Re: [PATCH V2 mlx5-next 14/14] vfio/mlx5: Use its own PCI reset_done error handler

[Alex Williamson <alex.williamson@redhat.com>]

On Wed, 20 Oct 2021 13:46:29 -0300
Jason Gunthorpe <jgg@nvidia.com> wrote:

> On Wed, Oct 20, 2021 at 11:46:07AM +0300, Yishai Hadas wrote:
> 
> > What is the expectation for a reasonable delay ? we may expect this system
> > WQ to run only short tasks and be very responsive.  
> 
> If the expectation is that qemu will see the error return and the turn
> around and issue FLR followed by another state operation then it does
> seem strange that there would be a delay.
> 
> On the other hand, this doesn't seem that useful. If qemu tries to
> migrate and the device fails then the migration operation is toast and
> possibly the device is wrecked. It can't really issue a FLR without
> coordinating with the VM, and it cannot resume the VM as the device is
> now irrecoverably messed up.
> 
> If we look at this from a RAS perspective would would be useful here
> is a way for qemu to request a fail safe migration data. This must
> always be available and cannot fail.
> 
> When the failsafe is loaded into the device it would trigger the
> device's built-in RAS features to co-ordinate with the VM driver and
> recover. Perhaps qemu would also have to inject an AER or something.
> 
> Basically instead of the device starting in an "empty ready to use
> state" it would start in a "failure detected, needs recovery" state.

The "fail-safe recovery state" is essentially the reset state of the
device.  If a device enters an error state during migration, I would
think the ultimate recovery procedure would be to abort the migration,
send an AER to the VM, whereby the guest would trigger a reset, and
the RAS capabilities of the guest would handle failing over to a
multipath device, ejecting the failing device, etc.

However, regardless of the migration recovery strategy, userspace needs
a means to get the device back into an initial state in a deterministic
way without closing and re-opening the device (or polling for an
arbitrary length of time).  That's the minimum viable product here.
Thanks,

Alex