freebsd stable - Aug 2016 - zfs recv causes nfs server to throw NFSERR_IO i/o errors

Hi everyone,

we've been tracing an issue where snapshot replication is causing
interruptions for the NFS serivce.

The problem is as follows:

Every time a zfs recv finishes, there is a chance for the NFS server to
return an NFSERR_IO for a GETATTR call. This shows up as input/output
errors on the nfs clients.

Here the tcpdump showing the NFS conversation:
https://nopaste.me/view/95d1a79d

NFS 202 V3 GETATTR Call (Reply In 6043), FH: 0x8c711a60
NFS 98  V3 GETATTR Reply (Call In 6042) Error: NFS3ERR_IO
NFS 222 V3 LOOKUP Call (Reply In 6046), DH: 0x6694634f/example.file.txt
NFS 102 V3 LOOKUP Reply (Call In 6045) Error: NFS3ERR_ACCES

We've been able to verify that there is a _direct_ correlation between
the zfs recv command and these NFS errors. For every input/output error
we can find a log entry of a replication just finishing (zfs recv exiting).

The receiving server is running 10.3-RELEASE

I've read about a VFS/ZFS deadlock issue which is to be included/fixed
in Freebsd 11.0-BETA4.

Could our issue be related?
Otherwise does anyone have any suggestions how to tackle this issue?


For the record, say we have two volumes:
tank/volumeA and tank/volumeB

If there is a zfs recv busy for tank/volumeA then tank/volumeB can get
these NFS "io" errors, it does not have to be the same volume.


Has anyone else seen/experience this as well?

Any insights are appreciated!

With kind regards,

Daniel

On Fri, Aug 5, 2016 at 7:22 AM, Daniel Genis <daniel.genis at gmx.de>
wrote:
> Hi everyone,
>
> we've been tracing an issue where snapshot replication is causing
> interruptions for the NFS serivce.
>
> The problem is as follows:
>
> Every time a zfs recv finishes, there is a chance for the NFS server to
> return an NFSERR_IO for a GETATTR call. This shows up as input/output
> errors on the nfs clients.
>
> Here the tcpdump showing the NFS conversation:
> https://nopaste.me/view/95d1a79d
>
> NFS 202 V3 GETATTR Call (Reply In 6043), FH: 0x8c711a60
> NFS 98  V3 GETATTR Reply (Call In 6042) Error: NFS3ERR_IO
> NFS 222 V3 LOOKUP Call (Reply In 6046), DH: 0x6694634f/example.file.txt
> NFS 102 V3 LOOKUP Reply (Call In 6045) Error: NFS3ERR_ACCES
>
> We've been able to verify that there is a _direct_ correlation between
> the zfs recv command and these NFS errors. For every input/output error
> we can find a log entry of a replication just finishing (zfs recv exiting).
>
> The receiving server is running 10.3-RELEASE
>
> I've read about a VFS/ZFS deadlock issue which is to be included/fixed
> in Freebsd 11.0-BETA4.
>
> Could our issue be related?
> Otherwise does anyone have any suggestions how to tackle this issue?
>
>
> For the record, say we have two volumes:
> tank/volumeA and tank/volumeB
>
> If there is a zfs recv busy for tank/volumeA then tank/volumeB can get
> these NFS "io" errors, it does not have to be the same volume.
>
>
> Has anyone else seen/experience this as well?
>
> Any insights are appreciated!
>
> With kind regards,
>
> Daniel
Try adding mountd_flags="-S" to /etc/rc.conf.