Re: [PATCH v2 1/2] virtio-pci: add error_detected for PCI AER recovery

Hi Michael,

Thanks for the review.  I will post a v3 with the following.

On Sun, Aug 02, 2026 at 05:10:21PM -0400, Michael S. Tsirkin wrote:
> On Fri, Jul 31, 2026 at 11:17:00AM +0800, Xixin Liu wrote:
> > virtio-pci only registered reset_prepare/reset_done.  The PCI error
> > recovery core treats devices without error_detected as NO_AER_DRIVER and
> > does not deliver pci_channel_io_perm_failure to the driver after a failed
> > recovery.
>
> Maybe it should?

Agreed that the current PCI core behaviour is awkward for drivers that
only wire reset_prepare/reset_done.

Without .error_detected the recovery core votes NO_AER_DRIVER on the
first broadcast, and report_perm_failure_detected skips the driver
callback whenever overall status is not RECOVERED.  That skip covers
NO_AER_DRIVER even when subordinate reset succeeded.  uevent DISCONNECT
is still emitted.

Changing the core would be a PCI-wide policy change.  This series
registers .error_detected and joins the existing ERS contract the same
way other ERS-aware drivers do.  Happy to follow up on the PCI side if
that is preferred.

> > +	 * permanent failure break virtqueues (follow-up may call driver
> > +	 * .shutdown for block teardown).
>
> block specifically? this is common code.

Right, that wording was wrong for virtio-pci.  On perm_failure the
common path only calls virtio_break_device() and returns DISCONNECT.
Same for every virtio device.

> > +		return PCI_ERS_RESULT_NEED_RESET;
>
> so I guess pci will do slot reset? but what will call driver reset
> later? confused.

NEED_RESET asks for a PCI bus or slot reset.  On AER that is
aer_root_reset() then pci_bus_error_reset().  That path does not call
.reset_prepare or .reset_done; those hooks still serve FLR.  Driver
restore after a successful reset is .slot_reset.  v3 adds that hook and
it restores config then calls virtio_device_reset_done().  Quiesce stays
in error_detected on frozen via virtio_device_reset_prepare().  If the
PCI reset fails, .slot_reset is skipped and perm_failure only breaks
the device.

> > +	default:
> > +		break;
> > +	}
> > +	return PCI_ERS_RESULT_NEED_RESET;
>
> here, too, and we didn't virtio_device_reset_prepare either.

Right.  pci_channel_state_t today is only normal, frozen, and
perm_failure.  In v3 frozen and default share one path: quiesce via
virtio_device_reset_prepare(), then NEED_RESET.

Thanks,
Xixin