Re: [PATCH v2 00/14] Crash consistency xfstest using dm-log-writes

[Josef Bacik <josef@toxicpanda.com>]

On Fri, Sep 01, 2017 at 09:52:18AM +0300, Amir Goldstein wrote:
> [CC list, Ted]
> 
> On Thu, Aug 31, 2017 at 11:54 PM, Josef Bacik <josef@toxicpanda.com> wrote:
> > On Thu, Aug 31, 2017 at 05:02:46PM +0300, Amir Goldstein wrote:
> >> On Thu, Aug 31, 2017 at 4:43 PM, Josef Bacik <josef@toxicpanda.com> wrote:
> >> > On Thu, Aug 31, 2017 at 03:48:44PM +0300, Amir Goldstein wrote:
> >> >>
> >> >> Josef,
> >> >>
> >> >> I am at lost with these log corruptions.
> >> >> I see log entry bios submitted and log_end_io report success,
> >> >> but then in the log I see old data on disk where that entry should be.
> >> >> This happens quite randomly and I assume it also happens on
> >> >> logged data, because tests sometime fail on checksum on ext4.
> >> >>
> >> >> Mean while I added some more log entry sanity checks and debug
> >> >> prints to replay-log to debug the corruption:
> >> >> https://github.com/amir73il/xfstests/commit/bb946deb0dc285867be394613ddb19ce281392cc
> >> >>
> >> >> This only happens to me when running in kvm, so maybe something
> >> >> with the virtio devices is fishy.
> >> >>
> >> >> Anyway, I ran out of time to work on this for now, so if you have
> >> >> any ideas and/or time to test this issue, let me know.
> >> >>
> >> >
> ...
> >>
> >
> > Alright I tested it and it's working fine for me.  I'm creating three lv's and
> > then doing
> >
> > -drive file=/dev/mapper/whatever,format=raw,cache=none,if=virtio,aio=native
> >
> > And I get /dev/vd[bcd] which I use for my test/scratch/log dev and it works out
> > fine.  What is your -drive option line and I'll duplicate what you are doing.
> > Thanks,
> >
> 
> I am using Ted's kvm-xfstests, so this is the qemu command line:
> https://github.com/tytso/xfstests-bld/blob/master/kvm-xfstests/kvm-xfstests#L104
> 
> The only difference in -drive command is no aio=native.
> BINGO! when I add aio-native there are no more log corruptions :)
> Please try to use aio=threads to see if you also get log corruptions.
> 
> Thing is we cannot change kvm-xfstests to always use aio=native because
> it is not recommended for sparse images:
> https://access.redhat.com/articles/41313
> I will try to work something out so that kvm-xfstest will use aio=native
> when using the recommended (by not default) LV setup.
> 
> However, why would aio=threads cause log corruption?
> Does it indicate a bug in kvm-qemu or in dm-log-writes??
> 
> Did you try to use kvm-xfstests? its quite convenient to deploy in masses,
> so I think it would be ideal to integrate crash tests with.
> It also helps unifying the environment between us fs developers
> when a bug can not be reproduced on another system. see:
> https://github.com/tytso/xfstests-bld/blob/master/Documentation/kvm-xfstests.md
> 
> Anyway, if you do end up using kvm-xfstests, you'l need this
> small patch to automatically define the log-writes device:
> 
> --- a/kvm-xfstests/test-appliance/files/root/runtests.sh
> +++ b/kvm-xfstests/test-appliance/files/root/runtests.sh
> @@ -269,9 +269,11 @@ do
>             if test "$SIZE" = "large" ; then
>                 export SCRATCH_DEV=$LG_SCR_DEV
>                 export SCRATCH_MNT=$LG_SCR_MNT
> +               export LOGWRITES_DEV=$SM_SCR_DEV
>             else
>                 export SCRATCH_DEV=$SM_SCR_DEV
>                 export SCRATCH_MNT=$SM_SCR_MNT
> +               export LOGWRITES_DEV=$LG_SCR_DEV
>             fi
>         fi
> 
> kvm-xfstests defined 2 sets of test/scratch a small and a large set
> and uses only one of those sets depending on command line,
> so I use the "other" scratch as the log writes device.
>

Cool I didn't know about kvm-xfstests, I'll give that a whirl.  The baby just
woke me up but when I get up for real I'll switch my config to use aio=threads
and see what happens, but I'm starting to suspect there's a bug in qemu.
Thanks,

Josef