Hi Mark,
Kindly ping...
I tried to fix it by decouple "host_irq" from vgic_irq when dealing
with vlpi. Because I think the semantics of vic_irq->host_irq is
vgic_irq bind 1:1 to the host-side LPI hwintid. But when I modified
the code, I found that "host_irq" was used multiple times in
vlpi related code and necessary...
Looking forward to your views on this qeustion.
Thanks,
Kunkun Jiang
On 2023/9/26 10:21, Kunkun Jiang wrote:
Hi everyone,
Sorry, yesterday's email was garbled. Please see this version.
Here is a very valuable question about the direct injection of vLPI.
Environment configuration:
1)A virtio_SCSI device is pass through to a small-scale VM,1U
2)Guest Kernel 4.19,Host Kernel 5.10
3)Enable GICv4/v4.1
The Guest will hang in the BIOS phase when it is restarted, and
report"Synchronous Exception at 0x280004654FF40".
Here's the analysis:
The virtio_SCSI device has six queues. The virtio driver may apply for
a vector for each queue of the device. It may also apply for one vector
for all queues. These queues share the vector.
In the problem scenario:
1.The host driver(vDPA or VFIO) applies for six vectors(LPI) for the
device.
2.The virtio driver applies for only one vector(vLPI) in the guest.
3.Only one vgic_irq is allocated by the vigic driver.In the current
vgic driver
implemention, when MAPTI/MAPI is executed in the Guest, it will be
trapped
to KVM. Therefore, the vgic driver allocates the same number of vgic_irq
to record these vectors which applied by device drivers in VM.
4.The kvm_vgic_v4_set_forwarding and its_map_vlpi is executed six times.
vgic_irq->host_irq equals the last linux interrupt ID(virq). The
result is
that six LPIs are mapped to one vLPI. The six LPIs of the device can
send interrupts. These interrupts will be injected into the guest
through the same vLPI.
5.When the Guest is restarted.The kvm_vgic_v4_unset_forwarding will
also be
executed six times. However, multiple call traces are generated. Since
there is only one vgic_irq, its_unmap_vlpi is executed only once.
WARN_ON(!(irq->hw && irq->host_irq == virq));
if (irq->hw) {
atomic_dec(&irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count);
irq->hw = false;
ret = its_unmap_vlpi(virq);
}
6.In the BIOS phase after the Guest restarted, the other five vectors
continue
to send interrupts. BIOS cannot handle these interrupts, so the Guest
hang.
This problem does not occur when the guest kernel is version 5.10,
because
this patch is incorporated.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=c66d4bd110a1f
I think there are other scenarios where the virtual machine will apply
for an
vector, and the host will apply for multiple vectors. There is still
value in
fixing this problem at the hypervisor layer. I think there are two
modification
methods here, but not sure if it is possible:
1)The vDPA or VFIO driver is aware of the behavior within the Guest
and only
apply for the same number of vectors.
2)Modify the vgic driver so that one vgic_irq can be bound to multiple
LPIs.
But I understand that the semantics of vigc_irq->host_irq is that
vgic_irq
is bound 1:1 to the host-side LPI hwintid.
If you have other ideas, we can discuss them together.
Looking forwarding to your reply.
Thanks,
Kunkun Jiang
.
