On 3/24/16 6:15 AM, jtulak@xxxxxxxxxx wrote:
> From: Dave Chinner <dchinner@xxxxxxxxxx>
>
> CHANGELOG:
> o Remove nonexistent headers from LIBHFILES in include/Makefile
> o Remove a useless assignment which was immediately overwritten
> on the next line.
> o Rename include/xfs_mkfs.h global header to include/xfs_multidisk.h,
> because we need it just for multidisk configuration
> o Fix AG count for size thresholds to keep consistency
>
> Both xfs_repair and mkfs.xfs need to agree on what is a "multidisk:
> configuration - mkfs for determining the AG count of the filesystem,
> repair for determining how to automatically parallelise it's
> execution. This requires a bunch of common defines that both mkfs
> and reapir need to share.
>
> In fact, most of the defines in xfs_mkfs.h could be shared with
> other programs (i.e. all the defaults mkfs uses) and so it is
> simplest to move xfs_mkfs.h to the shared include directory and add
> the new defines to it directly.
Minor comment stuff below, otherwise seems ok.
Reviewed-by: Eric Sandeen <sandeen@xxxxxxxxxx>
and feel free to add that if you make the suggested changes, too.
> diff --git a/include/xfs_multidisk.h b/include/xfs_multidisk.h
> new file mode 100644
> index 0000000..4006a01
> --- /dev/null
> +++ b/include/xfs_multidisk.h
<big snip this is mostly just a file rename, but...>
> +/*
> + * These values define what we consider a "multi-disk" filesystem. That is, a
> + * filesystem that is likely to be made up of multiple devices, and hence
> have
> + * some level of parallelism avoid to it at the IO level.
"avoid to it?" Maybe "available to it?"
...
> @@ -664,43 +664,45 @@ calc_default_ag_geometry(
> }
>
> /*
> - * For the remainder we choose an AG size based on the
> - * number of data blocks available, trying to keep the
> - * number of AGs relatively small (especially compared
> - * to the original algorithm). AG count is calculated
> - * based on the preferred AG size, not vice-versa - the
> - * count can be increased by growfs, so prefer to use
> - * smaller counts at mkfs time.
> - *
> - * For a single underlying storage device between 128MB
> - * and 4TB in size, just use 4 AGs, otherwise scale up
> - * smoothly between min/max AG sizes.
> + * For a single underlying storage device between 128MB and 4TB in size
> + * just use 4 AGs and scale up smoothly between min/max AG sizes.
I guess a comment about the *first* >= 4T case might make this more
clear; it's a little odd to document only 1 of the 2 cases below.
Maybe:
+ * For a single underlying storage device over 4TB in size
+ * use the maximum AG size. Between 128MB and 4TB, just use
+ * 4 AGs and scale up smoothly between min/max AG sizes.
> */
> -
> - if (!multidisk && dblocks >= MEGABYTES(128, blocklog)) {
> + if (!multidisk) {
> if (dblocks >= TERABYTES(4, blocklog)) {
> blocks = XFS_AG_MAX_BLOCKS(blocklog);
> goto done;
> + } else if (dblocks >= MEGABYTES(128, blocklog)) {
> + shift = XFS_NOMULTIDISK_AGLOG;
> + goto calc_blocks;
> }
> - shift = 2;
> - } else if (dblocks > GIGABYTES(512, blocklog))
> - shift = 5;
> - else if (dblocks > GIGABYTES(8, blocklog))
> - shift = 4;
> - else if (dblocks >= MEGABYTES(128, blocklog))
> - shift = 3;
> - else if (dblocks >= MEGABYTES(64, blocklog))
> - shift = 2;
> - else if (dblocks >= MEGABYTES(32, blocklog))
> - shift = 1;
> - else
> - shift = 0;
> + }
> +
> + /*
> + * For the multidisk configs we choose an AG count based on the number
> + * of data blocks available, trying to keep the number of AGs higher
> + * than the single disk configurations. This makes the assumption that
> + * larger filesystems have more parallelism available to them.
> + */
> + shift = XFS_MULTIDISK_AGLOG;
> + if (dblocks <= GIGABYTES(512, blocklog))
> + shift--;
> + if (dblocks <= GIGABYTES(8, blocklog))
> + shift--;
> + if (dblocks < MEGABYTES(128, blocklog))
> + shift--;
> + if (dblocks < MEGABYTES(64, blocklog))
> + shift--;
> + if (dblocks < MEGABYTES(32, blocklog))
> + shift--;
> +
> /*
> * If dblocks is not evenly divisible by the number of
> * desired AGs, round "blocks" up so we don't lose the
> * last bit of the filesystem. The same principle applies
> * to the AG count, so we don't lose the last AG!
> */
> +calc_blocks:
> + ASSERT(shift >= 0 && shift <= XFS_MULTIDISK_AGLOG);
> blocks = dblocks >> shift;
> if (dblocks & xfs_mask32lo(shift)) {
> if (blocks < XFS_AG_MAX_BLOCKS(blocklog))
> diff --git a/repair/xfs_repair.c b/repair/xfs_repair.c
> index 5d5f3aa..9d91f2d 100644
> --- a/repair/xfs_repair.c
> +++ b/repair/xfs_repair.c
> @@ -19,6 +19,7 @@
> #include "libxfs.h"
> #include "libxlog.h"
> #include <sys/resource.h>
> +#include "xfs_multidisk.h"
> #include "avl.h"
> #include "avl64.h"
> #include "globals.h"
> @@ -589,6 +590,33 @@ format_log_max_lsn(
> XLOG_FMT, new_cycle, true);
> }
>
> +/*
> + * mkfs increases the AG count for "multidisk" configurations, we want
> + * to target these for an increase in thread count. Hence check the superlock
> + * geometry information to determine if mkfs considered this a multidisk
> + * configuration.
> + */
> +static bool
> +is_multidisk_filesystem(
> + struct xfs_mount *mp)
> +{
> + struct xfs_sb *sbp = &mp->m_sb;
> +
> + /* High agcount filesystems are always considered "multidisk" */
> + if (sbp->sb_agcount >= XFS_MULTIDISK_AGCOUNT)
> + return true;
> +
> + /*
> + * If it doesn't have a sunit/swidth, mkfs didn't consider it a
> + * multi-disk array, so we don't either.
> + */
> + if (!sbp->sb_unit)
> + return false;
> +
> + ASSERT(sbp->sb_width);
> + return true;
> +}
> +
> int
> main(int argc, char **argv)
> {
> @@ -729,9 +757,21 @@ main(int argc, char **argv)
> * threads/CPU as this is enough threads to saturate a CPU on fast
> * devices, yet few enough that it will saturate but won't overload slow
> * devices.
> + *
> + * Multidisk filesystems can handle more IO parallelism so we should try
> + * to process multiple AGs at a time in such a configuration to try to
> + * saturate the underlying storage and speed the repair process. Only do
> + * this if prefetching is enabled.
> */
> - if (!ag_stride && glob_agcount >= 16 && do_prefetch)
> - ag_stride = 15;
> + if (!ag_stride && do_prefetch && is_multidisk_filesystem(mp)) {
> + /*
> + * For small agcount multidisk systems, just double the
> + * parallelism. For larger AG count filesystems (32 and above)
> + * use more parallelism, and linearly increase the parallelism
> + * with the number of AGs.
> + */
> + ag_stride = min(glob_agcount, XFS_MULTIDISK_AGCOUNT / 2) - 1;
> + }
>
> if (ag_stride) {
> int max_threads = platform_nproc() * 8;
>
|