lkml.org 
[lkml]   [2020]   [May]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v2 03/22] KVM: SVM: immediately inject INTR vmexit
From
Date
On 21/05/20 16:08, Paolo Bonzini wrote:
> On 21/05/20 14:50, Vitaly Kuznetsov wrote:
>> Sorry for reporting this late but I just found out that this commit
>> breaks Hyper-V 2016 on KVM on SVM completely (always hangs on boot). I
>> haven't investigated it yet (well, this is Windows, you know...) but
>> what's usually different about Hyper-V is that unlike KVM/Linux it has
>> handlers for some hardware interrupts in the guest and not in the
>> hypervisor.
>
> "Always hangs on boot" is easy. :) At this point I think it's easiest
> to debug it on top of the whole pending SVM patches that remove
> exit_required completely (and exit_required is not coming back anyway).

Ok so there could be two bugs, as the hang seems to happens much earlier
later in the series (try "grep int_ctl:.0x.....1.." on the trace).

As one could guess from the grep, one thing that is certainly different
between KVM and Hyper-V is that Hyper-V injects interrupts using
int_ctl; sometimes it also uses eventinj but presumably it's just
copying it from exitintinfo).

This could cause problems: for example, when L1 wants to inject a
virtual interrupt into L2 that has interrupts disabled or V_TPR >=
V_INTR_PRIO, and KVM also wants to inject an interrupt to L1, then KVM
might end up stomping on Hyper-V's int_ctl. However I cannot think
off-hand of a scenario where this could happen in this case, because
Hyper-V does set EXIT_INTR and therefore we should never get into
enable_irq_window while L2 is running. Still, that's one place where
I'd start adding some trace_printk's.

Also, if a uniprocessor guest also fails, it might be easier to debug.

Paolo

\
 
 \ /
  Last update: 2020-05-21 23:05    [W:0.063 / U:2.396 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site