On Mon, May 30, 2022 at 06:04:32PM +0300, Roman Kagan wrote: > On Mon, May 30, 2022 at 01:28:17PM +0200, Markus Armbruster wrote: > > Roman Kagan <rvka...@yandex-team.ru> writes: > > > > > On Wed, May 25, 2022 at 12:54:47PM +0200, Markus Armbruster wrote: > > >> Konstantin Khlebnikov <khlebni...@yandex-team.ru> writes: > > >> > > >> > This event represents device runtime errors to give time and > > >> > reason why device is broken. > > >> > > >> Can you give an or more examples of the "device runtime errors" you have > > >> in mind? > > > > > > Initially we wanted to address a situation when a vhost device > > > discovered an inconsistency during virtqueue processing and silently > > > stopped the virtqueue. This resulted in device stall (partial for > > > multiqueue devices) and we were the last to notice that. > > > > > > The solution appeared to be to employ errfd and, upon receiving a > > > notification through it, to emit a QMP event which is actionable in the > > > management layer or further up the stack. > > > > > > Then we observed that virtio (non-vhost) devices suffer from the same > > > issue: they only log the error but don't signal it to the management > > > layer. The case was very similar so we thought it would make sense to > > > share the infrastructure and the QMP event between virtio and vhost. > > > > > > Then Konstantin went a bit further and generalized the concept into > > > generic "device runtime error". I'm personally not completely convinced > > > this generalization is appropriate here; we'd appreciate the opinions > > > from the community on the matter. > > > > "Device emulation sending an even on entering certain error states, so > > that a management application can do something about it" feels > > reasonable enough to me as a general concept. > > > > The key point is of course "can do something": the event needs to be > > actionable. Can you describe possible actions for the cases you > > implement? > > The first one that we had in mind was informational, like triggering an > alert in the monitoring system and/or painting the VM as malfunctioning > in the owner's UI. > > There can be more advanced scenarios like autorecovery by resetting the > faulty VM, or fencing it if it's a cluster member.
The discussion kind of stalled here. Do you think the approach makes sense or not? Should we try and resubmit the series with a proper cover letter and possibly other improvements or is it a dead end? Thanks, Roman.