On Fri, Sep 11, 2015 at 02:55:32PM -0400, Brian Foster wrote:
> The LSN validation helper is called in the I/O verifier codepath for
> metadata that embed a last-modification LSN. While the codepath exists,
> this is not used in userspace as in the kernel because the former
> doesn't have an active log.
>
> xfs_repair does need to check the validity of the LSN metadata with
> respect to the on-disk log, however. Use the LSN validation mechanism to
> track the largest LSN that has been seen. Export the value so repair can
> use it once it has processed the entire filesystem. Note that the helper
> continues to always return true to preserve existing behavior.
>
> Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx>
....
> +xfs_lsn_t libxfs_max_lsn = 0;
> +pthread_mutex_t libxfs_max_lsn_lock = PTHREAD_MUTEX_INITIALIZER;
> +
> void
> xfs_detect_invalid_lsn(
> struct xfs_mount *mp,
> xfs_lsn_t lsn)
> {
> + int cycle = CYCLE_LSN(lsn);
> + int block = BLOCK_LSN(lsn);
> + int max_cycle;
> + int max_block;
> +
> + if (lsn == NULLCOMMITLSN)
> + return;
> +
> + pthread_mutex_lock(&libxfs_max_lsn_lock);
> +
> + max_cycle = CYCLE_LSN(libxfs_max_lsn);
> + max_block = BLOCK_LSN(libxfs_max_lsn);
> +
> + if ((cycle > max_cycle) ||
> + (cycle == max_cycle && block > max_block))
> + libxfs_max_lsn = lsn;
> +
> + pthread_mutex_unlock(&libxfs_max_lsn_lock);
This will have the same lock contention problems that the kernel
code would have had - my repair scalablity tests regularly reach
over 1GB/s of metadata being prefetched through tens of threads, so
this is going have a significant impact on performance in those
tests....
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
|