On 17/03/2022 06:28, Juergen Gross wrote:
On 16.03.22 19:38, Raphael Ning wrote:
From: Raphael Ning <raphn...@amazon.com>
Currently, evtchn_fifo_set_pending() will mark the event as PENDING even
if it fails to lock the FIFO event queue(s), or if the guest has not
initialized the FIFO control block for the target vCPU. A well-behaved
guest should never trigger either of these cases.
Is this true even for the resume case e.g. after a migration?
The guests starts on the new host with no FIFO control block for any
vcpu registered, so couldn't an event get lost with your patch in case
it was sent before the target vcpu's control block gets registered?
An event that is PENDING but not LINKED is not reachable by the guest so
it won't ever see such an event, so the event is lost whether the P bit
is set or not.
Guests ensure that event channels are not bound to VCPUs that don't
(yet) have FIFO control blocks.
For example, in Linux xen_irq_resume() reinitializes the control blocks
(in xen_evtchn_resume()) before restoring any of the event channels.
David
