraps phys. log */
1080{
1081	struct xlog_rec_header	*tmp_rhead;
1082	char			*tmp_buffer;
1083	xfs_daddr_t		first_bad;
1084	xfs_daddr_t		tmp_rhead_blk;
1085	int			found;
1086	int			error;
1087	bool			tmp_wrapped;
1088
1089	/*
1090	 * Check the head of the log for torn writes. Search backwards from the
1091	 * head until we hit the tail or the maximum number of log record I/Os
1092	 * that could have been in flight at one time. Use a temporary buffer so
1093	 * we don't trash the rhead/buffer pointers from the caller.
1094	 */
1095	tmp_buffer = xlog_alloc_buffer(log, 1);
1096	if (!tmp_buffer)
1097		return -ENOMEM;
1098	error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1099				      XLOG_MAX_ICLOGS, tmp_buffer,
1100				      &tmp_rhead_blk, &tmp_rhead, &tmp_wrapped);
1101	kmem_free(tmp_buffer);
1102	if (error < 0)
1103		return error;
1104
1105	/*
1106	 * Now run a CRC verification pass over the records starting at the
1107	 * block found above to the current head. If a CRC failure occurs, the
1108	 * log block of the first bad record is saved in first_bad.
1109	 */
1110	error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1111				      XLOG_RECOVER_CRCPASS, &first_bad);
1112	if ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1113		/*
1114		 * We've hit a potential torn write. Reset the error and warn
1115		 * about it.
1116		 */
1117		error = 0;
1118		xfs_warn(log->l_mp,
1119"Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1120			 first_bad, *head_blk);
1121
1122		/*
1123		 * Get the header block and buffer pointer for the last good
1124		 * record before the bad record.
1125		 *
1126		 * Note that xlog_find_tail() clears the blocks at the new head
1127		 * (i.e., the records with invalid CRC) if the cycle number
1128		 * matches the the current cycle.
1129		 */
1130		found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1,
1131				buffer, rhead_blk, rhead, wrapped);
1132		if (found < 0)
1133			return found;
1134		if (found == 0)		/* XXX: right thing to do here? */
1135			return -EIO;
1136
1137		/*
1138		 * Reset the head block to the starting block of the first bad
1139		 * log record and set the tail block based on the last good
1140		 * record.
1141		 *
1142		 * Bail out if the updated head/tail match as this indicates
1143		 * possible corruption outside of the acceptable
1144		 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1145		 */
1146		*head_blk = first_bad;
1147		*tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1148		if (*head_blk == *tail_blk) {
1149			ASSERT(0);
1150			return 0;
1151		}
1152	}
1153	if (error)
1154		return error;
1155
1156	return xlog_verify_tail(log, *head_blk, tail_blk,
1157				be32_to_cpu((*rhead)->h_size));
1158}
1159
1160/*
1161 * We need to make sure we handle log wrapping properly, so we can't use the
1162 * calculated logbno directly. Make sure it wraps to the correct bno inside the
1163 * log.
1164 *
1165 * The log is limited to 32 bit sizes, so we use the appropriate modulus
1166 * operation here and cast it back to a 64 bit daddr on return.
1167 */
1168static inline xfs_daddr_t
1169xlog_wrap_logbno(
1170	struct xlog		*log,
1171	xfs_daddr_t		bno)
1172{
1173	int			mod;
1174
1175	div_s64_rem(bno, log->l_logBBsize, &mod);
1176	return mod;
1177}
1178
1179/*
1180 * Check whether the head of the log points to an unmount record. In other
1181 * words, determine whether the log is clean. If so, update the in-core state
1182 * appropriately.
1183 */
1184static int
1185xlog_check_unmount_rec(
1186	struct xlog		*log,
1187	xfs_daddr_t		*head_blk,
1188	xfs_daddr_t		*tail_blk,
1189	struct xlog_rec_header	*rhead,
1190	xfs_daddr_t		rhead_blk,
1191	char			*buffer,
1192	bool			*clean)
1193{
1194	struct xlog_op_header	*op_head;
1195	xfs_daddr_t		umount_data_blk;
1196	xfs_daddr_t		after_umount_blk;
1197	int			hblks;
1198	int			error;
1199	char			*offset;
1200
1201	*clean = false;
1202
1203	/*
1204	 * Look for unmount record. If we find it, then we know there was a
1205	 * clean unmount. Since 'i' could be the last block in the physical
1206	 * log, we convert to a log block before comparing to the head_blk.
1207	 *
1208	 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1209	 * below. We won't want to clear the unmount record if there is one, so
1210	 * we pass the lsn of the unmount record rather than the block after it.
1211	 */
1212	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1213		int	h_size = be32_to_cpu(rhead->h_size);
1214		int	h_version = be32_to_cpu(rhead->h_version);
1215
1216		if ((h_version & XLOG_VERSION_2) &&
1217		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1218			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1219			if (h_size % XLOG_HEADER_CYCLE_SIZE)
1220				hblks++;
1221		} else {
1222			hblks = 1;
1223		}
1224	} else {
1225		hblks = 1;
1226	}
1227
1228	after_umount_blk = xlog_wrap_logbno(log,
1229			rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len)));
1230
1231	if (*head_blk == after_umount_blk &&
1232	    be32_to_cpu(rhead->h_num_logops) == 1) {
1233		umount_data_blk = xlog_wrap_logbno(log, rhead_blk + hblks);
1234		error = xlog_bread(log, umount_data_blk, 1, buffer, &offset);
1235		if (error)
1236			return error;
1237
1238		op_head = (struct xlog_op_header *)offset;
1239		if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1240			/*
1241			 * Set tail and last sync so that newly written log
1242			 * records will point recovery to after the current
1243			 * unmount record.
1244			 */
1245			xlog_assign_atomic_lsn(&log->l_tail_lsn,
1246					log->l_curr_cycle, after_umount_blk);
1247			xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1248					log->l_curr_cycle, after_umount_blk);
1249			*tail_blk = after_umount_blk;
1250
1251			*clean = true;
1252		}
1253	}
1254
1255	return 0;
1256}
1257
1258static void
1259xlog_set_state(
1260	struct xlog		*log,
1261	xfs_daddr_t		head_blk,
1262	struct xlog_rec_header	*rhead,
1263	xfs_daddr_t		rhead_blk,
1264	bool			bump_cycle)
1265{
1266	/*
1267	 * Reset log values according to the state of the log when we
1268	 * crashed.  In the case where head_blk == 0, we bump curr_cycle
1269	 * one because the next write starts a new cycle rather than
1270	 * continuing the cycle of the last good log record.  At this
1271	 * point we have guaranteed that all partial log records have been
1272	 * accounted for.  Therefore, we know that the last good log record
1273	 * written was complete and ended exactly on the end boundary
1274	 * of the physical log.
1275	 */
1276	log->l_prev_block = rhead_blk;
1277	log->l_curr_block = (int)head_blk;
1278	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1279	if (bump_cycle)
1280		log->l_curr_cycle++;
1281	atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1282	atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1283	xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1284					BBTOB(log->l_curr_block));
1285	xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1286					BBTOB(log->l_curr_block));
1287}
1288
1289/*
1290 * Find the sync block number or the tail of the log.
1291 *
1292 * This will be the block number of the last record to have its
1293 * associated buffers synced to disk.  Every log record header has
1294 * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
1295 * to get a sync block number.  The only concern is to figure out which
1296 * log record header to believe.
1297 *
1298 * The following algorithm uses the log record header with the largest
1299 * lsn.  The entire log record does not need to be valid.  We only care
1300 * that the header is valid.
1301 *
1302 * We could speed up search by using current head_blk buffer, but it is not
1303 * available.
1304 */
1305STATIC int
1306xlog_find_tail(
1307	struct xlog		*log,
1308	xfs_daddr_t		*head_blk,
1309	xfs_daddr_t		*tail_blk)
1310{
1311	xlog_rec_header_t	*rhead;
1312	char			*offset = NULL;
1313	char			*buffer;
1314	int			error;
1315	xfs_daddr_t		rhead_blk;
1316	xfs_lsn_t		tail_lsn;
1317	bool			wrapped = false;
1318	bool			clean = false;
1319
1320	/*
1321	 * Find previous log record
1322	 */
1323	if ((error = xlog_find_head(log, head_blk)))
1324		return error;
1325	ASSERT(*head_blk < INT_MAX);
1326
1327	buffer = xlog_alloc_buffer(log, 1);
1328	if (!buffer)
1329		return -ENOMEM;
1330	if (*head_blk == 0) {				/* special case */
1331		error = xlog_bread(log, 0, 1, buffer, &offset);
1332		if (error)
1333			goto done;
1334
1335		if (xlog_get_cycle(offset) == 0) {
1336			*tail_blk = 0;
1337			/* leave all other log inited values alone */
1338			goto done;
1339		}
1340	}
1341
1342	/*
1343	 * Search backwards through the log looking for the log record header
1344	 * block. This wraps all the way back around to the head so something is
1345	 * seriously wrong if we can't find it.
1346	 */
1347	error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, buffer,
1348				      &rhead_blk, &rhead, &wrapped);
1349	if (error < 0)
1350		return error;
1351	if (!error) {
1352		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1353		return -EIO;
 
1354	}
1355	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1356
1357	/*
1358	 * Set the log state based on the current head record.
1359	 */
1360	xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1361	tail_lsn = atomic64_read(&log->l_tail_lsn);
1362
1363	/*
1364	 * Look for an unmount record at the head of the log. This sets the log
1365	 * state to determine whether recovery is necessary.
1366	 */
1367	error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1368				       rhead_blk, buffer, &clean);
1369	if (error)
1370		goto done;
1371
1372	/*
1373	 * Verify the log head if the log is not clean (e.g., we have anything
1374	 * but an unmount record at the head). This uses CRC verification to
1375	 * detect and trim torn writes. If discovered, CRC failures are
1376	 * considered torn writes and the log head is trimmed accordingly.
1377	 *
1378	 * Note that we can only run CRC verification when the log is dirty
1379	 * because there's no guarantee that the log data behind an unmount
1380	 * record is compatible with the current architecture.
1381	 */
1382	if (!clean) {
1383		xfs_daddr_t	orig_head = *head_blk;
1384
1385		error = xlog_verify_head(log, head_blk, tail_blk, buffer,
1386					 &rhead_blk, &rhead, &wrapped);
1387		if (error)
1388			goto done;
1389
1390		/* update in-core state again if the head changed */
1391		if (*head_blk != orig_head) {
1392			xlog_set_state(log, *head_blk, rhead, rhead_blk,
1393				       wrapped);
1394			tail_lsn = atomic64_read(&log->l_tail_lsn);
1395			error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1396						       rhead, rhead_blk, buffer,
1397						       &clean);
1398			if (error)
1399				goto done;
1400		}
1401	}
1402
1403	/*
1404	 * Note that the unmount was clean. If the unmount was not clean, we
1405	 * need to know this to rebuild the superblock counters from the perag
1406	 * headers if we have a filesystem using non-persistent counters.
1407	 */
1408	if (clean)
1409		log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1410
1411	/*
1412	 * Make sure that there are no blocks in front of the head
1413	 * with the same cycle number as the head.  This can happen
1414	 * because we allow multiple outstanding log writes concurrently,
1415	 * and the later writes might make it out before earlier ones.
1416	 *
1417	 * We use the lsn from before modifying it so that we'll never
1418	 * overwrite the unmount record after a clean unmount.
1419	 *
1420	 * Do this only if we are going to recover the filesystem
1421	 *
1422	 * NOTE: This used to say "if (!readonly)"
1423	 * However on Linux, we can & do recover a read-only filesystem.
1424	 * We only skip recovery if NORECOVERY is specified on mount,
1425	 * in which case we would not be here.
1426	 *
1427	 * But... if the -device- itself is readonly, just skip this.
1428	 * We can't recover this device anyway, so it won't matter.
1429	 */
1430	if (!xfs_readonly_buftarg(log->l_targ))
1431		error = xlog_clear_stale_blocks(log, tail_lsn);
1432
1433done:
1434	kmem_free(buffer);
1435
1436	if (error)
1437		xfs_warn(log->l_mp, "failed to locate log tail");
1438	return error;
1439}
1440
1441/*
1442 * Is the log zeroed at all?
1443 *
1444 * The last binary search should be changed to perform an X block read
1445 * once X becomes small enough.  You can then search linearly through
1446 * the X blocks.  This will cut down on the number of reads we need to do.
1447 *
1448 * If the log is partially zeroed, this routine will pass back the blkno
1449 * of the first block with cycle number 0.  It won't have a complete LR
1450 * preceding it.
1451 *
1452 * Return:
1453 *	0  => the log is completely written to
1454 *	1 => use *blk_no as the first block of the log
1455 *	<0 => error has occurred
1456 */
1457STATIC int
1458xlog_find_zeroed(
1459	struct xlog	*log,
1460	xfs_daddr_t	*blk_no)
1461{
1462	char		*buffer;
1463	char		*offset;
1464	uint	        first_cycle, last_cycle;
1465	xfs_daddr_t	new_blk, last_blk, start_blk;
1466	xfs_daddr_t     num_scan_bblks;
1467	int	        error, log_bbnum = log->l_logBBsize;
1468
1469	*blk_no = 0;
1470
1471	/* check totally zeroed log */
1472	buffer = xlog_alloc_buffer(log, 1);
1473	if (!buffer)
1474		return -ENOMEM;
1475	error = xlog_bread(log, 0, 1, buffer, &offset);
1476	if (error)
1477		goto out_free_buffer;
1478
1479	first_cycle = xlog_get_cycle(offset);
1480	if (first_cycle == 0) {		/* completely zeroed log */
1481		*blk_no = 0;
1482		kmem_free(buffer);
1483		return 1;
1484	}
1485
1486	/* check partially zeroed log */
1487	error = xlog_bread(log, log_bbnum-1, 1, buffer, &offset);
1488	if (error)
1489		goto out_free_buffer;
1490
1491	last_cycle = xlog_get_cycle(offset);
1492	if (last_cycle != 0) {		/* log completely written to */
1493		kmem_free(buffer);
1494		return 0;
1495	}
1496
1497	/* we have a partially zeroed log */
1498	last_blk = log_bbnum-1;
1499	error = xlog_find_cycle_start(log, buffer, 0, &last_blk, 0);
1500	if (error)
1501		goto out_free_buffer;
1502
1503	/*
1504	 * Validate the answer.  Because there is no way to guarantee that
1505	 * the entire log is made up of log records which are the same size,
1506	 * we scan over the defined maximum blocks.  At this point, the maximum
1507	 * is not chosen to mean anything special.   XXXmiken
1508	 */
1509	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1510	ASSERT(num_scan_bblks <= INT_MAX);
1511
1512	if (last_blk < num_scan_bblks)
1513		num_scan_bblks = last_blk;
1514	start_blk = last_blk - num_scan_bblks;
1515
1516	/*
1517	 * We search for any instances of cycle number 0 that occur before
1518	 * our current estimate of the head.  What we're trying to detect is
1519	 *        1 ... | 0 | 1 | 0...
1520	 *                       ^ binary search ends here
1521	 */
1522	if ((error = xlog_find_verify_cycle(log, start_blk,
1523					 (int)num_scan_bblks, 0, &new_blk)))
1524		goto out_free_buffer;
1525	if (new_blk != -1)
1526		last_blk = new_blk;
1527
1528	/*
1529	 * Potentially backup over partial log record write.  We don't need
1530	 * to search the end of the log because we know it is zero.
1531	 */
1532	error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1533	if (error == 1)
1534		error = -EIO;
1535	if (error)
1536		goto out_free_buffer;
1537
1538	*blk_no = last_blk;
1539out_free_buffer:
1540	kmem_free(buffer);
1541	if (error)
1542		return error;
1543	return 1;
1544}
1545
1546/*
1547 * These are simple subroutines used by xlog_clear_stale_blocks() below
1548 * to initialize a buffer full of empty log record headers and write
1549 * them into the log.
1550 */
1551STATIC void
1552xlog_add_record(
1553	struct xlog		*log,
1554	char			*buf,
1555	int			cycle,
1556	int			block,
1557	int			tail_cycle,
1558	int			tail_block)
1559{
1560	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1561
1562	memset(buf, 0, BBSIZE);
1563	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1564	recp->h_cycle = cpu_to_be32(cycle);
1565	recp->h_version = cpu_to_be32(
1566			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1567	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1568	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1569	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1570	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1571}
1572
1573STATIC int
1574xlog_write_log_records(
1575	struct xlog	*log,
1576	int		cycle,
1577	int		start_block,
1578	int		blocks,
1579	int		tail_cycle,
1580	int		tail_block)
1581{
1582	char		*offset;
1583	char		*buffer;
1584	int		balign, ealign;
1585	int		sectbb = log->l_sectBBsize;
1586	int		end_block = start_block + blocks;
1587	int		bufblks;
1588	int		error = 0;
1589	int		i, j = 0;
1590
1591	/*
1592	 * Greedily allocate a buffer big enough to handle the full
1593	 * range of basic blocks to be written.  If that fails, try
1594	 * a smaller size.  We need to be able to write at least a
1595	 * log sector, or we're out of luck.
1596	 */
1597	bufblks = 1 << ffs(blocks);
1598	while (bufblks > log->l_logBBsize)
1599		bufblks >>= 1;
1600	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
1601		bufblks >>= 1;
1602		if (bufblks < sectbb)
1603			return -ENOMEM;
1604	}
1605
1606	/* We may need to do a read at the start to fill in part of
1607	 * the buffer in the starting sector not covered by the first
1608	 * write below.
1609	 */
1610	balign = round_down(start_block, sectbb);
1611	if (balign != start_block) {
1612		error = xlog_bread_noalign(log, start_block, 1, buffer);
1613		if (error)
1614			goto out_free_buffer;
1615
1616		j = start_block - balign;
1617	}
1618
1619	for (i = start_block; i < end_block; i += bufblks) {
1620		int		bcount, endcount;
1621
1622		bcount = min(bufblks, end_block - start_block);
1623		endcount = bcount - j;
1624
1625		/* We may need to do a read at the end to fill in part of
1626		 * the buffer in the final sector not covered by the write.
1627		 * If this is the same sector as the above read, skip it.
1628		 */
1629		ealign = round_down(end_block, sectbb);
1630		if (j == 0 && (start_block + endcount > ealign)) {
1631			error = xlog_bread_noalign(log, ealign, sectbb,
1632					buffer + BBTOB(ealign - start_block));
1633			if (error)
1634				break;
1635
1636		}
1637
1638		offset = buffer + xlog_align(log, start_block);
1639		for (; j < endcount; j++) {
1640			xlog_add_record(log, offset, cycle, i+j,
1641					tail_cycle, tail_block);
1642			offset += BBSIZE;
1643		}
1644		error = xlog_bwrite(log, start_block, endcount, buffer);
1645		if (error)
1646			break;
1647		start_block += endcount;
1648		j = 0;
1649	}
1650
1651out_free_buffer:
1652	kmem_free(buffer);
1653	return error;
1654}
1655
1656/*
1657 * This routine is called to blow away any incomplete log writes out
1658 * in front of the log head.  We do this so that we won't become confused
1659 * if we come up, write only a little bit more, and then crash again.
1660 * If we leave the partial log records out there, this situation could
1661 * cause us to think those partial writes are valid blocks since they
1662 * have the current cycle number.  We get rid of them by overwriting them
1663 * with empty log records with the old cycle number rather than the
1664 * current one.
1665 *
1666 * The tail lsn is passed in rather than taken from
1667 * the log so that we will not write over the unmount record after a
1668 * clean unmount in a 512 block log.  Doing so would leave the log without
1669 * any valid log records in it until a new one was written.  If we crashed
1670 * during that time we would not be able to recover.
1671 */
1672STATIC int
1673xlog_clear_stale_blocks(
1674	struct xlog	*log,
1675	xfs_lsn_t	tail_lsn)
1676{
1677	int		tail_cycle, head_cycle;
1678	int		tail_block, head_block;
1679	int		tail_distance, max_distance;
1680	int		distance;
1681	int		error;
1682
1683	tail_cycle = CYCLE_LSN(tail_lsn);
1684	tail_block = BLOCK_LSN(tail_lsn);
1685	head_cycle = log->l_curr_cycle;
1686	head_block = log->l_curr_block;
1687
1688	/*
1689	 * Figure out the distance between the new head of the log
1690	 * and the tail.  We want to write over any blocks beyond the
1691	 * head that we may have written just before the crash, but
1692	 * we don't want to overwrite the tail of the log.
1693	 */
1694	if (head_cycle == tail_cycle) {
1695		/*
1696		 * The tail is behind the head in the physical log,
1697		 * so the distance from the head to the tail is the
1698		 * distance from the head to the end of the log plus
1699		 * the distance from the beginning of the log to the
1700		 * tail.
1701		 */
1702		if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1703			XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1704					 XFS_ERRLEVEL_LOW, log->l_mp);
1705			return -EFSCORRUPTED;
1706		}
1707		tail_distance = tail_block + (log->l_logBBsize - head_block);
1708	} else {
1709		/*
1710		 * The head is behind the tail in the physical log,
1711		 * so the distance from the head to the tail is just
1712		 * the tail block minus the head block.
1713		 */
1714		if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1715			XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1716					 XFS_ERRLEVEL_LOW, log->l_mp);
1717			return -EFSCORRUPTED;
1718		}
1719		tail_distance = tail_block - head_block;
1720	}
1721
1722	/*
1723	 * If the head is right up against the tail, we can't clear
1724	 * anything.
1725	 */
1726	if (tail_distance <= 0) {
1727		ASSERT(tail_distance == 0);
1728		return 0;
1729	}
1730
1731	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1732	/*
1733	 * Take the smaller of the maximum amount of outstanding I/O
1734	 * we could have and the distance to the tail to clear out.
1735	 * We take the smaller so that we don't overwrite the tail and
1736	 * we don't waste all day writing from the head to the tail
1737	 * for no reason.
1738	 */
1739	max_distance = min(max_distance, tail_distance);
1740
1741	if ((head_block + max_distance) <= log->l_logBBsize) {
1742		/*
1743		 * We can stomp all the blocks we need to without
1744		 * wrapping around the end of the log.  Just do it
1745		 * in a single write.  Use the cycle number of the
1746		 * current cycle minus one so that the log will look like:
1747		 *     n ... | n - 1 ...
1748		 */
1749		error = xlog_write_log_records(log, (head_cycle - 1),
1750				head_block, max_distance, tail_cycle,
1751				tail_block);
1752		if (error)
1753			return error;
1754	} else {
1755		/*
1756		 * We need to wrap around the end of the physical log in
1757		 * order to clear all the blocks.  Do it in two separate
1758		 * I/Os.  The first write should be from the head to the
1759		 * end of the physical log, and it should use the current
1760		 * cycle number minus one just like above.
1761		 */
1762		distance = log->l_logBBsize - head_block;
1763		error = xlog_write_log_records(log, (head_cycle - 1),
1764				head_block, distance, tail_cycle,
1765				tail_block);
1766
1767		if (error)
1768			return error;
1769
1770		/*
1771		 * Now write the blocks at the start of the physical log.
1772		 * This writes the remainder of the blocks we want to clear.
1773		 * It uses the current cycle number since we're now on the
1774		 * same cycle as the head so that we get:
1775		 *    n ... n ... | n - 1 ...
1776		 *    ^^^^^ blocks we're writing
1777		 */
1778		distance = max_distance - (log->l_logBBsize - head_block);
1779		error = xlog_write_log_records(log, head_cycle, 0, distance,
1780				tail_cycle, tail_block);
1781		if (error)
1782			return error;
1783	}
1784
1785	return 0;
1786}
1787
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1788/******************************************************************************
1789 *
1790 *		Log recover routines
1791 *
1792 ******************************************************************************
1793 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1794
1795/*
1796 * Sort the log items in the transaction.
1797 *
1798 * The ordering constraints are defined by the inode allocation and unlink
1799 * behaviour. The rules are:
1800 *
1801 *	1. Every item is only logged once in a given transaction. Hence it
1802 *	   represents the last logged state of the item. Hence ordering is
1803 *	   dependent on the order in which operations need to be performed so
1804 *	   required initial conditions are always met.
1805 *
1806 *	2. Cancelled buffers are recorded in pass 1 in a separate table and
1807 *	   there's nothing to replay from them so we can simply cull them
1808 *	   from the transaction. However, we can't do that until after we've
1809 *	   replayed all the other items because they may be dependent on the
1810 *	   cancelled buffer and replaying the cancelled buffer can remove it
1811 *	   form the cancelled buffer table. Hence they have tobe done last.
1812 *
1813 *	3. Inode allocation buffers must be replayed before inode items that
1814 *	   read the buffer and replay changes into it. For filesystems using the
1815 *	   ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1816 *	   treated the same as inode allocation buffers as they create and
1817 *	   initialise the buffers directly.
1818 *
1819 *	4. Inode unlink buffers must be replayed after inode items are replayed.
1820 *	   This ensures that inodes are completely flushed to the inode buffer
1821 *	   in a "free" state before we remove the unlinked inode list pointer.
1822 *
1823 * Hence the ordering needs to be inode allocation buffers first, inode items
1824 * second, inode unlink buffers third and cancelled buffers last.
1825 *
1826 * But there's a problem with that - we can't tell an inode allocation buffer
1827 * apart from a regular buffer, so we can't separate them. We can, however,
1828 * tell an inode unlink buffer from the others, and so we can separate them out
1829 * from all the other buffers and move them to last.
1830 *
1831 * Hence, 4 lists, in order from head to tail:
1832 *	- buffer_list for all buffers except cancelled/inode unlink buffers
1833 *	- item_list for all non-buffer items
1834 *	- inode_buffer_list for inode unlink buffers
1835 *	- cancel_list for the cancelled buffers
1836 *
1837 * Note that we add objects to the tail of the lists so that first-to-last
1838 * ordering is preserved within the lists. Adding objects to the head of the
1839 * list means when we traverse from the head we walk them in last-to-first
1840 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1841 * but for all other items there may be specific ordering that we need to
1842 * preserve.
1843 */
1844STATIC int
1845xlog_recover_reorder_trans(
1846	struct xlog		*log,
1847	struct xlog_recover	*trans,
1848	int			pass)
1849{
1850	xlog_recover_item_t	*item, *n;
1851	int			error = 0;
1852	LIST_HEAD(sort_list);
1853	LIST_HEAD(cancel_list);
1854	LIST_HEAD(buffer_list);
1855	LIST_HEAD(inode_buffer_list);
1856	LIST_HEAD(inode_list);
1857
1858	list_splice_init(&trans->r_itemq, &sort_list);
1859	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1860		xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
1861
1862		switch (ITEM_TYPE(item)) {
1863		case XFS_LI_ICREATE:
1864			list_move_tail(&item->ri_list, &buffer_list);
1865			break;
1866		case XFS_LI_BUF:
1867			if (buf_f->blf_flags & XFS_BLF_CANCEL) {
1868				trace_xfs_log_recover_item_reorder_head(log,
1869							trans, item, pass);
1870				list_move(&item->ri_list, &cancel_list);
1871				break;
1872			}
1873			if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
1874				list_move(&item->ri_list, &inode_buffer_list);
1875				break;
1876			}
1877			list_move_tail(&item->ri_list, &buffer_list);
1878			break;
1879		case XFS_LI_INODE:
1880		case XFS_LI_DQUOT:
1881		case XFS_LI_QUOTAOFF:
1882		case XFS_LI_EFD:
1883		case XFS_LI_EFI:
1884		case XFS_LI_RUI:
1885		case XFS_LI_RUD:
1886		case XFS_LI_CUI:
1887		case XFS_LI_CUD:
1888		case XFS_LI_BUI:
1889		case XFS_LI_BUD:
1890			trace_xfs_log_recover_item_reorder_tail(log,
1891							trans, item, pass);
1892			list_move_tail(&item->ri_list, &inode_list);
1893			break;
1894		default:
1895			xfs_warn(log->l_mp,
1896				"%s: unrecognized type of log operation",
1897				__func__);
1898			ASSERT(0);
1899			/*
1900			 * return the remaining items back to the transaction
1901			 * item list so they can be freed in caller.
1902			 */
1903			if (!list_empty(&sort_list))
1904				list_splice_init(&sort_list, &trans->r_itemq);
1905			error = -EIO;
1906			goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1907		}
1908	}
1909out:
1910	ASSERT(list_empty(&sort_list));
1911	if (!list_empty(&buffer_list))
1912		list_splice(&buffer_list, &trans->r_itemq);
1913	if (!list_empty(&inode_list))
1914		list_splice_tail(&inode_list, &trans->r_itemq);
1915	if (!list_empty(&inode_buffer_list))
1916		list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1917	if (!list_empty(&cancel_list))
1918		list_splice_tail(&cancel_list, &trans->r_itemq);
1919	return error;
1920}
1921
1922/*
1923 * Build up the table of buf cancel records so that we don't replay
1924 * cancelled data in the second pass.  For buffer records that are
1925 * not cancel records, there is nothing to do here so we just return.
1926 *
1927 * If we get a cancel record which is already in the table, this indicates
1928 * that the buffer was cancelled multiple times.  In order to ensure
1929 * that during pass 2 we keep the record in the table until we reach its
1930 * last occurrence in the log, we keep a reference count in the cancel
1931 * record in the table to tell us how many times we expect to see this
1932 * record during the second pass.
1933 */
1934STATIC int
1935xlog_recover_buffer_pass1(
1936	struct xlog			*log,
1937	struct xlog_recover_item	*item)
1938{
1939	xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
1940	struct list_head	*bucket;
1941	struct xfs_buf_cancel	*bcp;
1942
1943	/*
1944	 * If this isn't a cancel buffer item, then just return.
1945	 */
1946	if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1947		trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1948		return 0;
1949	}
1950
1951	/*
1952	 * Insert an xfs_buf_cancel record into the hash table of them.
1953	 * If there is already an identical record, bump its reference count.
1954	 */
1955	bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1956	list_for_each_entry(bcp, bucket, bc_list) {
1957		if (bcp->bc_blkno == buf_f->blf_blkno &&
1958		    bcp->bc_len == buf_f->blf_len) {
1959			bcp->bc_refcount++;
1960			trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1961			return 0;
1962		}
1963	}
1964
1965	bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), 0);
1966	bcp->bc_blkno = buf_f->blf_blkno;
1967	bcp->bc_len = buf_f->blf_len;
1968	bcp->bc_refcount = 1;
1969	list_add_tail(&bcp->bc_list, bucket);
1970
1971	trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1972	return 0;
1973}
1974
1975/*
1976 * Check to see whether the buffer being recovered has a corresponding
1977 * entry in the buffer cancel record table. If it is, return the cancel
1978 * buffer structure to the caller.
1979 */
1980STATIC struct xfs_buf_cancel *
1981xlog_peek_buffer_cancelled(
1982	struct xlog		*log,
1983	xfs_daddr_t		blkno,
1984	uint			len,
1985	unsigned short			flags)
1986{
1987	struct list_head	*bucket;
1988	struct xfs_buf_cancel	*bcp;
1989
1990	if (!log->l_buf_cancel_table) {
1991		/* empty table means no cancelled buffers in the log */
1992		ASSERT(!(flags & XFS_BLF_CANCEL));
1993		return NULL;
1994	}
1995
1996	bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
1997	list_for_each_entry(bcp, bucket, bc_list) {
1998		if (bcp->bc_blkno == blkno && bcp->bc_len == len)
1999			return bcp;
2000	}
2001
2002	/*
2003	 * We didn't find a corresponding entry in the table, so return 0 so
2004	 * that the buffer is NOT cancelled.
2005	 */
2006	ASSERT(!(flags & XFS_BLF_CANCEL));
2007	return NULL;
2008}
2009
2010/*
2011 * If the buffer is being cancelled then return 1 so that it will be cancelled,
2012 * otherwise return 0.  If the buffer is actually a buffer cancel item
2013 * (XFS_BLF_CANCEL is set), then decrement the refcount on the entry in the
2014 * table and remove it from the table if this is the last reference.
2015 *
2016 * We remove the cancel record from the table when we encounter its last
2017 * occurrence in the log so that if the same buffer is re-used again after its
2018 * last cancellation we actually replay the changes made at that point.
2019 */
2020STATIC int
2021xlog_check_buffer_cancelled(
2022	struct xlog		*log,
2023	xfs_daddr_t		blkno,
2024	uint			len,
2025	unsigned short			flags)
2026{
2027	struct xfs_buf_cancel	*bcp;
2028
2029	bcp = xlog_peek_buffer_cancelled(log, blkno, len, flags);
2030	if (!bcp)
2031		return 0;
2032
2033	/*
2034	 * We've go a match, so return 1 so that the recovery of this buffer
2035	 * is cancelled.  If this buffer is actually a buffer cancel log
2036	 * item, then decrement the refcount on the one in the table and
2037	 * remove it if this is the last reference.
2038	 */
2039	if (flags & XFS_BLF_CANCEL) {
2040		if (--bcp->bc_refcount == 0) {
2041			list_del(&bcp->bc_list);
2042			kmem_free(bcp);
2043		}
2044	}
2045	return 1;
2046}
2047
2048/*
2049 * Perform recovery for a buffer full of inodes.  In these buffers, the only
2050 * data which should be recovered is that which corresponds to the
2051 * di_next_unlinked pointers in the on disk inode structures.  The rest of the
2052 * data for the inodes is always logged through the inodes themselves rather
2053 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
2054 *
2055 * The only time when buffers full of inodes are fully recovered is when the
2056 * buffer is full of newly allocated inodes.  In this case the buffer will
2057 * not be marked as an inode buffer and so will be sent to
2058 * xlog_recover_do_reg_buffer() below during recovery.
2059 */
2060STATIC int
2061xlog_recover_do_inode_buffer(
2062	struct xfs_mount	*mp,
2063	xlog_recover_item_t	*item,
2064	struct xfs_buf		*bp,
2065	xfs_buf_log_format_t	*buf_f)
2066{
2067	int			i;
2068	int			item_index = 0;
2069	int			bit = 0;
2070	int			nbits = 0;
2071	int			reg_buf_offset = 0;
2072	int			reg_buf_bytes = 0;
2073	int			next_unlinked_offset;
2074	int			inodes_per_buf;
2075	xfs_agino_t		*logged_nextp;
2076	xfs_agino_t		*buffer_nextp;
2077
2078	trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
2079
2080	/*
2081	 * Post recovery validation only works properly on CRC enabled
2082	 * filesystems.
2083	 */
2084	if (xfs_sb_version_hascrc(&mp->m_sb))
2085		bp->b_ops = &xfs_inode_buf_ops;
2086
2087	inodes_per_buf = BBTOB(bp->b_length) >> mp->m_sb.sb_inodelog;
2088	for (i = 0; i < inodes_per_buf; i++) {
2089		next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
2090			offsetof(xfs_dinode_t, di_next_unlinked);
2091
2092		while (next_unlinked_offset >=
2093		       (reg_buf_offset + reg_buf_bytes)) {
2094			/*
2095			 * The next di_next_unlinked field is beyond
2096			 * the current logged region.  Find the next
2097			 * logged region that contains or is beyond
2098			 * the current di_next_unlinked field.
2099			 */
2100			bit += nbits;
2101			bit = xfs_next_bit(buf_f->blf_data_map,
2102					   buf_f->blf_map_size, bit);
2103
2104			/*
2105			 * If there are no more logged regions in the
2106			 * buffer, then we're done.
2107			 */
2108			if (bit == -1)
2109				return 0;
2110
2111			nbits = xfs_contig_bits(buf_f->blf_data_map,
2112						buf_f->blf_map_size, bit);
2113			ASSERT(nbits > 0);
2114			reg_buf_offset = bit << XFS_BLF_SHIFT;
2115			reg_buf_bytes = nbits << XFS_BLF_SHIFT;
2116			item_index++;
2117		}
2118
2119		/*
2120		 * If the current logged region starts after the current
2121		 * di_next_unlinked field, then move on to the next
2122		 * di_next_unlinked field.
2123		 */
2124		if (next_unlinked_offset < reg_buf_offset)
2125			continue;
2126
2127		ASSERT(item->ri_buf[item_index].i_addr != NULL);
2128		ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
2129		ASSERT((reg_buf_offset + reg_buf_bytes) <= BBTOB(bp->b_length));
2130
2131		/*
2132		 * The current logged region contains a copy of the
2133		 * current di_next_unlinked field.  Extract its value
2134		 * and copy it to the buffer copy.
2135		 */
2136		logged_nextp = item->ri_buf[item_index].i_addr +
2137				next_unlinked_offset - reg_buf_offset;
2138		if (unlikely(*logged_nextp == 0)) {
2139			xfs_alert(mp,
2140		"Bad inode buffer log record (ptr = "PTR_FMT", bp = "PTR_FMT"). "
2141		"Trying to replay bad (0) inode di_next_unlinked field.",
2142				item, bp);
2143			XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
2144					 XFS_ERRLEVEL_LOW, mp);
2145			return -EFSCORRUPTED;
2146		}
2147
2148		buffer_nextp = xfs_buf_offset(bp, next_unlinked_offset);
2149		*buffer_nextp = *logged_nextp;
2150
2151		/*
2152		 * If necessary, recalculate the CRC in the on-disk inode. We
2153		 * have to leave the inode in a consistent state for whoever
2154		 * reads it next....
2155		 */
2156		xfs_dinode_calc_crc(mp,
2157				xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
2158
2159	}
2160
2161	return 0;
2162}
2163
2164/*
2165 * V5 filesystems know the age of the buffer on disk being recovered. We can
2166 * have newer objects on disk than we are replaying, and so for these cases we
2167 * don't want to replay the current change as that will make the buffer contents
2168 * temporarily invalid on disk.
2169 *
2170 * The magic number might not match the buffer type we are going to recover
2171 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags.  Hence
2172 * extract the LSN of the existing object in the buffer based on it's current
2173 * magic number.  If we don't recognise the magic number in the buffer, then
2174 * return a LSN of -1 so that the caller knows it was an unrecognised block and
2175 * so can recover the buffer.
2176 *
2177 * Note: we cannot rely solely on magic number matches to determine that the
2178 * buffer has a valid LSN - we also need to verify that it belongs to this
2179 * filesystem, so we need to extract the object's LSN and compare it to that
2180 * which we read from the superblock. If the UUIDs don't match, then we've got a
2181 * stale metadata block from an old filesystem instance that we need to recover
2182 * over the top of.
2183 */
2184static xfs_lsn_t
2185xlog_recover_get_buf_lsn(
2186	struct xfs_mount	*mp,
2187	struct xfs_buf		*bp)
2188{
2189	uint32_t		magic32;
2190	uint16_t		magic16;
2191	uint16_t		magicda;
2192	void			*blk = bp->b_addr;
2193	uuid_t			*uuid;
2194	xfs_lsn_t		lsn = -1;
2195
2196	/* v4 filesystems always recover immediately */
2197	if (!xfs_sb_version_hascrc(&mp->m_sb))
2198		goto recover_immediately;
2199
2200	magic32 = be32_to_cpu(*(__be32 *)blk);
2201	switch (magic32) {
2202	case XFS_ABTB_CRC_MAGIC:
2203	case XFS_ABTC_CRC_MAGIC:
2204	case XFS_ABTB_MAGIC:
2205	case XFS_ABTC_MAGIC:
2206	case XFS_RMAP_CRC_MAGIC:
2207	case XFS_REFC_CRC_MAGIC:
2208	case XFS_IBT_CRC_MAGIC:
2209	case XFS_IBT_MAGIC: {
2210		struct xfs_btree_block *btb = blk;
2211
2212		lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
2213		uuid = &btb->bb_u.s.bb_uuid;
2214		break;
2215	}
2216	case XFS_BMAP_CRC_MAGIC:
2217	case XFS_BMAP_MAGIC: {
2218		struct xfs_btree_block *btb = blk;
2219
2220		lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
2221		uuid = &btb->bb_u.l.bb_uuid;
2222		break;
2223	}
2224	case XFS_AGF_MAGIC:
2225		lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
2226		uuid = &((struct xfs_agf *)blk)->agf_uuid;
2227		break;
2228	case XFS_AGFL_MAGIC:
2229		lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
2230		uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
2231		break;
2232	case XFS_AGI_MAGIC:
2233		lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
2234		uuid = &((struct xfs_agi *)blk)->agi_uuid;
2235		break;
2236	case XFS_SYMLINK_MAGIC:
2237		lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
2238		uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
2239		break;
2240	case XFS_DIR3_BLOCK_MAGIC:
2241	case XFS_DIR3_DATA_MAGIC:
2242	case XFS_DIR3_FREE_MAGIC:
2243		lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
2244		uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
2245		break;
2246	case XFS_ATTR3_RMT_MAGIC:
2247		/*
2248		 * Remote attr blocks are written synchronously, rather than
2249		 * being logged. That means they do not contain a valid LSN
2250		 * (i.e. transactionally ordered) in them, and hence any time we
2251		 * see a buffer to replay over the top of a remote attribute
2252		 * block we should simply do so.
2253		 */
2254		goto recover_immediately;
2255	case XFS_SB_MAGIC:
2256		/*
2257		 * superblock uuids are magic. We may or may not have a
2258		 * sb_meta_uuid on disk, but it will be set in the in-core
2259		 * superblock. We set the uuid pointer for verification
2260		 * according to the superblock feature mask to ensure we check
2261		 * the relevant UUID in the superblock.
2262		 */
2263		lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
2264		if (xfs_sb_version_hasmetauuid(&mp->m_sb))
2265			uuid = &((struct xfs_dsb *)blk)->sb_meta_uuid;
2266		else
2267			uuid = &((struct xfs_dsb *)blk)->sb_uuid;
2268		break;
2269	default:
2270		break;
2271	}
2272
2273	if (lsn != (xfs_lsn_t)-1) {
2274		if (!uuid_equal(&mp->m_sb.sb_meta_uuid, uuid))
2275			goto recover_immediately;
2276		return lsn;
2277	}
2278
2279	magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
2280	switch (magicda) {
2281	case XFS_DIR3_LEAF1_MAGIC:
2282	case XFS_DIR3_LEAFN_MAGIC:
2283	case XFS_DA3_NODE_MAGIC:
2284		lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
2285		uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
2286		break;
2287	default:
2288		break;
2289	}
2290
2291	if (lsn != (xfs_lsn_t)-1) {
2292		if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
2293			goto recover_immediately;
2294		return lsn;
2295	}
2296
2297	/*
2298	 * We do individual object checks on dquot and inode buffers as they
2299	 * have their own individual LSN records. Also, we could have a stale
2300	 * buffer here, so we have to at least recognise these buffer types.
2301	 *
2302	 * A notd complexity here is inode unlinked list processing - it logs
2303	 * the inode directly in the buffer, but we don't know which inodes have
2304	 * been modified, and there is no global buffer LSN. Hence we need to
2305	 * recover all inode buffer types immediately. This problem will be
2306	 * fixed by logical logging of the unlinked list modifications.
2307	 */
2308	magic16 = be16_to_cpu(*(__be16 *)blk);
2309	switch (magic16) {
2310	case XFS_DQUOT_MAGIC:
2311	case XFS_DINODE_MAGIC:
2312		goto recover_immediately;
2313	default:
2314		break;
2315	}
2316
2317	/* unknown buffer contents, recover immediately */
2318
2319recover_immediately:
2320	return (xfs_lsn_t)-1;
2321
2322}
2323
2324/*
2325 * Validate the recovered buffer is of the correct type and attach the
2326 * appropriate buffer operations to them for writeback. Magic numbers are in a
2327 * few places:
2328 *	the first 16 bits of the buffer (inode buffer, dquot buffer),
2329 *	the first 32 bits of the buffer (most blocks),
2330 *	inside a struct xfs_da_blkinfo at the start of the buffer.
2331 */
2332static void
2333xlog_recover_validate_buf_type(
2334	struct xfs_mount	*mp,
2335	struct xfs_buf		*bp,
2336	xfs_buf_log_format_t	*buf_f,
2337	xfs_lsn_t		current_lsn)
2338{
2339	struct xfs_da_blkinfo	*info = bp->b_addr;
2340	uint32_t		magic32;
2341	uint16_t		magic16;
2342	uint16_t		magicda;
2343	char			*warnmsg = NULL;
2344
2345	/*
2346	 * We can only do post recovery validation on items on CRC enabled
2347	 * fielsystems as we need to know when the buffer was written to be able
2348	 * to determine if we should have replayed the item. If we replay old
2349	 * metadata over a newer buffer, then it will enter a temporarily
2350	 * inconsistent state resulting in verification failures. Hence for now
2351	 * just avoid the verification stage for non-crc filesystems
2352	 */
2353	if (!xfs_sb_version_hascrc(&mp->m_sb))
2354		return;
2355
2356	magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
2357	magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
2358	magicda = be16_to_cpu(info->magic);
2359	switch (xfs_blft_from_flags(buf_f)) {
2360	case XFS_BLFT_BTREE_BUF:
2361		switch (magic32) {
2362		case XFS_ABTB_CRC_MAGIC:
2363		case XFS_ABTB_MAGIC:
2364			bp->b_ops = &xfs_bnobt_buf_ops;
2365			break;
2366		case XFS_ABTC_CRC_MAGIC:
2367		case XFS_ABTC_MAGIC:
2368			bp->b_ops = &xfs_cntbt_buf_ops;
2369			break;
2370		case XFS_IBT_CRC_MAGIC:
2371		case XFS_IBT_MAGIC:
2372			bp->b_ops = &xfs_inobt_buf_ops;
2373			break;
2374		case XFS_FIBT_CRC_MAGIC:
2375		case XFS_FIBT_MAGIC:
2376			bp->b_ops = &xfs_finobt_buf_ops;
2377			break;
2378		case XFS_BMAP_CRC_MAGIC:
2379		case XFS_BMAP_MAGIC:
2380			bp->b_ops = &xfs_bmbt_buf_ops;
2381			break;
2382		case XFS_RMAP_CRC_MAGIC:
2383			bp->b_ops = &xfs_rmapbt_buf_ops;
2384			break;
2385		case XFS_REFC_CRC_MAGIC:
2386			bp->b_ops = &xfs_refcountbt_buf_ops;
2387			break;
2388		default:
2389			warnmsg = "Bad btree block magic!";
2390			break;
2391		}
2392		break;
2393	case XFS_BLFT_AGF_BUF:
2394		if (magic32 != XFS_AGF_MAGIC) {
2395			warnmsg = "Bad AGF block magic!";
2396			break;
2397		}
2398		bp->b_ops = &xfs_agf_buf_ops;
2399		break;
2400	case XFS_BLFT_AGFL_BUF:
2401		if (magic32 != XFS_AGFL_MAGIC) {
2402			warnmsg = "Bad AGFL block magic!";
2403			break;
2404		}
2405		bp->b_ops = &xfs_agfl_buf_ops;
2406		break;
2407	case XFS_BLFT_AGI_BUF:
2408		if (magic32 != XFS_AGI_MAGIC) {
2409			warnmsg = "Bad AGI block magic!";
2410			break;
2411		}
2412		bp->b_ops = &xfs_agi_buf_ops;
2413		break;
2414	case XFS_BLFT_UDQUOT_BUF:
2415	case XFS_BLFT_PDQUOT_BUF:
2416	case XFS_BLFT_GDQUOT_BUF:
2417#ifdef CONFIG_XFS_QUOTA
2418		if (magic16 != XFS_DQUOT_MAGIC) {
2419			warnmsg = "Bad DQUOT block magic!";
2420			break;
2421		}
2422		bp->b_ops = &xfs_dquot_buf_ops;
2423#else
2424		xfs_alert(mp,
2425	"Trying to recover dquots without QUOTA support built in!");
2426		ASSERT(0);
2427#endif
2428		break;
2429	case XFS_BLFT_DINO_BUF:
2430		if (magic16 != XFS_DINODE_MAGIC) {
2431			warnmsg = "Bad INODE block magic!";
2432			break;
2433		}
2434		bp->b_ops = &xfs_inode_buf_ops;
2435		break;
2436	case XFS_BLFT_SYMLINK_BUF:
2437		if (magic32 != XFS_SYMLINK_MAGIC) {
2438			warnmsg = "Bad symlink block magic!";
2439			break;
2440		}
2441		bp->b_ops = &xfs_symlink_buf_ops;
2442		break;
2443	case XFS_BLFT_DIR_BLOCK_BUF:
2444		if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
2445		    magic32 != XFS_DIR3_BLOCK_MAGIC) {
2446			warnmsg = "Bad dir block magic!";
2447			break;
2448		}
2449		bp->b_ops = &xfs_dir3_block_buf_ops;
2450		break;
2451	case XFS_BLFT_DIR_DATA_BUF:
2452		if (magic32 != XFS_DIR2_DATA_MAGIC &&
2453		    magic32 != XFS_DIR3_DATA_MAGIC) {
2454			warnmsg = "Bad dir data magic!";
2455			break;
2456		}
2457		bp->b_ops = &xfs_dir3_data_buf_ops;
2458		break;
2459	case XFS_BLFT_DIR_FREE_BUF:
2460		if (magic32 != XFS_DIR2_FREE_MAGIC &&
2461		    magic32 != XFS_DIR3_FREE_MAGIC) {
2462			warnmsg = "Bad dir3 free magic!";
2463			break;
2464		}
2465		bp->b_ops = &xfs_dir3_free_buf_ops;
2466		break;
2467	case XFS_BLFT_DIR_LEAF1_BUF:
2468		if (magicda != XFS_DIR2_LEAF1_MAGIC &&
2469		    magicda != XFS_DIR3_LEAF1_MAGIC) {
2470			warnmsg = "Bad dir leaf1 magic!";
2471			break;
2472		}
2473		bp->b_ops = &xfs_dir3_leaf1_buf_ops;
2474		break;
2475	case XFS_BLFT_DIR_LEAFN_BUF:
2476		if (magicda != XFS_DIR2_LEAFN_MAGIC &&
2477		    magicda != XFS_DIR3_LEAFN_MAGIC) {
2478			warnmsg = "Bad dir leafn magic!";
2479			break;
2480		}
2481		bp->b_ops = &xfs_dir3_leafn_buf_ops;
2482		break;
2483	case XFS_BLFT_DA_NODE_BUF:
2484		if (magicda != XFS_DA_NODE_MAGIC &&
2485		    magicda != XFS_DA3_NODE_MAGIC) {
2486			warnmsg = "Bad da node magic!";
2487			break;
2488		}
2489		bp->b_ops = &xfs_da3_node_buf_ops;
2490		break;
2491	case XFS_BLFT_ATTR_LEAF_BUF:
2492		if (magicda != XFS_ATTR_LEAF_MAGIC &&
2493		    magicda != XFS_ATTR3_LEAF_MAGIC) {
2494			warnmsg = "Bad attr leaf magic!";
2495			break;
2496		}
2497		bp->b_ops = &xfs_attr3_leaf_buf_ops;
2498		break;
2499	case XFS_BLFT_ATTR_RMT_BUF:
2500		if (magic32 != XFS_ATTR3_RMT_MAGIC) {
2501			warnmsg = "Bad attr remote magic!";
2502			break;
2503		}
2504		bp->b_ops = &xfs_attr3_rmt_buf_ops;
2505		break;
2506	case XFS_BLFT_SB_BUF:
2507		if (magic32 != XFS_SB_MAGIC) {
2508			warnmsg = "Bad SB block magic!";
2509			break;
2510		}
2511		bp->b_ops = &xfs_sb_buf_ops;
2512		break;
2513#ifdef CONFIG_XFS_RT
2514	case XFS_BLFT_RTBITMAP_BUF:
2515	case XFS_BLFT_RTSUMMARY_BUF:
2516		/* no magic numbers for verification of RT buffers */
2517		bp->b_ops = &xfs_rtbuf_ops;
2518		break;
2519#endif /* CONFIG_XFS_RT */
2520	default:
2521		xfs_warn(mp, "Unknown buffer type %d!",
2522			 xfs_blft_from_flags(buf_f));
2523		break;
2524	}
2525
2526	/*
2527	 * Nothing else to do in the case of a NULL current LSN as this means
2528	 * the buffer is more recent than the change in the log and will be
2529	 * skipped.
2530	 */
2531	if (current_lsn == NULLCOMMITLSN)
2532		return;
2533
2534	if (warnmsg) {
2535		xfs_warn(mp, warnmsg);
2536		ASSERT(0);
2537	}
2538
2539	/*
2540	 * We must update the metadata LSN of the buffer as it is written out to
2541	 * ensure that older transactions never replay over this one and corrupt
2542	 * the buffer. This can occur if log recovery is interrupted at some
2543	 * point after the current transaction completes, at which point a
2544	 * subsequent mount starts recovery from the beginning.
2545	 *
2546	 * Write verifiers update the metadata LSN from log items attached to
2547	 * the buffer. Therefore, initialize a bli purely to carry the LSN to
2548	 * the verifier. We'll clean it up in our ->iodone() callback.
2549	 */
2550	if (bp->b_ops) {
2551		struct xfs_buf_log_item	*bip;
2552
2553		ASSERT(!bp->b_iodone || bp->b_iodone == xlog_recover_iodone);
2554		bp->b_iodone = xlog_recover_iodone;
2555		xfs_buf_item_init(bp, mp);
2556		bip = bp->b_log_item;
2557		bip->bli_item.li_lsn = current_lsn;
2558	}
2559}
2560
2561/*
2562 * Perform a 'normal' buffer recovery.  Each logged region of the
2563 * buffer should be copied over the corresponding region in the
2564 * given buffer.  The bitmap in the buf log format structure indicates
2565 * where to place the logged data.
2566 */
2567STATIC void
2568xlog_recover_do_reg_buffer(
2569	struct xfs_mount	*mp,
2570	xlog_recover_item_t	*item,
2571	struct xfs_buf		*bp,
2572	xfs_buf_log_format_t	*buf_f,
2573	xfs_lsn_t		current_lsn)
2574{
2575	int			i;
2576	int			bit;
2577	int			nbits;
2578	xfs_failaddr_t		fa;
2579
2580	trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
2581
2582	bit = 0;
2583	i = 1;  /* 0 is the buf format structure */
2584	while (1) {
2585		bit = xfs_next_bit(buf_f->blf_data_map,
2586				   buf_f->blf_map_size, bit);
2587		if (bit == -1)
2588			break;
2589		nbits = xfs_contig_bits(buf_f->blf_data_map,
2590					buf_f->blf_map_size, bit);
2591		ASSERT(nbits > 0);
2592		ASSERT(item->ri_buf[i].i_addr != NULL);
2593		ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
2594		ASSERT(BBTOB(bp->b_length) >=
2595		       ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
2596
2597		/*
2598		 * The dirty regions logged in the buffer, even though
2599		 * contiguous, may span multiple chunks. This is because the
2600		 * dirty region may span a physical page boundary in a buffer
2601		 * and hence be split into two separate vectors for writing into
2602		 * the log. Hence we need to trim nbits back to the length of
2603		 * the current region being copied out of the log.
2604		 */
2605		if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
2606			nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
2607
2608		/*
2609		 * Do a sanity check if this is a dquot buffer. Just checking
2610		 * the first dquot in the buffer should do. XXXThis is
2611		 * probably a good thing to do for other buf types also.
2612		 */
2613		fa = NULL;
2614		if (buf_f->blf_flags &
2615		   (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2616			if (item->ri_buf[i].i_addr == NULL) {
2617				xfs_alert(mp,
2618					"XFS: NULL dquot in %s.", __func__);
2619				goto next;
2620			}
2621			if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
2622				xfs_alert(mp,
2623					"XFS: dquot too small (%d) in %s.",
2624					item->ri_buf[i].i_len, __func__);
2625				goto next;
2626			}
2627			fa = xfs_dquot_verify(mp, item->ri_buf[i].i_addr,
2628					       -1, 0);
2629			if (fa) {
2630				xfs_alert(mp,
2631	"dquot corrupt at %pS trying to replay into block 0x%llx",
2632					fa, bp->b_bn);
2633				goto next;
2634			}
2635		}
2636
2637		memcpy(xfs_buf_offset(bp,
2638			(uint)bit << XFS_BLF_SHIFT),	/* dest */
2639			item->ri_buf[i].i_addr,		/* source */
2640			nbits<<XFS_BLF_SHIFT);		/* length */
2641 next:
2642		i++;
2643		bit += nbits;
2644	}
2645
2646	/* Shouldn't be any more regions */
2647	ASSERT(i == item->ri_total);
2648
2649	xlog_recover_validate_buf_type(mp, bp, buf_f, current_lsn);
2650}
2651
2652/*
2653 * Perform a dquot buffer recovery.
2654 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
2655 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2656 * Else, treat it as a regular buffer and do recovery.
2657 *
2658 * Return false if the buffer was tossed and true if we recovered the buffer to
2659 * indicate to the caller if the buffer needs writing.
2660 */
2661STATIC bool
2662xlog_recover_do_dquot_buffer(
2663	struct xfs_mount		*mp,
2664	struct xlog			*log,
2665	struct xlog_recover_item	*item,
2666	struct xfs_buf			*bp,
2667	struct xfs_buf_log_format	*buf_f)
2668{
2669	uint			type;
2670
2671	trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2672
2673	/*
2674	 * Filesystems are required to send in quota flags at mount time.
2675	 */
2676	if (!mp->m_qflags)
2677		return false;
2678
2679	type = 0;
2680	if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2681		type |= XFS_DQ_USER;
2682	if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2683		type |= XFS_DQ_PROJ;
2684	if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2685		type |= XFS_DQ_GROUP;
2686	/*
2687	 * This type of quotas was turned off, so ignore this buffer
2688	 */
2689	if (log->l_quotaoffs_flag & type)
2690		return false;
2691
2692	xlog_recover_do_reg_buffer(mp, item, bp, buf_f, NULLCOMMITLSN);
2693	return true;
2694}
2695
2696/*
2697 * This routine replays a modification made to a buffer at runtime.
2698 * There are actually two types of buffer, regular and inode, which
2699 * are handled differently.  Inode buffers are handled differently
2700 * in that we only recover a specific set of data from them, namely
2701 * the inode di_next_unlinked fields.  This is because all other inode
2702 * data is actually logged via inode records and any data we replay
2703 * here which overlaps that may be stale.
2704 *
2705 * When meta-data buffers are freed at run time we log a buffer item
2706 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2707 * of the buffer in the log should not be replayed at recovery time.
2708 * This is so that if the blocks covered by the buffer are reused for
2709 * file data before we crash we don't end up replaying old, freed
2710 * meta-data into a user's file.
2711 *
2712 * To handle the cancellation of buffer log items, we make two passes
2713 * over the log during recovery.  During the first we build a table of
2714 * those buffers which have been cancelled, and during the second we
2715 * only replay those buffers which do not have corresponding cancel
2716 * records in the table.  See xlog_recover_buffer_pass[1,2] above
2717 * for more details on the implementation of the table of cancel records.
2718 */
2719STATIC int
2720xlog_recover_buffer_pass2(
2721	struct xlog			*log,
2722	struct list_head		*buffer_list,
2723	struct xlog_recover_item	*item,
2724	xfs_lsn_t			current_lsn)
2725{
2726	xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
2727	xfs_mount_t		*mp = log->l_mp;
2728	xfs_buf_t		*bp;
2729	int			error;
2730	uint			buf_flags;
2731	xfs_lsn_t		lsn;
2732
2733	/*
2734	 * In this pass we only want to recover all the buffers which have
2735	 * not been cancelled and are not cancellation buffers themselves.
2736	 */
2737	if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2738			buf_f->blf_len, buf_f->blf_flags)) {
2739		trace_xfs_log_recover_buf_cancel(log, buf_f);
2740		return 0;
2741	}
2742
2743	trace_xfs_log_recover_buf_recover(log, buf_f);
2744
2745	buf_flags = 0;
2746	if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
2747		buf_flags |= XBF_UNMAPPED;
2748
2749	bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2750			  buf_flags, NULL);
2751	if (!bp)
2752		return -ENOMEM;
2753	error = bp->b_error;
2754	if (error) {
2755		xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#1)");
2756		goto out_release;
2757	}
2758
2759	/*
2760	 * Recover the buffer only if we get an LSN from it and it's less than
2761	 * the lsn of the transaction we are replaying.
2762	 *
2763	 * Note that we have to be extremely careful of readahead here.
2764	 * Readahead does not attach verfiers to the buffers so if we don't
2765	 * actually do any replay after readahead because of the LSN we found
2766	 * in the buffer if more recent than that current transaction then we
2767	 * need to attach the verifier directly. Failure to do so can lead to
2768	 * future recovery actions (e.g. EFI and unlinked list recovery) can
2769	 * operate on the buffers and they won't get the verifier attached. This
2770	 * can lead to blocks on disk having the correct content but a stale
2771	 * CRC.
2772	 *
2773	 * It is safe to assume these clean buffers are currently up to date.
2774	 * If the buffer is dirtied by a later transaction being replayed, then
2775	 * the verifier will be reset to match whatever recover turns that
2776	 * buffer into.
2777	 */
2778	lsn = xlog_recover_get_buf_lsn(mp, bp);
2779	if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
2780		trace_xfs_log_recover_buf_skip(log, buf_f);
2781		xlog_recover_validate_buf_type(mp, bp, buf_f, NULLCOMMITLSN);
2782		goto out_release;
2783	}
2784
2785	if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2786		error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2787		if (error)
2788			goto out_release;
2789	} else if (buf_f->blf_flags &
2790		  (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2791		bool	dirty;
2792
2793		dirty = xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2794		if (!dirty)
2795			goto out_release;
2796	} else {
2797		xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn);
2798	}
2799
2800	/*
2801	 * Perform delayed write on the buffer.  Asynchronous writes will be
2802	 * slower when taking into account all the buffers to be flushed.
2803	 *
2804	 * Also make sure that only inode buffers with good sizes stay in
2805	 * the buffer cache.  The kernel moves inodes in buffers of 1 block
2806	 * or inode_cluster_size bytes, whichever is bigger.  The inode
2807	 * buffers in the log can be a different size if the log was generated
2808	 * by an older kernel using unclustered inode buffers or a newer kernel
2809	 * running with a different inode cluster size.  Regardless, if the
2810	 * the inode buffer size isn't max(blocksize, inode_cluster_size)
2811	 * for *our* value of inode_cluster_size, then we need to keep
2812	 * the buffer out of the buffer cache so that the buffer won't
2813	 * overlap with future reads of those inodes.
2814	 */
2815	if (XFS_DINODE_MAGIC ==
2816	    be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2817	    (BBTOB(bp->b_length) != M_IGEO(log->l_mp)->inode_cluster_size)) {
2818		xfs_buf_stale(bp);
2819		error = xfs_bwrite(bp);
2820	} else {
2821		ASSERT(bp->b_mount == mp);
2822		bp->b_iodone = xlog_recover_iodone;
2823		xfs_buf_delwri_queue(bp, buffer_list);
2824	}
2825
2826out_release:
2827	xfs_buf_relse(bp);
2828	return error;
2829}
2830
2831/*
2832 * Inode fork owner changes
2833 *
2834 * If we have been told that we have to reparent the inode fork, it's because an
2835 * extent swap operation on a CRC enabled filesystem has been done and we are
2836 * replaying it. We need to walk the BMBT of the appropriate fork and change the
2837 * owners of it.
2838 *
2839 * The complexity here is that we don't have an inode context to work with, so
2840 * after we've replayed the inode we need to instantiate one.  This is where the
2841 * fun begins.
2842 *
2843 * We are in the middle of log recovery, so we can't run transactions. That
2844 * means we cannot use cache coherent inode instantiation via xfs_iget(), as
2845 * that will result in the corresponding iput() running the inode through
2846 * xfs_inactive(). If we've just replayed an inode core that changes the link
2847 * count to zero (i.e. it's been unlinked), then xfs_inactive() will run
2848 * transactions (bad!).
2849 *
2850 * So, to avoid this, we instantiate an inode directly from the inode core we've
2851 * just recovered. We have the buffer still locked, and all we really need to
2852 * instantiate is the inode core and the forks being modified. We can do this
2853 * manually, then run the inode btree owner change, and then tear down the
2854 * xfs_inode without having to run any transactions at all.
2855 *
2856 * Also, because we don't have a transaction context available here but need to
2857 * gather all the buffers we modify for writeback so we pass the buffer_list
2858 * instead for the operation to use.
2859 */
2860
2861STATIC int
2862xfs_recover_inode_owner_change(
2863	struct xfs_mount	*mp,
2864	struct xfs_dinode	*dip,
2865	struct xfs_inode_log_format *in_f,
2866	struct list_head	*buffer_list)
2867{
2868	struct xfs_inode	*ip;
2869	int			error;
2870
2871	ASSERT(in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER));
2872
2873	ip = xfs_inode_alloc(mp, in_f->ilf_ino);
2874	if (!ip)
2875		return -ENOMEM;
2876
2877	/* instantiate the inode */
2878	xfs_inode_from_disk(ip, dip);
2879	ASSERT(ip->i_d.di_version >= 3);
2880
2881	error = xfs_iformat_fork(ip, dip);
2882	if (error)
2883		goto out_free_ip;
2884
2885	if (!xfs_inode_verify_forks(ip)) {
2886		error = -EFSCORRUPTED;
2887		goto out_free_ip;
2888	}
2889
2890	if (in_f->ilf_fields & XFS_ILOG_DOWNER) {
2891		ASSERT(in_f->ilf_fields & XFS_ILOG_DBROOT);
2892		error = xfs_bmbt_change_owner(NULL, ip, XFS_DATA_FORK,
2893					      ip->i_ino, buffer_list);
2894		if (error)
2895			goto out_free_ip;
2896	}
2897
2898	if (in_f->ilf_fields & XFS_ILOG_AOWNER) {
2899		ASSERT(in_f->ilf_fields & XFS_ILOG_ABROOT);
2900		error = xfs_bmbt_change_owner(NULL, ip, XFS_ATTR_FORK,
2901					      ip->i_ino, buffer_list);
2902		if (error)
2903			goto out_free_ip;
2904	}
2905
2906out_free_ip:
2907	xfs_inode_free(ip);
2908	return error;
2909}
2910
2911STATIC int
2912xlog_recover_inode_pass2(
2913	struct xlog			*log,
2914	struct list_head		*buffer_list,
2915	struct xlog_recover_item	*item,
2916	xfs_lsn_t			current_lsn)
2917{
2918	struct xfs_inode_log_format	*in_f;
2919	xfs_mount_t		*mp = log->l_mp;
2920	xfs_buf_t		*bp;
2921	xfs_dinode_t		*dip;
2922	int			len;
2923	char			*src;
2924	char			*dest;
2925	int			error;
2926	int			attr_index;
2927	uint			fields;
2928	struct xfs_log_dinode	*ldip;
2929	uint			isize;
2930	int			need_free = 0;
2931
2932	if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
2933		in_f = item->ri_buf[0].i_addr;
2934	} else {
2935		in_f = kmem_alloc(sizeof(struct xfs_inode_log_format), 0);
2936		need_free = 1;
2937		error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2938		if (error)
2939			goto error;
2940	}
2941
2942	/*
2943	 * Inode buffers can be freed, look out for it,
2944	 * and do not replay the inode.
2945	 */
2946	if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2947					in_f->ilf_len, 0)) {
2948		error = 0;
2949		trace_xfs_log_recover_inode_cancel(log, in_f);
2950		goto error;
2951	}
2952	trace_xfs_log_recover_inode_recover(log, in_f);
2953
2954	bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len, 0,
2955			  &xfs_inode_buf_ops);
2956	if (!bp) {
2957		error = -ENOMEM;
2958		goto error;
2959	}
2960	error = bp->b_error;
2961	if (error) {
2962		xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#2)");
2963		goto out_release;
2964	}
2965	ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2966	dip = xfs_buf_offset(bp, in_f->ilf_boffset);
2967
2968	/*
2969	 * Make sure the place we're flushing out to really looks
2970	 * like an inode!
2971	 */
2972	if (unlikely(!xfs_verify_magic16(bp, dip->di_magic))) {
2973		xfs_alert(mp,
2974	"%s: Bad inode magic number, dip = "PTR_FMT", dino bp = "PTR_FMT", ino = %Ld",
2975			__func__, dip, bp, in_f->ilf_ino);
2976		XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2977				 XFS_ERRLEVEL_LOW, mp);
2978		error = -EFSCORRUPTED;
2979		goto out_release;
2980	}
2981	ldip = item->ri_buf[1].i_addr;
2982	if (unlikely(ldip->di_magic != XFS_DINODE_MAGIC)) {
2983		xfs_alert(mp,
2984			"%s: Bad inode log record, rec ptr "PTR_FMT", ino %Ld",
2985			__func__, item, in_f->ilf_ino);
2986		XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2987				 XFS_ERRLEVEL_LOW, mp);
2988		error = -EFSCORRUPTED;
2989		goto out_release;
2990	}
2991
2992	/*
2993	 * If the inode has an LSN in it, recover the inode only if it's less
2994	 * than the lsn of the transaction we are replaying. Note: we still
2995	 * need to replay an owner change even though the inode is more recent
2996	 * than the transaction as there is no guarantee that all the btree
2997	 * blocks are more recent than this transaction, too.
2998	 */
2999	if (dip->di_version >= 3) {
3000		xfs_lsn_t	lsn = be64_to_cpu(dip->di_lsn);
3001
3002		if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3003			trace_xfs_log_recover_inode_skip(log, in_f);
3004			error = 0;
3005			goto out_owner_change;
3006		}
3007	}
3008
3009	/*
3010	 * di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
3011	 * are transactional and if ordering is necessary we can determine that
3012	 * more accurately by the LSN field in the V3 inode core. Don't trust
3013	 * the inode versions we might be changing them here - use the
3014	 * superblock flag to determine whether we need to look at di_flushiter
3015	 * to skip replay when the on disk inode is newer than the log one
3016	 */
3017	if (!xfs_sb_version_hascrc(&mp->m_sb) &&
3018	    ldip->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
3019		/*
3020		 * Deal with the wrap case, DI_MAX_FLUSH is less
3021		 * than smaller numbers
3022		 */
3023		if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
3024		    ldip->di_flushiter < (DI_MAX_FLUSH >> 1)) {
3025			/* do nothing */
3026		} else {
3027			trace_xfs_log_recover_inode_skip(log, in_f);
3028			error = 0;
3029			goto out_release;
3030		}
3031	}
3032
3033	/* Take the opportunity to reset the flush iteration count */
3034	ldip->di_flushiter = 0;
3035
3036	if (unlikely(S_ISREG(ldip->di_mode))) {
3037		if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3038		    (ldip->di_format != XFS_DINODE_FMT_BTREE)) {
3039			XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
3040					 XFS_ERRLEVEL_LOW, mp, ldip,
3041					 sizeof(*ldip));
3042			xfs_alert(mp,
3043		"%s: Bad regular inode log record, rec ptr "PTR_FMT", "
3044		"ino ptr = "PTR_FMT", ino bp = "PTR_FMT", ino %Ld",
3045				__func__, item, dip, bp, in_f->ilf_ino);
3046			error = -EFSCORRUPTED;
3047			goto out_release;
3048		}
3049	} else if (unlikely(S_ISDIR(ldip->di_mode))) {
3050		if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3051		    (ldip->di_format != XFS_DINODE_FMT_BTREE) &&
3052		    (ldip->di_format != XFS_DINODE_FMT_LOCAL)) {
3053			XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
3054					     XFS_ERRLEVEL_LOW, mp, ldip,
3055					     sizeof(*ldip));
3056			xfs_alert(mp,
3057		"%s: Bad dir inode log record, rec ptr "PTR_FMT", "
3058		"ino ptr = "PTR_FMT", ino bp = "PTR_FMT", ino %Ld",
3059				__func__, item, dip, bp, in_f->ilf_ino);
3060			error = -EFSCORRUPTED;
3061			goto out_release;
3062		}
3063	}
3064	if (unlikely(ldip->di_nextents + ldip->di_anextents > ldip->di_nblocks)){
3065		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
3066				     XFS_ERRLEVEL_LOW, mp, ldip,
3067				     sizeof(*ldip));
3068		xfs_alert(mp,
3069	"%s: Bad inode log record, rec ptr "PTR_FMT", dino ptr "PTR_FMT", "
3070	"dino bp "PTR_FMT", ino %Ld, total extents = %d, nblocks = %Ld",
3071			__func__, item, dip, bp, in_f->ilf_ino,
3072			ldip->di_nextents + ldip->di_anextents,
3073			ldip->di_nblocks);
3074		error = -EFSCORRUPTED;
3075		goto out_release;
3076	}
3077	if (unlikely(ldip->di_forkoff > mp->m_sb.sb_inodesize)) {
3078		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
3079				     XFS_ERRLEVEL_LOW, mp, ldip,
3080				     sizeof(*ldip));
3081		xfs_alert(mp,
3082	"%s: Bad inode log record, rec ptr "PTR_FMT", dino ptr "PTR_FMT", "
3083	"dino bp "PTR_FMT", ino %Ld, forkoff 0x%x", __func__,
3084			item, dip, bp, in_f->ilf_ino, ldip->di_forkoff);
3085		error = -EFSCORRUPTED;
3086		goto out_release;
3087	}
3088	isize = xfs_log_dinode_size(ldip->di_version);
3089	if (unlikely(item->ri_buf[1].i_len > isize)) {
3090		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
3091				     XFS_ERRLEVEL_LOW, mp, ldip,
3092				     sizeof(*ldip));
3093		xfs_alert(mp,
3094			"%s: Bad inode log record length %d, rec ptr "PTR_FMT,
3095			__func__, item->ri_buf[1].i_len, item);
3096		error = -EFSCORRUPTED;
3097		goto out_release;
3098	}
3099
3100	/* recover the log dinode inode into the on disk inode */
3101	xfs_log_dinode_to_disk(ldip, dip);
3102
3103	fields = in_f->ilf_fields;
3104	if (fields & XFS_ILOG_DEV)
3105		xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
3106
3107	if (in_f->ilf_size == 2)
3108		goto out_owner_change;
3109	len = item->ri_buf[2].i_len;
3110	src = item->ri_buf[2].i_addr;
3111	ASSERT(in_f->ilf_size <= 4);
3112	ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
3113	ASSERT(!(fields & XFS_ILOG_DFORK) ||
3114	       (len == in_f->ilf_dsize));
3115
3116	switch (fields & XFS_ILOG_DFORK) {
3117	case XFS_ILOG_DDATA:
3118	case XFS_ILOG_DEXT:
3119		memcpy(XFS_DFORK_DPTR(dip), src, len);
3120		break;
3121
3122	case XFS_ILOG_DBROOT:
3123		xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
3124				 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
3125				 XFS_DFORK_DSIZE(dip, mp));
3126		break;
3127
3128	default:
3129		/*
3130		 * There are no data fork flags set.
3131		 */
3132		ASSERT((fields & XFS_ILOG_DFORK) == 0);
3133		break;
3134	}
3135
3136	/*
3137	 * If we logged any attribute data, recover it.  There may or
3138	 * may not have been any other non-core data logged in this
3139	 * transaction.
3140	 */
3141	if (in_f->ilf_fields & XFS_ILOG_AFORK) {
3142		if (in_f->ilf_fields & XFS_ILOG_DFORK) {
3143			attr_index = 3;
3144		} else {
3145			attr_index = 2;
3146		}
3147		len = item->ri_buf[attr_index].i_len;
3148		src = item->ri_buf[attr_index].i_addr;
3149		ASSERT(len == in_f->ilf_asize);
3150
3151		switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
3152		case XFS_ILOG_ADATA:
3153		case XFS_ILOG_AEXT:
3154			dest = XFS_DFORK_APTR(dip);
3155			ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
3156			memcpy(dest, src, len);
3157			break;
3158
3159		case XFS_ILOG_ABROOT:
3160			dest = XFS_DFORK_APTR(dip);
3161			xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
3162					 len, (xfs_bmdr_block_t*)dest,
3163					 XFS_DFORK_ASIZE(dip, mp));
3164			break;
3165
3166		default:
3167			xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
3168			ASSERT(0);
3169			error = -EIO;
3170			goto out_release;
3171		}
3172	}
3173
3174out_owner_change:
3175	/* Recover the swapext owner change unless inode has been deleted */
3176	if ((in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER)) &&
3177	    (dip->di_mode != 0))
3178		error = xfs_recover_inode_owner_change(mp, dip, in_f,
3179						       buffer_list);
3180	/* re-generate the checksum. */
3181	xfs_dinode_calc_crc(log->l_mp, dip);
3182
3183	ASSERT(bp->b_mount == mp);
3184	bp->b_iodone = xlog_recover_iodone;
3185	xfs_buf_delwri_queue(bp, buffer_list);
3186
3187out_release:
3188	xfs_buf_relse(bp);
3189error:
3190	if (need_free)
3191		kmem_free(in_f);
3192	return error;
3193}
3194
3195/*
3196 * Recover QUOTAOFF records. We simply make a note of it in the xlog
3197 * structure, so that we know not to do any dquot item or dquot buffer recovery,
3198 * of that type.
3199 */
3200STATIC int
3201xlog_recover_quotaoff_pass1(
3202	struct xlog			*log,
3203	struct xlog_recover_item	*item)
3204{
3205	xfs_qoff_logformat_t	*qoff_f = item->ri_buf[0].i_addr;
3206	ASSERT(qoff_f);
3207
3208	/*
3209	 * The logitem format's flag tells us if this was user quotaoff,
3210	 * group/project quotaoff or both.
3211	 */
3212	if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
3213		log->l_quotaoffs_flag |= XFS_DQ_USER;
3214	if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
3215		log->l_quotaoffs_flag |= XFS_DQ_PROJ;
3216	if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
3217		log->l_quotaoffs_flag |= XFS_DQ_GROUP;
3218
3219	return 0;
3220}
3221
3222/*
3223 * Recover a dquot record
3224 */
3225STATIC int
3226xlog_recover_dquot_pass2(
3227	struct xlog			*log,
3228	struct list_head		*buffer_list,
3229	struct xlog_recover_item	*item,
3230	xfs_lsn_t			current_lsn)
3231{
3232	xfs_mount_t		*mp = log->l_mp;
3233	xfs_buf_t		*bp;
3234	struct xfs_disk_dquot	*ddq, *recddq;
3235	xfs_failaddr_t		fa;
3236	int			error;
3237	xfs_dq_logformat_t	*dq_f;
3238	uint			type;
3239
3240
3241	/*
3242	 * Filesystems are required to send in quota flags at mount time.
3243	 */
3244	if (mp->m_qflags == 0)
3245		return 0;
3246
3247	recddq = item->ri_buf[1].i_addr;
3248	if (recddq == NULL) {
3249		xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
3250		return -EIO;
3251	}
3252	if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
3253		xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
3254			item->ri_buf[1].i_len, __func__);
3255		return -EIO;
3256	}
3257
3258	/*
3259	 * This type of quotas was turned off, so ignore this record.
3260	 */
3261	type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3262	ASSERT(type);
3263	if (log->l_quotaoffs_flag & type)
3264		return 0;
3265
3266	/*
3267	 * At this point we know that quota was _not_ turned off.
3268	 * Since the mount flags are not indicating to us otherwise, this
3269	 * must mean that quota is on, and the dquot needs to be replayed.
3270	 * Remember that we may not have fully recovered the superblock yet,
3271	 * so we can't do the usual trick of looking at the SB quota bits.
3272	 *
3273	 * The other possibility, of course, is that the quota subsystem was
3274	 * removed since the last mount - ENOSYS.
3275	 */
3276	dq_f = item->ri_buf[0].i_addr;
3277	ASSERT(dq_f);
3278	fa = xfs_dquot_verify(mp, recddq, dq_f->qlf_id, 0);
3279	if (fa) {
3280		xfs_alert(mp, "corrupt dquot ID 0x%x in log at %pS",
3281				dq_f->qlf_id, fa);
3282		return -EIO;
3283	}
3284	ASSERT(dq_f->qlf_len == 1);
3285
3286	/*
3287	 * At this point we are assuming that the dquots have been allocated
3288	 * and hence the buffer has valid dquots stamped in it. It should,
3289	 * therefore, pass verifier validation. If the dquot is bad, then the
3290	 * we'll return an error here, so we don't need to specifically check
3291	 * the dquot in the buffer after the verifier has run.
3292	 */
3293	error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dq_f->qlf_blkno,
3294				   XFS_FSB_TO_BB(mp, dq_f->qlf_len), 0, &bp,
3295				   &xfs_dquot_buf_ops);
3296	if (error)
3297		return error;
3298
3299	ASSERT(bp);
3300	ddq = xfs_buf_offset(bp, dq_f->qlf_boffset);
3301
3302	/*
3303	 * If the dquot has an LSN in it, recover the dquot only if it's less
3304	 * than the lsn of the transaction we are replaying.
3305	 */
3306	if (xfs_sb_version_hascrc(&mp->m_sb)) {
3307		struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddq;
3308		xfs_lsn_t	lsn = be64_to_cpu(dqb->dd_lsn);
3309
3310		if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3311			goto out_release;
3312		}
3313	}
3314
3315	memcpy(ddq, recddq, item->ri_buf[1].i_len);
3316	if (xfs_sb_version_hascrc(&mp->m_sb)) {
3317		xfs_update_cksum((char *)ddq, sizeof(struct xfs_dqblk),
3318				 XFS_DQUOT_CRC_OFF);
3319	}
3320
3321	ASSERT(dq_f->qlf_size == 2);
3322	ASSERT(bp->b_mount == mp);
3323	bp->b_iodone = xlog_recover_iodone;
3324	xfs_buf_delwri_queue(bp, buffer_list);
3325
3326out_release:
3327	xfs_buf_relse(bp);
3328	return 0;
3329}
3330
3331/*
3332 * This routine is called to create an in-core extent free intent
3333 * item from the efi format structure which was logged on disk.
3334 * It allocates an in-core efi, copies the extents from the format
3335 * structure into it, and adds the efi to the AIL with the given
3336 * LSN.
3337 */
3338STATIC int
3339xlog_recover_efi_pass2(
3340	struct xlog			*log,
3341	struct xlog_recover_item	*item,
3342	xfs_lsn_t			lsn)
3343{
3344	int				error;
3345	struct xfs_mount		*mp = log->l_mp;
3346	struct xfs_efi_log_item		*efip;
3347	struct xfs_efi_log_format	*efi_formatp;
3348
3349	efi_formatp = item->ri_buf[0].i_addr;
3350
3351	efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
3352	error = xfs_efi_copy_format(&item->ri_buf[0], &efip->efi_format);
3353	if (error) {
3354		xfs_efi_item_free(efip);
3355		return error;
3356	}
3357	atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
3358
3359	spin_lock(&log->l_ailp->ail_lock);
3360	/*
3361	 * The EFI has two references. One for the EFD and one for EFI to ensure
3362	 * it makes it into the AIL. Insert the EFI into the AIL directly and
3363	 * drop the EFI reference. Note that xfs_trans_ail_update() drops the
3364	 * AIL lock.
3365	 */
3366	xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
3367	xfs_efi_release(efip);
3368	return 0;
3369}
3370
3371
3372/*
3373 * This routine is called when an EFD format structure is found in a committed
3374 * transaction in the log. Its purpose is to cancel the corresponding EFI if it
3375 * was still in the log. To do this it searches the AIL for the EFI with an id
3376 * equal to that in the EFD format structure. If we find it we drop the EFD
3377 * reference, which removes the EFI from the AIL and frees it.
3378 */
3379STATIC int
3380xlog_recover_efd_pass2(
3381	struct xlog			*log,
3382	struct xlog_recover_item	*item)
3383{
3384	xfs_efd_log_format_t	*efd_formatp;
3385	xfs_efi_log_item_t	*efip = NULL;
3386	struct xfs_log_item	*lip;
3387	uint64_t		efi_id;
3388	struct xfs_ail_cursor	cur;
3389	struct xfs_ail		*ailp = log->l_ailp;
3390
3391	efd_formatp = item->ri_buf[0].i_addr;
3392	ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
3393		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
3394	       (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
3395		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
3396	efi_id = efd_formatp->efd_efi_id;
3397
3398	/*
3399	 * Search for the EFI with the id in the EFD format structure in the
3400	 * AIL.
3401	 */
3402	spin_lock(&ailp->ail_lock);
3403	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3404	while (lip != NULL) {
3405		if (lip->li_type == XFS_LI_EFI) {
3406			efip = (xfs_efi_log_item_t *)lip;
3407			if (efip->efi_format.efi_id == efi_id) {
3408				/*
3409				 * Drop the EFD reference to the EFI. This
3410				 * removes the EFI from the AIL and frees it.
3411				 */
3412				spin_unlock(&ailp->ail_lock);
3413				xfs_efi_release(efip);
3414				spin_lock(&ailp->ail_lock);
3415				break;
3416			}
3417		}
3418		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3419	}
3420
3421	xfs_trans_ail_cursor_done(&cur);
3422	spin_unlock(&ailp->ail_lock);
3423
3424	return 0;
3425}
3426
3427/*
3428 * This routine is called to create an in-core extent rmap update
3429 * item from the rui format structure which was logged on disk.
3430 * It allocates an in-core rui, copies the extents from the format
3431 * structure into it, and adds the rui to the AIL with the given
3432 * LSN.
3433 */
3434STATIC int
3435xlog_recover_rui_pass2(
3436	struct xlog			*log,
3437	struct xlog_recover_item	*item,
3438	xfs_lsn_t			lsn)
3439{
3440	int				error;
3441	struct xfs_mount		*mp = log->l_mp;
3442	struct xfs_rui_log_item		*ruip;
3443	struct xfs_rui_log_format	*rui_formatp;
3444
3445	rui_formatp = item->ri_buf[0].i_addr;
3446
3447	ruip = xfs_rui_init(mp, rui_formatp->rui_nextents);
3448	error = xfs_rui_copy_format(&item->ri_buf[0], &ruip->rui_format);
3449	if (error) {
3450		xfs_rui_item_free(ruip);
3451		return error;
3452	}
3453	atomic_set(&ruip->rui_next_extent, rui_formatp->rui_nextents);
3454
3455	spin_lock(&log->l_ailp->ail_lock);
3456	/*
3457	 * The RUI has two references. One for the RUD and one for RUI to ensure
3458	 * it makes it into the AIL. Insert the RUI into the AIL directly and
3459	 * drop the RUI reference. Note that xfs_trans_ail_update() drops the
3460	 * AIL lock.
3461	 */
3462	xfs_trans_ail_update(log->l_ailp, &ruip->rui_item, lsn);
3463	xfs_rui_release(ruip);
3464	return 0;
3465}
3466
3467
3468/*
3469 * This routine is called when an RUD format structure is found in a committed
3470 * transaction in the log. Its purpose is to cancel the corresponding RUI if it
3471 * was still in the log. To do this it searches the AIL for the RUI with an id
3472 * equal to that in the RUD format structure. If we find it we drop the RUD
3473 * reference, which removes the RUI from the AIL and frees it.
3474 */
3475STATIC int
3476xlog_recover_rud_pass2(
3477	struct xlog			*log,
3478	struct xlog_recover_item	*item)
3479{
3480	struct xfs_rud_log_format	*rud_formatp;
3481	struct xfs_rui_log_item		*ruip = NULL;
3482	struct xfs_log_item		*lip;
3483	uint64_t			rui_id;
3484	struct xfs_ail_cursor		cur;
3485	struct xfs_ail			*ailp = log->l_ailp;
3486
3487	rud_formatp = item->ri_buf[0].i_addr;
3488	ASSERT(item->ri_buf[0].i_len == sizeof(struct xfs_rud_log_format));
3489	rui_id = rud_formatp->rud_rui_id;
3490
3491	/*
3492	 * Search for the RUI with the id in the RUD format structure in the
3493	 * AIL.
3494	 */
3495	spin_lock(&ailp->ail_lock);
3496	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3497	while (lip != NULL) {
3498		if (lip->li_type == XFS_LI_RUI) {
3499			ruip = (struct xfs_rui_log_item *)lip;
3500			if (ruip->rui_format.rui_id == rui_id) {
3501				/*
3502				 * Drop the RUD reference to the RUI. This
3503				 * removes the RUI from the AIL and frees it.
3504				 */
3505				spin_unlock(&ailp->ail_lock);
3506				xfs_rui_release(ruip);
3507				spin_lock(&ailp->ail_lock);
3508				break;
3509			}
3510		}
3511		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3512	}
3513
3514	xfs_trans_ail_cursor_done(&cur);
3515	spin_unlock(&ailp->ail_lock);
3516
3517	return 0;
3518}
3519
3520/*
3521 * Copy an CUI format buffer from the given buf, and into the destination
3522 * CUI format structure.  The CUI/CUD items were designed not to need any
3523 * special alignment handling.
3524 */
3525static int
3526xfs_cui_copy_format(
3527	struct xfs_log_iovec		*buf,
3528	struct xfs_cui_log_format	*dst_cui_fmt)
3529{
3530	struct xfs_cui_log_format	*src_cui_fmt;
3531	uint				len;
3532
3533	src_cui_fmt = buf->i_addr;
3534	len = xfs_cui_log_format_sizeof(src_cui_fmt->cui_nextents);
3535
3536	if (buf->i_len == len) {
3537		memcpy(dst_cui_fmt, src_cui_fmt, len);
3538		return 0;
3539	}
3540	return -EFSCORRUPTED;
3541}
3542
3543/*
3544 * This routine is called to create an in-core extent refcount update
3545 * item from the cui format structure which was logged on disk.
3546 * It allocates an in-core cui, copies the extents from the format
3547 * structure into it, and adds the cui to the AIL with the given
3548 * LSN.
3549 */
3550STATIC int
3551xlog_recover_cui_pass2(
3552	struct xlog			*log,
3553	struct xlog_recover_item	*item,
3554	xfs_lsn_t			lsn)
3555{
3556	int				error;
3557	struct xfs_mount		*mp = log->l_mp;
3558	struct xfs_cui_log_item		*cuip;
3559	struct xfs_cui_log_format	*cui_formatp;
3560
3561	cui_formatp = item->ri_buf[0].i_addr;
3562
3563	cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
3564	error = xfs_cui_copy_format(&item->ri_buf[0], &cuip->cui_format);
3565	if (error) {
3566		xfs_cui_item_free(cuip);
3567		return error;
3568	}
3569	atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
3570
3571	spin_lock(&log->l_ailp->ail_lock);
3572	/*
3573	 * The CUI has two references. One for the CUD and one for CUI to ensure
3574	 * it makes it into the AIL. Insert the CUI into the AIL directly and
3575	 * drop the CUI reference. Note that xfs_trans_ail_update() drops the
3576	 * AIL lock.
3577	 */
3578	xfs_trans_ail_update(log->l_ailp, &cuip->cui_item, lsn);
3579	xfs_cui_release(cuip);
3580	return 0;
3581}
3582
3583
3584/*
3585 * This routine is called when an CUD format structure is found in a committed
3586 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
3587 * was still in the log. To do this it searches the AIL for the CUI with an id
3588 * equal to that in the CUD format structure. If we find it we drop the CUD
3589 * reference, which removes the CUI from the AIL and frees it.
3590 */
3591STATIC int
3592xlog_recover_cud_pass2(
3593	struct xlog			*log,
3594	struct xlog_recover_item	*item)
3595{
3596	struct xfs_cud_log_format	*cud_formatp;
3597	struct xfs_cui_log_item		*cuip = NULL;
3598	struct xfs_log_item		*lip;
3599	uint64_t			cui_id;
3600	struct xfs_ail_cursor		cur;
3601	struct xfs_ail			*ailp = log->l_ailp;
3602
3603	cud_formatp = item->ri_buf[0].i_addr;
3604	if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format))
3605		return -EFSCORRUPTED;
3606	cui_id = cud_formatp->cud_cui_id;
3607
3608	/*
3609	 * Search for the CUI with the id in the CUD format structure in the
3610	 * AIL.
3611	 */
3612	spin_lock(&ailp->ail_lock);
3613	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3614	while (lip != NULL) {
3615		if (lip->li_type == XFS_LI_CUI) {
3616			cuip = (struct xfs_cui_log_item *)lip;
3617			if (cuip->cui_format.cui_id == cui_id) {
3618				/*
3619				 * Drop the CUD reference to the CUI. This
3620				 * removes the CUI from the AIL and frees it.
3621				 */
3622				spin_unlock(&ailp->ail_lock);
3623				xfs_cui_release(cuip);
3624				spin_lock(&ailp->ail_lock);
3625				break;
3626			}
3627		}
3628		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3629	}
3630
3631	xfs_trans_ail_cursor_done(&cur);
3632	spin_unlock(&ailp->ail_lock);
3633
3634	return 0;
3635}
3636
3637/*
3638 * Copy an BUI format buffer from the given buf, and into the destination
3639 * BUI format structure.  The BUI/BUD items were designed not to need any
3640 * special alignment handling.
3641 */
3642static int
3643xfs_bui_copy_format(
3644	struct xfs_log_iovec		*buf,
3645	struct xfs_bui_log_format	*dst_bui_fmt)
3646{
3647	struct xfs_bui_log_format	*src_bui_fmt;
3648	uint				len;
3649
3650	src_bui_fmt = buf->i_addr;
3651	len = xfs_bui_log_format_sizeof(src_bui_fmt->bui_nextents);
3652
3653	if (buf->i_len == len) {
3654		memcpy(dst_bui_fmt, src_bui_fmt, len);
3655		return 0;
3656	}
3657	return -EFSCORRUPTED;
3658}
3659
3660/*
3661 * This routine is called to create an in-core extent bmap update
3662 * item from the bui format structure which was logged on disk.
3663 * It allocates an in-core bui, copies the extents from the format
3664 * structure into it, and adds the bui to the AIL with the given
3665 * LSN.
3666 */
3667STATIC int
3668xlog_recover_bui_pass2(
3669	struct xlog			*log,
3670	struct xlog_recover_item	*item,
3671	xfs_lsn_t			lsn)
3672{
3673	int				error;
3674	struct xfs_mount		*mp = log->l_mp;
3675	struct xfs_bui_log_item		*buip;
3676	struct xfs_bui_log_format	*bui_formatp;
3677
3678	bui_formatp = item->ri_buf[0].i_addr;
3679
3680	if (bui_formatp->bui_nextents != XFS_BUI_MAX_FAST_EXTENTS)
3681		return -EFSCORRUPTED;
3682	buip = xfs_bui_init(mp);
3683	error = xfs_bui_copy_format(&item->ri_buf[0], &buip->bui_format);
3684	if (error) {
3685		xfs_bui_item_free(buip);
3686		return error;
3687	}
3688	atomic_set(&buip->bui_next_extent, bui_formatp->bui_nextents);
3689
3690	spin_lock(&log->l_ailp->ail_lock);
3691	/*
3692	 * The RUI has two references. One for the RUD and one for RUI to ensure
3693	 * it makes it into the AIL. Insert the RUI into the AIL directly and
3694	 * drop the RUI reference. Note that xfs_trans_ail_update() drops the
3695	 * AIL lock.
3696	 */
3697	xfs_trans_ail_update(log->l_ailp, &buip->bui_item, lsn);
3698	xfs_bui_release(buip);
3699	return 0;
3700}
3701
3702
3703/*
3704 * This routine is called when an BUD format structure is found in a committed
3705 * transaction in the log. Its purpose is to cancel the corresponding BUI if it
3706 * was still in the log. To do this it searches the AIL for the BUI with an id
3707 * equal to that in the BUD format structure. If we find it we drop the BUD
3708 * reference, which removes the BUI from the AIL and frees it.
3709 */
3710STATIC int
3711xlog_recover_bud_pass2(
3712	struct xlog			*log,
3713	struct xlog_recover_item	*item)
3714{
3715	struct xfs_bud_log_format	*bud_formatp;
3716	struct xfs_bui_log_item		*buip = NULL;
3717	struct xfs_log_item		*lip;
3718	uint64_t			bui_id;
3719	struct xfs_ail_cursor		cur;
3720	struct xfs_ail			*ailp = log->l_ailp;
3721
3722	bud_formatp = item->ri_buf[0].i_addr;
3723	if (item->ri_buf[0].i_len != sizeof(struct xfs_bud_log_format))
3724		return -EFSCORRUPTED;
3725	bui_id = bud_formatp->bud_bui_id;
3726
3727	/*
3728	 * Search for the BUI with the id in the BUD format structure in the
3729	 * AIL.
3730	 */
3731	spin_lock(&ailp->ail_lock);
3732	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3733	while (lip != NULL) {
3734		if (lip->li_type == XFS_LI_BUI) {
3735			buip = (struct xfs_bui_log_item *)lip;
3736			if (buip->bui_format.bui_id == bui_id) {
3737				/*
3738				 * Drop the BUD reference to the BUI. This
3739				 * removes the BUI from the AIL and frees it.
3740				 */
3741				spin_unlock(&ailp->ail_lock);
3742				xfs_bui_release(buip);
3743				spin_lock(&ailp->ail_lock);
3744				break;
3745			}
3746		}
3747		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3748	}
3749
3750	xfs_trans_ail_cursor_done(&cur);
3751	spin_unlock(&ailp->ail_lock);
3752
3753	return 0;
3754}
3755
3756/*
3757 * This routine is called when an inode create format structure is found in a
3758 * committed transaction in the log.  It's purpose is to initialise the inodes
3759 * being allocated on disk. This requires us to get inode cluster buffers that
3760 * match the range to be initialised, stamped with inode templates and written
3761 * by delayed write so that subsequent modifications will hit the cached buffer
3762 * and only need writing out at the end of recovery.
3763 */
3764STATIC int
3765xlog_recover_do_icreate_pass2(
3766	struct xlog		*log,
3767	struct list_head	*buffer_list,
3768	xlog_recover_item_t	*item)
3769{
3770	struct xfs_mount	*mp = log->l_mp;
3771	struct xfs_icreate_log	*icl;
3772	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
3773	xfs_agnumber_t		agno;
3774	xfs_agblock_t		agbno;
3775	unsigned int		count;
3776	unsigned int		isize;
3777	xfs_agblock_t		length;
3778	int			bb_per_cluster;
3779	int			cancel_count;
3780	int			nbufs;
3781	int			i;
3782
3783	icl = (struct xfs_icreate_log *)item->ri_buf[0].i_addr;
3784	if (icl->icl_type != XFS_LI_ICREATE) {
3785		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad type");
3786		return -EINVAL;
3787	}
3788
3789	if (icl->icl_size != 1) {
3790		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad icl size");
3791		return -EINVAL;
3792	}
3793
3794	agno = be32_to_cpu(icl->icl_ag);
3795	if (agno >= mp->m_sb.sb_agcount) {
3796		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agno");
3797		return -EINVAL;
3798	}
3799	agbno = be32_to_cpu(icl->icl_agbno);
3800	if (!agbno || agbno == NULLAGBLOCK || agbno >= mp->m_sb.sb_agblocks) {
3801		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agbno");
3802		return -EINVAL;
3803	}
3804	isize = be32_to_cpu(icl->icl_isize);
3805	if (isize != mp->m_sb.sb_inodesize) {
3806		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad isize");
3807		return -EINVAL;
3808	}
3809	count = be32_to_cpu(icl->icl_count);
3810	if (!count) {
3811		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count");
3812		return -EINVAL;
3813	}
3814	length = be32_to_cpu(icl->icl_length);
3815	if (!length || length >= mp->m_sb.sb_agblocks) {
3816		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad length");
3817		return -EINVAL;
3818	}
3819
3820	/*
3821	 * The inode chunk is either full or sparse and we only support
3822	 * m_ino_geo.ialloc_min_blks sized sparse allocations at this time.
3823	 */
3824	if (length != igeo->ialloc_blks &&
3825	    length != igeo->ialloc_min_blks) {
3826		xfs_warn(log->l_mp,
3827			 "%s: unsupported chunk length", __FUNCTION__);
3828		return -EINVAL;
3829	}
3830
3831	/* verify inode count is consistent with extent length */
3832	if ((count >> mp->m_sb.sb_inopblog) != length) {
3833		xfs_warn(log->l_mp,
3834			 "%s: inconsistent inode count and chunk length",
3835			 __FUNCTION__);
3836		return -EINVAL;
3837	}
3838
3839	/*
3840	 * The icreate transaction can cover multiple cluster buffers and these
3841	 * buffers could have been freed and reused. Check the individual
3842	 * buffers for cancellation so we don't overwrite anything written after
3843	 * a cancellation.
3844	 */
3845	bb_per_cluster = XFS_FSB_TO_BB(mp, igeo->blocks_per_cluster);
3846	nbufs = length / igeo->blocks_per_cluster;
3847	for (i = 0, cancel_count = 0; i < nbufs; i++) {
3848		xfs_daddr_t	daddr;
3849
3850		daddr = XFS_AGB_TO_DADDR(mp, agno,
3851				agbno + i * igeo->blocks_per_cluster);
3852		if (xlog_check_buffer_cancelled(log, daddr, bb_per_cluster, 0))
3853			cancel_count++;
3854	}
3855
3856	/*
3857	 * We currently only use icreate for a single allocation at a time. This
3858	 * means we should expect either all or none of the buffers to be
3859	 * cancelled. Be conservative and skip replay if at least one buffer is
3860	 * cancelled, but warn the user that something is awry if the buffers
3861	 * are not consistent.
3862	 *
3863	 * XXX: This must be refined to only skip cancelled clusters once we use
3864	 * icreate for multiple chunk allocations.
3865	 */
3866	ASSERT(!cancel_count || cancel_count == nbufs);
3867	if (cancel_count) {
3868		if (cancel_count != nbufs)
3869			xfs_warn(mp,
3870	"WARNING: partial inode chunk cancellation, skipped icreate.");
3871		trace_xfs_log_recover_icreate_cancel(log, icl);
3872		return 0;
3873	}
3874
3875	trace_xfs_log_recover_icreate_recover(log, icl);
3876	return xfs_ialloc_inode_init(mp, NULL, buffer_list, count, agno, agbno,
3877				     length, be32_to_cpu(icl->icl_gen));
3878}
3879
3880STATIC void
3881xlog_recover_buffer_ra_pass2(
3882	struct xlog                     *log,
3883	struct xlog_recover_item        *item)
3884{
3885	struct xfs_buf_log_format	*buf_f = item->ri_buf[0].i_addr;
3886	struct xfs_mount		*mp = log->l_mp;
3887
3888	if (xlog_peek_buffer_cancelled(log, buf_f->blf_blkno,
3889			buf_f->blf_len, buf_f->blf_flags)) {
3890		return;
3891	}
3892
3893	xfs_buf_readahead(mp->m_ddev_targp, buf_f->blf_blkno,
3894				buf_f->blf_len, NULL);
3895}
3896
3897STATIC void
3898xlog_recover_inode_ra_pass2(
3899	struct xlog                     *log,
3900	struct xlog_recover_item        *item)
3901{
3902	struct xfs_inode_log_format	ilf_buf;
3903	struct xfs_inode_log_format	*ilfp;
3904	struct xfs_mount		*mp = log->l_mp;
3905	int			error;
3906
3907	if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
3908		ilfp = item->ri_buf[0].i_addr;
3909	} else {
3910		ilfp = &ilf_buf;
3911		memset(ilfp, 0, sizeof(*ilfp));
3912		error = xfs_inode_item_format_convert(&item->ri_buf[0], ilfp);
3913		if (error)
3914			return;
3915	}
3916
3917	if (xlog_peek_buffer_cancelled(log, ilfp->ilf_blkno, ilfp->ilf_len, 0))
3918		return;
3919
3920	xfs_buf_readahead(mp->m_ddev_targp, ilfp->ilf_blkno,
3921				ilfp->ilf_len, &xfs_inode_buf_ra_ops);
3922}
3923
3924STATIC void
3925xlog_recover_dquot_ra_pass2(
3926	struct xlog			*log,
3927	struct xlog_recover_item	*item)
3928{
3929	struct xfs_mount	*mp = log->l_mp;
3930	struct xfs_disk_dquot	*recddq;
3931	struct xfs_dq_logformat	*dq_f;
3932	uint			type;
3933	int			len;
3934
3935
3936	if (mp->m_qflags == 0)
3937		return;
3938
3939	recddq = item->ri_buf[1].i_addr;
3940	if (recddq == NULL)
3941		return;
3942	if (item->ri_buf[1].i_len < sizeof(struct xfs_disk_dquot))
3943		return;
3944
3945	type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3946	ASSERT(type);
3947	if (log->l_quotaoffs_flag & type)
3948		return;
3949
3950	dq_f = item->ri_buf[0].i_addr;
3951	ASSERT(dq_f);
3952	ASSERT(dq_f->qlf_len == 1);
3953
3954	len = XFS_FSB_TO_BB(mp, dq_f->qlf_len);
3955	if (xlog_peek_buffer_cancelled(log, dq_f->qlf_blkno, len, 0))
3956		return;
3957
3958	xfs_buf_readahead(mp->m_ddev_targp, dq_f->qlf_blkno, len,
3959			  &xfs_dquot_buf_ra_ops);
3960}
3961
3962STATIC void
3963xlog_recover_ra_pass2(
3964	struct xlog			*log,
3965	struct xlog_recover_item	*item)
3966{
3967	switch (ITEM_TYPE(item)) {
3968	case XFS_LI_BUF:
3969		xlog_recover_buffer_ra_pass2(log, item);
3970		break;
3971	case XFS_LI_INODE:
3972		xlog_recover_inode_ra_pass2(log, item);
3973		break;
3974	case XFS_LI_DQUOT:
3975		xlog_recover_dquot_ra_pass2(log, item);
3976		break;
3977	case XFS_LI_EFI:
3978	case XFS_LI_EFD:
3979	case XFS_LI_QUOTAOFF:
3980	case XFS_LI_RUI:
3981	case XFS_LI_RUD:
3982	case XFS_LI_CUI:
3983	case XFS_LI_CUD:
3984	case XFS_LI_BUI:
3985	case XFS_LI_BUD:
3986	default:
3987		break;
3988	}
3989}
3990
3991STATIC int
3992xlog_recover_commit_pass1(
3993	struct xlog			*log,
3994	struct xlog_recover		*trans,
3995	struct xlog_recover_item	*item)
3996{
3997	trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
3998
3999	switch (ITEM_TYPE(item)) {
4000	case XFS_LI_BUF:
4001		return xlog_recover_buffer_pass1(log, item);
4002	case XFS_LI_QUOTAOFF:
4003		return xlog_recover_quotaoff_pass1(log, item);
4004	case XFS_LI_INODE:
4005	case XFS_LI_EFI:
4006	case XFS_LI_EFD:
4007	case XFS_LI_DQUOT:
4008	case XFS_LI_ICREATE:
4009	case XFS_LI_RUI:
4010	case XFS_LI_RUD:
4011	case XFS_LI_CUI:
4012	case XFS_LI_CUD:
4013	case XFS_LI_BUI:
4014	case XFS_LI_BUD:
4015		/* nothing to do in pass 1 */
4016		return 0;
4017	default:
4018		xfs_warn(log->l_mp, "%s: invalid item type (%d)",
4019			__func__, ITEM_TYPE(item));
4020		ASSERT(0);
4021		return -EIO;
4022	}
4023}
4024
4025STATIC int
4026xlog_recover_commit_pass2(
4027	struct xlog			*log,
4028	struct xlog_recover		*trans,
4029	struct list_head		*buffer_list,
4030	struct xlog_recover_item	*item)
4031{
4032	trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
4033
4034	switch (ITEM_TYPE(item)) {
4035	case XFS_LI_BUF:
4036		return xlog_recover_buffer_pass2(log, buffer_list, item,
4037						 trans->r_lsn);
4038	case XFS_LI_INODE:
4039		return xlog_recover_inode_pass2(log, buffer_list, item,
4040						 trans->r_lsn);
4041	case XFS_LI_EFI:
4042		return xlog_recover_efi_pass2(log, item, trans->r_lsn);
4043	case XFS_LI_EFD:
4044		return xlog_recover_efd_pass2(log, item);
4045	case XFS_LI_RUI:
4046		return xlog_recover_rui_pass2(log, item, trans->r_lsn);
4047	case XFS_LI_RUD:
4048		return xlog_recover_rud_pass2(log, item);
4049	case XFS_LI_CUI:
4050		return xlog_recover_cui_pass2(log, item, trans->r_lsn);
4051	case XFS_LI_CUD:
4052		return xlog_recover_cud_pass2(log, item);
4053	case XFS_LI_BUI:
4054		return xlog_recover_bui_pass2(log, item, trans->r_lsn);
4055	case XFS_LI_BUD:
4056		return xlog_recover_bud_pass2(log, item);
4057	case XFS_LI_DQUOT:
4058		return xlog_recover_dquot_pass2(log, buffer_list, item,
4059						trans->r_lsn);
4060	case XFS_LI_ICREATE:
4061		return xlog_recover_do_icreate_pass2(log, buffer_list, item);
4062	case XFS_LI_QUOTAOFF:
4063		/* nothing to do in pass2 */
4064		return 0;
4065	default:
4066		xfs_warn(log->l_mp, "%s: invalid item type (%d)",
4067			__func__, ITEM_TYPE(item));
4068		ASSERT(0);
4069		return -EIO;
4070	}
4071}
4072
4073STATIC int
4074xlog_recover_items_pass2(
4075	struct xlog                     *log,
4076	struct xlog_recover             *trans,
4077	struct list_head                *buffer_list,
4078	struct list_head                *item_list)
4079{
4080	struct xlog_recover_item	*item;
4081	int				error = 0;
4082
4083	list_for_each_entry(item, item_list, ri_list) {
4084		error = xlog_recover_commit_pass2(log, trans,
4085					  buffer_list, item);
 
 
 
 
4086		if (error)
4087			return error;
4088	}
4089
4090	return error;
4091}
4092
4093/*
4094 * Perform the transaction.
4095 *
4096 * If the transaction modifies a buffer or inode, do it now.  Otherwise,
4097 * EFIs and EFDs get queued up by adding entries into the AIL for them.
4098 */
4099STATIC int
4100xlog_recover_commit_trans(
4101	struct xlog		*log,
4102	struct xlog_recover	*trans,
4103	int			pass,
4104	struct list_head	*buffer_list)
4105{
4106	int				error = 0;
4107	int				items_queued = 0;
4108	struct xlog_recover_item	*item;
4109	struct xlog_recover_item	*next;
4110	LIST_HEAD			(ra_list);
4111	LIST_HEAD			(done_list);
4112
4113	#define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
4114
4115	hlist_del_init(&trans->r_list);
4116
4117	error = xlog_recover_reorder_trans(log, trans, pass);
4118	if (error)
4119		return error;
4120
4121	list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
 
 
4122		switch (pass) {
4123		case XLOG_RECOVER_PASS1:
4124			error = xlog_recover_commit_pass1(log, trans, item);
 
4125			break;
4126		case XLOG_RECOVER_PASS2:
4127			xlog_recover_ra_pass2(log, item);
 
4128			list_move_tail(&item->ri_list, &ra_list);
4129			items_queued++;
4130			if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
4131				error = xlog_recover_items_pass2(log, trans,
4132						buffer_list, &ra_list);
4133				list_splice_tail_init(&ra_list, &done_list);
4134				items_queued = 0;
4135			}
4136
4137			break;
4138		default:
4139			ASSERT(0);
4140		}
4141
4142		if (error)
4143			goto out;
4144	}
4145
4146out:
4147	if (!list_empty(&ra_list)) {
4148		if (!error)
4149			error = xlog_recover_items_pass2(log, trans,
4150					buffer_list, &ra_list);
4151		list_splice_tail_init(&ra_list, &done_list);
4152	}
4153
4154	if (!list_empty(&done_list))
4155		list_splice_init(&done_list, &trans->r_itemq);
4156
4157	return error;
4158}
4159
4160STATIC void
4161xlog_recover_add_item(
4162	struct list_head	*head)
4163{
4164	xlog_recover_item_t	*item;
4165
4166	item = kmem_zalloc(sizeof(xlog_recover_item_t), 0);
4167	INIT_LIST_HEAD(&item->ri_list);
4168	list_add_tail(&item->ri_list, head);
4169}
4170
4171STATIC int
4172xlog_recover_add_to_cont_trans(
4173	struct xlog		*log,
4174	struct xlog_recover	*trans,
4175	char			*dp,
4176	int			len)
4177{
4178	xlog_recover_item_t	*item;
4179	char			*ptr, *old_ptr;
4180	int			old_len;
4181
4182	/*
4183	 * If the transaction is empty, the header was split across this and the
4184	 * previous record. Copy the rest of the header.
4185	 */
4186	if (list_empty(&trans->r_itemq)) {
4187		ASSERT(len <= sizeof(struct xfs_trans_header));
4188		if (len > sizeof(struct xfs_trans_header)) {
4189			xfs_warn(log->l_mp, "%s: bad header length", __func__);
4190			return -EIO;
4191		}
4192
4193		xlog_recover_add_item(&trans->r_itemq);
4194		ptr = (char *)&trans->r_theader +
4195				sizeof(struct xfs_trans_header) - len;
4196		memcpy(ptr, dp, len);
4197		return 0;
4198	}
4199
4200	/* take the tail entry */
4201	item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
 
4202
4203	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
4204	old_len = item->ri_buf[item->ri_cnt-1].i_len;
4205
4206	ptr = kmem_realloc(old_ptr, len + old_len, 0);
 
 
4207	memcpy(&ptr[old_len], dp, len);
4208	item->ri_buf[item->ri_cnt-1].i_len += len;
4209	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
4210	trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
4211	return 0;
4212}
4213
4214/*
4215 * The next region to add is the start of a new region.  It could be
4216 * a whole region or it could be the first part of a new region.  Because
4217 * of this, the assumption here is that the type and size fields of all
4218 * format structures fit into the first 32 bits of the structure.
4219 *
4220 * This works because all regions must be 32 bit aligned.  Therefore, we
4221 * either have both fields or we have neither field.  In the case we have
4222 * neither field, the data part of the region is zero length.  We only have
4223 * a log_op_header and can throw away the header since a new one will appear
4224 * later.  If we have at least 4 bytes, then we can determine how many regions
4225 * will appear in the current log item.
4226 */
4227STATIC int
4228xlog_recover_add_to_trans(
4229	struct xlog		*log,
4230	struct xlog_recover	*trans,
4231	char			*dp,
4232	int			len)
4233{
4234	struct xfs_inode_log_format	*in_f;			/* any will do */
4235	xlog_recover_item_t	*item;
4236	char			*ptr;
4237
4238	if (!len)
4239		return 0;
4240	if (list_empty(&trans->r_itemq)) {
4241		/* we need to catch log corruptions here */
4242		if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
4243			xfs_warn(log->l_mp, "%s: bad header magic number",
4244				__func__);
4245			ASSERT(0);
4246			return -EIO;
4247		}
4248
4249		if (len > sizeof(struct xfs_trans_header)) {
4250			xfs_warn(log->l_mp, "%s: bad header length", __func__);
4251			ASSERT(0);
4252			return -EIO;
4253		}
4254
4255		/*
4256		 * The transaction header can be arbitrarily split across op
4257		 * records. If we don't have the whole thing here, copy what we
4258		 * do have and handle the rest in the next record.
4259		 */
4260		if (len == sizeof(struct xfs_trans_header))
4261			xlog_recover_add_item(&trans->r_itemq);
4262		memcpy(&trans->r_theader, dp, len);
4263		return 0;
4264	}
4265
4266	ptr = kmem_alloc(len, 0);
4267	memcpy(ptr, dp, len);
4268	in_f = (struct xfs_inode_log_format *)ptr;
4269
4270	/* take the tail entry */
4271	item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
 
4272	if (item->ri_total != 0 &&
4273	     item->ri_total == item->ri_cnt) {
4274		/* tail item is in use, get a new one */
4275		xlog_recover_add_item(&trans->r_itemq);
4276		item = list_entry(trans->r_itemq.prev,
4277					xlog_recover_item_t, ri_list);
4278	}
4279
4280	if (item->ri_total == 0) {		/* first region to be added */
4281		if (in_f->ilf_size == 0 ||
4282		    in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
4283			xfs_warn(log->l_mp,
4284		"bad number of regions (%d) in inode log format",
4285				  in_f->ilf_size);
4286			ASSERT(0);
4287			kmem_free(ptr);
4288			return -EIO;
4289		}
4290
4291		item->ri_total = in_f->ilf_size;
4292		item->ri_buf =
4293			kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
4294				    0);
4295	}
4296	ASSERT(item->ri_total > item->ri_cnt);
 
 
 
 
 
 
 
 
 
4297	/* Description region is ri_buf[0] */
4298	item->ri_buf[item->ri_cnt].i_addr = ptr;
4299	item->ri_buf[item->ri_cnt].i_len  = len;
4300	item->ri_cnt++;
4301	trace_xfs_log_recover_item_add(log, trans, item, 0);
4302	return 0;
4303}
4304
4305/*
4306 * Free up any resources allocated by the transaction
4307 *
4308 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
4309 */
4310STATIC void
4311xlog_recover_free_trans(
4312	struct xlog_recover	*trans)
4313{
4314	xlog_recover_item_t	*item, *n;
4315	int			i;
4316
4317	hlist_del_init(&trans->r_list);
4318
4319	list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
4320		/* Free the regions in the item. */
4321		list_del(&item->ri_list);
4322		for (i = 0; i < item->ri_cnt; i++)
4323			kmem_free(item->ri_buf[i].i_addr);
4324		/* Free the item itself */
4325		kmem_free(item->ri_buf);
4326		kmem_free(item);
4327	}
4328	/* Free the transaction recover structure */
4329	kmem_free(trans);
4330}
4331
4332/*
4333 * On error or completion, trans is freed.
4334 */
4335STATIC int
4336xlog_recovery_process_trans(
4337	struct xlog		*log,
4338	struct xlog_recover	*trans,
4339	char			*dp,
4340	unsigned int		len,
4341	unsigned int		flags,
4342	int			pass,
4343	struct list_head	*buffer_list)
4344{
4345	int			error = 0;
4346	bool			freeit = false;
4347
4348	/* mask off ophdr transaction container flags */
4349	flags &= ~XLOG_END_TRANS;
4350	if (flags & XLOG_WAS_CONT_TRANS)
4351		flags &= ~XLOG_CONTINUE_TRANS;
4352
4353	/*
4354	 * Callees must not free the trans structure. We'll decide if we need to
4355	 * free it or not based on the operation being done and it's result.
4356	 */
4357	switch (flags) {
4358	/* expected flag values */
4359	case 0:
4360	case XLOG_CONTINUE_TRANS:
4361		error = xlog_recover_add_to_trans(log, trans, dp, len);
4362		break;
4363	case XLOG_WAS_CONT_TRANS:
4364		error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
4365		break;
4366	case XLOG_COMMIT_TRANS:
4367		error = xlog_recover_commit_trans(log, trans, pass,
4368						  buffer_list);
4369		/* success or fail, we are now done with this transaction. */
4370		freeit = true;
4371		break;
4372
4373	/* unexpected flag values */
4374	case XLOG_UNMOUNT_TRANS:
4375		/* just skip trans */
4376		xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
4377		freeit = true;
4378		break;
4379	case XLOG_START_TRANS:
4380	default:
4381		xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
4382		ASSERT(0);
4383		error = -EIO;
4384		break;
4385	}
4386	if (error || freeit)
4387		xlog_recover_free_trans(trans);
4388	return error;
4389}
4390
4391/*
4392 * Lookup the transaction recovery structure associated with the ID in the
4393 * current ophdr. If the transaction doesn't exist and the start flag is set in
4394 * the ophdr, then allocate a new transaction for future ID matches to find.
4395 * Either way, return what we found during the lookup - an existing transaction
4396 * or nothing.
4397 */
4398STATIC struct xlog_recover *
4399xlog_recover_ophdr_to_trans(
4400	struct hlist_head	rhash[],
4401	struct xlog_rec_header	*rhead,
4402	struct xlog_op_header	*ohead)
4403{
4404	struct xlog_recover	*trans;
4405	xlog_tid_t		tid;
4406	struct hlist_head	*rhp;
4407
4408	tid = be32_to_cpu(ohead->oh_tid);
4409	rhp = &rhash[XLOG_RHASH(tid)];
4410	hlist_for_each_entry(trans, rhp, r_list) {
4411		if (trans->r_log_tid == tid)
4412			return trans;
4413	}
4414
4415	/*
4416	 * skip over non-start transaction headers - we could be
4417	 * processing slack space before the next transaction starts
4418	 */
4419	if (!(ohead->oh_flags & XLOG_START_TRANS))
4420		return NULL;
4421
4422	ASSERT(be32_to_cpu(ohead->oh_len) == 0);
4423
4424	/*
4425	 * This is a new transaction so allocate a new recovery container to
4426	 * hold the recovery ops that will follow.
4427	 */
4428	trans = kmem_zalloc(sizeof(struct xlog_recover), 0);
4429	trans->r_log_tid = tid;
4430	trans->r_lsn = be64_to_cpu(rhead->h_lsn);
4431	INIT_LIST_HEAD(&trans->r_itemq);
4432	INIT_HLIST_NODE(&trans->r_list);
4433	hlist_add_head(&trans->r_list, rhp);
4434
4435	/*
4436	 * Nothing more to do for this ophdr. Items to be added to this new
4437	 * transaction will be in subsequent ophdr containers.
4438	 */
4439	return NULL;
4440}
4441
4442STATIC int
4443xlog_recover_process_ophdr(
4444	struct xlog		*log,
4445	struct hlist_head	rhash[],
4446	struct xlog_rec_header	*rhead,
4447	struct xlog_op_header	*ohead,
4448	char			*dp,
4449	char			*end,
4450	int			pass,
4451	struct list_head	*buffer_list)
4452{
4453	struct xlog_recover	*trans;
4454	unsigned int		len;
4455	int			error;
4456
4457	/* Do we understand who wrote this op? */
4458	if (ohead->oh_clientid != XFS_TRANSACTION &&
4459	    ohead->oh_clientid != XFS_LOG) {
4460		xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
4461			__func__, ohead->oh_clientid);
4462		ASSERT(0);
4463		return -EIO;
4464	}
4465
4466	/*
4467	 * Check the ophdr contains all the data it is supposed to contain.
4468	 */
4469	len = be32_to_cpu(ohead->oh_len);
4470	if (dp + len > end) {
4471		xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
4472		WARN_ON(1);
4473		return -EIO;
4474	}
4475
4476	trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
4477	if (!trans) {
4478		/* nothing to do, so skip over this ophdr */
4479		return 0;
4480	}
4481
4482	/*
4483	 * The recovered buffer queue is drained only once we know that all
4484	 * recovery items for the current LSN have been processed. This is
4485	 * required because:
4486	 *
4487	 * - Buffer write submission updates the metadata LSN of the buffer.
4488	 * - Log recovery skips items with a metadata LSN >= the current LSN of
4489	 *   the recovery item.
4490	 * - Separate recovery items against the same metadata buffer can share
4491	 *   a current LSN. I.e., consider that the LSN of a recovery item is
4492	 *   defined as the starting LSN of the first record in which its
4493	 *   transaction appears, that a record can hold multiple transactions,
4494	 *   and/or that a transaction can span multiple records.
4495	 *
4496	 * In other words, we are allowed to submit a buffer from log recovery
4497	 * once per current LSN. Otherwise, we may incorrectly skip recovery
4498	 * items and cause corruption.
4499	 *
4500	 * We don't know up front whether buffers are updated multiple times per
4501	 * LSN. Therefore, track the current LSN of each commit log record as it
4502	 * is processed and drain the queue when it changes. Use commit records
4503	 * because they are ordered correctly by the logging code.
4504	 */
4505	if (log->l_recovery_lsn != trans->r_lsn &&
4506	    ohead->oh_flags & XLOG_COMMIT_TRANS) {
4507		error = xfs_buf_delwri_submit(buffer_list);
4508		if (error)
4509			return error;
4510		log->l_recovery_lsn = trans->r_lsn;
4511	}
4512
4513	return xlog_recovery_process_trans(log, trans, dp, len,
4514					   ohead->oh_flags, pass, buffer_list);
4515}
4516
4517/*
4518 * There are two valid states of the r_state field.  0 indicates that the
4519 * transaction structure is in a normal state.  We have either seen the
4520 * start of the transaction or the last operation we added was not a partial
4521 * operation.  If the last operation we added to the transaction was a
4522 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
4523 *
4524 * NOTE: skip LRs with 0 data length.
4525 */
4526STATIC int
4527xlog_recover_process_data(
4528	struct xlog		*log,
4529	struct hlist_head	rhash[],
4530	struct xlog_rec_header	*rhead,
4531	char			*dp,
4532	int			pass,
4533	struct list_head	*buffer_list)
4534{
4535	struct xlog_op_header	*ohead;
4536	char			*end;
4537	int			num_logops;
4538	int			error;
4539
4540	end = dp + be32_to_cpu(rhead->h_len);
4541	num_logops = be32_to_cpu(rhead->h_num_logops);
4542
4543	/* check the log format matches our own - else we can't recover */
4544	if (xlog_header_check_recover(log->l_mp, rhead))
4545		return -EIO;
4546
4547	trace_xfs_log_recover_record(log, rhead, pass);
4548	while ((dp < end) && num_logops) {
4549
4550		ohead = (struct xlog_op_header *)dp;
4551		dp += sizeof(*ohead);
4552		ASSERT(dp <= end);
4553
4554		/* errors will abort recovery */
4555		error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
4556						   dp, end, pass, buffer_list);
4557		if (error)
4558			return error;
4559
4560		dp += be32_to_cpu(ohead->oh_len);
4561		num_logops--;
4562	}
4563	return 0;
4564}
4565
4566/* Recover the EFI if necessary. */
4567STATIC int
4568xlog_recover_process_efi(
4569	struct xfs_mount		*mp,
4570	struct xfs_ail			*ailp,
4571	struct xfs_log_item		*lip)
4572{
4573	struct xfs_efi_log_item		*efip;
4574	int				error;
 
4575
4576	/*
4577	 * Skip EFIs that we've already processed.
4578	 */
4579	efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4580	if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags))
4581		return 0;
 
 
 
 
 
 
 
4582
4583	spin_unlock(&ailp->ail_lock);
4584	error = xfs_efi_recover(mp, efip);
4585	spin_lock(&ailp->ail_lock);
 
 
 
4586
4587	return error;
4588}
 
 
 
 
 
 
 
 
 
4589
4590/* Release the EFI since we're cancelling everything. */
4591STATIC void
4592xlog_recover_cancel_efi(
4593	struct xfs_mount		*mp,
4594	struct xfs_ail			*ailp,
4595	struct xfs_log_item		*lip)
4596{
4597	struct xfs_efi_log_item		*efip;
4598
4599	efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4600
4601	spin_unlock(&ailp->ail_lock);
4602	xfs_efi_release(efip);
4603	spin_lock(&ailp->ail_lock);
4604}
4605
4606/* Recover the RUI if necessary. */
4607STATIC int
4608xlog_recover_process_rui(
4609	struct xfs_mount		*mp,
4610	struct xfs_ail			*ailp,
4611	struct xfs_log_item		*lip)
4612{
4613	struct xfs_rui_log_item		*ruip;
4614	int				error;
4615
4616	/*
4617	 * Skip RUIs that we've already processed.
4618	 */
4619	ruip = container_of(lip, struct xfs_rui_log_item, rui_item);
4620	if (test_bit(XFS_RUI_RECOVERED, &ruip->rui_flags))
4621		return 0;
4622
4623	spin_unlock(&ailp->ail_lock);
4624	error = xfs_rui_recover(mp, ruip);
4625	spin_lock(&ailp->ail_lock);
4626
4627	return error;
4628}
4629
4630/* Release the RUI since we're cancelling everything. */
4631STATIC void
4632xlog_recover_cancel_rui(
4633	struct xfs_mount		*mp,
4634	struct xfs_ail			*ailp,
4635	struct xfs_log_item		*lip)
4636{
4637	struct xfs_rui_log_item		*ruip;
4638
4639	ruip = container_of(lip, struct xfs_rui_log_item, rui_item);
4640
4641	spin_unlock(&ailp->ail_lock);
4642	xfs_rui_release(ruip);
4643	spin_lock(&ailp->ail_lock);
4644}
4645
4646/* Recover the CUI if necessary. */
4647STATIC int
4648xlog_recover_process_cui(
4649	struct xfs_trans		*parent_tp,
4650	struct xfs_ail			*ailp,
4651	struct xfs_log_item		*lip)
4652{
4653	struct xfs_cui_log_item		*cuip;
4654	int				error;
4655
4656	/*
4657	 * Skip CUIs that we've already processed.
4658	 */
4659	cuip = container_of(lip, struct xfs_cui_log_item, cui_item);
4660	if (test_bit(XFS_CUI_RECOVERED, &cuip->cui_flags))
4661		return 0;
4662
4663	spin_unlock(&ailp->ail_lock);
4664	error = xfs_cui_recover(parent_tp, cuip);
4665	spin_lock(&ailp->ail_lock);
4666
4667	return error;
4668}
4669
4670/* Release the CUI since we're cancelling everything. */
4671STATIC void
4672xlog_recover_cancel_cui(
4673	struct xfs_mount		*mp,
4674	struct xfs_ail			*ailp,
4675	struct xfs_log_item		*lip)
4676{
4677	struct xfs_cui_log_item		*cuip;
4678
4679	cuip = container_of(lip, struct xfs_cui_log_item, cui_item);
4680
4681	spin_unlock(&ailp->ail_lock);
4682	xfs_cui_release(cuip);
4683	spin_lock(&ailp->ail_lock);
4684}
4685
4686/* Recover the BUI if necessary. */
4687STATIC int
4688xlog_recover_process_bui(
4689	struct xfs_trans		*parent_tp,
4690	struct xfs_ail			*ailp,
4691	struct xfs_log_item		*lip)
4692{
4693	struct xfs_bui_log_item		*buip;
4694	int				error;
4695
4696	/*
4697	 * Skip BUIs that we've already processed.
4698	 */
4699	buip = container_of(lip, struct xfs_bui_log_item, bui_item);
4700	if (test_bit(XFS_BUI_RECOVERED, &buip->bui_flags))
4701		return 0;
4702
4703	spin_unlock(&ailp->ail_lock);
4704	error = xfs_bui_recover(parent_tp, buip);
4705	spin_lock(&ailp->ail_lock);
4706
4707	return error;
4708}
4709
4710/* Release the BUI since we're cancelling everything. */
4711STATIC void
4712xlog_recover_cancel_bui(
4713	struct xfs_mount		*mp,
4714	struct xfs_ail			*ailp,
4715	struct xfs_log_item		*lip)
4716{
4717	struct xfs_bui_log_item		*buip;
4718
4719	buip = container_of(lip, struct xfs_bui_log_item, bui_item);
4720
4721	spin_unlock(&ailp->ail_lock);
4722	xfs_bui_release(buip);
4723	spin_lock(&ailp->ail_lock);
4724}
4725
4726/* Is this log item a deferred action intent? */
4727static inline bool xlog_item_is_intent(struct xfs_log_item *lip)
4728{
4729	switch (lip->li_type) {
4730	case XFS_LI_EFI:
4731	case XFS_LI_RUI:
4732	case XFS_LI_CUI:
4733	case XFS_LI_BUI:
4734		return true;
4735	default:
4736		return false;
4737	}
4738}
4739
4740/* Take all the collected deferred ops and finish them in order. */
4741static int
4742xlog_finish_defer_ops(
4743	struct xfs_trans	*parent_tp)
4744{
4745	struct xfs_mount	*mp = parent_tp->t_mountp;
4746	struct xfs_trans	*tp;
4747	int64_t			freeblks;
4748	uint			resblks;
4749	int			error;
4750
4751	/*
4752	 * We're finishing the defer_ops that accumulated as a result of
4753	 * recovering unfinished intent items during log recovery.  We
4754	 * reserve an itruncate transaction because it is the largest
4755	 * permanent transaction type.  Since we're the only user of the fs
4756	 * right now, take 93% (15/16) of the available free blocks.  Use
4757	 * weird math to avoid a 64-bit division.
4758	 */
4759	freeblks = percpu_counter_sum(&mp->m_fdblocks);
4760	if (freeblks <= 0)
4761		return -ENOSPC;
4762	resblks = min_t(int64_t, UINT_MAX, freeblks);
4763	resblks = (resblks * 15) >> 4;
4764	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, resblks,
4765			0, XFS_TRANS_RESERVE, &tp);
4766	if (error)
4767		return error;
4768	/* transfer all collected dfops to this transaction */
4769	xfs_defer_move(tp, parent_tp);
4770
4771	return xfs_trans_commit(tp);
4772}
4773
4774/*
4775 * When this is called, all of the log intent items which did not have
4776 * corresponding log done items should be in the AIL.  What we do now
4777 * is update the data structures associated with each one.
4778 *
4779 * Since we process the log intent items in normal transactions, they
4780 * will be removed at some point after the commit.  This prevents us
4781 * from just walking down the list processing each one.  We'll use a
4782 * flag in the intent item to skip those that we've already processed
4783 * and use the AIL iteration mechanism's generation count to try to
4784 * speed this up at least a bit.
4785 *
4786 * When we start, we know that the intents are the only things in the
4787 * AIL.  As we process them, however, other items are added to the
4788 * AIL.
 
 
 
 
 
 
 
4789 */
4790STATIC int
4791xlog_recover_process_intents(
4792	struct xlog		*log)
4793{
4794	struct xfs_trans	*parent_tp;
4795	struct xfs_ail_cursor	cur;
4796	struct xfs_log_item	*lip;
4797	struct xfs_ail		*ailp;
4798	int			error;
4799#if defined(DEBUG) || defined(XFS_WARN)
4800	xfs_lsn_t		last_lsn;
4801#endif
4802
4803	/*
4804	 * The intent recovery handlers commit transactions to complete recovery
4805	 * for individual intents, but any new deferred operations that are
4806	 * queued during that process are held off until the very end. The
4807	 * purpose of this transaction is to serve as a container for deferred
4808	 * operations. Each intent recovery handler must transfer dfops here
4809	 * before its local transaction commits, and we'll finish the entire
4810	 * list below.
4811	 */
4812	error = xfs_trans_alloc_empty(log->l_mp, &parent_tp);
4813	if (error)
4814		return error;
4815
4816	ailp = log->l_ailp;
4817	spin_lock(&ailp->ail_lock);
4818	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4819#if defined(DEBUG) || defined(XFS_WARN)
4820	last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
4821#endif
4822	while (lip != NULL) {
4823		/*
4824		 * We're done when we see something other than an intent.
4825		 * There should be no intents left in the AIL now.
4826		 */
4827		if (!xlog_item_is_intent(lip)) {
4828#ifdef DEBUG
4829			for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4830				ASSERT(!xlog_item_is_intent(lip));
4831#endif
4832			break;
4833		}
4834
4835		/*
4836		 * We should never see a redo item with a LSN higher than
4837		 * the last transaction we found in the log at the start
4838		 * of recovery.
4839		 */
4840		ASSERT(XFS_LSN_CMP(last_lsn, lip->li_lsn) >= 0);
4841
4842		/*
4843		 * NOTE: If your intent processing routine can create more
4844		 * deferred ops, you /must/ attach them to the dfops in this
4845		 * routine or else those subsequent intents will get
4846		 * replayed in the wrong order!
 
 
 
 
4847		 */
4848		switch (lip->li_type) {
4849		case XFS_LI_EFI:
4850			error = xlog_recover_process_efi(log->l_mp, ailp, lip);
4851			break;
4852		case XFS_LI_RUI:
4853			error = xlog_recover_process_rui(log->l_mp, ailp, lip);
4854			break;
4855		case XFS_LI_CUI:
4856			error = xlog_recover_process_cui(parent_tp, ailp, lip);
4857			break;
4858		case XFS_LI_BUI:
4859			error = xlog_recover_process_bui(parent_tp, ailp, lip);
4860			break;
4861		}
4862		if (error)
4863			goto out;
4864		lip = xfs_trans_ail_cursor_next(ailp, &cur);
4865	}
4866out:
4867	xfs_trans_ail_cursor_done(&cur);
4868	spin_unlock(&ailp->ail_lock);
4869	if (!error)
4870		error = xlog_finish_defer_ops(parent_tp);
4871	xfs_trans_cancel(parent_tp);
4872
 
 
 
4873	return error;
4874}
4875
4876/*
4877 * A cancel occurs when the mount has failed and we're bailing out.
4878 * Release all pending log intent items so they don't pin the AIL.
 
4879 */
4880STATIC void
4881xlog_recover_cancel_intents(
4882	struct xlog		*log)
4883{
4884	struct xfs_log_item	*lip;
4885	struct xfs_ail_cursor	cur;
4886	struct xfs_ail		*ailp;
4887
4888	ailp = log->l_ailp;
4889	spin_lock(&ailp->ail_lock);
4890	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4891	while (lip != NULL) {
4892		/*
4893		 * We're done when we see something other than an intent.
4894		 * There should be no intents left in the AIL now.
4895		 */
4896		if (!xlog_item_is_intent(lip)) {
4897#ifdef DEBUG
4898			for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4899				ASSERT(!xlog_item_is_intent(lip));
4900#endif
4901			break;
4902		}
4903
4904		switch (lip->li_type) {
4905		case XFS_LI_EFI:
4906			xlog_recover_cancel_efi(log->l_mp, ailp, lip);
4907			break;
4908		case XFS_LI_RUI:
4909			xlog_recover_cancel_rui(log->l_mp, ailp, lip);
4910			break;
4911		case XFS_LI_CUI:
4912			xlog_recover_cancel_cui(log->l_mp, ailp, lip);
4913			break;
4914		case XFS_LI_BUI:
4915			xlog_recover_cancel_bui(log->l_mp, ailp, lip);
4916			break;
4917		}
4918
4919		lip = xfs_trans_ail_cursor_next(ailp, &cur);
4920	}
 
 
 
 
 
 
 
 
 
 
 
 
 
4921
4922	xfs_trans_ail_cursor_done(&cur);
4923	spin_unlock(&ailp->ail_lock);
 
 
 
4924}
4925
4926/*
4927 * This routine performs a transaction to null out a bad inode pointer
4928 * in an agi unlinked inode hash bucket.
4929 */
4930STATIC void
4931xlog_recover_clear_agi_bucket(
4932	xfs_mount_t	*mp,
4933	xfs_agnumber_t	agno,
4934	int		bucket)
4935{
4936	xfs_trans_t	*tp;
4937	xfs_agi_t	*agi;
4938	xfs_buf_t	*agibp;
4939	int		offset;
4940	int		error;
4941
4942	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
4943	if (error)
4944		goto out_error;
4945
4946	error = xfs_read_agi(mp, tp, agno, &agibp);
4947	if (error)
4948		goto out_abort;
4949
4950	agi = XFS_BUF_TO_AGI(agibp);
4951	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
4952	offset = offsetof(xfs_agi_t, agi_unlinked) +
4953		 (sizeof(xfs_agino_t) * bucket);
4954	xfs_trans_log_buf(tp, agibp, offset,
4955			  (offset + sizeof(xfs_agino_t) - 1));
4956
4957	error = xfs_trans_commit(tp);
4958	if (error)
4959		goto out_error;
4960	return;
4961
4962out_abort:
4963	xfs_trans_cancel(tp);
4964out_error:
4965	xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
 
4966	return;
4967}
4968
4969STATIC xfs_agino_t
4970xlog_recover_process_one_iunlink(
4971	struct xfs_mount		*mp,
4972	xfs_agnumber_t			agno,
4973	xfs_agino_t			agino,
4974	int				bucket)
4975{
4976	struct xfs_buf			*ibp;
4977	struct xfs_dinode		*dip;
4978	struct xfs_inode		*ip;
4979	xfs_ino_t			ino;
4980	int				error;
4981
4982	ino = XFS_AGINO_TO_INO(mp, agno, agino);
4983	error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
4984	if (error)
4985		goto fail;
 
 
 
4986
4987	/*
4988	 * Get the on disk inode to find the next inode in the bucket.
4989	 */
4990	error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0, 0);
4991	if (error)
4992		goto fail_iput;
 
 
4993
4994	xfs_iflags_clear(ip, XFS_IRECOVERY);
4995	ASSERT(VFS_I(ip)->i_nlink == 0);
4996	ASSERT(VFS_I(ip)->i_mode != 0);
 
 
 
 
 
 
 
 
 
 
4997
4998	/* setup for the next pass */
4999	agino = be32_to_cpu(dip->di_next_unlinked);
5000	xfs_buf_relse(ibp);
5001
5002	/*
5003	 * Prevent any DMAPI event from being sent when the reference on
5004	 * the inode is dropped.
5005	 */
5006	ip->i_d.di_dmevmask = 0;
5007
5008	xfs_irele(ip);
5009	return agino;
5010
5011 fail_iput:
5012	xfs_irele(ip);
5013 fail:
5014	/*
5015	 * We can't read in the inode this bucket points to, or this inode
5016	 * is messed up.  Just ditch this bucket of inodes.  We will lose
5017	 * some inodes and space, but at least we won't hang.
5018	 *
5019	 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
5020	 * clear the inode pointer in the bucket.
5021	 */
5022	xlog_recover_clear_agi_bucket(mp, agno, bucket);
5023	return NULLAGINO;
5024}
5025
5026/*
5027 * Recover AGI unlinked lists
5028 *
5029 * This is called during recovery to process any inodes which we unlinked but
5030 * not freed when the system crashed.  These inodes will be on the lists in the
5031 * AGI blocks. What we do here is scan all the AGIs and fully truncate and free
5032 * any inodes found on the lists. Each inode is removed from the lists when it
5033 * has been fully truncated and is freed. The freeing of the inode and its
5034 * removal from the list must be atomic.
5035 *
5036 * If everything we touch in the agi processing loop is already in memory, this
5037 * loop can hold the cpu for a long time. It runs without lock contention,
5038 * memory allocation contention, the need wait for IO, etc, and so will run
5039 * until we either run out of inodes to process, run low on memory or we run out
5040 * of log space.
5041 *
5042 * This behaviour is bad for latency on single CPU and non-preemptible kernels,
5043 * and can prevent other filesytem work (such as CIL pushes) from running. This
5044 * can lead to deadlocks if the recovery process runs out of log reservation
5045 * space. Hence we need to yield the CPU when there is other kernel work
5046 * scheduled on this CPU to ensure other scheduled work can run without undue
5047 * latency.
5048 */
5049STATIC void
5050xlog_recover_process_iunlinks(
5051	struct xlog	*log)
5052{
5053	xfs_mount_t	*mp;
5054	xfs_agnumber_t	agno;
5055	xfs_agi_t	*agi;
5056	xfs_buf_t	*agibp;
5057	xfs_agino_t	agino;
5058	int		bucket;
5059	int		error;
5060
5061	mp = log->l_mp;
5062
5063	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5064		/*
5065		 * Find the agi for this ag.
 
 
 
5066		 */
5067		error = xfs_read_agi(mp, NULL, agno, &agibp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5068		if (error) {
5069			/*
5070			 * AGI is b0rked. Don't process it.
5071			 *
5072			 * We should probably mark the filesystem as corrupt
5073			 * after we've recovered all the ag's we can....
5074			 */
5075			continue;
5076		}
5077		/*
5078		 * Unlock the buffer so that it can be acquired in the normal
5079		 * course of the transaction to truncate and free each inode.
5080		 * Because we are not racing with anyone else here for the AGI
5081		 * buffer, we don't even need to hold it locked to read the
5082		 * initial unlinked bucket entries out of the buffer. We keep
5083		 * buffer reference though, so that it stays pinned in memory
5084		 * while we need the buffer.
5085		 */
5086		agi = XFS_BUF_TO_AGI(agibp);
5087		xfs_buf_unlock(agibp);
5088
5089		for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
5090			agino = be32_to_cpu(agi->agi_unlinked[bucket]);
5091			while (agino != NULLAGINO) {
5092				agino = xlog_recover_process_one_iunlink(mp,
5093							agno, agino, bucket);
5094				cond_resched();
5095			}
5096		}
5097		xfs_buf_rele(agibp);
5098	}
5099}
5100
5101STATIC void
5102xlog_unpack_data(
5103	struct xlog_rec_header	*rhead,
5104	char			*dp,
5105	struct xlog		*log)
5106{
5107	int			i, j, k;
5108
5109	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
5110		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
5111		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
5112		dp += BBSIZE;
5113	}
5114
5115	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
5116		xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
5117		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
5118			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
5119			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
5120			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
5121			dp += BBSIZE;
5122		}
5123	}
5124}
5125
5126/*
5127 * CRC check, unpack and process a log record.
5128 */
5129STATIC int
5130xlog_recover_process(
5131	struct xlog		*log,
5132	struct hlist_head	rhash[],
5133	struct xlog_rec_header	*rhead,
5134	char			*dp,
5135	int			pass,
5136	struct list_head	*buffer_list)
5137{
5138	__le32			old_crc = rhead->h_crc;
5139	__le32			crc;
5140
5141	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
5142
5143	/*
5144	 * Nothing else to do if this is a CRC verification pass. Just return
5145	 * if this a record with a non-zero crc. Unfortunately, mkfs always
5146	 * sets old_crc to 0 so we must consider this valid even on v5 supers.
5147	 * Otherwise, return EFSBADCRC on failure so the callers up the stack
5148	 * know precisely what failed.
5149	 */
5150	if (pass == XLOG_RECOVER_CRCPASS) {
5151		if (old_crc && crc != old_crc)
5152			return -EFSBADCRC;
5153		return 0;
5154	}
5155
5156	/*
5157	 * We're in the normal recovery path. Issue a warning if and only if the
5158	 * CRC in the header is non-zero. This is an advisory warning and the
5159	 * zero CRC check prevents warnings from being emitted when upgrading
5160	 * the kernel from one that does not add CRCs by default.
5161	 */
5162	if (crc != old_crc) {
5163		if (old_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
5164			xfs_alert(log->l_mp,
5165		"log record CRC mismatch: found 0x%x, expected 0x%x.",
5166					le32_to_cpu(old_crc),
5167					le32_to_cpu(crc));
5168			xfs_hex_dump(dp, 32);
5169		}
5170
5171		/*
5172		 * If the filesystem is CRC enabled, this mismatch becomes a
5173		 * fatal log corruption failure.
5174		 */
5175		if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
 
5176			return -EFSCORRUPTED;
 
5177	}
5178
5179	xlog_unpack_data(rhead, dp, log);
5180
5181	return xlog_recover_process_data(log, rhash, rhead, dp, pass,
5182					 buffer_list);
5183}
5184
5185STATIC int
5186xlog_valid_rec_header(
5187	struct xlog		*log,
5188	struct xlog_rec_header	*rhead,
5189	xfs_daddr_t		blkno)
 
5190{
5191	int			hlen;
5192
5193	if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
5194		XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
5195				XFS_ERRLEVEL_LOW, log->l_mp);
5196		return -EFSCORRUPTED;
5197	}
5198	if (unlikely(
5199	    (!rhead->h_version ||
5200	    (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
5201		xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
5202			__func__, be32_to_cpu(rhead->h_version));
5203		return -EIO;
5204	}
5205
5206	/* LR body must have data or it wouldn't have been written */
 
 
 
5207	hlen = be32_to_cpu(rhead->h_len);
5208	if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
5209		XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
5210				XFS_ERRLEVEL_LOW, log->l_mp);
5211		return -EFSCORRUPTED;
5212	}
5213	if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
5214		XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
5215				XFS_ERRLEVEL_LOW, log->l_mp);
5216		return -EFSCORRUPTED;
5217	}
5218	return 0;
5219}
5220
5221/*
5222 * Read the log from tail to head and process the log records found.
5223 * Handle the two cases where the tail and head are in the same cycle
5224 * and where the active portion of the log wraps around the end of
5225 * the physical log separately.  The pass parameter is passed through
5226 * to the routines called to process the data and is not looked at
5227 * here.
5228 */
5229STATIC int
5230xlog_do_recovery_pass(
5231	struct xlog		*log,
5232	xfs_daddr_t		head_blk,
5233	xfs_daddr_t		tail_blk,
5234	int			pass,
5235	xfs_daddr_t		*first_bad)	/* out: first bad log rec */
5236{
5237	xlog_rec_header_t	*rhead;
5238	xfs_daddr_t		blk_no, rblk_no;
5239	xfs_daddr_t		rhead_blk;
5240	char			*offset;
5241	char			*hbp, *dbp;
5242	int			error = 0, h_size, h_len;
5243	int			error2 = 0;
5244	int			bblks, split_bblks;
5245	int			hblks, split_hblks, wrapped_hblks;
5246	int			i;
5247	struct hlist_head	rhash[XLOG_RHASH_SIZE];
5248	LIST_HEAD		(buffer_list);
5249
5250	ASSERT(head_blk != tail_blk);
5251	blk_no = rhead_blk = tail_blk;
5252
5253	for (i = 0; i < XLOG_RHASH_SIZE; i++)
5254		INIT_HLIST_HEAD(&rhash[i]);
5255
5256	/*
5257	 * Read the header of the tail block and get the iclog buffer size from
5258	 * h_size.  Use this to tell how many sectors make up the log header.
5259	 */
5260	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
5261		/*
5262		 * When using variable length iclogs, read first sector of
5263		 * iclog header and extract the header size from it.  Get a
5264		 * new hbp that is the correct size.
5265		 */
5266		hbp = xlog_alloc_buffer(log, 1);
5267		if (!hbp)
5268			return -ENOMEM;
5269
5270		error = xlog_bread(log, tail_blk, 1, hbp, &offset);
5271		if (error)
5272			goto bread_err1;
5273
5274		rhead = (xlog_rec_header_t *)offset;
5275		error = xlog_valid_rec_header(log, rhead, tail_blk);
5276		if (error)
5277			goto bread_err1;
5278
5279		/*
5280		 * xfsprogs has a bug where record length is based on lsunit but
5281		 * h_size (iclog size) is hardcoded to 32k. Now that we
5282		 * unconditionally CRC verify the unmount record, this means the
5283		 * log buffer can be too small for the record and cause an
5284		 * overrun.
5285		 *
5286		 * Detect this condition here. Use lsunit for the buffer size as
5287		 * long as this looks like the mkfs case. Otherwise, return an
5288		 * error to avoid a buffer overrun.
5289		 */
5290		h_size = be32_to_cpu(rhead->h_size);
5291		h_len = be32_to_cpu(rhead->h_len);
5292		if (h_len > h_size) {
5293			if (h_len <= log->l_mp->m_logbsize &&
5294			    be32_to_cpu(rhead->h_num_logops) == 1) {
5295				xfs_warn(log->l_mp,
5296		"invalid iclog size (%d bytes), using lsunit (%d bytes)",
5297					 h_size, log->l_mp->m_logbsize);
5298				h_size = log->l_mp->m_logbsize;
5299			} else
5300				return -EFSCORRUPTED;
5301		}
5302
5303		if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
5304		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
5305			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
5306			if (h_size % XLOG_HEADER_CYCLE_SIZE)
5307				hblks++;
 
5308			kmem_free(hbp);
5309			hbp = xlog_alloc_buffer(log, hblks);
5310		} else {
5311			hblks = 1;
5312		}
5313	} else {
5314		ASSERT(log->l_sectBBsize == 1);
5315		hblks = 1;
5316		hbp = xlog_alloc_buffer(log, 1);
5317		h_size = XLOG_BIG_RECORD_BSIZE;
5318	}
5319
5320	if (!hbp)
5321		return -ENOMEM;
5322	dbp = xlog_alloc_buffer(log, BTOBB(h_size));
5323	if (!dbp) {
5324		kmem_free(hbp);
5325		return -ENOMEM;
5326	}
5327
5328	memset(rhash, 0, sizeof(rhash));
5329	if (tail_blk > head_blk) {
5330		/*
5331		 * Perform recovery around the end of the physical log.
5332		 * When the head is not on the same cycle number as the tail,
5333		 * we can't do a sequential recovery.
5334		 */
5335		while (blk_no < log->l_logBBsize) {
5336			/*
5337			 * Check for header wrapping around physical end-of-log
5338			 */
5339			offset = hbp;
5340			split_hblks = 0;
5341			wrapped_hblks = 0;
5342			if (blk_no + hblks <= log->l_logBBsize) {
5343				/* Read header in one read */
5344				error = xlog_bread(log, blk_no, hblks, hbp,
5345						   &offset);
5346				if (error)
5347					goto bread_err2;
5348			} else {
5349				/* This LR is split across physical log end */
5350				if (blk_no != log->l_logBBsize) {
5351					/* some data before physical log end */
5352					ASSERT(blk_no <= INT_MAX);
5353					split_hblks = log->l_logBBsize - (int)blk_no;
5354					ASSERT(split_hblks > 0);
5355					error = xlog_bread(log, blk_no,
5356							   split_hblks, hbp,
5357							   &offset);
5358					if (error)
5359						goto bread_err2;
5360				}
5361
5362				/*
5363				 * Note: this black magic still works with
5364				 * large sector sizes (non-512) only because:
5365				 * - we increased the buffer size originally
5366				 *   by 1 sector giving us enough extra space
5367				 *   for the second read;
5368				 * - the log start is guaranteed to be sector
5369				 *   aligned;
5370				 * - we read the log end (LR header start)
5371				 *   _first_, then the log start (LR header end)
5372				 *   - order is important.
5373				 */
5374				wrapped_hblks = hblks - split_hblks;
5375				error = xlog_bread_noalign(log, 0,
5376						wrapped_hblks,
5377						offset + BBTOB(split_hblks));
5378				if (error)
5379					goto bread_err2;
5380			}
5381			rhead = (xlog_rec_header_t *)offset;
5382			error = xlog_valid_rec_header(log, rhead,
5383						split_hblks ? blk_no : 0);
5384			if (error)
5385				goto bread_err2;
5386
5387			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
5388			blk_no += hblks;
5389
5390			/*
5391			 * Read the log record data in multiple reads if it
5392			 * wraps around the end of the log. Note that if the
5393			 * header already wrapped, blk_no could point past the
5394			 * end of the log. The record data is contiguous in
5395			 * that case.
5396			 */
5397			if (blk_no + bblks <= log->l_logBBsize ||
5398			    blk_no >= log->l_logBBsize) {
5399				rblk_no = xlog_wrap_logbno(log, blk_no);
5400				error = xlog_bread(log, rblk_no, bblks, dbp,
5401						   &offset);
5402				if (error)
5403					goto bread_err2;
5404			} else {
5405				/* This log record is split across the
5406				 * physical end of log */
5407				offset = dbp;
5408				split_bblks = 0;
5409				if (blk_no != log->l_logBBsize) {
5410					/* some data is before the physical
5411					 * end of log */
5412					ASSERT(!wrapped_hblks);
5413					ASSERT(blk_no <= INT_MAX);
5414					split_bblks =
5415						log->l_logBBsize - (int)blk_no;
5416					ASSERT(split_bblks > 0);
5417					error = xlog_bread(log, blk_no,
5418							split_bblks, dbp,
5419							&offset);
5420					if (error)
5421						goto bread_err2;
5422				}
5423
5424				/*
5425				 * Note: this black magic still works with
5426				 * large sector sizes (non-512) only because:
5427				 * - we increased the buffer size originally
5428				 *   by 1 sector giving us enough extra space
5429				 *   for the second read;
5430				 * - the log start is guaranteed to be sector
5431				 *   aligned;
5432				 * - we read the log end (LR header start)
5433				 *   _first_, then the log start (LR header end)
5434				 *   - order is important.
5435				 */
5436				error = xlog_bread_noalign(log, 0,
5437						bblks - split_bblks,
5438						offset + BBTOB(split_bblks));
5439				if (error)
5440					goto bread_err2;
5441			}
5442
5443			error = xlog_recover_process(log, rhash, rhead, offset,
5444						     pass, &buffer_list);
5445			if (error)
5446				goto bread_err2;
5447
5448			blk_no += bblks;
5449			rhead_blk = blk_no;
5450		}
5451
5452		ASSERT(blk_no >= log->l_logBBsize);
5453		blk_no -= log->l_logBBsize;
5454		rhead_blk = blk_no;
5455	}
5456
5457	/* read first part of physical log */
5458	while (blk_no < head_blk) {
5459		error = xlog_bread(log, blk_no, hblks, hbp, &offset);
5460		if (error)
5461			goto bread_err2;
5462
5463		rhead = (xlog_rec_header_t *)offset;
5464		error = xlog_valid_rec_header(log, rhead, blk_no);
5465		if (error)
5466			goto bread_err2;
5467
5468		/* blocks in data section */
5469		bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
5470		error = xlog_bread(log, blk_no+hblks, bblks, dbp,
5471				   &offset);
5472		if (error)
5473			goto bread_err2;
5474
5475		error = xlog_recover_process(log, rhash, rhead, offset, pass,
5476					     &buffer_list);
5477		if (error)
5478			goto bread_err2;
5479
5480		blk_no += bblks + hblks;
5481		rhead_blk = blk_no;
5482	}
5483
5484 bread_err2:
5485	kmem_free(dbp);
5486 bread_err1:
5487	kmem_free(hbp);
5488
5489	/*
5490	 * Submit buffers that have been added from the last record processed,
5491	 * regardless of error status.
5492	 */
5493	if (!list_empty(&buffer_list))
5494		error2 = xfs_buf_delwri_submit(&buffer_list);
5495
5496	if (error && first_bad)
5497		*first_bad = rhead_blk;
5498
5499	/*
5500	 * Transactions are freed at commit time but transactions without commit
5501	 * records on disk are never committed. Free any that may be left in the
5502	 * hash table.
5503	 */
5504	for (i = 0; i < XLOG_RHASH_SIZE; i++) {
5505		struct hlist_node	*tmp;
5506		struct xlog_recover	*trans;
5507
5508		hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
5509			xlog_recover_free_trans(trans);
5510	}
5511
5512	return error ? error : error2;
5513}
5514
5515/*
5516 * Do the recovery of the log.  We actually do this in two phases.
5517 * The two passes are necessary in order to implement the function
5518 * of cancelling a record written into the log.  The first pass
5519 * determines those things which have been cancelled, and the
5520 * second pass replays log items normally except for those which
5521 * have been cancelled.  The handling of the replay and cancellations
5522 * takes place in the log item type specific routines.
5523 *
5524 * The table of items which have cancel records in the log is allocated
5525 * and freed at this level, since only here do we know when all of
5526 * the log recovery has been completed.
5527 */
5528STATIC int
5529xlog_do_log_recovery(
5530	struct xlog	*log,
5531	xfs_daddr_t	head_blk,
5532	xfs_daddr_t	tail_blk)
5533{
5534	int		error, i;
5535
5536	ASSERT(head_blk != tail_blk);
5537
5538	/*
5539	 * First do a pass to find all of the cancelled buf log items.
5540	 * Store them in the buf_cancel_table for use in the second pass.
5541	 */
5542	log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
5543						 sizeof(struct list_head),
5544						 0);
5545	for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
5546		INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
5547
5548	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
5549				      XLOG_RECOVER_PASS1, NULL);
5550	if (error != 0) {
5551		kmem_free(log->l_buf_cancel_table);
5552		log->l_buf_cancel_table = NULL;
5553		return error;
5554	}
5555	/*
5556	 * Then do a second pass to actually recover the items in the log.
5557	 * When it is complete free the table of buf cancel items.
5558	 */
5559	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
5560				      XLOG_RECOVER_PASS2, NULL);
5561#ifdef DEBUG
5562	if (!error) {
5563		int	i;
5564
5565		for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
5566			ASSERT(list_empty(&log->l_buf_cancel_table[i]));
5567	}
5568#endif	/* DEBUG */
5569
5570	kmem_free(log->l_buf_cancel_table);
5571	log->l_buf_cancel_table = NULL;
5572
5573	return error;
5574}
5575
5576/*
5577 * Do the actual recovery
5578 */
5579STATIC int
5580xlog_do_recover(
5581	struct xlog	*log,
5582	xfs_daddr_t	head_blk,
5583	xfs_daddr_t	tail_blk)
5584{
5585	struct xfs_mount *mp = log->l_mp;
5586	int		error;
5587	xfs_buf_t	*bp;
5588	xfs_sb_t	*sbp;
5589
5590	trace_xfs_log_recover(log, head_blk, tail_blk);
5591
5592	/*
5593	 * First replay the images in the log.
5594	 */
5595	error = xlog_do_log_recovery(log, head_blk, tail_blk);
5596	if (error)
5597		return error;
5598
5599	/*
5600	 * If IO errors happened during recovery, bail out.
5601	 */
5602	if (XFS_FORCED_SHUTDOWN(mp)) {
5603		return -EIO;
5604	}
5605
5606	/*
5607	 * We now update the tail_lsn since much of the recovery has completed
5608	 * and there may be space available to use.  If there were no extent
5609	 * or iunlinks, we can free up the entire log and set the tail_lsn to
5610	 * be the last_sync_lsn.  This was set in xlog_find_tail to be the
5611	 * lsn of the last known good LR on disk.  If there are extent frees
5612	 * or iunlinks they will have some entries in the AIL; so we look at
5613	 * the AIL to determine how to set the tail_lsn.
5614	 */
5615	xlog_assign_tail_lsn(mp);
5616
5617	/*
5618	 * Now that we've finished replaying all buffer and inode
5619	 * updates, re-read in the superblock and reverify it.
5620	 */
5621	bp = xfs_getsb(mp);
5622	bp->b_flags &= ~(XBF_DONE | XBF_ASYNC);
5623	ASSERT(!(bp->b_flags & XBF_WRITE));
5624	bp->b_flags |= XBF_READ;
5625	bp->b_ops = &xfs_sb_buf_ops;
5626
5627	error = xfs_buf_submit(bp);
5628	if (error) {
5629		if (!XFS_FORCED_SHUTDOWN(mp)) {
5630			xfs_buf_ioerror_alert(bp, __func__);
5631			ASSERT(0);
5632		}
5633		xfs_buf_relse(bp);
5634		return error;
5635	}
5636
5637	/* Convert superblock from on-disk format */
5638	sbp = &mp->m_sb;
5639	xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
5640	xfs_buf_relse(bp);
5641
5642	/* re-initialise in-core superblock and geometry structures */
 
5643	xfs_reinit_percpu_counters(mp);
5644	error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
 
5645	if (error) {
5646		xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
5647		return error;
5648	}
5649	mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
5650
5651	xlog_recover_check_summary(log);
5652
5653	/* Normal transactions can now occur */
5654	log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
5655	return 0;
5656}
5657
5658/*
5659 * Perform recovery and re-initialize some log variables in xlog_find_tail.
5660 *
5661 * Return error or zero.
5662 */
5663int
5664xlog_recover(
5665	struct xlog	*log)
5666{
5667	xfs_daddr_t	head_blk, tail_blk;
5668	int		error;
5669
5670	/* find the tail of the log */
5671	error = xlog_find_tail(log, &head_blk, &tail_blk);
5672	if (error)
5673		return error;
5674
5675	/*
5676	 * The superblock was read before the log was available and thus the LSN
5677	 * could not be verified. Check the superblock LSN against the current
5678	 * LSN now that it's known.
5679	 */
5680	if (xfs_sb_version_hascrc(&log->l_mp->m_sb) &&
5681	    !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
5682		return -EINVAL;
5683
5684	if (tail_blk != head_blk) {
5685		/* There used to be a comment here:
5686		 *
5687		 * disallow recovery on read-only mounts.  note -- mount
5688		 * checks for ENOSPC and turns it into an intelligent
5689		 * error message.
5690		 * ...but this is no longer true.  Now, unless you specify
5691		 * NORECOVERY (in which case this function would never be
5692		 * called), we just go ahead and recover.  We do this all
5693		 * under the vfs layer, so we can get away with it unless
5694		 * the device itself is read-only, in which case we fail.
5695		 */
5696		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
5697			return error;
5698		}
5699
5700		/*
5701		 * Version 5 superblock log feature mask validation. We know the
5702		 * log is dirty so check if there are any unknown log features
5703		 * in what we need to recover. If there are unknown features
5704		 * (e.g. unsupported transactions, then simply reject the
5705		 * attempt at recovery before touching anything.
5706		 */
5707		if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
5708		    xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
5709					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
5710			xfs_warn(log->l_mp,
5711"Superblock has unknown incompatible log features (0x%x) enabled.",
5712				(log->l_mp->m_sb.sb_features_log_incompat &
5713					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
5714			xfs_warn(log->l_mp,
5715"The log can not be fully and/or safely recovered by this kernel.");
5716			xfs_warn(log->l_mp,
5717"Please recover the log on a kernel that supports the unknown features.");
5718			return -EINVAL;
5719		}
5720
5721		/*
5722		 * Delay log recovery if the debug hook is set. This is debug
5723		 * instrumention to coordinate simulation of I/O failures with
5724		 * log recovery.
5725		 */
5726		if (xfs_globals.log_recovery_delay) {
5727			xfs_notice(log->l_mp,
5728				"Delaying log recovery for %d seconds.",
5729				xfs_globals.log_recovery_delay);
5730			msleep(xfs_globals.log_recovery_delay * 1000);
5731		}
5732
5733		xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
5734				log->l_mp->m_logname ? log->l_mp->m_logname
5735						     : "internal");
5736
5737		error = xlog_do_recover(log, head_blk, tail_blk);
5738		log->l_flags |= XLOG_RECOVERY_NEEDED;
5739	}
5740	return error;
5741}
5742
5743/*
5744 * In the first part of recovery we replay inodes and buffers and build
5745 * up the list of extent free items which need to be processed.  Here
5746 * we process the extent free items and clean up the on disk unlinked
5747 * inode lists.  This is separated from the first part of recovery so
5748 * that the root and real-time bitmap inodes can be read in from disk in
5749 * between the two stages.  This is necessary so that we can free space
5750 * in the real-time portion of the file system.
5751 */
5752int
5753xlog_recover_finish(
5754	struct xlog	*log)
5755{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5756	/*
5757	 * Now we're ready to do the transactions needed for the
5758	 * rest of recovery.  Start with completing all the extent
5759	 * free intent records and then process the unlinked inode
5760	 * lists.  At this point, we essentially run in normal mode
5761	 * except that we're still performing recovery actions
5762	 * rather than accepting new requests.
5763	 */
5764	if (log->l_flags & XLOG_RECOVERY_NEEDED) {
5765		int	error;
5766		error = xlog_recover_process_intents(log);
5767		if (error) {
5768			xfs_alert(log->l_mp, "Failed to recover intents");
5769			return error;
5770		}
 
5771
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5772		/*
5773		 * Sync the log to get all the intents out of the AIL.
5774		 * This isn't absolutely necessary, but it helps in
5775		 * case the unlink transactions would have problems
5776		 * pushing the intents out of the way.
5777		 */
5778		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
5779
5780		xlog_recover_process_iunlinks(log);
5781
5782		xlog_recover_check_summary(log);
5783
5784		xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
5785				log->l_mp->m_logname ? log->l_mp->m_logname
5786						     : "internal");
5787		log->l_flags &= ~XLOG_RECOVERY_NEEDED;
5788	} else {
5789		xfs_info(log->l_mp, "Ending clean mount");
5790	}
5791	return 0;
5792}
5793
5794void
5795xlog_recover_cancel(
5796	struct xlog	*log)
5797{
5798	if (log->l_flags & XLOG_RECOVERY_NEEDED)
5799		xlog_recover_cancel_intents(log);
5800}
5801
5802#if defined(DEBUG)
5803/*
5804 * Read all of the agf and agi counters and check that they
5805 * are consistent with the superblock counters.
5806 */
5807STATIC void
5808xlog_recover_check_summary(
5809	struct xlog	*log)
5810{
5811	xfs_mount_t	*mp;
5812	xfs_agf_t	*agfp;
5813	xfs_buf_t	*agfbp;
5814	xfs_buf_t	*agibp;
5815	xfs_agnumber_t	agno;
5816	uint64_t	freeblks;
5817	uint64_t	itotal;
5818	uint64_t	ifree;
5819	int		error;
5820
5821	mp = log->l_mp;
5822
5823	freeblks = 0LL;
5824	itotal = 0LL;
5825	ifree = 0LL;
5826	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5827		error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
5828		if (error) {
5829			xfs_alert(mp, "%s agf read failed agno %d error %d",
5830						__func__, agno, error);
5831		} else {
5832			agfp = XFS_BUF_TO_AGF(agfbp);
5833			freeblks += be32_to_cpu(agfp->agf_freeblks) +
5834				    be32_to_cpu(agfp->agf_flcount);
5835			xfs_buf_relse(agfbp);
5836		}
5837
5838		error = xfs_read_agi(mp, NULL, agno, &agibp);
5839		if (error) {
5840			xfs_alert(mp, "%s agi read failed agno %d error %d",
5841						__func__, agno, error);
5842		} else {
5843			struct xfs_agi	*agi = XFS_BUF_TO_AGI(agibp);
5844
5845			itotal += be32_to_cpu(agi->agi_count);
5846			ifree += be32_to_cpu(agi->agi_freecount);
5847			xfs_buf_relse(agibp);
5848		}
5849	}
5850}
5851#endif /* DEBUG */