So there is no default interrupt affinity
for any physical IRQ in Xen? Is IOAPIC programmed to deliver interrupts in
round robin fashion or all interrupts go to one processor only?
From: Kay, Allen M
Sent: Wednesday, October 24, 2007
5:22 PM
To: Agarwal, Lomesh;
xen-devel@xxxxxxxxxxxxxxxxxxx
Cc: Han, Weidong
Subject: RE: [Xen-devel] interrupt
affinity question
The dma_msi_* stuff in intel-iommu.c is
not related to this. It looks like an area that needs to be cleaned up a
bit.
The call to request_irq() is for setting
up vt-d fault handler - linking vector with iommu_page_fault(). It is
only used when there is a iommu page fault which should not happen if
everything is setup correctly.
Passthru device interrupt handling is via
do_IRQ->do_IRQ_guest->hvm_do_IRQ_dpci path. The ioapic programming
for the passthru device was originally setup by the dom0 pci driver. The
interrupt of the passthru device always gets handled by xen first and
then gets re-inject to the guest via virtual ioapic/lapic models.
There is a interrupt latency
between the point where physical interrupt occurs and the point
virtual interrupt interrupt is injected to the guest - especially if
guest's vcpu is not running. We are still investigating on how to lower
this latency.
Allen
From:
xen-devel-bounces@xxxxxxxxxxxxxxxxxxx [mailto:xen-devel-bounces@xxxxxxxxxxxxxxxxxxx]
On Behalf Of Agarwal, Lomesh
Sent: Wednesday, October 24, 2007
3:55 PM
To: xen-devel@xxxxxxxxxxxxxxxxxxx
Subject: [Xen-devel] interrupt
affinity question
From looking at the code it looks like that interrupt
affinity will be set for all physical IRQs and it will be set to the physical
processor on which VCPU is running which called request_irq.
Can somebody confirm my understanding?
Pirq_guest_bind (in arch/x86/irq.c) calls set_affinity
(which will translate to dma_msi_set_affinity function in
arch/x86/hvm/vmx/vtd/intel-iommu.c for VTd).
So that means if request_irq for NIC interrupt is called
when a domain with single VCPU is scheduled on physical CPU 1 then NIC
interrupt will be bind to physical CPU 1 and later if the same domain is
scheduled to physical CPU 0 it won’t get the interrupt until it does a
VMEXIT.
So for lower interrupt latency we are should pin domain VCPU
also.
