KVM: s390: add and wire function gib_alert_irq_handler()
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 73 lines · Michael Mueller
Quoted 56 lines · Pierre Morel
Quoted 47 lines · Michael Mueller
Quoted 15 lines · Pierre Morel
Quoted 14 lines · Michael Mueller
The patch implements a handler for GIB alert interruptions on the host. Its task is to alert storage backed guests that interrupts are pending for them.
A GIB alert interrupt statistic counter is added as well:
$ cat /proc/interrupts ���������� CPU0������ CPU1 �� ... �� GAL:������ 0��������� 0�� [I/O] GIB Alert �� ...Signed-off-by Michael Mueller <[email protected]>---
...
Quoted 9 lines
Quoted 9 lines
arch/s390/kvm/kvm-s390.c +1
-0@@ -3531,6 +3531,10 @@ static int vcpu_post_run(struct kvm_vcpu *vcpu, int exit_reason) ����� vcpu->run->s.regs.gprs[15] = vcpu->arch.sie_block->gg15; ����� atomic_dec(&vcpu->kvm->arch.vcpus_in_sie); +��� if (vcpu->kvm->arch.gib_in_use &&Quoted 8 lines
arch/s390/kvm/kvm-s390.c static int vcpu_post_run(struct kvm_vcpu · line 3537
Quoted 3 lines
arch/s390/kvm/kvm-s390.c +3
-0+������� !in_alert_list(vcpu->kvm->arch.gisa) && +������� !atomic_read(&vcpu->kvm->arch.vcpus_in_sie)) +������� vcpu->kvm->arch.gisa->iam = vcpu->kvm->arch.iam;Here, AFAIU, if the IAM is set, we must take care that not bit from IPM for which IAM is set is also set otherwise we won't get an interruption.
My view is that we have a short time window between SIE exit and the restoration of the IAM in the GISA, when an interruption might be indicated by setting the respective ISC bit in the IPM. We will not get an GAL interruption in the host for that. I think until here we are on the same page, right?
In that situation the host needs to take action.
Quoted 7 lines
And it is taken, see the trace. These interruptions are delivered outside of SIE because they are observed by:
kvm_s390_deliver_pending_interrupts()
��deliverable_irqs(vcpu)
����� IPM != 0
�������__deliver_io()Quoted 8 lines
If I don't make mistake, this code is only executed during
vcpu_pre_run().What I understand is that the next time the vcpu is scheduled, for example due to another interrupt, the IPM interrupts are delivered through this channel.
In the case the vcpu does not get scheduled they are not delivered.
Idle vcpus will always get scheduled by the hrtimer function
kvm_s390_idle_wakeup()which callskvm_s390_vcpu_wakeup()which then go throughkvm_s390_deliver_pending_interrupts()and deliver the interruption.
Quoted 5 lines
That's what we see in the trace below. (Well that's an assumption but I know that only 5 interruptions were generated by the device)
The question is if the interruption can't be delivered more early instead of waiting for one of the
ckc_timersto expire.
Yes, what if the timer takes a very long time.
Think of NO_HZ_IDLE configuration
Quoted 25 lines
Quoted 25 lines
Quoted 24 lines
00 01543426770:725504 4 - 01 00000000e280ea16� inject: I/O (AI/gisa) isc 3
00 01543426770:727023 4 - 0200000000676fd4ca00[0706c00180000000-00000000008e9428]: deliver: I/O (AI/gisa) isc 6 00 01543426770:727197 4 - 0200000000676fd4ca00[0706c00180000000-00000000008e9428]: deliver: I/O (AI/gisa) isc 6 00 01543426770:727353 4 - 0200000000676fd4ca00[0706c00180000000-00000000008e9428]: deliver: I/O (AI/gisa) isc 6 00 01543426770:727506 4 - 0200000000676fd4ca00[0706c00180000000-00000000008e9428]: deliver: I/O (AI/gisa) isc 6 00 01543426770:731878 4 - 0200000000676fd4ca00[0706c00180000000-00000000008e9428]: deliver: I/O (AI/gisa) isc 6 00 01543426770:731987 4 - 01 00000000e280ea16� inject: I/O (AI/gisa) isc 3Quoted 2 lines
Quoted 4 lines
Quoted 2 lines
AFAIU you do not need to deliver IRQ which IPM is set in GISA during
vcpu_pre_run(), they will be delivered to the guest on SIE entry.I know that is the patch that I drew back, but for other reasons.
Yes but this is not what is done here.
Here you deliver the interrupt on the next schedule but you do not provoque the schedule, just wait for the next interrupt. If it comes.
contined here under....
Quoted 10 lines
Quoted 8 lines
I tested this Code example in:
Message-Id: <[email protected]>Back to
vcpu_post_run(), I am not sure that this is the better place to reset the IAM.What better place you have in mind?
...continued from above
yes, when the vCPU issues a WAIT (sorry not HALT), in post_run(), kvm_handle_sie_intercept()
Quoted 6 lines
Quoted 2 lines
Shouldn't we set the IAM only if the last active vcpu issues a halt?
Can you be little more specific with "vcpu issues a halt"? My view is the none of the vcpus of a guest are in SIE context. Currently I identify this situation with the
vcpus_in_siemetric.
in __set_cpu_idle()
seems to me to excatly follow the documentation.
There would be two metrics:
- 1: are vCPU running
- 2: is one vCPU in wait state
You need to setup the IAM only in the case
(NO vCPU running && at least one CPU wait)
Quoted 5 lines
Quoted 3 lines
Is there other SIE exit reasons to activate the alert list?
I don't think there is. Have you read something?
No, I think there is no other reason to activate the alert list than when the condition above becomes true.
In all other SIE exit reason the vCPU will be rescheduled after the event has been processed.
STOP is special, but we surely do not want to send an IRQ to a stopped vCPU.
Quoted 7 lines
Quoted 6 lines
Regards,
Pierre
Signature · 2 lines
Pierre Morel Linux/KVM/QEMU in B�blingen - Germany