On 12/12/2013 02:22 AM, Dave Chinner wrote:
> From: Dave Chinner <dchinner@xxxxxxxxxx>
>
> It's possible to have filesystems with hundreds of AGs on systems
> with little concurrency and resources. In this case, we can easily
> exhaust memory and fail to create threads and have all sorts of
> interesting problems.
>
> xfs/250 can cause this to occur, with failures like:
>
> - agno = 707
> - agno = 692
> fatal error -- cannot create worker threads, error = [11] Resource
> temporarily unavailable
>
> And this:
>
> - agno = 484
> - agno = 782
> failed to create prefetch thread: Resource temporarily unavailable
>
> Because it's trying to create more threads than a poor little 512MB
> single CPU ia32 box can handle.
>
> So, limit concurrency to a maximum of numcpus * 8 to prevent this.
>
> Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx>
> ---
Reviewed-by: Brian Foster <bfoster@xxxxxxxxxx>
> include/libxfs.h | 1 +
> libxfs/init.h | 1 -
> repair/xfs_repair.c | 18 +++++++++++++++++-
> 3 files changed, 18 insertions(+), 2 deletions(-)
>
> diff --git a/include/libxfs.h b/include/libxfs.h
> index 4bf331c..39e3d85 100644
> --- a/include/libxfs.h
> +++ b/include/libxfs.h
> @@ -144,6 +144,7 @@ extern void libxfs_device_close (dev_t);
> extern int libxfs_device_alignment (void);
> extern void libxfs_report(FILE *);
> extern void platform_findsizes(char *path, int fd, long long *sz, int *bsz);
> +extern int platform_nproc(void);
>
> /* check or write log footer: specify device, log size in blocks & uuid */
> typedef xfs_caddr_t (libxfs_get_block_t)(xfs_caddr_t, int, void *);
> diff --git a/libxfs/init.h b/libxfs/init.h
> index f0b8cb6..112febb 100644
> --- a/libxfs/init.h
> +++ b/libxfs/init.h
> @@ -31,7 +31,6 @@ extern char *platform_findrawpath (char *path);
> extern char *platform_findblockpath (char *path);
> extern int platform_direct_blockdev (void);
> extern int platform_align_blockdev (void);
> -extern int platform_nproc(void);
> extern unsigned long platform_physmem(void); /* in kilobytes */
> extern int platform_has_uuid;
>
> diff --git a/repair/xfs_repair.c b/repair/xfs_repair.c
> index 7beffcb..0d006ae 100644
> --- a/repair/xfs_repair.c
> +++ b/repair/xfs_repair.c
> @@ -627,13 +627,29 @@ main(int argc, char **argv)
> * to target these for an increase in thread count. Hence a stride value
> * of 15 is chosen to ensure we get at least 2 AGs being scanned at once
> * on such filesystems.
> + *
> + * Limit the maximum thread count based on the available CPU power that
> + * is available. If we use too many threads, we might run out of memory
> + * and CPU power before we run out of IO concurrency.
> */
> if (!ag_stride && glob_agcount >= 16 && do_prefetch)
> ag_stride = 15;
>
> if (ag_stride) {
> + int max_threads = platform_nproc() * 8;
> +
> thread_count = (glob_agcount + ag_stride - 1) / ag_stride;
> - thread_init();
> + while (thread_count > max_threads) {
> + ag_stride *= 2;
> + thread_count = (glob_agcount + ag_stride - 1) /
> + ag_stride;
> + }
> + if (thread_count > 0)
> + thread_init();
> + else {
> + thread_count = 1;
> + ag_stride = 0;
> + }
> }
>
> if (ag_stride && report_interval) {
>
|