Linux Virtualization - Jan 2020 - vhost changes (batched) in linux-next after 12/13 trigger random crashes in KVM guests after reboot

On 18.12.19 16:10, Michael S. Tsirkin wrote:
> On Wed, Dec 18, 2019 at 03:43:43PM +0100, Christian Borntraeger wrote:
>> Michael,
>>
>> with 
>> commit db7286b100b503ef80612884453bed53d74c9a16
(refs/bisect/skip-db7286b100b503ef80612884453bed53d74c9a16)
>>     vhost: use batched version by default
>> plus
>> commit 6bd262d5eafcdf8cdfae491e2e748e4e434dcda6 (HEAD, refs/bisect/bad)
>>     Revert "vhost/net: add an option to test new code"
>> to make things compile (your next tree is not easily bisectable, can
you fix that as well?).
> 
> I'll try.
> 
>>
>> I get random crashes in my s390 KVM guests after reboot.
>> Reverting both patches together with commit decd9b8 "vhost: use
vhost_desc instead of vhost_log" to
>> make it compile again) on top of linux-next-1218 makes the problem go
away.
>>
>> Looks like the batched version is not yet ready for prime time. Can you
drop these patches until
>> we have fixed the issues?
>>
>> Christian
>>
> 
> Will do, thanks for letting me know.
I have confirmed with the initial reporter (internal test team) that <driver
name='qemu'/>
with a known to be broken linux next kernel also fixes the problem, so it is
really the
vhost changes.

On Wed, Dec 18, 2019 at 04:59:02PM +0100, Christian Borntraeger
wrote:
> On 18.12.19 16:10, Michael S. Tsirkin wrote:
> > On Wed, Dec 18, 2019 at 03:43:43PM +0100, Christian Borntraeger wrote:
> >> Michael,
> >>
> >> with 
> >> commit db7286b100b503ef80612884453bed53d74c9a16
(refs/bisect/skip-db7286b100b503ef80612884453bed53d74c9a16)
> >>     vhost: use batched version by default
> >> plus
> >> commit 6bd262d5eafcdf8cdfae491e2e748e4e434dcda6 (HEAD,
refs/bisect/bad)
> >>     Revert "vhost/net: add an option to test new code"
> >> to make things compile (your next tree is not easily bisectable,
can you fix that as well?).
> > 
> > I'll try.
> > 
> >>
> >> I get random crashes in my s390 KVM guests after reboot.
> >> Reverting both patches together with commit decd9b8 "vhost:
use vhost_desc instead of vhost_log" to
> >> make it compile again) on top of linux-next-1218 makes the problem
go away.
> >>
> >> Looks like the batched version is not yet ready for prime time.
Can you drop these patches until
> >> we have fixed the issues?
> >>
> >> Christian
> >>
> > 
> > Will do, thanks for letting me know.
> 
> I have confirmed with the initial reporter (internal test team) that
<driver name='qemu'/>
> with a known to be broken linux next kernel also fixes the problem, so it
is really the
> vhost changes.
OK I'm back and trying to make it more bisectable.

I pushed a new tag "batch-v2".
It's same code but with this bisect should get more information.


I suspect one of the following:

commit 1414d7ee3d10d2ec2bc4ee652d1d90ec91da1c79
Author: Michael S. Tsirkin <mst at redhat.com>
Date:   Mon Oct 7 06:11:18 2019 -0400

    vhost: batching fetches
    
    With this patch applied, new and old code perform identically.
    
    Lots of extra optimizations are now possible, e.g.
    we can fetch multiple heads with copy_from/to_user now.
    We can get rid of maintaining the log array.  Etc etc.
    
    Signed-off-by: Michael S. Tsirkin <mst at redhat.com>

commit 50297a8480b439efc5f3f23088cb2d90b799acef
Author: Michael S. Tsirkin <mst at redhat.com>
Date:   Wed Dec 11 12:19:26 2019 -0500

    vhost: use batched version by default
    
    As testing shows no performance change, switch to that now.
    
    Signed-off-by: Michael S. Tsirkin <mst at redhat.com>


and would like to know which.

Thanks!

On 06.01.20 11:50, Michael S. Tsirkin wrote:
> On Wed, Dec 18, 2019 at 04:59:02PM +0100, Christian Borntraeger wrote:
>> On 18.12.19 16:10, Michael S. Tsirkin wrote:
>>> On Wed, Dec 18, 2019 at 03:43:43PM +0100, Christian Borntraeger
wrote:
>>>> Michael,
>>>>
>>>> with 
>>>> commit db7286b100b503ef80612884453bed53d74c9a16
(refs/bisect/skip-db7286b100b503ef80612884453bed53d74c9a16)
>>>>     vhost: use batched version by default
>>>> plus
>>>> commit 6bd262d5eafcdf8cdfae491e2e748e4e434dcda6 (HEAD,
refs/bisect/bad)
>>>>     Revert "vhost/net: add an option to test new
code"
>>>> to make things compile (your next tree is not easily
bisectable, can you fix that as well?).
>>>
>>> I'll try.
>>>
>>>>
>>>> I get random crashes in my s390 KVM guests after reboot.
>>>> Reverting both patches together with commit decd9b8
"vhost: use vhost_desc instead of vhost_log" to
>>>> make it compile again) on top of linux-next-1218 makes the
problem go away.
>>>>
>>>> Looks like the batched version is not yet ready for prime time.
Can you drop these patches until
>>>> we have fixed the issues?
>>>>
>>>> Christian
>>>>
>>>
>>> Will do, thanks for letting me know.
>>
>> I have confirmed with the initial reporter (internal test team) that
<driver name='qemu'/>
>> with a known to be broken linux next kernel also fixes the problem, so
it is really the
>> vhost changes.
> 
> OK I'm back and trying to make it more bisectable.
> 
> I pushed a new tag "batch-v2".
> It's same code but with this bisect should get more information.
I get the following with this tag

drivers/vhost/net.c: In function ?vhost_net_tx_get_vq_desc?:
drivers/vhost/net.c:574:7: error: implicit declaration of function
?vhost_get_vq_desc_batch?; did you mean ?vhost_get_vq_desc??
[-Werror=implicit-function-declaration]
  574 |   r = vhost_get_vq_desc_batch(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
      |       ^~~~~~~~~~~~~~~~~~~~~~~
      |       vhost_get_vq_desc