Re: [Qemu-devel] [PATCH] block: extend BLOCK_IO_ERROR event with nospace indicator

[Luiz Capitulino]

On Tue, 9 Sep 2014 15:23:50 +0200
Kevin Wolf <address@hidden> wrote:

> Am 09.09.2014 um 14:53 hat Eric Blake geschrieben:
> > On 09/09/2014 06:43 AM, Luiz Capitulino wrote:
> > 
> > >> Enhancing query-block in addition to the event makes sense, if it is
> > >> easy enough to do.  At this point, we are talking about debugging aids,
> > >> so as long as they are documented appropriately, I won't be too fussy.
> > > 
> > > OK, but I'm wondering if we need to add the string field to both,
> > > BLOCK_IO_ERROR and query-block, or only to one to the other.
> > > 
> > > In my opinion, we should only add it to BLOCK_IO_ERROR if libvirt is
> > > going to consume. Otherwise, it makes more sense to add it to query-block
> > > because that's where we'll meet the user.
> > > 
> > > Btw, by "consume" I mean read it and make it available to libvirt clients
> > > so that they can print it to their users. If we don't want libvirt to
> > > consume that field then I think we should only add it to query-block and
> > > info block.
> > 
> > [For those not aware, qemu built for downstream RHEL already has an
> > error string in the __com.redhat_ namespace; we're trying to figure out
> > what upstream should have so that downstream doesn't have to perpetually
> > maintain an extension]
> > 
> > Downstream libvirt does not currently consume any error string.  With
> > downstream qemu, the _only_ way to get the error string in the event is
> > to parse libvirt logs, or use upstream libvirt with its backdoor of
> > 'virsh qemu-monitor-event' (through the explicitly unsupported
> > libvirt-qemu.so) to get at the raw event information.  Changing libvirt
> > to expose such an error string to the end user would require a new
> > libvirt event number (the existing libvirt event is not extensible), so
> > existing clients would not be able to get at the information without
> > being recompiled to a new libvirt.
> > 
> > Since the whole point of this field is for debugging, I think that it is
> > sufficient to add it to JUST query-block, and not to the event.  That
> > is, if the app on top of libvirt gets an error, in the common case, they
> > won't care about the message (it won't change how they act), and in the
> > debug case, a developer trying to learn more about what happened can do
> > their own query-block directly (via 'virsh qemu-monitor-command', also
> > in libvirt-qemu.so) rather than trying to wire up libvirt to pass
> > through an error string through a new event.
> 
> You mention libvirt logs and I think that's an important point: If we
> move it to query-block, will we still get a log entry so that we can
> check this after the fact when only a log file is attached to a bug
> report, but we don't have a running guest?

Good point. I think this is call to have it in the event or both.