Re: [REGRESSION] vfio gpu passthrough stopped working

[Harshit Mogalapalli <harshit.m.mogalapalli@oracle.com>]

Hi,

On 07/08/25 21:22, Greg KH wrote:
> On Thu, Aug 07, 2025 at 03:31:17PM +0000, cat wrote:
>> #regzbot introduced: v6.12.34..v6.12.35
>>
>> After upgrade to kernel 6.12.35, vfio passthrough for my GPU has stopped working within a windows VM, it sees device in device manager but reports that it did not start correctly. I compared lspci logs in the vm before and after upgrade to 6.12.35, and here are the changes I noticed:
>>
>> - the reported link speed for the passthrough GPU has changed from 2.5 to 16GT/s
>> - the passthrough GPU has lost it's 'BusMaster' and MSI enable flags
>> - latency measurement feature appeared
>>
>> These entries also began appearing within the vm in dmesg when host kernel is 6.12.35 or above:
>>
>> [    1.963177] nouveau 0000:01:00.0: sec2(gsp): mbox 1c503000 00000001
>> [    1.963296] nouveau 0000:01:00.0: sec2(gsp):booter-load: boot failed: -5
>> ...
>> [    1.964580] nouveau 0000:01:00.0: gsp: init failed, -5
>> [    1.964641] nouveau 0000:01:00.0: init failed with -5
>> [    1.964681] nouveau: drm:00000000:00000080: init failed with -5
>> [    1.964721] nouveau 0000:01:00.0: drm: Device allocation failed: -5
>> [    1.966318] nouveau 0000:01:00.0: probe with driver nouveau failed with error -5
>>
>>
>> 6.12.34 worked fine, and latest 6.12 LTS does not work either. I am using intel CPU and nvidia GPU (for passthrough, and as my GPU on linux system).
> 
 > Can you use git bisect to find the offending commit?>

Additional notes: I looked at the log and am listing probably relevant 
commit, if bisection is too costly:

68e58f579121 PCI: dwc: ep: Correct PBA offset in .set_msix() callback
523815857b1e PCI: cadence-ep: Correct PBA offset in .set_msix() callback

These two might be interesting ones to consider. Please ignore this note 
if bisection is already in progress as these are pure guesses.


Thanks,
Harshit

> 
>