Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
v3.1
   1/*
   2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3 * All Rights Reserved.
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it would be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write the Free Software Foundation,
  16 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18#include "xfs.h"
  19#include "xfs_fs.h"
  20#include "xfs_types.h"
 
 
 
  21#include "xfs_bit.h"
  22#include "xfs_log.h"
  23#include "xfs_inum.h"
  24#include "xfs_trans.h"
  25#include "xfs_sb.h"
  26#include "xfs_ag.h"
  27#include "xfs_mount.h"
  28#include "xfs_error.h"
  29#include "xfs_bmap_btree.h"
  30#include "xfs_alloc_btree.h"
  31#include "xfs_ialloc_btree.h"
  32#include "xfs_dinode.h"
  33#include "xfs_inode.h"
  34#include "xfs_inode_item.h"
  35#include "xfs_alloc.h"
  36#include "xfs_ialloc.h"
  37#include "xfs_log_priv.h"
  38#include "xfs_buf_item.h"
  39#include "xfs_log_recover.h"
 
  40#include "xfs_extfree_item.h"
  41#include "xfs_trans_priv.h"
 
 
  42#include "xfs_quota.h"
  43#include "xfs_rw.h"
  44#include "xfs_utils.h"
  45#include "xfs_trace.h"
 
 
 
 
 
 
 
  46
  47STATIC int	xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
  48STATIC int	xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
 
 
 
 
 
 
  49#if defined(DEBUG)
  50STATIC void	xlog_recover_check_summary(xlog_t *);
 
 
  51#else
  52#define	xlog_recover_check_summary(log)
  53#endif
  54
  55/*
  56 * This structure is used during recovery to record the buf log items which
  57 * have been canceled and should not be replayed.
  58 */
  59struct xfs_buf_cancel {
  60	xfs_daddr_t		bc_blkno;
  61	uint			bc_len;
  62	int			bc_refcount;
  63	struct list_head	bc_list;
  64};
  65
  66/*
  67 * Sector aligned buffer routines for buffer create/read/write/access
  68 */
  69
  70/*
  71 * Verify the given count of basic blocks is valid number of blocks
  72 * to specify for an operation involving the given XFS log buffer.
  73 * Returns nonzero if the count is valid, 0 otherwise.
  74 */
  75
  76static inline int
  77xlog_buf_bbcount_valid(
  78	xlog_t		*log,
  79	int		bbcount)
  80{
  81	return bbcount > 0 && bbcount <= log->l_logBBsize;
  82}
  83
  84/*
  85 * Allocate a buffer to hold log data.  The buffer needs to be able
  86 * to map to a range of nbblks basic blocks at any valid (basic
  87 * block) offset within the log.
  88 */
  89STATIC xfs_buf_t *
  90xlog_get_bp(
  91	xlog_t		*log,
  92	int		nbblks)
  93{
  94	struct xfs_buf	*bp;
  95
  96	if (!xlog_buf_bbcount_valid(log, nbblks)) {
  97		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
  98			nbblks);
  99		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 100		return NULL;
 101	}
 102
 103	/*
 104	 * We do log I/O in units of log sectors (a power-of-2
 105	 * multiple of the basic block size), so we round up the
 106	 * requested size to accommodate the basic blocks required
 107	 * for complete log sectors.
 108	 *
 109	 * In addition, the buffer may be used for a non-sector-
 110	 * aligned block offset, in which case an I/O of the
 111	 * requested size could extend beyond the end of the
 112	 * buffer.  If the requested size is only 1 basic block it
 113	 * will never straddle a sector boundary, so this won't be
 114	 * an issue.  Nor will this be a problem if the log I/O is
 115	 * done in basic blocks (sector size 1).  But otherwise we
 116	 * extend the buffer by one extra log sector to ensure
 117	 * there's space to accommodate this possibility.
 118	 */
 119	if (nbblks > 1 && log->l_sectBBsize > 1)
 120		nbblks += log->l_sectBBsize;
 121	nbblks = round_up(nbblks, log->l_sectBBsize);
 122
 123	bp = xfs_buf_get_uncached(log->l_mp->m_logdev_targp, BBTOB(nbblks), 0);
 124	if (bp)
 125		xfs_buf_unlock(bp);
 126	return bp;
 127}
 128
 129STATIC void
 130xlog_put_bp(
 131	xfs_buf_t	*bp)
 132{
 133	xfs_buf_free(bp);
 134}
 135
 136/*
 137 * Return the address of the start of the given block number's data
 138 * in a log buffer.  The buffer covers a log sector-aligned region.
 139 */
 140STATIC xfs_caddr_t
 141xlog_align(
 142	xlog_t		*log,
 143	xfs_daddr_t	blk_no,
 144	int		nbblks,
 145	xfs_buf_t	*bp)
 146{
 147	xfs_daddr_t	offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
 148
 149	ASSERT(BBTOB(offset + nbblks) <= XFS_BUF_SIZE(bp));
 150	return bp->b_addr + BBTOB(offset);
 151}
 152
 153
 154/*
 155 * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
 156 */
 157STATIC int
 158xlog_bread_noalign(
 159	xlog_t		*log,
 160	xfs_daddr_t	blk_no,
 161	int		nbblks,
 162	xfs_buf_t	*bp)
 163{
 164	int		error;
 165
 166	if (!xlog_buf_bbcount_valid(log, nbblks)) {
 167		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 168			nbblks);
 169		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 170		return EFSCORRUPTED;
 171	}
 172
 173	blk_no = round_down(blk_no, log->l_sectBBsize);
 174	nbblks = round_up(nbblks, log->l_sectBBsize);
 175
 176	ASSERT(nbblks > 0);
 177	ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
 178
 179	XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
 180	XFS_BUF_READ(bp);
 181	XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
 
 
 
 
 182
 183	xfsbdstrat(log->l_mp, bp);
 184	error = xfs_buf_iowait(bp);
 185	if (error)
 186		xfs_ioerror_alert("xlog_bread", log->l_mp,
 187				  bp, XFS_BUF_ADDR(bp));
 188	return error;
 189}
 190
 191STATIC int
 192xlog_bread(
 193	xlog_t		*log,
 194	xfs_daddr_t	blk_no,
 195	int		nbblks,
 196	xfs_buf_t	*bp,
 197	xfs_caddr_t	*offset)
 198{
 199	int		error;
 200
 201	error = xlog_bread_noalign(log, blk_no, nbblks, bp);
 202	if (error)
 203		return error;
 204
 205	*offset = xlog_align(log, blk_no, nbblks, bp);
 206	return 0;
 207}
 208
 209/*
 210 * Read at an offset into the buffer. Returns with the buffer in it's original
 211 * state regardless of the result of the read.
 212 */
 213STATIC int
 214xlog_bread_offset(
 215	xlog_t		*log,
 216	xfs_daddr_t	blk_no,		/* block to read from */
 217	int		nbblks,		/* blocks to read */
 218	xfs_buf_t	*bp,
 219	xfs_caddr_t	offset)
 220{
 221	xfs_caddr_t	orig_offset = bp->b_addr;
 222	int		orig_len = bp->b_buffer_length;
 223	int		error, error2;
 224
 225	error = xfs_buf_associate_memory(bp, offset, BBTOB(nbblks));
 226	if (error)
 227		return error;
 228
 229	error = xlog_bread_noalign(log, blk_no, nbblks, bp);
 230
 231	/* must reset buffer pointer even on error */
 232	error2 = xfs_buf_associate_memory(bp, orig_offset, orig_len);
 233	if (error)
 234		return error;
 235	return error2;
 236}
 237
 238/*
 239 * Write out the buffer at the given block for the given number of blocks.
 240 * The buffer is kept locked across the write and is returned locked.
 241 * This can only be used for synchronous log writes.
 242 */
 243STATIC int
 244xlog_bwrite(
 245	xlog_t		*log,
 246	xfs_daddr_t	blk_no,
 247	int		nbblks,
 248	xfs_buf_t	*bp)
 249{
 250	int		error;
 251
 252	if (!xlog_buf_bbcount_valid(log, nbblks)) {
 253		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 254			nbblks);
 255		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 256		return EFSCORRUPTED;
 257	}
 258
 259	blk_no = round_down(blk_no, log->l_sectBBsize);
 260	nbblks = round_up(nbblks, log->l_sectBBsize);
 261
 262	ASSERT(nbblks > 0);
 263	ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
 264
 265	XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
 266	XFS_BUF_ZEROFLAGS(bp);
 267	xfs_buf_hold(bp);
 268	xfs_buf_lock(bp);
 269	XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
 
 270
 271	if ((error = xfs_bwrite(log->l_mp, bp)))
 272		xfs_ioerror_alert("xlog_bwrite", log->l_mp,
 273				  bp, XFS_BUF_ADDR(bp));
 
 274	return error;
 275}
 276
 277#ifdef DEBUG
 278/*
 279 * dump debug superblock and log record information
 280 */
 281STATIC void
 282xlog_header_check_dump(
 283	xfs_mount_t		*mp,
 284	xlog_rec_header_t	*head)
 285{
 286	xfs_debug(mp, "%s:  SB : uuid = %pU, fmt = %d\n",
 287		__func__, &mp->m_sb.sb_uuid, XLOG_FMT);
 288	xfs_debug(mp, "    log : uuid = %pU, fmt = %d\n",
 289		&head->h_fs_uuid, be32_to_cpu(head->h_fmt));
 290}
 291#else
 292#define xlog_header_check_dump(mp, head)
 293#endif
 294
 295/*
 296 * check log record header for recovery
 297 */
 298STATIC int
 299xlog_header_check_recover(
 300	xfs_mount_t		*mp,
 301	xlog_rec_header_t	*head)
 302{
 303	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 304
 305	/*
 306	 * IRIX doesn't write the h_fmt field and leaves it zeroed
 307	 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
 308	 * a dirty log created in IRIX.
 309	 */
 310	if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
 311		xfs_warn(mp,
 312	"dirty log written in incompatible format - can't recover");
 313		xlog_header_check_dump(mp, head);
 314		XFS_ERROR_REPORT("xlog_header_check_recover(1)",
 315				 XFS_ERRLEVEL_HIGH, mp);
 316		return XFS_ERROR(EFSCORRUPTED);
 317	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 318		xfs_warn(mp,
 319	"dirty log entry has mismatched uuid - can't recover");
 320		xlog_header_check_dump(mp, head);
 321		XFS_ERROR_REPORT("xlog_header_check_recover(2)",
 322				 XFS_ERRLEVEL_HIGH, mp);
 323		return XFS_ERROR(EFSCORRUPTED);
 324	}
 325	return 0;
 326}
 327
 328/*
 329 * read the head block of the log and check the header
 330 */
 331STATIC int
 332xlog_header_check_mount(
 333	xfs_mount_t		*mp,
 334	xlog_rec_header_t	*head)
 335{
 336	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 337
 338	if (uuid_is_nil(&head->h_fs_uuid)) {
 339		/*
 340		 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
 341		 * h_fs_uuid is nil, we assume this log was last mounted
 342		 * by IRIX and continue.
 343		 */
 344		xfs_warn(mp, "nil uuid in log - IRIX style log");
 345	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 346		xfs_warn(mp, "log has mismatched uuid - can't recover");
 347		xlog_header_check_dump(mp, head);
 348		XFS_ERROR_REPORT("xlog_header_check_mount",
 349				 XFS_ERRLEVEL_HIGH, mp);
 350		return XFS_ERROR(EFSCORRUPTED);
 351	}
 352	return 0;
 353}
 354
 355STATIC void
 356xlog_recover_iodone(
 357	struct xfs_buf	*bp)
 358{
 359	if (bp->b_error) {
 360		/*
 361		 * We're not going to bother about retrying
 362		 * this during recovery. One strike!
 363		 */
 364		xfs_ioerror_alert("xlog_recover_iodone",
 365					bp->b_target->bt_mount, bp,
 366					XFS_BUF_ADDR(bp));
 367		xfs_force_shutdown(bp->b_target->bt_mount,
 368					SHUTDOWN_META_IO_ERROR);
 369	}
 370	bp->b_iodone = NULL;
 371	xfs_buf_ioend(bp, 0);
 372}
 373
 374/*
 375 * This routine finds (to an approximation) the first block in the physical
 376 * log which contains the given cycle.  It uses a binary search algorithm.
 377 * Note that the algorithm can not be perfect because the disk will not
 378 * necessarily be perfect.
 379 */
 380STATIC int
 381xlog_find_cycle_start(
 382	xlog_t		*log,
 383	xfs_buf_t	*bp,
 384	xfs_daddr_t	first_blk,
 385	xfs_daddr_t	*last_blk,
 386	uint		cycle)
 387{
 388	xfs_caddr_t	offset;
 389	xfs_daddr_t	mid_blk;
 390	xfs_daddr_t	end_blk;
 391	uint		mid_cycle;
 392	int		error;
 393
 394	end_blk = *last_blk;
 395	mid_blk = BLK_AVG(first_blk, end_blk);
 396	while (mid_blk != first_blk && mid_blk != end_blk) {
 397		error = xlog_bread(log, mid_blk, 1, bp, &offset);
 398		if (error)
 399			return error;
 400		mid_cycle = xlog_get_cycle(offset);
 401		if (mid_cycle == cycle)
 402			end_blk = mid_blk;   /* last_half_cycle == mid_cycle */
 403		else
 404			first_blk = mid_blk; /* first_half_cycle == mid_cycle */
 405		mid_blk = BLK_AVG(first_blk, end_blk);
 406	}
 407	ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
 408	       (mid_blk == end_blk && mid_blk-1 == first_blk));
 409
 410	*last_blk = end_blk;
 411
 412	return 0;
 413}
 414
 415/*
 416 * Check that a range of blocks does not contain stop_on_cycle_no.
 417 * Fill in *new_blk with the block offset where such a block is
 418 * found, or with -1 (an invalid block number) if there is no such
 419 * block in the range.  The scan needs to occur from front to back
 420 * and the pointer into the region must be updated since a later
 421 * routine will need to perform another test.
 422 */
 423STATIC int
 424xlog_find_verify_cycle(
 425	xlog_t		*log,
 426	xfs_daddr_t	start_blk,
 427	int		nbblks,
 428	uint		stop_on_cycle_no,
 429	xfs_daddr_t	*new_blk)
 430{
 431	xfs_daddr_t	i, j;
 432	uint		cycle;
 433	xfs_buf_t	*bp;
 434	xfs_daddr_t	bufblks;
 435	xfs_caddr_t	buf = NULL;
 436	int		error = 0;
 437
 438	/*
 439	 * Greedily allocate a buffer big enough to handle the full
 440	 * range of basic blocks we'll be examining.  If that fails,
 441	 * try a smaller size.  We need to be able to read at least
 442	 * a log sector, or we're out of luck.
 443	 */
 444	bufblks = 1 << ffs(nbblks);
 
 
 445	while (!(bp = xlog_get_bp(log, bufblks))) {
 446		bufblks >>= 1;
 447		if (bufblks < log->l_sectBBsize)
 448			return ENOMEM;
 449	}
 450
 451	for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
 452		int	bcount;
 453
 454		bcount = min(bufblks, (start_blk + nbblks - i));
 455
 456		error = xlog_bread(log, i, bcount, bp, &buf);
 457		if (error)
 458			goto out;
 459
 460		for (j = 0; j < bcount; j++) {
 461			cycle = xlog_get_cycle(buf);
 462			if (cycle == stop_on_cycle_no) {
 463				*new_blk = i+j;
 464				goto out;
 465			}
 466
 467			buf += BBSIZE;
 468		}
 469	}
 470
 471	*new_blk = -1;
 472
 473out:
 474	xlog_put_bp(bp);
 475	return error;
 476}
 477
 478/*
 479 * Potentially backup over partial log record write.
 480 *
 481 * In the typical case, last_blk is the number of the block directly after
 482 * a good log record.  Therefore, we subtract one to get the block number
 483 * of the last block in the given buffer.  extra_bblks contains the number
 484 * of blocks we would have read on a previous read.  This happens when the
 485 * last log record is split over the end of the physical log.
 486 *
 487 * extra_bblks is the number of blocks potentially verified on a previous
 488 * call to this routine.
 489 */
 490STATIC int
 491xlog_find_verify_log_record(
 492	xlog_t			*log,
 493	xfs_daddr_t		start_blk,
 494	xfs_daddr_t		*last_blk,
 495	int			extra_bblks)
 496{
 497	xfs_daddr_t		i;
 498	xfs_buf_t		*bp;
 499	xfs_caddr_t		offset = NULL;
 500	xlog_rec_header_t	*head = NULL;
 501	int			error = 0;
 502	int			smallmem = 0;
 503	int			num_blks = *last_blk - start_blk;
 504	int			xhdrs;
 505
 506	ASSERT(start_blk != 0 || *last_blk != start_blk);
 507
 508	if (!(bp = xlog_get_bp(log, num_blks))) {
 509		if (!(bp = xlog_get_bp(log, 1)))
 510			return ENOMEM;
 511		smallmem = 1;
 512	} else {
 513		error = xlog_bread(log, start_blk, num_blks, bp, &offset);
 514		if (error)
 515			goto out;
 516		offset += ((num_blks - 1) << BBSHIFT);
 517	}
 518
 519	for (i = (*last_blk) - 1; i >= 0; i--) {
 520		if (i < start_blk) {
 521			/* valid log record not found */
 522			xfs_warn(log->l_mp,
 523		"Log inconsistent (didn't find previous header)");
 524			ASSERT(0);
 525			error = XFS_ERROR(EIO);
 526			goto out;
 527		}
 528
 529		if (smallmem) {
 530			error = xlog_bread(log, i, 1, bp, &offset);
 531			if (error)
 532				goto out;
 533		}
 534
 535		head = (xlog_rec_header_t *)offset;
 536
 537		if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
 538			break;
 539
 540		if (!smallmem)
 541			offset -= BBSIZE;
 542	}
 543
 544	/*
 545	 * We hit the beginning of the physical log & still no header.  Return
 546	 * to caller.  If caller can handle a return of -1, then this routine
 547	 * will be called again for the end of the physical log.
 548	 */
 549	if (i == -1) {
 550		error = -1;
 551		goto out;
 552	}
 553
 554	/*
 555	 * We have the final block of the good log (the first block
 556	 * of the log record _before_ the head. So we check the uuid.
 557	 */
 558	if ((error = xlog_header_check_mount(log->l_mp, head)))
 559		goto out;
 560
 561	/*
 562	 * We may have found a log record header before we expected one.
 563	 * last_blk will be the 1st block # with a given cycle #.  We may end
 564	 * up reading an entire log record.  In this case, we don't want to
 565	 * reset last_blk.  Only when last_blk points in the middle of a log
 566	 * record do we update last_blk.
 567	 */
 568	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
 569		uint	h_size = be32_to_cpu(head->h_size);
 570
 571		xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
 572		if (h_size % XLOG_HEADER_CYCLE_SIZE)
 573			xhdrs++;
 574	} else {
 575		xhdrs = 1;
 576	}
 577
 578	if (*last_blk - i + extra_bblks !=
 579	    BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
 580		*last_blk = i;
 581
 582out:
 583	xlog_put_bp(bp);
 584	return error;
 585}
 586
 587/*
 588 * Head is defined to be the point of the log where the next log write
 589 * write could go.  This means that incomplete LR writes at the end are
 590 * eliminated when calculating the head.  We aren't guaranteed that previous
 591 * LR have complete transactions.  We only know that a cycle number of
 592 * current cycle number -1 won't be present in the log if we start writing
 593 * from our current block number.
 594 *
 595 * last_blk contains the block number of the first block with a given
 596 * cycle number.
 597 *
 598 * Return: zero if normal, non-zero if error.
 599 */
 600STATIC int
 601xlog_find_head(
 602	xlog_t 		*log,
 603	xfs_daddr_t	*return_head_blk)
 604{
 605	xfs_buf_t	*bp;
 606	xfs_caddr_t	offset;
 607	xfs_daddr_t	new_blk, first_blk, start_blk, last_blk, head_blk;
 608	int		num_scan_bblks;
 609	uint		first_half_cycle, last_half_cycle;
 610	uint		stop_on_cycle;
 611	int		error, log_bbnum = log->l_logBBsize;
 612
 613	/* Is the end of the log device zeroed? */
 614	if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
 615		*return_head_blk = first_blk;
 616
 617		/* Is the whole lot zeroed? */
 618		if (!first_blk) {
 619			/* Linux XFS shouldn't generate totally zeroed logs -
 620			 * mkfs etc write a dummy unmount record to a fresh
 621			 * log so we can store the uuid in there
 622			 */
 623			xfs_warn(log->l_mp, "totally zeroed log");
 624		}
 625
 626		return 0;
 627	} else if (error) {
 628		xfs_warn(log->l_mp, "empty log check failed");
 629		return error;
 630	}
 631
 632	first_blk = 0;			/* get cycle # of 1st block */
 633	bp = xlog_get_bp(log, 1);
 634	if (!bp)
 635		return ENOMEM;
 636
 637	error = xlog_bread(log, 0, 1, bp, &offset);
 638	if (error)
 639		goto bp_err;
 640
 641	first_half_cycle = xlog_get_cycle(offset);
 642
 643	last_blk = head_blk = log_bbnum - 1;	/* get cycle # of last block */
 644	error = xlog_bread(log, last_blk, 1, bp, &offset);
 645	if (error)
 646		goto bp_err;
 647
 648	last_half_cycle = xlog_get_cycle(offset);
 649	ASSERT(last_half_cycle != 0);
 650
 651	/*
 652	 * If the 1st half cycle number is equal to the last half cycle number,
 653	 * then the entire log is stamped with the same cycle number.  In this
 654	 * case, head_blk can't be set to zero (which makes sense).  The below
 655	 * math doesn't work out properly with head_blk equal to zero.  Instead,
 656	 * we set it to log_bbnum which is an invalid block number, but this
 657	 * value makes the math correct.  If head_blk doesn't changed through
 658	 * all the tests below, *head_blk is set to zero at the very end rather
 659	 * than log_bbnum.  In a sense, log_bbnum and zero are the same block
 660	 * in a circular file.
 661	 */
 662	if (first_half_cycle == last_half_cycle) {
 663		/*
 664		 * In this case we believe that the entire log should have
 665		 * cycle number last_half_cycle.  We need to scan backwards
 666		 * from the end verifying that there are no holes still
 667		 * containing last_half_cycle - 1.  If we find such a hole,
 668		 * then the start of that hole will be the new head.  The
 669		 * simple case looks like
 670		 *        x | x ... | x - 1 | x
 671		 * Another case that fits this picture would be
 672		 *        x | x + 1 | x ... | x
 673		 * In this case the head really is somewhere at the end of the
 674		 * log, as one of the latest writes at the beginning was
 675		 * incomplete.
 676		 * One more case is
 677		 *        x | x + 1 | x ... | x - 1 | x
 678		 * This is really the combination of the above two cases, and
 679		 * the head has to end up at the start of the x-1 hole at the
 680		 * end of the log.
 681		 *
 682		 * In the 256k log case, we will read from the beginning to the
 683		 * end of the log and search for cycle numbers equal to x-1.
 684		 * We don't worry about the x+1 blocks that we encounter,
 685		 * because we know that they cannot be the head since the log
 686		 * started with x.
 687		 */
 688		head_blk = log_bbnum;
 689		stop_on_cycle = last_half_cycle - 1;
 690	} else {
 691		/*
 692		 * In this case we want to find the first block with cycle
 693		 * number matching last_half_cycle.  We expect the log to be
 694		 * some variation on
 695		 *        x + 1 ... | x ... | x
 696		 * The first block with cycle number x (last_half_cycle) will
 697		 * be where the new head belongs.  First we do a binary search
 698		 * for the first occurrence of last_half_cycle.  The binary
 699		 * search may not be totally accurate, so then we scan back
 700		 * from there looking for occurrences of last_half_cycle before
 701		 * us.  If that backwards scan wraps around the beginning of
 702		 * the log, then we look for occurrences of last_half_cycle - 1
 703		 * at the end of the log.  The cases we're looking for look
 704		 * like
 705		 *                               v binary search stopped here
 706		 *        x + 1 ... | x | x + 1 | x ... | x
 707		 *                   ^ but we want to locate this spot
 708		 * or
 709		 *        <---------> less than scan distance
 710		 *        x + 1 ... | x ... | x - 1 | x
 711		 *                           ^ we want to locate this spot
 712		 */
 713		stop_on_cycle = last_half_cycle;
 714		if ((error = xlog_find_cycle_start(log, bp, first_blk,
 715						&head_blk, last_half_cycle)))
 716			goto bp_err;
 717	}
 718
 719	/*
 720	 * Now validate the answer.  Scan back some number of maximum possible
 721	 * blocks and make sure each one has the expected cycle number.  The
 722	 * maximum is determined by the total possible amount of buffering
 723	 * in the in-core log.  The following number can be made tighter if
 724	 * we actually look at the block size of the filesystem.
 725	 */
 726	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
 727	if (head_blk >= num_scan_bblks) {
 728		/*
 729		 * We are guaranteed that the entire check can be performed
 730		 * in one buffer.
 731		 */
 732		start_blk = head_blk - num_scan_bblks;
 733		if ((error = xlog_find_verify_cycle(log,
 734						start_blk, num_scan_bblks,
 735						stop_on_cycle, &new_blk)))
 736			goto bp_err;
 737		if (new_blk != -1)
 738			head_blk = new_blk;
 739	} else {		/* need to read 2 parts of log */
 740		/*
 741		 * We are going to scan backwards in the log in two parts.
 742		 * First we scan the physical end of the log.  In this part
 743		 * of the log, we are looking for blocks with cycle number
 744		 * last_half_cycle - 1.
 745		 * If we find one, then we know that the log starts there, as
 746		 * we've found a hole that didn't get written in going around
 747		 * the end of the physical log.  The simple case for this is
 748		 *        x + 1 ... | x ... | x - 1 | x
 749		 *        <---------> less than scan distance
 750		 * If all of the blocks at the end of the log have cycle number
 751		 * last_half_cycle, then we check the blocks at the start of
 752		 * the log looking for occurrences of last_half_cycle.  If we
 753		 * find one, then our current estimate for the location of the
 754		 * first occurrence of last_half_cycle is wrong and we move
 755		 * back to the hole we've found.  This case looks like
 756		 *        x + 1 ... | x | x + 1 | x ...
 757		 *                               ^ binary search stopped here
 758		 * Another case we need to handle that only occurs in 256k
 759		 * logs is
 760		 *        x + 1 ... | x ... | x+1 | x ...
 761		 *                   ^ binary search stops here
 762		 * In a 256k log, the scan at the end of the log will see the
 763		 * x + 1 blocks.  We need to skip past those since that is
 764		 * certainly not the head of the log.  By searching for
 765		 * last_half_cycle-1 we accomplish that.
 766		 */
 767		ASSERT(head_blk <= INT_MAX &&
 768			(xfs_daddr_t) num_scan_bblks >= head_blk);
 769		start_blk = log_bbnum - (num_scan_bblks - head_blk);
 770		if ((error = xlog_find_verify_cycle(log, start_blk,
 771					num_scan_bblks - (int)head_blk,
 772					(stop_on_cycle - 1), &new_blk)))
 773			goto bp_err;
 774		if (new_blk != -1) {
 775			head_blk = new_blk;
 776			goto validate_head;
 777		}
 778
 779		/*
 780		 * Scan beginning of log now.  The last part of the physical
 781		 * log is good.  This scan needs to verify that it doesn't find
 782		 * the last_half_cycle.
 783		 */
 784		start_blk = 0;
 785		ASSERT(head_blk <= INT_MAX);
 786		if ((error = xlog_find_verify_cycle(log,
 787					start_blk, (int)head_blk,
 788					stop_on_cycle, &new_blk)))
 789			goto bp_err;
 790		if (new_blk != -1)
 791			head_blk = new_blk;
 792	}
 793
 794validate_head:
 795	/*
 796	 * Now we need to make sure head_blk is not pointing to a block in
 797	 * the middle of a log record.
 798	 */
 799	num_scan_bblks = XLOG_REC_SHIFT(log);
 800	if (head_blk >= num_scan_bblks) {
 801		start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
 802
 803		/* start ptr at last block ptr before head_blk */
 804		if ((error = xlog_find_verify_log_record(log, start_blk,
 805							&head_blk, 0)) == -1) {
 806			error = XFS_ERROR(EIO);
 807			goto bp_err;
 808		} else if (error)
 809			goto bp_err;
 810	} else {
 811		start_blk = 0;
 812		ASSERT(head_blk <= INT_MAX);
 813		if ((error = xlog_find_verify_log_record(log, start_blk,
 814							&head_blk, 0)) == -1) {
 815			/* We hit the beginning of the log during our search */
 816			start_blk = log_bbnum - (num_scan_bblks - head_blk);
 817			new_blk = log_bbnum;
 818			ASSERT(start_blk <= INT_MAX &&
 819				(xfs_daddr_t) log_bbnum-start_blk >= 0);
 820			ASSERT(head_blk <= INT_MAX);
 821			if ((error = xlog_find_verify_log_record(log,
 822							start_blk, &new_blk,
 823							(int)head_blk)) == -1) {
 824				error = XFS_ERROR(EIO);
 825				goto bp_err;
 826			} else if (error)
 827				goto bp_err;
 828			if (new_blk != log_bbnum)
 829				head_blk = new_blk;
 830		} else if (error)
 831			goto bp_err;
 832	}
 833
 834	xlog_put_bp(bp);
 835	if (head_blk == log_bbnum)
 836		*return_head_blk = 0;
 837	else
 838		*return_head_blk = head_blk;
 839	/*
 840	 * When returning here, we have a good block number.  Bad block
 841	 * means that during a previous crash, we didn't have a clean break
 842	 * from cycle number N to cycle number N-1.  In this case, we need
 843	 * to find the first block with cycle number N-1.
 844	 */
 845	return 0;
 846
 847 bp_err:
 848	xlog_put_bp(bp);
 849
 850	if (error)
 851		xfs_warn(log->l_mp, "failed to find log head");
 852	return error;
 853}
 854
 855/*
 856 * Find the sync block number or the tail of the log.
 857 *
 858 * This will be the block number of the last record to have its
 859 * associated buffers synced to disk.  Every log record header has
 860 * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
 861 * to get a sync block number.  The only concern is to figure out which
 862 * log record header to believe.
 863 *
 864 * The following algorithm uses the log record header with the largest
 865 * lsn.  The entire log record does not need to be valid.  We only care
 866 * that the header is valid.
 867 *
 868 * We could speed up search by using current head_blk buffer, but it is not
 869 * available.
 870 */
 871STATIC int
 872xlog_find_tail(
 873	xlog_t			*log,
 874	xfs_daddr_t		*head_blk,
 875	xfs_daddr_t		*tail_blk)
 876{
 877	xlog_rec_header_t	*rhead;
 878	xlog_op_header_t	*op_head;
 879	xfs_caddr_t		offset = NULL;
 880	xfs_buf_t		*bp;
 881	int			error, i, found;
 882	xfs_daddr_t		umount_data_blk;
 883	xfs_daddr_t		after_umount_blk;
 884	xfs_lsn_t		tail_lsn;
 885	int			hblks;
 886
 887	found = 0;
 888
 889	/*
 890	 * Find previous log record
 891	 */
 892	if ((error = xlog_find_head(log, head_blk)))
 893		return error;
 894
 895	bp = xlog_get_bp(log, 1);
 896	if (!bp)
 897		return ENOMEM;
 898	if (*head_blk == 0) {				/* special case */
 899		error = xlog_bread(log, 0, 1, bp, &offset);
 900		if (error)
 901			goto done;
 902
 903		if (xlog_get_cycle(offset) == 0) {
 904			*tail_blk = 0;
 905			/* leave all other log inited values alone */
 906			goto done;
 907		}
 908	}
 909
 910	/*
 911	 * Search backwards looking for log record header block
 912	 */
 913	ASSERT(*head_blk < INT_MAX);
 914	for (i = (int)(*head_blk) - 1; i >= 0; i--) {
 915		error = xlog_bread(log, i, 1, bp, &offset);
 916		if (error)
 917			goto done;
 918
 919		if (*(__be32 *)offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 920			found = 1;
 921			break;
 922		}
 923	}
 924	/*
 925	 * If we haven't found the log record header block, start looking
 926	 * again from the end of the physical log.  XXXmiken: There should be
 927	 * a check here to make sure we didn't search more than N blocks in
 928	 * the previous code.
 929	 */
 930	if (!found) {
 931		for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
 932			error = xlog_bread(log, i, 1, bp, &offset);
 933			if (error)
 934				goto done;
 935
 936			if (*(__be32 *)offset ==
 937			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 938				found = 2;
 939				break;
 940			}
 941		}
 942	}
 943	if (!found) {
 944		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
 
 945		ASSERT(0);
 946		return XFS_ERROR(EIO);
 947	}
 948
 949	/* find blk_no of tail of log */
 950	rhead = (xlog_rec_header_t *)offset;
 951	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
 952
 953	/*
 954	 * Reset log values according to the state of the log when we
 955	 * crashed.  In the case where head_blk == 0, we bump curr_cycle
 956	 * one because the next write starts a new cycle rather than
 957	 * continuing the cycle of the last good log record.  At this
 958	 * point we have guaranteed that all partial log records have been
 959	 * accounted for.  Therefore, we know that the last good log record
 960	 * written was complete and ended exactly on the end boundary
 961	 * of the physical log.
 962	 */
 963	log->l_prev_block = i;
 964	log->l_curr_block = (int)*head_blk;
 965	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
 966	if (found == 2)
 967		log->l_curr_cycle++;
 968	atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
 969	atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
 970	xlog_assign_grant_head(&log->l_grant_reserve_head, log->l_curr_cycle,
 971					BBTOB(log->l_curr_block));
 972	xlog_assign_grant_head(&log->l_grant_write_head, log->l_curr_cycle,
 973					BBTOB(log->l_curr_block));
 974
 975	/*
 976	 * Look for unmount record.  If we find it, then we know there
 977	 * was a clean unmount.  Since 'i' could be the last block in
 978	 * the physical log, we convert to a log block before comparing
 979	 * to the head_blk.
 980	 *
 981	 * Save the current tail lsn to use to pass to
 982	 * xlog_clear_stale_blocks() below.  We won't want to clear the
 983	 * unmount record if there is one, so we pass the lsn of the
 984	 * unmount record rather than the block after it.
 985	 */
 986	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
 987		int	h_size = be32_to_cpu(rhead->h_size);
 988		int	h_version = be32_to_cpu(rhead->h_version);
 989
 990		if ((h_version & XLOG_VERSION_2) &&
 991		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
 992			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
 993			if (h_size % XLOG_HEADER_CYCLE_SIZE)
 994				hblks++;
 995		} else {
 996			hblks = 1;
 997		}
 998	} else {
 999		hblks = 1;
1000	}
1001	after_umount_blk = (i + hblks + (int)
1002		BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
1003	tail_lsn = atomic64_read(&log->l_tail_lsn);
1004	if (*head_blk == after_umount_blk &&
1005	    be32_to_cpu(rhead->h_num_logops) == 1) {
1006		umount_data_blk = (i + hblks) % log->l_logBBsize;
1007		error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
1008		if (error)
1009			goto done;
1010
1011		op_head = (xlog_op_header_t *)offset;
1012		if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1013			/*
1014			 * Set tail and last sync so that newly written
1015			 * log records will point recovery to after the
1016			 * current unmount record.
1017			 */
1018			xlog_assign_atomic_lsn(&log->l_tail_lsn,
1019					log->l_curr_cycle, after_umount_blk);
1020			xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1021					log->l_curr_cycle, after_umount_blk);
1022			*tail_blk = after_umount_blk;
1023
1024			/*
1025			 * Note that the unmount was clean. If the unmount
1026			 * was not clean, we need to know this to rebuild the
1027			 * superblock counters from the perag headers if we
1028			 * have a filesystem using non-persistent counters.
1029			 */
1030			log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1031		}
1032	}
1033
1034	/*
1035	 * Make sure that there are no blocks in front of the head
1036	 * with the same cycle number as the head.  This can happen
1037	 * because we allow multiple outstanding log writes concurrently,
1038	 * and the later writes might make it out before earlier ones.
1039	 *
1040	 * We use the lsn from before modifying it so that we'll never
1041	 * overwrite the unmount record after a clean unmount.
1042	 *
1043	 * Do this only if we are going to recover the filesystem
1044	 *
1045	 * NOTE: This used to say "if (!readonly)"
1046	 * However on Linux, we can & do recover a read-only filesystem.
1047	 * We only skip recovery if NORECOVERY is specified on mount,
1048	 * in which case we would not be here.
1049	 *
1050	 * But... if the -device- itself is readonly, just skip this.
1051	 * We can't recover this device anyway, so it won't matter.
1052	 */
1053	if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
1054		error = xlog_clear_stale_blocks(log, tail_lsn);
1055
1056done:
1057	xlog_put_bp(bp);
1058
1059	if (error)
1060		xfs_warn(log->l_mp, "failed to locate log tail");
1061	return error;
1062}
1063
1064/*
1065 * Is the log zeroed at all?
1066 *
1067 * The last binary search should be changed to perform an X block read
1068 * once X becomes small enough.  You can then search linearly through
1069 * the X blocks.  This will cut down on the number of reads we need to do.
1070 *
1071 * If the log is partially zeroed, this routine will pass back the blkno
1072 * of the first block with cycle number 0.  It won't have a complete LR
1073 * preceding it.
1074 *
1075 * Return:
1076 *	0  => the log is completely written to
1077 *	-1 => use *blk_no as the first block of the log
1078 *	>0 => error has occurred
1079 */
1080STATIC int
1081xlog_find_zeroed(
1082	xlog_t		*log,
1083	xfs_daddr_t	*blk_no)
1084{
1085	xfs_buf_t	*bp;
1086	xfs_caddr_t	offset;
1087	uint	        first_cycle, last_cycle;
1088	xfs_daddr_t	new_blk, last_blk, start_blk;
1089	xfs_daddr_t     num_scan_bblks;
1090	int	        error, log_bbnum = log->l_logBBsize;
1091
1092	*blk_no = 0;
1093
1094	/* check totally zeroed log */
1095	bp = xlog_get_bp(log, 1);
1096	if (!bp)
1097		return ENOMEM;
1098	error = xlog_bread(log, 0, 1, bp, &offset);
1099	if (error)
1100		goto bp_err;
1101
1102	first_cycle = xlog_get_cycle(offset);
1103	if (first_cycle == 0) {		/* completely zeroed log */
1104		*blk_no = 0;
1105		xlog_put_bp(bp);
1106		return -1;
1107	}
1108
1109	/* check partially zeroed log */
1110	error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1111	if (error)
1112		goto bp_err;
1113
1114	last_cycle = xlog_get_cycle(offset);
1115	if (last_cycle != 0) {		/* log completely written to */
1116		xlog_put_bp(bp);
1117		return 0;
1118	} else if (first_cycle != 1) {
1119		/*
1120		 * If the cycle of the last block is zero, the cycle of
1121		 * the first block must be 1. If it's not, maybe we're
1122		 * not looking at a log... Bail out.
1123		 */
1124		xfs_warn(log->l_mp,
1125			"Log inconsistent or not a log (last==0, first!=1)");
1126		return XFS_ERROR(EINVAL);
 
1127	}
1128
1129	/* we have a partially zeroed log */
1130	last_blk = log_bbnum-1;
1131	if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1132		goto bp_err;
1133
1134	/*
1135	 * Validate the answer.  Because there is no way to guarantee that
1136	 * the entire log is made up of log records which are the same size,
1137	 * we scan over the defined maximum blocks.  At this point, the maximum
1138	 * is not chosen to mean anything special.   XXXmiken
1139	 */
1140	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1141	ASSERT(num_scan_bblks <= INT_MAX);
1142
1143	if (last_blk < num_scan_bblks)
1144		num_scan_bblks = last_blk;
1145	start_blk = last_blk - num_scan_bblks;
1146
1147	/*
1148	 * We search for any instances of cycle number 0 that occur before
1149	 * our current estimate of the head.  What we're trying to detect is
1150	 *        1 ... | 0 | 1 | 0...
1151	 *                       ^ binary search ends here
1152	 */
1153	if ((error = xlog_find_verify_cycle(log, start_blk,
1154					 (int)num_scan_bblks, 0, &new_blk)))
1155		goto bp_err;
1156	if (new_blk != -1)
1157		last_blk = new_blk;
1158
1159	/*
1160	 * Potentially backup over partial log record write.  We don't need
1161	 * to search the end of the log because we know it is zero.
1162	 */
1163	if ((error = xlog_find_verify_log_record(log, start_blk,
1164				&last_blk, 0)) == -1) {
1165	    error = XFS_ERROR(EIO);
1166	    goto bp_err;
1167	} else if (error)
1168	    goto bp_err;
1169
1170	*blk_no = last_blk;
1171bp_err:
1172	xlog_put_bp(bp);
1173	if (error)
1174		return error;
1175	return -1;
1176}
1177
1178/*
1179 * These are simple subroutines used by xlog_clear_stale_blocks() below
1180 * to initialize a buffer full of empty log record headers and write
1181 * them into the log.
1182 */
1183STATIC void
1184xlog_add_record(
1185	xlog_t			*log,
1186	xfs_caddr_t		buf,
1187	int			cycle,
1188	int			block,
1189	int			tail_cycle,
1190	int			tail_block)
1191{
1192	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1193
1194	memset(buf, 0, BBSIZE);
1195	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1196	recp->h_cycle = cpu_to_be32(cycle);
1197	recp->h_version = cpu_to_be32(
1198			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1199	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1200	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1201	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1202	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1203}
1204
1205STATIC int
1206xlog_write_log_records(
1207	xlog_t		*log,
1208	int		cycle,
1209	int		start_block,
1210	int		blocks,
1211	int		tail_cycle,
1212	int		tail_block)
1213{
1214	xfs_caddr_t	offset;
1215	xfs_buf_t	*bp;
1216	int		balign, ealign;
1217	int		sectbb = log->l_sectBBsize;
1218	int		end_block = start_block + blocks;
1219	int		bufblks;
1220	int		error = 0;
1221	int		i, j = 0;
1222
1223	/*
1224	 * Greedily allocate a buffer big enough to handle the full
1225	 * range of basic blocks to be written.  If that fails, try
1226	 * a smaller size.  We need to be able to write at least a
1227	 * log sector, or we're out of luck.
1228	 */
1229	bufblks = 1 << ffs(blocks);
 
 
1230	while (!(bp = xlog_get_bp(log, bufblks))) {
1231		bufblks >>= 1;
1232		if (bufblks < sectbb)
1233			return ENOMEM;
1234	}
1235
1236	/* We may need to do a read at the start to fill in part of
1237	 * the buffer in the starting sector not covered by the first
1238	 * write below.
1239	 */
1240	balign = round_down(start_block, sectbb);
1241	if (balign != start_block) {
1242		error = xlog_bread_noalign(log, start_block, 1, bp);
1243		if (error)
1244			goto out_put_bp;
1245
1246		j = start_block - balign;
1247	}
1248
1249	for (i = start_block; i < end_block; i += bufblks) {
1250		int		bcount, endcount;
1251
1252		bcount = min(bufblks, end_block - start_block);
1253		endcount = bcount - j;
1254
1255		/* We may need to do a read at the end to fill in part of
1256		 * the buffer in the final sector not covered by the write.
1257		 * If this is the same sector as the above read, skip it.
1258		 */
1259		ealign = round_down(end_block, sectbb);
1260		if (j == 0 && (start_block + endcount > ealign)) {
1261			offset = bp->b_addr + BBTOB(ealign - start_block);
1262			error = xlog_bread_offset(log, ealign, sectbb,
1263							bp, offset);
1264			if (error)
1265				break;
1266
1267		}
1268
1269		offset = xlog_align(log, start_block, endcount, bp);
1270		for (; j < endcount; j++) {
1271			xlog_add_record(log, offset, cycle, i+j,
1272					tail_cycle, tail_block);
1273			offset += BBSIZE;
1274		}
1275		error = xlog_bwrite(log, start_block, endcount, bp);
1276		if (error)
1277			break;
1278		start_block += endcount;
1279		j = 0;
1280	}
1281
1282 out_put_bp:
1283	xlog_put_bp(bp);
1284	return error;
1285}
1286
1287/*
1288 * This routine is called to blow away any incomplete log writes out
1289 * in front of the log head.  We do this so that we won't become confused
1290 * if we come up, write only a little bit more, and then crash again.
1291 * If we leave the partial log records out there, this situation could
1292 * cause us to think those partial writes are valid blocks since they
1293 * have the current cycle number.  We get rid of them by overwriting them
1294 * with empty log records with the old cycle number rather than the
1295 * current one.
1296 *
1297 * The tail lsn is passed in rather than taken from
1298 * the log so that we will not write over the unmount record after a
1299 * clean unmount in a 512 block log.  Doing so would leave the log without
1300 * any valid log records in it until a new one was written.  If we crashed
1301 * during that time we would not be able to recover.
1302 */
1303STATIC int
1304xlog_clear_stale_blocks(
1305	xlog_t		*log,
1306	xfs_lsn_t	tail_lsn)
1307{
1308	int		tail_cycle, head_cycle;
1309	int		tail_block, head_block;
1310	int		tail_distance, max_distance;
1311	int		distance;
1312	int		error;
1313
1314	tail_cycle = CYCLE_LSN(tail_lsn);
1315	tail_block = BLOCK_LSN(tail_lsn);
1316	head_cycle = log->l_curr_cycle;
1317	head_block = log->l_curr_block;
1318
1319	/*
1320	 * Figure out the distance between the new head of the log
1321	 * and the tail.  We want to write over any blocks beyond the
1322	 * head that we may have written just before the crash, but
1323	 * we don't want to overwrite the tail of the log.
1324	 */
1325	if (head_cycle == tail_cycle) {
1326		/*
1327		 * The tail is behind the head in the physical log,
1328		 * so the distance from the head to the tail is the
1329		 * distance from the head to the end of the log plus
1330		 * the distance from the beginning of the log to the
1331		 * tail.
1332		 */
1333		if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1334			XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1335					 XFS_ERRLEVEL_LOW, log->l_mp);
1336			return XFS_ERROR(EFSCORRUPTED);
1337		}
1338		tail_distance = tail_block + (log->l_logBBsize - head_block);
1339	} else {
1340		/*
1341		 * The head is behind the tail in the physical log,
1342		 * so the distance from the head to the tail is just
1343		 * the tail block minus the head block.
1344		 */
1345		if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1346			XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1347					 XFS_ERRLEVEL_LOW, log->l_mp);
1348			return XFS_ERROR(EFSCORRUPTED);
1349		}
1350		tail_distance = tail_block - head_block;
1351	}
1352
1353	/*
1354	 * If the head is right up against the tail, we can't clear
1355	 * anything.
1356	 */
1357	if (tail_distance <= 0) {
1358		ASSERT(tail_distance == 0);
1359		return 0;
1360	}
1361
1362	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1363	/*
1364	 * Take the smaller of the maximum amount of outstanding I/O
1365	 * we could have and the distance to the tail to clear out.
1366	 * We take the smaller so that we don't overwrite the tail and
1367	 * we don't waste all day writing from the head to the tail
1368	 * for no reason.
1369	 */
1370	max_distance = MIN(max_distance, tail_distance);
1371
1372	if ((head_block + max_distance) <= log->l_logBBsize) {
1373		/*
1374		 * We can stomp all the blocks we need to without
1375		 * wrapping around the end of the log.  Just do it
1376		 * in a single write.  Use the cycle number of the
1377		 * current cycle minus one so that the log will look like:
1378		 *     n ... | n - 1 ...
1379		 */
1380		error = xlog_write_log_records(log, (head_cycle - 1),
1381				head_block, max_distance, tail_cycle,
1382				tail_block);
1383		if (error)
1384			return error;
1385	} else {
1386		/*
1387		 * We need to wrap around the end of the physical log in
1388		 * order to clear all the blocks.  Do it in two separate
1389		 * I/Os.  The first write should be from the head to the
1390		 * end of the physical log, and it should use the current
1391		 * cycle number minus one just like above.
1392		 */
1393		distance = log->l_logBBsize - head_block;
1394		error = xlog_write_log_records(log, (head_cycle - 1),
1395				head_block, distance, tail_cycle,
1396				tail_block);
1397
1398		if (error)
1399			return error;
1400
1401		/*
1402		 * Now write the blocks at the start of the physical log.
1403		 * This writes the remainder of the blocks we want to clear.
1404		 * It uses the current cycle number since we're now on the
1405		 * same cycle as the head so that we get:
1406		 *    n ... n ... | n - 1 ...
1407		 *    ^^^^^ blocks we're writing
1408		 */
1409		distance = max_distance - (log->l_logBBsize - head_block);
1410		error = xlog_write_log_records(log, head_cycle, 0, distance,
1411				tail_cycle, tail_block);
1412		if (error)
1413			return error;
1414	}
1415
1416	return 0;
1417}
1418
1419/******************************************************************************
1420 *
1421 *		Log recover routines
1422 *
1423 ******************************************************************************
1424 */
1425
1426STATIC xlog_recover_t *
1427xlog_recover_find_tid(
1428	struct hlist_head	*head,
1429	xlog_tid_t		tid)
1430{
1431	xlog_recover_t		*trans;
1432	struct hlist_node	*n;
1433
1434	hlist_for_each_entry(trans, n, head, r_list) {
1435		if (trans->r_log_tid == tid)
1436			return trans;
1437	}
1438	return NULL;
1439}
1440
1441STATIC void
1442xlog_recover_new_tid(
1443	struct hlist_head	*head,
1444	xlog_tid_t		tid,
1445	xfs_lsn_t		lsn)
1446{
1447	xlog_recover_t		*trans;
1448
1449	trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1450	trans->r_log_tid   = tid;
1451	trans->r_lsn	   = lsn;
1452	INIT_LIST_HEAD(&trans->r_itemq);
1453
1454	INIT_HLIST_NODE(&trans->r_list);
1455	hlist_add_head(&trans->r_list, head);
1456}
1457
1458STATIC void
1459xlog_recover_add_item(
1460	struct list_head	*head)
1461{
1462	xlog_recover_item_t	*item;
1463
1464	item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1465	INIT_LIST_HEAD(&item->ri_list);
1466	list_add_tail(&item->ri_list, head);
1467}
1468
1469STATIC int
1470xlog_recover_add_to_cont_trans(
1471	struct log		*log,
1472	xlog_recover_t		*trans,
1473	xfs_caddr_t		dp,
1474	int			len)
1475{
1476	xlog_recover_item_t	*item;
1477	xfs_caddr_t		ptr, old_ptr;
1478	int			old_len;
1479
1480	if (list_empty(&trans->r_itemq)) {
1481		/* finish copying rest of trans header */
1482		xlog_recover_add_item(&trans->r_itemq);
1483		ptr = (xfs_caddr_t) &trans->r_theader +
1484				sizeof(xfs_trans_header_t) - len;
1485		memcpy(ptr, dp, len); /* d, s, l */
1486		return 0;
1487	}
1488	/* take the tail entry */
1489	item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1490
1491	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1492	old_len = item->ri_buf[item->ri_cnt-1].i_len;
1493
1494	ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1495	memcpy(&ptr[old_len], dp, len); /* d, s, l */
1496	item->ri_buf[item->ri_cnt-1].i_len += len;
1497	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1498	trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
1499	return 0;
1500}
1501
1502/*
1503 * The next region to add is the start of a new region.  It could be
1504 * a whole region or it could be the first part of a new region.  Because
1505 * of this, the assumption here is that the type and size fields of all
1506 * format structures fit into the first 32 bits of the structure.
1507 *
1508 * This works because all regions must be 32 bit aligned.  Therefore, we
1509 * either have both fields or we have neither field.  In the case we have
1510 * neither field, the data part of the region is zero length.  We only have
1511 * a log_op_header and can throw away the header since a new one will appear
1512 * later.  If we have at least 4 bytes, then we can determine how many regions
1513 * will appear in the current log item.
1514 */
1515STATIC int
1516xlog_recover_add_to_trans(
1517	struct log		*log,
1518	xlog_recover_t		*trans,
1519	xfs_caddr_t		dp,
1520	int			len)
1521{
1522	xfs_inode_log_format_t	*in_f;			/* any will do */
1523	xlog_recover_item_t	*item;
1524	xfs_caddr_t		ptr;
1525
1526	if (!len)
1527		return 0;
1528	if (list_empty(&trans->r_itemq)) {
1529		/* we need to catch log corruptions here */
1530		if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1531			xfs_warn(log->l_mp, "%s: bad header magic number",
1532				__func__);
1533			ASSERT(0);
1534			return XFS_ERROR(EIO);
1535		}
1536		if (len == sizeof(xfs_trans_header_t))
1537			xlog_recover_add_item(&trans->r_itemq);
1538		memcpy(&trans->r_theader, dp, len); /* d, s, l */
1539		return 0;
1540	}
1541
1542	ptr = kmem_alloc(len, KM_SLEEP);
1543	memcpy(ptr, dp, len);
1544	in_f = (xfs_inode_log_format_t *)ptr;
1545
1546	/* take the tail entry */
1547	item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1548	if (item->ri_total != 0 &&
1549	     item->ri_total == item->ri_cnt) {
1550		/* tail item is in use, get a new one */
1551		xlog_recover_add_item(&trans->r_itemq);
1552		item = list_entry(trans->r_itemq.prev,
1553					xlog_recover_item_t, ri_list);
1554	}
1555
1556	if (item->ri_total == 0) {		/* first region to be added */
1557		if (in_f->ilf_size == 0 ||
1558		    in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
1559			xfs_warn(log->l_mp,
1560		"bad number of regions (%d) in inode log format",
1561				  in_f->ilf_size);
1562			ASSERT(0);
 
1563			return XFS_ERROR(EIO);
1564		}
1565
1566		item->ri_total = in_f->ilf_size;
1567		item->ri_buf =
1568			kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
1569				    KM_SLEEP);
1570	}
1571	ASSERT(item->ri_total > item->ri_cnt);
1572	/* Description region is ri_buf[0] */
1573	item->ri_buf[item->ri_cnt].i_addr = ptr;
1574	item->ri_buf[item->ri_cnt].i_len  = len;
1575	item->ri_cnt++;
1576	trace_xfs_log_recover_item_add(log, trans, item, 0);
1577	return 0;
1578}
1579
1580/*
1581 * Sort the log items in the transaction. Cancelled buffers need
1582 * to be put first so they are processed before any items that might
1583 * modify the buffers. If they are cancelled, then the modifications
1584 * don't need to be replayed.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1585 */
1586STATIC int
1587xlog_recover_reorder_trans(
1588	struct log		*log,
1589	xlog_recover_t		*trans,
1590	int			pass)
1591{
1592	xlog_recover_item_t	*item, *n;
 
1593	LIST_HEAD(sort_list);
 
 
 
 
1594
1595	list_splice_init(&trans->r_itemq, &sort_list);
1596	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1597		xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
1598
1599		switch (ITEM_TYPE(item)) {
 
 
 
1600		case XFS_LI_BUF:
1601			if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1602				trace_xfs_log_recover_item_reorder_head(log,
1603							trans, item, pass);
1604				list_move(&item->ri_list, &trans->r_itemq);
 
 
 
 
1605				break;
1606			}
 
 
1607		case XFS_LI_INODE:
1608		case XFS_LI_DQUOT:
1609		case XFS_LI_QUOTAOFF:
1610		case XFS_LI_EFD:
1611		case XFS_LI_EFI:
1612			trace_xfs_log_recover_item_reorder_tail(log,
1613							trans, item, pass);
1614			list_move_tail(&item->ri_list, &trans->r_itemq);
1615			break;
1616		default:
1617			xfs_warn(log->l_mp,
1618				"%s: unrecognized type of log operation",
1619				__func__);
1620			ASSERT(0);
1621			return XFS_ERROR(EIO);
 
 
 
 
 
 
 
1622		}
1623	}
 
1624	ASSERT(list_empty(&sort_list));
1625	return 0;
 
 
 
 
 
 
 
 
1626}
1627
1628/*
1629 * Build up the table of buf cancel records so that we don't replay
1630 * cancelled data in the second pass.  For buffer records that are
1631 * not cancel records, there is nothing to do here so we just return.
1632 *
1633 * If we get a cancel record which is already in the table, this indicates
1634 * that the buffer was cancelled multiple times.  In order to ensure
1635 * that during pass 2 we keep the record in the table until we reach its
1636 * last occurrence in the log, we keep a reference count in the cancel
1637 * record in the table to tell us how many times we expect to see this
1638 * record during the second pass.
1639 */
1640STATIC int
1641xlog_recover_buffer_pass1(
1642	struct log		*log,
1643	xlog_recover_item_t	*item)
1644{
1645	xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
1646	struct list_head	*bucket;
1647	struct xfs_buf_cancel	*bcp;
1648
1649	/*
1650	 * If this isn't a cancel buffer item, then just return.
1651	 */
1652	if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1653		trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1654		return 0;
1655	}
1656
1657	/*
1658	 * Insert an xfs_buf_cancel record into the hash table of them.
1659	 * If there is already an identical record, bump its reference count.
1660	 */
1661	bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1662	list_for_each_entry(bcp, bucket, bc_list) {
1663		if (bcp->bc_blkno == buf_f->blf_blkno &&
1664		    bcp->bc_len == buf_f->blf_len) {
1665			bcp->bc_refcount++;
1666			trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1667			return 0;
1668		}
1669	}
1670
1671	bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
1672	bcp->bc_blkno = buf_f->blf_blkno;
1673	bcp->bc_len = buf_f->blf_len;
1674	bcp->bc_refcount = 1;
1675	list_add_tail(&bcp->bc_list, bucket);
1676
1677	trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1678	return 0;
1679}
1680
1681/*
1682 * Check to see whether the buffer being recovered has a corresponding
1683 * entry in the buffer cancel record table.  If it does then return 1
1684 * so that it will be cancelled, otherwise return 0.  If the buffer is
1685 * actually a buffer cancel item (XFS_BLF_CANCEL is set), then decrement
1686 * the refcount on the entry in the table and remove it from the table
1687 * if this is the last reference.
1688 *
1689 * We remove the cancel record from the table when we encounter its
1690 * last occurrence in the log so that if the same buffer is re-used
1691 * again after its last cancellation we actually replay the changes
1692 * made at that point.
1693 */
1694STATIC int
1695xlog_check_buffer_cancelled(
1696	struct log		*log,
1697	xfs_daddr_t		blkno,
1698	uint			len,
1699	ushort			flags)
1700{
1701	struct list_head	*bucket;
1702	struct xfs_buf_cancel	*bcp;
1703
1704	if (log->l_buf_cancel_table == NULL) {
1705		/*
1706		 * There is nothing in the table built in pass one,
1707		 * so this buffer must not be cancelled.
1708		 */
1709		ASSERT(!(flags & XFS_BLF_CANCEL));
1710		return 0;
1711	}
1712
1713	/*
1714	 * Search for an entry in the  cancel table that matches our buffer.
1715	 */
1716	bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
1717	list_for_each_entry(bcp, bucket, bc_list) {
1718		if (bcp->bc_blkno == blkno && bcp->bc_len == len)
1719			goto found;
1720	}
1721
1722	/*
1723	 * We didn't find a corresponding entry in the table, so return 0 so
1724	 * that the buffer is NOT cancelled.
1725	 */
1726	ASSERT(!(flags & XFS_BLF_CANCEL));
1727	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1728
1729found:
1730	/*
1731	 * We've go a match, so return 1 so that the recovery of this buffer
1732	 * is cancelled.  If this buffer is actually a buffer cancel log
1733	 * item, then decrement the refcount on the one in the table and
1734	 * remove it if this is the last reference.
1735	 */
1736	if (flags & XFS_BLF_CANCEL) {
1737		if (--bcp->bc_refcount == 0) {
1738			list_del(&bcp->bc_list);
1739			kmem_free(bcp);
1740		}
1741	}
1742	return 1;
1743}
1744
1745/*
1746 * Perform recovery for a buffer full of inodes.  In these buffers, the only
1747 * data which should be recovered is that which corresponds to the
1748 * di_next_unlinked pointers in the on disk inode structures.  The rest of the
1749 * data for the inodes is always logged through the inodes themselves rather
1750 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
1751 *
1752 * The only time when buffers full of inodes are fully recovered is when the
1753 * buffer is full of newly allocated inodes.  In this case the buffer will
1754 * not be marked as an inode buffer and so will be sent to
1755 * xlog_recover_do_reg_buffer() below during recovery.
1756 */
1757STATIC int
1758xlog_recover_do_inode_buffer(
1759	struct xfs_mount	*mp,
1760	xlog_recover_item_t	*item,
1761	struct xfs_buf		*bp,
1762	xfs_buf_log_format_t	*buf_f)
1763{
1764	int			i;
1765	int			item_index = 0;
1766	int			bit = 0;
1767	int			nbits = 0;
1768	int			reg_buf_offset = 0;
1769	int			reg_buf_bytes = 0;
1770	int			next_unlinked_offset;
1771	int			inodes_per_buf;
1772	xfs_agino_t		*logged_nextp;
1773	xfs_agino_t		*buffer_nextp;
1774
1775	trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
1776
1777	inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
 
 
 
 
 
 
 
1778	for (i = 0; i < inodes_per_buf; i++) {
1779		next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1780			offsetof(xfs_dinode_t, di_next_unlinked);
1781
1782		while (next_unlinked_offset >=
1783		       (reg_buf_offset + reg_buf_bytes)) {
1784			/*
1785			 * The next di_next_unlinked field is beyond
1786			 * the current logged region.  Find the next
1787			 * logged region that contains or is beyond
1788			 * the current di_next_unlinked field.
1789			 */
1790			bit += nbits;
1791			bit = xfs_next_bit(buf_f->blf_data_map,
1792					   buf_f->blf_map_size, bit);
1793
1794			/*
1795			 * If there are no more logged regions in the
1796			 * buffer, then we're done.
1797			 */
1798			if (bit == -1)
1799				return 0;
1800
1801			nbits = xfs_contig_bits(buf_f->blf_data_map,
1802						buf_f->blf_map_size, bit);
1803			ASSERT(nbits > 0);
1804			reg_buf_offset = bit << XFS_BLF_SHIFT;
1805			reg_buf_bytes = nbits << XFS_BLF_SHIFT;
1806			item_index++;
1807		}
1808
1809		/*
1810		 * If the current logged region starts after the current
1811		 * di_next_unlinked field, then move on to the next
1812		 * di_next_unlinked field.
1813		 */
1814		if (next_unlinked_offset < reg_buf_offset)
1815			continue;
1816
1817		ASSERT(item->ri_buf[item_index].i_addr != NULL);
1818		ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
1819		ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
 
1820
1821		/*
1822		 * The current logged region contains a copy of the
1823		 * current di_next_unlinked field.  Extract its value
1824		 * and copy it to the buffer copy.
1825		 */
1826		logged_nextp = item->ri_buf[item_index].i_addr +
1827				next_unlinked_offset - reg_buf_offset;
1828		if (unlikely(*logged_nextp == 0)) {
1829			xfs_alert(mp,
1830		"Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
1831		"Trying to replay bad (0) inode di_next_unlinked field.",
1832				item, bp);
1833			XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1834					 XFS_ERRLEVEL_LOW, mp);
1835			return XFS_ERROR(EFSCORRUPTED);
1836		}
1837
1838		buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1839					      next_unlinked_offset);
1840		*buffer_nextp = *logged_nextp;
 
 
 
 
 
 
 
 
 
1841	}
1842
1843	return 0;
1844}
1845
1846/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1847 * Perform a 'normal' buffer recovery.  Each logged region of the
1848 * buffer should be copied over the corresponding region in the
1849 * given buffer.  The bitmap in the buf log format structure indicates
1850 * where to place the logged data.
1851 */
1852STATIC void
1853xlog_recover_do_reg_buffer(
1854	struct xfs_mount	*mp,
1855	xlog_recover_item_t	*item,
1856	struct xfs_buf		*bp,
1857	xfs_buf_log_format_t	*buf_f)
1858{
1859	int			i;
1860	int			bit;
1861	int			nbits;
1862	int                     error;
1863
1864	trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
1865
1866	bit = 0;
1867	i = 1;  /* 0 is the buf format structure */
1868	while (1) {
1869		bit = xfs_next_bit(buf_f->blf_data_map,
1870				   buf_f->blf_map_size, bit);
1871		if (bit == -1)
1872			break;
1873		nbits = xfs_contig_bits(buf_f->blf_data_map,
1874					buf_f->blf_map_size, bit);
1875		ASSERT(nbits > 0);
1876		ASSERT(item->ri_buf[i].i_addr != NULL);
1877		ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
1878		ASSERT(XFS_BUF_COUNT(bp) >=
1879		       ((uint)bit << XFS_BLF_SHIFT)+(nbits<<XFS_BLF_SHIFT));
 
 
 
 
 
 
 
 
 
 
 
1880
1881		/*
1882		 * Do a sanity check if this is a dquot buffer. Just checking
1883		 * the first dquot in the buffer should do. XXXThis is
1884		 * probably a good thing to do for other buf types also.
1885		 */
1886		error = 0;
1887		if (buf_f->blf_flags &
1888		   (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
1889			if (item->ri_buf[i].i_addr == NULL) {
1890				xfs_alert(mp,
1891					"XFS: NULL dquot in %s.", __func__);
1892				goto next;
1893			}
1894			if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
1895				xfs_alert(mp,
1896					"XFS: dquot too small (%d) in %s.",
1897					item->ri_buf[i].i_len, __func__);
1898				goto next;
1899			}
1900			error = xfs_qm_dqcheck(mp, item->ri_buf[i].i_addr,
1901					       -1, 0, XFS_QMOPT_DOWARN,
1902					       "dquot_buf_recover");
1903			if (error)
1904				goto next;
1905		}
1906
1907		memcpy(xfs_buf_offset(bp,
1908			(uint)bit << XFS_BLF_SHIFT),	/* dest */
1909			item->ri_buf[i].i_addr,		/* source */
1910			nbits<<XFS_BLF_SHIFT);		/* length */
1911 next:
1912		i++;
1913		bit += nbits;
1914	}
1915
1916	/* Shouldn't be any more regions */
1917	ASSERT(i == item->ri_total);
1918}
1919
1920/*
1921 * Do some primitive error checking on ondisk dquot data structures.
1922 */
1923int
1924xfs_qm_dqcheck(
1925	struct xfs_mount *mp,
1926	xfs_disk_dquot_t *ddq,
1927	xfs_dqid_t	 id,
1928	uint		 type,	  /* used only when IO_dorepair is true */
1929	uint		 flags,
1930	char		 *str)
1931{
1932	xfs_dqblk_t	 *d = (xfs_dqblk_t *)ddq;
1933	int		errs = 0;
1934
1935	/*
1936	 * We can encounter an uninitialized dquot buffer for 2 reasons:
1937	 * 1. If we crash while deleting the quotainode(s), and those blks got
1938	 *    used for user data. This is because we take the path of regular
1939	 *    file deletion; however, the size field of quotainodes is never
1940	 *    updated, so all the tricks that we play in itruncate_finish
1941	 *    don't quite matter.
1942	 *
1943	 * 2. We don't play the quota buffers when there's a quotaoff logitem.
1944	 *    But the allocation will be replayed so we'll end up with an
1945	 *    uninitialized quota block.
1946	 *
1947	 * This is all fine; things are still consistent, and we haven't lost
1948	 * any quota information. Just don't complain about bad dquot blks.
1949	 */
1950	if (ddq->d_magic != cpu_to_be16(XFS_DQUOT_MAGIC)) {
1951		if (flags & XFS_QMOPT_DOWARN)
1952			xfs_alert(mp,
1953			"%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1954			str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
1955		errs++;
1956	}
1957	if (ddq->d_version != XFS_DQUOT_VERSION) {
1958		if (flags & XFS_QMOPT_DOWARN)
1959			xfs_alert(mp,
1960			"%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1961			str, id, ddq->d_version, XFS_DQUOT_VERSION);
1962		errs++;
1963	}
1964
1965	if (ddq->d_flags != XFS_DQ_USER &&
1966	    ddq->d_flags != XFS_DQ_PROJ &&
1967	    ddq->d_flags != XFS_DQ_GROUP) {
1968		if (flags & XFS_QMOPT_DOWARN)
1969			xfs_alert(mp,
1970			"%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1971			str, id, ddq->d_flags);
1972		errs++;
1973	}
1974
1975	if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
1976		if (flags & XFS_QMOPT_DOWARN)
1977			xfs_alert(mp,
1978			"%s : ondisk-dquot 0x%p, ID mismatch: "
1979			"0x%x expected, found id 0x%x",
1980			str, ddq, id, be32_to_cpu(ddq->d_id));
1981		errs++;
1982	}
1983
1984	if (!errs && ddq->d_id) {
1985		if (ddq->d_blk_softlimit &&
1986		    be64_to_cpu(ddq->d_bcount) >=
1987				be64_to_cpu(ddq->d_blk_softlimit)) {
1988			if (!ddq->d_btimer) {
1989				if (flags & XFS_QMOPT_DOWARN)
1990					xfs_alert(mp,
1991			"%s : Dquot ID 0x%x (0x%p) BLK TIMER NOT STARTED",
1992					str, (int)be32_to_cpu(ddq->d_id), ddq);
1993				errs++;
1994			}
1995		}
1996		if (ddq->d_ino_softlimit &&
1997		    be64_to_cpu(ddq->d_icount) >=
1998				be64_to_cpu(ddq->d_ino_softlimit)) {
1999			if (!ddq->d_itimer) {
2000				if (flags & XFS_QMOPT_DOWARN)
2001					xfs_alert(mp,
2002			"%s : Dquot ID 0x%x (0x%p) INODE TIMER NOT STARTED",
2003					str, (int)be32_to_cpu(ddq->d_id), ddq);
2004				errs++;
2005			}
2006		}
2007		if (ddq->d_rtb_softlimit &&
2008		    be64_to_cpu(ddq->d_rtbcount) >=
2009				be64_to_cpu(ddq->d_rtb_softlimit)) {
2010			if (!ddq->d_rtbtimer) {
2011				if (flags & XFS_QMOPT_DOWARN)
2012					xfs_alert(mp,
2013			"%s : Dquot ID 0x%x (0x%p) RTBLK TIMER NOT STARTED",
2014					str, (int)be32_to_cpu(ddq->d_id), ddq);
2015				errs++;
2016			}
2017		}
2018	}
2019
2020	if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2021		return errs;
2022
2023	if (flags & XFS_QMOPT_DOWARN)
2024		xfs_notice(mp, "Re-initializing dquot ID 0x%x", id);
2025
2026	/*
2027	 * Typically, a repair is only requested by quotacheck.
 
 
 
 
 
2028	 */
2029	ASSERT(id != -1);
2030	ASSERT(flags & XFS_QMOPT_DQREPAIR);
2031	memset(d, 0, sizeof(xfs_dqblk_t));
2032
2033	d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2034	d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2035	d->dd_diskdq.d_flags = type;
2036	d->dd_diskdq.d_id = cpu_to_be32(id);
2037
2038	return errs;
2039}
2040
2041/*
2042 * Perform a dquot buffer recovery.
2043 * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2044 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2045 * Else, treat it as a regular buffer and do recovery.
2046 */
2047STATIC void
2048xlog_recover_do_dquot_buffer(
2049	xfs_mount_t		*mp,
2050	xlog_t			*log,
2051	xlog_recover_item_t	*item,
2052	xfs_buf_t		*bp,
2053	xfs_buf_log_format_t	*buf_f)
2054{
2055	uint			type;
2056
2057	trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2058
2059	/*
2060	 * Filesystems are required to send in quota flags at mount time.
2061	 */
2062	if (mp->m_qflags == 0) {
2063		return;
2064	}
2065
2066	type = 0;
2067	if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2068		type |= XFS_DQ_USER;
2069	if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2070		type |= XFS_DQ_PROJ;
2071	if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2072		type |= XFS_DQ_GROUP;
2073	/*
2074	 * This type of quotas was turned off, so ignore this buffer
2075	 */
2076	if (log->l_quotaoffs_flag & type)
2077		return;
2078
2079	xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2080}
2081
2082/*
2083 * This routine replays a modification made to a buffer at runtime.
2084 * There are actually two types of buffer, regular and inode, which
2085 * are handled differently.  Inode buffers are handled differently
2086 * in that we only recover a specific set of data from them, namely
2087 * the inode di_next_unlinked fields.  This is because all other inode
2088 * data is actually logged via inode records and any data we replay
2089 * here which overlaps that may be stale.
2090 *
2091 * When meta-data buffers are freed at run time we log a buffer item
2092 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2093 * of the buffer in the log should not be replayed at recovery time.
2094 * This is so that if the blocks covered by the buffer are reused for
2095 * file data before we crash we don't end up replaying old, freed
2096 * meta-data into a user's file.
2097 *
2098 * To handle the cancellation of buffer log items, we make two passes
2099 * over the log during recovery.  During the first we build a table of
2100 * those buffers which have been cancelled, and during the second we
2101 * only replay those buffers which do not have corresponding cancel
2102 * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2103 * for more details on the implementation of the table of cancel records.
2104 */
2105STATIC int
2106xlog_recover_buffer_pass2(
2107	xlog_t			*log,
2108	xlog_recover_item_t	*item)
 
 
2109{
2110	xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
2111	xfs_mount_t		*mp = log->l_mp;
2112	xfs_buf_t		*bp;
2113	int			error;
2114	uint			buf_flags;
 
2115
2116	/*
2117	 * In this pass we only want to recover all the buffers which have
2118	 * not been cancelled and are not cancellation buffers themselves.
2119	 */
2120	if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2121			buf_f->blf_len, buf_f->blf_flags)) {
2122		trace_xfs_log_recover_buf_cancel(log, buf_f);
2123		return 0;
2124	}
2125
2126	trace_xfs_log_recover_buf_recover(log, buf_f);
2127
2128	buf_flags = XBF_LOCK;
2129	if (!(buf_f->blf_flags & XFS_BLF_INODE_BUF))
2130		buf_flags |= XBF_MAPPED;
2131
2132	bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2133			  buf_flags);
2134	if (!bp)
2135		return XFS_ERROR(ENOMEM);
2136	error = bp->b_error;
2137	if (error) {
2138		xfs_ioerror_alert("xlog_recover_do..(read#1)", mp,
2139				  bp, buf_f->blf_blkno);
2140		xfs_buf_relse(bp);
2141		return error;
2142	}
2143
 
 
 
 
 
 
 
 
2144	if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2145		error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2146	} else if (buf_f->blf_flags &
2147		  (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2148		xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2149	} else {
2150		xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2151	}
2152	if (error)
2153		return XFS_ERROR(error);
2154
2155	/*
2156	 * Perform delayed write on the buffer.  Asynchronous writes will be
2157	 * slower when taking into account all the buffers to be flushed.
2158	 *
2159	 * Also make sure that only inode buffers with good sizes stay in
2160	 * the buffer cache.  The kernel moves inodes in buffers of 1 block
2161	 * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2162	 * buffers in the log can be a different size if the log was generated
2163	 * by an older kernel using unclustered inode buffers or a newer kernel
2164	 * running with a different inode cluster size.  Regardless, if the
2165	 * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2166	 * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2167	 * the buffer out of the buffer cache so that the buffer won't
2168	 * overlap with future reads of those inodes.
2169	 */
2170	if (XFS_DINODE_MAGIC ==
2171	    be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2172	    (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2173			(__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2174		XFS_BUF_STALE(bp);
2175		error = xfs_bwrite(mp, bp);
2176	} else {
2177		ASSERT(bp->b_target->bt_mount == mp);
2178		bp->b_iodone = xlog_recover_iodone;
2179		xfs_bdwrite(mp, bp);
2180	}
2181
2182	return (error);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2183}
2184
2185STATIC int
2186xlog_recover_inode_pass2(
2187	xlog_t			*log,
2188	xlog_recover_item_t	*item)
 
 
2189{
2190	xfs_inode_log_format_t	*in_f;
2191	xfs_mount_t		*mp = log->l_mp;
2192	xfs_buf_t		*bp;
2193	xfs_dinode_t		*dip;
2194	int			len;
2195	xfs_caddr_t		src;
2196	xfs_caddr_t		dest;
2197	int			error;
2198	int			attr_index;
2199	uint			fields;
2200	xfs_icdinode_t		*dicp;
 
2201	int			need_free = 0;
2202
2203	if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2204		in_f = item->ri_buf[0].i_addr;
2205	} else {
2206		in_f = kmem_alloc(sizeof(xfs_inode_log_format_t), KM_SLEEP);
2207		need_free = 1;
2208		error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2209		if (error)
2210			goto error;
2211	}
2212
2213	/*
2214	 * Inode buffers can be freed, look out for it,
2215	 * and do not replay the inode.
2216	 */
2217	if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2218					in_f->ilf_len, 0)) {
2219		error = 0;
2220		trace_xfs_log_recover_inode_cancel(log, in_f);
2221		goto error;
2222	}
2223	trace_xfs_log_recover_inode_recover(log, in_f);
2224
2225	bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len,
2226			  XBF_LOCK);
2227	if (!bp) {
2228		error = ENOMEM;
2229		goto error;
2230	}
2231	error = bp->b_error;
2232	if (error) {
2233		xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2234				  bp, in_f->ilf_blkno);
2235		xfs_buf_relse(bp);
2236		goto error;
2237	}
2238	ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2239	dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
2240
2241	/*
2242	 * Make sure the place we're flushing out to really looks
2243	 * like an inode!
2244	 */
2245	if (unlikely(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC))) {
2246		xfs_buf_relse(bp);
2247		xfs_alert(mp,
2248	"%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
2249			__func__, dip, bp, in_f->ilf_ino);
2250		XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2251				 XFS_ERRLEVEL_LOW, mp);
2252		error = EFSCORRUPTED;
2253		goto error;
2254	}
2255	dicp = item->ri_buf[1].i_addr;
2256	if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2257		xfs_buf_relse(bp);
2258		xfs_alert(mp,
2259			"%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
2260			__func__, item, in_f->ilf_ino);
2261		XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2262				 XFS_ERRLEVEL_LOW, mp);
2263		error = EFSCORRUPTED;
2264		goto error;
2265	}
2266
2267	/* Skip replay when the on disk inode is newer than the log one */
2268	if (dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2269		/*
2270		 * Deal with the wrap case, DI_MAX_FLUSH is less
2271		 * than smaller numbers
2272		 */
2273		if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
2274		    dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2275			/* do nothing */
2276		} else {
2277			xfs_buf_relse(bp);
2278			trace_xfs_log_recover_inode_skip(log, in_f);
2279			error = 0;
2280			goto error;
2281		}
2282	}
 
2283	/* Take the opportunity to reset the flush iteration count */
2284	dicp->di_flushiter = 0;
2285
2286	if (unlikely(S_ISREG(dicp->di_mode))) {
2287		if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2288		    (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2289			XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
2290					 XFS_ERRLEVEL_LOW, mp, dicp);
2291			xfs_buf_relse(bp);
2292			xfs_alert(mp,
2293		"%s: Bad regular inode log record, rec ptr 0x%p, "
2294		"ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2295				__func__, item, dip, bp, in_f->ilf_ino);
2296			error = EFSCORRUPTED;
2297			goto error;
2298		}
2299	} else if (unlikely(S_ISDIR(dicp->di_mode))) {
2300		if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2301		    (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2302		    (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2303			XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
2304					     XFS_ERRLEVEL_LOW, mp, dicp);
2305			xfs_buf_relse(bp);
2306			xfs_alert(mp,
2307		"%s: Bad dir inode log record, rec ptr 0x%p, "
2308		"ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2309				__func__, item, dip, bp, in_f->ilf_ino);
2310			error = EFSCORRUPTED;
2311			goto error;
2312		}
2313	}
2314	if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2315		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
2316				     XFS_ERRLEVEL_LOW, mp, dicp);
2317		xfs_buf_relse(bp);
2318		xfs_alert(mp,
2319	"%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2320	"dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2321			__func__, item, dip, bp, in_f->ilf_ino,
2322			dicp->di_nextents + dicp->di_anextents,
2323			dicp->di_nblocks);
2324		error = EFSCORRUPTED;
2325		goto error;
2326	}
2327	if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2328		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
2329				     XFS_ERRLEVEL_LOW, mp, dicp);
2330		xfs_buf_relse(bp);
2331		xfs_alert(mp,
2332	"%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2333	"dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__,
2334			item, dip, bp, in_f->ilf_ino, dicp->di_forkoff);
2335		error = EFSCORRUPTED;
2336		goto error;
2337	}
2338	if (unlikely(item->ri_buf[1].i_len > sizeof(struct xfs_icdinode))) {
 
2339		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
2340				     XFS_ERRLEVEL_LOW, mp, dicp);
2341		xfs_buf_relse(bp);
2342		xfs_alert(mp,
2343			"%s: Bad inode log record length %d, rec ptr 0x%p",
2344			__func__, item->ri_buf[1].i_len, item);
2345		error = EFSCORRUPTED;
2346		goto error;
2347	}
2348
2349	/* The core is in in-core format */
2350	xfs_dinode_to_disk(dip, item->ri_buf[1].i_addr);
2351
2352	/* the rest is in on-disk format */
2353	if (item->ri_buf[1].i_len > sizeof(struct xfs_icdinode)) {
2354		memcpy((xfs_caddr_t) dip + sizeof(struct xfs_icdinode),
2355			item->ri_buf[1].i_addr + sizeof(struct xfs_icdinode),
2356			item->ri_buf[1].i_len  - sizeof(struct xfs_icdinode));
2357	}
2358
2359	fields = in_f->ilf_fields;
2360	switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2361	case XFS_ILOG_DEV:
2362		xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
2363		break;
2364	case XFS_ILOG_UUID:
2365		memcpy(XFS_DFORK_DPTR(dip),
2366		       &in_f->ilf_u.ilfu_uuid,
2367		       sizeof(uuid_t));
2368		break;
2369	}
2370
2371	if (in_f->ilf_size == 2)
2372		goto write_inode_buffer;
2373	len = item->ri_buf[2].i_len;
2374	src = item->ri_buf[2].i_addr;
2375	ASSERT(in_f->ilf_size <= 4);
2376	ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2377	ASSERT(!(fields & XFS_ILOG_DFORK) ||
2378	       (len == in_f->ilf_dsize));
2379
2380	switch (fields & XFS_ILOG_DFORK) {
2381	case XFS_ILOG_DDATA:
2382	case XFS_ILOG_DEXT:
2383		memcpy(XFS_DFORK_DPTR(dip), src, len);
2384		break;
2385
2386	case XFS_ILOG_DBROOT:
2387		xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2388				 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
2389				 XFS_DFORK_DSIZE(dip, mp));
2390		break;
2391
2392	default:
2393		/*
2394		 * There are no data fork flags set.
2395		 */
2396		ASSERT((fields & XFS_ILOG_DFORK) == 0);
2397		break;
2398	}
2399
2400	/*
2401	 * If we logged any attribute data, recover it.  There may or
2402	 * may not have been any other non-core data logged in this
2403	 * transaction.
2404	 */
2405	if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2406		if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2407			attr_index = 3;
2408		} else {
2409			attr_index = 2;
2410		}
2411		len = item->ri_buf[attr_index].i_len;
2412		src = item->ri_buf[attr_index].i_addr;
2413		ASSERT(len == in_f->ilf_asize);
2414
2415		switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2416		case XFS_ILOG_ADATA:
2417		case XFS_ILOG_AEXT:
2418			dest = XFS_DFORK_APTR(dip);
2419			ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2420			memcpy(dest, src, len);
2421			break;
2422
2423		case XFS_ILOG_ABROOT:
2424			dest = XFS_DFORK_APTR(dip);
2425			xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2426					 len, (xfs_bmdr_block_t*)dest,
2427					 XFS_DFORK_ASIZE(dip, mp));
2428			break;
2429
2430		default:
2431			xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
2432			ASSERT(0);
2433			xfs_buf_relse(bp);
2434			error = EIO;
2435			goto error;
2436		}
2437	}
2438
2439write_inode_buffer:
 
 
 
 
 
 
2440	ASSERT(bp->b_target->bt_mount == mp);
2441	bp->b_iodone = xlog_recover_iodone;
2442	xfs_bdwrite(mp, bp);
 
 
 
2443error:
2444	if (need_free)
2445		kmem_free(in_f);
2446	return XFS_ERROR(error);
2447}
2448
2449/*
2450 * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2451 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2452 * of that type.
2453 */
2454STATIC int
2455xlog_recover_quotaoff_pass1(
2456	xlog_t			*log,
2457	xlog_recover_item_t	*item)
2458{
2459	xfs_qoff_logformat_t	*qoff_f = item->ri_buf[0].i_addr;
2460	ASSERT(qoff_f);
2461
2462	/*
2463	 * The logitem format's flag tells us if this was user quotaoff,
2464	 * group/project quotaoff or both.
2465	 */
2466	if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2467		log->l_quotaoffs_flag |= XFS_DQ_USER;
2468	if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2469		log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2470	if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2471		log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2472
2473	return (0);
2474}
2475
2476/*
2477 * Recover a dquot record
2478 */
2479STATIC int
2480xlog_recover_dquot_pass2(
2481	xlog_t			*log,
2482	xlog_recover_item_t	*item)
 
 
2483{
2484	xfs_mount_t		*mp = log->l_mp;
2485	xfs_buf_t		*bp;
2486	struct xfs_disk_dquot	*ddq, *recddq;
2487	int			error;
2488	xfs_dq_logformat_t	*dq_f;
2489	uint			type;
2490
2491
2492	/*
2493	 * Filesystems are required to send in quota flags at mount time.
2494	 */
2495	if (mp->m_qflags == 0)
2496		return (0);
2497
2498	recddq = item->ri_buf[1].i_addr;
2499	if (recddq == NULL) {
2500		xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
2501		return XFS_ERROR(EIO);
2502	}
2503	if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
2504		xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
2505			item->ri_buf[1].i_len, __func__);
2506		return XFS_ERROR(EIO);
2507	}
2508
2509	/*
2510	 * This type of quotas was turned off, so ignore this record.
2511	 */
2512	type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2513	ASSERT(type);
2514	if (log->l_quotaoffs_flag & type)
2515		return (0);
2516
2517	/*
2518	 * At this point we know that quota was _not_ turned off.
2519	 * Since the mount flags are not indicating to us otherwise, this
2520	 * must mean that quota is on, and the dquot needs to be replayed.
2521	 * Remember that we may not have fully recovered the superblock yet,
2522	 * so we can't do the usual trick of looking at the SB quota bits.
2523	 *
2524	 * The other possibility, of course, is that the quota subsystem was
2525	 * removed since the last mount - ENOSYS.
2526	 */
2527	dq_f = item->ri_buf[0].i_addr;
2528	ASSERT(dq_f);
2529	error = xfs_qm_dqcheck(mp, recddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2530			   "xlog_recover_dquot_pass2 (log copy)");
2531	if (error)
2532		return XFS_ERROR(EIO);
2533	ASSERT(dq_f->qlf_len == 1);
2534
2535	error = xfs_read_buf(mp, mp->m_ddev_targp,
2536			     dq_f->qlf_blkno,
2537			     XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2538			     0, &bp);
2539	if (error) {
2540		xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2541				  bp, dq_f->qlf_blkno);
2542		return error;
2543	}
2544	ASSERT(bp);
2545	ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2546
2547	/*
2548	 * At least the magic num portion should be on disk because this
2549	 * was among a chunk of dquots created earlier, and we did some
2550	 * minimal initialization then.
2551	 */
2552	error = xfs_qm_dqcheck(mp, ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2553			   "xlog_recover_dquot_pass2");
2554	if (error) {
2555		xfs_buf_relse(bp);
2556		return XFS_ERROR(EIO);
2557	}
2558
 
 
 
 
 
 
 
 
 
 
 
 
 
2559	memcpy(ddq, recddq, item->ri_buf[1].i_len);
 
 
 
 
2560
2561	ASSERT(dq_f->qlf_size == 2);
2562	ASSERT(bp->b_target->bt_mount == mp);
2563	bp->b_iodone = xlog_recover_iodone;
2564	xfs_bdwrite(mp, bp);
2565
2566	return (0);
 
 
2567}
2568
2569/*
2570 * This routine is called to create an in-core extent free intent
2571 * item from the efi format structure which was logged on disk.
2572 * It allocates an in-core efi, copies the extents from the format
2573 * structure into it, and adds the efi to the AIL with the given
2574 * LSN.
2575 */
2576STATIC int
2577xlog_recover_efi_pass2(
2578	xlog_t			*log,
2579	xlog_recover_item_t	*item,
2580	xfs_lsn_t		lsn)
2581{
2582	int			error;
2583	xfs_mount_t		*mp = log->l_mp;
2584	xfs_efi_log_item_t	*efip;
2585	xfs_efi_log_format_t	*efi_formatp;
2586
2587	efi_formatp = item->ri_buf[0].i_addr;
2588
2589	efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2590	if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2591					 &(efip->efi_format)))) {
2592		xfs_efi_item_free(efip);
2593		return error;
2594	}
2595	atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
2596
2597	spin_lock(&log->l_ailp->xa_lock);
2598	/*
2599	 * xfs_trans_ail_update() drops the AIL lock.
2600	 */
2601	xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
2602	return 0;
2603}
2604
2605
2606/*
2607 * This routine is called when an efd format structure is found in
2608 * a committed transaction in the log.  It's purpose is to cancel
2609 * the corresponding efi if it was still in the log.  To do this
2610 * it searches the AIL for the efi with an id equal to that in the
2611 * efd format structure.  If we find it, we remove the efi from the
2612 * AIL and free it.
2613 */
2614STATIC int
2615xlog_recover_efd_pass2(
2616	xlog_t			*log,
2617	xlog_recover_item_t	*item)
2618{
2619	xfs_efd_log_format_t	*efd_formatp;
2620	xfs_efi_log_item_t	*efip = NULL;
2621	xfs_log_item_t		*lip;
2622	__uint64_t		efi_id;
2623	struct xfs_ail_cursor	cur;
2624	struct xfs_ail		*ailp = log->l_ailp;
2625
2626	efd_formatp = item->ri_buf[0].i_addr;
2627	ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2628		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2629	       (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2630		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2631	efi_id = efd_formatp->efd_efi_id;
2632
2633	/*
2634	 * Search for the efi with the id in the efd format structure
2635	 * in the AIL.
2636	 */
2637	spin_lock(&ailp->xa_lock);
2638	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2639	while (lip != NULL) {
2640		if (lip->li_type == XFS_LI_EFI) {
2641			efip = (xfs_efi_log_item_t *)lip;
2642			if (efip->efi_format.efi_id == efi_id) {
2643				/*
2644				 * xfs_trans_ail_delete() drops the
2645				 * AIL lock.
2646				 */
2647				xfs_trans_ail_delete(ailp, lip);
 
2648				xfs_efi_item_free(efip);
2649				spin_lock(&ailp->xa_lock);
2650				break;
2651			}
2652		}
2653		lip = xfs_trans_ail_cursor_next(ailp, &cur);
2654	}
2655	xfs_trans_ail_cursor_done(ailp, &cur);
2656	spin_unlock(&ailp->xa_lock);
2657
2658	return 0;
2659}
2660
2661/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2662 * Free up any resources allocated by the transaction
2663 *
2664 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2665 */
2666STATIC void
2667xlog_recover_free_trans(
2668	struct xlog_recover	*trans)
2669{
2670	xlog_recover_item_t	*item, *n;
2671	int			i;
2672
2673	list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2674		/* Free the regions in the item. */
2675		list_del(&item->ri_list);
2676		for (i = 0; i < item->ri_cnt; i++)
2677			kmem_free(item->ri_buf[i].i_addr);
2678		/* Free the item itself */
2679		kmem_free(item->ri_buf);
2680		kmem_free(item);
2681	}
2682	/* Free the transaction recover structure */
2683	kmem_free(trans);
2684}
2685
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2686STATIC int
2687xlog_recover_commit_pass1(
2688	struct log		*log,
2689	struct xlog_recover	*trans,
2690	xlog_recover_item_t	*item)
2691{
2692	trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
2693
2694	switch (ITEM_TYPE(item)) {
2695	case XFS_LI_BUF:
2696		return xlog_recover_buffer_pass1(log, item);
2697	case XFS_LI_QUOTAOFF:
2698		return xlog_recover_quotaoff_pass1(log, item);
2699	case XFS_LI_INODE:
2700	case XFS_LI_EFI:
2701	case XFS_LI_EFD:
2702	case XFS_LI_DQUOT:
 
2703		/* nothing to do in pass 1 */
2704		return 0;
2705	default:
2706		xfs_warn(log->l_mp, "%s: invalid item type (%d)",
2707			__func__, ITEM_TYPE(item));
2708		ASSERT(0);
2709		return XFS_ERROR(EIO);
2710	}
2711}
2712
2713STATIC int
2714xlog_recover_commit_pass2(
2715	struct log		*log,
2716	struct xlog_recover	*trans,
2717	xlog_recover_item_t	*item)
 
2718{
2719	trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
2720
2721	switch (ITEM_TYPE(item)) {
2722	case XFS_LI_BUF:
2723		return xlog_recover_buffer_pass2(log, item);
 
2724	case XFS_LI_INODE:
2725		return xlog_recover_inode_pass2(log, item);
 
2726	case XFS_LI_EFI:
2727		return xlog_recover_efi_pass2(log, item, trans->r_lsn);
2728	case XFS_LI_EFD:
2729		return xlog_recover_efd_pass2(log, item);
2730	case XFS_LI_DQUOT:
2731		return xlog_recover_dquot_pass2(log, item);
 
 
 
2732	case XFS_LI_QUOTAOFF:
2733		/* nothing to do in pass2 */
2734		return 0;
2735	default:
2736		xfs_warn(log->l_mp, "%s: invalid item type (%d)",
2737			__func__, ITEM_TYPE(item));
2738		ASSERT(0);
2739		return XFS_ERROR(EIO);
2740	}
2741}
2742
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2743/*
2744 * Perform the transaction.
2745 *
2746 * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2747 * EFIs and EFDs get queued up by adding entries into the AIL for them.
2748 */
2749STATIC int
2750xlog_recover_commit_trans(
2751	struct log		*log,
2752	struct xlog_recover	*trans,
2753	int			pass)
2754{
2755	int			error = 0;
2756	xlog_recover_item_t	*item;
 
 
 
 
 
 
 
 
2757
2758	hlist_del(&trans->r_list);
2759
2760	error = xlog_recover_reorder_trans(log, trans, pass);
2761	if (error)
2762		return error;
2763
2764	list_for_each_entry(item, &trans->r_itemq, ri_list) {
2765		if (pass == XLOG_RECOVER_PASS1)
 
2766			error = xlog_recover_commit_pass1(log, trans, item);
2767		else
2768			error = xlog_recover_commit_pass2(log, trans, item);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2769		if (error)
2770			return error;
2771	}
2772
 
 
 
 
 
 
 
 
 
 
 
2773	xlog_recover_free_trans(trans);
2774	return 0;
 
 
2775}
2776
2777STATIC int
2778xlog_recover_unmount_trans(
2779	struct log		*log,
2780	xlog_recover_t		*trans)
2781{
2782	/* Do nothing now */
2783	xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
2784	return 0;
2785}
2786
2787/*
2788 * There are two valid states of the r_state field.  0 indicates that the
2789 * transaction structure is in a normal state.  We have either seen the
2790 * start of the transaction or the last operation we added was not a partial
2791 * operation.  If the last operation we added to the transaction was a
2792 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2793 *
2794 * NOTE: skip LRs with 0 data length.
2795 */
2796STATIC int
2797xlog_recover_process_data(
2798	xlog_t			*log,
2799	struct hlist_head	rhash[],
2800	xlog_rec_header_t	*rhead,
2801	xfs_caddr_t		dp,
2802	int			pass)
2803{
2804	xfs_caddr_t		lp;
2805	int			num_logops;
2806	xlog_op_header_t	*ohead;
2807	xlog_recover_t		*trans;
2808	xlog_tid_t		tid;
2809	int			error;
2810	unsigned long		hash;
2811	uint			flags;
2812
2813	lp = dp + be32_to_cpu(rhead->h_len);
2814	num_logops = be32_to_cpu(rhead->h_num_logops);
2815
2816	/* check the log format matches our own - else we can't recover */
2817	if (xlog_header_check_recover(log->l_mp, rhead))
2818		return (XFS_ERROR(EIO));
2819
2820	while ((dp < lp) && num_logops) {
2821		ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2822		ohead = (xlog_op_header_t *)dp;
2823		dp += sizeof(xlog_op_header_t);
2824		if (ohead->oh_clientid != XFS_TRANSACTION &&
2825		    ohead->oh_clientid != XFS_LOG) {
2826			xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
2827					__func__, ohead->oh_clientid);
2828			ASSERT(0);
2829			return (XFS_ERROR(EIO));
2830		}
2831		tid = be32_to_cpu(ohead->oh_tid);
2832		hash = XLOG_RHASH(tid);
2833		trans = xlog_recover_find_tid(&rhash[hash], tid);
2834		if (trans == NULL) {		   /* not found; add new tid */
2835			if (ohead->oh_flags & XLOG_START_TRANS)
2836				xlog_recover_new_tid(&rhash[hash], tid,
2837					be64_to_cpu(rhead->h_lsn));
2838		} else {
2839			if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2840				xfs_warn(log->l_mp, "%s: bad length 0x%x",
2841					__func__, be32_to_cpu(ohead->oh_len));
2842				WARN_ON(1);
2843				return (XFS_ERROR(EIO));
2844			}
2845			flags = ohead->oh_flags & ~XLOG_END_TRANS;
2846			if (flags & XLOG_WAS_CONT_TRANS)
2847				flags &= ~XLOG_CONTINUE_TRANS;
2848			switch (flags) {
2849			case XLOG_COMMIT_TRANS:
2850				error = xlog_recover_commit_trans(log,
2851								trans, pass);
2852				break;
2853			case XLOG_UNMOUNT_TRANS:
2854				error = xlog_recover_unmount_trans(log, trans);
2855				break;
2856			case XLOG_WAS_CONT_TRANS:
2857				error = xlog_recover_add_to_cont_trans(log,
2858						trans, dp,
2859						be32_to_cpu(ohead->oh_len));
2860				break;
2861			case XLOG_START_TRANS:
2862				xfs_warn(log->l_mp, "%s: bad transaction",
2863					__func__);
2864				ASSERT(0);
2865				error = XFS_ERROR(EIO);
2866				break;
2867			case 0:
2868			case XLOG_CONTINUE_TRANS:
2869				error = xlog_recover_add_to_trans(log, trans,
2870						dp, be32_to_cpu(ohead->oh_len));
2871				break;
2872			default:
2873				xfs_warn(log->l_mp, "%s: bad flag 0x%x",
2874					__func__, flags);
2875				ASSERT(0);
2876				error = XFS_ERROR(EIO);
2877				break;
2878			}
2879			if (error)
 
2880				return error;
 
2881		}
2882		dp += be32_to_cpu(ohead->oh_len);
2883		num_logops--;
2884	}
2885	return 0;
2886}
2887
2888/*
2889 * Process an extent free intent item that was recovered from
2890 * the log.  We need to free the extents that it describes.
2891 */
2892STATIC int
2893xlog_recover_process_efi(
2894	xfs_mount_t		*mp,
2895	xfs_efi_log_item_t	*efip)
2896{
2897	xfs_efd_log_item_t	*efdp;
2898	xfs_trans_t		*tp;
2899	int			i;
2900	int			error = 0;
2901	xfs_extent_t		*extp;
2902	xfs_fsblock_t		startblock_fsb;
2903
2904	ASSERT(!test_bit(XFS_EFI_RECOVERED, &efip->efi_flags));
2905
2906	/*
2907	 * First check the validity of the extents described by the
2908	 * EFI.  If any are bad, then assume that all are bad and
2909	 * just toss the EFI.
2910	 */
2911	for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2912		extp = &(efip->efi_format.efi_extents[i]);
2913		startblock_fsb = XFS_BB_TO_FSB(mp,
2914				   XFS_FSB_TO_DADDR(mp, extp->ext_start));
2915		if ((startblock_fsb == 0) ||
2916		    (extp->ext_len == 0) ||
2917		    (startblock_fsb >= mp->m_sb.sb_dblocks) ||
2918		    (extp->ext_len >= mp->m_sb.sb_agblocks)) {
2919			/*
2920			 * This will pull the EFI from the AIL and
2921			 * free the memory associated with it.
2922			 */
 
2923			xfs_efi_release(efip, efip->efi_format.efi_nextents);
2924			return XFS_ERROR(EIO);
2925		}
2926	}
2927
2928	tp = xfs_trans_alloc(mp, 0);
2929	error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
2930	if (error)
2931		goto abort_error;
2932	efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
2933
2934	for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2935		extp = &(efip->efi_format.efi_extents[i]);
2936		error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
2937		if (error)
2938			goto abort_error;
2939		xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
2940					 extp->ext_len);
2941	}
2942
2943	set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
2944	error = xfs_trans_commit(tp, 0);
2945	return error;
2946
2947abort_error:
2948	xfs_trans_cancel(tp, XFS_TRANS_ABORT);
2949	return error;
2950}
2951
2952/*
2953 * When this is called, all of the EFIs which did not have
2954 * corresponding EFDs should be in the AIL.  What we do now
2955 * is free the extents associated with each one.
2956 *
2957 * Since we process the EFIs in normal transactions, they
2958 * will be removed at some point after the commit.  This prevents
2959 * us from just walking down the list processing each one.
2960 * We'll use a flag in the EFI to skip those that we've already
2961 * processed and use the AIL iteration mechanism's generation
2962 * count to try to speed this up at least a bit.
2963 *
2964 * When we start, we know that the EFIs are the only things in
2965 * the AIL.  As we process them, however, other items are added
2966 * to the AIL.  Since everything added to the AIL must come after
2967 * everything already in the AIL, we stop processing as soon as
2968 * we see something other than an EFI in the AIL.
2969 */
2970STATIC int
2971xlog_recover_process_efis(
2972	xlog_t			*log)
2973{
2974	xfs_log_item_t		*lip;
2975	xfs_efi_log_item_t	*efip;
2976	int			error = 0;
2977	struct xfs_ail_cursor	cur;
2978	struct xfs_ail		*ailp;
2979
2980	ailp = log->l_ailp;
2981	spin_lock(&ailp->xa_lock);
2982	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2983	while (lip != NULL) {
2984		/*
2985		 * We're done when we see something other than an EFI.
2986		 * There should be no EFIs left in the AIL now.
2987		 */
2988		if (lip->li_type != XFS_LI_EFI) {
2989#ifdef DEBUG
2990			for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2991				ASSERT(lip->li_type != XFS_LI_EFI);
2992#endif
2993			break;
2994		}
2995
2996		/*
2997		 * Skip EFIs that we've already processed.
2998		 */
2999		efip = (xfs_efi_log_item_t *)lip;
3000		if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags)) {
3001			lip = xfs_trans_ail_cursor_next(ailp, &cur);
3002			continue;
3003		}
3004
3005		spin_unlock(&ailp->xa_lock);
3006		error = xlog_recover_process_efi(log->l_mp, efip);
3007		spin_lock(&ailp->xa_lock);
3008		if (error)
3009			goto out;
3010		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3011	}
3012out:
3013	xfs_trans_ail_cursor_done(ailp, &cur);
3014	spin_unlock(&ailp->xa_lock);
3015	return error;
3016}
3017
3018/*
3019 * This routine performs a transaction to null out a bad inode pointer
3020 * in an agi unlinked inode hash bucket.
3021 */
3022STATIC void
3023xlog_recover_clear_agi_bucket(
3024	xfs_mount_t	*mp,
3025	xfs_agnumber_t	agno,
3026	int		bucket)
3027{
3028	xfs_trans_t	*tp;
3029	xfs_agi_t	*agi;
3030	xfs_buf_t	*agibp;
3031	int		offset;
3032	int		error;
3033
3034	tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3035	error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp),
3036				  0, 0, 0);
3037	if (error)
3038		goto out_abort;
3039
3040	error = xfs_read_agi(mp, tp, agno, &agibp);
3041	if (error)
3042		goto out_abort;
3043
3044	agi = XFS_BUF_TO_AGI(agibp);
3045	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3046	offset = offsetof(xfs_agi_t, agi_unlinked) +
3047		 (sizeof(xfs_agino_t) * bucket);
3048	xfs_trans_log_buf(tp, agibp, offset,
3049			  (offset + sizeof(xfs_agino_t) - 1));
3050
3051	error = xfs_trans_commit(tp, 0);
3052	if (error)
3053		goto out_error;
3054	return;
3055
3056out_abort:
3057	xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3058out_error:
3059	xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
3060	return;
3061}
3062
3063STATIC xfs_agino_t
3064xlog_recover_process_one_iunlink(
3065	struct xfs_mount		*mp,
3066	xfs_agnumber_t			agno,
3067	xfs_agino_t			agino,
3068	int				bucket)
3069{
3070	struct xfs_buf			*ibp;
3071	struct xfs_dinode		*dip;
3072	struct xfs_inode		*ip;
3073	xfs_ino_t			ino;
3074	int				error;
3075
3076	ino = XFS_AGINO_TO_INO(mp, agno, agino);
3077	error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
3078	if (error)
3079		goto fail;
3080
3081	/*
3082	 * Get the on disk inode to find the next inode in the bucket.
3083	 */
3084	error = xfs_itobp(mp, NULL, ip, &dip, &ibp, XBF_LOCK);
3085	if (error)
3086		goto fail_iput;
3087
3088	ASSERT(ip->i_d.di_nlink == 0);
3089	ASSERT(ip->i_d.di_mode != 0);
3090
3091	/* setup for the next pass */
3092	agino = be32_to_cpu(dip->di_next_unlinked);
3093	xfs_buf_relse(ibp);
3094
3095	/*
3096	 * Prevent any DMAPI event from being sent when the reference on
3097	 * the inode is dropped.
3098	 */
3099	ip->i_d.di_dmevmask = 0;
3100
3101	IRELE(ip);
3102	return agino;
3103
3104 fail_iput:
3105	IRELE(ip);
3106 fail:
3107	/*
3108	 * We can't read in the inode this bucket points to, or this inode
3109	 * is messed up.  Just ditch this bucket of inodes.  We will lose
3110	 * some inodes and space, but at least we won't hang.
3111	 *
3112	 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3113	 * clear the inode pointer in the bucket.
3114	 */
3115	xlog_recover_clear_agi_bucket(mp, agno, bucket);
3116	return NULLAGINO;
3117}
3118
3119/*
3120 * xlog_iunlink_recover
3121 *
3122 * This is called during recovery to process any inodes which
3123 * we unlinked but not freed when the system crashed.  These
3124 * inodes will be on the lists in the AGI blocks.  What we do
3125 * here is scan all the AGIs and fully truncate and free any
3126 * inodes found on the lists.  Each inode is removed from the
3127 * lists when it has been fully truncated and is freed.  The
3128 * freeing of the inode and its removal from the list must be
3129 * atomic.
3130 */
3131STATIC void
3132xlog_recover_process_iunlinks(
3133	xlog_t		*log)
3134{
3135	xfs_mount_t	*mp;
3136	xfs_agnumber_t	agno;
3137	xfs_agi_t	*agi;
3138	xfs_buf_t	*agibp;
3139	xfs_agino_t	agino;
3140	int		bucket;
3141	int		error;
3142	uint		mp_dmevmask;
3143
3144	mp = log->l_mp;
3145
3146	/*
3147	 * Prevent any DMAPI event from being sent while in this function.
3148	 */
3149	mp_dmevmask = mp->m_dmevmask;
3150	mp->m_dmevmask = 0;
3151
3152	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3153		/*
3154		 * Find the agi for this ag.
3155		 */
3156		error = xfs_read_agi(mp, NULL, agno, &agibp);
3157		if (error) {
3158			/*
3159			 * AGI is b0rked. Don't process it.
3160			 *
3161			 * We should probably mark the filesystem as corrupt
3162			 * after we've recovered all the ag's we can....
3163			 */
3164			continue;
3165		}
 
 
 
 
 
 
 
 
 
3166		agi = XFS_BUF_TO_AGI(agibp);
 
3167
3168		for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3169			agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3170			while (agino != NULLAGINO) {
3171				/*
3172				 * Release the agi buffer so that it can
3173				 * be acquired in the normal course of the
3174				 * transaction to truncate and free the inode.
3175				 */
3176				xfs_buf_relse(agibp);
3177
3178				agino = xlog_recover_process_one_iunlink(mp,
3179							agno, agino, bucket);
3180
3181				/*
3182				 * Reacquire the agibuffer and continue around
3183				 * the loop. This should never fail as we know
3184				 * the buffer was good earlier on.
3185				 */
3186				error = xfs_read_agi(mp, NULL, agno, &agibp);
3187				ASSERT(error == 0);
3188				agi = XFS_BUF_TO_AGI(agibp);
3189			}
3190		}
3191
3192		/*
3193		 * Release the buffer for the current agi so we can
3194		 * go on to the next one.
3195		 */
3196		xfs_buf_relse(agibp);
3197	}
3198
3199	mp->m_dmevmask = mp_dmevmask;
3200}
3201
3202
3203#ifdef DEBUG
3204STATIC void
3205xlog_pack_data_checksum(
3206	xlog_t		*log,
3207	xlog_in_core_t	*iclog,
3208	int		size)
3209{
3210	int		i;
3211	__be32		*up;
3212	uint		chksum = 0;
3213
3214	up = (__be32 *)iclog->ic_datap;
3215	/* divide length by 4 to get # words */
3216	for (i = 0; i < (size >> 2); i++) {
3217		chksum ^= be32_to_cpu(*up);
3218		up++;
3219	}
3220	iclog->ic_header.h_chksum = cpu_to_be32(chksum);
3221}
3222#else
3223#define xlog_pack_data_checksum(log, iclog, size)
3224#endif
3225
3226/*
3227 * Stamp cycle number in every block
 
 
 
 
 
 
 
3228 */
3229void
3230xlog_pack_data(
3231	xlog_t			*log,
3232	xlog_in_core_t		*iclog,
3233	int			roundoff)
3234{
3235	int			i, j, k;
3236	int			size = iclog->ic_offset + roundoff;
3237	__be32			cycle_lsn;
3238	xfs_caddr_t		dp;
3239
3240	xlog_pack_data_checksum(log, iclog, size);
3241
3242	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3243
3244	dp = iclog->ic_datap;
3245	for (i = 0; i < BTOBB(size) &&
3246		i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3247		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
3248		*(__be32 *)dp = cycle_lsn;
3249		dp += BBSIZE;
3250	}
3251
3252	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3253		xlog_in_core_2_t *xhdr = iclog->ic_data;
3254
3255		for ( ; i < BTOBB(size); i++) {
3256			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3257			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3258			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
3259			*(__be32 *)dp = cycle_lsn;
3260			dp += BBSIZE;
 
 
3261		}
3262
3263		for (i = 1; i < log->l_iclog_heads; i++) {
3264			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3265		}
 
 
 
 
3266	}
 
 
3267}
3268
3269STATIC void
3270xlog_unpack_data(
3271	xlog_rec_header_t	*rhead,
3272	xfs_caddr_t		dp,
3273	xlog_t			*log)
3274{
3275	int			i, j, k;
 
 
 
 
 
3276
3277	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3278		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3279		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3280		dp += BBSIZE;
3281	}
3282
3283	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3284		xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
3285		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3286			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3287			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3288			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3289			dp += BBSIZE;
3290		}
3291	}
 
 
3292}
3293
3294STATIC int
3295xlog_valid_rec_header(
3296	xlog_t			*log,
3297	xlog_rec_header_t	*rhead,
3298	xfs_daddr_t		blkno)
3299{
3300	int			hlen;
3301
3302	if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
3303		XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3304				XFS_ERRLEVEL_LOW, log->l_mp);
3305		return XFS_ERROR(EFSCORRUPTED);
3306	}
3307	if (unlikely(
3308	    (!rhead->h_version ||
3309	    (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3310		xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
3311			__func__, be32_to_cpu(rhead->h_version));
3312		return XFS_ERROR(EIO);
3313	}
3314
3315	/* LR body must have data or it wouldn't have been written */
3316	hlen = be32_to_cpu(rhead->h_len);
3317	if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3318		XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3319				XFS_ERRLEVEL_LOW, log->l_mp);
3320		return XFS_ERROR(EFSCORRUPTED);
3321	}
3322	if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3323		XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3324				XFS_ERRLEVEL_LOW, log->l_mp);
3325		return XFS_ERROR(EFSCORRUPTED);
3326	}
3327	return 0;
3328}
3329
3330/*
3331 * Read the log from tail to head and process the log records found.
3332 * Handle the two cases where the tail and head are in the same cycle
3333 * and where the active portion of the log wraps around the end of
3334 * the physical log separately.  The pass parameter is passed through
3335 * to the routines called to process the data and is not looked at
3336 * here.
3337 */
3338STATIC int
3339xlog_do_recovery_pass(
3340	xlog_t			*log,
3341	xfs_daddr_t		head_blk,
3342	xfs_daddr_t		tail_blk,
3343	int			pass)
3344{
3345	xlog_rec_header_t	*rhead;
3346	xfs_daddr_t		blk_no;
3347	xfs_caddr_t		offset;
3348	xfs_buf_t		*hbp, *dbp;
3349	int			error = 0, h_size;
3350	int			bblks, split_bblks;
3351	int			hblks, split_hblks, wrapped_hblks;
3352	struct hlist_head	rhash[XLOG_RHASH_SIZE];
3353
3354	ASSERT(head_blk != tail_blk);
3355
3356	/*
3357	 * Read the header of the tail block and get the iclog buffer size from
3358	 * h_size.  Use this to tell how many sectors make up the log header.
3359	 */
3360	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3361		/*
3362		 * When using variable length iclogs, read first sector of
3363		 * iclog header and extract the header size from it.  Get a
3364		 * new hbp that is the correct size.
3365		 */
3366		hbp = xlog_get_bp(log, 1);
3367		if (!hbp)
3368			return ENOMEM;
3369
3370		error = xlog_bread(log, tail_blk, 1, hbp, &offset);
3371		if (error)
3372			goto bread_err1;
3373
3374		rhead = (xlog_rec_header_t *)offset;
3375		error = xlog_valid_rec_header(log, rhead, tail_blk);
3376		if (error)
3377			goto bread_err1;
3378		h_size = be32_to_cpu(rhead->h_size);
3379		if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3380		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3381			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3382			if (h_size % XLOG_HEADER_CYCLE_SIZE)
3383				hblks++;
3384			xlog_put_bp(hbp);
3385			hbp = xlog_get_bp(log, hblks);
3386		} else {
3387			hblks = 1;
3388		}
3389	} else {
3390		ASSERT(log->l_sectBBsize == 1);
3391		hblks = 1;
3392		hbp = xlog_get_bp(log, 1);
3393		h_size = XLOG_BIG_RECORD_BSIZE;
3394	}
3395
3396	if (!hbp)
3397		return ENOMEM;
3398	dbp = xlog_get_bp(log, BTOBB(h_size));
3399	if (!dbp) {
3400		xlog_put_bp(hbp);
3401		return ENOMEM;
3402	}
3403
3404	memset(rhash, 0, sizeof(rhash));
3405	if (tail_blk <= head_blk) {
3406		for (blk_no = tail_blk; blk_no < head_blk; ) {
3407			error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3408			if (error)
3409				goto bread_err2;
3410
3411			rhead = (xlog_rec_header_t *)offset;
3412			error = xlog_valid_rec_header(log, rhead, blk_no);
3413			if (error)
3414				goto bread_err2;
3415
3416			/* blocks in data section */
3417			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3418			error = xlog_bread(log, blk_no + hblks, bblks, dbp,
3419					   &offset);
3420			if (error)
3421				goto bread_err2;
3422
3423			xlog_unpack_data(rhead, offset, log);
3424			if ((error = xlog_recover_process_data(log,
3425						rhash, rhead, offset, pass)))
 
 
 
 
3426				goto bread_err2;
3427			blk_no += bblks + hblks;
3428		}
3429	} else {
3430		/*
3431		 * Perform recovery around the end of the physical log.
3432		 * When the head is not on the same cycle number as the tail,
3433		 * we can't do a sequential recovery as above.
3434		 */
3435		blk_no = tail_blk;
3436		while (blk_no < log->l_logBBsize) {
3437			/*
3438			 * Check for header wrapping around physical end-of-log
3439			 */
3440			offset = hbp->b_addr;
3441			split_hblks = 0;
3442			wrapped_hblks = 0;
3443			if (blk_no + hblks <= log->l_logBBsize) {
3444				/* Read header in one read */
3445				error = xlog_bread(log, blk_no, hblks, hbp,
3446						   &offset);
3447				if (error)
3448					goto bread_err2;
3449			} else {
3450				/* This LR is split across physical log end */
3451				if (blk_no != log->l_logBBsize) {
3452					/* some data before physical log end */
3453					ASSERT(blk_no <= INT_MAX);
3454					split_hblks = log->l_logBBsize - (int)blk_no;
3455					ASSERT(split_hblks > 0);
3456					error = xlog_bread(log, blk_no,
3457							   split_hblks, hbp,
3458							   &offset);
3459					if (error)
3460						goto bread_err2;
3461				}
3462
3463				/*
3464				 * Note: this black magic still works with
3465				 * large sector sizes (non-512) only because:
3466				 * - we increased the buffer size originally
3467				 *   by 1 sector giving us enough extra space
3468				 *   for the second read;
3469				 * - the log start is guaranteed to be sector
3470				 *   aligned;
3471				 * - we read the log end (LR header start)
3472				 *   _first_, then the log start (LR header end)
3473				 *   - order is important.
3474				 */
3475				wrapped_hblks = hblks - split_hblks;
3476				error = xlog_bread_offset(log, 0,
3477						wrapped_hblks, hbp,
3478						offset + BBTOB(split_hblks));
3479				if (error)
3480					goto bread_err2;
3481			}
3482			rhead = (xlog_rec_header_t *)offset;
3483			error = xlog_valid_rec_header(log, rhead,
3484						split_hblks ? blk_no : 0);
3485			if (error)
3486				goto bread_err2;
3487
3488			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3489			blk_no += hblks;
3490
3491			/* Read in data for log record */
3492			if (blk_no + bblks <= log->l_logBBsize) {
3493				error = xlog_bread(log, blk_no, bblks, dbp,
3494						   &offset);
3495				if (error)
3496					goto bread_err2;
3497			} else {
3498				/* This log record is split across the
3499				 * physical end of log */
3500				offset = dbp->b_addr;
3501				split_bblks = 0;
3502				if (blk_no != log->l_logBBsize) {
3503					/* some data is before the physical
3504					 * end of log */
3505					ASSERT(!wrapped_hblks);
3506					ASSERT(blk_no <= INT_MAX);
3507					split_bblks =
3508						log->l_logBBsize - (int)blk_no;
3509					ASSERT(split_bblks > 0);
3510					error = xlog_bread(log, blk_no,
3511							split_bblks, dbp,
3512							&offset);
3513					if (error)
3514						goto bread_err2;
3515				}
3516
3517				/*
3518				 * Note: this black magic still works with
3519				 * large sector sizes (non-512) only because:
3520				 * - we increased the buffer size originally
3521				 *   by 1 sector giving us enough extra space
3522				 *   for the second read;
3523				 * - the log start is guaranteed to be sector
3524				 *   aligned;
3525				 * - we read the log end (LR header start)
3526				 *   _first_, then the log start (LR header end)
3527				 *   - order is important.
3528				 */
3529				error = xlog_bread_offset(log, 0,
3530						bblks - split_bblks, hbp,
3531						offset + BBTOB(split_bblks));
3532				if (error)
3533					goto bread_err2;
3534			}
3535			xlog_unpack_data(rhead, offset, log);
3536			if ((error = xlog_recover_process_data(log, rhash,
3537							rhead, offset, pass)))
 
 
 
 
 
3538				goto bread_err2;
3539			blk_no += bblks;
3540		}
3541
3542		ASSERT(blk_no >= log->l_logBBsize);
3543		blk_no -= log->l_logBBsize;
3544
3545		/* read first part of physical log */
3546		while (blk_no < head_blk) {
3547			error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3548			if (error)
3549				goto bread_err2;
3550
3551			rhead = (xlog_rec_header_t *)offset;
3552			error = xlog_valid_rec_header(log, rhead, blk_no);
3553			if (error)
3554				goto bread_err2;
3555
3556			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3557			error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3558					   &offset);
3559			if (error)
3560				goto bread_err2;
3561
3562			xlog_unpack_data(rhead, offset, log);
3563			if ((error = xlog_recover_process_data(log, rhash,
3564							rhead, offset, pass)))
 
 
 
 
3565				goto bread_err2;
3566			blk_no += bblks + hblks;
3567		}
3568	}
3569
3570 bread_err2:
3571	xlog_put_bp(dbp);
3572 bread_err1:
3573	xlog_put_bp(hbp);
3574	return error;
3575}
3576
3577/*
3578 * Do the recovery of the log.  We actually do this in two phases.
3579 * The two passes are necessary in order to implement the function
3580 * of cancelling a record written into the log.  The first pass
3581 * determines those things which have been cancelled, and the
3582 * second pass replays log items normally except for those which
3583 * have been cancelled.  The handling of the replay and cancellations
3584 * takes place in the log item type specific routines.
3585 *
3586 * The table of items which have cancel records in the log is allocated
3587 * and freed at this level, since only here do we know when all of
3588 * the log recovery has been completed.
3589 */
3590STATIC int
3591xlog_do_log_recovery(
3592	xlog_t		*log,
3593	xfs_daddr_t	head_blk,
3594	xfs_daddr_t	tail_blk)
3595{
3596	int		error, i;
3597
3598	ASSERT(head_blk != tail_blk);
3599
3600	/*
3601	 * First do a pass to find all of the cancelled buf log items.
3602	 * Store them in the buf_cancel_table for use in the second pass.
3603	 */
3604	log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
3605						 sizeof(struct list_head),
3606						 KM_SLEEP);
3607	for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3608		INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
3609
3610	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3611				      XLOG_RECOVER_PASS1);
3612	if (error != 0) {
3613		kmem_free(log->l_buf_cancel_table);
3614		log->l_buf_cancel_table = NULL;
3615		return error;
3616	}
3617	/*
3618	 * Then do a second pass to actually recover the items in the log.
3619	 * When it is complete free the table of buf cancel items.
3620	 */
3621	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3622				      XLOG_RECOVER_PASS2);
3623#ifdef DEBUG
3624	if (!error) {
3625		int	i;
3626
3627		for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3628			ASSERT(list_empty(&log->l_buf_cancel_table[i]));
3629	}
3630#endif	/* DEBUG */
3631
3632	kmem_free(log->l_buf_cancel_table);
3633	log->l_buf_cancel_table = NULL;
3634
3635	return error;
3636}
3637
3638/*
3639 * Do the actual recovery
3640 */
3641STATIC int
3642xlog_do_recover(
3643	xlog_t		*log,
3644	xfs_daddr_t	head_blk,
3645	xfs_daddr_t	tail_blk)
3646{
3647	int		error;
3648	xfs_buf_t	*bp;
3649	xfs_sb_t	*sbp;
3650
3651	/*
3652	 * First replay the images in the log.
3653	 */
3654	error = xlog_do_log_recovery(log, head_blk, tail_blk);
3655	if (error) {
3656		return error;
3657	}
3658
3659	XFS_bflush(log->l_mp->m_ddev_targp);
3660
3661	/*
3662	 * If IO errors happened during recovery, bail out.
3663	 */
3664	if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3665		return (EIO);
3666	}
3667
3668	/*
3669	 * We now update the tail_lsn since much of the recovery has completed
3670	 * and there may be space available to use.  If there were no extent
3671	 * or iunlinks, we can free up the entire log and set the tail_lsn to
3672	 * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3673	 * lsn of the last known good LR on disk.  If there are extent frees
3674	 * or iunlinks they will have some entries in the AIL; so we look at
3675	 * the AIL to determine how to set the tail_lsn.
3676	 */
3677	xlog_assign_tail_lsn(log->l_mp);
3678
3679	/*
3680	 * Now that we've finished replaying all buffer and inode
3681	 * updates, re-read in the superblock.
3682	 */
3683	bp = xfs_getsb(log->l_mp, 0);
3684	XFS_BUF_UNDONE(bp);
3685	ASSERT(!(XFS_BUF_ISWRITE(bp)));
3686	ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
3687	XFS_BUF_READ(bp);
3688	XFS_BUF_UNASYNC(bp);
3689	xfsbdstrat(log->l_mp, bp);
 
 
 
 
 
 
 
3690	error = xfs_buf_iowait(bp);
3691	if (error) {
3692		xfs_ioerror_alert("xlog_do_recover",
3693				  log->l_mp, bp, XFS_BUF_ADDR(bp));
3694		ASSERT(0);
3695		xfs_buf_relse(bp);
3696		return error;
3697	}
3698
3699	/* Convert superblock from on-disk format */
3700	sbp = &log->l_mp->m_sb;
3701	xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3702	ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3703	ASSERT(xfs_sb_good_version(sbp));
3704	xfs_buf_relse(bp);
3705
3706	/* We've re-read the superblock so re-initialize per-cpu counters */
3707	xfs_icsb_reinit_counters(log->l_mp);
3708
3709	xlog_recover_check_summary(log);
3710
3711	/* Normal transactions can now occur */
3712	log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3713	return 0;
3714}
3715
3716/*
3717 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3718 *
3719 * Return error or zero.
3720 */
3721int
3722xlog_recover(
3723	xlog_t		*log)
3724{
3725	xfs_daddr_t	head_blk, tail_blk;
3726	int		error;
3727
3728	/* find the tail of the log */
3729	if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3730		return error;
3731
3732	if (tail_blk != head_blk) {
3733		/* There used to be a comment here:
3734		 *
3735		 * disallow recovery on read-only mounts.  note -- mount
3736		 * checks for ENOSPC and turns it into an intelligent
3737		 * error message.
3738		 * ...but this is no longer true.  Now, unless you specify
3739		 * NORECOVERY (in which case this function would never be
3740		 * called), we just go ahead and recover.  We do this all
3741		 * under the vfs layer, so we can get away with it unless
3742		 * the device itself is read-only, in which case we fail.
3743		 */
3744		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3745			return error;
3746		}
3747
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3748		xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
3749				log->l_mp->m_logname ? log->l_mp->m_logname
3750						     : "internal");
3751
3752		error = xlog_do_recover(log, head_blk, tail_blk);
3753		log->l_flags |= XLOG_RECOVERY_NEEDED;
3754	}
3755	return error;
3756}
3757
3758/*
3759 * In the first part of recovery we replay inodes and buffers and build
3760 * up the list of extent free items which need to be processed.  Here
3761 * we process the extent free items and clean up the on disk unlinked
3762 * inode lists.  This is separated from the first part of recovery so
3763 * that the root and real-time bitmap inodes can be read in from disk in
3764 * between the two stages.  This is necessary so that we can free space
3765 * in the real-time portion of the file system.
3766 */
3767int
3768xlog_recover_finish(
3769	xlog_t		*log)
3770{
3771	/*
3772	 * Now we're ready to do the transactions needed for the
3773	 * rest of recovery.  Start with completing all the extent
3774	 * free intent records and then process the unlinked inode
3775	 * lists.  At this point, we essentially run in normal mode
3776	 * except that we're still performing recovery actions
3777	 * rather than accepting new requests.
3778	 */
3779	if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3780		int	error;
3781		error = xlog_recover_process_efis(log);
3782		if (error) {
3783			xfs_alert(log->l_mp, "Failed to recover EFIs");
3784			return error;
3785		}
3786		/*
3787		 * Sync the log to get all the EFIs out of the AIL.
3788		 * This isn't absolutely necessary, but it helps in
3789		 * case the unlink transactions would have problems
3790		 * pushing the EFIs out of the way.
3791		 */
3792		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3793
3794		xlog_recover_process_iunlinks(log);
3795
3796		xlog_recover_check_summary(log);
3797
3798		xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
3799				log->l_mp->m_logname ? log->l_mp->m_logname
3800						     : "internal");
3801		log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3802	} else {
3803		xfs_info(log->l_mp, "Ending clean mount");
3804	}
3805	return 0;
3806}
3807
3808
3809#if defined(DEBUG)
3810/*
3811 * Read all of the agf and agi counters and check that they
3812 * are consistent with the superblock counters.
3813 */
3814void
3815xlog_recover_check_summary(
3816	xlog_t		*log)
3817{
3818	xfs_mount_t	*mp;
3819	xfs_agf_t	*agfp;
3820	xfs_buf_t	*agfbp;
3821	xfs_buf_t	*agibp;
3822	xfs_agnumber_t	agno;
3823	__uint64_t	freeblks;
3824	__uint64_t	itotal;
3825	__uint64_t	ifree;
3826	int		error;
3827
3828	mp = log->l_mp;
3829
3830	freeblks = 0LL;
3831	itotal = 0LL;
3832	ifree = 0LL;
3833	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3834		error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
3835		if (error) {
3836			xfs_alert(mp, "%s agf read failed agno %d error %d",
3837						__func__, agno, error);
3838		} else {
3839			agfp = XFS_BUF_TO_AGF(agfbp);
3840			freeblks += be32_to_cpu(agfp->agf_freeblks) +
3841				    be32_to_cpu(agfp->agf_flcount);
3842			xfs_buf_relse(agfbp);
3843		}
3844
3845		error = xfs_read_agi(mp, NULL, agno, &agibp);
3846		if (error) {
3847			xfs_alert(mp, "%s agi read failed agno %d error %d",
3848						__func__, agno, error);
3849		} else {
3850			struct xfs_agi	*agi = XFS_BUF_TO_AGI(agibp);
3851
3852			itotal += be32_to_cpu(agi->agi_count);
3853			ifree += be32_to_cpu(agi->agi_freecount);
3854			xfs_buf_relse(agibp);
3855		}
3856	}
3857}
3858#endif /* DEBUG */
v3.15
   1/*
   2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3 * All Rights Reserved.
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it would be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write the Free Software Foundation,
  16 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18#include "xfs.h"
  19#include "xfs_fs.h"
  20#include "xfs_shared.h"
  21#include "xfs_format.h"
  22#include "xfs_log_format.h"
  23#include "xfs_trans_resv.h"
  24#include "xfs_bit.h"
 
  25#include "xfs_inum.h"
 
  26#include "xfs_sb.h"
  27#include "xfs_ag.h"
  28#include "xfs_mount.h"
  29#include "xfs_da_format.h"
 
 
 
 
  30#include "xfs_inode.h"
  31#include "xfs_trans.h"
  32#include "xfs_log.h"
 
  33#include "xfs_log_priv.h"
 
  34#include "xfs_log_recover.h"
  35#include "xfs_inode_item.h"
  36#include "xfs_extfree_item.h"
  37#include "xfs_trans_priv.h"
  38#include "xfs_alloc.h"
  39#include "xfs_ialloc.h"
  40#include "xfs_quota.h"
  41#include "xfs_cksum.h"
 
  42#include "xfs_trace.h"
  43#include "xfs_icache.h"
  44#include "xfs_bmap_btree.h"
  45#include "xfs_dinode.h"
  46#include "xfs_error.h"
  47#include "xfs_dir2.h"
  48
  49#define BLK_AVG(blk1, blk2)	((blk1+blk2) >> 1)
  50
  51STATIC int
  52xlog_find_zeroed(
  53	struct xlog	*,
  54	xfs_daddr_t	*);
  55STATIC int
  56xlog_clear_stale_blocks(
  57	struct xlog	*,
  58	xfs_lsn_t);
  59#if defined(DEBUG)
  60STATIC void
  61xlog_recover_check_summary(
  62	struct xlog *);
  63#else
  64#define	xlog_recover_check_summary(log)
  65#endif
  66
  67/*
  68 * This structure is used during recovery to record the buf log items which
  69 * have been canceled and should not be replayed.
  70 */
  71struct xfs_buf_cancel {
  72	xfs_daddr_t		bc_blkno;
  73	uint			bc_len;
  74	int			bc_refcount;
  75	struct list_head	bc_list;
  76};
  77
  78/*
  79 * Sector aligned buffer routines for buffer create/read/write/access
  80 */
  81
  82/*
  83 * Verify the given count of basic blocks is valid number of blocks
  84 * to specify for an operation involving the given XFS log buffer.
  85 * Returns nonzero if the count is valid, 0 otherwise.
  86 */
  87
  88static inline int
  89xlog_buf_bbcount_valid(
  90	struct xlog	*log,
  91	int		bbcount)
  92{
  93	return bbcount > 0 && bbcount <= log->l_logBBsize;
  94}
  95
  96/*
  97 * Allocate a buffer to hold log data.  The buffer needs to be able
  98 * to map to a range of nbblks basic blocks at any valid (basic
  99 * block) offset within the log.
 100 */
 101STATIC xfs_buf_t *
 102xlog_get_bp(
 103	struct xlog	*log,
 104	int		nbblks)
 105{
 106	struct xfs_buf	*bp;
 107
 108	if (!xlog_buf_bbcount_valid(log, nbblks)) {
 109		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 110			nbblks);
 111		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 112		return NULL;
 113	}
 114
 115	/*
 116	 * We do log I/O in units of log sectors (a power-of-2
 117	 * multiple of the basic block size), so we round up the
 118	 * requested size to accommodate the basic blocks required
 119	 * for complete log sectors.
 120	 *
 121	 * In addition, the buffer may be used for a non-sector-
 122	 * aligned block offset, in which case an I/O of the
 123	 * requested size could extend beyond the end of the
 124	 * buffer.  If the requested size is only 1 basic block it
 125	 * will never straddle a sector boundary, so this won't be
 126	 * an issue.  Nor will this be a problem if the log I/O is
 127	 * done in basic blocks (sector size 1).  But otherwise we
 128	 * extend the buffer by one extra log sector to ensure
 129	 * there's space to accommodate this possibility.
 130	 */
 131	if (nbblks > 1 && log->l_sectBBsize > 1)
 132		nbblks += log->l_sectBBsize;
 133	nbblks = round_up(nbblks, log->l_sectBBsize);
 134
 135	bp = xfs_buf_get_uncached(log->l_mp->m_logdev_targp, nbblks, 0);
 136	if (bp)
 137		xfs_buf_unlock(bp);
 138	return bp;
 139}
 140
 141STATIC void
 142xlog_put_bp(
 143	xfs_buf_t	*bp)
 144{
 145	xfs_buf_free(bp);
 146}
 147
 148/*
 149 * Return the address of the start of the given block number's data
 150 * in a log buffer.  The buffer covers a log sector-aligned region.
 151 */
 152STATIC xfs_caddr_t
 153xlog_align(
 154	struct xlog	*log,
 155	xfs_daddr_t	blk_no,
 156	int		nbblks,
 157	struct xfs_buf	*bp)
 158{
 159	xfs_daddr_t	offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
 160
 161	ASSERT(offset + nbblks <= bp->b_length);
 162	return bp->b_addr + BBTOB(offset);
 163}
 164
 165
 166/*
 167 * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
 168 */
 169STATIC int
 170xlog_bread_noalign(
 171	struct xlog	*log,
 172	xfs_daddr_t	blk_no,
 173	int		nbblks,
 174	struct xfs_buf	*bp)
 175{
 176	int		error;
 177
 178	if (!xlog_buf_bbcount_valid(log, nbblks)) {
 179		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 180			nbblks);
 181		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 182		return EFSCORRUPTED;
 183	}
 184
 185	blk_no = round_down(blk_no, log->l_sectBBsize);
 186	nbblks = round_up(nbblks, log->l_sectBBsize);
 187
 188	ASSERT(nbblks > 0);
 189	ASSERT(nbblks <= bp->b_length);
 190
 191	XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
 192	XFS_BUF_READ(bp);
 193	bp->b_io_length = nbblks;
 194	bp->b_error = 0;
 195
 196	if (XFS_FORCED_SHUTDOWN(log->l_mp))
 197		return XFS_ERROR(EIO);
 198
 199	xfs_buf_iorequest(bp);
 200	error = xfs_buf_iowait(bp);
 201	if (error)
 202		xfs_buf_ioerror_alert(bp, __func__);
 
 203	return error;
 204}
 205
 206STATIC int
 207xlog_bread(
 208	struct xlog	*log,
 209	xfs_daddr_t	blk_no,
 210	int		nbblks,
 211	struct xfs_buf	*bp,
 212	xfs_caddr_t	*offset)
 213{
 214	int		error;
 215
 216	error = xlog_bread_noalign(log, blk_no, nbblks, bp);
 217	if (error)
 218		return error;
 219
 220	*offset = xlog_align(log, blk_no, nbblks, bp);
 221	return 0;
 222}
 223
 224/*
 225 * Read at an offset into the buffer. Returns with the buffer in it's original
 226 * state regardless of the result of the read.
 227 */
 228STATIC int
 229xlog_bread_offset(
 230	struct xlog	*log,
 231	xfs_daddr_t	blk_no,		/* block to read from */
 232	int		nbblks,		/* blocks to read */
 233	struct xfs_buf	*bp,
 234	xfs_caddr_t	offset)
 235{
 236	xfs_caddr_t	orig_offset = bp->b_addr;
 237	int		orig_len = BBTOB(bp->b_length);
 238	int		error, error2;
 239
 240	error = xfs_buf_associate_memory(bp, offset, BBTOB(nbblks));
 241	if (error)
 242		return error;
 243
 244	error = xlog_bread_noalign(log, blk_no, nbblks, bp);
 245
 246	/* must reset buffer pointer even on error */
 247	error2 = xfs_buf_associate_memory(bp, orig_offset, orig_len);
 248	if (error)
 249		return error;
 250	return error2;
 251}
 252
 253/*
 254 * Write out the buffer at the given block for the given number of blocks.
 255 * The buffer is kept locked across the write and is returned locked.
 256 * This can only be used for synchronous log writes.
 257 */
 258STATIC int
 259xlog_bwrite(
 260	struct xlog	*log,
 261	xfs_daddr_t	blk_no,
 262	int		nbblks,
 263	struct xfs_buf	*bp)
 264{
 265	int		error;
 266
 267	if (!xlog_buf_bbcount_valid(log, nbblks)) {
 268		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
 269			nbblks);
 270		XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
 271		return EFSCORRUPTED;
 272	}
 273
 274	blk_no = round_down(blk_no, log->l_sectBBsize);
 275	nbblks = round_up(nbblks, log->l_sectBBsize);
 276
 277	ASSERT(nbblks > 0);
 278	ASSERT(nbblks <= bp->b_length);
 279
 280	XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
 281	XFS_BUF_ZEROFLAGS(bp);
 282	xfs_buf_hold(bp);
 283	xfs_buf_lock(bp);
 284	bp->b_io_length = nbblks;
 285	bp->b_error = 0;
 286
 287	error = xfs_bwrite(bp);
 288	if (error)
 289		xfs_buf_ioerror_alert(bp, __func__);
 290	xfs_buf_relse(bp);
 291	return error;
 292}
 293
 294#ifdef DEBUG
 295/*
 296 * dump debug superblock and log record information
 297 */
 298STATIC void
 299xlog_header_check_dump(
 300	xfs_mount_t		*mp,
 301	xlog_rec_header_t	*head)
 302{
 303	xfs_debug(mp, "%s:  SB : uuid = %pU, fmt = %d",
 304		__func__, &mp->m_sb.sb_uuid, XLOG_FMT);
 305	xfs_debug(mp, "    log : uuid = %pU, fmt = %d",
 306		&head->h_fs_uuid, be32_to_cpu(head->h_fmt));
 307}
 308#else
 309#define xlog_header_check_dump(mp, head)
 310#endif
 311
 312/*
 313 * check log record header for recovery
 314 */
 315STATIC int
 316xlog_header_check_recover(
 317	xfs_mount_t		*mp,
 318	xlog_rec_header_t	*head)
 319{
 320	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 321
 322	/*
 323	 * IRIX doesn't write the h_fmt field and leaves it zeroed
 324	 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
 325	 * a dirty log created in IRIX.
 326	 */
 327	if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
 328		xfs_warn(mp,
 329	"dirty log written in incompatible format - can't recover");
 330		xlog_header_check_dump(mp, head);
 331		XFS_ERROR_REPORT("xlog_header_check_recover(1)",
 332				 XFS_ERRLEVEL_HIGH, mp);
 333		return XFS_ERROR(EFSCORRUPTED);
 334	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 335		xfs_warn(mp,
 336	"dirty log entry has mismatched uuid - can't recover");
 337		xlog_header_check_dump(mp, head);
 338		XFS_ERROR_REPORT("xlog_header_check_recover(2)",
 339				 XFS_ERRLEVEL_HIGH, mp);
 340		return XFS_ERROR(EFSCORRUPTED);
 341	}
 342	return 0;
 343}
 344
 345/*
 346 * read the head block of the log and check the header
 347 */
 348STATIC int
 349xlog_header_check_mount(
 350	xfs_mount_t		*mp,
 351	xlog_rec_header_t	*head)
 352{
 353	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
 354
 355	if (uuid_is_nil(&head->h_fs_uuid)) {
 356		/*
 357		 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
 358		 * h_fs_uuid is nil, we assume this log was last mounted
 359		 * by IRIX and continue.
 360		 */
 361		xfs_warn(mp, "nil uuid in log - IRIX style log");
 362	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 363		xfs_warn(mp, "log has mismatched uuid - can't recover");
 364		xlog_header_check_dump(mp, head);
 365		XFS_ERROR_REPORT("xlog_header_check_mount",
 366				 XFS_ERRLEVEL_HIGH, mp);
 367		return XFS_ERROR(EFSCORRUPTED);
 368	}
 369	return 0;
 370}
 371
 372STATIC void
 373xlog_recover_iodone(
 374	struct xfs_buf	*bp)
 375{
 376	if (bp->b_error) {
 377		/*
 378		 * We're not going to bother about retrying
 379		 * this during recovery. One strike!
 380		 */
 381		xfs_buf_ioerror_alert(bp, __func__);
 
 
 382		xfs_force_shutdown(bp->b_target->bt_mount,
 383					SHUTDOWN_META_IO_ERROR);
 384	}
 385	bp->b_iodone = NULL;
 386	xfs_buf_ioend(bp, 0);
 387}
 388
 389/*
 390 * This routine finds (to an approximation) the first block in the physical
 391 * log which contains the given cycle.  It uses a binary search algorithm.
 392 * Note that the algorithm can not be perfect because the disk will not
 393 * necessarily be perfect.
 394 */
 395STATIC int
 396xlog_find_cycle_start(
 397	struct xlog	*log,
 398	struct xfs_buf	*bp,
 399	xfs_daddr_t	first_blk,
 400	xfs_daddr_t	*last_blk,
 401	uint		cycle)
 402{
 403	xfs_caddr_t	offset;
 404	xfs_daddr_t	mid_blk;
 405	xfs_daddr_t	end_blk;
 406	uint		mid_cycle;
 407	int		error;
 408
 409	end_blk = *last_blk;
 410	mid_blk = BLK_AVG(first_blk, end_blk);
 411	while (mid_blk != first_blk && mid_blk != end_blk) {
 412		error = xlog_bread(log, mid_blk, 1, bp, &offset);
 413		if (error)
 414			return error;
 415		mid_cycle = xlog_get_cycle(offset);
 416		if (mid_cycle == cycle)
 417			end_blk = mid_blk;   /* last_half_cycle == mid_cycle */
 418		else
 419			first_blk = mid_blk; /* first_half_cycle == mid_cycle */
 420		mid_blk = BLK_AVG(first_blk, end_blk);
 421	}
 422	ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
 423	       (mid_blk == end_blk && mid_blk-1 == first_blk));
 424
 425	*last_blk = end_blk;
 426
 427	return 0;
 428}
 429
 430/*
 431 * Check that a range of blocks does not contain stop_on_cycle_no.
 432 * Fill in *new_blk with the block offset where such a block is
 433 * found, or with -1 (an invalid block number) if there is no such
 434 * block in the range.  The scan needs to occur from front to back
 435 * and the pointer into the region must be updated since a later
 436 * routine will need to perform another test.
 437 */
 438STATIC int
 439xlog_find_verify_cycle(
 440	struct xlog	*log,
 441	xfs_daddr_t	start_blk,
 442	int		nbblks,
 443	uint		stop_on_cycle_no,
 444	xfs_daddr_t	*new_blk)
 445{
 446	xfs_daddr_t	i, j;
 447	uint		cycle;
 448	xfs_buf_t	*bp;
 449	xfs_daddr_t	bufblks;
 450	xfs_caddr_t	buf = NULL;
 451	int		error = 0;
 452
 453	/*
 454	 * Greedily allocate a buffer big enough to handle the full
 455	 * range of basic blocks we'll be examining.  If that fails,
 456	 * try a smaller size.  We need to be able to read at least
 457	 * a log sector, or we're out of luck.
 458	 */
 459	bufblks = 1 << ffs(nbblks);
 460	while (bufblks > log->l_logBBsize)
 461		bufblks >>= 1;
 462	while (!(bp = xlog_get_bp(log, bufblks))) {
 463		bufblks >>= 1;
 464		if (bufblks < log->l_sectBBsize)
 465			return ENOMEM;
 466	}
 467
 468	for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
 469		int	bcount;
 470
 471		bcount = min(bufblks, (start_blk + nbblks - i));
 472
 473		error = xlog_bread(log, i, bcount, bp, &buf);
 474		if (error)
 475			goto out;
 476
 477		for (j = 0; j < bcount; j++) {
 478			cycle = xlog_get_cycle(buf);
 479			if (cycle == stop_on_cycle_no) {
 480				*new_blk = i+j;
 481				goto out;
 482			}
 483
 484			buf += BBSIZE;
 485		}
 486	}
 487
 488	*new_blk = -1;
 489
 490out:
 491	xlog_put_bp(bp);
 492	return error;
 493}
 494
 495/*
 496 * Potentially backup over partial log record write.
 497 *
 498 * In the typical case, last_blk is the number of the block directly after
 499 * a good log record.  Therefore, we subtract one to get the block number
 500 * of the last block in the given buffer.  extra_bblks contains the number
 501 * of blocks we would have read on a previous read.  This happens when the
 502 * last log record is split over the end of the physical log.
 503 *
 504 * extra_bblks is the number of blocks potentially verified on a previous
 505 * call to this routine.
 506 */
 507STATIC int
 508xlog_find_verify_log_record(
 509	struct xlog		*log,
 510	xfs_daddr_t		start_blk,
 511	xfs_daddr_t		*last_blk,
 512	int			extra_bblks)
 513{
 514	xfs_daddr_t		i;
 515	xfs_buf_t		*bp;
 516	xfs_caddr_t		offset = NULL;
 517	xlog_rec_header_t	*head = NULL;
 518	int			error = 0;
 519	int			smallmem = 0;
 520	int			num_blks = *last_blk - start_blk;
 521	int			xhdrs;
 522
 523	ASSERT(start_blk != 0 || *last_blk != start_blk);
 524
 525	if (!(bp = xlog_get_bp(log, num_blks))) {
 526		if (!(bp = xlog_get_bp(log, 1)))
 527			return ENOMEM;
 528		smallmem = 1;
 529	} else {
 530		error = xlog_bread(log, start_blk, num_blks, bp, &offset);
 531		if (error)
 532			goto out;
 533		offset += ((num_blks - 1) << BBSHIFT);
 534	}
 535
 536	for (i = (*last_blk) - 1; i >= 0; i--) {
 537		if (i < start_blk) {
 538			/* valid log record not found */
 539			xfs_warn(log->l_mp,
 540		"Log inconsistent (didn't find previous header)");
 541			ASSERT(0);
 542			error = XFS_ERROR(EIO);
 543			goto out;
 544		}
 545
 546		if (smallmem) {
 547			error = xlog_bread(log, i, 1, bp, &offset);
 548			if (error)
 549				goto out;
 550		}
 551
 552		head = (xlog_rec_header_t *)offset;
 553
 554		if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
 555			break;
 556
 557		if (!smallmem)
 558			offset -= BBSIZE;
 559	}
 560
 561	/*
 562	 * We hit the beginning of the physical log & still no header.  Return
 563	 * to caller.  If caller can handle a return of -1, then this routine
 564	 * will be called again for the end of the physical log.
 565	 */
 566	if (i == -1) {
 567		error = -1;
 568		goto out;
 569	}
 570
 571	/*
 572	 * We have the final block of the good log (the first block
 573	 * of the log record _before_ the head. So we check the uuid.
 574	 */
 575	if ((error = xlog_header_check_mount(log->l_mp, head)))
 576		goto out;
 577
 578	/*
 579	 * We may have found a log record header before we expected one.
 580	 * last_blk will be the 1st block # with a given cycle #.  We may end
 581	 * up reading an entire log record.  In this case, we don't want to
 582	 * reset last_blk.  Only when last_blk points in the middle of a log
 583	 * record do we update last_blk.
 584	 */
 585	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
 586		uint	h_size = be32_to_cpu(head->h_size);
 587
 588		xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
 589		if (h_size % XLOG_HEADER_CYCLE_SIZE)
 590			xhdrs++;
 591	} else {
 592		xhdrs = 1;
 593	}
 594
 595	if (*last_blk - i + extra_bblks !=
 596	    BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
 597		*last_blk = i;
 598
 599out:
 600	xlog_put_bp(bp);
 601	return error;
 602}
 603
 604/*
 605 * Head is defined to be the point of the log where the next log write
 606 * could go.  This means that incomplete LR writes at the end are
 607 * eliminated when calculating the head.  We aren't guaranteed that previous
 608 * LR have complete transactions.  We only know that a cycle number of
 609 * current cycle number -1 won't be present in the log if we start writing
 610 * from our current block number.
 611 *
 612 * last_blk contains the block number of the first block with a given
 613 * cycle number.
 614 *
 615 * Return: zero if normal, non-zero if error.
 616 */
 617STATIC int
 618xlog_find_head(
 619	struct xlog	*log,
 620	xfs_daddr_t	*return_head_blk)
 621{
 622	xfs_buf_t	*bp;
 623	xfs_caddr_t	offset;
 624	xfs_daddr_t	new_blk, first_blk, start_blk, last_blk, head_blk;
 625	int		num_scan_bblks;
 626	uint		first_half_cycle, last_half_cycle;
 627	uint		stop_on_cycle;
 628	int		error, log_bbnum = log->l_logBBsize;
 629
 630	/* Is the end of the log device zeroed? */
 631	if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
 632		*return_head_blk = first_blk;
 633
 634		/* Is the whole lot zeroed? */
 635		if (!first_blk) {
 636			/* Linux XFS shouldn't generate totally zeroed logs -
 637			 * mkfs etc write a dummy unmount record to a fresh
 638			 * log so we can store the uuid in there
 639			 */
 640			xfs_warn(log->l_mp, "totally zeroed log");
 641		}
 642
 643		return 0;
 644	} else if (error) {
 645		xfs_warn(log->l_mp, "empty log check failed");
 646		return error;
 647	}
 648
 649	first_blk = 0;			/* get cycle # of 1st block */
 650	bp = xlog_get_bp(log, 1);
 651	if (!bp)
 652		return ENOMEM;
 653
 654	error = xlog_bread(log, 0, 1, bp, &offset);
 655	if (error)
 656		goto bp_err;
 657
 658	first_half_cycle = xlog_get_cycle(offset);
 659
 660	last_blk = head_blk = log_bbnum - 1;	/* get cycle # of last block */
 661	error = xlog_bread(log, last_blk, 1, bp, &offset);
 662	if (error)
 663		goto bp_err;
 664
 665	last_half_cycle = xlog_get_cycle(offset);
 666	ASSERT(last_half_cycle != 0);
 667
 668	/*
 669	 * If the 1st half cycle number is equal to the last half cycle number,
 670	 * then the entire log is stamped with the same cycle number.  In this
 671	 * case, head_blk can't be set to zero (which makes sense).  The below
 672	 * math doesn't work out properly with head_blk equal to zero.  Instead,
 673	 * we set it to log_bbnum which is an invalid block number, but this
 674	 * value makes the math correct.  If head_blk doesn't changed through
 675	 * all the tests below, *head_blk is set to zero at the very end rather
 676	 * than log_bbnum.  In a sense, log_bbnum and zero are the same block
 677	 * in a circular file.
 678	 */
 679	if (first_half_cycle == last_half_cycle) {
 680		/*
 681		 * In this case we believe that the entire log should have
 682		 * cycle number last_half_cycle.  We need to scan backwards
 683		 * from the end verifying that there are no holes still
 684		 * containing last_half_cycle - 1.  If we find such a hole,
 685		 * then the start of that hole will be the new head.  The
 686		 * simple case looks like
 687		 *        x | x ... | x - 1 | x
 688		 * Another case that fits this picture would be
 689		 *        x | x + 1 | x ... | x
 690		 * In this case the head really is somewhere at the end of the
 691		 * log, as one of the latest writes at the beginning was
 692		 * incomplete.
 693		 * One more case is
 694		 *        x | x + 1 | x ... | x - 1 | x
 695		 * This is really the combination of the above two cases, and
 696		 * the head has to end up at the start of the x-1 hole at the
 697		 * end of the log.
 698		 *
 699		 * In the 256k log case, we will read from the beginning to the
 700		 * end of the log and search for cycle numbers equal to x-1.
 701		 * We don't worry about the x+1 blocks that we encounter,
 702		 * because we know that they cannot be the head since the log
 703		 * started with x.
 704		 */
 705		head_blk = log_bbnum;
 706		stop_on_cycle = last_half_cycle - 1;
 707	} else {
 708		/*
 709		 * In this case we want to find the first block with cycle
 710		 * number matching last_half_cycle.  We expect the log to be
 711		 * some variation on
 712		 *        x + 1 ... | x ... | x
 713		 * The first block with cycle number x (last_half_cycle) will
 714		 * be where the new head belongs.  First we do a binary search
 715		 * for the first occurrence of last_half_cycle.  The binary
 716		 * search may not be totally accurate, so then we scan back
 717		 * from there looking for occurrences of last_half_cycle before
 718		 * us.  If that backwards scan wraps around the beginning of
 719		 * the log, then we look for occurrences of last_half_cycle - 1
 720		 * at the end of the log.  The cases we're looking for look
 721		 * like
 722		 *                               v binary search stopped here
 723		 *        x + 1 ... | x | x + 1 | x ... | x
 724		 *                   ^ but we want to locate this spot
 725		 * or
 726		 *        <---------> less than scan distance
 727		 *        x + 1 ... | x ... | x - 1 | x
 728		 *                           ^ we want to locate this spot
 729		 */
 730		stop_on_cycle = last_half_cycle;
 731		if ((error = xlog_find_cycle_start(log, bp, first_blk,
 732						&head_blk, last_half_cycle)))
 733			goto bp_err;
 734	}
 735
 736	/*
 737	 * Now validate the answer.  Scan back some number of maximum possible
 738	 * blocks and make sure each one has the expected cycle number.  The
 739	 * maximum is determined by the total possible amount of buffering
 740	 * in the in-core log.  The following number can be made tighter if
 741	 * we actually look at the block size of the filesystem.
 742	 */
 743	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
 744	if (head_blk >= num_scan_bblks) {
 745		/*
 746		 * We are guaranteed that the entire check can be performed
 747		 * in one buffer.
 748		 */
 749		start_blk = head_blk - num_scan_bblks;
 750		if ((error = xlog_find_verify_cycle(log,
 751						start_blk, num_scan_bblks,
 752						stop_on_cycle, &new_blk)))
 753			goto bp_err;
 754		if (new_blk != -1)
 755			head_blk = new_blk;
 756	} else {		/* need to read 2 parts of log */
 757		/*
 758		 * We are going to scan backwards in the log in two parts.
 759		 * First we scan the physical end of the log.  In this part
 760		 * of the log, we are looking for blocks with cycle number
 761		 * last_half_cycle - 1.
 762		 * If we find one, then we know that the log starts there, as
 763		 * we've found a hole that didn't get written in going around
 764		 * the end of the physical log.  The simple case for this is
 765		 *        x + 1 ... | x ... | x - 1 | x
 766		 *        <---------> less than scan distance
 767		 * If all of the blocks at the end of the log have cycle number
 768		 * last_half_cycle, then we check the blocks at the start of
 769		 * the log looking for occurrences of last_half_cycle.  If we
 770		 * find one, then our current estimate for the location of the
 771		 * first occurrence of last_half_cycle is wrong and we move
 772		 * back to the hole we've found.  This case looks like
 773		 *        x + 1 ... | x | x + 1 | x ...
 774		 *                               ^ binary search stopped here
 775		 * Another case we need to handle that only occurs in 256k
 776		 * logs is
 777		 *        x + 1 ... | x ... | x+1 | x ...
 778		 *                   ^ binary search stops here
 779		 * In a 256k log, the scan at the end of the log will see the
 780		 * x + 1 blocks.  We need to skip past those since that is
 781		 * certainly not the head of the log.  By searching for
 782		 * last_half_cycle-1 we accomplish that.
 783		 */
 784		ASSERT(head_blk <= INT_MAX &&
 785			(xfs_daddr_t) num_scan_bblks >= head_blk);
 786		start_blk = log_bbnum - (num_scan_bblks - head_blk);
 787		if ((error = xlog_find_verify_cycle(log, start_blk,
 788					num_scan_bblks - (int)head_blk,
 789					(stop_on_cycle - 1), &new_blk)))
 790			goto bp_err;
 791		if (new_blk != -1) {
 792			head_blk = new_blk;
 793			goto validate_head;
 794		}
 795
 796		/*
 797		 * Scan beginning of log now.  The last part of the physical
 798		 * log is good.  This scan needs to verify that it doesn't find
 799		 * the last_half_cycle.
 800		 */
 801		start_blk = 0;
 802		ASSERT(head_blk <= INT_MAX);
 803		if ((error = xlog_find_verify_cycle(log,
 804					start_blk, (int)head_blk,
 805					stop_on_cycle, &new_blk)))
 806			goto bp_err;
 807		if (new_blk != -1)
 808			head_blk = new_blk;
 809	}
 810
 811validate_head:
 812	/*
 813	 * Now we need to make sure head_blk is not pointing to a block in
 814	 * the middle of a log record.
 815	 */
 816	num_scan_bblks = XLOG_REC_SHIFT(log);
 817	if (head_blk >= num_scan_bblks) {
 818		start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
 819
 820		/* start ptr at last block ptr before head_blk */
 821		if ((error = xlog_find_verify_log_record(log, start_blk,
 822							&head_blk, 0)) == -1) {
 823			error = XFS_ERROR(EIO);
 824			goto bp_err;
 825		} else if (error)
 826			goto bp_err;
 827	} else {
 828		start_blk = 0;
 829		ASSERT(head_blk <= INT_MAX);
 830		if ((error = xlog_find_verify_log_record(log, start_blk,
 831							&head_blk, 0)) == -1) {
 832			/* We hit the beginning of the log during our search */
 833			start_blk = log_bbnum - (num_scan_bblks - head_blk);
 834			new_blk = log_bbnum;
 835			ASSERT(start_blk <= INT_MAX &&
 836				(xfs_daddr_t) log_bbnum-start_blk >= 0);
 837			ASSERT(head_blk <= INT_MAX);
 838			if ((error = xlog_find_verify_log_record(log,
 839							start_blk, &new_blk,
 840							(int)head_blk)) == -1) {
 841				error = XFS_ERROR(EIO);
 842				goto bp_err;
 843			} else if (error)
 844				goto bp_err;
 845			if (new_blk != log_bbnum)
 846				head_blk = new_blk;
 847		} else if (error)
 848			goto bp_err;
 849	}
 850
 851	xlog_put_bp(bp);
 852	if (head_blk == log_bbnum)
 853		*return_head_blk = 0;
 854	else
 855		*return_head_blk = head_blk;
 856	/*
 857	 * When returning here, we have a good block number.  Bad block
 858	 * means that during a previous crash, we didn't have a clean break
 859	 * from cycle number N to cycle number N-1.  In this case, we need
 860	 * to find the first block with cycle number N-1.
 861	 */
 862	return 0;
 863
 864 bp_err:
 865	xlog_put_bp(bp);
 866
 867	if (error)
 868		xfs_warn(log->l_mp, "failed to find log head");
 869	return error;
 870}
 871
 872/*
 873 * Find the sync block number or the tail of the log.
 874 *
 875 * This will be the block number of the last record to have its
 876 * associated buffers synced to disk.  Every log record header has
 877 * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
 878 * to get a sync block number.  The only concern is to figure out which
 879 * log record header to believe.
 880 *
 881 * The following algorithm uses the log record header with the largest
 882 * lsn.  The entire log record does not need to be valid.  We only care
 883 * that the header is valid.
 884 *
 885 * We could speed up search by using current head_blk buffer, but it is not
 886 * available.
 887 */
 888STATIC int
 889xlog_find_tail(
 890	struct xlog		*log,
 891	xfs_daddr_t		*head_blk,
 892	xfs_daddr_t		*tail_blk)
 893{
 894	xlog_rec_header_t	*rhead;
 895	xlog_op_header_t	*op_head;
 896	xfs_caddr_t		offset = NULL;
 897	xfs_buf_t		*bp;
 898	int			error, i, found;
 899	xfs_daddr_t		umount_data_blk;
 900	xfs_daddr_t		after_umount_blk;
 901	xfs_lsn_t		tail_lsn;
 902	int			hblks;
 903
 904	found = 0;
 905
 906	/*
 907	 * Find previous log record
 908	 */
 909	if ((error = xlog_find_head(log, head_blk)))
 910		return error;
 911
 912	bp = xlog_get_bp(log, 1);
 913	if (!bp)
 914		return ENOMEM;
 915	if (*head_blk == 0) {				/* special case */
 916		error = xlog_bread(log, 0, 1, bp, &offset);
 917		if (error)
 918			goto done;
 919
 920		if (xlog_get_cycle(offset) == 0) {
 921			*tail_blk = 0;
 922			/* leave all other log inited values alone */
 923			goto done;
 924		}
 925	}
 926
 927	/*
 928	 * Search backwards looking for log record header block
 929	 */
 930	ASSERT(*head_blk < INT_MAX);
 931	for (i = (int)(*head_blk) - 1; i >= 0; i--) {
 932		error = xlog_bread(log, i, 1, bp, &offset);
 933		if (error)
 934			goto done;
 935
 936		if (*(__be32 *)offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 937			found = 1;
 938			break;
 939		}
 940	}
 941	/*
 942	 * If we haven't found the log record header block, start looking
 943	 * again from the end of the physical log.  XXXmiken: There should be
 944	 * a check here to make sure we didn't search more than N blocks in
 945	 * the previous code.
 946	 */
 947	if (!found) {
 948		for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
 949			error = xlog_bread(log, i, 1, bp, &offset);
 950			if (error)
 951				goto done;
 952
 953			if (*(__be32 *)offset ==
 954			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
 955				found = 2;
 956				break;
 957			}
 958		}
 959	}
 960	if (!found) {
 961		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
 962		xlog_put_bp(bp);
 963		ASSERT(0);
 964		return XFS_ERROR(EIO);
 965	}
 966
 967	/* find blk_no of tail of log */
 968	rhead = (xlog_rec_header_t *)offset;
 969	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
 970
 971	/*
 972	 * Reset log values according to the state of the log when we
 973	 * crashed.  In the case where head_blk == 0, we bump curr_cycle
 974	 * one because the next write starts a new cycle rather than
 975	 * continuing the cycle of the last good log record.  At this
 976	 * point we have guaranteed that all partial log records have been
 977	 * accounted for.  Therefore, we know that the last good log record
 978	 * written was complete and ended exactly on the end boundary
 979	 * of the physical log.
 980	 */
 981	log->l_prev_block = i;
 982	log->l_curr_block = (int)*head_blk;
 983	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
 984	if (found == 2)
 985		log->l_curr_cycle++;
 986	atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
 987	atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
 988	xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
 989					BBTOB(log->l_curr_block));
 990	xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
 991					BBTOB(log->l_curr_block));
 992
 993	/*
 994	 * Look for unmount record.  If we find it, then we know there
 995	 * was a clean unmount.  Since 'i' could be the last block in
 996	 * the physical log, we convert to a log block before comparing
 997	 * to the head_blk.
 998	 *
 999	 * Save the current tail lsn to use to pass to
1000	 * xlog_clear_stale_blocks() below.  We won't want to clear the
1001	 * unmount record if there is one, so we pass the lsn of the
1002	 * unmount record rather than the block after it.
1003	 */
1004	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1005		int	h_size = be32_to_cpu(rhead->h_size);
1006		int	h_version = be32_to_cpu(rhead->h_version);
1007
1008		if ((h_version & XLOG_VERSION_2) &&
1009		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1010			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1011			if (h_size % XLOG_HEADER_CYCLE_SIZE)
1012				hblks++;
1013		} else {
1014			hblks = 1;
1015		}
1016	} else {
1017		hblks = 1;
1018	}
1019	after_umount_blk = (i + hblks + (int)
1020		BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
1021	tail_lsn = atomic64_read(&log->l_tail_lsn);
1022	if (*head_blk == after_umount_blk &&
1023	    be32_to_cpu(rhead->h_num_logops) == 1) {
1024		umount_data_blk = (i + hblks) % log->l_logBBsize;
1025		error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
1026		if (error)
1027			goto done;
1028
1029		op_head = (xlog_op_header_t *)offset;
1030		if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1031			/*
1032			 * Set tail and last sync so that newly written
1033			 * log records will point recovery to after the
1034			 * current unmount record.
1035			 */
1036			xlog_assign_atomic_lsn(&log->l_tail_lsn,
1037					log->l_curr_cycle, after_umount_blk);
1038			xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1039					log->l_curr_cycle, after_umount_blk);
1040			*tail_blk = after_umount_blk;
1041
1042			/*
1043			 * Note that the unmount was clean. If the unmount
1044			 * was not clean, we need to know this to rebuild the
1045			 * superblock counters from the perag headers if we
1046			 * have a filesystem using non-persistent counters.
1047			 */
1048			log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1049		}
1050	}
1051
1052	/*
1053	 * Make sure that there are no blocks in front of the head
1054	 * with the same cycle number as the head.  This can happen
1055	 * because we allow multiple outstanding log writes concurrently,
1056	 * and the later writes might make it out before earlier ones.
1057	 *
1058	 * We use the lsn from before modifying it so that we'll never
1059	 * overwrite the unmount record after a clean unmount.
1060	 *
1061	 * Do this only if we are going to recover the filesystem
1062	 *
1063	 * NOTE: This used to say "if (!readonly)"
1064	 * However on Linux, we can & do recover a read-only filesystem.
1065	 * We only skip recovery if NORECOVERY is specified on mount,
1066	 * in which case we would not be here.
1067	 *
1068	 * But... if the -device- itself is readonly, just skip this.
1069	 * We can't recover this device anyway, so it won't matter.
1070	 */
1071	if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
1072		error = xlog_clear_stale_blocks(log, tail_lsn);
1073
1074done:
1075	xlog_put_bp(bp);
1076
1077	if (error)
1078		xfs_warn(log->l_mp, "failed to locate log tail");
1079	return error;
1080}
1081
1082/*
1083 * Is the log zeroed at all?
1084 *
1085 * The last binary search should be changed to perform an X block read
1086 * once X becomes small enough.  You can then search linearly through
1087 * the X blocks.  This will cut down on the number of reads we need to do.
1088 *
1089 * If the log is partially zeroed, this routine will pass back the blkno
1090 * of the first block with cycle number 0.  It won't have a complete LR
1091 * preceding it.
1092 *
1093 * Return:
1094 *	0  => the log is completely written to
1095 *	-1 => use *blk_no as the first block of the log
1096 *	>0 => error has occurred
1097 */
1098STATIC int
1099xlog_find_zeroed(
1100	struct xlog	*log,
1101	xfs_daddr_t	*blk_no)
1102{
1103	xfs_buf_t	*bp;
1104	xfs_caddr_t	offset;
1105	uint	        first_cycle, last_cycle;
1106	xfs_daddr_t	new_blk, last_blk, start_blk;
1107	xfs_daddr_t     num_scan_bblks;
1108	int	        error, log_bbnum = log->l_logBBsize;
1109
1110	*blk_no = 0;
1111
1112	/* check totally zeroed log */
1113	bp = xlog_get_bp(log, 1);
1114	if (!bp)
1115		return ENOMEM;
1116	error = xlog_bread(log, 0, 1, bp, &offset);
1117	if (error)
1118		goto bp_err;
1119
1120	first_cycle = xlog_get_cycle(offset);
1121	if (first_cycle == 0) {		/* completely zeroed log */
1122		*blk_no = 0;
1123		xlog_put_bp(bp);
1124		return -1;
1125	}
1126
1127	/* check partially zeroed log */
1128	error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1129	if (error)
1130		goto bp_err;
1131
1132	last_cycle = xlog_get_cycle(offset);
1133	if (last_cycle != 0) {		/* log completely written to */
1134		xlog_put_bp(bp);
1135		return 0;
1136	} else if (first_cycle != 1) {
1137		/*
1138		 * If the cycle of the last block is zero, the cycle of
1139		 * the first block must be 1. If it's not, maybe we're
1140		 * not looking at a log... Bail out.
1141		 */
1142		xfs_warn(log->l_mp,
1143			"Log inconsistent or not a log (last==0, first!=1)");
1144		error = XFS_ERROR(EINVAL);
1145		goto bp_err;
1146	}
1147
1148	/* we have a partially zeroed log */
1149	last_blk = log_bbnum-1;
1150	if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1151		goto bp_err;
1152
1153	/*
1154	 * Validate the answer.  Because there is no way to guarantee that
1155	 * the entire log is made up of log records which are the same size,
1156	 * we scan over the defined maximum blocks.  At this point, the maximum
1157	 * is not chosen to mean anything special.   XXXmiken
1158	 */
1159	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1160	ASSERT(num_scan_bblks <= INT_MAX);
1161
1162	if (last_blk < num_scan_bblks)
1163		num_scan_bblks = last_blk;
1164	start_blk = last_blk - num_scan_bblks;
1165
1166	/*
1167	 * We search for any instances of cycle number 0 that occur before
1168	 * our current estimate of the head.  What we're trying to detect is
1169	 *        1 ... | 0 | 1 | 0...
1170	 *                       ^ binary search ends here
1171	 */
1172	if ((error = xlog_find_verify_cycle(log, start_blk,
1173					 (int)num_scan_bblks, 0, &new_blk)))
1174		goto bp_err;
1175	if (new_blk != -1)
1176		last_blk = new_blk;
1177
1178	/*
1179	 * Potentially backup over partial log record write.  We don't need
1180	 * to search the end of the log because we know it is zero.
1181	 */
1182	if ((error = xlog_find_verify_log_record(log, start_blk,
1183				&last_blk, 0)) == -1) {
1184	    error = XFS_ERROR(EIO);
1185	    goto bp_err;
1186	} else if (error)
1187	    goto bp_err;
1188
1189	*blk_no = last_blk;
1190bp_err:
1191	xlog_put_bp(bp);
1192	if (error)
1193		return error;
1194	return -1;
1195}
1196
1197/*
1198 * These are simple subroutines used by xlog_clear_stale_blocks() below
1199 * to initialize a buffer full of empty log record headers and write
1200 * them into the log.
1201 */
1202STATIC void
1203xlog_add_record(
1204	struct xlog		*log,
1205	xfs_caddr_t		buf,
1206	int			cycle,
1207	int			block,
1208	int			tail_cycle,
1209	int			tail_block)
1210{
1211	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1212
1213	memset(buf, 0, BBSIZE);
1214	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1215	recp->h_cycle = cpu_to_be32(cycle);
1216	recp->h_version = cpu_to_be32(
1217			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1218	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1219	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1220	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1221	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1222}
1223
1224STATIC int
1225xlog_write_log_records(
1226	struct xlog	*log,
1227	int		cycle,
1228	int		start_block,
1229	int		blocks,
1230	int		tail_cycle,
1231	int		tail_block)
1232{
1233	xfs_caddr_t	offset;
1234	xfs_buf_t	*bp;
1235	int		balign, ealign;
1236	int		sectbb = log->l_sectBBsize;
1237	int		end_block = start_block + blocks;
1238	int		bufblks;
1239	int		error = 0;
1240	int		i, j = 0;
1241
1242	/*
1243	 * Greedily allocate a buffer big enough to handle the full
1244	 * range of basic blocks to be written.  If that fails, try
1245	 * a smaller size.  We need to be able to write at least a
1246	 * log sector, or we're out of luck.
1247	 */
1248	bufblks = 1 << ffs(blocks);
1249	while (bufblks > log->l_logBBsize)
1250		bufblks >>= 1;
1251	while (!(bp = xlog_get_bp(log, bufblks))) {
1252		bufblks >>= 1;
1253		if (bufblks < sectbb)
1254			return ENOMEM;
1255	}
1256
1257	/* We may need to do a read at the start to fill in part of
1258	 * the buffer in the starting sector not covered by the first
1259	 * write below.
1260	 */
1261	balign = round_down(start_block, sectbb);
1262	if (balign != start_block) {
1263		error = xlog_bread_noalign(log, start_block, 1, bp);
1264		if (error)
1265			goto out_put_bp;
1266
1267		j = start_block - balign;
1268	}
1269
1270	for (i = start_block; i < end_block; i += bufblks) {
1271		int		bcount, endcount;
1272
1273		bcount = min(bufblks, end_block - start_block);
1274		endcount = bcount - j;
1275
1276		/* We may need to do a read at the end to fill in part of
1277		 * the buffer in the final sector not covered by the write.
1278		 * If this is the same sector as the above read, skip it.
1279		 */
1280		ealign = round_down(end_block, sectbb);
1281		if (j == 0 && (start_block + endcount > ealign)) {
1282			offset = bp->b_addr + BBTOB(ealign - start_block);
1283			error = xlog_bread_offset(log, ealign, sectbb,
1284							bp, offset);
1285			if (error)
1286				break;
1287
1288		}
1289
1290		offset = xlog_align(log, start_block, endcount, bp);
1291		for (; j < endcount; j++) {
1292			xlog_add_record(log, offset, cycle, i+j,
1293					tail_cycle, tail_block);
1294			offset += BBSIZE;
1295		}
1296		error = xlog_bwrite(log, start_block, endcount, bp);
1297		if (error)
1298			break;
1299		start_block += endcount;
1300		j = 0;
1301	}
1302
1303 out_put_bp:
1304	xlog_put_bp(bp);
1305	return error;
1306}
1307
1308/*
1309 * This routine is called to blow away any incomplete log writes out
1310 * in front of the log head.  We do this so that we won't become confused
1311 * if we come up, write only a little bit more, and then crash again.
1312 * If we leave the partial log records out there, this situation could
1313 * cause us to think those partial writes are valid blocks since they
1314 * have the current cycle number.  We get rid of them by overwriting them
1315 * with empty log records with the old cycle number rather than the
1316 * current one.
1317 *
1318 * The tail lsn is passed in rather than taken from
1319 * the log so that we will not write over the unmount record after a
1320 * clean unmount in a 512 block log.  Doing so would leave the log without
1321 * any valid log records in it until a new one was written.  If we crashed
1322 * during that time we would not be able to recover.
1323 */
1324STATIC int
1325xlog_clear_stale_blocks(
1326	struct xlog	*log,
1327	xfs_lsn_t	tail_lsn)
1328{
1329	int		tail_cycle, head_cycle;
1330	int		tail_block, head_block;
1331	int		tail_distance, max_distance;
1332	int		distance;
1333	int		error;
1334
1335	tail_cycle = CYCLE_LSN(tail_lsn);
1336	tail_block = BLOCK_LSN(tail_lsn);
1337	head_cycle = log->l_curr_cycle;
1338	head_block = log->l_curr_block;
1339
1340	/*
1341	 * Figure out the distance between the new head of the log
1342	 * and the tail.  We want to write over any blocks beyond the
1343	 * head that we may have written just before the crash, but
1344	 * we don't want to overwrite the tail of the log.
1345	 */
1346	if (head_cycle == tail_cycle) {
1347		/*
1348		 * The tail is behind the head in the physical log,
1349		 * so the distance from the head to the tail is the
1350		 * distance from the head to the end of the log plus
1351		 * the distance from the beginning of the log to the
1352		 * tail.
1353		 */
1354		if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1355			XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1356					 XFS_ERRLEVEL_LOW, log->l_mp);
1357			return XFS_ERROR(EFSCORRUPTED);
1358		}
1359		tail_distance = tail_block + (log->l_logBBsize - head_block);
1360	} else {
1361		/*
1362		 * The head is behind the tail in the physical log,
1363		 * so the distance from the head to the tail is just
1364		 * the tail block minus the head block.
1365		 */
1366		if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1367			XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1368					 XFS_ERRLEVEL_LOW, log->l_mp);
1369			return XFS_ERROR(EFSCORRUPTED);
1370		}
1371		tail_distance = tail_block - head_block;
1372	}
1373
1374	/*
1375	 * If the head is right up against the tail, we can't clear
1376	 * anything.
1377	 */
1378	if (tail_distance <= 0) {
1379		ASSERT(tail_distance == 0);
1380		return 0;
1381	}
1382
1383	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1384	/*
1385	 * Take the smaller of the maximum amount of outstanding I/O
1386	 * we could have and the distance to the tail to clear out.
1387	 * We take the smaller so that we don't overwrite the tail and
1388	 * we don't waste all day writing from the head to the tail
1389	 * for no reason.
1390	 */
1391	max_distance = MIN(max_distance, tail_distance);
1392
1393	if ((head_block + max_distance) <= log->l_logBBsize) {
1394		/*
1395		 * We can stomp all the blocks we need to without
1396		 * wrapping around the end of the log.  Just do it
1397		 * in a single write.  Use the cycle number of the
1398		 * current cycle minus one so that the log will look like:
1399		 *     n ... | n - 1 ...
1400		 */
1401		error = xlog_write_log_records(log, (head_cycle - 1),
1402				head_block, max_distance, tail_cycle,
1403				tail_block);
1404		if (error)
1405			return error;
1406	} else {
1407		/*
1408		 * We need to wrap around the end of the physical log in
1409		 * order to clear all the blocks.  Do it in two separate
1410		 * I/Os.  The first write should be from the head to the
1411		 * end of the physical log, and it should use the current
1412		 * cycle number minus one just like above.
1413		 */
1414		distance = log->l_logBBsize - head_block;
1415		error = xlog_write_log_records(log, (head_cycle - 1),
1416				head_block, distance, tail_cycle,
1417				tail_block);
1418
1419		if (error)
1420			return error;
1421
1422		/*
1423		 * Now write the blocks at the start of the physical log.
1424		 * This writes the remainder of the blocks we want to clear.
1425		 * It uses the current cycle number since we're now on the
1426		 * same cycle as the head so that we get:
1427		 *    n ... n ... | n - 1 ...
1428		 *    ^^^^^ blocks we're writing
1429		 */
1430		distance = max_distance - (log->l_logBBsize - head_block);
1431		error = xlog_write_log_records(log, head_cycle, 0, distance,
1432				tail_cycle, tail_block);
1433		if (error)
1434			return error;
1435	}
1436
1437	return 0;
1438}
1439
1440/******************************************************************************
1441 *
1442 *		Log recover routines
1443 *
1444 ******************************************************************************
1445 */
1446
1447STATIC xlog_recover_t *
1448xlog_recover_find_tid(
1449	struct hlist_head	*head,
1450	xlog_tid_t		tid)
1451{
1452	xlog_recover_t		*trans;
 
1453
1454	hlist_for_each_entry(trans, head, r_list) {
1455		if (trans->r_log_tid == tid)
1456			return trans;
1457	}
1458	return NULL;
1459}
1460
1461STATIC void
1462xlog_recover_new_tid(
1463	struct hlist_head	*head,
1464	xlog_tid_t		tid,
1465	xfs_lsn_t		lsn)
1466{
1467	xlog_recover_t		*trans;
1468
1469	trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1470	trans->r_log_tid   = tid;
1471	trans->r_lsn	   = lsn;
1472	INIT_LIST_HEAD(&trans->r_itemq);
1473
1474	INIT_HLIST_NODE(&trans->r_list);
1475	hlist_add_head(&trans->r_list, head);
1476}
1477
1478STATIC void
1479xlog_recover_add_item(
1480	struct list_head	*head)
1481{
1482	xlog_recover_item_t	*item;
1483
1484	item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1485	INIT_LIST_HEAD(&item->ri_list);
1486	list_add_tail(&item->ri_list, head);
1487}
1488
1489STATIC int
1490xlog_recover_add_to_cont_trans(
1491	struct xlog		*log,
1492	struct xlog_recover	*trans,
1493	xfs_caddr_t		dp,
1494	int			len)
1495{
1496	xlog_recover_item_t	*item;
1497	xfs_caddr_t		ptr, old_ptr;
1498	int			old_len;
1499
1500	if (list_empty(&trans->r_itemq)) {
1501		/* finish copying rest of trans header */
1502		xlog_recover_add_item(&trans->r_itemq);
1503		ptr = (xfs_caddr_t) &trans->r_theader +
1504				sizeof(xfs_trans_header_t) - len;
1505		memcpy(ptr, dp, len); /* d, s, l */
1506		return 0;
1507	}
1508	/* take the tail entry */
1509	item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1510
1511	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1512	old_len = item->ri_buf[item->ri_cnt-1].i_len;
1513
1514	ptr = kmem_realloc(old_ptr, len+old_len, old_len, KM_SLEEP);
1515	memcpy(&ptr[old_len], dp, len); /* d, s, l */
1516	item->ri_buf[item->ri_cnt-1].i_len += len;
1517	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1518	trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
1519	return 0;
1520}
1521
1522/*
1523 * The next region to add is the start of a new region.  It could be
1524 * a whole region or it could be the first part of a new region.  Because
1525 * of this, the assumption here is that the type and size fields of all
1526 * format structures fit into the first 32 bits of the structure.
1527 *
1528 * This works because all regions must be 32 bit aligned.  Therefore, we
1529 * either have both fields or we have neither field.  In the case we have
1530 * neither field, the data part of the region is zero length.  We only have
1531 * a log_op_header and can throw away the header since a new one will appear
1532 * later.  If we have at least 4 bytes, then we can determine how many regions
1533 * will appear in the current log item.
1534 */
1535STATIC int
1536xlog_recover_add_to_trans(
1537	struct xlog		*log,
1538	struct xlog_recover	*trans,
1539	xfs_caddr_t		dp,
1540	int			len)
1541{
1542	xfs_inode_log_format_t	*in_f;			/* any will do */
1543	xlog_recover_item_t	*item;
1544	xfs_caddr_t		ptr;
1545
1546	if (!len)
1547		return 0;
1548	if (list_empty(&trans->r_itemq)) {
1549		/* we need to catch log corruptions here */
1550		if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1551			xfs_warn(log->l_mp, "%s: bad header magic number",
1552				__func__);
1553			ASSERT(0);
1554			return XFS_ERROR(EIO);
1555		}
1556		if (len == sizeof(xfs_trans_header_t))
1557			xlog_recover_add_item(&trans->r_itemq);
1558		memcpy(&trans->r_theader, dp, len); /* d, s, l */
1559		return 0;
1560	}
1561
1562	ptr = kmem_alloc(len, KM_SLEEP);
1563	memcpy(ptr, dp, len);
1564	in_f = (xfs_inode_log_format_t *)ptr;
1565
1566	/* take the tail entry */
1567	item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1568	if (item->ri_total != 0 &&
1569	     item->ri_total == item->ri_cnt) {
1570		/* tail item is in use, get a new one */
1571		xlog_recover_add_item(&trans->r_itemq);
1572		item = list_entry(trans->r_itemq.prev,
1573					xlog_recover_item_t, ri_list);
1574	}
1575
1576	if (item->ri_total == 0) {		/* first region to be added */
1577		if (in_f->ilf_size == 0 ||
1578		    in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
1579			xfs_warn(log->l_mp,
1580		"bad number of regions (%d) in inode log format",
1581				  in_f->ilf_size);
1582			ASSERT(0);
1583			kmem_free(ptr);
1584			return XFS_ERROR(EIO);
1585		}
1586
1587		item->ri_total = in_f->ilf_size;
1588		item->ri_buf =
1589			kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
1590				    KM_SLEEP);
1591	}
1592	ASSERT(item->ri_total > item->ri_cnt);
1593	/* Description region is ri_buf[0] */
1594	item->ri_buf[item->ri_cnt].i_addr = ptr;
1595	item->ri_buf[item->ri_cnt].i_len  = len;
1596	item->ri_cnt++;
1597	trace_xfs_log_recover_item_add(log, trans, item, 0);
1598	return 0;
1599}
1600
1601/*
1602 * Sort the log items in the transaction.
1603 *
1604 * The ordering constraints are defined by the inode allocation and unlink
1605 * behaviour. The rules are:
1606 *
1607 *	1. Every item is only logged once in a given transaction. Hence it
1608 *	   represents the last logged state of the item. Hence ordering is
1609 *	   dependent on the order in which operations need to be performed so
1610 *	   required initial conditions are always met.
1611 *
1612 *	2. Cancelled buffers are recorded in pass 1 in a separate table and
1613 *	   there's nothing to replay from them so we can simply cull them
1614 *	   from the transaction. However, we can't do that until after we've
1615 *	   replayed all the other items because they may be dependent on the
1616 *	   cancelled buffer and replaying the cancelled buffer can remove it
1617 *	   form the cancelled buffer table. Hence they have tobe done last.
1618 *
1619 *	3. Inode allocation buffers must be replayed before inode items that
1620 *	   read the buffer and replay changes into it. For filesystems using the
1621 *	   ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1622 *	   treated the same as inode allocation buffers as they create and
1623 *	   initialise the buffers directly.
1624 *
1625 *	4. Inode unlink buffers must be replayed after inode items are replayed.
1626 *	   This ensures that inodes are completely flushed to the inode buffer
1627 *	   in a "free" state before we remove the unlinked inode list pointer.
1628 *
1629 * Hence the ordering needs to be inode allocation buffers first, inode items
1630 * second, inode unlink buffers third and cancelled buffers last.
1631 *
1632 * But there's a problem with that - we can't tell an inode allocation buffer
1633 * apart from a regular buffer, so we can't separate them. We can, however,
1634 * tell an inode unlink buffer from the others, and so we can separate them out
1635 * from all the other buffers and move them to last.
1636 *
1637 * Hence, 4 lists, in order from head to tail:
1638 *	- buffer_list for all buffers except cancelled/inode unlink buffers
1639 *	- item_list for all non-buffer items
1640 *	- inode_buffer_list for inode unlink buffers
1641 *	- cancel_list for the cancelled buffers
1642 *
1643 * Note that we add objects to the tail of the lists so that first-to-last
1644 * ordering is preserved within the lists. Adding objects to the head of the
1645 * list means when we traverse from the head we walk them in last-to-first
1646 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1647 * but for all other items there may be specific ordering that we need to
1648 * preserve.
1649 */
1650STATIC int
1651xlog_recover_reorder_trans(
1652	struct xlog		*log,
1653	struct xlog_recover	*trans,
1654	int			pass)
1655{
1656	xlog_recover_item_t	*item, *n;
1657	int			error = 0;
1658	LIST_HEAD(sort_list);
1659	LIST_HEAD(cancel_list);
1660	LIST_HEAD(buffer_list);
1661	LIST_HEAD(inode_buffer_list);
1662	LIST_HEAD(inode_list);
1663
1664	list_splice_init(&trans->r_itemq, &sort_list);
1665	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1666		xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
1667
1668		switch (ITEM_TYPE(item)) {
1669		case XFS_LI_ICREATE:
1670			list_move_tail(&item->ri_list, &buffer_list);
1671			break;
1672		case XFS_LI_BUF:
1673			if (buf_f->blf_flags & XFS_BLF_CANCEL) {
1674				trace_xfs_log_recover_item_reorder_head(log,
1675							trans, item, pass);
1676				list_move(&item->ri_list, &cancel_list);
1677				break;
1678			}
1679			if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
1680				list_move(&item->ri_list, &inode_buffer_list);
1681				break;
1682			}
1683			list_move_tail(&item->ri_list, &buffer_list);
1684			break;
1685		case XFS_LI_INODE:
1686		case XFS_LI_DQUOT:
1687		case XFS_LI_QUOTAOFF:
1688		case XFS_LI_EFD:
1689		case XFS_LI_EFI:
1690			trace_xfs_log_recover_item_reorder_tail(log,
1691							trans, item, pass);
1692			list_move_tail(&item->ri_list, &inode_list);
1693			break;
1694		default:
1695			xfs_warn(log->l_mp,
1696				"%s: unrecognized type of log operation",
1697				__func__);
1698			ASSERT(0);
1699			/*
1700			 * return the remaining items back to the transaction
1701			 * item list so they can be freed in caller.
1702			 */
1703			if (!list_empty(&sort_list))
1704				list_splice_init(&sort_list, &trans->r_itemq);
1705			error = XFS_ERROR(EIO);
1706			goto out;
1707		}
1708	}
1709out:
1710	ASSERT(list_empty(&sort_list));
1711	if (!list_empty(&buffer_list))
1712		list_splice(&buffer_list, &trans->r_itemq);
1713	if (!list_empty(&inode_list))
1714		list_splice_tail(&inode_list, &trans->r_itemq);
1715	if (!list_empty(&inode_buffer_list))
1716		list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1717	if (!list_empty(&cancel_list))
1718		list_splice_tail(&cancel_list, &trans->r_itemq);
1719	return error;
1720}
1721
1722/*
1723 * Build up the table of buf cancel records so that we don't replay
1724 * cancelled data in the second pass.  For buffer records that are
1725 * not cancel records, there is nothing to do here so we just return.
1726 *
1727 * If we get a cancel record which is already in the table, this indicates
1728 * that the buffer was cancelled multiple times.  In order to ensure
1729 * that during pass 2 we keep the record in the table until we reach its
1730 * last occurrence in the log, we keep a reference count in the cancel
1731 * record in the table to tell us how many times we expect to see this
1732 * record during the second pass.
1733 */
1734STATIC int
1735xlog_recover_buffer_pass1(
1736	struct xlog			*log,
1737	struct xlog_recover_item	*item)
1738{
1739	xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
1740	struct list_head	*bucket;
1741	struct xfs_buf_cancel	*bcp;
1742
1743	/*
1744	 * If this isn't a cancel buffer item, then just return.
1745	 */
1746	if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1747		trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1748		return 0;
1749	}
1750
1751	/*
1752	 * Insert an xfs_buf_cancel record into the hash table of them.
1753	 * If there is already an identical record, bump its reference count.
1754	 */
1755	bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1756	list_for_each_entry(bcp, bucket, bc_list) {
1757		if (bcp->bc_blkno == buf_f->blf_blkno &&
1758		    bcp->bc_len == buf_f->blf_len) {
1759			bcp->bc_refcount++;
1760			trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1761			return 0;
1762		}
1763	}
1764
1765	bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
1766	bcp->bc_blkno = buf_f->blf_blkno;
1767	bcp->bc_len = buf_f->blf_len;
1768	bcp->bc_refcount = 1;
1769	list_add_tail(&bcp->bc_list, bucket);
1770
1771	trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1772	return 0;
1773}
1774
1775/*
1776 * Check to see whether the buffer being recovered has a corresponding
1777 * entry in the buffer cancel record table. If it is, return the cancel
1778 * buffer structure to the caller.
 
 
 
 
 
 
 
 
1779 */
1780STATIC struct xfs_buf_cancel *
1781xlog_peek_buffer_cancelled(
1782	struct xlog		*log,
1783	xfs_daddr_t		blkno,
1784	uint			len,
1785	ushort			flags)
1786{
1787	struct list_head	*bucket;
1788	struct xfs_buf_cancel	*bcp;
1789
1790	if (!log->l_buf_cancel_table) {
1791		/* empty table means no cancelled buffers in the log */
 
 
 
1792		ASSERT(!(flags & XFS_BLF_CANCEL));
1793		return NULL;
1794	}
1795
 
 
 
1796	bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
1797	list_for_each_entry(bcp, bucket, bc_list) {
1798		if (bcp->bc_blkno == blkno && bcp->bc_len == len)
1799			return bcp;
1800	}
1801
1802	/*
1803	 * We didn't find a corresponding entry in the table, so return 0 so
1804	 * that the buffer is NOT cancelled.
1805	 */
1806	ASSERT(!(flags & XFS_BLF_CANCEL));
1807	return NULL;
1808}
1809
1810/*
1811 * If the buffer is being cancelled then return 1 so that it will be cancelled,
1812 * otherwise return 0.  If the buffer is actually a buffer cancel item
1813 * (XFS_BLF_CANCEL is set), then decrement the refcount on the entry in the
1814 * table and remove it from the table if this is the last reference.
1815 *
1816 * We remove the cancel record from the table when we encounter its last
1817 * occurrence in the log so that if the same buffer is re-used again after its
1818 * last cancellation we actually replay the changes made at that point.
1819 */
1820STATIC int
1821xlog_check_buffer_cancelled(
1822	struct xlog		*log,
1823	xfs_daddr_t		blkno,
1824	uint			len,
1825	ushort			flags)
1826{
1827	struct xfs_buf_cancel	*bcp;
1828
1829	bcp = xlog_peek_buffer_cancelled(log, blkno, len, flags);
1830	if (!bcp)
1831		return 0;
1832
 
1833	/*
1834	 * We've go a match, so return 1 so that the recovery of this buffer
1835	 * is cancelled.  If this buffer is actually a buffer cancel log
1836	 * item, then decrement the refcount on the one in the table and
1837	 * remove it if this is the last reference.
1838	 */
1839	if (flags & XFS_BLF_CANCEL) {
1840		if (--bcp->bc_refcount == 0) {
1841			list_del(&bcp->bc_list);
1842			kmem_free(bcp);
1843		}
1844	}
1845	return 1;
1846}
1847
1848/*
1849 * Perform recovery for a buffer full of inodes.  In these buffers, the only
1850 * data which should be recovered is that which corresponds to the
1851 * di_next_unlinked pointers in the on disk inode structures.  The rest of the
1852 * data for the inodes is always logged through the inodes themselves rather
1853 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
1854 *
1855 * The only time when buffers full of inodes are fully recovered is when the
1856 * buffer is full of newly allocated inodes.  In this case the buffer will
1857 * not be marked as an inode buffer and so will be sent to
1858 * xlog_recover_do_reg_buffer() below during recovery.
1859 */
1860STATIC int
1861xlog_recover_do_inode_buffer(
1862	struct xfs_mount	*mp,
1863	xlog_recover_item_t	*item,
1864	struct xfs_buf		*bp,
1865	xfs_buf_log_format_t	*buf_f)
1866{
1867	int			i;
1868	int			item_index = 0;
1869	int			bit = 0;
1870	int			nbits = 0;
1871	int			reg_buf_offset = 0;
1872	int			reg_buf_bytes = 0;
1873	int			next_unlinked_offset;
1874	int			inodes_per_buf;
1875	xfs_agino_t		*logged_nextp;
1876	xfs_agino_t		*buffer_nextp;
1877
1878	trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
1879
1880	/*
1881	 * Post recovery validation only works properly on CRC enabled
1882	 * filesystems.
1883	 */
1884	if (xfs_sb_version_hascrc(&mp->m_sb))
1885		bp->b_ops = &xfs_inode_buf_ops;
1886
1887	inodes_per_buf = BBTOB(bp->b_io_length) >> mp->m_sb.sb_inodelog;
1888	for (i = 0; i < inodes_per_buf; i++) {
1889		next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1890			offsetof(xfs_dinode_t, di_next_unlinked);
1891
1892		while (next_unlinked_offset >=
1893		       (reg_buf_offset + reg_buf_bytes)) {
1894			/*
1895			 * The next di_next_unlinked field is beyond
1896			 * the current logged region.  Find the next
1897			 * logged region that contains or is beyond
1898			 * the current di_next_unlinked field.
1899			 */
1900			bit += nbits;
1901			bit = xfs_next_bit(buf_f->blf_data_map,
1902					   buf_f->blf_map_size, bit);
1903
1904			/*
1905			 * If there are no more logged regions in the
1906			 * buffer, then we're done.
1907			 */
1908			if (bit == -1)
1909				return 0;
1910
1911			nbits = xfs_contig_bits(buf_f->blf_data_map,
1912						buf_f->blf_map_size, bit);
1913			ASSERT(nbits > 0);
1914			reg_buf_offset = bit << XFS_BLF_SHIFT;
1915			reg_buf_bytes = nbits << XFS_BLF_SHIFT;
1916			item_index++;
1917		}
1918
1919		/*
1920		 * If the current logged region starts after the current
1921		 * di_next_unlinked field, then move on to the next
1922		 * di_next_unlinked field.
1923		 */
1924		if (next_unlinked_offset < reg_buf_offset)
1925			continue;
1926
1927		ASSERT(item->ri_buf[item_index].i_addr != NULL);
1928		ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
1929		ASSERT((reg_buf_offset + reg_buf_bytes) <=
1930							BBTOB(bp->b_io_length));
1931
1932		/*
1933		 * The current logged region contains a copy of the
1934		 * current di_next_unlinked field.  Extract its value
1935		 * and copy it to the buffer copy.
1936		 */
1937		logged_nextp = item->ri_buf[item_index].i_addr +
1938				next_unlinked_offset - reg_buf_offset;
1939		if (unlikely(*logged_nextp == 0)) {
1940			xfs_alert(mp,
1941		"Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
1942		"Trying to replay bad (0) inode di_next_unlinked field.",
1943				item, bp);
1944			XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1945					 XFS_ERRLEVEL_LOW, mp);
1946			return XFS_ERROR(EFSCORRUPTED);
1947		}
1948
1949		buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1950					      next_unlinked_offset);
1951		*buffer_nextp = *logged_nextp;
1952
1953		/*
1954		 * If necessary, recalculate the CRC in the on-disk inode. We
1955		 * have to leave the inode in a consistent state for whoever
1956		 * reads it next....
1957		 */
1958		xfs_dinode_calc_crc(mp, (struct xfs_dinode *)
1959				xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
1960
1961	}
1962
1963	return 0;
1964}
1965
1966/*
1967 * V5 filesystems know the age of the buffer on disk being recovered. We can
1968 * have newer objects on disk than we are replaying, and so for these cases we
1969 * don't want to replay the current change as that will make the buffer contents
1970 * temporarily invalid on disk.
1971 *
1972 * The magic number might not match the buffer type we are going to recover
1973 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags.  Hence
1974 * extract the LSN of the existing object in the buffer based on it's current
1975 * magic number.  If we don't recognise the magic number in the buffer, then
1976 * return a LSN of -1 so that the caller knows it was an unrecognised block and
1977 * so can recover the buffer.
1978 *
1979 * Note: we cannot rely solely on magic number matches to determine that the
1980 * buffer has a valid LSN - we also need to verify that it belongs to this
1981 * filesystem, so we need to extract the object's LSN and compare it to that
1982 * which we read from the superblock. If the UUIDs don't match, then we've got a
1983 * stale metadata block from an old filesystem instance that we need to recover
1984 * over the top of.
1985 */
1986static xfs_lsn_t
1987xlog_recover_get_buf_lsn(
1988	struct xfs_mount	*mp,
1989	struct xfs_buf		*bp)
1990{
1991	__uint32_t		magic32;
1992	__uint16_t		magic16;
1993	__uint16_t		magicda;
1994	void			*blk = bp->b_addr;
1995	uuid_t			*uuid;
1996	xfs_lsn_t		lsn = -1;
1997
1998	/* v4 filesystems always recover immediately */
1999	if (!xfs_sb_version_hascrc(&mp->m_sb))
2000		goto recover_immediately;
2001
2002	magic32 = be32_to_cpu(*(__be32 *)blk);
2003	switch (magic32) {
2004	case XFS_ABTB_CRC_MAGIC:
2005	case XFS_ABTC_CRC_MAGIC:
2006	case XFS_ABTB_MAGIC:
2007	case XFS_ABTC_MAGIC:
2008	case XFS_IBT_CRC_MAGIC:
2009	case XFS_IBT_MAGIC: {
2010		struct xfs_btree_block *btb = blk;
2011
2012		lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
2013		uuid = &btb->bb_u.s.bb_uuid;
2014		break;
2015	}
2016	case XFS_BMAP_CRC_MAGIC:
2017	case XFS_BMAP_MAGIC: {
2018		struct xfs_btree_block *btb = blk;
2019
2020		lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
2021		uuid = &btb->bb_u.l.bb_uuid;
2022		break;
2023	}
2024	case XFS_AGF_MAGIC:
2025		lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
2026		uuid = &((struct xfs_agf *)blk)->agf_uuid;
2027		break;
2028	case XFS_AGFL_MAGIC:
2029		lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
2030		uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
2031		break;
2032	case XFS_AGI_MAGIC:
2033		lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
2034		uuid = &((struct xfs_agi *)blk)->agi_uuid;
2035		break;
2036	case XFS_SYMLINK_MAGIC:
2037		lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
2038		uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
2039		break;
2040	case XFS_DIR3_BLOCK_MAGIC:
2041	case XFS_DIR3_DATA_MAGIC:
2042	case XFS_DIR3_FREE_MAGIC:
2043		lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
2044		uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
2045		break;
2046	case XFS_ATTR3_RMT_MAGIC:
2047		lsn = be64_to_cpu(((struct xfs_attr3_rmt_hdr *)blk)->rm_lsn);
2048		uuid = &((struct xfs_attr3_rmt_hdr *)blk)->rm_uuid;
2049		break;
2050	case XFS_SB_MAGIC:
2051		lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
2052		uuid = &((struct xfs_dsb *)blk)->sb_uuid;
2053		break;
2054	default:
2055		break;
2056	}
2057
2058	if (lsn != (xfs_lsn_t)-1) {
2059		if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
2060			goto recover_immediately;
2061		return lsn;
2062	}
2063
2064	magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
2065	switch (magicda) {
2066	case XFS_DIR3_LEAF1_MAGIC:
2067	case XFS_DIR3_LEAFN_MAGIC:
2068	case XFS_DA3_NODE_MAGIC:
2069		lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
2070		uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
2071		break;
2072	default:
2073		break;
2074	}
2075
2076	if (lsn != (xfs_lsn_t)-1) {
2077		if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
2078			goto recover_immediately;
2079		return lsn;
2080	}
2081
2082	/*
2083	 * We do individual object checks on dquot and inode buffers as they
2084	 * have their own individual LSN records. Also, we could have a stale
2085	 * buffer here, so we have to at least recognise these buffer types.
2086	 *
2087	 * A notd complexity here is inode unlinked list processing - it logs
2088	 * the inode directly in the buffer, but we don't know which inodes have
2089	 * been modified, and there is no global buffer LSN. Hence we need to
2090	 * recover all inode buffer types immediately. This problem will be
2091	 * fixed by logical logging of the unlinked list modifications.
2092	 */
2093	magic16 = be16_to_cpu(*(__be16 *)blk);
2094	switch (magic16) {
2095	case XFS_DQUOT_MAGIC:
2096	case XFS_DINODE_MAGIC:
2097		goto recover_immediately;
2098	default:
2099		break;
2100	}
2101
2102	/* unknown buffer contents, recover immediately */
2103
2104recover_immediately:
2105	return (xfs_lsn_t)-1;
2106
2107}
2108
2109/*
2110 * Validate the recovered buffer is of the correct type and attach the
2111 * appropriate buffer operations to them for writeback. Magic numbers are in a
2112 * few places:
2113 *	the first 16 bits of the buffer (inode buffer, dquot buffer),
2114 *	the first 32 bits of the buffer (most blocks),
2115 *	inside a struct xfs_da_blkinfo at the start of the buffer.
2116 */
2117static void
2118xlog_recover_validate_buf_type(
2119	struct xfs_mount	*mp,
2120	struct xfs_buf		*bp,
2121	xfs_buf_log_format_t	*buf_f)
2122{
2123	struct xfs_da_blkinfo	*info = bp->b_addr;
2124	__uint32_t		magic32;
2125	__uint16_t		magic16;
2126	__uint16_t		magicda;
2127
2128	magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
2129	magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
2130	magicda = be16_to_cpu(info->magic);
2131	switch (xfs_blft_from_flags(buf_f)) {
2132	case XFS_BLFT_BTREE_BUF:
2133		switch (magic32) {
2134		case XFS_ABTB_CRC_MAGIC:
2135		case XFS_ABTC_CRC_MAGIC:
2136		case XFS_ABTB_MAGIC:
2137		case XFS_ABTC_MAGIC:
2138			bp->b_ops = &xfs_allocbt_buf_ops;
2139			break;
2140		case XFS_IBT_CRC_MAGIC:
2141		case XFS_IBT_MAGIC:
2142			bp->b_ops = &xfs_inobt_buf_ops;
2143			break;
2144		case XFS_BMAP_CRC_MAGIC:
2145		case XFS_BMAP_MAGIC:
2146			bp->b_ops = &xfs_bmbt_buf_ops;
2147			break;
2148		default:
2149			xfs_warn(mp, "Bad btree block magic!");
2150			ASSERT(0);
2151			break;
2152		}
2153		break;
2154	case XFS_BLFT_AGF_BUF:
2155		if (magic32 != XFS_AGF_MAGIC) {
2156			xfs_warn(mp, "Bad AGF block magic!");
2157			ASSERT(0);
2158			break;
2159		}
2160		bp->b_ops = &xfs_agf_buf_ops;
2161		break;
2162	case XFS_BLFT_AGFL_BUF:
2163		if (!xfs_sb_version_hascrc(&mp->m_sb))
2164			break;
2165		if (magic32 != XFS_AGFL_MAGIC) {
2166			xfs_warn(mp, "Bad AGFL block magic!");
2167			ASSERT(0);
2168			break;
2169		}
2170		bp->b_ops = &xfs_agfl_buf_ops;
2171		break;
2172	case XFS_BLFT_AGI_BUF:
2173		if (magic32 != XFS_AGI_MAGIC) {
2174			xfs_warn(mp, "Bad AGI block magic!");
2175			ASSERT(0);
2176			break;
2177		}
2178		bp->b_ops = &xfs_agi_buf_ops;
2179		break;
2180	case XFS_BLFT_UDQUOT_BUF:
2181	case XFS_BLFT_PDQUOT_BUF:
2182	case XFS_BLFT_GDQUOT_BUF:
2183#ifdef CONFIG_XFS_QUOTA
2184		if (magic16 != XFS_DQUOT_MAGIC) {
2185			xfs_warn(mp, "Bad DQUOT block magic!");
2186			ASSERT(0);
2187			break;
2188		}
2189		bp->b_ops = &xfs_dquot_buf_ops;
2190#else
2191		xfs_alert(mp,
2192	"Trying to recover dquots without QUOTA support built in!");
2193		ASSERT(0);
2194#endif
2195		break;
2196	case XFS_BLFT_DINO_BUF:
2197		/*
2198		 * we get here with inode allocation buffers, not buffers that
2199		 * track unlinked list changes.
2200		 */
2201		if (magic16 != XFS_DINODE_MAGIC) {
2202			xfs_warn(mp, "Bad INODE block magic!");
2203			ASSERT(0);
2204			break;
2205		}
2206		bp->b_ops = &xfs_inode_buf_ops;
2207		break;
2208	case XFS_BLFT_SYMLINK_BUF:
2209		if (magic32 != XFS_SYMLINK_MAGIC) {
2210			xfs_warn(mp, "Bad symlink block magic!");
2211			ASSERT(0);
2212			break;
2213		}
2214		bp->b_ops = &xfs_symlink_buf_ops;
2215		break;
2216	case XFS_BLFT_DIR_BLOCK_BUF:
2217		if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
2218		    magic32 != XFS_DIR3_BLOCK_MAGIC) {
2219			xfs_warn(mp, "Bad dir block magic!");
2220			ASSERT(0);
2221			break;
2222		}
2223		bp->b_ops = &xfs_dir3_block_buf_ops;
2224		break;
2225	case XFS_BLFT_DIR_DATA_BUF:
2226		if (magic32 != XFS_DIR2_DATA_MAGIC &&
2227		    magic32 != XFS_DIR3_DATA_MAGIC) {
2228			xfs_warn(mp, "Bad dir data magic!");
2229			ASSERT(0);
2230			break;
2231		}
2232		bp->b_ops = &xfs_dir3_data_buf_ops;
2233		break;
2234	case XFS_BLFT_DIR_FREE_BUF:
2235		if (magic32 != XFS_DIR2_FREE_MAGIC &&
2236		    magic32 != XFS_DIR3_FREE_MAGIC) {
2237			xfs_warn(mp, "Bad dir3 free magic!");
2238			ASSERT(0);
2239			break;
2240		}
2241		bp->b_ops = &xfs_dir3_free_buf_ops;
2242		break;
2243	case XFS_BLFT_DIR_LEAF1_BUF:
2244		if (magicda != XFS_DIR2_LEAF1_MAGIC &&
2245		    magicda != XFS_DIR3_LEAF1_MAGIC) {
2246			xfs_warn(mp, "Bad dir leaf1 magic!");
2247			ASSERT(0);
2248			break;
2249		}
2250		bp->b_ops = &xfs_dir3_leaf1_buf_ops;
2251		break;
2252	case XFS_BLFT_DIR_LEAFN_BUF:
2253		if (magicda != XFS_DIR2_LEAFN_MAGIC &&
2254		    magicda != XFS_DIR3_LEAFN_MAGIC) {
2255			xfs_warn(mp, "Bad dir leafn magic!");
2256			ASSERT(0);
2257			break;
2258		}
2259		bp->b_ops = &xfs_dir3_leafn_buf_ops;
2260		break;
2261	case XFS_BLFT_DA_NODE_BUF:
2262		if (magicda != XFS_DA_NODE_MAGIC &&
2263		    magicda != XFS_DA3_NODE_MAGIC) {
2264			xfs_warn(mp, "Bad da node magic!");
2265			ASSERT(0);
2266			break;
2267		}
2268		bp->b_ops = &xfs_da3_node_buf_ops;
2269		break;
2270	case XFS_BLFT_ATTR_LEAF_BUF:
2271		if (magicda != XFS_ATTR_LEAF_MAGIC &&
2272		    magicda != XFS_ATTR3_LEAF_MAGIC) {
2273			xfs_warn(mp, "Bad attr leaf magic!");
2274			ASSERT(0);
2275			break;
2276		}
2277		bp->b_ops = &xfs_attr3_leaf_buf_ops;
2278		break;
2279	case XFS_BLFT_ATTR_RMT_BUF:
2280		if (!xfs_sb_version_hascrc(&mp->m_sb))
2281			break;
2282		if (magic32 != XFS_ATTR3_RMT_MAGIC) {
2283			xfs_warn(mp, "Bad attr remote magic!");
2284			ASSERT(0);
2285			break;
2286		}
2287		bp->b_ops = &xfs_attr3_rmt_buf_ops;
2288		break;
2289	case XFS_BLFT_SB_BUF:
2290		if (magic32 != XFS_SB_MAGIC) {
2291			xfs_warn(mp, "Bad SB block magic!");
2292			ASSERT(0);
2293			break;
2294		}
2295		bp->b_ops = &xfs_sb_buf_ops;
2296		break;
2297	default:
2298		xfs_warn(mp, "Unknown buffer type %d!",
2299			 xfs_blft_from_flags(buf_f));
2300		break;
2301	}
2302}
2303
2304/*
2305 * Perform a 'normal' buffer recovery.  Each logged region of the
2306 * buffer should be copied over the corresponding region in the
2307 * given buffer.  The bitmap in the buf log format structure indicates
2308 * where to place the logged data.
2309 */
2310STATIC void
2311xlog_recover_do_reg_buffer(
2312	struct xfs_mount	*mp,
2313	xlog_recover_item_t	*item,
2314	struct xfs_buf		*bp,
2315	xfs_buf_log_format_t	*buf_f)
2316{
2317	int			i;
2318	int			bit;
2319	int			nbits;
2320	int                     error;
2321
2322	trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
2323
2324	bit = 0;
2325	i = 1;  /* 0 is the buf format structure */
2326	while (1) {
2327		bit = xfs_next_bit(buf_f->blf_data_map,
2328				   buf_f->blf_map_size, bit);
2329		if (bit == -1)
2330			break;
2331		nbits = xfs_contig_bits(buf_f->blf_data_map,
2332					buf_f->blf_map_size, bit);
2333		ASSERT(nbits > 0);
2334		ASSERT(item->ri_buf[i].i_addr != NULL);
2335		ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
2336		ASSERT(BBTOB(bp->b_io_length) >=
2337		       ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
2338
2339		/*
2340		 * The dirty regions logged in the buffer, even though
2341		 * contiguous, may span multiple chunks. This is because the
2342		 * dirty region may span a physical page boundary in a buffer
2343		 * and hence be split into two separate vectors for writing into
2344		 * the log. Hence we need to trim nbits back to the length of
2345		 * the current region being copied out of the log.
2346		 */
2347		if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
2348			nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
2349
2350		/*
2351		 * Do a sanity check if this is a dquot buffer. Just checking
2352		 * the first dquot in the buffer should do. XXXThis is
2353		 * probably a good thing to do for other buf types also.
2354		 */
2355		error = 0;
2356		if (buf_f->blf_flags &
2357		   (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2358			if (item->ri_buf[i].i_addr == NULL) {
2359				xfs_alert(mp,
2360					"XFS: NULL dquot in %s.", __func__);
2361				goto next;
2362			}
2363			if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
2364				xfs_alert(mp,
2365					"XFS: dquot too small (%d) in %s.",
2366					item->ri_buf[i].i_len, __func__);
2367				goto next;
2368			}
2369			error = xfs_dqcheck(mp, item->ri_buf[i].i_addr,
2370					       -1, 0, XFS_QMOPT_DOWARN,
2371					       "dquot_buf_recover");
2372			if (error)
2373				goto next;
2374		}
2375
2376		memcpy(xfs_buf_offset(bp,
2377			(uint)bit << XFS_BLF_SHIFT),	/* dest */
2378			item->ri_buf[i].i_addr,		/* source */
2379			nbits<<XFS_BLF_SHIFT);		/* length */
2380 next:
2381		i++;
2382		bit += nbits;
2383	}
2384
2385	/* Shouldn't be any more regions */
2386	ASSERT(i == item->ri_total);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2387
2388	/*
2389	 * We can only do post recovery validation on items on CRC enabled
2390	 * fielsystems as we need to know when the buffer was written to be able
2391	 * to determine if we should have replayed the item. If we replay old
2392	 * metadata over a newer buffer, then it will enter a temporarily
2393	 * inconsistent state resulting in verification failures. Hence for now
2394	 * just avoid the verification stage for non-crc filesystems
2395	 */
2396	if (xfs_sb_version_hascrc(&mp->m_sb))
2397		xlog_recover_validate_buf_type(mp, bp, buf_f);
 
 
 
 
 
 
 
 
2398}
2399
2400/*
2401 * Perform a dquot buffer recovery.
2402 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
2403 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2404 * Else, treat it as a regular buffer and do recovery.
2405 */
2406STATIC void
2407xlog_recover_do_dquot_buffer(
2408	struct xfs_mount		*mp,
2409	struct xlog			*log,
2410	struct xlog_recover_item	*item,
2411	struct xfs_buf			*bp,
2412	struct xfs_buf_log_format	*buf_f)
2413{
2414	uint			type;
2415
2416	trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2417
2418	/*
2419	 * Filesystems are required to send in quota flags at mount time.
2420	 */
2421	if (mp->m_qflags == 0) {
2422		return;
2423	}
2424
2425	type = 0;
2426	if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2427		type |= XFS_DQ_USER;
2428	if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2429		type |= XFS_DQ_PROJ;
2430	if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2431		type |= XFS_DQ_GROUP;
2432	/*
2433	 * This type of quotas was turned off, so ignore this buffer
2434	 */
2435	if (log->l_quotaoffs_flag & type)
2436		return;
2437
2438	xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2439}
2440
2441/*
2442 * This routine replays a modification made to a buffer at runtime.
2443 * There are actually two types of buffer, regular and inode, which
2444 * are handled differently.  Inode buffers are handled differently
2445 * in that we only recover a specific set of data from them, namely
2446 * the inode di_next_unlinked fields.  This is because all other inode
2447 * data is actually logged via inode records and any data we replay
2448 * here which overlaps that may be stale.
2449 *
2450 * When meta-data buffers are freed at run time we log a buffer item
2451 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2452 * of the buffer in the log should not be replayed at recovery time.
2453 * This is so that if the blocks covered by the buffer are reused for
2454 * file data before we crash we don't end up replaying old, freed
2455 * meta-data into a user's file.
2456 *
2457 * To handle the cancellation of buffer log items, we make two passes
2458 * over the log during recovery.  During the first we build a table of
2459 * those buffers which have been cancelled, and during the second we
2460 * only replay those buffers which do not have corresponding cancel
2461 * records in the table.  See xlog_recover_buffer_pass[1,2] above
2462 * for more details on the implementation of the table of cancel records.
2463 */
2464STATIC int
2465xlog_recover_buffer_pass2(
2466	struct xlog			*log,
2467	struct list_head		*buffer_list,
2468	struct xlog_recover_item	*item,
2469	xfs_lsn_t			current_lsn)
2470{
2471	xfs_buf_log_format_t	*buf_f = item->ri_buf[0].i_addr;
2472	xfs_mount_t		*mp = log->l_mp;
2473	xfs_buf_t		*bp;
2474	int			error;
2475	uint			buf_flags;
2476	xfs_lsn_t		lsn;
2477
2478	/*
2479	 * In this pass we only want to recover all the buffers which have
2480	 * not been cancelled and are not cancellation buffers themselves.
2481	 */
2482	if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2483			buf_f->blf_len, buf_f->blf_flags)) {
2484		trace_xfs_log_recover_buf_cancel(log, buf_f);
2485		return 0;
2486	}
2487
2488	trace_xfs_log_recover_buf_recover(log, buf_f);
2489
2490	buf_flags = 0;
2491	if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
2492		buf_flags |= XBF_UNMAPPED;
2493
2494	bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2495			  buf_flags, NULL);
2496	if (!bp)
2497		return XFS_ERROR(ENOMEM);
2498	error = bp->b_error;
2499	if (error) {
2500		xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#1)");
2501		goto out_release;
 
 
2502	}
2503
2504	/*
2505	 * recover the buffer only if we get an LSN from it and it's less than
2506	 * the lsn of the transaction we are replaying.
2507	 */
2508	lsn = xlog_recover_get_buf_lsn(mp, bp);
2509	if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0)
2510		goto out_release;
2511
2512	if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2513		error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2514	} else if (buf_f->blf_flags &
2515		  (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2516		xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2517	} else {
2518		xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2519	}
2520	if (error)
2521		goto out_release;
2522
2523	/*
2524	 * Perform delayed write on the buffer.  Asynchronous writes will be
2525	 * slower when taking into account all the buffers to be flushed.
2526	 *
2527	 * Also make sure that only inode buffers with good sizes stay in
2528	 * the buffer cache.  The kernel moves inodes in buffers of 1 block
2529	 * or mp->m_inode_cluster_size bytes, whichever is bigger.  The inode
2530	 * buffers in the log can be a different size if the log was generated
2531	 * by an older kernel using unclustered inode buffers or a newer kernel
2532	 * running with a different inode cluster size.  Regardless, if the
2533	 * the inode buffer size isn't MAX(blocksize, mp->m_inode_cluster_size)
2534	 * for *our* value of mp->m_inode_cluster_size, then we need to keep
2535	 * the buffer out of the buffer cache so that the buffer won't
2536	 * overlap with future reads of those inodes.
2537	 */
2538	if (XFS_DINODE_MAGIC ==
2539	    be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2540	    (BBTOB(bp->b_io_length) != MAX(log->l_mp->m_sb.sb_blocksize,
2541			(__uint32_t)log->l_mp->m_inode_cluster_size))) {
2542		xfs_buf_stale(bp);
2543		error = xfs_bwrite(bp);
2544	} else {
2545		ASSERT(bp->b_target->bt_mount == mp);
2546		bp->b_iodone = xlog_recover_iodone;
2547		xfs_buf_delwri_queue(bp, buffer_list);
2548	}
2549
2550out_release:
2551	xfs_buf_relse(bp);
2552	return error;
2553}
2554
2555/*
2556 * Inode fork owner changes
2557 *
2558 * If we have been told that we have to reparent the inode fork, it's because an
2559 * extent swap operation on a CRC enabled filesystem has been done and we are
2560 * replaying it. We need to walk the BMBT of the appropriate fork and change the
2561 * owners of it.
2562 *
2563 * The complexity here is that we don't have an inode context to work with, so
2564 * after we've replayed the inode we need to instantiate one.  This is where the
2565 * fun begins.
2566 *
2567 * We are in the middle of log recovery, so we can't run transactions. That
2568 * means we cannot use cache coherent inode instantiation via xfs_iget(), as
2569 * that will result in the corresponding iput() running the inode through
2570 * xfs_inactive(). If we've just replayed an inode core that changes the link
2571 * count to zero (i.e. it's been unlinked), then xfs_inactive() will run
2572 * transactions (bad!).
2573 *
2574 * So, to avoid this, we instantiate an inode directly from the inode core we've
2575 * just recovered. We have the buffer still locked, and all we really need to
2576 * instantiate is the inode core and the forks being modified. We can do this
2577 * manually, then run the inode btree owner change, and then tear down the
2578 * xfs_inode without having to run any transactions at all.
2579 *
2580 * Also, because we don't have a transaction context available here but need to
2581 * gather all the buffers we modify for writeback so we pass the buffer_list
2582 * instead for the operation to use.
2583 */
2584
2585STATIC int
2586xfs_recover_inode_owner_change(
2587	struct xfs_mount	*mp,
2588	struct xfs_dinode	*dip,
2589	struct xfs_inode_log_format *in_f,
2590	struct list_head	*buffer_list)
2591{
2592	struct xfs_inode	*ip;
2593	int			error;
2594
2595	ASSERT(in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER));
2596
2597	ip = xfs_inode_alloc(mp, in_f->ilf_ino);
2598	if (!ip)
2599		return ENOMEM;
2600
2601	/* instantiate the inode */
2602	xfs_dinode_from_disk(&ip->i_d, dip);
2603	ASSERT(ip->i_d.di_version >= 3);
2604
2605	error = xfs_iformat_fork(ip, dip);
2606	if (error)
2607		goto out_free_ip;
2608
2609
2610	if (in_f->ilf_fields & XFS_ILOG_DOWNER) {
2611		ASSERT(in_f->ilf_fields & XFS_ILOG_DBROOT);
2612		error = xfs_bmbt_change_owner(NULL, ip, XFS_DATA_FORK,
2613					      ip->i_ino, buffer_list);
2614		if (error)
2615			goto out_free_ip;
2616	}
2617
2618	if (in_f->ilf_fields & XFS_ILOG_AOWNER) {
2619		ASSERT(in_f->ilf_fields & XFS_ILOG_ABROOT);
2620		error = xfs_bmbt_change_owner(NULL, ip, XFS_ATTR_FORK,
2621					      ip->i_ino, buffer_list);
2622		if (error)
2623			goto out_free_ip;
2624	}
2625
2626out_free_ip:
2627	xfs_inode_free(ip);
2628	return error;
2629}
2630
2631STATIC int
2632xlog_recover_inode_pass2(
2633	struct xlog			*log,
2634	struct list_head		*buffer_list,
2635	struct xlog_recover_item	*item,
2636	xfs_lsn_t			current_lsn)
2637{
2638	xfs_inode_log_format_t	*in_f;
2639	xfs_mount_t		*mp = log->l_mp;
2640	xfs_buf_t		*bp;
2641	xfs_dinode_t		*dip;
2642	int			len;
2643	xfs_caddr_t		src;
2644	xfs_caddr_t		dest;
2645	int			error;
2646	int			attr_index;
2647	uint			fields;
2648	xfs_icdinode_t		*dicp;
2649	uint			isize;
2650	int			need_free = 0;
2651
2652	if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2653		in_f = item->ri_buf[0].i_addr;
2654	} else {
2655		in_f = kmem_alloc(sizeof(xfs_inode_log_format_t), KM_SLEEP);
2656		need_free = 1;
2657		error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2658		if (error)
2659			goto error;
2660	}
2661
2662	/*
2663	 * Inode buffers can be freed, look out for it,
2664	 * and do not replay the inode.
2665	 */
2666	if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2667					in_f->ilf_len, 0)) {
2668		error = 0;
2669		trace_xfs_log_recover_inode_cancel(log, in_f);
2670		goto error;
2671	}
2672	trace_xfs_log_recover_inode_recover(log, in_f);
2673
2674	bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len, 0,
2675			  &xfs_inode_buf_ops);
2676	if (!bp) {
2677		error = ENOMEM;
2678		goto error;
2679	}
2680	error = bp->b_error;
2681	if (error) {
2682		xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#2)");
2683		goto out_release;
 
 
2684	}
2685	ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2686	dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
2687
2688	/*
2689	 * Make sure the place we're flushing out to really looks
2690	 * like an inode!
2691	 */
2692	if (unlikely(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC))) {
 
2693		xfs_alert(mp,
2694	"%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
2695			__func__, dip, bp, in_f->ilf_ino);
2696		XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2697				 XFS_ERRLEVEL_LOW, mp);
2698		error = EFSCORRUPTED;
2699		goto out_release;
2700	}
2701	dicp = item->ri_buf[1].i_addr;
2702	if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
 
2703		xfs_alert(mp,
2704			"%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
2705			__func__, item, in_f->ilf_ino);
2706		XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2707				 XFS_ERRLEVEL_LOW, mp);
2708		error = EFSCORRUPTED;
2709		goto out_release;
2710	}
2711
2712	/*
2713	 * If the inode has an LSN in it, recover the inode only if it's less
2714	 * than the lsn of the transaction we are replaying. Note: we still
2715	 * need to replay an owner change even though the inode is more recent
2716	 * than the transaction as there is no guarantee that all the btree
2717	 * blocks are more recent than this transaction, too.
2718	 */
2719	if (dip->di_version >= 3) {
2720		xfs_lsn_t	lsn = be64_to_cpu(dip->di_lsn);
2721
2722		if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
2723			trace_xfs_log_recover_inode_skip(log, in_f);
2724			error = 0;
2725			goto out_owner_change;
2726		}
2727	}
2728
2729	/*
2730	 * di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
2731	 * are transactional and if ordering is necessary we can determine that
2732	 * more accurately by the LSN field in the V3 inode core. Don't trust
2733	 * the inode versions we might be changing them here - use the
2734	 * superblock flag to determine whether we need to look at di_flushiter
2735	 * to skip replay when the on disk inode is newer than the log one
2736	 */
2737	if (!xfs_sb_version_hascrc(&mp->m_sb) &&
2738	    dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2739		/*
2740		 * Deal with the wrap case, DI_MAX_FLUSH is less
2741		 * than smaller numbers
2742		 */
2743		if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
2744		    dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2745			/* do nothing */
2746		} else {
 
2747			trace_xfs_log_recover_inode_skip(log, in_f);
2748			error = 0;
2749			goto out_release;
2750		}
2751	}
2752
2753	/* Take the opportunity to reset the flush iteration count */
2754	dicp->di_flushiter = 0;
2755
2756	if (unlikely(S_ISREG(dicp->di_mode))) {
2757		if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2758		    (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2759			XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
2760					 XFS_ERRLEVEL_LOW, mp, dicp);
 
2761			xfs_alert(mp,
2762		"%s: Bad regular inode log record, rec ptr 0x%p, "
2763		"ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2764				__func__, item, dip, bp, in_f->ilf_ino);
2765			error = EFSCORRUPTED;
2766			goto out_release;
2767		}
2768	} else if (unlikely(S_ISDIR(dicp->di_mode))) {
2769		if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2770		    (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2771		    (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2772			XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
2773					     XFS_ERRLEVEL_LOW, mp, dicp);
 
2774			xfs_alert(mp,
2775		"%s: Bad dir inode log record, rec ptr 0x%p, "
2776		"ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2777				__func__, item, dip, bp, in_f->ilf_ino);
2778			error = EFSCORRUPTED;
2779			goto out_release;
2780		}
2781	}
2782	if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2783		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
2784				     XFS_ERRLEVEL_LOW, mp, dicp);
 
2785		xfs_alert(mp,
2786	"%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2787	"dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2788			__func__, item, dip, bp, in_f->ilf_ino,
2789			dicp->di_nextents + dicp->di_anextents,
2790			dicp->di_nblocks);
2791		error = EFSCORRUPTED;
2792		goto out_release;
2793	}
2794	if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2795		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
2796				     XFS_ERRLEVEL_LOW, mp, dicp);
 
2797		xfs_alert(mp,
2798	"%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2799	"dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__,
2800			item, dip, bp, in_f->ilf_ino, dicp->di_forkoff);
2801		error = EFSCORRUPTED;
2802		goto out_release;
2803	}
2804	isize = xfs_icdinode_size(dicp->di_version);
2805	if (unlikely(item->ri_buf[1].i_len > isize)) {
2806		XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
2807				     XFS_ERRLEVEL_LOW, mp, dicp);
 
2808		xfs_alert(mp,
2809			"%s: Bad inode log record length %d, rec ptr 0x%p",
2810			__func__, item->ri_buf[1].i_len, item);
2811		error = EFSCORRUPTED;
2812		goto out_release;
2813	}
2814
2815	/* The core is in in-core format */
2816	xfs_dinode_to_disk(dip, dicp);
2817
2818	/* the rest is in on-disk format */
2819	if (item->ri_buf[1].i_len > isize) {
2820		memcpy((char *)dip + isize,
2821			item->ri_buf[1].i_addr + isize,
2822			item->ri_buf[1].i_len - isize);
2823	}
2824
2825	fields = in_f->ilf_fields;
2826	switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2827	case XFS_ILOG_DEV:
2828		xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
2829		break;
2830	case XFS_ILOG_UUID:
2831		memcpy(XFS_DFORK_DPTR(dip),
2832		       &in_f->ilf_u.ilfu_uuid,
2833		       sizeof(uuid_t));
2834		break;
2835	}
2836
2837	if (in_f->ilf_size == 2)
2838		goto out_owner_change;
2839	len = item->ri_buf[2].i_len;
2840	src = item->ri_buf[2].i_addr;
2841	ASSERT(in_f->ilf_size <= 4);
2842	ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2843	ASSERT(!(fields & XFS_ILOG_DFORK) ||
2844	       (len == in_f->ilf_dsize));
2845
2846	switch (fields & XFS_ILOG_DFORK) {
2847	case XFS_ILOG_DDATA:
2848	case XFS_ILOG_DEXT:
2849		memcpy(XFS_DFORK_DPTR(dip), src, len);
2850		break;
2851
2852	case XFS_ILOG_DBROOT:
2853		xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2854				 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
2855				 XFS_DFORK_DSIZE(dip, mp));
2856		break;
2857
2858	default:
2859		/*
2860		 * There are no data fork flags set.
2861		 */
2862		ASSERT((fields & XFS_ILOG_DFORK) == 0);
2863		break;
2864	}
2865
2866	/*
2867	 * If we logged any attribute data, recover it.  There may or
2868	 * may not have been any other non-core data logged in this
2869	 * transaction.
2870	 */
2871	if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2872		if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2873			attr_index = 3;
2874		} else {
2875			attr_index = 2;
2876		}
2877		len = item->ri_buf[attr_index].i_len;
2878		src = item->ri_buf[attr_index].i_addr;
2879		ASSERT(len == in_f->ilf_asize);
2880
2881		switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2882		case XFS_ILOG_ADATA:
2883		case XFS_ILOG_AEXT:
2884			dest = XFS_DFORK_APTR(dip);
2885			ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2886			memcpy(dest, src, len);
2887			break;
2888
2889		case XFS_ILOG_ABROOT:
2890			dest = XFS_DFORK_APTR(dip);
2891			xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2892					 len, (xfs_bmdr_block_t*)dest,
2893					 XFS_DFORK_ASIZE(dip, mp));
2894			break;
2895
2896		default:
2897			xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
2898			ASSERT(0);
 
2899			error = EIO;
2900			goto out_release;
2901		}
2902	}
2903
2904out_owner_change:
2905	if (in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER))
2906		error = xfs_recover_inode_owner_change(mp, dip, in_f,
2907						       buffer_list);
2908	/* re-generate the checksum. */
2909	xfs_dinode_calc_crc(log->l_mp, dip);
2910
2911	ASSERT(bp->b_target->bt_mount == mp);
2912	bp->b_iodone = xlog_recover_iodone;
2913	xfs_buf_delwri_queue(bp, buffer_list);
2914
2915out_release:
2916	xfs_buf_relse(bp);
2917error:
2918	if (need_free)
2919		kmem_free(in_f);
2920	return XFS_ERROR(error);
2921}
2922
2923/*
2924 * Recover QUOTAOFF records. We simply make a note of it in the xlog
2925 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2926 * of that type.
2927 */
2928STATIC int
2929xlog_recover_quotaoff_pass1(
2930	struct xlog			*log,
2931	struct xlog_recover_item	*item)
2932{
2933	xfs_qoff_logformat_t	*qoff_f = item->ri_buf[0].i_addr;
2934	ASSERT(qoff_f);
2935
2936	/*
2937	 * The logitem format's flag tells us if this was user quotaoff,
2938	 * group/project quotaoff or both.
2939	 */
2940	if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2941		log->l_quotaoffs_flag |= XFS_DQ_USER;
2942	if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2943		log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2944	if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2945		log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2946
2947	return (0);
2948}
2949
2950/*
2951 * Recover a dquot record
2952 */
2953STATIC int
2954xlog_recover_dquot_pass2(
2955	struct xlog			*log,
2956	struct list_head		*buffer_list,
2957	struct xlog_recover_item	*item,
2958	xfs_lsn_t			current_lsn)
2959{
2960	xfs_mount_t		*mp = log->l_mp;
2961	xfs_buf_t		*bp;
2962	struct xfs_disk_dquot	*ddq, *recddq;
2963	int			error;
2964	xfs_dq_logformat_t	*dq_f;
2965	uint			type;
2966
2967
2968	/*
2969	 * Filesystems are required to send in quota flags at mount time.
2970	 */
2971	if (mp->m_qflags == 0)
2972		return (0);
2973
2974	recddq = item->ri_buf[1].i_addr;
2975	if (recddq == NULL) {
2976		xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
2977		return XFS_ERROR(EIO);
2978	}
2979	if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
2980		xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
2981			item->ri_buf[1].i_len, __func__);
2982		return XFS_ERROR(EIO);
2983	}
2984
2985	/*
2986	 * This type of quotas was turned off, so ignore this record.
2987	 */
2988	type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2989	ASSERT(type);
2990	if (log->l_quotaoffs_flag & type)
2991		return (0);
2992
2993	/*
2994	 * At this point we know that quota was _not_ turned off.
2995	 * Since the mount flags are not indicating to us otherwise, this
2996	 * must mean that quota is on, and the dquot needs to be replayed.
2997	 * Remember that we may not have fully recovered the superblock yet,
2998	 * so we can't do the usual trick of looking at the SB quota bits.
2999	 *
3000	 * The other possibility, of course, is that the quota subsystem was
3001	 * removed since the last mount - ENOSYS.
3002	 */
3003	dq_f = item->ri_buf[0].i_addr;
3004	ASSERT(dq_f);
3005	error = xfs_dqcheck(mp, recddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
3006			   "xlog_recover_dquot_pass2 (log copy)");
3007	if (error)
3008		return XFS_ERROR(EIO);
3009	ASSERT(dq_f->qlf_len == 1);
3010
3011	error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dq_f->qlf_blkno,
3012				   XFS_FSB_TO_BB(mp, dq_f->qlf_len), 0, &bp,
3013				   NULL);
3014	if (error)
 
 
 
3015		return error;
3016
3017	ASSERT(bp);
3018	ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
3019
3020	/*
3021	 * At least the magic num portion should be on disk because this
3022	 * was among a chunk of dquots created earlier, and we did some
3023	 * minimal initialization then.
3024	 */
3025	error = xfs_dqcheck(mp, ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
3026			   "xlog_recover_dquot_pass2");
3027	if (error) {
3028		xfs_buf_relse(bp);
3029		return XFS_ERROR(EIO);
3030	}
3031
3032	/*
3033	 * If the dquot has an LSN in it, recover the dquot only if it's less
3034	 * than the lsn of the transaction we are replaying.
3035	 */
3036	if (xfs_sb_version_hascrc(&mp->m_sb)) {
3037		struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddq;
3038		xfs_lsn_t	lsn = be64_to_cpu(dqb->dd_lsn);
3039
3040		if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3041			goto out_release;
3042		}
3043	}
3044
3045	memcpy(ddq, recddq, item->ri_buf[1].i_len);
3046	if (xfs_sb_version_hascrc(&mp->m_sb)) {
3047		xfs_update_cksum((char *)ddq, sizeof(struct xfs_dqblk),
3048				 XFS_DQUOT_CRC_OFF);
3049	}
3050
3051	ASSERT(dq_f->qlf_size == 2);
3052	ASSERT(bp->b_target->bt_mount == mp);
3053	bp->b_iodone = xlog_recover_iodone;
3054	xfs_buf_delwri_queue(bp, buffer_list);
3055
3056out_release:
3057	xfs_buf_relse(bp);
3058	return 0;
3059}
3060
3061/*
3062 * This routine is called to create an in-core extent free intent
3063 * item from the efi format structure which was logged on disk.
3064 * It allocates an in-core efi, copies the extents from the format
3065 * structure into it, and adds the efi to the AIL with the given
3066 * LSN.
3067 */
3068STATIC int
3069xlog_recover_efi_pass2(
3070	struct xlog			*log,
3071	struct xlog_recover_item	*item,
3072	xfs_lsn_t			lsn)
3073{
3074	int			error;
3075	xfs_mount_t		*mp = log->l_mp;
3076	xfs_efi_log_item_t	*efip;
3077	xfs_efi_log_format_t	*efi_formatp;
3078
3079	efi_formatp = item->ri_buf[0].i_addr;
3080
3081	efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
3082	if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
3083					 &(efip->efi_format)))) {
3084		xfs_efi_item_free(efip);
3085		return error;
3086	}
3087	atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
3088
3089	spin_lock(&log->l_ailp->xa_lock);
3090	/*
3091	 * xfs_trans_ail_update() drops the AIL lock.
3092	 */
3093	xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
3094	return 0;
3095}
3096
3097
3098/*
3099 * This routine is called when an efd format structure is found in
3100 * a committed transaction in the log.  It's purpose is to cancel
3101 * the corresponding efi if it was still in the log.  To do this
3102 * it searches the AIL for the efi with an id equal to that in the
3103 * efd format structure.  If we find it, we remove the efi from the
3104 * AIL and free it.
3105 */
3106STATIC int
3107xlog_recover_efd_pass2(
3108	struct xlog			*log,
3109	struct xlog_recover_item	*item)
3110{
3111	xfs_efd_log_format_t	*efd_formatp;
3112	xfs_efi_log_item_t	*efip = NULL;
3113	xfs_log_item_t		*lip;
3114	__uint64_t		efi_id;
3115	struct xfs_ail_cursor	cur;
3116	struct xfs_ail		*ailp = log->l_ailp;
3117
3118	efd_formatp = item->ri_buf[0].i_addr;
3119	ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
3120		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
3121	       (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
3122		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
3123	efi_id = efd_formatp->efd_efi_id;
3124
3125	/*
3126	 * Search for the efi with the id in the efd format structure
3127	 * in the AIL.
3128	 */
3129	spin_lock(&ailp->xa_lock);
3130	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3131	while (lip != NULL) {
3132		if (lip->li_type == XFS_LI_EFI) {
3133			efip = (xfs_efi_log_item_t *)lip;
3134			if (efip->efi_format.efi_id == efi_id) {
3135				/*
3136				 * xfs_trans_ail_delete() drops the
3137				 * AIL lock.
3138				 */
3139				xfs_trans_ail_delete(ailp, lip,
3140						     SHUTDOWN_CORRUPT_INCORE);
3141				xfs_efi_item_free(efip);
3142				spin_lock(&ailp->xa_lock);
3143				break;
3144			}
3145		}
3146		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3147	}
3148	xfs_trans_ail_cursor_done(ailp, &cur);
3149	spin_unlock(&ailp->xa_lock);
3150
3151	return 0;
3152}
3153
3154/*
3155 * This routine is called when an inode create format structure is found in a
3156 * committed transaction in the log.  It's purpose is to initialise the inodes
3157 * being allocated on disk. This requires us to get inode cluster buffers that
3158 * match the range to be intialised, stamped with inode templates and written
3159 * by delayed write so that subsequent modifications will hit the cached buffer
3160 * and only need writing out at the end of recovery.
3161 */
3162STATIC int
3163xlog_recover_do_icreate_pass2(
3164	struct xlog		*log,
3165	struct list_head	*buffer_list,
3166	xlog_recover_item_t	*item)
3167{
3168	struct xfs_mount	*mp = log->l_mp;
3169	struct xfs_icreate_log	*icl;
3170	xfs_agnumber_t		agno;
3171	xfs_agblock_t		agbno;
3172	unsigned int		count;
3173	unsigned int		isize;
3174	xfs_agblock_t		length;
3175
3176	icl = (struct xfs_icreate_log *)item->ri_buf[0].i_addr;
3177	if (icl->icl_type != XFS_LI_ICREATE) {
3178		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad type");
3179		return EINVAL;
3180	}
3181
3182	if (icl->icl_size != 1) {
3183		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad icl size");
3184		return EINVAL;
3185	}
3186
3187	agno = be32_to_cpu(icl->icl_ag);
3188	if (agno >= mp->m_sb.sb_agcount) {
3189		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agno");
3190		return EINVAL;
3191	}
3192	agbno = be32_to_cpu(icl->icl_agbno);
3193	if (!agbno || agbno == NULLAGBLOCK || agbno >= mp->m_sb.sb_agblocks) {
3194		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agbno");
3195		return EINVAL;
3196	}
3197	isize = be32_to_cpu(icl->icl_isize);
3198	if (isize != mp->m_sb.sb_inodesize) {
3199		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad isize");
3200		return EINVAL;
3201	}
3202	count = be32_to_cpu(icl->icl_count);
3203	if (!count) {
3204		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count");
3205		return EINVAL;
3206	}
3207	length = be32_to_cpu(icl->icl_length);
3208	if (!length || length >= mp->m_sb.sb_agblocks) {
3209		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad length");
3210		return EINVAL;
3211	}
3212
3213	/* existing allocation is fixed value */
3214	ASSERT(count == mp->m_ialloc_inos);
3215	ASSERT(length == mp->m_ialloc_blks);
3216	if (count != mp->m_ialloc_inos ||
3217	     length != mp->m_ialloc_blks) {
3218		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count 2");
3219		return EINVAL;
3220	}
3221
3222	/*
3223	 * Inode buffers can be freed. Do not replay the inode initialisation as
3224	 * we could be overwriting something written after this inode buffer was
3225	 * cancelled.
3226	 *
3227	 * XXX: we need to iterate all buffers and only init those that are not
3228	 * cancelled. I think that a more fine grained factoring of
3229	 * xfs_ialloc_inode_init may be appropriate here to enable this to be
3230	 * done easily.
3231	 */
3232	if (xlog_check_buffer_cancelled(log,
3233			XFS_AGB_TO_DADDR(mp, agno, agbno), length, 0))
3234		return 0;
3235
3236	xfs_ialloc_inode_init(mp, NULL, buffer_list, agno, agbno, length,
3237					be32_to_cpu(icl->icl_gen));
3238	return 0;
3239}
3240
3241/*
3242 * Free up any resources allocated by the transaction
3243 *
3244 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
3245 */
3246STATIC void
3247xlog_recover_free_trans(
3248	struct xlog_recover	*trans)
3249{
3250	xlog_recover_item_t	*item, *n;
3251	int			i;
3252
3253	list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
3254		/* Free the regions in the item. */
3255		list_del(&item->ri_list);
3256		for (i = 0; i < item->ri_cnt; i++)
3257			kmem_free(item->ri_buf[i].i_addr);
3258		/* Free the item itself */
3259		kmem_free(item->ri_buf);
3260		kmem_free(item);
3261	}
3262	/* Free the transaction recover structure */
3263	kmem_free(trans);
3264}
3265
3266STATIC void
3267xlog_recover_buffer_ra_pass2(
3268	struct xlog                     *log,
3269	struct xlog_recover_item        *item)
3270{
3271	struct xfs_buf_log_format	*buf_f = item->ri_buf[0].i_addr;
3272	struct xfs_mount		*mp = log->l_mp;
3273
3274	if (xlog_peek_buffer_cancelled(log, buf_f->blf_blkno,
3275			buf_f->blf_len, buf_f->blf_flags)) {
3276		return;
3277	}
3278
3279	xfs_buf_readahead(mp->m_ddev_targp, buf_f->blf_blkno,
3280				buf_f->blf_len, NULL);
3281}
3282
3283STATIC void
3284xlog_recover_inode_ra_pass2(
3285	struct xlog                     *log,
3286	struct xlog_recover_item        *item)
3287{
3288	struct xfs_inode_log_format	ilf_buf;
3289	struct xfs_inode_log_format	*ilfp;
3290	struct xfs_mount		*mp = log->l_mp;
3291	int			error;
3292
3293	if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
3294		ilfp = item->ri_buf[0].i_addr;
3295	} else {
3296		ilfp = &ilf_buf;
3297		memset(ilfp, 0, sizeof(*ilfp));
3298		error = xfs_inode_item_format_convert(&item->ri_buf[0], ilfp);
3299		if (error)
3300			return;
3301	}
3302
3303	if (xlog_peek_buffer_cancelled(log, ilfp->ilf_blkno, ilfp->ilf_len, 0))
3304		return;
3305
3306	xfs_buf_readahead(mp->m_ddev_targp, ilfp->ilf_blkno,
3307				ilfp->ilf_len, &xfs_inode_buf_ra_ops);
3308}
3309
3310STATIC void
3311xlog_recover_dquot_ra_pass2(
3312	struct xlog			*log,
3313	struct xlog_recover_item	*item)
3314{
3315	struct xfs_mount	*mp = log->l_mp;
3316	struct xfs_disk_dquot	*recddq;
3317	struct xfs_dq_logformat	*dq_f;
3318	uint			type;
3319
3320
3321	if (mp->m_qflags == 0)
3322		return;
3323
3324	recddq = item->ri_buf[1].i_addr;
3325	if (recddq == NULL)
3326		return;
3327	if (item->ri_buf[1].i_len < sizeof(struct xfs_disk_dquot))
3328		return;
3329
3330	type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3331	ASSERT(type);
3332	if (log->l_quotaoffs_flag & type)
3333		return;
3334
3335	dq_f = item->ri_buf[0].i_addr;
3336	ASSERT(dq_f);
3337	ASSERT(dq_f->qlf_len == 1);
3338
3339	xfs_buf_readahead(mp->m_ddev_targp, dq_f->qlf_blkno,
3340			  XFS_FSB_TO_BB(mp, dq_f->qlf_len), NULL);
3341}
3342
3343STATIC void
3344xlog_recover_ra_pass2(
3345	struct xlog			*log,
3346	struct xlog_recover_item	*item)
3347{
3348	switch (ITEM_TYPE(item)) {
3349	case XFS_LI_BUF:
3350		xlog_recover_buffer_ra_pass2(log, item);
3351		break;
3352	case XFS_LI_INODE:
3353		xlog_recover_inode_ra_pass2(log, item);
3354		break;
3355	case XFS_LI_DQUOT:
3356		xlog_recover_dquot_ra_pass2(log, item);
3357		break;
3358	case XFS_LI_EFI:
3359	case XFS_LI_EFD:
3360	case XFS_LI_QUOTAOFF:
3361	default:
3362		break;
3363	}
3364}
3365
3366STATIC int
3367xlog_recover_commit_pass1(
3368	struct xlog			*log,
3369	struct xlog_recover		*trans,
3370	struct xlog_recover_item	*item)
3371{
3372	trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
3373
3374	switch (ITEM_TYPE(item)) {
3375	case XFS_LI_BUF:
3376		return xlog_recover_buffer_pass1(log, item);
3377	case XFS_LI_QUOTAOFF:
3378		return xlog_recover_quotaoff_pass1(log, item);
3379	case XFS_LI_INODE:
3380	case XFS_LI_EFI:
3381	case XFS_LI_EFD:
3382	case XFS_LI_DQUOT:
3383	case XFS_LI_ICREATE:
3384		/* nothing to do in pass 1 */
3385		return 0;
3386	default:
3387		xfs_warn(log->l_mp, "%s: invalid item type (%d)",
3388			__func__, ITEM_TYPE(item));
3389		ASSERT(0);
3390		return XFS_ERROR(EIO);
3391	}
3392}
3393
3394STATIC int
3395xlog_recover_commit_pass2(
3396	struct xlog			*log,
3397	struct xlog_recover		*trans,
3398	struct list_head		*buffer_list,
3399	struct xlog_recover_item	*item)
3400{
3401	trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
3402
3403	switch (ITEM_TYPE(item)) {
3404	case XFS_LI_BUF:
3405		return xlog_recover_buffer_pass2(log, buffer_list, item,
3406						 trans->r_lsn);
3407	case XFS_LI_INODE:
3408		return xlog_recover_inode_pass2(log, buffer_list, item,
3409						 trans->r_lsn);
3410	case XFS_LI_EFI:
3411		return xlog_recover_efi_pass2(log, item, trans->r_lsn);
3412	case XFS_LI_EFD:
3413		return xlog_recover_efd_pass2(log, item);
3414	case XFS_LI_DQUOT:
3415		return xlog_recover_dquot_pass2(log, buffer_list, item,
3416						trans->r_lsn);
3417	case XFS_LI_ICREATE:
3418		return xlog_recover_do_icreate_pass2(log, buffer_list, item);
3419	case XFS_LI_QUOTAOFF:
3420		/* nothing to do in pass2 */
3421		return 0;
3422	default:
3423		xfs_warn(log->l_mp, "%s: invalid item type (%d)",
3424			__func__, ITEM_TYPE(item));
3425		ASSERT(0);
3426		return XFS_ERROR(EIO);
3427	}
3428}
3429
3430STATIC int
3431xlog_recover_items_pass2(
3432	struct xlog                     *log,
3433	struct xlog_recover             *trans,
3434	struct list_head                *buffer_list,
3435	struct list_head                *item_list)
3436{
3437	struct xlog_recover_item	*item;
3438	int				error = 0;
3439
3440	list_for_each_entry(item, item_list, ri_list) {
3441		error = xlog_recover_commit_pass2(log, trans,
3442					  buffer_list, item);
3443		if (error)
3444			return error;
3445	}
3446
3447	return error;
3448}
3449
3450/*
3451 * Perform the transaction.
3452 *
3453 * If the transaction modifies a buffer or inode, do it now.  Otherwise,
3454 * EFIs and EFDs get queued up by adding entries into the AIL for them.
3455 */
3456STATIC int
3457xlog_recover_commit_trans(
3458	struct xlog		*log,
3459	struct xlog_recover	*trans,
3460	int			pass)
3461{
3462	int				error = 0;
3463	int				error2;
3464	int				items_queued = 0;
3465	struct xlog_recover_item	*item;
3466	struct xlog_recover_item	*next;
3467	LIST_HEAD			(buffer_list);
3468	LIST_HEAD			(ra_list);
3469	LIST_HEAD			(done_list);
3470
3471	#define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
3472
3473	hlist_del(&trans->r_list);
3474
3475	error = xlog_recover_reorder_trans(log, trans, pass);
3476	if (error)
3477		return error;
3478
3479	list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
3480		switch (pass) {
3481		case XLOG_RECOVER_PASS1:
3482			error = xlog_recover_commit_pass1(log, trans, item);
3483			break;
3484		case XLOG_RECOVER_PASS2:
3485			xlog_recover_ra_pass2(log, item);
3486			list_move_tail(&item->ri_list, &ra_list);
3487			items_queued++;
3488			if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
3489				error = xlog_recover_items_pass2(log, trans,
3490						&buffer_list, &ra_list);
3491				list_splice_tail_init(&ra_list, &done_list);
3492				items_queued = 0;
3493			}
3494
3495			break;
3496		default:
3497			ASSERT(0);
3498		}
3499
3500		if (error)
3501			goto out;
3502	}
3503
3504out:
3505	if (!list_empty(&ra_list)) {
3506		if (!error)
3507			error = xlog_recover_items_pass2(log, trans,
3508					&buffer_list, &ra_list);
3509		list_splice_tail_init(&ra_list, &done_list);
3510	}
3511
3512	if (!list_empty(&done_list))
3513		list_splice_init(&done_list, &trans->r_itemq);
3514
3515	xlog_recover_free_trans(trans);
3516
3517	error2 = xfs_buf_delwri_submit(&buffer_list);
3518	return error ? error : error2;
3519}
3520
3521STATIC int
3522xlog_recover_unmount_trans(
3523	struct xlog		*log,
3524	struct xlog_recover	*trans)
3525{
3526	/* Do nothing now */
3527	xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
3528	return 0;
3529}
3530
3531/*
3532 * There are two valid states of the r_state field.  0 indicates that the
3533 * transaction structure is in a normal state.  We have either seen the
3534 * start of the transaction or the last operation we added was not a partial
3535 * operation.  If the last operation we added to the transaction was a
3536 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
3537 *
3538 * NOTE: skip LRs with 0 data length.
3539 */
3540STATIC int
3541xlog_recover_process_data(
3542	struct xlog		*log,
3543	struct hlist_head	rhash[],
3544	struct xlog_rec_header	*rhead,
3545	xfs_caddr_t		dp,
3546	int			pass)
3547{
3548	xfs_caddr_t		lp;
3549	int			num_logops;
3550	xlog_op_header_t	*ohead;
3551	xlog_recover_t		*trans;
3552	xlog_tid_t		tid;
3553	int			error;
3554	unsigned long		hash;
3555	uint			flags;
3556
3557	lp = dp + be32_to_cpu(rhead->h_len);
3558	num_logops = be32_to_cpu(rhead->h_num_logops);
3559
3560	/* check the log format matches our own - else we can't recover */
3561	if (xlog_header_check_recover(log->l_mp, rhead))
3562		return (XFS_ERROR(EIO));
3563
3564	while ((dp < lp) && num_logops) {
3565		ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
3566		ohead = (xlog_op_header_t *)dp;
3567		dp += sizeof(xlog_op_header_t);
3568		if (ohead->oh_clientid != XFS_TRANSACTION &&
3569		    ohead->oh_clientid != XFS_LOG) {
3570			xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
3571					__func__, ohead->oh_clientid);
3572			ASSERT(0);
3573			return (XFS_ERROR(EIO));
3574		}
3575		tid = be32_to_cpu(ohead->oh_tid);
3576		hash = XLOG_RHASH(tid);
3577		trans = xlog_recover_find_tid(&rhash[hash], tid);
3578		if (trans == NULL) {		   /* not found; add new tid */
3579			if (ohead->oh_flags & XLOG_START_TRANS)
3580				xlog_recover_new_tid(&rhash[hash], tid,
3581					be64_to_cpu(rhead->h_lsn));
3582		} else {
3583			if (dp + be32_to_cpu(ohead->oh_len) > lp) {
3584				xfs_warn(log->l_mp, "%s: bad length 0x%x",
3585					__func__, be32_to_cpu(ohead->oh_len));
3586				WARN_ON(1);
3587				return (XFS_ERROR(EIO));
3588			}
3589			flags = ohead->oh_flags & ~XLOG_END_TRANS;
3590			if (flags & XLOG_WAS_CONT_TRANS)
3591				flags &= ~XLOG_CONTINUE_TRANS;
3592			switch (flags) {
3593			case XLOG_COMMIT_TRANS:
3594				error = xlog_recover_commit_trans(log,
3595								trans, pass);
3596				break;
3597			case XLOG_UNMOUNT_TRANS:
3598				error = xlog_recover_unmount_trans(log, trans);
3599				break;
3600			case XLOG_WAS_CONT_TRANS:
3601				error = xlog_recover_add_to_cont_trans(log,
3602						trans, dp,
3603						be32_to_cpu(ohead->oh_len));
3604				break;
3605			case XLOG_START_TRANS:
3606				xfs_warn(log->l_mp, "%s: bad transaction",
3607					__func__);
3608				ASSERT(0);
3609				error = XFS_ERROR(EIO);
3610				break;
3611			case 0:
3612			case XLOG_CONTINUE_TRANS:
3613				error = xlog_recover_add_to_trans(log, trans,
3614						dp, be32_to_cpu(ohead->oh_len));
3615				break;
3616			default:
3617				xfs_warn(log->l_mp, "%s: bad flag 0x%x",
3618					__func__, flags);
3619				ASSERT(0);
3620				error = XFS_ERROR(EIO);
3621				break;
3622			}
3623			if (error) {
3624				xlog_recover_free_trans(trans);
3625				return error;
3626			}
3627		}
3628		dp += be32_to_cpu(ohead->oh_len);
3629		num_logops--;
3630	}
3631	return 0;
3632}
3633
3634/*
3635 * Process an extent free intent item that was recovered from
3636 * the log.  We need to free the extents that it describes.
3637 */
3638STATIC int
3639xlog_recover_process_efi(
3640	xfs_mount_t		*mp,
3641	xfs_efi_log_item_t	*efip)
3642{
3643	xfs_efd_log_item_t	*efdp;
3644	xfs_trans_t		*tp;
3645	int			i;
3646	int			error = 0;
3647	xfs_extent_t		*extp;
3648	xfs_fsblock_t		startblock_fsb;
3649
3650	ASSERT(!test_bit(XFS_EFI_RECOVERED, &efip->efi_flags));
3651
3652	/*
3653	 * First check the validity of the extents described by the
3654	 * EFI.  If any are bad, then assume that all are bad and
3655	 * just toss the EFI.
3656	 */
3657	for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3658		extp = &(efip->efi_format.efi_extents[i]);
3659		startblock_fsb = XFS_BB_TO_FSB(mp,
3660				   XFS_FSB_TO_DADDR(mp, extp->ext_start));
3661		if ((startblock_fsb == 0) ||
3662		    (extp->ext_len == 0) ||
3663		    (startblock_fsb >= mp->m_sb.sb_dblocks) ||
3664		    (extp->ext_len >= mp->m_sb.sb_agblocks)) {
3665			/*
3666			 * This will pull the EFI from the AIL and
3667			 * free the memory associated with it.
3668			 */
3669			set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
3670			xfs_efi_release(efip, efip->efi_format.efi_nextents);
3671			return XFS_ERROR(EIO);
3672		}
3673	}
3674
3675	tp = xfs_trans_alloc(mp, 0);
3676	error = xfs_trans_reserve(tp, &M_RES(mp)->tr_itruncate, 0, 0);
3677	if (error)
3678		goto abort_error;
3679	efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3680
3681	for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3682		extp = &(efip->efi_format.efi_extents[i]);
3683		error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3684		if (error)
3685			goto abort_error;
3686		xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3687					 extp->ext_len);
3688	}
3689
3690	set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
3691	error = xfs_trans_commit(tp, 0);
3692	return error;
3693
3694abort_error:
3695	xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3696	return error;
3697}
3698
3699/*
3700 * When this is called, all of the EFIs which did not have
3701 * corresponding EFDs should be in the AIL.  What we do now
3702 * is free the extents associated with each one.
3703 *
3704 * Since we process the EFIs in normal transactions, they
3705 * will be removed at some point after the commit.  This prevents
3706 * us from just walking down the list processing each one.
3707 * We'll use a flag in the EFI to skip those that we've already
3708 * processed and use the AIL iteration mechanism's generation
3709 * count to try to speed this up at least a bit.
3710 *
3711 * When we start, we know that the EFIs are the only things in
3712 * the AIL.  As we process them, however, other items are added
3713 * to the AIL.  Since everything added to the AIL must come after
3714 * everything already in the AIL, we stop processing as soon as
3715 * we see something other than an EFI in the AIL.
3716 */
3717STATIC int
3718xlog_recover_process_efis(
3719	struct xlog	*log)
3720{
3721	xfs_log_item_t		*lip;
3722	xfs_efi_log_item_t	*efip;
3723	int			error = 0;
3724	struct xfs_ail_cursor	cur;
3725	struct xfs_ail		*ailp;
3726
3727	ailp = log->l_ailp;
3728	spin_lock(&ailp->xa_lock);
3729	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3730	while (lip != NULL) {
3731		/*
3732		 * We're done when we see something other than an EFI.
3733		 * There should be no EFIs left in the AIL now.
3734		 */
3735		if (lip->li_type != XFS_LI_EFI) {
3736#ifdef DEBUG
3737			for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
3738				ASSERT(lip->li_type != XFS_LI_EFI);
3739#endif
3740			break;
3741		}
3742
3743		/*
3744		 * Skip EFIs that we've already processed.
3745		 */
3746		efip = (xfs_efi_log_item_t *)lip;
3747		if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags)) {
3748			lip = xfs_trans_ail_cursor_next(ailp, &cur);
3749			continue;
3750		}
3751
3752		spin_unlock(&ailp->xa_lock);
3753		error = xlog_recover_process_efi(log->l_mp, efip);
3754		spin_lock(&ailp->xa_lock);
3755		if (error)
3756			goto out;
3757		lip = xfs_trans_ail_cursor_next(ailp, &cur);
3758	}
3759out:
3760	xfs_trans_ail_cursor_done(ailp, &cur);
3761	spin_unlock(&ailp->xa_lock);
3762	return error;
3763}
3764
3765/*
3766 * This routine performs a transaction to null out a bad inode pointer
3767 * in an agi unlinked inode hash bucket.
3768 */
3769STATIC void
3770xlog_recover_clear_agi_bucket(
3771	xfs_mount_t	*mp,
3772	xfs_agnumber_t	agno,
3773	int		bucket)
3774{
3775	xfs_trans_t	*tp;
3776	xfs_agi_t	*agi;
3777	xfs_buf_t	*agibp;
3778	int		offset;
3779	int		error;
3780
3781	tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3782	error = xfs_trans_reserve(tp, &M_RES(mp)->tr_clearagi, 0, 0);
 
3783	if (error)
3784		goto out_abort;
3785
3786	error = xfs_read_agi(mp, tp, agno, &agibp);
3787	if (error)
3788		goto out_abort;
3789
3790	agi = XFS_BUF_TO_AGI(agibp);
3791	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3792	offset = offsetof(xfs_agi_t, agi_unlinked) +
3793		 (sizeof(xfs_agino_t) * bucket);
3794	xfs_trans_log_buf(tp, agibp, offset,
3795			  (offset + sizeof(xfs_agino_t) - 1));
3796
3797	error = xfs_trans_commit(tp, 0);
3798	if (error)
3799		goto out_error;
3800	return;
3801
3802out_abort:
3803	xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3804out_error:
3805	xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
3806	return;
3807}
3808
3809STATIC xfs_agino_t
3810xlog_recover_process_one_iunlink(
3811	struct xfs_mount		*mp,
3812	xfs_agnumber_t			agno,
3813	xfs_agino_t			agino,
3814	int				bucket)
3815{
3816	struct xfs_buf			*ibp;
3817	struct xfs_dinode		*dip;
3818	struct xfs_inode		*ip;
3819	xfs_ino_t			ino;
3820	int				error;
3821
3822	ino = XFS_AGINO_TO_INO(mp, agno, agino);
3823	error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
3824	if (error)
3825		goto fail;
3826
3827	/*
3828	 * Get the on disk inode to find the next inode in the bucket.
3829	 */
3830	error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0, 0);
3831	if (error)
3832		goto fail_iput;
3833
3834	ASSERT(ip->i_d.di_nlink == 0);
3835	ASSERT(ip->i_d.di_mode != 0);
3836
3837	/* setup for the next pass */
3838	agino = be32_to_cpu(dip->di_next_unlinked);
3839	xfs_buf_relse(ibp);
3840
3841	/*
3842	 * Prevent any DMAPI event from being sent when the reference on
3843	 * the inode is dropped.
3844	 */
3845	ip->i_d.di_dmevmask = 0;
3846
3847	IRELE(ip);
3848	return agino;
3849
3850 fail_iput:
3851	IRELE(ip);
3852 fail:
3853	/*
3854	 * We can't read in the inode this bucket points to, or this inode
3855	 * is messed up.  Just ditch this bucket of inodes.  We will lose
3856	 * some inodes and space, but at least we won't hang.
3857	 *
3858	 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3859	 * clear the inode pointer in the bucket.
3860	 */
3861	xlog_recover_clear_agi_bucket(mp, agno, bucket);
3862	return NULLAGINO;
3863}
3864
3865/*
3866 * xlog_iunlink_recover
3867 *
3868 * This is called during recovery to process any inodes which
3869 * we unlinked but not freed when the system crashed.  These
3870 * inodes will be on the lists in the AGI blocks.  What we do
3871 * here is scan all the AGIs and fully truncate and free any
3872 * inodes found on the lists.  Each inode is removed from the
3873 * lists when it has been fully truncated and is freed.  The
3874 * freeing of the inode and its removal from the list must be
3875 * atomic.
3876 */
3877STATIC void
3878xlog_recover_process_iunlinks(
3879	struct xlog	*log)
3880{
3881	xfs_mount_t	*mp;
3882	xfs_agnumber_t	agno;
3883	xfs_agi_t	*agi;
3884	xfs_buf_t	*agibp;
3885	xfs_agino_t	agino;
3886	int		bucket;
3887	int		error;
3888	uint		mp_dmevmask;
3889
3890	mp = log->l_mp;
3891
3892	/*
3893	 * Prevent any DMAPI event from being sent while in this function.
3894	 */
3895	mp_dmevmask = mp->m_dmevmask;
3896	mp->m_dmevmask = 0;
3897
3898	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3899		/*
3900		 * Find the agi for this ag.
3901		 */
3902		error = xfs_read_agi(mp, NULL, agno, &agibp);
3903		if (error) {
3904			/*
3905			 * AGI is b0rked. Don't process it.
3906			 *
3907			 * We should probably mark the filesystem as corrupt
3908			 * after we've recovered all the ag's we can....
3909			 */
3910			continue;
3911		}
3912		/*
3913		 * Unlock the buffer so that it can be acquired in the normal
3914		 * course of the transaction to truncate and free each inode.
3915		 * Because we are not racing with anyone else here for the AGI
3916		 * buffer, we don't even need to hold it locked to read the
3917		 * initial unlinked bucket entries out of the buffer. We keep
3918		 * buffer reference though, so that it stays pinned in memory
3919		 * while we need the buffer.
3920		 */
3921		agi = XFS_BUF_TO_AGI(agibp);
3922		xfs_buf_unlock(agibp);
3923
3924		for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3925			agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3926			while (agino != NULLAGINO) {
 
 
 
 
 
 
 
3927				agino = xlog_recover_process_one_iunlink(mp,
3928							agno, agino, bucket);
 
 
 
 
 
 
 
 
 
3929			}
3930		}
3931		xfs_buf_rele(agibp);
 
 
 
 
 
3932	}
3933
3934	mp->m_dmevmask = mp_dmevmask;
3935}
3936
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3937/*
3938 * Upack the log buffer data and crc check it. If the check fails, issue a
3939 * warning if and only if the CRC in the header is non-zero. This makes the
3940 * check an advisory warning, and the zero CRC check will prevent failure
3941 * warnings from being emitted when upgrading the kernel from one that does not
3942 * add CRCs by default.
3943 *
3944 * When filesystems are CRC enabled, this CRC mismatch becomes a fatal log
3945 * corruption failure
3946 */
3947STATIC int
3948xlog_unpack_data_crc(
3949	struct xlog_rec_header	*rhead,
3950	xfs_caddr_t		dp,
3951	struct xlog		*log)
3952{
3953	__le32			crc;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3954
3955	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
3956	if (crc != rhead->h_crc) {
3957		if (rhead->h_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
3958			xfs_alert(log->l_mp,
3959		"log record CRC mismatch: found 0x%x, expected 0x%x.",
3960					le32_to_cpu(rhead->h_crc),
3961					le32_to_cpu(crc));
3962			xfs_hex_dump(dp, 32);
3963		}
3964
3965		/*
3966		 * If we've detected a log record corruption, then we can't
3967		 * recover past this point. Abort recovery if we are enforcing
3968		 * CRC protection by punting an error back up the stack.
3969		 */
3970		if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
3971			return EFSCORRUPTED;
3972	}
3973
3974	return 0;
3975}
3976
3977STATIC int
3978xlog_unpack_data(
3979	struct xlog_rec_header	*rhead,
3980	xfs_caddr_t		dp,
3981	struct xlog		*log)
3982{
3983	int			i, j, k;
3984	int			error;
3985
3986	error = xlog_unpack_data_crc(rhead, dp, log);
3987	if (error)
3988		return error;
3989
3990	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3991		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3992		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3993		dp += BBSIZE;
3994	}
3995
3996	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3997		xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
3998		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3999			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
4000			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
4001			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
4002			dp += BBSIZE;
4003		}
4004	}
4005
4006	return 0;
4007}
4008
4009STATIC int
4010xlog_valid_rec_header(
4011	struct xlog		*log,
4012	struct xlog_rec_header	*rhead,
4013	xfs_daddr_t		blkno)
4014{
4015	int			hlen;
4016
4017	if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
4018		XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
4019				XFS_ERRLEVEL_LOW, log->l_mp);
4020		return XFS_ERROR(EFSCORRUPTED);
4021	}
4022	if (unlikely(
4023	    (!rhead->h_version ||
4024	    (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
4025		xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
4026			__func__, be32_to_cpu(rhead->h_version));
4027		return XFS_ERROR(EIO);
4028	}
4029
4030	/* LR body must have data or it wouldn't have been written */
4031	hlen = be32_to_cpu(rhead->h_len);
4032	if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
4033		XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
4034				XFS_ERRLEVEL_LOW, log->l_mp);
4035		return XFS_ERROR(EFSCORRUPTED);
4036	}
4037	if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
4038		XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
4039				XFS_ERRLEVEL_LOW, log->l_mp);
4040		return XFS_ERROR(EFSCORRUPTED);
4041	}
4042	return 0;
4043}
4044
4045/*
4046 * Read the log from tail to head and process the log records found.
4047 * Handle the two cases where the tail and head are in the same cycle
4048 * and where the active portion of the log wraps around the end of
4049 * the physical log separately.  The pass parameter is passed through
4050 * to the routines called to process the data and is not looked at
4051 * here.
4052 */
4053STATIC int
4054xlog_do_recovery_pass(
4055	struct xlog		*log,
4056	xfs_daddr_t		head_blk,
4057	xfs_daddr_t		tail_blk,
4058	int			pass)
4059{
4060	xlog_rec_header_t	*rhead;
4061	xfs_daddr_t		blk_no;
4062	xfs_caddr_t		offset;
4063	xfs_buf_t		*hbp, *dbp;
4064	int			error = 0, h_size;
4065	int			bblks, split_bblks;
4066	int			hblks, split_hblks, wrapped_hblks;
4067	struct hlist_head	rhash[XLOG_RHASH_SIZE];
4068
4069	ASSERT(head_blk != tail_blk);
4070
4071	/*
4072	 * Read the header of the tail block and get the iclog buffer size from
4073	 * h_size.  Use this to tell how many sectors make up the log header.
4074	 */
4075	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
4076		/*
4077		 * When using variable length iclogs, read first sector of
4078		 * iclog header and extract the header size from it.  Get a
4079		 * new hbp that is the correct size.
4080		 */
4081		hbp = xlog_get_bp(log, 1);
4082		if (!hbp)
4083			return ENOMEM;
4084
4085		error = xlog_bread(log, tail_blk, 1, hbp, &offset);
4086		if (error)
4087			goto bread_err1;
4088
4089		rhead = (xlog_rec_header_t *)offset;
4090		error = xlog_valid_rec_header(log, rhead, tail_blk);
4091		if (error)
4092			goto bread_err1;
4093		h_size = be32_to_cpu(rhead->h_size);
4094		if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
4095		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
4096			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
4097			if (h_size % XLOG_HEADER_CYCLE_SIZE)
4098				hblks++;
4099			xlog_put_bp(hbp);
4100			hbp = xlog_get_bp(log, hblks);
4101		} else {
4102			hblks = 1;
4103		}
4104	} else {
4105		ASSERT(log->l_sectBBsize == 1);
4106		hblks = 1;
4107		hbp = xlog_get_bp(log, 1);
4108		h_size = XLOG_BIG_RECORD_BSIZE;
4109	}
4110
4111	if (!hbp)
4112		return ENOMEM;
4113	dbp = xlog_get_bp(log, BTOBB(h_size));
4114	if (!dbp) {
4115		xlog_put_bp(hbp);
4116		return ENOMEM;
4117	}
4118
4119	memset(rhash, 0, sizeof(rhash));
4120	if (tail_blk <= head_blk) {
4121		for (blk_no = tail_blk; blk_no < head_blk; ) {
4122			error = xlog_bread(log, blk_no, hblks, hbp, &offset);
4123			if (error)
4124				goto bread_err2;
4125
4126			rhead = (xlog_rec_header_t *)offset;
4127			error = xlog_valid_rec_header(log, rhead, blk_no);
4128			if (error)
4129				goto bread_err2;
4130
4131			/* blocks in data section */
4132			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
4133			error = xlog_bread(log, blk_no + hblks, bblks, dbp,
4134					   &offset);
4135			if (error)
4136				goto bread_err2;
4137
4138			error = xlog_unpack_data(rhead, offset, log);
4139			if (error)
4140				goto bread_err2;
4141
4142			error = xlog_recover_process_data(log,
4143						rhash, rhead, offset, pass);
4144			if (error)
4145				goto bread_err2;
4146			blk_no += bblks + hblks;
4147		}
4148	} else {
4149		/*
4150		 * Perform recovery around the end of the physical log.
4151		 * When the head is not on the same cycle number as the tail,
4152		 * we can't do a sequential recovery as above.
4153		 */
4154		blk_no = tail_blk;
4155		while (blk_no < log->l_logBBsize) {
4156			/*
4157			 * Check for header wrapping around physical end-of-log
4158			 */
4159			offset = hbp->b_addr;
4160			split_hblks = 0;
4161			wrapped_hblks = 0;
4162			if (blk_no + hblks <= log->l_logBBsize) {
4163				/* Read header in one read */
4164				error = xlog_bread(log, blk_no, hblks, hbp,
4165						   &offset);
4166				if (error)
4167					goto bread_err2;
4168			} else {
4169				/* This LR is split across physical log end */
4170				if (blk_no != log->l_logBBsize) {
4171					/* some data before physical log end */
4172					ASSERT(blk_no <= INT_MAX);
4173					split_hblks = log->l_logBBsize - (int)blk_no;
4174					ASSERT(split_hblks > 0);
4175					error = xlog_bread(log, blk_no,
4176							   split_hblks, hbp,
4177							   &offset);
4178					if (error)
4179						goto bread_err2;
4180				}
4181
4182				/*
4183				 * Note: this black magic still works with
4184				 * large sector sizes (non-512) only because:
4185				 * - we increased the buffer size originally
4186				 *   by 1 sector giving us enough extra space
4187				 *   for the second read;
4188				 * - the log start is guaranteed to be sector
4189				 *   aligned;
4190				 * - we read the log end (LR header start)
4191				 *   _first_, then the log start (LR header end)
4192				 *   - order is important.
4193				 */
4194				wrapped_hblks = hblks - split_hblks;
4195				error = xlog_bread_offset(log, 0,
4196						wrapped_hblks, hbp,
4197						offset + BBTOB(split_hblks));
4198				if (error)
4199					goto bread_err2;
4200			}
4201			rhead = (xlog_rec_header_t *)offset;
4202			error = xlog_valid_rec_header(log, rhead,
4203						split_hblks ? blk_no : 0);
4204			if (error)
4205				goto bread_err2;
4206
4207			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
4208			blk_no += hblks;
4209
4210			/* Read in data for log record */
4211			if (blk_no + bblks <= log->l_logBBsize) {
4212				error = xlog_bread(log, blk_no, bblks, dbp,
4213						   &offset);
4214				if (error)
4215					goto bread_err2;
4216			} else {
4217				/* This log record is split across the
4218				 * physical end of log */
4219				offset = dbp->b_addr;
4220				split_bblks = 0;
4221				if (blk_no != log->l_logBBsize) {
4222					/* some data is before the physical
4223					 * end of log */
4224					ASSERT(!wrapped_hblks);
4225					ASSERT(blk_no <= INT_MAX);
4226					split_bblks =
4227						log->l_logBBsize - (int)blk_no;
4228					ASSERT(split_bblks > 0);
4229					error = xlog_bread(log, blk_no,
4230							split_bblks, dbp,
4231							&offset);
4232					if (error)
4233						goto bread_err2;
4234				}
4235
4236				/*
4237				 * Note: this black magic still works with
4238				 * large sector sizes (non-512) only because:
4239				 * - we increased the buffer size originally
4240				 *   by 1 sector giving us enough extra space
4241				 *   for the second read;
4242				 * - the log start is guaranteed to be sector
4243				 *   aligned;
4244				 * - we read the log end (LR header start)
4245				 *   _first_, then the log start (LR header end)
4246				 *   - order is important.
4247				 */
4248				error = xlog_bread_offset(log, 0,
4249						bblks - split_bblks, dbp,
4250						offset + BBTOB(split_bblks));
4251				if (error)
4252					goto bread_err2;
4253			}
4254
4255			error = xlog_unpack_data(rhead, offset, log);
4256			if (error)
4257				goto bread_err2;
4258
4259			error = xlog_recover_process_data(log, rhash,
4260							rhead, offset, pass);
4261			if (error)
4262				goto bread_err2;
4263			blk_no += bblks;
4264		}
4265
4266		ASSERT(blk_no >= log->l_logBBsize);
4267		blk_no -= log->l_logBBsize;
4268
4269		/* read first part of physical log */
4270		while (blk_no < head_blk) {
4271			error = xlog_bread(log, blk_no, hblks, hbp, &offset);
4272			if (error)
4273				goto bread_err2;
4274
4275			rhead = (xlog_rec_header_t *)offset;
4276			error = xlog_valid_rec_header(log, rhead, blk_no);
4277			if (error)
4278				goto bread_err2;
4279
4280			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
4281			error = xlog_bread(log, blk_no+hblks, bblks, dbp,
4282					   &offset);
4283			if (error)
4284				goto bread_err2;
4285
4286			error = xlog_unpack_data(rhead, offset, log);
4287			if (error)
4288				goto bread_err2;
4289
4290			error = xlog_recover_process_data(log, rhash,
4291							rhead, offset, pass);
4292			if (error)
4293				goto bread_err2;
4294			blk_no += bblks + hblks;
4295		}
4296	}
4297
4298 bread_err2:
4299	xlog_put_bp(dbp);
4300 bread_err1:
4301	xlog_put_bp(hbp);
4302	return error;
4303}
4304
4305/*
4306 * Do the recovery of the log.  We actually do this in two phases.
4307 * The two passes are necessary in order to implement the function
4308 * of cancelling a record written into the log.  The first pass
4309 * determines those things which have been cancelled, and the
4310 * second pass replays log items normally except for those which
4311 * have been cancelled.  The handling of the replay and cancellations
4312 * takes place in the log item type specific routines.
4313 *
4314 * The table of items which have cancel records in the log is allocated
4315 * and freed at this level, since only here do we know when all of
4316 * the log recovery has been completed.
4317 */
4318STATIC int
4319xlog_do_log_recovery(
4320	struct xlog	*log,
4321	xfs_daddr_t	head_blk,
4322	xfs_daddr_t	tail_blk)
4323{
4324	int		error, i;
4325
4326	ASSERT(head_blk != tail_blk);
4327
4328	/*
4329	 * First do a pass to find all of the cancelled buf log items.
4330	 * Store them in the buf_cancel_table for use in the second pass.
4331	 */
4332	log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
4333						 sizeof(struct list_head),
4334						 KM_SLEEP);
4335	for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
4336		INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
4337
4338	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
4339				      XLOG_RECOVER_PASS1);
4340	if (error != 0) {
4341		kmem_free(log->l_buf_cancel_table);
4342		log->l_buf_cancel_table = NULL;
4343		return error;
4344	}
4345	/*
4346	 * Then do a second pass to actually recover the items in the log.
4347	 * When it is complete free the table of buf cancel items.
4348	 */
4349	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
4350				      XLOG_RECOVER_PASS2);
4351#ifdef DEBUG
4352	if (!error) {
4353		int	i;
4354
4355		for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
4356			ASSERT(list_empty(&log->l_buf_cancel_table[i]));
4357	}
4358#endif	/* DEBUG */
4359
4360	kmem_free(log->l_buf_cancel_table);
4361	log->l_buf_cancel_table = NULL;
4362
4363	return error;
4364}
4365
4366/*
4367 * Do the actual recovery
4368 */
4369STATIC int
4370xlog_do_recover(
4371	struct xlog	*log,
4372	xfs_daddr_t	head_blk,
4373	xfs_daddr_t	tail_blk)
4374{
4375	int		error;
4376	xfs_buf_t	*bp;
4377	xfs_sb_t	*sbp;
4378
4379	/*
4380	 * First replay the images in the log.
4381	 */
4382	error = xlog_do_log_recovery(log, head_blk, tail_blk);
4383	if (error)
4384		return error;
 
 
 
4385
4386	/*
4387	 * If IO errors happened during recovery, bail out.
4388	 */
4389	if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
4390		return (EIO);
4391	}
4392
4393	/*
4394	 * We now update the tail_lsn since much of the recovery has completed
4395	 * and there may be space available to use.  If there were no extent
4396	 * or iunlinks, we can free up the entire log and set the tail_lsn to
4397	 * be the last_sync_lsn.  This was set in xlog_find_tail to be the
4398	 * lsn of the last known good LR on disk.  If there are extent frees
4399	 * or iunlinks they will have some entries in the AIL; so we look at
4400	 * the AIL to determine how to set the tail_lsn.
4401	 */
4402	xlog_assign_tail_lsn(log->l_mp);
4403
4404	/*
4405	 * Now that we've finished replaying all buffer and inode
4406	 * updates, re-read in the superblock and reverify it.
4407	 */
4408	bp = xfs_getsb(log->l_mp, 0);
4409	XFS_BUF_UNDONE(bp);
4410	ASSERT(!(XFS_BUF_ISWRITE(bp)));
 
4411	XFS_BUF_READ(bp);
4412	XFS_BUF_UNASYNC(bp);
4413	bp->b_ops = &xfs_sb_buf_ops;
4414
4415	if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
4416		xfs_buf_relse(bp);
4417		return XFS_ERROR(EIO);
4418	}
4419
4420	xfs_buf_iorequest(bp);
4421	error = xfs_buf_iowait(bp);
4422	if (error) {
4423		xfs_buf_ioerror_alert(bp, __func__);
 
4424		ASSERT(0);
4425		xfs_buf_relse(bp);
4426		return error;
4427	}
4428
4429	/* Convert superblock from on-disk format */
4430	sbp = &log->l_mp->m_sb;
4431	xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
4432	ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
4433	ASSERT(xfs_sb_good_version(sbp));
4434	xfs_buf_relse(bp);
4435
4436	/* We've re-read the superblock so re-initialize per-cpu counters */
4437	xfs_icsb_reinit_counters(log->l_mp);
4438
4439	xlog_recover_check_summary(log);
4440
4441	/* Normal transactions can now occur */
4442	log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
4443	return 0;
4444}
4445
4446/*
4447 * Perform recovery and re-initialize some log variables in xlog_find_tail.
4448 *
4449 * Return error or zero.
4450 */
4451int
4452xlog_recover(
4453	struct xlog	*log)
4454{
4455	xfs_daddr_t	head_blk, tail_blk;
4456	int		error;
4457
4458	/* find the tail of the log */
4459	if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
4460		return error;
4461
4462	if (tail_blk != head_blk) {
4463		/* There used to be a comment here:
4464		 *
4465		 * disallow recovery on read-only mounts.  note -- mount
4466		 * checks for ENOSPC and turns it into an intelligent
4467		 * error message.
4468		 * ...but this is no longer true.  Now, unless you specify
4469		 * NORECOVERY (in which case this function would never be
4470		 * called), we just go ahead and recover.  We do this all
4471		 * under the vfs layer, so we can get away with it unless
4472		 * the device itself is read-only, in which case we fail.
4473		 */
4474		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
4475			return error;
4476		}
4477
4478		/*
4479		 * Version 5 superblock log feature mask validation. We know the
4480		 * log is dirty so check if there are any unknown log features
4481		 * in what we need to recover. If there are unknown features
4482		 * (e.g. unsupported transactions, then simply reject the
4483		 * attempt at recovery before touching anything.
4484		 */
4485		if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
4486		    xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
4487					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
4488			xfs_warn(log->l_mp,
4489"Superblock has unknown incompatible log features (0x%x) enabled.\n"
4490"The log can not be fully and/or safely recovered by this kernel.\n"
4491"Please recover the log on a kernel that supports the unknown features.",
4492				(log->l_mp->m_sb.sb_features_log_incompat &
4493					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
4494			return EINVAL;
4495		}
4496
4497		xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
4498				log->l_mp->m_logname ? log->l_mp->m_logname
4499						     : "internal");
4500
4501		error = xlog_do_recover(log, head_blk, tail_blk);
4502		log->l_flags |= XLOG_RECOVERY_NEEDED;
4503	}
4504	return error;
4505}
4506
4507/*
4508 * In the first part of recovery we replay inodes and buffers and build
4509 * up the list of extent free items which need to be processed.  Here
4510 * we process the extent free items and clean up the on disk unlinked
4511 * inode lists.  This is separated from the first part of recovery so
4512 * that the root and real-time bitmap inodes can be read in from disk in
4513 * between the two stages.  This is necessary so that we can free space
4514 * in the real-time portion of the file system.
4515 */
4516int
4517xlog_recover_finish(
4518	struct xlog	*log)
4519{
4520	/*
4521	 * Now we're ready to do the transactions needed for the
4522	 * rest of recovery.  Start with completing all the extent
4523	 * free intent records and then process the unlinked inode
4524	 * lists.  At this point, we essentially run in normal mode
4525	 * except that we're still performing recovery actions
4526	 * rather than accepting new requests.
4527	 */
4528	if (log->l_flags & XLOG_RECOVERY_NEEDED) {
4529		int	error;
4530		error = xlog_recover_process_efis(log);
4531		if (error) {
4532			xfs_alert(log->l_mp, "Failed to recover EFIs");
4533			return error;
4534		}
4535		/*
4536		 * Sync the log to get all the EFIs out of the AIL.
4537		 * This isn't absolutely necessary, but it helps in
4538		 * case the unlink transactions would have problems
4539		 * pushing the EFIs out of the way.
4540		 */
4541		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
4542
4543		xlog_recover_process_iunlinks(log);
4544
4545		xlog_recover_check_summary(log);
4546
4547		xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
4548				log->l_mp->m_logname ? log->l_mp->m_logname
4549						     : "internal");
4550		log->l_flags &= ~XLOG_RECOVERY_NEEDED;
4551	} else {
4552		xfs_info(log->l_mp, "Ending clean mount");
4553	}
4554	return 0;
4555}
4556
4557
4558#if defined(DEBUG)
4559/*
4560 * Read all of the agf and agi counters and check that they
4561 * are consistent with the superblock counters.
4562 */
4563void
4564xlog_recover_check_summary(
4565	struct xlog	*log)
4566{
4567	xfs_mount_t	*mp;
4568	xfs_agf_t	*agfp;
4569	xfs_buf_t	*agfbp;
4570	xfs_buf_t	*agibp;
4571	xfs_agnumber_t	agno;
4572	__uint64_t	freeblks;
4573	__uint64_t	itotal;
4574	__uint64_t	ifree;
4575	int		error;
4576
4577	mp = log->l_mp;
4578
4579	freeblks = 0LL;
4580	itotal = 0LL;
4581	ifree = 0LL;
4582	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4583		error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
4584		if (error) {
4585			xfs_alert(mp, "%s agf read failed agno %d error %d",
4586						__func__, agno, error);
4587		} else {
4588			agfp = XFS_BUF_TO_AGF(agfbp);
4589			freeblks += be32_to_cpu(agfp->agf_freeblks) +
4590				    be32_to_cpu(agfp->agf_flcount);
4591			xfs_buf_relse(agfbp);
4592		}
4593
4594		error = xfs_read_agi(mp, NULL, agno, &agibp);
4595		if (error) {
4596			xfs_alert(mp, "%s agi read failed agno %d error %d",
4597						__func__, agno, error);
4598		} else {
4599			struct xfs_agi	*agi = XFS_BUF_TO_AGI(agibp);
4600
4601			itotal += be32_to_cpu(agi->agi_count);
4602			ifree += be32_to_cpu(agi->agi_freecount);
4603			xfs_buf_relse(agibp);
4604		}
4605	}
4606}
4607#endif /* DEBUG */