Re: [RFC PATCH 0/1] vfio/pci: Wait for GB202 firmware after reset
From: Gustavo Andrade
Date: Mon Nov 16 2026 - 18:50:18 EST
Hi Alex,
I completely agree with you. I wasn't aware that QEMU was the proper
place for this, but I've made the change there and it is working
without any issues.
Thank you very much, Alex and Jason, for pointing me in the right direction.
Cheers,
Gustavo Andrade
On Mon, Nov 16, 2026 at 12:40 PM Alex Williamson <alex@xxxxxxxxxxx> wrote:
>
> On Sat, 14 Nov 2026 22:47:34 -0300
> Gustavo Andrade <gustavogyn@xxxxxxxxx> wrote:
>
> > Hi Jason,
> >
> > Thanks for the feedback. I submitted this as an RFC because, while the
> > workaround solved the issue for me, it seemed unusual that only my
> > card required it, since I haven't seen recent reports of this issue.
> >
> > How would you suggest I proceed? I am currently running the latest
> > VBIOS available for my card (Palit RTX 5090 GameRock, version
> > 98.02.2E.80.10), which is also the most recent version listed in the
> > TechPowerUp database.
>
> I imagine a better shape for a quirk is a resetfn in QEMU.
>
> The fact that nova knows about this issue and has code to handle it
> suggests to me that it's only the guest initial boot and reboot, where
> the VBIOS is run before the FSP is initialized, that's causing the
> problem. We don't want to pollute every vfio-pci reset with this
> wait-for-device action (if we did, a reset_done callback would be a
> better solution) and I also don't think we want to implement a
> kernel PCI-core device specific reset. Either of those would make the
> device behave differently in a VM, or modify the current bare metal
> behavior that's clearly expected by drivers like nova.
>
> A resetfn in QEMU can be device specific, wait for FSP to initialize,
> and not affect runtime resets where it would be the driver's
> responsibility to wait for FSP, as nova already does. For example
> something similar to vfio_radeon_reset(), but calling the
> VFIO_DEVICE_RESET ioctl and then stalling for the FSP initialization
> to complete. Thanks,
>
> Alex