Re: [PATCH 0/3] KVM: Fix lost IRQ acks for RTC

[Paolo Bonzini]

On 29/02/2016 16:04, Joerg Roedel wrote:
> Hi,
> 
> here is a small patch-set to fix a race condition which
> happens when an RTC-IRQ is migrated to another VCPU while it
> is being handled by the guest.
> 
> The RTC-EOI handling in KVM requires that all sent interrupt
> messages to the VCPUs need to be acked before another
> RTC-IRQ can be sent. When an EOI signal from the guest is
> lost, it will never see an RTC interrupt again (until it
> reboots).
> 
> This is easily reproducible with a Linux guest executing
> this loop:
> 
> 	$ while true;do time hwclock --show --test --debug;done
> 
> When the guest has multiple vcpus and the RTC-IRQ is
> regularily migrated (e.g. by irqbalance), the race condition
> will be hit after some time and the hwclock tool will fail
> with:
> 
> 	select() to /dev/rtc to wait for clock tick timed out...synchronization failed
> 
> The race condition happens because of the way the EOI
> backtracking between local APIC and IOAPIC works in KVM. The
> destination VCPU and vector is part of the IOAPIC state.
> When the guest sends an EOI to the local APIC the vector is
> matched against the destinations stored in the IOAPIC and
> ACKed there too if it matches.
> 
> The problem begins when a VCPU handles an RTC interrupt and
> at the same time another VCPU migrates the RTC-IRQ away from
> that VCPU. This updates the IOAPIC state in KVM to
> the new destination, so that the EOI sent from the first
> VCPU does not match anymore in the IOAPIC, hence losing the
> RTC-EOI.
> 
> This patch-set fixes the race-condition by adding explicit
> back-tracking information for RTC-IRQs. The rtc_status
> struct already holds a dest_map bitmap to store which VCPUs
> receveived an RTC-IRQ. This is extended to also hold the
> vector that was sent to this VCPU.
> 
> This information is then used to match EOI signals from the
> guest to the RTC. This explicit back-tracking fixes the
> issue.
> 
> Regards,

Nice patches, really.  Ok to wait until 4.6?

Paolo