Re: [PATCH] monitor/qmp: fix race with clients disconnecting early

[Markus Armbruster]

Stefan Reiter <s.reiter@proxmox.com> writes:

> On 10/12/21 11:27 AM, Markus Armbruster wrote:
>> Stefan, any thoughts on this?
>> 
>
> Sorry, I didn't get to work on implementing this. Idea 3 does seem very
> reasonable, though I suppose the question is what all should go into the
> per-session state, and also how it is passed down.

Let's start with the state you have shown to be problematic.  To decide
what else to move from monitor state to session state, we'll want to
review monitor state one by one.  Can be done in review of patches
creating the session state.

Most users need the current session state.  So have struct MonitorQMP
hold a pointer to it.

To execute an in-band command in the main thread, we need the command's
session state.  I'd try adding a pointer to QMPRequest.

Then use reference counting to keep the session alive until all its
commands are processed.

Makes sense?

> We did roll out my initial patch to our users in the meantime and got
> some positive feedback (that specific error disappeared), however another
> one (reported at the same time) still exists, so I was trying to diagnose
> - it might also turn out to be monitor related and resolved by the more
> thorough fix proposed here, but unsure.

That would be lovely.

Thanks!