Re: [Qemu-devel] [PATCH RFC 1/6] io: only allow return path for socket typed

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

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Fri, May 19, 2017 at 03:33:12PM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > > shutdown() is safe, in that it stops any other threads accessing the fd
> > > > but doesn't allow it's reallocation until the close;  We perform the
> > > > close only when we've joined all other threads that were using the fd.
> > > > Any of the threads that do new calls on the fd get an error and quickly
> > > > fall down their error paths.
> > > 
> > > Ahh that's certainly an interesting scenario. That would certainly be
> > > a problem with the migration code when this was originally written.
> > > It had two QEMUFile structs each with an 'int fd' field, so when you
> > > close 'fd' on one QEMUFile struct, it wouldn't update the other QEMUFile
> > > used by another thread.
> > > 
> > > Since we switched over to use QIOChannel though, I think the thread
> > > scenario you describe should be avoided entirely. When you have multiple
> > > QEMUFile objects, they each have a reference counted pointer to the same
> > > underlying QIOChannel object instance. So when QEMUFile triggers a call
> > > to qio_channel_close() in one thread, that'll set fd=-1 in the QIOChannel.
> > > Since the other threads have a reference to the same QIOChannel object,
> > > they'll now see this fd == -1 straightaway.
> > > 
> > > So, IIUC, this should make the need for shutdown() redundant (at least
> > > for the thread race conditions you describe).
> > 
> > That's not thread safe unless you're doing some very careful locking.
> > Consider:
> >   T1                      T2       
> >      oldfd=fd               tmp=fd
> >      fd=-1
> >      close(oldfd)
> >      unrelated open()
> >                             read(tmp,...
> > 
> > In practice every use of fd will be a copy into a tmp and then the
> > syscall; the unrelated open() could happen in another thread.
> > (OK, the gap between the tmp and the read is tiny, although if we're
> > doing multiple operations chances are the compiler will optimise
> > it to the top of a loop).
> > 
> > There's no way to make that code safe.
> 
> Urgh, yes, I see what you mean.
> 
> Currently the QIOChannelCommand implementation, uses a pair of anonymous
> pipes for stdin/out to the child process. I wonder if we could switch
> that to use socketpair() instead, thus letting us shutdown() on it too.
> 
> Though I guess it would be sufficient for qio_channel_shutdown() to
> merely kill the child PID, while leaving the FDs open, as then you'd
> get EOF and/or EPIPE on the read/writes.

Yes, I guess it's a question of which one is more likely to actually
kill the exec child off; the socketpair is more likely to cause the
source side migration code to cancel cleanly, although a kill -9 
should sort out a wayward exec child.

Dave

> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK