Fix: sessiond: off-by-one poll check when draining an event notifier

When an event source is removed (on the death of an application), the
notification thread "drains" any remaining notifications from the
event notifier pipe.

In doing so, it creates a new poll set containing the event source to
check if messages are left in the event notification pipe.

The invocation of `LTTNG_POLL_GETEV(&events, 1)` means to check the
events pending for the first (and only) fd in the poll set. This check
is off by one since `0` should be used.

For some reason, this worked everywhere except when using a 32-bit
userland on a 64-bit kernel (on x86_64).

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>
Change-Id: I6f274fdd7c80d5676fd48ae20a14adb3cc010142

diff --git a/src/bin/lttng-sessiond/notification-thread-events.c b/src/bin/lttng-sessiond/notification-thread-events.c
index 6bd799279b3ff6cead92c77bb2553c6117ed5afc..747902612d93059705cd9d987cde39e1cc757a31 100644 (file)
--- a/src/bin/lttng-sessiond/notification-thread-events.c
+++ b/src/bin/lttng-sessiond/notification-thread-events.c
@@ -2017,7 +2017,7 @@ int drain_event_notifier_notification_pipe(
                 * the pipe is closed but empty.
                 */
                ret = lttng_poll_wait_interruptible(&events, 0);
-               if (ret == 0 || (LTTNG_POLL_GETEV(&events, 1) & LPOLLIN) == 0) {
+               if (ret == 0 || (LTTNG_POLL_GETEV(&events, 0) & LPOLLIN) == 0) {
                        /* No more notification to be read on this pipe. */
                        ret = 0;
                        goto end;