AioContext: optimize clearing the EventNotifier

It is pretty rare for aio_notify to actually set the EventNotifier.  It
can happen with worker threads such as thread-pool.c's, but otherwise it
should never be set thanks to the ctx->notify_me optimization.  The
previous patch, unfortunately, added an unconditional call to
event_notifier_test_and_clear; now add a userspace fast path that
avoids the call.

Note that it is not possible to do the same with event_notifier_set;
it would break, as proved (again) by the included formal model.

This patch survived over 3000 reboots on aarch64 KVM.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Tested-by: Richard W.M. Jones <rjones@redhat.com>
Message-id: 1437487673-23740-7-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>

1 files changed, 31 insertions, 1 deletions

diff --git a/include/block/aio.h b/include/block/aio.h
index be91e3f..9dd32e0 100644
--- a/include/block/aio.h
+++ b/include/block/aio.h
@@ -99,7 +99,19 @@ struct AioContext {
      */
     int walking_bh;
 
-    /* Used for aio_notify.  */
+    /* Used by aio_notify.
+     *
+     * "notified" is used to avoid expensive event_notifier_test_and_clear
+     * calls.  When it is clear, the EventNotifier is clear, or one thread
+     * is going to clear "notified" before processing more events.  False
+     * positives are possible, i.e. "notified" could be set even though the
+     * EventNotifier is clear.
+     *
+     * Note that event_notifier_set *cannot* be optimized the same way.  For
+     * more information on the problem that would result, see "#ifdef BUG2"
+     * in the docs/aio_notify_accept.promela formal model.
+     */
+    bool notified;
     EventNotifier notifier;
 
     /* Thread pool for performing work and receiving completion callbacks */
@@ -174,6 +186,24 @@ QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque);
 void aio_notify(AioContext *ctx);
 
 /**
+ * aio_notify_accept: Acknowledge receiving an aio_notify.
+ *
+ * aio_notify() uses an EventNotifier in order to wake up a sleeping
+ * aio_poll() or g_main_context_iteration().  Calls to aio_notify() are
+ * usually rare, but the AioContext has to clear the EventNotifier on
+ * every aio_poll() or g_main_context_iteration() in order to avoid
+ * busy waiting.  This event_notifier_test_and_clear() cannot be done
+ * using the usual aio_context_set_event_notifier(), because it must
+ * be done before processing all events (file descriptors, bottom halves,
+ * timers).
+ *
+ * aio_notify_accept() is an optimized event_notifier_test_and_clear()
+ * that is specific to an AioContext's notifier; it is used internally
+ * to clear the EventNotifier only if aio_notify() had been called.
+ */
+void aio_notify_accept(AioContext *ctx);
+
+/**
  * aio_bh_poll: Poll bottom halves for an AioContext.
  *
  * These are internal functions used by the QEMU main loop.