linux/fs/xfs/xfs_log_recover.c

4609 lines
127 KiB
C
Raw Normal View History

/*
* Copyright (c) 2000-2006 Silicon Graphics, Inc.
* All Rights Reserved.
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License as
* published by the Free Software Foundation.
*
* This program is distributed in the hope that it would be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "xfs.h"
#include "xfs_fs.h"
#include "xfs_shared.h"
#include "xfs_format.h"
#include "xfs_log_format.h"
#include "xfs_trans_resv.h"
#include "xfs_bit.h"
#include "xfs_inum.h"
#include "xfs_sb.h"
#include "xfs_ag.h"
#include "xfs_mount.h"
#include "xfs_da_format.h"
#include "xfs_inode.h"
#include "xfs_trans.h"
#include "xfs_log.h"
#include "xfs_log_priv.h"
#include "xfs_log_recover.h"
#include "xfs_inode_item.h"
#include "xfs_extfree_item.h"
#include "xfs_trans_priv.h"
#include "xfs_alloc.h"
#include "xfs_ialloc.h"
#include "xfs_quota.h"
2012-11-12 11:54:24 +00:00
#include "xfs_cksum.h"
xfs: event tracing support Convert the old xfs tracing support that could only be used with the out of tree kdb and xfsidbg patches to use the generic event tracer. To use it make sure CONFIG_EVENT_TRACING is enabled and then enable all xfs trace channels by: echo 1 > /sys/kernel/debug/tracing/events/xfs/enable or alternatively enable single events by just doing the same in one event subdirectory, e.g. echo 1 > /sys/kernel/debug/tracing/events/xfs/xfs_ihold/enable or set more complex filters, etc. In Documentation/trace/events.txt all this is desctribed in more detail. To reads the events do a cat /sys/kernel/debug/tracing/trace Compared to the last posting this patch converts the tracing mostly to the one tracepoint per callsite model that other users of the new tracing facility also employ. This allows a very fine-grained control of the tracing, a cleaner output of the traces and also enables the perf tool to use each tracepoint as a virtual performance counter, allowing us to e.g. count how often certain workloads git various spots in XFS. Take a look at http://lwn.net/Articles/346470/ for some examples. Also the btree tracing isn't included at all yet, as it will require additional core tracing features not in mainline yet, I plan to deliver it later. And the really nice thing about this patch is that it actually removes many lines of code while adding this nice functionality: fs/xfs/Makefile | 8 fs/xfs/linux-2.6/xfs_acl.c | 1 fs/xfs/linux-2.6/xfs_aops.c | 52 - fs/xfs/linux-2.6/xfs_aops.h | 2 fs/xfs/linux-2.6/xfs_buf.c | 117 +-- fs/xfs/linux-2.6/xfs_buf.h | 33 fs/xfs/linux-2.6/xfs_fs_subr.c | 3 fs/xfs/linux-2.6/xfs_ioctl.c | 1 fs/xfs/linux-2.6/xfs_ioctl32.c | 1 fs/xfs/linux-2.6/xfs_iops.c | 1 fs/xfs/linux-2.6/xfs_linux.h | 1 fs/xfs/linux-2.6/xfs_lrw.c | 87 -- fs/xfs/linux-2.6/xfs_lrw.h | 45 - fs/xfs/linux-2.6/xfs_super.c | 104 --- fs/xfs/linux-2.6/xfs_super.h | 7 fs/xfs/linux-2.6/xfs_sync.c | 1 fs/xfs/linux-2.6/xfs_trace.c | 75 ++ fs/xfs/linux-2.6/xfs_trace.h | 1369 +++++++++++++++++++++++++++++++++++++++++ fs/xfs/linux-2.6/xfs_vnode.h | 4 fs/xfs/quota/xfs_dquot.c | 110 --- fs/xfs/quota/xfs_dquot.h | 21 fs/xfs/quota/xfs_qm.c | 40 - fs/xfs/quota/xfs_qm_syscalls.c | 4 fs/xfs/support/ktrace.c | 323 --------- fs/xfs/support/ktrace.h | 85 -- fs/xfs/xfs.h | 16 fs/xfs/xfs_ag.h | 14 fs/xfs/xfs_alloc.c | 230 +----- fs/xfs/xfs_alloc.h | 27 fs/xfs/xfs_alloc_btree.c | 1 fs/xfs/xfs_attr.c | 107 --- fs/xfs/xfs_attr.h | 10 fs/xfs/xfs_attr_leaf.c | 14 fs/xfs/xfs_attr_sf.h | 40 - fs/xfs/xfs_bmap.c | 507 +++------------ fs/xfs/xfs_bmap.h | 49 - fs/xfs/xfs_bmap_btree.c | 6 fs/xfs/xfs_btree.c | 5 fs/xfs/xfs_btree_trace.h | 17 fs/xfs/xfs_buf_item.c | 87 -- fs/xfs/xfs_buf_item.h | 20 fs/xfs/xfs_da_btree.c | 3 fs/xfs/xfs_da_btree.h | 7 fs/xfs/xfs_dfrag.c | 2 fs/xfs/xfs_dir2.c | 8 fs/xfs/xfs_dir2_block.c | 20 fs/xfs/xfs_dir2_leaf.c | 21 fs/xfs/xfs_dir2_node.c | 27 fs/xfs/xfs_dir2_sf.c | 26 fs/xfs/xfs_dir2_trace.c | 216 ------ fs/xfs/xfs_dir2_trace.h | 72 -- fs/xfs/xfs_filestream.c | 8 fs/xfs/xfs_fsops.c | 2 fs/xfs/xfs_iget.c | 111 --- fs/xfs/xfs_inode.c | 67 -- fs/xfs/xfs_inode.h | 76 -- fs/xfs/xfs_inode_item.c | 5 fs/xfs/xfs_iomap.c | 85 -- fs/xfs/xfs_iomap.h | 8 fs/xfs/xfs_log.c | 181 +---- fs/xfs/xfs_log_priv.h | 20 fs/xfs/xfs_log_recover.c | 1 fs/xfs/xfs_mount.c | 2 fs/xfs/xfs_quota.h | 8 fs/xfs/xfs_rename.c | 1 fs/xfs/xfs_rtalloc.c | 1 fs/xfs/xfs_rw.c | 3 fs/xfs/xfs_trans.h | 47 + fs/xfs/xfs_trans_buf.c | 62 - fs/xfs/xfs_vnodeops.c | 8 70 files changed, 2151 insertions(+), 2592 deletions(-) Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Alex Elder <aelder@sgi.com>
2009-12-14 23:14:59 +00:00
#include "xfs_trace.h"
#include "xfs_icache.h"
#include "xfs_bmap_btree.h"
#include "xfs_dinode.h"
#include "xfs_error.h"
#include "xfs_dir2.h"
#define BLK_AVG(blk1, blk2) ((blk1+blk2) >> 1)
STATIC int
xlog_find_zeroed(
struct xlog *,
xfs_daddr_t *);
STATIC int
xlog_clear_stale_blocks(
struct xlog *,
xfs_lsn_t);
#if defined(DEBUG)
STATIC void
xlog_recover_check_summary(
struct xlog *);
#else
#define xlog_recover_check_summary(log)
#endif
/*
* This structure is used during recovery to record the buf log items which
* have been canceled and should not be replayed.
*/
struct xfs_buf_cancel {
xfs_daddr_t bc_blkno;
uint bc_len;
int bc_refcount;
struct list_head bc_list;
};
/*
* Sector aligned buffer routines for buffer create/read/write/access
*/
/*
* Verify the given count of basic blocks is valid number of blocks
* to specify for an operation involving the given XFS log buffer.
* Returns nonzero if the count is valid, 0 otherwise.
*/
static inline int
xlog_buf_bbcount_valid(
struct xlog *log,
int bbcount)
{
return bbcount > 0 && bbcount <= log->l_logBBsize;
}
/*
* Allocate a buffer to hold log data. The buffer needs to be able
* to map to a range of nbblks basic blocks at any valid (basic
* block) offset within the log.
*/
STATIC xfs_buf_t *
xlog_get_bp(
struct xlog *log,
int nbblks)
{
struct xfs_buf *bp;
if (!xlog_buf_bbcount_valid(log, nbblks)) {
xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
nbblks);
XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
return NULL;
}
/*
* We do log I/O in units of log sectors (a power-of-2
* multiple of the basic block size), so we round up the
* requested size to accommodate the basic blocks required
* for complete log sectors.
*
* In addition, the buffer may be used for a non-sector-
* aligned block offset, in which case an I/O of the
* requested size could extend beyond the end of the
* buffer. If the requested size is only 1 basic block it
* will never straddle a sector boundary, so this won't be
* an issue. Nor will this be a problem if the log I/O is
* done in basic blocks (sector size 1). But otherwise we
* extend the buffer by one extra log sector to ensure
* there's space to accommodate this possibility.
*/
if (nbblks > 1 && log->l_sectBBsize > 1)
nbblks += log->l_sectBBsize;
nbblks = round_up(nbblks, log->l_sectBBsize);
bp = xfs_buf_get_uncached(log->l_mp->m_logdev_targp, nbblks, 0);
if (bp)
xfs_buf_unlock(bp);
return bp;
}
STATIC void
xlog_put_bp(
xfs_buf_t *bp)
{
xfs_buf_free(bp);
}
/*
* Return the address of the start of the given block number's data
* in a log buffer. The buffer covers a log sector-aligned region.
*/
STATIC xfs_caddr_t
xlog_align(
struct xlog *log,
xfs_daddr_t blk_no,
int nbblks,
struct xfs_buf *bp)
{
xfs_daddr_t offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
ASSERT(offset + nbblks <= bp->b_length);
return bp->b_addr + BBTOB(offset);
}
/*
* nbblks should be uint, but oh well. Just want to catch that 32-bit length.
*/
STATIC int
xlog_bread_noalign(
struct xlog *log,
xfs_daddr_t blk_no,
int nbblks,
struct xfs_buf *bp)
{
int error;
if (!xlog_buf_bbcount_valid(log, nbblks)) {
xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
nbblks);
XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
return EFSCORRUPTED;
}
blk_no = round_down(blk_no, log->l_sectBBsize);
nbblks = round_up(nbblks, log->l_sectBBsize);
ASSERT(nbblks > 0);
ASSERT(nbblks <= bp->b_length);
XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
XFS_BUF_READ(bp);
bp->b_io_length = nbblks;
bp->b_error = 0;
if (XFS_FORCED_SHUTDOWN(log->l_mp))
return XFS_ERROR(EIO);
xfs_buf_iorequest(bp);
error = xfs_buf_iowait(bp);
if (error)
xfs_buf_ioerror_alert(bp, __func__);
return error;
}
STATIC int
xlog_bread(
struct xlog *log,
xfs_daddr_t blk_no,
int nbblks,
struct xfs_buf *bp,
xfs_caddr_t *offset)
{
int error;
error = xlog_bread_noalign(log, blk_no, nbblks, bp);
if (error)
return error;
*offset = xlog_align(log, blk_no, nbblks, bp);
return 0;
}
/*
* Read at an offset into the buffer. Returns with the buffer in it's original
* state regardless of the result of the read.
*/
STATIC int
xlog_bread_offset(
struct xlog *log,
xfs_daddr_t blk_no, /* block to read from */
int nbblks, /* blocks to read */
struct xfs_buf *bp,
xfs_caddr_t offset)
{
xfs_caddr_t orig_offset = bp->b_addr;
int orig_len = BBTOB(bp->b_length);
int error, error2;
error = xfs_buf_associate_memory(bp, offset, BBTOB(nbblks));
if (error)
return error;
error = xlog_bread_noalign(log, blk_no, nbblks, bp);
/* must reset buffer pointer even on error */
error2 = xfs_buf_associate_memory(bp, orig_offset, orig_len);
if (error)
return error;
return error2;
}
/*
* Write out the buffer at the given block for the given number of blocks.
* The buffer is kept locked across the write and is returned locked.
* This can only be used for synchronous log writes.
*/
STATIC int
xlog_bwrite(
struct xlog *log,
xfs_daddr_t blk_no,
int nbblks,
struct xfs_buf *bp)
{
int error;
if (!xlog_buf_bbcount_valid(log, nbblks)) {
xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
nbblks);
XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
return EFSCORRUPTED;
}
blk_no = round_down(blk_no, log->l_sectBBsize);
nbblks = round_up(nbblks, log->l_sectBBsize);
ASSERT(nbblks > 0);
ASSERT(nbblks <= bp->b_length);
XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
XFS_BUF_ZEROFLAGS(bp);
xfs_buf_hold(bp);
xfs_buf_lock(bp);
bp->b_io_length = nbblks;
bp->b_error = 0;
error = xfs_bwrite(bp);
if (error)
xfs_buf_ioerror_alert(bp, __func__);
xfs_buf_relse(bp);
return error;
}
#ifdef DEBUG
/*
* dump debug superblock and log record information
*/
STATIC void
xlog_header_check_dump(
xfs_mount_t *mp,
xlog_rec_header_t *head)
{
xfs_debug(mp, "%s: SB : uuid = %pU, fmt = %d",
__func__, &mp->m_sb.sb_uuid, XLOG_FMT);
xfs_debug(mp, " log : uuid = %pU, fmt = %d",
&head->h_fs_uuid, be32_to_cpu(head->h_fmt));
}
#else
#define xlog_header_check_dump(mp, head)
#endif
/*
* check log record header for recovery
*/
STATIC int
xlog_header_check_recover(
xfs_mount_t *mp,
xlog_rec_header_t *head)
{
ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
/*
* IRIX doesn't write the h_fmt field and leaves it zeroed
* (XLOG_FMT_UNKNOWN). This stops us from trying to recover
* a dirty log created in IRIX.
*/
if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
xfs_warn(mp,
"dirty log written in incompatible format - can't recover");
xlog_header_check_dump(mp, head);
XFS_ERROR_REPORT("xlog_header_check_recover(1)",
XFS_ERRLEVEL_HIGH, mp);
return XFS_ERROR(EFSCORRUPTED);
} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
xfs_warn(mp,
"dirty log entry has mismatched uuid - can't recover");
xlog_header_check_dump(mp, head);
XFS_ERROR_REPORT("xlog_header_check_recover(2)",
XFS_ERRLEVEL_HIGH, mp);
return XFS_ERROR(EFSCORRUPTED);
}
return 0;
}
/*
* read the head block of the log and check the header
*/
STATIC int
xlog_header_check_mount(
xfs_mount_t *mp,
xlog_rec_header_t *head)
{
ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
if (uuid_is_nil(&head->h_fs_uuid)) {
/*
* IRIX doesn't write the h_fs_uuid or h_fmt fields. If
* h_fs_uuid is nil, we assume this log was last mounted
* by IRIX and continue.
*/
xfs_warn(mp, "nil uuid in log - IRIX style log");
} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
xfs_warn(mp, "log has mismatched uuid - can't recover");
xlog_header_check_dump(mp, head);
XFS_ERROR_REPORT("xlog_header_check_mount",
XFS_ERRLEVEL_HIGH, mp);
return XFS_ERROR(EFSCORRUPTED);
}
return 0;
}
STATIC void
xlog_recover_iodone(
struct xfs_buf *bp)
{
if (bp->b_error) {
/*
* We're not going to bother about retrying
* this during recovery. One strike!
*/
xfs_buf_ioerror_alert(bp, __func__);
xfs_force_shutdown(bp->b_target->bt_mount,
SHUTDOWN_META_IO_ERROR);
}
bp->b_iodone = NULL;
xfs_buf_ioend(bp, 0);
}
/*
* This routine finds (to an approximation) the first block in the physical
* log which contains the given cycle. It uses a binary search algorithm.
* Note that the algorithm can not be perfect because the disk will not
* necessarily be perfect.
*/
STATIC int
xlog_find_cycle_start(
struct xlog *log,
struct xfs_buf *bp,
xfs_daddr_t first_blk,
xfs_daddr_t *last_blk,
uint cycle)
{
xfs_caddr_t offset;
xfs_daddr_t mid_blk;
xfs_daddr_t end_blk;
uint mid_cycle;
int error;
end_blk = *last_blk;
mid_blk = BLK_AVG(first_blk, end_blk);
while (mid_blk != first_blk && mid_blk != end_blk) {
error = xlog_bread(log, mid_blk, 1, bp, &offset);
if (error)
return error;
mid_cycle = xlog_get_cycle(offset);
if (mid_cycle == cycle)
end_blk = mid_blk; /* last_half_cycle == mid_cycle */
else
first_blk = mid_blk; /* first_half_cycle == mid_cycle */
mid_blk = BLK_AVG(first_blk, end_blk);
}
ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
(mid_blk == end_blk && mid_blk-1 == first_blk));
*last_blk = end_blk;
return 0;
}
/*
* Check that a range of blocks does not contain stop_on_cycle_no.
* Fill in *new_blk with the block offset where such a block is
* found, or with -1 (an invalid block number) if there is no such
* block in the range. The scan needs to occur from front to back
* and the pointer into the region must be updated since a later
* routine will need to perform another test.
*/
STATIC int
xlog_find_verify_cycle(
struct xlog *log,
xfs_daddr_t start_blk,
int nbblks,
uint stop_on_cycle_no,
xfs_daddr_t *new_blk)
{
xfs_daddr_t i, j;
uint cycle;
xfs_buf_t *bp;
xfs_daddr_t bufblks;
xfs_caddr_t buf = NULL;
int error = 0;
/*
* Greedily allocate a buffer big enough to handle the full
* range of basic blocks we'll be examining. If that fails,
* try a smaller size. We need to be able to read at least
* a log sector, or we're out of luck.
*/
bufblks = 1 << ffs(nbblks);
while (bufblks > log->l_logBBsize)
bufblks >>= 1;
while (!(bp = xlog_get_bp(log, bufblks))) {
bufblks >>= 1;
if (bufblks < log->l_sectBBsize)
return ENOMEM;
}
for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
int bcount;
bcount = min(bufblks, (start_blk + nbblks - i));
error = xlog_bread(log, i, bcount, bp, &buf);
if (error)
goto out;
for (j = 0; j < bcount; j++) {
cycle = xlog_get_cycle(buf);
if (cycle == stop_on_cycle_no) {
*new_blk = i+j;
goto out;
}
buf += BBSIZE;
}
}
*new_blk = -1;
out:
xlog_put_bp(bp);
return error;
}
/*
* Potentially backup over partial log record write.
*
* In the typical case, last_blk is the number of the block directly after
* a good log record. Therefore, we subtract one to get the block number
* of the last block in the given buffer. extra_bblks contains the number
* of blocks we would have read on a previous read. This happens when the
* last log record is split over the end of the physical log.
*
* extra_bblks is the number of blocks potentially verified on a previous
* call to this routine.
*/
STATIC int
xlog_find_verify_log_record(
struct xlog *log,
xfs_daddr_t start_blk,
xfs_daddr_t *last_blk,
int extra_bblks)
{
xfs_daddr_t i;
xfs_buf_t *bp;
xfs_caddr_t offset = NULL;
xlog_rec_header_t *head = NULL;
int error = 0;
int smallmem = 0;
int num_blks = *last_blk - start_blk;
int xhdrs;
ASSERT(start_blk != 0 || *last_blk != start_blk);
if (!(bp = xlog_get_bp(log, num_blks))) {
if (!(bp = xlog_get_bp(log, 1)))
return ENOMEM;
smallmem = 1;
} else {
error = xlog_bread(log, start_blk, num_blks, bp, &offset);
if (error)
goto out;
offset += ((num_blks - 1) << BBSHIFT);
}
for (i = (*last_blk) - 1; i >= 0; i--) {
if (i < start_blk) {
/* valid log record not found */
xfs_warn(log->l_mp,
"Log inconsistent (didn't find previous header)");
ASSERT(0);
error = XFS_ERROR(EIO);
goto out;
}
if (smallmem) {
error = xlog_bread(log, i, 1, bp, &offset);
if (error)
goto out;
}
head = (xlog_rec_header_t *)offset;
if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
break;
if (!smallmem)
offset -= BBSIZE;
}
/*
* We hit the beginning of the physical log & still no header. Return
* to caller. If caller can handle a return of -1, then this routine
* will be called again for the end of the physical log.
*/
if (i == -1) {
error = -1;
goto out;
}
/*
* We have the final block of the good log (the first block
* of the log record _before_ the head. So we check the uuid.
*/
if ((error = xlog_header_check_mount(log->l_mp, head)))
goto out;
/*
* We may have found a log record header before we expected one.
* last_blk will be the 1st block # with a given cycle #. We may end
* up reading an entire log record. In this case, we don't want to
* reset last_blk. Only when last_blk points in the middle of a log
* record do we update last_blk.
*/
if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
uint h_size = be32_to_cpu(head->h_size);
xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
if (h_size % XLOG_HEADER_CYCLE_SIZE)
xhdrs++;
} else {
xhdrs = 1;
}
if (*last_blk - i + extra_bblks !=
BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
*last_blk = i;
out:
xlog_put_bp(bp);
return error;
}
/*
* Head is defined to be the point of the log where the next log write
* could go. This means that incomplete LR writes at the end are
* eliminated when calculating the head. We aren't guaranteed that previous
* LR have complete transactions. We only know that a cycle number of
* current cycle number -1 won't be present in the log if we start writing
* from our current block number.
*
* last_blk contains the block number of the first block with a given
* cycle number.
*
* Return: zero if normal, non-zero if error.
*/
STATIC int
xlog_find_head(
struct xlog *log,
xfs_daddr_t *return_head_blk)
{
xfs_buf_t *bp;
xfs_caddr_t offset;
xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
int num_scan_bblks;
uint first_half_cycle, last_half_cycle;
uint stop_on_cycle;
int error, log_bbnum = log->l_logBBsize;
/* Is the end of the log device zeroed? */
if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
*return_head_blk = first_blk;
/* Is the whole lot zeroed? */
if (!first_blk) {
/* Linux XFS shouldn't generate totally zeroed logs -
* mkfs etc write a dummy unmount record to a fresh
* log so we can store the uuid in there
*/
xfs_warn(log->l_mp, "totally zeroed log");
}
return 0;
} else if (error) {
xfs_warn(log->l_mp, "empty log check failed");
return error;
}
first_blk = 0; /* get cycle # of 1st block */
bp = xlog_get_bp(log, 1);
if (!bp)
return ENOMEM;
error = xlog_bread(log, 0, 1, bp, &offset);
if (error)
goto bp_err;
first_half_cycle = xlog_get_cycle(offset);
last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
error = xlog_bread(log, last_blk, 1, bp, &offset);
if (error)
goto bp_err;
last_half_cycle = xlog_get_cycle(offset);
ASSERT(last_half_cycle != 0);
/*
* If the 1st half cycle number is equal to the last half cycle number,
* then the entire log is stamped with the same cycle number. In this
* case, head_blk can't be set to zero (which makes sense). The below
* math doesn't work out properly with head_blk equal to zero. Instead,
* we set it to log_bbnum which is an invalid block number, but this
* value makes the math correct. If head_blk doesn't changed through
* all the tests below, *head_blk is set to zero at the very end rather
* than log_bbnum. In a sense, log_bbnum and zero are the same block
* in a circular file.
*/
if (first_half_cycle == last_half_cycle) {
/*
* In this case we believe that the entire log should have
* cycle number last_half_cycle. We need to scan backwards
* from the end verifying that there are no holes still
* containing last_half_cycle - 1. If we find such a hole,
* then the start of that hole will be the new head. The
* simple case looks like
* x | x ... | x - 1 | x
* Another case that fits this picture would be
* x | x + 1 | x ... | x
* In this case the head really is somewhere at the end of the
* log, as one of the latest writes at the beginning was
* incomplete.
* One more case is
* x | x + 1 | x ... | x - 1 | x
* This is really the combination of the above two cases, and
* the head has to end up at the start of the x-1 hole at the
* end of the log.
*
* In the 256k log case, we will read from the beginning to the
* end of the log and search for cycle numbers equal to x-1.
* We don't worry about the x+1 blocks that we encounter,
* because we know that they cannot be the head since the log
* started with x.
*/
head_blk = log_bbnum;
stop_on_cycle = last_half_cycle - 1;
} else {
/*
* In this case we want to find the first block with cycle
* number matching last_half_cycle. We expect the log to be
* some variation on
* x + 1 ... | x ... | x
* The first block with cycle number x (last_half_cycle) will
* be where the new head belongs. First we do a binary search
* for the first occurrence of last_half_cycle. The binary
* search may not be totally accurate, so then we scan back
* from there looking for occurrences of last_half_cycle before
* us. If that backwards scan wraps around the beginning of
* the log, then we look for occurrences of last_half_cycle - 1
* at the end of the log. The cases we're looking for look
* like
* v binary search stopped here
* x + 1 ... | x | x + 1 | x ... | x
* ^ but we want to locate this spot
* or
* <---------> less than scan distance
* x + 1 ... | x ... | x - 1 | x
* ^ we want to locate this spot
*/
stop_on_cycle = last_half_cycle;
if ((error = xlog_find_cycle_start(log, bp, first_blk,
&head_blk, last_half_cycle)))
goto bp_err;
}
/*
* Now validate the answer. Scan back some number of maximum possible
* blocks and make sure each one has the expected cycle number. The
* maximum is determined by the total possible amount of buffering
* in the in-core log. The following number can be made tighter if
* we actually look at the block size of the filesystem.
*/
num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
if (head_blk >= num_scan_bblks) {
/*
* We are guaranteed that the entire check can be performed
* in one buffer.
*/
start_blk = head_blk - num_scan_bblks;
if ((error = xlog_find_verify_cycle(log,
start_blk, num_scan_bblks,
stop_on_cycle, &new_blk)))
goto bp_err;
if (new_blk != -1)
head_blk = new_blk;
} else { /* need to read 2 parts of log */
/*
* We are going to scan backwards in the log in two parts.
* First we scan the physical end of the log. In this part
* of the log, we are looking for blocks with cycle number
* last_half_cycle - 1.
* If we find one, then we know that the log starts there, as
* we've found a hole that didn't get written in going around
* the end of the physical log. The simple case for this is
* x + 1 ... | x ... | x - 1 | x
* <---------> less than scan distance
* If all of the blocks at the end of the log have cycle number
* last_half_cycle, then we check the blocks at the start of
* the log looking for occurrences of last_half_cycle. If we
* find one, then our current estimate for the location of the
* first occurrence of last_half_cycle is wrong and we move
* back to the hole we've found. This case looks like
* x + 1 ... | x | x + 1 | x ...
* ^ binary search stopped here
* Another case we need to handle that only occurs in 256k
* logs is
* x + 1 ... | x ... | x+1 | x ...
* ^ binary search stops here
* In a 256k log, the scan at the end of the log will see the
* x + 1 blocks. We need to skip past those since that is
* certainly not the head of the log. By searching for
* last_half_cycle-1 we accomplish that.
*/
ASSERT(head_blk <= INT_MAX &&
(xfs_daddr_t) num_scan_bblks >= head_blk);
start_blk = log_bbnum - (num_scan_bblks - head_blk);
if ((error = xlog_find_verify_cycle(log, start_blk,
num_scan_bblks - (int)head_blk,
(stop_on_cycle - 1), &new_blk)))
goto bp_err;
if (new_blk != -1) {
head_blk = new_blk;
goto validate_head;
}
/*
* Scan beginning of log now. The last part of the physical
* log is good. This scan needs to verify that it doesn't find
* the last_half_cycle.
*/
start_blk = 0;
ASSERT(head_blk <= INT_MAX);
if ((error = xlog_find_verify_cycle(log,
start_blk, (int)head_blk,
stop_on_cycle, &new_blk)))
goto bp_err;
if (new_blk != -1)
head_blk = new_blk;
}
validate_head:
/*
* Now we need to make sure head_blk is not pointing to a block in
* the middle of a log record.
*/
num_scan_bblks = XLOG_REC_SHIFT(log);
if (head_blk >= num_scan_bblks) {
start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
/* start ptr at last block ptr before head_blk */
if ((error = xlog_find_verify_log_record(log, start_blk,
&head_blk, 0)) == -1) {
error = XFS_ERROR(EIO);
goto bp_err;
} else if (error)
goto bp_err;
} else {
start_blk = 0;
ASSERT(head_blk <= INT_MAX);
if ((error = xlog_find_verify_log_record(log, start_blk,
&head_blk, 0)) == -1) {
/* We hit the beginning of the log during our search */
start_blk = log_bbnum - (num_scan_bblks - head_blk);
new_blk = log_bbnum;
ASSERT(start_blk <= INT_MAX &&
(xfs_daddr_t) log_bbnum-start_blk >= 0);
ASSERT(head_blk <= INT_MAX);
if ((error = xlog_find_verify_log_record(log,
start_blk, &new_blk,
(int)head_blk)) == -1) {
error = XFS_ERROR(EIO);
goto bp_err;
} else if (error)
goto bp_err;
if (new_blk != log_bbnum)
head_blk = new_blk;
} else if (error)
goto bp_err;
}
xlog_put_bp(bp);
if (head_blk == log_bbnum)
*return_head_blk = 0;
else
*return_head_blk = head_blk;
/*
* When returning here, we have a good block number. Bad block
* means that during a previous crash, we didn't have a clean break
* from cycle number N to cycle number N-1. In this case, we need
* to find the first block with cycle number N-1.
*/
return 0;
bp_err:
xlog_put_bp(bp);
if (error)
xfs_warn(log->l_mp, "failed to find log head");
return error;
}
/*
* Find the sync block number or the tail of the log.
*
* This will be the block number of the last record to have its
* associated buffers synced to disk. Every log record header has
* a sync lsn embedded in it. LSNs hold block numbers, so it is easy
* to get a sync block number. The only concern is to figure out which
* log record header to believe.
*
* The following algorithm uses the log record header with the largest
* lsn. The entire log record does not need to be valid. We only care
* that the header is valid.
*
* We could speed up search by using current head_blk buffer, but it is not
* available.
*/
STATIC int
xlog_find_tail(
struct xlog *log,
xfs_daddr_t *head_blk,
xfs_daddr_t *tail_blk)
{
xlog_rec_header_t *rhead;
xlog_op_header_t *op_head;
xfs_caddr_t offset = NULL;
xfs_buf_t *bp;
int error, i, found;
xfs_daddr_t umount_data_blk;
xfs_daddr_t after_umount_blk;
xfs_lsn_t tail_lsn;
int hblks;
found = 0;
/*
* Find previous log record
*/
if ((error = xlog_find_head(log, head_blk)))
return error;
bp = xlog_get_bp(log, 1);
if (!bp)
return ENOMEM;
if (*head_blk == 0) { /* special case */
error = xlog_bread(log, 0, 1, bp, &offset);
if (error)
goto done;
if (xlog_get_cycle(offset) == 0) {
*tail_blk = 0;
/* leave all other log inited values alone */
goto done;
}
}
/*
* Search backwards looking for log record header block
*/
ASSERT(*head_blk < INT_MAX);
for (i = (int)(*head_blk) - 1; i >= 0; i--) {
error = xlog_bread(log, i, 1, bp, &offset);
if (error)
goto done;
if (*(__be32 *)offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
found = 1;
break;
}
}
/*
* If we haven't found the log record header block, start looking
* again from the end of the physical log. XXXmiken: There should be
* a check here to make sure we didn't search more than N blocks in
* the previous code.
*/
if (!found) {
for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
error = xlog_bread(log, i, 1, bp, &offset);
if (error)
goto done;
if (*(__be32 *)offset ==
cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
found = 2;
break;
}
}
}
if (!found) {
xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
xlog_put_bp(bp);
ASSERT(0);
return XFS_ERROR(EIO);
}
/* find blk_no of tail of log */
rhead = (xlog_rec_header_t *)offset;
*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
/*
* Reset log values according to the state of the log when we
* crashed. In the case where head_blk == 0, we bump curr_cycle
* one because the next write starts a new cycle rather than
* continuing the cycle of the last good log record. At this
* point we have guaranteed that all partial log records have been
* accounted for. Therefore, we know that the last good log record
* written was complete and ended exactly on the end boundary
* of the physical log.
*/
log->l_prev_block = i;
log->l_curr_block = (int)*head_blk;
log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
if (found == 2)
log->l_curr_cycle++;
atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
BBTOB(log->l_curr_block));
xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
BBTOB(log->l_curr_block));
/*
* Look for unmount record. If we find it, then we know there
* was a clean unmount. Since 'i' could be the last block in
* the physical log, we convert to a log block before comparing
* to the head_blk.
*
* Save the current tail lsn to use to pass to
* xlog_clear_stale_blocks() below. We won't want to clear the
* unmount record if there is one, so we pass the lsn of the
* unmount record rather than the block after it.
*/
if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
int h_size = be32_to_cpu(rhead->h_size);
int h_version = be32_to_cpu(rhead->h_version);
if ((h_version & XLOG_VERSION_2) &&
(h_size > XLOG_HEADER_CYCLE_SIZE)) {
hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
if (h_size % XLOG_HEADER_CYCLE_SIZE)
hblks++;
} else {
hblks = 1;
}
} else {
hblks = 1;
}
after_umount_blk = (i + hblks + (int)
BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
tail_lsn = atomic64_read(&log->l_tail_lsn);
if (*head_blk == after_umount_blk &&
be32_to_cpu(rhead->h_num_logops) == 1) {
umount_data_blk = (i + hblks) % log->l_logBBsize;
error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
if (error)
goto done;
op_head = (xlog_op_header_t *)offset;
if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
/*
* Set tail and last sync so that newly written
* log records will point recovery to after the
* current unmount record.
*/
xlog_assign_atomic_lsn(&log->l_tail_lsn,
log->l_curr_cycle, after_umount_blk);
xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
log->l_curr_cycle, after_umount_blk);
*tail_blk = after_umount_blk;
[XFS] Lazy Superblock Counters When we have a couple of hundred transactions on the fly at once, they all typically modify the on disk superblock in some way. create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify free block counts. When these counts are modified in a transaction, they must eventually lock the superblock buffer and apply the mods. The buffer then remains locked until the transaction is committed into the incore log buffer. The result of this is that with enough transactions on the fly the incore superblock buffer becomes a bottleneck. The result of contention on the incore superblock buffer is that transaction rates fall - the more pressure that is put on the superblock buffer, the slower things go. The key to removing the contention is to not require the superblock fields in question to be locked. We do that by not marking the superblock dirty in the transaction. IOWs, we modify the incore superblock but do not modify the cached superblock buffer. In short, we do not log superblock modifications to critical fields in the superblock on every transaction. In fact we only do it just before we write the superblock to disk every sync period or just before unmount. This creates an interesting problem - if we don't log or write out the fields in every transaction, then how do the values get recovered after a crash? the answer is simple - we keep enough duplicate, logged information in other structures that we can reconstruct the correct count after log recovery has been performed. It is the AGF and AGI structures that contain the duplicate information; after recovery, we walk every AGI and AGF and sum their individual counters to get the correct value, and we do a transaction into the log to correct them. An optimisation of this is that if we have a clean unmount record, we know the value in the superblock is correct, so we can avoid the summation walk under normal conditions and so mount/recovery times do not change under normal operation. One wrinkle that was discovered during development was that the blocks used in the freespace btrees are never accounted for in the AGF counters. This was once a valid optimisation to make; when the filesystem is full, the free space btrees are empty and consume no space. Hence when it matters, the "accounting" is correct. But that means the when we do the AGF summations, we would not have a correct count and xfs_check would complain. Hence a new counter was added to track the number of blocks used by the free space btrees. This is an *on-disk format change*. As a result of this, lazy superblock counters are a mkfs option and at the moment on linux there is no way to convert an old filesystem. This is possible - xfs_db can be used to twiddle the right bits and then xfs_repair will do the format conversion for you. Similarly, you can convert backwards as well. At some point we'll add functionality to xfs_admin to do the bit twiddling easily.... SGI-PV: 964999 SGI-Modid: xfs-linux-melb:xfs-kern:28652a Signed-off-by: David Chinner <dgc@sgi.com> Signed-off-by: Christoph Hellwig <hch@infradead.org> Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 05:26:31 +00:00
/*
* Note that the unmount was clean. If the unmount
* was not clean, we need to know this to rebuild the
* superblock counters from the perag headers if we
* have a filesystem using non-persistent counters.
*/
log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
}
}
/*
* Make sure that there are no blocks in front of the head
* with the same cycle number as the head. This can happen
* because we allow multiple outstanding log writes concurrently,
* and the later writes might make it out before earlier ones.
*
* We use the lsn from before modifying it so that we'll never
* overwrite the unmount record after a clean unmount.
*
* Do this only if we are going to recover the filesystem
*
* NOTE: This used to say "if (!readonly)"
* However on Linux, we can & do recover a read-only filesystem.
* We only skip recovery if NORECOVERY is specified on mount,
* in which case we would not be here.
*
* But... if the -device- itself is readonly, just skip this.
* We can't recover this device anyway, so it won't matter.
*/
if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
error = xlog_clear_stale_blocks(log, tail_lsn);
done:
xlog_put_bp(bp);
if (error)
xfs_warn(log->l_mp, "failed to locate log tail");
return error;
}
/*
* Is the log zeroed at all?
*
* The last binary search should be changed to perform an X block read
* once X becomes small enough. You can then search linearly through
* the X blocks. This will cut down on the number of reads we need to do.
*
* If the log is partially zeroed, this routine will pass back the blkno
* of the first block with cycle number 0. It won't have a complete LR
* preceding it.
*
* Return:
* 0 => the log is completely written to
* -1 => use *blk_no as the first block of the log
* >0 => error has occurred
*/
STATIC int
xlog_find_zeroed(
struct xlog *log,
xfs_daddr_t *blk_no)
{
xfs_buf_t *bp;
xfs_caddr_t offset;
uint first_cycle, last_cycle;
xfs_daddr_t new_blk, last_blk, start_blk;
xfs_daddr_t num_scan_bblks;
int error, log_bbnum = log->l_logBBsize;
*blk_no = 0;
/* check totally zeroed log */
bp = xlog_get_bp(log, 1);
if (!bp)
return ENOMEM;
error = xlog_bread(log, 0, 1, bp, &offset);
if (error)
goto bp_err;
first_cycle = xlog_get_cycle(offset);
if (first_cycle == 0) { /* completely zeroed log */
*blk_no = 0;
xlog_put_bp(bp);
return -1;
}
/* check partially zeroed log */
error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
if (error)
goto bp_err;
last_cycle = xlog_get_cycle(offset);
if (last_cycle != 0) { /* log completely written to */
xlog_put_bp(bp);
return 0;
} else if (first_cycle != 1) {
/*
* If the cycle of the last block is zero, the cycle of
* the first block must be 1. If it's not, maybe we're
* not looking at a log... Bail out.
*/
xfs_warn(log->l_mp,
"Log inconsistent or not a log (last==0, first!=1)");
error = XFS_ERROR(EINVAL);
goto bp_err;
}
/* we have a partially zeroed log */
last_blk = log_bbnum-1;
if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
goto bp_err;
/*
* Validate the answer. Because there is no way to guarantee that
* the entire log is made up of log records which are the same size,
* we scan over the defined maximum blocks. At this point, the maximum
* is not chosen to mean anything special. XXXmiken
*/
num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
ASSERT(num_scan_bblks <= INT_MAX);
if (last_blk < num_scan_bblks)
num_scan_bblks = last_blk;
start_blk = last_blk - num_scan_bblks;
/*
* We search for any instances of cycle number 0 that occur before
* our current estimate of the head. What we're trying to detect is
* 1 ... | 0 | 1 | 0...
* ^ binary search ends here
*/
if ((error = xlog_find_verify_cycle(log, start_blk,
(int)num_scan_bblks, 0, &new_blk)))
goto bp_err;
if (new_blk != -1)
last_blk = new_blk;
/*
* Potentially backup over partial log record write. We don't need
* to search the end of the log because we know it is zero.
*/
if ((error = xlog_find_verify_log_record(log, start_blk,
&last_blk, 0)) == -1) {
error = XFS_ERROR(EIO);
goto bp_err;
} else if (error)
goto bp_err;
*blk_no = last_blk;
bp_err:
xlog_put_bp(bp);
if (error)
return error;
return -1;
}
/*
* These are simple subroutines used by xlog_clear_stale_blocks() below
* to initialize a buffer full of empty log record headers and write
* them into the log.
*/
STATIC void
xlog_add_record(
struct xlog *log,
xfs_caddr_t buf,
int cycle,
int block,
int tail_cycle,
int tail_block)
{
xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
memset(buf, 0, BBSIZE);
recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
recp->h_cycle = cpu_to_be32(cycle);
recp->h_version = cpu_to_be32(
xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
recp->h_fmt = cpu_to_be32(XLOG_FMT);
memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
}
STATIC int
xlog_write_log_records(
struct xlog *log,
int cycle,
int start_block,
int blocks,
int tail_cycle,
int tail_block)
{
xfs_caddr_t offset;
xfs_buf_t *bp;
int balign, ealign;
int sectbb = log->l_sectBBsize;
int end_block = start_block + blocks;
int bufblks;
int error = 0;
int i, j = 0;
/*
* Greedily allocate a buffer big enough to handle the full
* range of basic blocks to be written. If that fails, try
* a smaller size. We need to be able to write at least a
* log sector, or we're out of luck.
*/
bufblks = 1 << ffs(blocks);
while (bufblks > log->l_logBBsize)
bufblks >>= 1;
while (!(bp = xlog_get_bp(log, bufblks))) {
bufblks >>= 1;
if (bufblks < sectbb)
return ENOMEM;
}
/* We may need to do a read at the start to fill in part of
* the buffer in the starting sector not covered by the first
* write below.
*/
balign = round_down(start_block, sectbb);
if (balign != start_block) {
error = xlog_bread_noalign(log, start_block, 1, bp);
if (error)
goto out_put_bp;
j = start_block - balign;
}
for (i = start_block; i < end_block; i += bufblks) {
int bcount, endcount;
bcount = min(bufblks, end_block - start_block);
endcount = bcount - j;
/* We may need to do a read at the end to fill in part of
* the buffer in the final sector not covered by the write.
* If this is the same sector as the above read, skip it.
*/
ealign = round_down(end_block, sectbb);
if (j == 0 && (start_block + endcount > ealign)) {
offset = bp->b_addr + BBTOB(ealign - start_block);
error = xlog_bread_offset(log, ealign, sectbb,
bp, offset);
if (error)
break;
}
offset = xlog_align(log, start_block, endcount, bp);
for (; j < endcount; j++) {
xlog_add_record(log, offset, cycle, i+j,
tail_cycle, tail_block);
offset += BBSIZE;
}
error = xlog_bwrite(log, start_block, endcount, bp);
if (error)
break;
start_block += endcount;
j = 0;
}
out_put_bp:
xlog_put_bp(bp);
return error;
}
/*
* This routine is called to blow away any incomplete log writes out
* in front of the log head. We do this so that we won't become confused
* if we come up, write only a little bit more, and then crash again.
* If we leave the partial log records out there, this situation could
* cause us to think those partial writes are valid blocks since they
* have the current cycle number. We get rid of them by overwriting them
* with empty log records with the old cycle number rather than the
* current one.
*
* The tail lsn is passed in rather than taken from
* the log so that we will not write over the unmount record after a
* clean unmount in a 512 block log. Doing so would leave the log without
* any valid log records in it until a new one was written. If we crashed
* during that time we would not be able to recover.
*/
STATIC int
xlog_clear_stale_blocks(
struct xlog *log,
xfs_lsn_t tail_lsn)
{
int tail_cycle, head_cycle;
int tail_block, head_block;
int tail_distance, max_distance;
int distance;
int error;
tail_cycle = CYCLE_LSN(tail_lsn);
tail_block = BLOCK_LSN(tail_lsn);
head_cycle = log->l_curr_cycle;
head_block = log->l_curr_block;
/*
* Figure out the distance between the new head of the log
* and the tail. We want to write over any blocks beyond the
* head that we may have written just before the crash, but
* we don't want to overwrite the tail of the log.
*/
if (head_cycle == tail_cycle) {
/*
* The tail is behind the head in the physical log,
* so the distance from the head to the tail is the
* distance from the head to the end of the log plus
* the distance from the beginning of the log to the
* tail.
*/
if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
XFS_ERRLEVEL_LOW, log->l_mp);
return XFS_ERROR(EFSCORRUPTED);
}
tail_distance = tail_block + (log->l_logBBsize - head_block);
} else {
/*
* The head is behind the tail in the physical log,
* so the distance from the head to the tail is just
* the tail block minus the head block.
*/
if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
XFS_ERRLEVEL_LOW, log->l_mp);
return XFS_ERROR(EFSCORRUPTED);
}
tail_distance = tail_block - head_block;
}
/*
* If the head is right up against the tail, we can't clear
* anything.
*/
if (tail_distance <= 0) {
ASSERT(tail_distance == 0);
return 0;
}
max_distance = XLOG_TOTAL_REC_SHIFT(log);
/*
* Take the smaller of the maximum amount of outstanding I/O
* we could have and the distance to the tail to clear out.
* We take the smaller so that we don't overwrite the tail and
* we don't waste all day writing from the head to the tail
* for no reason.
*/
max_distance = MIN(max_distance, tail_distance);
if ((head_block + max_distance) <= log->l_logBBsize) {
/*
* We can stomp all the blocks we need to without
* wrapping around the end of the log. Just do it
* in a single write. Use the cycle number of the
* current cycle minus one so that the log will look like:
* n ... | n - 1 ...
*/
error = xlog_write_log_records(log, (head_cycle - 1),
head_block, max_distance, tail_cycle,
tail_block);
if (error)
return error;
} else {
/*
* We need to wrap around the end of the physical log in
* order to clear all the blocks. Do it in two separate
* I/Os. The first write should be from the head to the
* end of the physical log, and it should use the current
* cycle number minus one just like above.
*/
distance = log->l_logBBsize - head_block;
error = xlog_write_log_records(log, (head_cycle - 1),
head_block, distance, tail_cycle,
tail_block);
if (error)
return error;
/*
* Now write the blocks at the start of the physical log.
* This writes the remainder of the blocks we want to clear.
* It uses the current cycle number since we're now on the
* same cycle as the head so that we get:
* n ... n ... | n - 1 ...
* ^^^^^ blocks we're writing
*/
distance = max_distance - (log->l_logBBsize - head_block);
error = xlog_write_log_records(log, head_cycle, 0, distance,
tail_cycle, tail_block);
if (error)
return error;
}
return 0;
}
/******************************************************************************
*
* Log recover routines
*
******************************************************************************
*/
STATIC xlog_recover_t *
xlog_recover_find_tid(
struct hlist_head *head,
xlog_tid_t tid)
{
xlog_recover_t *trans;
hlist: drop the node parameter from iterators I'm not sure why, but the hlist for each entry iterators were conceived list_for_each_entry(pos, head, member) The hlist ones were greedy and wanted an extra parameter: hlist_for_each_entry(tpos, pos, head, member) Why did they need an extra pos parameter? I'm not quite sure. Not only they don't really need it, it also prevents the iterator from looking exactly like the list iterator, which is unfortunate. Besides the semantic patch, there was some manual work required: - Fix up the actual hlist iterators in linux/list.h - Fix up the declaration of other iterators based on the hlist ones. - A very small amount of places were using the 'node' parameter, this was modified to use 'obj->member' instead. - Coccinelle didn't handle the hlist_for_each_entry_safe iterator properly, so those had to be fixed up manually. The semantic patch which is mostly the work of Peter Senna Tschudin is here: @@ iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host; type T; expression a,c,d,e; identifier b; statement S; @@ -T b; <+... when != b ( hlist_for_each_entry(a, - b, c, d) S | hlist_for_each_entry_continue(a, - b, c) S | hlist_for_each_entry_from(a, - b, c) S | hlist_for_each_entry_rcu(a, - b, c, d) S | hlist_for_each_entry_rcu_bh(a, - b, c, d) S | hlist_for_each_entry_continue_rcu_bh(a, - b, c) S | for_each_busy_worker(a, c, - b, d) S | ax25_uid_for_each(a, - b, c) S | ax25_for_each(a, - b, c) S | inet_bind_bucket_for_each(a, - b, c) S | sctp_for_each_hentry(a, - b, c) S | sk_for_each(a, - b, c) S | sk_for_each_rcu(a, - b, c) S | sk_for_each_from -(a, b) +(a) S + sk_for_each_from(a) S | sk_for_each_safe(a, - b, c, d) S | sk_for_each_bound(a, - b, c) S | hlist_for_each_entry_safe(a, - b, c, d, e) S | hlist_for_each_entry_continue_rcu(a, - b, c) S | nr_neigh_for_each(a, - b, c) S | nr_neigh_for_each_safe(a, - b, c, d) S | nr_node_for_each(a, - b, c) S | nr_node_for_each_safe(a, - b, c, d) S | - for_each_gfn_sp(a, c, d, b) S + for_each_gfn_sp(a, c, d) S | - for_each_gfn_indirect_valid_sp(a, c, d, b) S + for_each_gfn_indirect_valid_sp(a, c, d) S | for_each_host(a, - b, c) S | for_each_host_safe(a, - b, c, d) S | for_each_mesh_entry(a, - b, c, d) S ) ...+> [akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c] [akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c] [akpm@linux-foundation.org: checkpatch fixes] [akpm@linux-foundation.org: fix warnings] [akpm@linux-foudnation.org: redo intrusive kvm changes] Tested-by: Peter Senna Tschudin <peter.senna@gmail.com> Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Sasha Levin <sasha.levin@oracle.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Gleb Natapov <gleb@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-28 01:06:00 +00:00
hlist_for_each_entry(trans, head, r_list) {
if (trans->r_log_tid == tid)
return trans;
}
return NULL;
}
STATIC void
xlog_recover_new_tid(
struct hlist_head *head,
xlog_tid_t tid,
xfs_lsn_t lsn)
{
xlog_recover_t *trans;
trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
trans->r_log_tid = tid;
trans->r_lsn = lsn;
INIT_LIST_HEAD(&trans->r_itemq);
INIT_HLIST_NODE(&trans->r_list);
hlist_add_head(&trans->r_list, head);
}
STATIC void
xlog_recover_add_item(
struct list_head *head)
{
xlog_recover_item_t *item;
item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
INIT_LIST_HEAD(&item->ri_list);
list_add_tail(&item->ri_list, head);
}
STATIC int
xlog_recover_add_to_cont_trans(
struct xlog *log,
struct xlog_recover *trans,
xfs_caddr_t dp,
int len)
{
xlog_recover_item_t *item;
xfs_caddr_t ptr, old_ptr;
int old_len;
if (list_empty(&trans->r_itemq)) {
/* finish copying rest of trans header */
xlog_recover_add_item(&trans->r_itemq);
ptr = (xfs_caddr_t) &trans->r_theader +
sizeof(xfs_trans_header_t) - len;
memcpy(ptr, dp, len); /* d, s, l */
return 0;
}
/* take the tail entry */
item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
old_len = item->ri_buf[item->ri_cnt-1].i_len;
ptr = kmem_realloc(old_ptr, len+old_len, old_len, KM_SLEEP);
memcpy(&ptr[old_len], dp, len); /* d, s, l */
item->ri_buf[item->ri_cnt-1].i_len += len;
item->ri_buf[item->ri_cnt-1].i_addr = ptr;
trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
return 0;
}
/*
* The next region to add is the start of a new region. It could be
* a whole region or it could be the first part of a new region. Because
* of this, the assumption here is that the type and size fields of all
* format structures fit into the first 32 bits of the structure.
*
* This works because all regions must be 32 bit aligned. Therefore, we
* either have both fields or we have neither field. In the case we have
* neither field, the data part of the region is zero length. We only have
* a log_op_header and can throw away the header since a new one will appear
* later. If we have at least 4 bytes, then we can determine how many regions
* will appear in the current log item.
*/
STATIC int
xlog_recover_add_to_trans(
struct xlog *log,
struct xlog_recover *trans,
xfs_caddr_t dp,
int len)
{
xfs_inode_log_format_t *in_f; /* any will do */
xlog_recover_item_t *item;
xfs_caddr_t ptr;
if (!len)
return 0;
if (list_empty(&trans->r_itemq)) {
/* we need to catch log corruptions here */
if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
xfs_warn(log->l_mp, "%s: bad header magic number",
__func__);
ASSERT(0);
return XFS_ERROR(EIO);
}
if (len == sizeof(xfs_trans_header_t))
xlog_recover_add_item(&trans->r_itemq);
memcpy(&trans->r_theader, dp, len); /* d, s, l */
return 0;
}
ptr = kmem_alloc(len, KM_SLEEP);
memcpy(ptr, dp, len);
in_f = (xfs_inode_log_format_t *)ptr;
/* take the tail entry */
item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
if (item->ri_total != 0 &&
item->ri_total == item->ri_cnt) {
/* tail item is in use, get a new one */
xlog_recover_add_item(&trans->r_itemq);
item = list_entry(trans->r_itemq.prev,
xlog_recover_item_t, ri_list);
}
if (item->ri_total == 0) { /* first region to be added */
if (in_f->ilf_size == 0 ||
in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
xfs_warn(log->l_mp,
"bad number of regions (%d) in inode log format",
in_f->ilf_size);
ASSERT(0);
kmem_free(ptr);
return XFS_ERROR(EIO);
}
item->ri_total = in_f->ilf_size;
item->ri_buf =
kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
KM_SLEEP);
}
ASSERT(item->ri_total > item->ri_cnt);
/* Description region is ri_buf[0] */
item->ri_buf[item->ri_cnt].i_addr = ptr;
item->ri_buf[item->ri_cnt].i_len = len;
item->ri_cnt++;
trace_xfs_log_recover_item_add(log, trans, item, 0);
return 0;
}
/*
* Sort the log items in the transaction.
*
* The ordering constraints are defined by the inode allocation and unlink
* behaviour. The rules are:
*
* 1. Every item is only logged once in a given transaction. Hence it
* represents the last logged state of the item. Hence ordering is
* dependent on the order in which operations need to be performed so
* required initial conditions are always met.
*
* 2. Cancelled buffers are recorded in pass 1 in a separate table and
* there's nothing to replay from them so we can simply cull them
* from the transaction. However, we can't do that until after we've
* replayed all the other items because they may be dependent on the
* cancelled buffer and replaying the cancelled buffer can remove it
* form the cancelled buffer table. Hence they have tobe done last.
*
* 3. Inode allocation buffers must be replayed before inode items that
* read the buffer and replay changes into it. For filesystems using the
* ICREATE transactions, this means XFS_LI_ICREATE objects need to get
* treated the same as inode allocation buffers as they create and
* initialise the buffers directly.
*
* 4. Inode unlink buffers must be replayed after inode items are replayed.
* This ensures that inodes are completely flushed to the inode buffer
* in a "free" state before we remove the unlinked inode list pointer.
*
* Hence the ordering needs to be inode allocation buffers first, inode items
* second, inode unlink buffers third and cancelled buffers last.
*
* But there's a problem with that - we can't tell an inode allocation buffer
* apart from a regular buffer, so we can't separate them. We can, however,
* tell an inode unlink buffer from the others, and so we can separate them out
* from all the other buffers and move them to last.
*
* Hence, 4 lists, in order from head to tail:
* - buffer_list for all buffers except cancelled/inode unlink buffers
* - item_list for all non-buffer items
* - inode_buffer_list for inode unlink buffers
* - cancel_list for the cancelled buffers
*
* Note that we add objects to the tail of the lists so that first-to-last
* ordering is preserved within the lists. Adding objects to the head of the
* list means when we traverse from the head we walk them in last-to-first
* order. For cancelled buffers and inode unlink buffers this doesn't matter,
* but for all other items there may be specific ordering that we need to
* preserve.
*/
STATIC int
xlog_recover_reorder_trans(
struct xlog *log,
struct xlog_recover *trans,
int pass)
{
xlog_recover_item_t *item, *n;
int error = 0;
LIST_HEAD(sort_list);
LIST_HEAD(cancel_list);
LIST_HEAD(buffer_list);
LIST_HEAD(inode_buffer_list);
LIST_HEAD(inode_list);
list_splice_init(&trans->r_itemq, &sort_list);
list_for_each_entry_safe(item, n, &sort_list, ri_list) {
xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
switch (ITEM_TYPE(item)) {
case XFS_LI_ICREATE:
list_move_tail(&item->ri_list, &buffer_list);
break;
case XFS_LI_BUF:
if (buf_f->blf_flags & XFS_BLF_CANCEL) {
trace_xfs_log_recover_item_reorder_head(log,
trans, item, pass);
list_move(&item->ri_list, &cancel_list);
break;
}
if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
list_move(&item->ri_list, &inode_buffer_list);
break;
}
list_move_tail(&item->ri_list, &buffer_list);
break;
case XFS_LI_INODE:
case XFS_LI_DQUOT:
case XFS_LI_QUOTAOFF:
case XFS_LI_EFD:
case XFS_LI_EFI:
trace_xfs_log_recover_item_reorder_tail(log,
trans, item, pass);
list_move_tail(&item->ri_list, &inode_list);
break;
default:
xfs_warn(log->l_mp,
"%s: unrecognized type of log operation",
__func__);
ASSERT(0);
/*
* return the remaining items back to the transaction
* item list so they can be freed in caller.
*/
if (!list_empty(&sort_list))
list_splice_init(&sort_list, &trans->r_itemq);
error = XFS_ERROR(EIO);
goto out;
}
}
out:
ASSERT(list_empty(&sort_list));
if (!list_empty(&buffer_list))
list_splice(&buffer_list, &trans->r_itemq);
if (!list_empty(&inode_list))
list_splice_tail(&inode_list, &trans->r_itemq);
if (!list_empty(&inode_buffer_list))
list_splice_tail(&inode_buffer_list, &trans->r_itemq);
if (!list_empty(&cancel_list))
list_splice_tail(&cancel_list, &trans->r_itemq);
return error;
}
/*
* Build up the table of buf cancel records so that we don't replay
* cancelled data in the second pass. For buffer records that are
* not cancel records, there is nothing to do here so we just return.
*
* If we get a cancel record which is already in the table, this indicates
* that the buffer was cancelled multiple times. In order to ensure
* that during pass 2 we keep the record in the table until we reach its
* last occurrence in the log, we keep a reference count in the cancel
* record in the table to tell us how many times we expect to see this
* record during the second pass.
*/
STATIC int
xlog_recover_buffer_pass1(
struct xlog *log,
struct xlog_recover_item *item)
{
xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
struct list_head *bucket;
struct xfs_buf_cancel *bcp;
/*
* If this isn't a cancel buffer item, then just return.
*/
if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
trace_xfs_log_recover_buf_not_cancel(log, buf_f);
return 0;
}
/*
* Insert an xfs_buf_cancel record into the hash table of them.
* If there is already an identical record, bump its reference count.
*/
bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
list_for_each_entry(bcp, bucket, bc_list) {
if (bcp->bc_blkno == buf_f->blf_blkno &&
bcp->bc_len == buf_f->blf_len) {
bcp->bc_refcount++;
trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
return 0;
}
}
bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
bcp->bc_blkno = buf_f->blf_blkno;
bcp->bc_len = buf_f->blf_len;
bcp->bc_refcount = 1;
list_add_tail(&bcp->bc_list, bucket);
trace_xfs_log_recover_buf_cancel_add(log, buf_f);
return 0;
}
/*
* Check to see whether the buffer being recovered has a corresponding
* entry in the buffer cancel record table. If it is, return the cancel
* buffer structure to the caller.
*/
STATIC struct xfs_buf_cancel *
xlog_peek_buffer_cancelled(
struct xlog *log,
xfs_daddr_t blkno,
uint len,
ushort flags)
{
struct list_head *bucket;
struct xfs_buf_cancel *bcp;
if (!log->l_buf_cancel_table) {
/* empty table means no cancelled buffers in the log */
ASSERT(!(flags & XFS_BLF_CANCEL));
return NULL;
}
bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
list_for_each_entry(bcp, bucket, bc_list) {
if (bcp->bc_blkno == blkno && bcp->bc_len == len)
return bcp;
}
/*
* We didn't find a corresponding entry in the table, so return 0 so
* that the buffer is NOT cancelled.
*/
ASSERT(!(flags & XFS_BLF_CANCEL));
return NULL;
}
/*
* If the buffer is being cancelled then return 1 so that it will be cancelled,
* otherwise return 0. If the buffer is actually a buffer cancel item
* (XFS_BLF_CANCEL is set), then decrement the refcount on the entry in the
* table and remove it from the table if this is the last reference.
*
* We remove the cancel record from the table when we encounter its last
* occurrence in the log so that if the same buffer is re-used again after its
* last cancellation we actually replay the changes made at that point.
*/
STATIC int
xlog_check_buffer_cancelled(
struct xlog *log,
xfs_daddr_t blkno,
uint len,
ushort flags)
{
struct xfs_buf_cancel *bcp;
bcp = xlog_peek_buffer_cancelled(log, blkno, len, flags);
if (!bcp)
return 0;
/*
* We've go a match, so return 1 so that the recovery of this buffer
* is cancelled. If this buffer is actually a buffer cancel log
* item, then decrement the refcount on the one in the table and
* remove it if this is the last reference.
*/
if (flags & XFS_BLF_CANCEL) {
if (--bcp->bc_refcount == 0) {
list_del(&bcp->bc_list);
kmem_free(bcp);
}
}
return 1;
}
/*
* Perform recovery for a buffer full of inodes. In these buffers, the only
* data which should be recovered is that which corresponds to the
* di_next_unlinked pointers in the on disk inode structures. The rest of the
* data for the inodes is always logged through the inodes themselves rather
* than the inode buffer and is recovered in xlog_recover_inode_pass2().
*
* The only time when buffers full of inodes are fully recovered is when the
* buffer is full of newly allocated inodes. In this case the buffer will
* not be marked as an inode buffer and so will be sent to
* xlog_recover_do_reg_buffer() below during recovery.
*/
STATIC int
xlog_recover_do_inode_buffer(
struct xfs_mount *mp,
xlog_recover_item_t *item,
struct xfs_buf *bp,
xfs_buf_log_format_t *buf_f)
{
int i;
int item_index = 0;
int bit = 0;
int nbits = 0;
int reg_buf_offset = 0;
int reg_buf_bytes = 0;
int next_unlinked_offset;
int inodes_per_buf;
xfs_agino_t *logged_nextp;
xfs_agino_t *buffer_nextp;
trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
/*
* Post recovery validation only works properly on CRC enabled
* filesystems.
*/
if (xfs_sb_version_hascrc(&mp->m_sb))
bp->b_ops = &xfs_inode_buf_ops;
inodes_per_buf = BBTOB(bp->b_io_length) >> mp->m_sb.sb_inodelog;
for (i = 0; i < inodes_per_buf; i++) {
next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
offsetof(xfs_dinode_t, di_next_unlinked);
while (next_unlinked_offset >=
(reg_buf_offset + reg_buf_bytes)) {
/*
* The next di_next_unlinked field is beyond
* the current logged region. Find the next
* logged region that contains or is beyond
* the current di_next_unlinked field.
*/
bit += nbits;
bit = xfs_next_bit(buf_f->blf_data_map,
buf_f->blf_map_size, bit);
/*
* If there are no more logged regions in the
* buffer, then we're done.
*/
if (bit == -1)
return 0;
nbits = xfs_contig_bits(buf_f->blf_data_map,
buf_f->blf_map_size, bit);
ASSERT(nbits > 0);
reg_buf_offset = bit << XFS_BLF_SHIFT;
reg_buf_bytes = nbits << XFS_BLF_SHIFT;
item_index++;
}
/*
* If the current logged region starts after the current
* di_next_unlinked field, then move on to the next
* di_next_unlinked field.
*/
if (next_unlinked_offset < reg_buf_offset)
continue;
ASSERT(item->ri_buf[item_index].i_addr != NULL);
ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
ASSERT((reg_buf_offset + reg_buf_bytes) <=
BBTOB(bp->b_io_length));
/*
* The current logged region contains a copy of the
* current di_next_unlinked field. Extract its value
* and copy it to the buffer copy.
*/
logged_nextp = item->ri_buf[item_index].i_addr +
next_unlinked_offset - reg_buf_offset;
if (unlikely(*logged_nextp == 0)) {
xfs_alert(mp,
"Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
"Trying to replay bad (0) inode di_next_unlinked field.",
item, bp);
XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
XFS_ERRLEVEL_LOW, mp);
return XFS_ERROR(EFSCORRUPTED);
}
buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
next_unlinked_offset);
*buffer_nextp = *logged_nextp;
/*
* If necessary, recalculate the CRC in the on-disk inode. We
* have to leave the inode in a consistent state for whoever
* reads it next....
*/
xfs_dinode_calc_crc(mp, (struct xfs_dinode *)
xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
}
return 0;
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
/*
* V5 filesystems know the age of the buffer on disk being recovered. We can
* have newer objects on disk than we are replaying, and so for these cases we
* don't want to replay the current change as that will make the buffer contents
* temporarily invalid on disk.
*
* The magic number might not match the buffer type we are going to recover
* (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags. Hence
* extract the LSN of the existing object in the buffer based on it's current
* magic number. If we don't recognise the magic number in the buffer, then
* return a LSN of -1 so that the caller knows it was an unrecognised block and
* so can recover the buffer.
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
*
* Note: we cannot rely solely on magic number matches to determine that the
* buffer has a valid LSN - we also need to verify that it belongs to this
* filesystem, so we need to extract the object's LSN and compare it to that
* which we read from the superblock. If the UUIDs don't match, then we've got a
* stale metadata block from an old filesystem instance that we need to recover
* over the top of.
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
*/
static xfs_lsn_t
xlog_recover_get_buf_lsn(
struct xfs_mount *mp,
struct xfs_buf *bp)
{
__uint32_t magic32;
__uint16_t magic16;
__uint16_t magicda;
void *blk = bp->b_addr;
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
uuid_t *uuid;
xfs_lsn_t lsn = -1;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
/* v4 filesystems always recover immediately */
if (!xfs_sb_version_hascrc(&mp->m_sb))
goto recover_immediately;
magic32 = be32_to_cpu(*(__be32 *)blk);
switch (magic32) {
case XFS_ABTB_CRC_MAGIC:
case XFS_ABTC_CRC_MAGIC:
case XFS_ABTB_MAGIC:
case XFS_ABTC_MAGIC:
case XFS_IBT_CRC_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
case XFS_IBT_MAGIC: {
struct xfs_btree_block *btb = blk;
lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
uuid = &btb->bb_u.s.bb_uuid;
break;
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
case XFS_BMAP_CRC_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
case XFS_BMAP_MAGIC: {
struct xfs_btree_block *btb = blk;
lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
uuid = &btb->bb_u.l.bb_uuid;
break;
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
case XFS_AGF_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
uuid = &((struct xfs_agf *)blk)->agf_uuid;
break;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
case XFS_AGFL_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
break;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
case XFS_AGI_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
uuid = &((struct xfs_agi *)blk)->agi_uuid;
break;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
case XFS_SYMLINK_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
break;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
case XFS_DIR3_BLOCK_MAGIC:
case XFS_DIR3_DATA_MAGIC:
case XFS_DIR3_FREE_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
break;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
case XFS_ATTR3_RMT_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
lsn = be64_to_cpu(((struct xfs_attr3_rmt_hdr *)blk)->rm_lsn);
uuid = &((struct xfs_attr3_rmt_hdr *)blk)->rm_uuid;
break;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
case XFS_SB_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
uuid = &((struct xfs_dsb *)blk)->sb_uuid;
break;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
default:
break;
}
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
if (lsn != (xfs_lsn_t)-1) {
if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
goto recover_immediately;
return lsn;
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
switch (magicda) {
case XFS_DIR3_LEAF1_MAGIC:
case XFS_DIR3_LEAFN_MAGIC:
case XFS_DA3_NODE_MAGIC:
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
break;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
default:
break;
}
xfs: log recovery lsn ordering needs uuid check After a fair number of xfstests runs, xfs/182 started to fail regularly with a corrupted directory - a directory read verifier was failing after recovery because it found a block with a XARM magic number (remote attribute block) rather than a directory data block. The first time I saw this repeated failure I did /something/ and the problem went away, so I was never able to find the underlying problem. Test xfs/182 failed again today, and I found the root cause before I did /something else/ that made it go away. Tracing indicated that the block in question was being correctly logged, the log was being flushed by sync, but the buffer was not being written back before the shutdown occurred. Tracing also indicated that log recovery was also reading the block, but then never writing it before log recovery invalidated the cache, indicating that it was not modified by log recovery. More detailed analysis of the corpse indicated that the filesystem had a uuid of "a4131074-1872-4cac-9323-2229adbcb886" but the XARM block had a uuid of "8f32f043-c3c9-e7f8-f947-4e7f989c05d3", which indicated it was a block from an older filesystem. The reason that log recovery didn't replay it was that the LSN in the XARM block was larger than the LSN of the transaction being replayed, and so the block was not overwritten by log recovery. Hence, log recovery cant blindly trust the magic number and LSN in the block - it must verify that it belongs to the filesystem being recovered before using the LSN. i.e. if the UUIDs don't match, we need to unconditionally recovery the change held in the log. This patch was first tested on a block device that was repeatedly causing xfs/182 to fail with the same failure on the same block with the same directory read corruption signature (i.e. XARM block). It did not fail, and hasn't failed since. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-09-24 06:01:16 +00:00
if (lsn != (xfs_lsn_t)-1) {
if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
goto recover_immediately;
return lsn;
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
/*
* We do individual object checks on dquot and inode buffers as they
* have their own individual LSN records. Also, we could have a stale
* buffer here, so we have to at least recognise these buffer types.
*
* A notd complexity here is inode unlinked list processing - it logs
* the inode directly in the buffer, but we don't know which inodes have
* been modified, and there is no global buffer LSN. Hence we need to
* recover all inode buffer types immediately. This problem will be
* fixed by logical logging of the unlinked list modifications.
*/
magic16 = be16_to_cpu(*(__be16 *)blk);
switch (magic16) {
case XFS_DQUOT_MAGIC:
case XFS_DINODE_MAGIC:
goto recover_immediately;
default:
break;
}
/* unknown buffer contents, recover immediately */
recover_immediately:
return (xfs_lsn_t)-1;
}
/*
* Validate the recovered buffer is of the correct type and attach the
* appropriate buffer operations to them for writeback. Magic numbers are in a
* few places:
* the first 16 bits of the buffer (inode buffer, dquot buffer),
* the first 32 bits of the buffer (most blocks),
* inside a struct xfs_da_blkinfo at the start of the buffer.
*/
static void
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
xlog_recover_validate_buf_type(
struct xfs_mount *mp,
struct xfs_buf *bp,
xfs_buf_log_format_t *buf_f)
{
struct xfs_da_blkinfo *info = bp->b_addr;
__uint32_t magic32;
__uint16_t magic16;
__uint16_t magicda;
magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
magicda = be16_to_cpu(info->magic);
switch (xfs_blft_from_flags(buf_f)) {
case XFS_BLFT_BTREE_BUF:
switch (magic32) {
case XFS_ABTB_CRC_MAGIC:
case XFS_ABTC_CRC_MAGIC:
case XFS_ABTB_MAGIC:
case XFS_ABTC_MAGIC:
bp->b_ops = &xfs_allocbt_buf_ops;
break;
case XFS_IBT_CRC_MAGIC:
case XFS_FIBT_CRC_MAGIC:
case XFS_IBT_MAGIC:
case XFS_FIBT_MAGIC:
bp->b_ops = &xfs_inobt_buf_ops;
break;
case XFS_BMAP_CRC_MAGIC:
case XFS_BMAP_MAGIC:
bp->b_ops = &xfs_bmbt_buf_ops;
break;
default:
xfs_warn(mp, "Bad btree block magic!");
ASSERT(0);
break;
}
break;
case XFS_BLFT_AGF_BUF:
if (magic32 != XFS_AGF_MAGIC) {
xfs_warn(mp, "Bad AGF block magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_agf_buf_ops;
break;
case XFS_BLFT_AGFL_BUF:
if (!xfs_sb_version_hascrc(&mp->m_sb))
break;
if (magic32 != XFS_AGFL_MAGIC) {
xfs_warn(mp, "Bad AGFL block magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_agfl_buf_ops;
break;
case XFS_BLFT_AGI_BUF:
if (magic32 != XFS_AGI_MAGIC) {
xfs_warn(mp, "Bad AGI block magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_agi_buf_ops;
break;
case XFS_BLFT_UDQUOT_BUF:
case XFS_BLFT_PDQUOT_BUF:
case XFS_BLFT_GDQUOT_BUF:
#ifdef CONFIG_XFS_QUOTA
if (magic16 != XFS_DQUOT_MAGIC) {
xfs_warn(mp, "Bad DQUOT block magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_dquot_buf_ops;
#else
xfs_alert(mp,
"Trying to recover dquots without QUOTA support built in!");
ASSERT(0);
#endif
break;
case XFS_BLFT_DINO_BUF:
/*
* we get here with inode allocation buffers, not buffers that
* track unlinked list changes.
*/
if (magic16 != XFS_DINODE_MAGIC) {
xfs_warn(mp, "Bad INODE block magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_inode_buf_ops;
break;
case XFS_BLFT_SYMLINK_BUF:
if (magic32 != XFS_SYMLINK_MAGIC) {
xfs_warn(mp, "Bad symlink block magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_symlink_buf_ops;
break;
case XFS_BLFT_DIR_BLOCK_BUF:
if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
magic32 != XFS_DIR3_BLOCK_MAGIC) {
xfs_warn(mp, "Bad dir block magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_dir3_block_buf_ops;
break;
case XFS_BLFT_DIR_DATA_BUF:
if (magic32 != XFS_DIR2_DATA_MAGIC &&
magic32 != XFS_DIR3_DATA_MAGIC) {
xfs_warn(mp, "Bad dir data magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_dir3_data_buf_ops;
break;
case XFS_BLFT_DIR_FREE_BUF:
if (magic32 != XFS_DIR2_FREE_MAGIC &&
magic32 != XFS_DIR3_FREE_MAGIC) {
xfs_warn(mp, "Bad dir3 free magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_dir3_free_buf_ops;
break;
case XFS_BLFT_DIR_LEAF1_BUF:
if (magicda != XFS_DIR2_LEAF1_MAGIC &&
magicda != XFS_DIR3_LEAF1_MAGIC) {
xfs_warn(mp, "Bad dir leaf1 magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_dir3_leaf1_buf_ops;
break;
case XFS_BLFT_DIR_LEAFN_BUF:
if (magicda != XFS_DIR2_LEAFN_MAGIC &&
magicda != XFS_DIR3_LEAFN_MAGIC) {
xfs_warn(mp, "Bad dir leafn magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_dir3_leafn_buf_ops;
break;
case XFS_BLFT_DA_NODE_BUF:
if (magicda != XFS_DA_NODE_MAGIC &&
magicda != XFS_DA3_NODE_MAGIC) {
xfs_warn(mp, "Bad da node magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_da3_node_buf_ops;
break;
case XFS_BLFT_ATTR_LEAF_BUF:
if (magicda != XFS_ATTR_LEAF_MAGIC &&
magicda != XFS_ATTR3_LEAF_MAGIC) {
xfs_warn(mp, "Bad attr leaf magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_attr3_leaf_buf_ops;
break;
case XFS_BLFT_ATTR_RMT_BUF:
if (!xfs_sb_version_hascrc(&mp->m_sb))
break;
if (magic32 != XFS_ATTR3_RMT_MAGIC) {
xfs_warn(mp, "Bad attr remote magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_attr3_rmt_buf_ops;
break;
case XFS_BLFT_SB_BUF:
if (magic32 != XFS_SB_MAGIC) {
xfs_warn(mp, "Bad SB block magic!");
ASSERT(0);
break;
}
bp->b_ops = &xfs_sb_buf_ops;
break;
default:
xfs_warn(mp, "Unknown buffer type %d!",
xfs_blft_from_flags(buf_f));
break;
}
}
/*
* Perform a 'normal' buffer recovery. Each logged region of the
* buffer should be copied over the corresponding region in the
* given buffer. The bitmap in the buf log format structure indicates
* where to place the logged data.
*/
STATIC void
xlog_recover_do_reg_buffer(
struct xfs_mount *mp,
xlog_recover_item_t *item,
struct xfs_buf *bp,
xfs_buf_log_format_t *buf_f)
{
int i;
int bit;
int nbits;
int error;
trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
bit = 0;
i = 1; /* 0 is the buf format structure */
while (1) {
bit = xfs_next_bit(buf_f->blf_data_map,
buf_f->blf_map_size, bit);
if (bit == -1)
break;
nbits = xfs_contig_bits(buf_f->blf_data_map,
buf_f->blf_map_size, bit);
ASSERT(nbits > 0);
ASSERT(item->ri_buf[i].i_addr != NULL);
ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
ASSERT(BBTOB(bp->b_io_length) >=
((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
xfs: fix split buffer vector log recovery support A long time ago in a galaxy far away.... .. the was a commit made to fix some ilinux specific "fragmented buffer" log recovery problem: http://oss.sgi.com/cgi-bin/gitweb.cgi?p=archive/xfs-import.git;a=commitdiff;h=b29c0bece51da72fb3ff3b61391a391ea54e1603 That problem occurred when a contiguous dirty region of a buffer was split across across two pages of an unmapped buffer. It's been a long time since that has been done in XFS, and the changes to log the entire inode buffers for CRC enabled filesystems has re-introduced that corner case. And, of course, it turns out that the above commit didn't actually fix anything - it just ensured that log recovery is guaranteed to fail when this situation occurs. And now for the gory details. xfstest xfs/085 is failing with this assert: XFS (vdb): bad number of regions (0) in inode log format XFS: Assertion failed: 0, file: fs/xfs/xfs_log_recover.c, line: 1583 Largely undocumented factoid #1: Log recovery depends on all log buffer format items starting with this format: struct foo_log_format { __uint16_t type; __uint16_t size; .... As recoery uses the size field and assumptions about 32 bit alignment in decoding format items. So don't pay much attention to the fact log recovery thinks that it decoding an inode log format item - it just uses them to determine what the size of the item is. But why would it see a log format item with a zero size? Well, luckily enough xfs_logprint uses the same code and gives the same error, so with a bit of gdb magic, it turns out that it isn't a log format that is being decoded. What logprint tells us is this: Oper (130): tid: a0375e1a len: 28 clientid: TRANS flags: none BUF: #regs: 2 start blkno: 144 (0x90) len: 16 bmap size: 2 flags: 0x4000 Oper (131): tid: a0375e1a len: 4096 clientid: TRANS flags: none BUF DATA ---------------------------------------------------------------------------- Oper (132): tid: a0375e1a len: 4096 clientid: TRANS flags: none xfs_logprint: unknown log operation type (4e49) ********************************************************************** * ERROR: data block=2 * ********************************************************************** That we've got a buffer format item (oper 130) that has two regions; the format item itself and one dirty region. The subsequent region after the buffer format item and it's data is them what we are tripping over, and the first bytes of it at an inode magic number. Not a log opheader like there is supposed to be. That means there's a problem with the buffer format item. It's dirty data region is 4096 bytes, and it contains - you guessed it - initialised inodes. But inode buffers are 8k, not 4k, and we log them in their entirety. So something is wrong here. The buffer format item contains: (gdb) p /x *(struct xfs_buf_log_format *)in_f $22 = {blf_type = 0x123c, blf_size = 0x2, blf_flags = 0x4000, blf_len = 0x10, blf_blkno = 0x90, blf_map_size = 0x2, blf_data_map = {0xffffffff, 0xffffffff, .... }} Two regions, and a signle dirty contiguous region of 64 bits. 64 * 128 = 8k, so this should be followed by a single 8k region of data. And the blf_flags tell us that the type of buffer is a XFS_BLFT_DINO_BUF. It contains inodes. And because it doesn't have the XFS_BLF_INODE_BUF flag set, that means it's an inode allocation buffer. So, it should be followed by 8k of inode data. But we know that the next region has a header of: (gdb) p /x *ohead $25 = {oh_tid = 0x1a5e37a0, oh_len = 0x100000, oh_clientid = 0x69, oh_flags = 0x0, oh_res2 = 0x0} and so be32_to_cpu(oh_len) = 0x1000 = 4096 bytes. It's simply not long enough to hold all the logged data. There must be another region. There is - there's a following opheader for another 4k of data that contains the other half of the inode cluster data - the one we assert fail on because it's not a log format header. So why is the second part of the data not being accounted to the correct buffer log format structure? It took a little more work with gdb to work out that the buffer log format structure was both expecting it to be there but hadn't accounted for it. It was at that point I went to the kernel code, as clearly this wasn't a bug in xfs_logprint and the kernel was writing bad stuff to the log. First port of call was the buffer item formatting code, and the discontiguous memory/contiguous dirty region handling code immediately stood out. I've wondered for a long time why the code had this comment in it: vecp->i_addr = xfs_buf_offset(bp, buffer_offset); vecp->i_len = nbits * XFS_BLF_CHUNK; vecp->i_type = XLOG_REG_TYPE_BCHUNK; /* * You would think we need to bump the nvecs here too, but we do not * this number is used by recovery, and it gets confused by the boundary * split here * nvecs++; */ vecp++; And it didn't account for the extra vector pointer. The case being handled here is that a contiguous dirty region lies across a boundary that cannot be memcpy()d across, and so has to be split into two separate operations for xlog_write() to perform. What this code assumes is that what is written to the log is two consecutive blocks of data that are accounted in the buf log format item as the same contiguous dirty region and so will get decoded as such by the log recovery code. The thing is, xlog_write() knows nothing about this, and so just does it's normal thing of adding an opheader for each vector. That means the 8k region gets written to the log as two separate regions of 4k each, but because nvecs has not been incremented, the buf log format item accounts for only one of them. Hence when we come to log recovery, we process the first 4k region and then expect to come across a new item that starts with a log format structure of some kind that tells us whenteh next data is going to be. Instead, we hit raw buffer data and things go bad real quick. So, the commit from 2002 that commented out nvecs++ is just plain wrong. It breaks log recovery completely, and it would seem the only reason this hasn't been since then is that we don't log large contigous regions of multi-page unmapped buffers very often. Never would be a closer estimate, at least until the CRC code came along.... So, lets fix that by restoring the nvecs accounting for the extra region when we hit this case..... .... and there's the problemin log recovery it is apparently working around: XFS: Assertion failed: i == item->ri_total, file: fs/xfs/xfs_log_recover.c, line: 2135 Yup, xlog_recover_do_reg_buffer() doesn't handle contigous dirty regions being broken up into multiple regions by the log formatting code. That's an easy fix, though - if the number of contiguous dirty bits exceeds the length of the region being copied out of the log, only account for the number of dirty bits that region covers, and then loop again and copy more from the next region. It's a 2 line fix. Now xfstests xfs/085 passes, we have one less piece of mystery code, and one more important piece of knowledge about how to structure new log format items.. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-05-27 06:38:23 +00:00
/*
* The dirty regions logged in the buffer, even though
* contiguous, may span multiple chunks. This is because the
* dirty region may span a physical page boundary in a buffer
* and hence be split into two separate vectors for writing into
* the log. Hence we need to trim nbits back to the length of
* the current region being copied out of the log.
*/
if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
/*
* Do a sanity check if this is a dquot buffer. Just checking
* the first dquot in the buffer should do. XXXThis is
* probably a good thing to do for other buf types also.
*/
error = 0;
if (buf_f->blf_flags &
(XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
if (item->ri_buf[i].i_addr == NULL) {
xfs_alert(mp,
"XFS: NULL dquot in %s.", __func__);
goto next;
}
if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
xfs_alert(mp,
"XFS: dquot too small (%d) in %s.",
item->ri_buf[i].i_len, __func__);
goto next;
}
error = xfs_dqcheck(mp, item->ri_buf[i].i_addr,
-1, 0, XFS_QMOPT_DOWARN,
"dquot_buf_recover");
if (error)
goto next;
}
memcpy(xfs_buf_offset(bp,
(uint)bit << XFS_BLF_SHIFT), /* dest */
item->ri_buf[i].i_addr, /* source */
nbits<<XFS_BLF_SHIFT); /* length */
next:
i++;
bit += nbits;
}
/* Shouldn't be any more regions */
ASSERT(i == item->ri_total);
/*
* We can only do post recovery validation on items on CRC enabled
* fielsystems as we need to know when the buffer was written to be able
* to determine if we should have replayed the item. If we replay old
* metadata over a newer buffer, then it will enter a temporarily
* inconsistent state resulting in verification failures. Hence for now
* just avoid the verification stage for non-crc filesystems
*/
if (xfs_sb_version_hascrc(&mp->m_sb))
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
xlog_recover_validate_buf_type(mp, bp, buf_f);
}
/*
* Perform a dquot buffer recovery.
* Simple algorithm: if we have found a QUOTAOFF log item of the same type
* (ie. USR or GRP), then just toss this buffer away; don't recover it.
* Else, treat it as a regular buffer and do recovery.
*/
STATIC void
xlog_recover_do_dquot_buffer(
struct xfs_mount *mp,
struct xlog *log,
struct xlog_recover_item *item,
struct xfs_buf *bp,
struct xfs_buf_log_format *buf_f)
{
uint type;
trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
/*
* Filesystems are required to send in quota flags at mount time.
*/
if (mp->m_qflags == 0) {
return;
}
type = 0;
if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
type |= XFS_DQ_USER;
if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
type |= XFS_DQ_PROJ;
if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
type |= XFS_DQ_GROUP;
/*
* This type of quotas was turned off, so ignore this buffer
*/
if (log->l_quotaoffs_flag & type)
return;
xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
}
/*
* This routine replays a modification made to a buffer at runtime.
* There are actually two types of buffer, regular and inode, which
* are handled differently. Inode buffers are handled differently
* in that we only recover a specific set of data from them, namely
* the inode di_next_unlinked fields. This is because all other inode
* data is actually logged via inode records and any data we replay
* here which overlaps that may be stale.
*
* When meta-data buffers are freed at run time we log a buffer item
* with the XFS_BLF_CANCEL bit set to indicate that previous copies
* of the buffer in the log should not be replayed at recovery time.
* This is so that if the blocks covered by the buffer are reused for
* file data before we crash we don't end up replaying old, freed
* meta-data into a user's file.
*
* To handle the cancellation of buffer log items, we make two passes
* over the log during recovery. During the first we build a table of
* those buffers which have been cancelled, and during the second we
* only replay those buffers which do not have corresponding cancel
* records in the table. See xlog_recover_buffer_pass[1,2] above
* for more details on the implementation of the table of cancel records.
*/
STATIC int
xlog_recover_buffer_pass2(
struct xlog *log,
struct list_head *buffer_list,
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
struct xlog_recover_item *item,
xfs_lsn_t current_lsn)
{
xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
xfs_mount_t *mp = log->l_mp;
xfs_buf_t *bp;
int error;
uint buf_flags;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
xfs_lsn_t lsn;
/*
* In this pass we only want to recover all the buffers which have
* not been cancelled and are not cancellation buffers themselves.
*/
if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
buf_f->blf_len, buf_f->blf_flags)) {
trace_xfs_log_recover_buf_cancel(log, buf_f);
return 0;
}
trace_xfs_log_recover_buf_recover(log, buf_f);
buf_flags = 0;
if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
buf_flags |= XBF_UNMAPPED;
bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
buf_flags, NULL);
if (!bp)
return XFS_ERROR(ENOMEM);
error = bp->b_error;
if (error) {
xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#1)");
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
goto out_release;
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
/*
* recover the buffer only if we get an LSN from it and it's less than
* the lsn of the transaction we are replaying.
*/
lsn = xlog_recover_get_buf_lsn(mp, bp);
if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0)
goto out_release;
if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
} else if (buf_f->blf_flags &
(XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
} else {
xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
}
if (error)
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
goto out_release;
/*
* Perform delayed write on the buffer. Asynchronous writes will be
* slower when taking into account all the buffers to be flushed.
*
* Also make sure that only inode buffers with good sizes stay in
* the buffer cache. The kernel moves inodes in buffers of 1 block
* or mp->m_inode_cluster_size bytes, whichever is bigger. The inode
* buffers in the log can be a different size if the log was generated
* by an older kernel using unclustered inode buffers or a newer kernel
* running with a different inode cluster size. Regardless, if the
* the inode buffer size isn't MAX(blocksize, mp->m_inode_cluster_size)
* for *our* value of mp->m_inode_cluster_size, then we need to keep
* the buffer out of the buffer cache so that the buffer won't
* overlap with future reads of those inodes.
*/
if (XFS_DINODE_MAGIC ==
be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
(BBTOB(bp->b_io_length) != MAX(log->l_mp->m_sb.sb_blocksize,
(__uint32_t)log->l_mp->m_inode_cluster_size))) {
xfs_buf_stale(bp);
error = xfs_bwrite(bp);
} else {
ASSERT(bp->b_target->bt_mount == mp);
bp->b_iodone = xlog_recover_iodone;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
xfs_buf_delwri_queue(bp, buffer_list);
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
out_release:
xfs_buf_relse(bp);
return error;
}
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
/*
* Inode fork owner changes
*
* If we have been told that we have to reparent the inode fork, it's because an
* extent swap operation on a CRC enabled filesystem has been done and we are
* replaying it. We need to walk the BMBT of the appropriate fork and change the
* owners of it.
*
* The complexity here is that we don't have an inode context to work with, so
* after we've replayed the inode we need to instantiate one. This is where the
* fun begins.
*
* We are in the middle of log recovery, so we can't run transactions. That
* means we cannot use cache coherent inode instantiation via xfs_iget(), as
* that will result in the corresponding iput() running the inode through
* xfs_inactive(). If we've just replayed an inode core that changes the link
* count to zero (i.e. it's been unlinked), then xfs_inactive() will run
* transactions (bad!).
*
* So, to avoid this, we instantiate an inode directly from the inode core we've
* just recovered. We have the buffer still locked, and all we really need to
* instantiate is the inode core and the forks being modified. We can do this
* manually, then run the inode btree owner change, and then tear down the
* xfs_inode without having to run any transactions at all.
*
* Also, because we don't have a transaction context available here but need to
* gather all the buffers we modify for writeback so we pass the buffer_list
* instead for the operation to use.
*/
STATIC int
xfs_recover_inode_owner_change(
struct xfs_mount *mp,
struct xfs_dinode *dip,
struct xfs_inode_log_format *in_f,
struct list_head *buffer_list)
{
struct xfs_inode *ip;
int error;
ASSERT(in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER));
ip = xfs_inode_alloc(mp, in_f->ilf_ino);
if (!ip)
return ENOMEM;
/* instantiate the inode */
xfs_dinode_from_disk(&ip->i_d, dip);
ASSERT(ip->i_d.di_version >= 3);
error = xfs_iformat_fork(ip, dip);
if (error)
goto out_free_ip;
if (in_f->ilf_fields & XFS_ILOG_DOWNER) {
ASSERT(in_f->ilf_fields & XFS_ILOG_DBROOT);
error = xfs_bmbt_change_owner(NULL, ip, XFS_DATA_FORK,
ip->i_ino, buffer_list);
if (error)
goto out_free_ip;
}
if (in_f->ilf_fields & XFS_ILOG_AOWNER) {
ASSERT(in_f->ilf_fields & XFS_ILOG_ABROOT);
error = xfs_bmbt_change_owner(NULL, ip, XFS_ATTR_FORK,
ip->i_ino, buffer_list);
if (error)
goto out_free_ip;
}
out_free_ip:
xfs_inode_free(ip);
return error;
}
STATIC int
xlog_recover_inode_pass2(
struct xlog *log,
struct list_head *buffer_list,
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
struct xlog_recover_item *item,
xfs_lsn_t current_lsn)
{
xfs_inode_log_format_t *in_f;
xfs_mount_t *mp = log->l_mp;
xfs_buf_t *bp;
xfs_dinode_t *dip;
int len;
xfs_caddr_t src;
xfs_caddr_t dest;
int error;
int attr_index;
uint fields;
xfs_icdinode_t *dicp;
uint isize;
int need_free = 0;
if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
in_f = item->ri_buf[0].i_addr;
} else {
in_f = kmem_alloc(sizeof(xfs_inode_log_format_t), KM_SLEEP);
need_free = 1;
error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
if (error)
goto error;
}
/*
* Inode buffers can be freed, look out for it,
* and do not replay the inode.
*/
if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
in_f->ilf_len, 0)) {
error = 0;
trace_xfs_log_recover_inode_cancel(log, in_f);
goto error;
}
trace_xfs_log_recover_inode_recover(log, in_f);
bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len, 0,
&xfs_inode_buf_ops);
if (!bp) {
error = ENOMEM;
goto error;
}
error = bp->b_error;
if (error) {
xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#2)");
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
/*
* Make sure the place we're flushing out to really looks
* like an inode!
*/
if (unlikely(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC))) {
xfs_alert(mp,
"%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
__func__, dip, bp, in_f->ilf_ino);
XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
XFS_ERRLEVEL_LOW, mp);
error = EFSCORRUPTED;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
dicp = item->ri_buf[1].i_addr;
if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
xfs_alert(mp,
"%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
__func__, item, in_f->ilf_ino);
XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
XFS_ERRLEVEL_LOW, mp);
error = EFSCORRUPTED;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
/*
* If the inode has an LSN in it, recover the inode only if it's less
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
* than the lsn of the transaction we are replaying. Note: we still
* need to replay an owner change even though the inode is more recent
* than the transaction as there is no guarantee that all the btree
* blocks are more recent than this transaction, too.
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
*/
if (dip->di_version >= 3) {
xfs_lsn_t lsn = be64_to_cpu(dip->di_lsn);
if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
trace_xfs_log_recover_inode_skip(log, in_f);
error = 0;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_owner_change;
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
}
}
xfs: di_flushiter considered harmful When we made all inode updates transactional, we no longer needed the log recovery detection for inodes being newer on disk than the transaction being replayed - it was redundant as replay of the log would always result in the latest version of the inode would be on disk. It was redundant, but left in place because it wasn't considered to be a problem. However, with the new "don't read inodes on create" optimisation, flushiter has come back to bite us. Essentially, the optimisation made always initialises flushiter to zero in the create transaction, and so if we then crash and run recovery and the inode already on disk has a non-zero flushiter it will skip recovery of that inode. As a result, log recovery does the wrong thing and we end up with a corrupt filesystem. Because we have to support old kernel to new kernel upgrades, we can't just get rid of the flushiter support in log recovery as we might be upgrading from a kernel that doesn't have fully transactional inode updates. Unfortunately, for v4 superblocks there is no way to guarantee that log recovery knows about this fact. We cannot add a new inode format flag to say it's a "special inode create" because it won't be understood by older kernels and so recovery could do the wrong thing on downgrade. We cannot specially detect the combination of zero mode/non-zero flushiter on disk to non-zero mode, zero flushiter in the log item during recovery because wrapping of the flushiter can result in false detection. Hence that makes this "don't use flushiter" optimisation limited to a disk format that guarantees that we don't need it. And that means the only fix here is to limit the "no read IO on create" optimisation to version 5 superblocks.... Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-07-24 05:47:30 +00:00
/*
* di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
* are transactional and if ordering is necessary we can determine that
* more accurately by the LSN field in the V3 inode core. Don't trust
* the inode versions we might be changing them here - use the
* superblock flag to determine whether we need to look at di_flushiter
* to skip replay when the on disk inode is newer than the log one
*/
if (!xfs_sb_version_hascrc(&mp->m_sb) &&
dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
/*
* Deal with the wrap case, DI_MAX_FLUSH is less
* than smaller numbers
*/
if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
/* do nothing */
} else {
trace_xfs_log_recover_inode_skip(log, in_f);
error = 0;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
}
xfs: di_flushiter considered harmful When we made all inode updates transactional, we no longer needed the log recovery detection for inodes being newer on disk than the transaction being replayed - it was redundant as replay of the log would always result in the latest version of the inode would be on disk. It was redundant, but left in place because it wasn't considered to be a problem. However, with the new "don't read inodes on create" optimisation, flushiter has come back to bite us. Essentially, the optimisation made always initialises flushiter to zero in the create transaction, and so if we then crash and run recovery and the inode already on disk has a non-zero flushiter it will skip recovery of that inode. As a result, log recovery does the wrong thing and we end up with a corrupt filesystem. Because we have to support old kernel to new kernel upgrades, we can't just get rid of the flushiter support in log recovery as we might be upgrading from a kernel that doesn't have fully transactional inode updates. Unfortunately, for v4 superblocks there is no way to guarantee that log recovery knows about this fact. We cannot add a new inode format flag to say it's a "special inode create" because it won't be understood by older kernels and so recovery could do the wrong thing on downgrade. We cannot specially detect the combination of zero mode/non-zero flushiter on disk to non-zero mode, zero flushiter in the log item during recovery because wrapping of the flushiter can result in false detection. Hence that makes this "don't use flushiter" optimisation limited to a disk format that guarantees that we don't need it. And that means the only fix here is to limit the "no read IO on create" optimisation to version 5 superblocks.... Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-07-24 05:47:30 +00:00
/* Take the opportunity to reset the flush iteration count */
dicp->di_flushiter = 0;
if (unlikely(S_ISREG(dicp->di_mode))) {
if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
(dicp->di_format != XFS_DINODE_FMT_BTREE)) {
XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
XFS_ERRLEVEL_LOW, mp, dicp);
xfs_alert(mp,
"%s: Bad regular inode log record, rec ptr 0x%p, "
"ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
__func__, item, dip, bp, in_f->ilf_ino);
error = EFSCORRUPTED;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
} else if (unlikely(S_ISDIR(dicp->di_mode))) {
if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
(dicp->di_format != XFS_DINODE_FMT_BTREE) &&
(dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
XFS_ERRLEVEL_LOW, mp, dicp);
xfs_alert(mp,
"%s: Bad dir inode log record, rec ptr 0x%p, "
"ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
__func__, item, dip, bp, in_f->ilf_ino);
error = EFSCORRUPTED;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
}
if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
XFS_ERRLEVEL_LOW, mp, dicp);
xfs_alert(mp,
"%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
"dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
__func__, item, dip, bp, in_f->ilf_ino,
dicp->di_nextents + dicp->di_anextents,
dicp->di_nblocks);
error = EFSCORRUPTED;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
XFS_ERRLEVEL_LOW, mp, dicp);
xfs_alert(mp,
"%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
"dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__,
item, dip, bp, in_f->ilf_ino, dicp->di_forkoff);
error = EFSCORRUPTED;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
isize = xfs_icdinode_size(dicp->di_version);
if (unlikely(item->ri_buf[1].i_len > isize)) {
XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
XFS_ERRLEVEL_LOW, mp, dicp);
xfs_alert(mp,
"%s: Bad inode log record length %d, rec ptr 0x%p",
__func__, item->ri_buf[1].i_len, item);
error = EFSCORRUPTED;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
/* The core is in in-core format */
xfs_dinode_to_disk(dip, dicp);
/* the rest is in on-disk format */
if (item->ri_buf[1].i_len > isize) {
memcpy((char *)dip + isize,
item->ri_buf[1].i_addr + isize,
item->ri_buf[1].i_len - isize);
}
fields = in_f->ilf_fields;
switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
case XFS_ILOG_DEV:
xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
break;
case XFS_ILOG_UUID:
memcpy(XFS_DFORK_DPTR(dip),
&in_f->ilf_u.ilfu_uuid,
sizeof(uuid_t));
break;
}
if (in_f->ilf_size == 2)
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_owner_change;
len = item->ri_buf[2].i_len;
src = item->ri_buf[2].i_addr;
ASSERT(in_f->ilf_size <= 4);
ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
ASSERT(!(fields & XFS_ILOG_DFORK) ||
(len == in_f->ilf_dsize));
switch (fields & XFS_ILOG_DFORK) {
case XFS_ILOG_DDATA:
case XFS_ILOG_DEXT:
memcpy(XFS_DFORK_DPTR(dip), src, len);
break;
case XFS_ILOG_DBROOT:
xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
(xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
XFS_DFORK_DSIZE(dip, mp));
break;
default:
/*
* There are no data fork flags set.
*/
ASSERT((fields & XFS_ILOG_DFORK) == 0);
break;
}
/*
* If we logged any attribute data, recover it. There may or
* may not have been any other non-core data logged in this
* transaction.
*/
if (in_f->ilf_fields & XFS_ILOG_AFORK) {
if (in_f->ilf_fields & XFS_ILOG_DFORK) {
attr_index = 3;
} else {
attr_index = 2;
}
len = item->ri_buf[attr_index].i_len;
src = item->ri_buf[attr_index].i_addr;
ASSERT(len == in_f->ilf_asize);
switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
case XFS_ILOG_ADATA:
case XFS_ILOG_AEXT:
dest = XFS_DFORK_APTR(dip);
ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
memcpy(dest, src, len);
break;
case XFS_ILOG_ABROOT:
dest = XFS_DFORK_APTR(dip);
xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
len, (xfs_bmdr_block_t*)dest,
XFS_DFORK_ASIZE(dip, mp));
break;
default:
xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
ASSERT(0);
error = EIO;
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
goto out_release;
}
}
xfs: recovery of swap extents operations for CRC filesystems This is the recovery side of the btree block owner change operation performed by swapext on CRC enabled filesystems. We detect that an owner change is needed by the flag that has been placed on the inode log format flag field. Because the inode recovery is being replayed after the buffers that make up the BMBT in the given checkpoint, we can walk all the buffers and directly modify them when we see the flag set on an inode. Because the inode can be relogged and hence present in multiple chekpoints with the "change owner" flag set, we could do multiple passes across the inode to do this change. While this isn't optimal, we can't directly ignore the flag as there may be multiple independent swap extent operations being replayed on the same inode in different checkpoints so we can't ignore them. Further, because the owner change operation uses ordered buffers, we might have buffers that are newer on disk than the current checkpoint and so already have the owner changed in them. Hence we cannot just peek at a buffer in the tree and check that it has the correct owner and assume that the change was completed. So, for the moment just brute force the owner change every time we see an inode with the flag set. Note that we have to be careful here because the owner of the buffers may point to either the old owner or the new owner. Currently the verifier can't verify the owner directly, so there is no failure case here right now. If we verify the owner exactly in future, then we'll have to take this into account. This was tested in terms of normal operation via xfstests - all of the fsr tests now pass without failure. however, we really need to modify xfs/227 to stress v3 inodes correctly to ensure we fully cover this case for v5 filesystems. In terms of recovery testing, I used a hacked version of xfs_fsr that held the temp inode open for a few seconds before exiting so that the filesystem could be shut down with an open owner change recovery flags set on at least the temp inode. fsr leaves the temp inode unlinked and in btree format, so this was necessary for the owner change to be reliably replayed. logprint confirmed the tmp inode in the log had the correct flag set: INO: cnt:3 total:3 a:0x69e9e0 len:56 a:0x69ea20 len:176 a:0x69eae0 len:88 INODE: #regs:3 ino:0x44 flags:0x209 dsize:88 ^^^^^ 0x200 is set, indicating a data fork owner change needed to be replayed on inode 0x44. A printk in the revoery code confirmed that the inode change was recovered: XFS (vdc): Mounting Filesystem XFS (vdc): Starting recovery (logdev: internal) recovering owner change ino 0x44 XFS (vdc): Version 5 superblock detected. This kernel L support enabled! Use of these features in this kernel is at your own risk! XFS (vdc): Ending recovery (logdev: internal) The script used to test this was: $ cat ./recovery-fsr.sh #!/bin/bash dev=/dev/vdc mntpt=/mnt/scratch testfile=$mntpt/testfile umount $mntpt mkfs.xfs -f -m crc=1 $dev mount $dev $mntpt chmod 777 $mntpt for i in `seq 10000 -1 0`; do xfs_io -f -d -c "pwrite $(($i * 4096)) 4096" $testfile > /dev/null 2>&1 done xfs_bmap -vp $testfile |head -20 xfs_fsr -d -v $testfile & sleep 10 /home/dave/src/xfstests-dev/src/godown -f $mntpt wait umount $mntpt xfs_logprint -t $dev |tail -20 time mount $dev $mntpt xfs_bmap -vp $testfile umount $mntpt $ Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-30 00:23:45 +00:00
out_owner_change:
if (in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER))
error = xfs_recover_inode_owner_change(mp, dip, in_f,
buffer_list);
/* re-generate the checksum. */
xfs_dinode_calc_crc(log->l_mp, dip);
ASSERT(bp->b_target->bt_mount == mp);
bp->b_iodone = xlog_recover_iodone;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
xfs_buf_delwri_queue(bp, buffer_list);
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
out_release:
xfs_buf_relse(bp);
error:
if (need_free)
kmem_free(in_f);
return XFS_ERROR(error);
}
/*
* Recover QUOTAOFF records. We simply make a note of it in the xlog
* structure, so that we know not to do any dquot item or dquot buffer recovery,
* of that type.
*/
STATIC int
xlog_recover_quotaoff_pass1(
struct xlog *log,
struct xlog_recover_item *item)
{
xfs_qoff_logformat_t *qoff_f = item->ri_buf[0].i_addr;
ASSERT(qoff_f);
/*
* The logitem format's flag tells us if this was user quotaoff,
* group/project quotaoff or both.
*/
if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
log->l_quotaoffs_flag |= XFS_DQ_USER;
if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
log->l_quotaoffs_flag |= XFS_DQ_PROJ;
if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
log->l_quotaoffs_flag |= XFS_DQ_GROUP;
return (0);
}
/*
* Recover a dquot record
*/
STATIC int
xlog_recover_dquot_pass2(
struct xlog *log,
struct list_head *buffer_list,
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
struct xlog_recover_item *item,
xfs_lsn_t current_lsn)
{
xfs_mount_t *mp = log->l_mp;
xfs_buf_t *bp;
struct xfs_disk_dquot *ddq, *recddq;
int error;
xfs_dq_logformat_t *dq_f;
uint type;
/*
* Filesystems are required to send in quota flags at mount time.
*/
if (mp->m_qflags == 0)
return (0);
recddq = item->ri_buf[1].i_addr;
if (recddq == NULL) {
xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
return XFS_ERROR(EIO);
}
if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
item->ri_buf[1].i_len, __func__);
return XFS_ERROR(EIO);
}
/*
* This type of quotas was turned off, so ignore this record.
*/
type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
ASSERT(type);
if (log->l_quotaoffs_flag & type)
return (0);
/*
* At this point we know that quota was _not_ turned off.
* Since the mount flags are not indicating to us otherwise, this
* must mean that quota is on, and the dquot needs to be replayed.
* Remember that we may not have fully recovered the superblock yet,
* so we can't do the usual trick of looking at the SB quota bits.
*
* The other possibility, of course, is that the quota subsystem was
* removed since the last mount - ENOSYS.
*/
dq_f = item->ri_buf[0].i_addr;
ASSERT(dq_f);
error = xfs_dqcheck(mp, recddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
"xlog_recover_dquot_pass2 (log copy)");
if (error)
return XFS_ERROR(EIO);
ASSERT(dq_f->qlf_len == 1);
error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dq_f->qlf_blkno,
XFS_FSB_TO_BB(mp, dq_f->qlf_len), 0, &bp,
NULL);
if (error)
return error;
ASSERT(bp);
ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
/*
* At least the magic num portion should be on disk because this
* was among a chunk of dquots created earlier, and we did some
* minimal initialization then.
*/
error = xfs_dqcheck(mp, ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
"xlog_recover_dquot_pass2");
if (error) {
xfs_buf_relse(bp);
return XFS_ERROR(EIO);
}
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
/*
* If the dquot has an LSN in it, recover the dquot only if it's less
* than the lsn of the transaction we are replaying.
*/
if (xfs_sb_version_hascrc(&mp->m_sb)) {
struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddq;
xfs_lsn_t lsn = be64_to_cpu(dqb->dd_lsn);
if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
goto out_release;
}
}
memcpy(ddq, recddq, item->ri_buf[1].i_len);
if (xfs_sb_version_hascrc(&mp->m_sb)) {
xfs_update_cksum((char *)ddq, sizeof(struct xfs_dqblk),
XFS_DQUOT_CRC_OFF);
}
ASSERT(dq_f->qlf_size == 2);
ASSERT(bp->b_target->bt_mount == mp);
bp->b_iodone = xlog_recover_iodone;
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
xfs_buf_delwri_queue(bp, buffer_list);
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
out_release:
xfs_buf_relse(bp);
return 0;
}
/*
* This routine is called to create an in-core extent free intent
* item from the efi format structure which was logged on disk.
* It allocates an in-core efi, copies the extents from the format
* structure into it, and adds the efi to the AIL with the given
* LSN.
*/
STATIC int
xlog_recover_efi_pass2(
struct xlog *log,
struct xlog_recover_item *item,
xfs_lsn_t lsn)
{
int error;
xfs_mount_t *mp = log->l_mp;
xfs_efi_log_item_t *efip;
xfs_efi_log_format_t *efi_formatp;
efi_formatp = item->ri_buf[0].i_addr;
efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
&(efip->efi_format)))) {
xfs_efi_item_free(efip);
return error;
}
atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
spin_lock(&log->l_ailp->xa_lock);
/*
* xfs_trans_ail_update() drops the AIL lock.
*/
xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
return 0;
}
/*
* This routine is called when an efd format structure is found in
* a committed transaction in the log. It's purpose is to cancel
* the corresponding efi if it was still in the log. To do this
* it searches the AIL for the efi with an id equal to that in the
* efd format structure. If we find it, we remove the efi from the
* AIL and free it.
*/
STATIC int
xlog_recover_efd_pass2(
struct xlog *log,
struct xlog_recover_item *item)
{
xfs_efd_log_format_t *efd_formatp;
xfs_efi_log_item_t *efip = NULL;
xfs_log_item_t *lip;
__uint64_t efi_id;
struct xfs_ail_cursor cur;
struct xfs_ail *ailp = log->l_ailp;
efd_formatp = item->ri_buf[0].i_addr;
ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
(item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
efi_id = efd_formatp->efd_efi_id;
/*
* Search for the efi with the id in the efd format structure
* in the AIL.
*/
spin_lock(&ailp->xa_lock);
lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
while (lip != NULL) {
if (lip->li_type == XFS_LI_EFI) {
efip = (xfs_efi_log_item_t *)lip;
if (efip->efi_format.efi_id == efi_id) {
/*
* xfs_trans_ail_delete() drops the
* AIL lock.
*/
xfs_trans_ail_delete(ailp, lip,
SHUTDOWN_CORRUPT_INCORE);
xfs_efi_item_free(efip);
spin_lock(&ailp->xa_lock);
break;
}
}
lip = xfs_trans_ail_cursor_next(ailp, &cur);
}
xfs_trans_ail_cursor_done(&cur);
spin_unlock(&ailp->xa_lock);
return 0;
}
/*
* This routine is called when an inode create format structure is found in a
* committed transaction in the log. It's purpose is to initialise the inodes
* being allocated on disk. This requires us to get inode cluster buffers that
* match the range to be intialised, stamped with inode templates and written
* by delayed write so that subsequent modifications will hit the cached buffer
* and only need writing out at the end of recovery.
*/
STATIC int
xlog_recover_do_icreate_pass2(
struct xlog *log,
struct list_head *buffer_list,
xlog_recover_item_t *item)
{
struct xfs_mount *mp = log->l_mp;
struct xfs_icreate_log *icl;
xfs_agnumber_t agno;
xfs_agblock_t agbno;
unsigned int count;
unsigned int isize;
xfs_agblock_t length;
icl = (struct xfs_icreate_log *)item->ri_buf[0].i_addr;
if (icl->icl_type != XFS_LI_ICREATE) {
xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad type");
return EINVAL;
}
if (icl->icl_size != 1) {
xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad icl size");
return EINVAL;
}
agno = be32_to_cpu(icl->icl_ag);
if (agno >= mp->m_sb.sb_agcount) {
xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agno");
return EINVAL;
}
agbno = be32_to_cpu(icl->icl_agbno);
if (!agbno || agbno == NULLAGBLOCK || agbno >= mp->m_sb.sb_agblocks) {
xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agbno");
return EINVAL;
}
isize = be32_to_cpu(icl->icl_isize);
if (isize != mp->m_sb.sb_inodesize) {
xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad isize");
return EINVAL;
}
count = be32_to_cpu(icl->icl_count);
if (!count) {
xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count");
return EINVAL;
}
length = be32_to_cpu(icl->icl_length);
if (!length || length >= mp->m_sb.sb_agblocks) {
xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad length");
return EINVAL;
}
/* existing allocation is fixed value */
ASSERT(count == mp->m_ialloc_inos);
ASSERT(length == mp->m_ialloc_blks);
if (count != mp->m_ialloc_inos ||
length != mp->m_ialloc_blks) {
xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count 2");
return EINVAL;
}
/*
* Inode buffers can be freed. Do not replay the inode initialisation as
* we could be overwriting something written after this inode buffer was
* cancelled.
*
* XXX: we need to iterate all buffers and only init those that are not
* cancelled. I think that a more fine grained factoring of
* xfs_ialloc_inode_init may be appropriate here to enable this to be
* done easily.
*/
if (xlog_check_buffer_cancelled(log,
XFS_AGB_TO_DADDR(mp, agno, agbno), length, 0))
return 0;
xfs_ialloc_inode_init(mp, NULL, buffer_list, agno, agbno, length,
be32_to_cpu(icl->icl_gen));
return 0;
}
/*
* Free up any resources allocated by the transaction
*
* Remember that EFIs, EFDs, and IUNLINKs are handled later.
*/
STATIC void
xlog_recover_free_trans(
struct xlog_recover *trans)
{
xlog_recover_item_t *item, *n;
int i;
list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
/* Free the regions in the item. */
list_del(&item->ri_list);
for (i = 0; i < item->ri_cnt; i++)
kmem_free(item->ri_buf[i].i_addr);
/* Free the item itself */
kmem_free(item->ri_buf);
kmem_free(item);
}
/* Free the transaction recover structure */
kmem_free(trans);
}
STATIC void
xlog_recover_buffer_ra_pass2(
struct xlog *log,
struct xlog_recover_item *item)
{
struct xfs_buf_log_format *buf_f = item->ri_buf[0].i_addr;
struct xfs_mount *mp = log->l_mp;
if (xlog_peek_buffer_cancelled(log, buf_f->blf_blkno,
buf_f->blf_len, buf_f->blf_flags)) {
return;
}
xfs_buf_readahead(mp->m_ddev_targp, buf_f->blf_blkno,
buf_f->blf_len, NULL);
}
STATIC void
xlog_recover_inode_ra_pass2(
struct xlog *log,
struct xlog_recover_item *item)
{
struct xfs_inode_log_format ilf_buf;
struct xfs_inode_log_format *ilfp;
struct xfs_mount *mp = log->l_mp;
int error;
if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
ilfp = item->ri_buf[0].i_addr;
} else {
ilfp = &ilf_buf;
memset(ilfp, 0, sizeof(*ilfp));
error = xfs_inode_item_format_convert(&item->ri_buf[0], ilfp);
if (error)
return;
}
if (xlog_peek_buffer_cancelled(log, ilfp->ilf_blkno, ilfp->ilf_len, 0))
return;
xfs_buf_readahead(mp->m_ddev_targp, ilfp->ilf_blkno,
ilfp->ilf_len, &xfs_inode_buf_ra_ops);
}
STATIC void
xlog_recover_dquot_ra_pass2(
struct xlog *log,
struct xlog_recover_item *item)
{
struct xfs_mount *mp = log->l_mp;
struct xfs_disk_dquot *recddq;
struct xfs_dq_logformat *dq_f;
uint type;
if (mp->m_qflags == 0)
return;
recddq = item->ri_buf[1].i_addr;
if (recddq == NULL)
return;
if (item->ri_buf[1].i_len < sizeof(struct xfs_disk_dquot))
return;
type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
ASSERT(type);
if (log->l_quotaoffs_flag & type)
return;
dq_f = item->ri_buf[0].i_addr;
ASSERT(dq_f);
ASSERT(dq_f->qlf_len == 1);
xfs_buf_readahead(mp->m_ddev_targp, dq_f->qlf_blkno,
XFS_FSB_TO_BB(mp, dq_f->qlf_len), NULL);
}
STATIC void
xlog_recover_ra_pass2(
struct xlog *log,
struct xlog_recover_item *item)
{
switch (ITEM_TYPE(item)) {
case XFS_LI_BUF:
xlog_recover_buffer_ra_pass2(log, item);
break;
case XFS_LI_INODE:
xlog_recover_inode_ra_pass2(log, item);
break;
case XFS_LI_DQUOT:
xlog_recover_dquot_ra_pass2(log, item);
break;
case XFS_LI_EFI:
case XFS_LI_EFD:
case XFS_LI_QUOTAOFF:
default:
break;
}
}
STATIC int
xlog_recover_commit_pass1(
struct xlog *log,
struct xlog_recover *trans,
struct xlog_recover_item *item)
{
trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
switch (ITEM_TYPE(item)) {
case XFS_LI_BUF:
return xlog_recover_buffer_pass1(log, item);
case XFS_LI_QUOTAOFF:
return xlog_recover_quotaoff_pass1(log, item);
case XFS_LI_INODE:
case XFS_LI_EFI:
case XFS_LI_EFD:
case XFS_LI_DQUOT:
case XFS_LI_ICREATE:
/* nothing to do in pass 1 */
return 0;
default:
xfs_warn(log->l_mp, "%s: invalid item type (%d)",
__func__, ITEM_TYPE(item));
ASSERT(0);
return XFS_ERROR(EIO);
}
}
STATIC int
xlog_recover_commit_pass2(
struct xlog *log,
struct xlog_recover *trans,
struct list_head *buffer_list,
struct xlog_recover_item *item)
{
trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
switch (ITEM_TYPE(item)) {
case XFS_LI_BUF:
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
return xlog_recover_buffer_pass2(log, buffer_list, item,
trans->r_lsn);
case XFS_LI_INODE:
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
return xlog_recover_inode_pass2(log, buffer_list, item,
trans->r_lsn);
case XFS_LI_EFI:
return xlog_recover_efi_pass2(log, item, trans->r_lsn);
case XFS_LI_EFD:
return xlog_recover_efd_pass2(log, item);
case XFS_LI_DQUOT:
xfs: check LSN ordering for v5 superblocks during recovery Log recovery has some strict ordering requirements which unordered or reordered metadata writeback can defeat. This can occur when an item is logged in a transaction, written back to disk, and then logged in a new transaction before the tail of the log is moved past the original modification. The result of this is that when we read an object off disk for recovery purposes, the buffer that we read may not contain the object type that recovery is expecting and hence at the end of the checkpoint being recovered we have an invalid object in memory. This isn't usually a problem, as recovery will then replay all the other checkpoints and that brings the object back to a valid and correct state, but the issue is that while the object is in the invalid state it can be flushed to disk. This results in the object verifier failing and triggering a corruption shutdown of log recover. This is correct behaviour for the verifiers - the problem is that we are not detecting that the object we've read off disk is newer than the transaction we are replaying. All metadata in v5 filesystems has the LSN of it's last modification stamped in it. This enabled log recover to read that field and determine the age of the object on disk correctly. If the LSN of the object on disk is older than the transaction being replayed, then we replay the modification. If the LSN of the object matches or is more recent than the transaction's LSN, then we should avoid overwriting the object as that is what leads to the transient corrupt state. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-08-28 11:22:47 +00:00
return xlog_recover_dquot_pass2(log, buffer_list, item,
trans->r_lsn);
case XFS_LI_ICREATE:
return xlog_recover_do_icreate_pass2(log, buffer_list, item);
case XFS_LI_QUOTAOFF:
/* nothing to do in pass2 */
return 0;
default:
xfs_warn(log->l_mp, "%s: invalid item type (%d)",
__func__, ITEM_TYPE(item));
ASSERT(0);
return XFS_ERROR(EIO);
}
}
STATIC int
xlog_recover_items_pass2(
struct xlog *log,
struct xlog_recover *trans,
struct list_head *buffer_list,
struct list_head *item_list)
{
struct xlog_recover_item *item;
int error = 0;
list_for_each_entry(item, item_list, ri_list) {
error = xlog_recover_commit_pass2(log, trans,
buffer_list, item);
if (error)
return error;
}
return error;
}
/*
* Perform the transaction.
*
* If the transaction modifies a buffer or inode, do it now. Otherwise,
* EFIs and EFDs get queued up by adding entries into the AIL for them.
*/
STATIC int
xlog_recover_commit_trans(
struct xlog *log,
struct xlog_recover *trans,
int pass)
{
int error = 0;
int error2;
int items_queued = 0;
struct xlog_recover_item *item;
struct xlog_recover_item *next;
LIST_HEAD (buffer_list);
LIST_HEAD (ra_list);
LIST_HEAD (done_list);
#define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
hlist_del(&trans->r_list);
error = xlog_recover_reorder_trans(log, trans, pass);
if (error)
return error;
list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
switch (pass) {
case XLOG_RECOVER_PASS1:
error = xlog_recover_commit_pass1(log, trans, item);
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
break;
case XLOG_RECOVER_PASS2:
xlog_recover_ra_pass2(log, item);
list_move_tail(&item->ri_list, &ra_list);
items_queued++;
if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
error = xlog_recover_items_pass2(log, trans,
&buffer_list, &ra_list);
list_splice_tail_init(&ra_list, &done_list);
items_queued = 0;
}
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
break;
default:
ASSERT(0);
}
if (error)
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
goto out;
}
out:
if (!list_empty(&ra_list)) {
if (!error)
error = xlog_recover_items_pass2(log, trans,
&buffer_list, &ra_list);
list_splice_tail_init(&ra_list, &done_list);
}
if (!list_empty(&done_list))
list_splice_init(&done_list, &trans->r_itemq);
xlog_recover_free_trans(trans);
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
error2 = xfs_buf_delwri_submit(&buffer_list);
return error ? error : error2;
}
STATIC int
xlog_recover_unmount_trans(
struct xlog *log)
{
/* Do nothing now */
xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
return 0;
}
/*
* There are two valid states of the r_state field. 0 indicates that the
* transaction structure is in a normal state. We have either seen the
* start of the transaction or the last operation we added was not a partial
* operation. If the last operation we added to the transaction was a
* partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
*
* NOTE: skip LRs with 0 data length.
*/
STATIC int
xlog_recover_process_data(
struct xlog *log,
struct hlist_head rhash[],
struct xlog_rec_header *rhead,
xfs_caddr_t dp,
int pass)
{
xfs_caddr_t lp;
int num_logops;
xlog_op_header_t *ohead;
xlog_recover_t *trans;
xlog_tid_t tid;
int error;
unsigned long hash;
uint flags;
lp = dp + be32_to_cpu(rhead->h_len);
num_logops = be32_to_cpu(rhead->h_num_logops);
/* check the log format matches our own - else we can't recover */
if (xlog_header_check_recover(log->l_mp, rhead))
return (XFS_ERROR(EIO));
while ((dp < lp) && num_logops) {
ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
ohead = (xlog_op_header_t *)dp;
dp += sizeof(xlog_op_header_t);
if (ohead->oh_clientid != XFS_TRANSACTION &&
ohead->oh_clientid != XFS_LOG) {
xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
__func__, ohead->oh_clientid);
ASSERT(0);
return (XFS_ERROR(EIO));
}
tid = be32_to_cpu(ohead->oh_tid);
hash = XLOG_RHASH(tid);
trans = xlog_recover_find_tid(&rhash[hash], tid);
if (trans == NULL) { /* not found; add new tid */
if (ohead->oh_flags & XLOG_START_TRANS)
xlog_recover_new_tid(&rhash[hash], tid,
be64_to_cpu(rhead->h_lsn));
} else {
if (dp + be32_to_cpu(ohead->oh_len) > lp) {
xfs_warn(log->l_mp, "%s: bad length 0x%x",
__func__, be32_to_cpu(ohead->oh_len));
WARN_ON(1);
return (XFS_ERROR(EIO));
}
flags = ohead->oh_flags & ~XLOG_END_TRANS;
if (flags & XLOG_WAS_CONT_TRANS)
flags &= ~XLOG_CONTINUE_TRANS;
switch (flags) {
case XLOG_COMMIT_TRANS:
error = xlog_recover_commit_trans(log,
trans, pass);
break;
case XLOG_UNMOUNT_TRANS:
error = xlog_recover_unmount_trans(log);
break;
case XLOG_WAS_CONT_TRANS:
error = xlog_recover_add_to_cont_trans(log,
trans, dp,
be32_to_cpu(ohead->oh_len));
break;
case XLOG_START_TRANS:
xfs_warn(log->l_mp, "%s: bad transaction",
__func__);
ASSERT(0);
error = XFS_ERROR(EIO);
break;
case 0:
case XLOG_CONTINUE_TRANS:
error = xlog_recover_add_to_trans(log, trans,
dp, be32_to_cpu(ohead->oh_len));
break;
default:
xfs_warn(log->l_mp, "%s: bad flag 0x%x",
__func__, flags);
ASSERT(0);
error = XFS_ERROR(EIO);
break;
}
if (error) {
xlog_recover_free_trans(trans);
return error;
}
}
dp += be32_to_cpu(ohead->oh_len);
num_logops--;
}
return 0;
}
/*
* Process an extent free intent item that was recovered from
* the log. We need to free the extents that it describes.
*/
STATIC int
xlog_recover_process_efi(
xfs_mount_t *mp,
xfs_efi_log_item_t *efip)
{
xfs_efd_log_item_t *efdp;
xfs_trans_t *tp;
int i;
int error = 0;
xfs_extent_t *extp;
xfs_fsblock_t startblock_fsb;
ASSERT(!test_bit(XFS_EFI_RECOVERED, &efip->efi_flags));
/*
* First check the validity of the extents described by the
* EFI. If any are bad, then assume that all are bad and
* just toss the EFI.
*/
for (i = 0; i < efip->efi_format.efi_nextents; i++) {
extp = &(efip->efi_format.efi_extents[i]);
startblock_fsb = XFS_BB_TO_FSB(mp,
XFS_FSB_TO_DADDR(mp, extp->ext_start));
if ((startblock_fsb == 0) ||
(extp->ext_len == 0) ||
(startblock_fsb >= mp->m_sb.sb_dblocks) ||
(extp->ext_len >= mp->m_sb.sb_agblocks)) {
/*
* This will pull the EFI from the AIL and
* free the memory associated with it.
*/
xfs: don't free EFIs before the EFDs are committed Filesystems are occasionally being shut down with this error: xfs_trans_ail_delete_bulk: attempting to delete a log item that is not in the AIL. It was diagnosed to be related to the EFI/EFD commit order when the EFI and EFD are in different checkpoints and the EFD is committed before the EFI here: http://oss.sgi.com/archives/xfs/2013-01/msg00082.html The real problem is that a single bit cannot fully describe the states that the EFI/EFD processing can be in. These completion states are: EFI EFI in AIL EFD Result committed/unpinned Yes committed OK committed/pinned No committed Shutdown uncommitted No committed Shutdown Note that the "result" field is what should happen, not what does happen. The current logic is broken and handles the first two cases correctly by luck. That is, the code will free the EFI if the XFS_EFI_COMMITTED bit is *not* set, rather than if it is set. The inverted logic "works" because if both EFI and EFD are committed, then the first __xfs_efi_release() call clears the XFS_EFI_COMMITTED bit, and the second frees the EFI item. Hence as long as xfs_efi_item_committed() has been called, everything appears to be fine. It is the third case where the logic fails - where xfs_efd_item_committed() is called before xfs_efi_item_committed(), and that results in the EFI being freed before it has been committed. That is the bug that triggered the shutdown, and hence keeping track of whether the EFI has been committed or not is insufficient to correctly order the EFI/EFD operations w.r.t. the AIL. What we really want is this: the EFI is always placed into the AIL before the last reference goes away. The only way to guarantee that is that the EFI is not freed until after it has been unpinned *and* the EFD has been committed. That is, restructure the logic so that the only case that can occur is the first case. This can be done easily by replacing the XFS_EFI_COMMITTED with an EFI reference count. The EFI is initialised with it's own count, and that is not released until it is unpinned. However, there is a complication to this method - the high level EFI/EFD code in xfs_bmap_finish() does not hold direct references to the EFI structure, and runs a transaction commit between the EFI and EFD processing. Hence the EFI can be freed even before the EFD is created using such a method. Further, log recovery uses the AIL for tracking EFI/EFDs that need to be recovered, but it uses the AIL *differently* to the EFI transaction commit. Hence log recovery never pins or unpins EFIs, so we can't drop the EFI reference count indirectly to free the EFI. However, this doesn't prevent us from using a reference count here. There is a 1:1 relationship between EFIs and EFDs, so when we initialise the EFI we can take a reference count for the EFD as well. This solves the xfs_bmap_finish() issue - the EFI will never be freed until the EFD is processed. In terms of log recovery, during the committing of the EFD we can look for the XFS_EFI_RECOVERED bit being set and drop the EFI reference as well, thereby ensuring everything works correctly there as well. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2013-04-03 03:09:21 +00:00
set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
xfs_efi_release(efip, efip->efi_format.efi_nextents);
return XFS_ERROR(EIO);
}
}
tp = xfs_trans_alloc(mp, 0);
error = xfs_trans_reserve(tp, &M_RES(mp)->tr_itruncate, 0, 0);
if (error)
goto abort_error;
efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
for (i = 0; i < efip->efi_format.efi_nextents; i++) {
extp = &(efip->efi_format.efi_extents[i]);
error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
if (error)
goto abort_error;
xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
extp->ext_len);
}
set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
error = xfs_trans_commit(tp, 0);
return error;
abort_error:
xfs_trans_cancel(tp, XFS_TRANS_ABORT);
return error;
}
/*
* When this is called, all of the EFIs which did not have
* corresponding EFDs should be in the AIL. What we do now
* is free the extents associated with each one.
*
* Since we process the EFIs in normal transactions, they
* will be removed at some point after the commit. This prevents
* us from just walking down the list processing each one.
* We'll use a flag in the EFI to skip those that we've already
* processed and use the AIL iteration mechanism's generation
* count to try to speed this up at least a bit.
*
* When we start, we know that the EFIs are the only things in
* the AIL. As we process them, however, other items are added
* to the AIL. Since everything added to the AIL must come after
* everything already in the AIL, we stop processing as soon as
* we see something other than an EFI in the AIL.
*/
STATIC int
xlog_recover_process_efis(
struct xlog *log)
{
xfs_log_item_t *lip;
xfs_efi_log_item_t *efip;
int error = 0;
struct xfs_ail_cursor cur;
struct xfs_ail *ailp;
ailp = log->l_ailp;
spin_lock(&ailp->xa_lock);
lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
while (lip != NULL) {
/*
* We're done when we see something other than an EFI.
* There should be no EFIs left in the AIL now.
*/
if (lip->li_type != XFS_LI_EFI) {
#ifdef DEBUG
for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
ASSERT(lip->li_type != XFS_LI_EFI);
#endif
break;
}
/*
* Skip EFIs that we've already processed.
*/
efip = (xfs_efi_log_item_t *)lip;
if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags)) {
lip = xfs_trans_ail_cursor_next(ailp, &cur);
continue;
}
spin_unlock(&ailp->xa_lock);
error = xlog_recover_process_efi(log->l_mp, efip);
spin_lock(&ailp->xa_lock);
if (error)
goto out;
lip = xfs_trans_ail_cursor_next(ailp, &cur);
}
out:
xfs_trans_ail_cursor_done(&cur);
spin_unlock(&ailp->xa_lock);
return error;
}
/*
* This routine performs a transaction to null out a bad inode pointer
* in an agi unlinked inode hash bucket.
*/
STATIC void
xlog_recover_clear_agi_bucket(
xfs_mount_t *mp,
xfs_agnumber_t agno,
int bucket)
{
xfs_trans_t *tp;
xfs_agi_t *agi;
xfs_buf_t *agibp;
int offset;
int error;
tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
error = xfs_trans_reserve(tp, &M_RES(mp)->tr_clearagi, 0, 0);
if (error)
goto out_abort;
error = xfs_read_agi(mp, tp, agno, &agibp);
if (error)
goto out_abort;
agi = XFS_BUF_TO_AGI(agibp);
agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
offset = offsetof(xfs_agi_t, agi_unlinked) +
(sizeof(xfs_agino_t) * bucket);
xfs_trans_log_buf(tp, agibp, offset,
(offset + sizeof(xfs_agino_t) - 1));
error = xfs_trans_commit(tp, 0);
if (error)
goto out_error;
return;
out_abort:
xfs_trans_cancel(tp, XFS_TRANS_ABORT);
out_error:
xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
return;
}
STATIC xfs_agino_t
xlog_recover_process_one_iunlink(
struct xfs_mount *mp,
xfs_agnumber_t agno,
xfs_agino_t agino,
int bucket)
{
struct xfs_buf *ibp;
struct xfs_dinode *dip;
struct xfs_inode *ip;
xfs_ino_t ino;
int error;
ino = XFS_AGINO_TO_INO(mp, agno, agino);
error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
if (error)
goto fail;
/*
* Get the on disk inode to find the next inode in the bucket.
*/
error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0, 0);
if (error)
goto fail_iput;
ASSERT(ip->i_d.di_nlink == 0);
ASSERT(ip->i_d.di_mode != 0);
/* setup for the next pass */
agino = be32_to_cpu(dip->di_next_unlinked);
xfs_buf_relse(ibp);
/*
* Prevent any DMAPI event from being sent when the reference on
* the inode is dropped.
*/
ip->i_d.di_dmevmask = 0;
IRELE(ip);
return agino;
fail_iput:
IRELE(ip);
fail:
/*
* We can't read in the inode this bucket points to, or this inode
* is messed up. Just ditch this bucket of inodes. We will lose
* some inodes and space, but at least we won't hang.
*
* Call xlog_recover_clear_agi_bucket() to perform a transaction to
* clear the inode pointer in the bucket.
*/
xlog_recover_clear_agi_bucket(mp, agno, bucket);
return NULLAGINO;
}
/*
* xlog_iunlink_recover
*
* This is called during recovery to process any inodes which
* we unlinked but not freed when the system crashed. These
* inodes will be on the lists in the AGI blocks. What we do
* here is scan all the AGIs and fully truncate and free any
* inodes found on the lists. Each inode is removed from the
* lists when it has been fully truncated and is freed. The
* freeing of the inode and its removal from the list must be
* atomic.
*/
STATIC void
xlog_recover_process_iunlinks(
struct xlog *log)
{
xfs_mount_t *mp;
xfs_agnumber_t agno;
xfs_agi_t *agi;
xfs_buf_t *agibp;
xfs_agino_t agino;
int bucket;
int error;
uint mp_dmevmask;
mp = log->l_mp;
/*
* Prevent any DMAPI event from being sent while in this function.
*/
mp_dmevmask = mp->m_dmevmask;
mp->m_dmevmask = 0;
for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
/*
* Find the agi for this ag.
*/
error = xfs_read_agi(mp, NULL, agno, &agibp);
if (error) {
/*
* AGI is b0rked. Don't process it.
*
* We should probably mark the filesystem as corrupt
* after we've recovered all the ag's we can....
*/
continue;
}
/*
* Unlock the buffer so that it can be acquired in the normal
* course of the transaction to truncate and free each inode.
* Because we are not racing with anyone else here for the AGI
* buffer, we don't even need to hold it locked to read the
* initial unlinked bucket entries out of the buffer. We keep
* buffer reference though, so that it stays pinned in memory
* while we need the buffer.
*/
agi = XFS_BUF_TO_AGI(agibp);
xfs_buf_unlock(agibp);
for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
agino = be32_to_cpu(agi->agi_unlinked[bucket]);
while (agino != NULLAGINO) {
agino = xlog_recover_process_one_iunlink(mp,
agno, agino, bucket);
}
}
xfs_buf_rele(agibp);
}
mp->m_dmevmask = mp_dmevmask;
}
/*
2012-11-12 11:54:24 +00:00
* Upack the log buffer data and crc check it. If the check fails, issue a
* warning if and only if the CRC in the header is non-zero. This makes the
* check an advisory warning, and the zero CRC check will prevent failure
* warnings from being emitted when upgrading the kernel from one that does not
* add CRCs by default.
*
* When filesystems are CRC enabled, this CRC mismatch becomes a fatal log
* corruption failure
*/
2012-11-12 11:54:24 +00:00
STATIC int
xlog_unpack_data_crc(
struct xlog_rec_header *rhead,
xfs_caddr_t dp,
struct xlog *log)
{
__le32 crc;
2012-11-12 11:54:24 +00:00
crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
if (crc != rhead->h_crc) {
if (rhead->h_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
xfs_alert(log->l_mp,
"log record CRC mismatch: found 0x%x, expected 0x%x.",
le32_to_cpu(rhead->h_crc),
le32_to_cpu(crc));
2012-11-12 11:54:24 +00:00
xfs_hex_dump(dp, 32);
}
2012-11-12 11:54:24 +00:00
/*
* If we've detected a log record corruption, then we can't
* recover past this point. Abort recovery if we are enforcing
* CRC protection by punting an error back up the stack.
*/
if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
return EFSCORRUPTED;
}
2012-11-12 11:54:24 +00:00
return 0;
}
2012-11-12 11:54:24 +00:00
STATIC int
xlog_unpack_data(
struct xlog_rec_header *rhead,
xfs_caddr_t dp,
struct xlog *log)
{
int i, j, k;
2012-11-12 11:54:24 +00:00
int error;
error = xlog_unpack_data_crc(rhead, dp, log);
if (error)
return error;
for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
dp += BBSIZE;
}
if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
dp += BBSIZE;
}
}
2012-11-12 11:54:24 +00:00
return 0;
}
STATIC int
xlog_valid_rec_header(
struct xlog *log,
struct xlog_rec_header *rhead,
xfs_daddr_t blkno)
{
int hlen;
if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
XFS_ERRLEVEL_LOW, log->l_mp);
return XFS_ERROR(EFSCORRUPTED);
}
if (unlikely(
(!rhead->h_version ||
(be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
__func__, be32_to_cpu(rhead->h_version));
return XFS_ERROR(EIO);
}
/* LR body must have data or it wouldn't have been written */
hlen = be32_to_cpu(rhead->h_len);
if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
XFS_ERRLEVEL_LOW, log->l_mp);
return XFS_ERROR(EFSCORRUPTED);
}
if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
XFS_ERRLEVEL_LOW, log->l_mp);
return XFS_ERROR(EFSCORRUPTED);
}
return 0;
}
/*
* Read the log from tail to head and process the log records found.
* Handle the two cases where the tail and head are in the same cycle
* and where the active portion of the log wraps around the end of
* the physical log separately. The pass parameter is passed through
* to the routines called to process the data and is not looked at
* here.
*/
STATIC int
xlog_do_recovery_pass(
struct xlog *log,
xfs_daddr_t head_blk,
xfs_daddr_t tail_blk,
int pass)
{
xlog_rec_header_t *rhead;
xfs_daddr_t blk_no;
xfs: Wrapped journal record corruption on read at recovery Summary of problem: If a journal record wraps at the physical end of the journal, it has to be read in two parts in xlog_do_recovery_pass(): a read at the physical end and a read at the physical beginning. If xlog_bread() has to re-align the first read, the second read request does not take that re-alignment into account. If the first read was re-aligned, the second read over-writes the end of the data from the first read, effectively corrupting it. This can happen either when reading the record header or reading the record data. The first sanity check in xlog_recover_process_data() is to check for a valid clientid, so that is the error reported. Summary of fix: If there was a first read at the physical end, XFS_BUF_PTR() returns where the data was requested to begin. Conversely, because it is the result of xlog_align(), offset indicates where the requested data for the first read actually begins - whether or not xlog_bread() has re-aligned it. Using offset as the base for the calculation of where to place the second read data ensures that it will be correctly placed immediately following the data from the first read instead of sometimes over-writing the end of it. The attached patch has resolved the reported problem of occasional inability to recover the journal (reporting "bad clientid"). Signed-off-by: Andy Poling <andy@realbig.com> Reviewed-by: Alex Elder <aelder@sgi.com> Signed-off-by: Alex Elder <aelder@sgi.com>
2009-11-03 17:26:47 +00:00
xfs_caddr_t offset;
xfs_buf_t *hbp, *dbp;
int error = 0, h_size;
int bblks, split_bblks;
int hblks, split_hblks, wrapped_hblks;
struct hlist_head rhash[XLOG_RHASH_SIZE];
ASSERT(head_blk != tail_blk);
/*
* Read the header of the tail block and get the iclog buffer size from
* h_size. Use this to tell how many sectors make up the log header.
*/
if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
/*
* When using variable length iclogs, read first sector of
* iclog header and extract the header size from it. Get a
* new hbp that is the correct size.
*/
hbp = xlog_get_bp(log, 1);
if (!hbp)
return ENOMEM;
error = xlog_bread(log, tail_blk, 1, hbp, &offset);
if (error)
goto bread_err1;
rhead = (xlog_rec_header_t *)offset;
error = xlog_valid_rec_header(log, rhead, tail_blk);
if (error)
goto bread_err1;
h_size = be32_to_cpu(rhead->h_size);
if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
(h_size > XLOG_HEADER_CYCLE_SIZE)) {
hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
if (h_size % XLOG_HEADER_CYCLE_SIZE)
hblks++;
xlog_put_bp(hbp);
hbp = xlog_get_bp(log, hblks);
} else {
hblks = 1;
}
} else {
ASSERT(log->l_sectBBsize == 1);
hblks = 1;
hbp = xlog_get_bp(log, 1);
h_size = XLOG_BIG_RECORD_BSIZE;
}
if (!hbp)
return ENOMEM;
dbp = xlog_get_bp(log, BTOBB(h_size));
if (!dbp) {
xlog_put_bp(hbp);
return ENOMEM;
}
memset(rhash, 0, sizeof(rhash));
if (tail_blk <= head_blk) {
for (blk_no = tail_blk; blk_no < head_blk; ) {
error = xlog_bread(log, blk_no, hblks, hbp, &offset);
if (error)
goto bread_err2;
rhead = (xlog_rec_header_t *)offset;
error = xlog_valid_rec_header(log, rhead, blk_no);
if (error)
goto bread_err2;
/* blocks in data section */
bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
error = xlog_bread(log, blk_no + hblks, bblks, dbp,
&offset);
if (error)
goto bread_err2;
2012-11-12 11:54:24 +00:00
error = xlog_unpack_data(rhead, offset, log);
if (error)
goto bread_err2;
error = xlog_recover_process_data(log,
rhash, rhead, offset, pass);
if (error)
goto bread_err2;
blk_no += bblks + hblks;
}
} else {
/*
* Perform recovery around the end of the physical log.
* When the head is not on the same cycle number as the tail,
* we can't do a sequential recovery as above.
*/
blk_no = tail_blk;
while (blk_no < log->l_logBBsize) {
/*
* Check for header wrapping around physical end-of-log
*/
offset = hbp->b_addr;
split_hblks = 0;
wrapped_hblks = 0;
if (blk_no + hblks <= log->l_logBBsize) {
/* Read header in one read */
error = xlog_bread(log, blk_no, hblks, hbp,
&offset);
if (error)
goto bread_err2;
} else {
/* This LR is split across physical log end */
if (blk_no != log->l_logBBsize) {
/* some data before physical log end */
ASSERT(blk_no <= INT_MAX);
split_hblks = log->l_logBBsize - (int)blk_no;
ASSERT(split_hblks > 0);
error = xlog_bread(log, blk_no,
split_hblks, hbp,
&offset);
if (error)
goto bread_err2;
}
/*
* Note: this black magic still works with
* large sector sizes (non-512) only because:
* - we increased the buffer size originally
* by 1 sector giving us enough extra space
* for the second read;
* - the log start is guaranteed to be sector
* aligned;
* - we read the log end (LR header start)
* _first_, then the log start (LR header end)
* - order is important.
*/
wrapped_hblks = hblks - split_hblks;
error = xlog_bread_offset(log, 0,
wrapped_hblks, hbp,
offset + BBTOB(split_hblks));
if (error)
goto bread_err2;
}
rhead = (xlog_rec_header_t *)offset;
error = xlog_valid_rec_header(log, rhead,
split_hblks ? blk_no : 0);
if (error)
goto bread_err2;
bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
blk_no += hblks;
/* Read in data for log record */
if (blk_no + bblks <= log->l_logBBsize) {
error = xlog_bread(log, blk_no, bblks, dbp,
&offset);
if (error)
goto bread_err2;
} else {
/* This log record is split across the
* physical end of log */
offset = dbp->b_addr;
split_bblks = 0;
if (blk_no != log->l_logBBsize) {
/* some data is before the physical
* end of log */
ASSERT(!wrapped_hblks);
ASSERT(blk_no <= INT_MAX);
split_bblks =
log->l_logBBsize - (int)blk_no;
ASSERT(split_bblks > 0);
error = xlog_bread(log, blk_no,
split_bblks, dbp,
&offset);
if (error)
goto bread_err2;
}
/*
* Note: this black magic still works with
* large sector sizes (non-512) only because:
* - we increased the buffer size originally
* by 1 sector giving us enough extra space
* for the second read;
* - the log start is guaranteed to be sector
* aligned;
* - we read the log end (LR header start)
* _first_, then the log start (LR header end)
* - order is important.
*/
error = xlog_bread_offset(log, 0,
bblks - split_bblks, dbp,
offset + BBTOB(split_bblks));
if (error)
goto bread_err2;
}
2012-11-12 11:54:24 +00:00
error = xlog_unpack_data(rhead, offset, log);
if (error)
goto bread_err2;
error = xlog_recover_process_data(log, rhash,
rhead, offset, pass);
if (error)
goto bread_err2;
blk_no += bblks;
}
ASSERT(blk_no >= log->l_logBBsize);
blk_no -= log->l_logBBsize;
/* read first part of physical log */
while (blk_no < head_blk) {
error = xlog_bread(log, blk_no, hblks, hbp, &offset);
if (error)
goto bread_err2;
rhead = (xlog_rec_header_t *)offset;
error = xlog_valid_rec_header(log, rhead, blk_no);
if (error)
goto bread_err2;
bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
error = xlog_bread(log, blk_no+hblks, bblks, dbp,
&offset);
if (error)
goto bread_err2;
2012-11-12 11:54:24 +00:00
error = xlog_unpack_data(rhead, offset, log);
if (error)
goto bread_err2;
error = xlog_recover_process_data(log, rhash,
rhead, offset, pass);
if (error)
goto bread_err2;
blk_no += bblks + hblks;
}
}
bread_err2:
xlog_put_bp(dbp);
bread_err1:
xlog_put_bp(hbp);
return error;
}
/*
* Do the recovery of the log. We actually do this in two phases.
* The two passes are necessary in order to implement the function
* of cancelling a record written into the log. The first pass
* determines those things which have been cancelled, and the
* second pass replays log items normally except for those which
* have been cancelled. The handling of the replay and cancellations
* takes place in the log item type specific routines.
*
* The table of items which have cancel records in the log is allocated
* and freed at this level, since only here do we know when all of
* the log recovery has been completed.
*/
STATIC int
xlog_do_log_recovery(
struct xlog *log,
xfs_daddr_t head_blk,
xfs_daddr_t tail_blk)
{
int error, i;
ASSERT(head_blk != tail_blk);
/*
* First do a pass to find all of the cancelled buf log items.
* Store them in the buf_cancel_table for use in the second pass.
*/
log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
sizeof(struct list_head),
KM_SLEEP);
for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
error = xlog_do_recovery_pass(log, head_blk, tail_blk,
XLOG_RECOVER_PASS1);
if (error != 0) {
kmem_free(log->l_buf_cancel_table);
log->l_buf_cancel_table = NULL;
return error;
}
/*
* Then do a second pass to actually recover the items in the log.
* When it is complete free the table of buf cancel items.
*/
error = xlog_do_recovery_pass(log, head_blk, tail_blk,
XLOG_RECOVER_PASS2);
#ifdef DEBUG
if (!error) {
int i;
for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
ASSERT(list_empty(&log->l_buf_cancel_table[i]));
}
#endif /* DEBUG */
kmem_free(log->l_buf_cancel_table);
log->l_buf_cancel_table = NULL;
return error;
}
/*
* Do the actual recovery
*/
STATIC int
xlog_do_recover(
struct xlog *log,
xfs_daddr_t head_blk,
xfs_daddr_t tail_blk)
{
int error;
xfs_buf_t *bp;
xfs_sb_t *sbp;
/*
* First replay the images in the log.
*/
error = xlog_do_log_recovery(log, head_blk, tail_blk);
xfs: on-stack delayed write buffer lists Queue delwri buffers on a local on-stack list instead of a per-buftarg one, and write back the buffers per-process instead of by waking up xfsbufd. This is now easily doable given that we have very few places left that write delwri buffers: - log recovery: Only done at mount time, and already forcing out the buffers synchronously using xfs_flush_buftarg - quotacheck: Same story. - dquot reclaim: Writes out dirty dquots on the LRU under memory pressure. We might want to look into doing more of this via xfsaild, but it's already more optimal than the synchronous inode reclaim that writes each buffer synchronously. - xfsaild: This is the main beneficiary of the change. By keeping a local list of buffers to write we reduce latency of writing out buffers, and more importably we can remove all the delwri list promotions which were hitting the buffer cache hard under sustained metadata loads. The implementation is very straight forward - xfs_buf_delwri_queue now gets a new list_head pointer that it adds the delwri buffers to, and all callers need to eventually submit the list using xfs_buf_delwi_submit or xfs_buf_delwi_submit_nowait. Buffers that already are on a delwri list are skipped in xfs_buf_delwri_queue, assuming they already are on another delwri list. The biggest change to pass down the buffer list was done to the AIL pushing. Now that we operate on buffers the trylock, push and pushbuf log item methods are merged into a single push routine, which tries to lock the item, and if possible add the buffer that needs writeback to the buffer list. This leads to much simpler code than the previous split but requires the individual IOP_PUSH instances to unlock and reacquire the AIL around calls to blocking routines. Given that xfsailds now also handle writing out buffers, the conditions for log forcing and the sleep times needed some small changes. The most important one is that we consider an AIL busy as long we still have buffers to push, and the other one is that we do increment the pushed LSN for buffers that are under flushing at this moment, but still count them towards the stuck items for restart purposes. Without this we could hammer on stuck items without ever forcing the log and not make progress under heavy random delete workloads on fast flash storage devices. [ Dave Chinner: - rebase on previous patches. - improved comments for XBF_DELWRI_Q handling - fix XBF_ASYNC handling in queue submission (test 106 failure) - rename delwri submit function buffer list parameters for clarity - xfs_efd_item_push() should return XFS_ITEM_PINNED ] Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Mark Tinguely <tinguely@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>
2012-04-23 05:58:39 +00:00
if (error)
return error;
/*
* If IO errors happened during recovery, bail out.
*/
if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
return (EIO);
}
/*
* We now update the tail_lsn since much of the recovery has completed
* and there may be space available to use. If there were no extent
* or iunlinks, we can free up the entire log and set the tail_lsn to
* be the last_sync_lsn. This was set in xlog_find_tail to be the
* lsn of the last known good LR on disk. If there are extent frees
* or iunlinks they will have some entries in the AIL; so we look at
* the AIL to determine how to set the tail_lsn.
*/
xlog_assign_tail_lsn(log->l_mp);
/*
* Now that we've finished replaying all buffer and inode
* updates, re-read in the superblock and reverify it.
*/
bp = xfs_getsb(log->l_mp, 0);
XFS_BUF_UNDONE(bp);
ASSERT(!(XFS_BUF_ISWRITE(bp)));
XFS_BUF_READ(bp);
XFS_BUF_UNASYNC(bp);
bp->b_ops = &xfs_sb_buf_ops;
if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
xfs_buf_relse(bp);
return XFS_ERROR(EIO);
}
xfs_buf_iorequest(bp);
error = xfs_buf_iowait(bp);
if (error) {
xfs_buf_ioerror_alert(bp, __func__);
ASSERT(0);
xfs_buf_relse(bp);
return error;
}
/* Convert superblock from on-disk format */
sbp = &log->l_mp->m_sb;
xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
ASSERT(xfs_sb_good_version(sbp));
xfs_buf_relse(bp);
/* We've re-read the superblock so re-initialize per-cpu counters */
xfs_icsb_reinit_counters(log->l_mp);
xlog_recover_check_summary(log);
/* Normal transactions can now occur */
log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
return 0;
}
/*
* Perform recovery and re-initialize some log variables in xlog_find_tail.
*
* Return error or zero.
*/
int
xlog_recover(
struct xlog *log)
{
xfs_daddr_t head_blk, tail_blk;
int error;
/* find the tail of the log */
if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
return error;
if (tail_blk != head_blk) {
/* There used to be a comment here:
*
* disallow recovery on read-only mounts. note -- mount
* checks for ENOSPC and turns it into an intelligent
* error message.
* ...but this is no longer true. Now, unless you specify
* NORECOVERY (in which case this function would never be
* called), we just go ahead and recover. We do this all
* under the vfs layer, so we can get away with it unless
* the device itself is read-only, in which case we fail.
*/
if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
return error;
}
/*
* Version 5 superblock log feature mask validation. We know the
* log is dirty so check if there are any unknown log features
* in what we need to recover. If there are unknown features
* (e.g. unsupported transactions, then simply reject the
* attempt at recovery before touching anything.
*/
if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
xfs_warn(log->l_mp,
"Superblock has unknown incompatible log features (0x%x) enabled.\n"
"The log can not be fully and/or safely recovered by this kernel.\n"
"Please recover the log on a kernel that supports the unknown features.",
(log->l_mp->m_sb.sb_features_log_incompat &
XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
return EINVAL;
}
xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
log->l_mp->m_logname ? log->l_mp->m_logname
: "internal");
error = xlog_do_recover(log, head_blk, tail_blk);
log->l_flags |= XLOG_RECOVERY_NEEDED;
}
return error;
}
/*
* In the first part of recovery we replay inodes and buffers and build
* up the list of extent free items which need to be processed. Here
* we process the extent free items and clean up the on disk unlinked
* inode lists. This is separated from the first part of recovery so
* that the root and real-time bitmap inodes can be read in from disk in
* between the two stages. This is necessary so that we can free space
* in the real-time portion of the file system.
*/
int
xlog_recover_finish(
struct xlog *log)
{
/*
* Now we're ready to do the transactions needed for the
* rest of recovery. Start with completing all the extent
* free intent records and then process the unlinked inode
* lists. At this point, we essentially run in normal mode
* except that we're still performing recovery actions
* rather than accepting new requests.
*/
if (log->l_flags & XLOG_RECOVERY_NEEDED) {
int error;
error = xlog_recover_process_efis(log);
if (error) {
xfs_alert(log->l_mp, "Failed to recover EFIs");
return error;
}
/*
* Sync the log to get all the EFIs out of the AIL.
* This isn't absolutely necessary, but it helps in
* case the unlink transactions would have problems
* pushing the EFIs out of the way.
*/
xfs_log_force(log->l_mp, XFS_LOG_SYNC);
xlog_recover_process_iunlinks(log);
xlog_recover_check_summary(log);
xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
log->l_mp->m_logname ? log->l_mp->m_logname
: "internal");
log->l_flags &= ~XLOG_RECOVERY_NEEDED;
} else {
xfs_info(log->l_mp, "Ending clean mount");
}
return 0;
}
#if defined(DEBUG)
/*
* Read all of the agf and agi counters and check that they
* are consistent with the superblock counters.
*/
void
xlog_recover_check_summary(
struct xlog *log)
{
xfs_mount_t *mp;
xfs_agf_t *agfp;
xfs_buf_t *agfbp;
xfs_buf_t *agibp;
xfs_agnumber_t agno;
__uint64_t freeblks;
__uint64_t itotal;
__uint64_t ifree;
int error;
mp = log->l_mp;
freeblks = 0LL;
itotal = 0LL;
ifree = 0LL;
for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
if (error) {
xfs_alert(mp, "%s agf read failed agno %d error %d",
__func__, agno, error);
} else {
agfp = XFS_BUF_TO_AGF(agfbp);
freeblks += be32_to_cpu(agfp->agf_freeblks) +
be32_to_cpu(agfp->agf_flcount);
xfs_buf_relse(agfbp);
}
error = xfs_read_agi(mp, NULL, agno, &agibp);
if (error) {
xfs_alert(mp, "%s agi read failed agno %d error %d",
__func__, agno, error);
} else {
struct xfs_agi *agi = XFS_BUF_TO_AGI(agibp);
itotal += be32_to_cpu(agi->agi_count);
ifree += be32_to_cpu(agi->agi_freecount);
xfs_buf_relse(agibp);
}
}
}
#endif /* DEBUG */