Re: [Xen-devel] [PATCH] xen: make sure stop_machine_run() is always called in a tasklet

[Jürgen Groß <jgross@xxxxxxxx>]

On 14.02.20 18:34, Igor Druzhinin wrote:
> On 14/02/2020 16:39, Jürgen Groß wrote:
> > On 14.02.20 15:06, Igor Druzhinin wrote:
> > > On 11/02/2020 09:35, Juergen Gross wrote:
> > > > With core scheduling active it is mandatory for stop_machine_run() to
> > > > be called in a tasklet only, as otherwise a scheduling deadlock would
> > > > occur: stop_machine_run() does a cpu rendezvous by activating a tasklet
> > > > on all other cpus. In case stop_machine_run() was not called in an idle
> > > > vcpu it would block scheduling the idle vcpu on its siblings with core
> > > > scheduling being active, resulting in a hang.
> > > I suppose rcu_barrier() is fine due to process_pending_softirqs() being
> > > called inside? I'm a little concerned by imposing is_vcpu_idle() restriction
> > > in that case as rcu_barrier() could be technically called from a non-tasklet
> > > context.
> > No, stop_machine_run() with core scheduling active can only work when
> > called in an idle vcpu.
> >
> > OTOH it would be fairly easy to add another softirq for a similar
> > purpose and have a sync_machine_run() using that instead of tasklets.
> > This could be used for ucode loading, too.
> >
> > stop_machine_run() and sync_machine_run() could use a common main
> > function. The patch should be rather simple.
> >
> > Thoughts?
> I have a patch on the list (which I was planning to send a v2 for) that
> fixes another issue with rcu_barrier():
> https://lists.xenproject.org/archives/html/xen-devel/2020-01/msg02273.html
>
> As I understand it now that wouldn't work with core-scheduling. Do you think
> it's possible to synchronously wait for tasklets to finish in non-tasklet
> context (because that's what the purpose of rcu_barrier() is)?
No, won't work, unless we add preemption (basically would need per-vcpu
stacks instead of per-pcpu ones).

What might work IMO would be to do rcu_process_callbacks() no longer
during idle, but to have a specific softirq for that purpose. This would
remove the need to involve scheduling for rcu_barrier(). A brief check
of process_pending_softirqs() callers seems to allow that, but I'd like
to have a second opinion from someone having more rcu knowledge than me.
Single problematic users of process_pending_softirqs() could still be
switched to a variant not allowing the new rcu softirq.


Juergen

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel