Re: [PATCH] console: make QMP screendump use coroutine

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Markus Armbruster (armbru@redhat.com) wrote:
> "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
> 
> > * Markus Armbruster (armbru@redhat.com) wrote:
> [...]
> >> Collecting several users before building infrastructure makes sense when
> >> the design of the infrastructure isn't obvious, or when the need for it
> >> is in doubt.
> >> 
> >> Neither is the case for running QMP handlers in a coroutine: QMP
> >> commands blocking the main loop is without doubt a problem we need to
> >> solve, and the way to solve it was obvious enough for Kevin to do it
> >> with one user: block_resize.  A second one quickly followed: screendump.
> >> 
> >> The only part that's different for HMP, I think, is "need".
> >> 
> >> Is HMP blocking the main loop a problem?
> >> 
> >> If yes, is it serious enough to justify solving it?
> >
> > I don't mind if HMP blocks for a small time while doing something, but
> > not if it can hang if the guest (or something else like it) misbehaves.
> > Not if it's something you might need to issue another command to recover
> > from.
> 
> The issue isn't HMP being unavailable while a command executes.  The
> issue is HMP stopping the main loop while a command executes.
> 
> Stopping the main loop not only stops everything running there, it can
> also stop other threads when they synchronize with the main loop via the
> Big QEMU Lock.

Yep.

> The obvious example is a command accessing a remote filesystem.  Special
> case: NFS with the hard option can hang indefinitely.

That I don't worry about too much for HMP; if your host is hosed, fix your host.

> screendump does that, and also waits for asynchronous gfx_update() with
> qxl devices.  Networking again, with a different peer.

That I would worry about since that's probably got interactions with the
guest and spice, and all the type of things you might be trying to debug
or test.

> We already decided that QMP commands stopping the main loop is serious.
> 
> To say it's not serious for HMP amounts to "don't do that then, use
> QMP".  Which may be fair.  Not for me to decide, though.

It's certainly more important for QMP; you don't want the main lock
being held for everyday type of interactions with management layers.

Dave

--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK