win32: fix main-loop busy loop on socket/fd event

Commit 05e514b1d4d5bd4209e2c8bbc76ff05c85a235f3 introduced an AIO
context optimization to avoid calling event_notifier_test_and_clear() on
ctx->notifier. On Windows, the same notifier is being used to wakeup the
wait on socket events (see commit
d3385eb448e38f828c78f8f68ec5d79c66a58b5d).

The ctx->notifier event is added to the gpoll sources in
aio_set_event_notifier(), aio_ctx_check() should clear the event
regardless of ctx->notified, since Windows sets the event by itself,
bypassing the aio->notified. This fixes qemu not clearing the event
resulting in a busy loop.

Paolo suggested to me on irc to call event_notifier_test_and_clear()
after select() >0 from aio-win32.c's aio_prepare. Unfortunately, not all
fds associated with ctx->notifiers are in AIO fd handlers set.
(qemu_set_nonblock() in util/oslib-win32.c calls qemu_fd_register()).

Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
---
 include/block/aio.h | 2 ++
 async.c             | 6 ++++++
 2 files changed, 8 insertions(+)

> Paolo suggested to me on irc to call event_notifier_test_and_clear()
> after select() >0 from aio-win32.c's aio_prepare. Unfortunately, not all
> fds associated with ctx->notifiers are in AIO fd handlers set.
> (qemu_set_nonblock() in util/oslib-win32.c calls qemu_fd_register()).

That makes sense.  Out of curiosity, what is a practical case of a socket
that is nonblocking but doesn't have an attached handler?

Another possibility (this one requires much more attention to avoid missing
some edge case; however, it should be easy to verify if it fixes the busy
loop) could be to move aio_notify_accept to just before setting ctx->notify_me.
This would work for both aio-posix and aio-win32.

Paolo

> Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
> ---
>  include/block/aio.h | 2 ++
>  async.c             | 6 ++++++
>  2 files changed, 8 insertions(+)
> 
> diff --git a/include/block/aio.h b/include/block/aio.h
> index ca551e346f..a6da135bf3 100644
> --- a/include/block/aio.h
> +++ b/include/block/aio.h
> @@ -100,6 +100,7 @@ struct AioContext {
>       */
>      int walking_bh;
>  
> +#ifndef _WIN32
>      /* Used by aio_notify.
>       *
>       * "notified" is used to avoid expensive event_notifier_test_and_clear
> @@ -113,6 +114,7 @@ struct AioContext {
>       * in the docs/aio_notify_accept.promela formal model.
>       */
>      bool notified;
> +#endif
>      EventNotifier notifier;
>  
>      /* Thread pool for performing work and receiving completion callbacks */
> diff --git a/async.c b/async.c
> index b2de360c23..8c2a68b6cc 100644
> --- a/async.c
> +++ b/async.c
> @@ -329,15 +329,21 @@ void aio_notify(AioContext *ctx)
>      smp_mb();
>      if (ctx->notify_me) {
>          event_notifier_set(&ctx->notifier);
> +#ifndef _WIN32
>          atomic_mb_set(&ctx->notified, true);
> +#endif
>      }
>  }
>  
>  void aio_notify_accept(AioContext *ctx)
>  {
> +#ifndef _WIN32
>      if (atomic_xchg(&ctx->notified, false)) {
> +#endif
>          event_notifier_test_and_clear(&ctx->notifier);
> +#ifndef _WIN32
>      }
> +#endif
>  }
>  
>  static void aio_timerlist_notify(void *opaque)
> --
> 2.11.0
> 
>

Hi

On Wed, Jan 4, 2017 at 10:23 PM Paolo Bonzini <pbonzini@redhat.com> wrote:

>
> > Paolo suggested to me on irc to call event_notifier_test_and_clear()
> > after select() >0 from aio-win32.c's aio_prepare. Unfortunately, not all
> > fds associated with ctx->notifiers are in AIO fd handlers set.
> > (qemu_set_nonblock() in util/oslib-win32.c calls qemu_fd_register()).
>
> That makes sense.  Out of curiosity, what is a practical case of a socket
> that is nonblocking but doesn't have an attached handler?
>
>
Good question, qio_channel_socket_set_blocking() calls qemu_set_nonblock().
But it seems to also use its own source handler
qio_channel_create_socket_watch(), so the AIO notifier is probably needless
here.

Another possibility (this one requires much more attention to avoid missing
> some edge case; however, it should be easy to verify if it fixes the busy
> loop) could be to move aio_notify_accept to just before setting
> ctx->notify_me.
> This would work for both aio-posix and aio-win32.
>
>
 --- a/async.c
+++ b/async.c
@@ -232,8 +232,8 @@ aio_ctx_check(GSource *source)
     AioContext *ctx = (AioContext *) source;
     QEMUBH *bh;

-    atomic_and(&ctx->notify_me, ~1);
     aio_notify_accept(ctx);
+    atomic_and(&ctx->notify_me, ~1);

That? doesn't work here.

thanks