Messages in this thread | | | From | Jürgen Groß <> | Subject | Re: [PATCH 0/7] xen/events: bug fixes and some diagnostic aids | Date | Mon, 8 Feb 2021 10:41:00 +0100 |
| |
On 08.02.21 10:11, Julien Grall wrote: > Hi Juergen, > > On 07/02/2021 12:58, Jürgen Groß wrote: >> On 06.02.21 19:46, Julien Grall wrote: >>> Hi Juergen, >>> >>> On 06/02/2021 10:49, Juergen Gross wrote: >>>> The first three patches are fixes for XSA-332. The avoid WARN splats >>>> and a performance issue with interdomain events. >>> >>> Thanks for helping to figure out the problem. Unfortunately, I still >>> see reliably the WARN splat with the latest Linux master >>> (1e0d27fce010) + your first 3 patches. >>> >>> I am using Xen 4.11 (1c7d984645f9) and dom0 is forced to use the 2L >>> events ABI. >>> >>> After some debugging, I think I have an idea what's went wrong. The >>> problem happens when the event is initially bound from vCPU0 to a >>> different vCPU. >>> >>> From the comment in xen_rebind_evtchn_to_cpu(), we are masking the >>> event to prevent it being delivered on an unexpected vCPU. However, I >>> believe the following can happen: >>> >>> vCPU0 | vCPU1 >>> | >>> | Call xen_rebind_evtchn_to_cpu() >>> receive event X | >>> | mask event X >>> | bind to vCPU1 >>> <vCPU descheduled> | unmask event X >>> | >>> | receive event X >>> | >>> | handle_edge_irq(X) >>> handle_edge_irq(X) | -> handle_irq_event() >>> | -> set IRQD_IN_PROGRESS >>> -> set IRQS_PENDING | >>> | -> evtchn_interrupt() >>> | -> clear IRQD_IN_PROGRESS >>> | -> IRQS_PENDING is set >>> | -> handle_irq_event() >>> | -> evtchn_interrupt() >>> | -> WARN() >>> | >>> >>> All the lateeoi handlers expect a ONESHOT semantic and >>> evtchn_interrupt() is doesn't tolerate any deviation. >>> >>> I think the problem was introduced by 7f874a0447a9 ("xen/events: fix >>> lateeoi irq acknowledgment") because the interrupt was disabled >>> previously. Therefore we wouldn't do another iteration in >>> handle_edge_irq(). >> >> I think you picked the wrong commit for blaming, as this is just >> the last patch of the three patches you were testing. > > I actually found the right commit for blaming but I copied the > information from the wrong shell :/. The bug was introduced by: > > c44b849cee8c ("xen/events: switch user event channels to lateeoi model") > >> >>> Aside the handlers, I think it may impact the defer EOI mitigation >>> because in theory if a 3rd vCPU is joining the party (let say vCPU A >>> migrate the event from vCPU B to vCPU C). So info->{eoi_cpu, >>> irq_epoch, eoi_time} could possibly get mangled? >>> >>> For a fix, we may want to consider to hold evtchn_rwlock with the >>> write permission. Although, I am not 100% sure this is going to >>> prevent everything. >> >> It will make things worse, as it would violate the locking hierarchy >> (xen_rebind_evtchn_to_cpu() is called with the IRQ-desc lock held). > > Ah, right. > >> >> On a first glance I think we'll need a 3rd masking state ("temporarily >> masked") in the second patch in order to avoid a race with lateeoi. >> >> In order to avoid the race you outlined above we need an "event is being >> handled" indicator checked via test_and_set() semantics in >> handle_irq_for_port() and reset only when calling clear_evtchn(). > > It feels like we are trying to workaround the IRQ flow we are using > (i.e. handle_edge_irq()).
I'm not really sure this is the main problem here. According to your analysis the main problem is occurring when handling the event, not when handling the IRQ: the event is being received on two vcpus.
Our problem isn't due to the IRQ still being pending, but due it being raised again, which should happen for a one shot IRQ the same way.
But maybe I'm misunderstanding your idea.
Juergen [unhandled content-type:application/pgp-keys][unhandled content-type:application/pgp-signature] | |