lkml.org 
[lkml]   [2023]   [Jan]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/2] vhost: improve livepatch switching for heavily loaded vhost worker kthreads
On Fri, Jan 27, 2023 at 12:19:03PM +0100, Petr Mladek wrote:
> On Thu 2023-01-26 15:12:35, Seth Forshee (DigitalOcean) wrote:
> > On Thu, Jan 26, 2023 at 06:03:16PM +0100, Petr Mladek wrote:
> > > On Fri 2023-01-20 16:12:20, Seth Forshee (DigitalOcean) wrote:
> > > > We've fairly regularaly seen liveptches which cannot transition within kpatch's
> > > > timeout period due to busy vhost worker kthreads.
> > >
> > > I have missed this detail. Miroslav told me that we have solved
> > > something similar some time ago, see
> > > https://lore.kernel.org/all/20220507174628.2086373-1-song@kernel.org/
> >
> > Interesting thread. I had thought about something along the lines of the
> > original patch, but there are some ideas in there that I hadn't
> > considered.
>
> Could you please provide some more details about the test system?
> Is there anything important to make it reproducible?
>
> The following aspects come to my mind. It might require:
>
> + more workers running on the same system
> + have a dedicated CPU for the worker
> + livepatching the function called by work->fn()
> + running the same work again and again
> + huge and overloaded system

I'm isolating a CPU, starting a KVM guest with a virtio-net device, and
setting the affinity of the vhost worker thread to only the isolated
CPU. Thus the vhost-worker thread has a dedicated CPU, as you say. (I'll
note that in real-world cases the systems have many CPUs, and while the
vhost threads aren't each given a dedicated CPU, if the system load is
light enough a thread can end up with exlusive use of a CPU).

Then all I do is run iperf between the guest and the host with several
parallel streams. I seem to be hitting the limits of the guest vCPUs
before the vhost thread is fully saturated, as this gets it to about 90%
CPU utilization by the vhost thread.

> > > Honestly, kpatch's timeout 1 minute looks incredible low to me. Note
> > > that the transition is tried only once per minute. It means that there
> > > are "only" 60 attempts.
> > >
> > > Just by chance, does it help you to increase the timeout, please?
> >
> > To be honest my test setup reproduces the problem well enough to make
> > KLP wait significant time due to vhost threads, but it seldom causes it
> > to hit kpatch's timeout.
> >
> > Our system management software will try to load a patch tens of times in
> > a day, and we've seen real-world cases where patches couldn't load
> > within kpatch's timeout for multiple days. But I don't have such an
> > environment readily accessible for my own testing. I can try to refine
> > my test case and see if I can get it to that point.
>
> My understanding is that you try to load the patch repeatedly but
> it always fails after the 1 minute timeout. It means that it always
> starts from the beginning (no livepatched process).
>
> Is there any chance to try it with a longer timeout, for example, one
> hour? It should increase the chance if there are more problematic kthreads.

Yes, I can try it. But I think I already mentioned that we are somewhat
limited by our system management software and how livepatch loading is
currently implemented there. I'd need to consult with others about how
long we could make the timeout, but 1 hour is definitely too long under
our current system.

> > > This low timeout might be useful for testing. But in practice, it does
> > > not matter when the transition is lasting one hour or even longer.
> > > It takes much longer time to prepare the livepatch.
> >
> > Agreed. And to be clear, we cope with the fact that patches may take
> > hours or even days to get applied in some cases. The patches I sent are
> > just about improving the only case I've identified which has lead to
> > kpatch failing to load a patch for a day or longer.
>
> If it is acceptable to wait hours or even days then the 1 minute
> timeout is quite contra-productive. We actually do not use any timeout
> at all in livepatches provided by SUSE.

I agree, though I'd still prefer it didn't take days. Based on this
discussion I do plan to look at changing how we load livepatches to make
this possible, but it will take some time.

Thanks,
Seth

\
 
 \ /
  Last update: 2023-03-26 23:58    [W:0.195 / U:0.348 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site