On Mon, Jul 06, 2015 at 02:26:34PM -0400, Brian Foster wrote:
> Persistent memory without block translation table (btt) support provides
> a block device suitable for filesystems, but does not provide the sector
> write atomicity guarantees typical of block storage. This is a problem
> for log recovery on XFS. The on-disk log record structure already
> includes a CRC and thus can detect torn writes. The problem is that such
> a failure isn't detected until log recovery is already in progress and
> therefore results in a hard error and mount failure.
>
> Update the log head/tail discovery algorithm to detect and trim off a
> torn log record from the end (from a recovery perspective) of the log.
> Once the head is determined from the log cycle information and we have a
> pointer to the last record to be recovered, read and verify the CRC of
> said record before we initiate actual recovery. If CRC verification
> fails, assume the record is torn and reset the head to the start of the
> torn record. We reverse seek for the previous log record header from
> that point and attempt recovery up through that record.
>
> Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx>
> ---
>
> Hi all,
>
> As far as I'm aware, the torn log write problem is the only major gap we
> have to safely run XFS w/ crc=1 on non-btt pmem devices. This is an RFC
> to attempt to address that problem. I used a custom hack to actually
> reproduce such torn writes on a ramdisk to primarily help me understand
> the problem, but I was able to use the same hack to sanity test this
> approach under the basic corruption case (this is generally untested,
> otherwise).
>
> This could obviously use some cleanup, but is the approach sane? I'm
> curious if this should also be tied to DAX enablement so as to not
> interfere with unrelated corruption handling (IOW, with what logic is it
> reasonable enough to assume a crc failure of the final record == torn
> write?). I'm also wondering if this should be tied to a new mount option
> rather than make this behavior implicit. Any other thoughts?
Seems good fom a conceptual point of view - the fact that we CRC all
log writes now means this will work on both v4 and v5 filesystems.
The CRC validation needs to be done on more than just the head
record. We can have up to 8 IOs in flight at the time power goes
down, so we need to validate the CRCs for at least the previous 8
log writes...
We also need to validate the tail record, as we could be in a
situation where the head is overwriting the tail and it tears and
so by discarding the head record we discard the tail update and so
now the tail record is also corrupt. In that case, we need to abort
recovery because the log is unrecoverable.
Also: should we zero out the torn sections before starting recovery,
like we do with the call to xlog_clear_stale_blocks() later in
xlog_find_tail()?
> diff --git a/fs/xfs/xfs_log_recover.c b/fs/xfs/xfs_log_recover.c
> index 01dd228..6015d02 100644
> --- a/fs/xfs/xfs_log_recover.c
> +++ b/fs/xfs/xfs_log_recover.c
> @@ -62,6 +62,9 @@ xlog_recover_check_summary(
> #define xlog_recover_check_summary(log)
> #endif
>
> +STATIC int
> +xlog_validate_logrec_crc(struct xlog *, xfs_daddr_t);
> +
Put the function here, don't use forward declarations....
> /*
> * This structure is used during recovery to record the buf log items which
> * have been canceled and should not be replayed.
> @@ -898,6 +901,7 @@ xlog_find_tail(
> xfs_daddr_t after_umount_blk;
> xfs_lsn_t tail_lsn;
> int hblks;
> + bool skipped_last = false;
>
> found = 0;
>
> @@ -925,6 +929,7 @@ xlog_find_tail(
> /*
> * Search backwards looking for log record header block
> */
> +retry:
> ASSERT(*head_blk < INT_MAX);
> for (i = (int)(*head_blk) - 1; i >= 0; i--) {
> error = xlog_bread(log, i, 1, bp, &offset);
> @@ -962,6 +967,31 @@ xlog_find_tail(
> return -EIO;
> }
>
> + /*
> + * Now that we think we've found the log head, we have to check that the
> + * last log record wasn't torn when written out to the log. This is
> + * possible on devices without sector atomicity guarantees (e.g., pmem).
> + *
> + * Verify the CRC of the last log record that was written. If the CRC is
> + * invalid, point the head at the start of this record and retry the
> + * above backwards log record header search. We'll try the recovery up
> + * through this record. Note that we only walk backwards once since this
> + * is only intended to handle the torn write on power loss case.
> + *
> + * TODO: mount option? tied to DAX?
> + */
Always do it - if any hardware is tearing writes, then we should
handle it sanely.
> + if (xfs_sb_version_hascrc(&log->l_mp->m_sb) && !skipped_last) {
No need to be tied to xfs_sb_version_hascrc(), even v4 filesystems
now have a CRC calculated on the log. Only skip it on v4 filesystes
when the h_crc field is zero (indicating the log came from an old
kernel that didn't calculate CRCs).
> + error = xlog_validate_logrec_crc(log, i);
> + if (error == -EFSBADCRC) {
> + skipped_last = true;
> + *head_blk = i;
> + xfs_warn(log->l_mp,
> + "WARNING: Torn write? Attempting recovery up to previous record.");
> + goto retry;
> + } else if (error)
> + goto done;
> + }
Better to do a loop and factor the code a bit?
> +
> /* find blk_no of tail of log */
> rhead = (xlog_rec_header_t *)offset;
> *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
> @@ -4607,6 +4637,63 @@ xlog_recover_finish(
> return 0;
> }
>
> +/*
> + * Read and CRC validate a full log record. This is used to detect torn log
> + * writes during head/tail discovery.
> + */
> +STATIC int
> +xlog_validate_logrec_crc(
> + struct xlog *log,
> + xfs_daddr_t rec_blk)
> +{
> + int hblks, bblks;
> + struct xlog_rec_header *rhead;
> + struct xfs_buf *hbp = NULL;
> + struct xfs_buf *dbp = NULL;
> + int error;
> + char *offset;
> +
> + hblks = 1; /* XXX */
> +
> + /* read and validate the log record header */
> + hbp = xlog_get_bp(log, 1);
> + if (!hbp)
> + return -ENOMEM;
> + error = xlog_bread(log, rec_blk, hblks, hbp, &offset);
> + if (error)
> + goto out;
> +
> + rhead = (struct xlog_rec_header *) offset;
> +
> + error = xlog_valid_rec_header(log, rhead, rec_blk);
> + if (error)
> + goto out;
> +
> + /* read the full record and verify the CRC */
> + /* XXX: factor out from do_recovery_pass() ? */
Yes, please do ;)
> + dbp = xlog_get_bp(log, BTOBB(be32_to_cpu(rhead->h_size)));
> + if (!dbp)
> + goto out;
> +
> + bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
> + error = xlog_bread(log, rec_blk+hblks, bblks, dbp, &offset);
> + if (error)
> + goto out;
> +
> + /*
> + * If the CRC validation fails, convert the return code so the caller
> + * can distinguish from unrelated errors.
> + */
> + error = xlog_unpack_data_crc(rhead, offset, log);
> + if (error)
> + error = -EFSBADCRC;
Note that this will only return an error on v5 filesystems so some
changes will be needed here to handle v4 filesystems. It will
also log a CRC mismatch error, so perhaps we want to make the error
reporting in this case a little bit less noisy....
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
_______________________________________________
xfs mailing list
xfs@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/xfs
