Re: [Qemu-stable] Recent patches for 2.4

[Peter Lieven]

Am 04.08.2015 um 13:53 schrieb Paolo Bonzini:
> On 04/08/2015 11:22, Peter Lieven wrote:
> > > > > > edec47c main-loop: fix qemu_notify_event for aio_notify optimization
> > > > > Part of the above AioContext series.
> > > > So either the whole series or none of them I guess?
> > > It's a separate bug, and theoretically it's there in 2.3.1 as well, but
> > > no one ever reproduced it (it would hang in make check) so not
> > > worthwhile.
> > Can you give me a pointer what the symtoms where?
> If a thread tries to wake up the main thread using qemu_notify_event(),
> the main thread will never wake up.  This for example could happen if
> the first thread calls qemu_set_fd_handler() or timer_mod().
>
> > I have a qemu-img convert job on x86_64 that reproducibly hangs on
> > bdrv_drain_all at the end of the convert process.
> > I convert from nfs:// to local storage here. I try to figure out which BS
> > reports busy. Qemu here is still 2.2.1.
> qemu-img does not use main-loop, so this cannot be the cause.
>
> The AioContext bugs only happen when you have a thread executing the
> main loop and one thread executing aio_poll, so they can also be
> excluded as the cause of qemu-img problems.

Okay, what I found out is that in aio_poll I get revents = POLLIN for
the nfs file descriptor. But there is no data available on the socket.
But as a consequence progress is true and we loop here forever.

I have seen that is a common bug in Linux to return POLLIN on a fd
even there is no data available. I don't have this problem in general,
in this case no qemu-img or qemu process would ever terminate when
nfs is involved, but in this special case it happens reproducible.

Peter