zfs_trim_enabled destroys zio_free() performance
Steven Hartland
steven at multiplay.co.uk
Sun Sep 13 12:55:14 UTC 2015
On 11/09/2015 17:07, Matthew Ahrens wrote:
> I discovered that when destroying a ZFS snapshot, we can end up using
> several seconds of CPU via this stack trace:
>
> kernel`spinlock_exit+0x2d
> kernel`taskqueue_enqueue+0x12c
> zfs.ko`zio_issue_async+0x7c
> zfs.ko`zio_execute+0x162
> zfs.ko`dsl_scan_free_block_cb+0x15f
> zfs.ko`bpobj_iterate_impl+0x25d
> zfs.ko`bpobj_iterate_impl+0x46e
> zfs.ko`dsl_scan_sync+0x152
> zfs.ko`spa_sync+0x5c1
> zfs.ko`txg_sync_thread+0x3a6
> kernel`fork_exit+0x9a
> kernel`0xffffffff80d0acbe
> 6558 ms
>
> This is not good for performance since, in addition to the CPU cost, it
> doesn't allow the sync thread to do anything else, and this is observable
> as periods where we don't do any write i/o to disk for several seconds.
>
> The problem is that when zfs_trim_enabled is set (which it is by default),
> zio_free_sync() always sets ZIO_STAGE_ISSUE_ASYNC, causing the free to be
> dispatched to a taskq. Since each task completes very quickly, there is a
> large locking and context switching overhead -- we would be better off just
> processing the free in the caller's context.
>
> I'm not sure exactly why we need to go async when trim is enabled, but it
> seems like at least we should not bother going async if trim is not
> actually being used (e.g. with an all-spinning-disk pool). It would also
> be worth investigating not going async even when trim is useful (e.g. on
> SSD-based pools).
>
> Here is the relevant code:
>
> zio_free_sync():
> if (zfs_trim_enabled)
> stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
> ZIO_STAGE_VDEV_IO_ASSESS;
> /*
> * GANG and DEDUP blocks can induce a read (for the gang block
> header,
> * or the DDT), so issue them asynchronously so that this thread is
> * not tied up.
> */
> else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
> stage |= ZIO_STAGE_ISSUE_ASYNC;
TRIM requests are queued, combined and only actioned after time in the
TRIM thread as they are quite expensive which why I believe it was
thought async was required, however given all this will do is trigger a
call to trim_map_free for leaf vdev's which will be either:
1. A no-op if vdev_notrim is set (spinning rust)
2. An insert into the trim AVL
The processing of the zio should always be quick I don't see why we
couldn't execute it sync.
I've set a test going on my head box removing ZIO_STAGE_ISSUE_ASYNC to
see if I get any strange behaviour.
Regards
Steve
More information about the freebsd-fs
mailing list