Re: [PATCH] OvmfPkg: End timer interrupt later to avoid stack overflow under load

["Laszlo Ersek" <lersek@redhat.com>]

On 06/17/20 15:51, Paolo Bonzini wrote:
> On 16/06/20 20:42, Laszlo Ersek wrote:
>> (Hmmm... maybe the hypervisor *has* to queue the timer interrupts,
>> otherwise some of them would simply be lost, and the guest would lose
>> track of time.)
> 
> Yes, there are various kinds of coalescing of interrupts that
> hypervisors perform to help the guest keep track of time.  This is
> especially true of the PIT and RTC; newer OSes track time directly from
> the TSC, the HPET or the APIC timer so they tolerate lost ticks much better.
> 
> That said, Igor's patch seems correct to me.  In fact, I'd even move
> DisableInterrupts before gBS->RestoreTPL unless there's a good reason
> not to do so.

OK, thank you!

Igor, please confirm if you'd like to submit v2 with the update
suggested by Paolo, or if you prefer the current version. We're at the
beginning of the current development cycle, so I guess we can apply the
patch and see how it works in practice. If it ends up wreaking havoc on
some platforms, we can always revert the patch in time for the next
stable tag (edk2-stable202008).

Perhaps we should also file a TianoCore BZ for this issue, with a clear
problem statement, and the solution outline. The commit message is not
lacking, but I think a TianoCore BZ could be easier to find with a web
search, if users (not developers) want to comment after the patch is
merged. It's also easier to round up (possibly) important changes, for
stable tag content review, when there are BZs.

Thanks!
Laszlo