Re: [Qemu-devel] [RFC PATCH 0/2] establish nesting rule of BQL vs cpu-exclusive

[Roman Kagan]

On Mon, May 27, 2019 at 11:05:38AM +0000, Roman Kagan wrote:
> On Thu, May 23, 2019 at 12:31:16PM +0100, Alex Bennée wrote:
> > 
> > Roman Kagan <address@hidden> writes:
> > 
> > > I came across the following AB-BA deadlock:
> > >
> > >     vCPU thread                             main thread
> > >     -----------                             -----------
> > > async_safe_run_on_cpu(self,
> > >                       async_synic_update)
> > > ...                                         [cpu hot-add]
> > > process_queued_cpu_work()
> > >   qemu_mutex_unlock_iothread()
> > >                                             [grab BQL]
> > >   start_exclusive()                         cpu_list_add()
> > >   async_synic_update()                        finish_safe_work()
> > >     qemu_mutex_lock_iothread()                  cpu_exec_start()
> > >
> > > ATM async_synic_update seems to be the only async safe work item that
> > > grabs BQL.  However it isn't quite obvious that it shouldn't; in the
> > > past there were more examples of this (e.g.
> > > memory_region_do_invalidate_mmio_ptr).
> > >
> > > It looks like the problem is generally in the lack of the nesting rule
> > > for cpu-exclusive sections against BQL, so I thought I would try to
> > > address that.  This patchset is my feeble attempt at this; I'm not sure
> > > I fully comprehend all the consequences (rather, I'm sure I don't) hence
> > > RFC.
> > 
> > Hmm I think this is an area touched by:
> > 
> >   Subject: [PATCH v7 00/73] per-CPU locks
> >   Date: Mon,  4 Mar 2019 13:17:00 -0500
> >   Message-Id: <address@hidden>
> > 
> > which has stalled on it's path into the tree. Last time I checked it
> > explicitly handled the concept of work that needed the BQL and work that
> > didn't.
> 
> I'm still trying to get my head around that patchset, but it looks like
> it changes nothing in regards to cpu-exclusive sections and safe work,
> so it doesn't make the problem go.
> 
> > How do you trigger your deadlock? Just hot-pluging CPUs?
> 
> Yes.  The window is pretty narrow so I only saw it once although this
> test (where the vms are started and stopped and the cpus are plugged in
> and out) is in our test loop for quite a bit (probably 2+ years).
> 
> Roman.

ping?