hast and zfs trim possibly causing some problems in 9.2

Mikolaj Golub trociny at FreeBSD.org
Thu Oct 10 18:09:42 UTC 2013


On Wed, Oct 09, 2013 at 03:47:29PM +0100, Steven Hartland wrote:

> ZFS will try to send DELETE requests to the underlying storage to
> support TRIM. If that fails then it will disable TRIM support for
> that vdev.
> 
> My guess would be you're just seeing hast being a bit verbose
> when these initial batch failures happen.

If the device on the secondary node does not supports DELETE, but the
device on the primary does, HAST will report to ZFS that DELETE
succeeded (although it failed on the secondary), and ZFS will not
disable TRIM. Pete, isn't this your case?

> From: "Pete French" <petefrench at ingresso.co.uk>
> 
> >I just had a machine fall over on my for the first time in ages - one
> > of a pair of machine we have running hast with zfs on top. I havent
> > got any concrete evidence of what made it die as yet, but I
> > did notice the logifles filling up with thoursands of lines like this
> > just prior to the crash:
> > 
> > serpentine-active hastd[1522]: [serp1] (primary) Remote request failed (Operation not supported): DELETE(26847744000, 1536).
> > 
> > so I am guessing taht is ZFS trying to send a trim command to hast, and hast
> > does not support it. Have disabled zfs trim now, but thought it was
> > worth mentioning - I would have not expected zfs to be trying to issue
> > a trim command to an underlying device which doesnt support it. These
> > machines were rock solid under 8, and the only chnage I can see with 9 is
> > the trim support being added.

Another important change that comes to mind is the default replication
mode, changed from fullsync to memsync. Do you have the replication
mode explicitly set in your config?

-- 
Mikolaj Golub


More information about the freebsd-stable mailing list