event/Loop: don't wait for io_uring if epoll is still ready

This fixes a starvation bug if epoll is getting hammered with
completions.  This fills the io_uring completion queue until it
overflows, but these overflows are never picked up because we never
invoke io_uring_enter().

We avoid this by calling io_uring_enter() in any case, but with a zero
timeout.

Author: MaxKellermann

src/event/Loop.cxx

@@ -411,7 +411,12 @@ EventLoop::UringWait(Event::Duration timeout) noexcept
            io_uring_prep_poll_multishot() is edge-triggered, so we
            have to consume all events to rearm it */
 
-	if (!epoll_ready) {
+	if (epoll_ready)
+		/* don't wait for io_uring completions if epoll is
+		   still ready */
+		timeout = Event::Duration{0};
+
+	{
 		struct __kernel_timespec timeout_buffer;
 		auto *kernel_timeout = ExportTimeoutKernelTimespec(timeout, timeout_buffer);
 		Uring::Queue &uring_queue = *uring;