Re: [Qemu-devel] [PULL 3/3] vhost-user: Attempt to fix a race with set_mem_table.

[Michael S. Tsirkin]

On Sat, Aug 13, 2016 at 02:13:46AM -0400, Marc-André Lureau wrote:
> Hi
> 
> ----- Original Message -----
> > On Fri, Aug 12, 2016 at 11:54:54AM -0400, Marc-André Lureau wrote:
> > > Hi
> > > 
> > > ----- Original Message -----
> > > > On Fri, Aug 12, 2016 at 03:20:56AM -0400, Marc-André Lureau wrote:
> > > > > Hi
> > > > > 
> > > > > ----- Original Message -----
> > > > > > sent a follow-up response to GET_FEATURES), I am now wondering if
> > > > > > this
> > > > > > patch
> > > > > > may break existing vhost applications too ? If so, reverting it
> > > > > > possibly
> > > > > > better.
> > > > > > What confuses me is why it doesn’t fail all the time, but only about
> > > > > > 20%
> > > > > > to
> > > > > > 30% time as Fam reports.
> > > > > > 
> > > > > > Thoughts : Michael, Fam, MarcAndre ?
> > > > > 
> > > > > Indeed, I didn't ack that patch in the first place for that kind of
> > > > > reasons, so I would revert it.
> > > > > 
> > > > > thanks
> > > > 
> > > > I guess that's the safest thing to do for 2.7.
> > > > At least that's not any worse than 2.6.
> > > > I still think it's a good idea long term and test should be fixed,
> > > > but let's revert for now.
> > > > 
> > > 
> > > What about other backends that may have similar expectations from the
> > > protocol.
> > > 
> > > This patch is a hack, there is no reason to have it upstream.
> > 
> > The reason is to avoid crashes with existing backends.
> 
> Which backend? I had a similar issue, it wasn't about crashes, and Prerna 
> didn't mention crashes either, but anyway there is not guarantee that adding 
> a GET_FEATURES message will solve it...
> 


So what bothers me the most about dropping this is that
there are no backends with the new feature implemented
at the moment.

The GET_FEATURES hack at least makes it easy to test
with existing backends ...




> > > The solution is provided with the REPLY_ACK patch.
> > 
> > It needs a backend update though.
> > 
> > But the issue is old, it's not a regression. I think we lose nothing
> > by pushing the work-around out until after 2.7.
> > 
> > --
> > MST
> >