[Bug 207489] Kernel panic due to Lazy update IOAPIC EOI on an x86_64 *host*, when two (or more) PCI devices from different IOMMU groups are passed to Windows 10 guest, upon guest boot into Windows, with more than 4 VCPUs

[bugzilla-daemon@bugzilla.kernel.org]

https://bugzilla.kernel.org/show_bug.cgi?id=207489

Pol Van Aubel (linux-kernel@polvanaubel.com) changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |CODE_FIX

--- Comment #17 from Pol Van Aubel (linux-kernel@polvanaubel.com) ---
(In reply to Yani Stoyanov from comment #16)
> Kernel 5.8+
> on Fedora 32
> 
> I have the same issue with macos (sierra and catalina) virtual machines with
> passed through gpus. 
> 
> What I do to identify the problem: 
> 
> Try adding different cards to the mac vm's no matter amd/nvidia the vm hangs
> and if you try to restart/force restart it the whole libvirt process hangs. 
> 
> I get the same kernel overflow as mentioned in the bug. 
> 
> How I am sure its the same problem: 
> 
> I checked out the kernel source comment the problematic: 
> 
> if (edge && kvm_apicv_activated(ioapic->kvm))
>               ioapic_lazy_update_eoi(ioapic, irq);
> 
> Rebuild the kernel and try with my version and it was working fine. I also
> try kernel 5.5.19 where it was fine also since it didn't have the check. 
> 
> I assume macos have some special way of treating the interrupts but since I
> am new to the linux kernel and linux as hole I can not identify what exactly
> case the problem.

The fix to the bug I reported works for me and has made it into kernel 5.7
since v5.7-rc5. I'm therefore updating the status of this particular bug to
resolved.

Your problem seems to be with the ioapic_lazy_update_eoi mechanism regardless
of how it is triggered, and you should open a new report -- with all the
relevant information like debug logs -- for your issue if it's not been
resolved yet.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
You are watching the assignee of the bug.