IoContext: register the worker io_uring fds on Linux >= 5.18

Linux 5.18 introduces IORING_REGISTER_RING_FDS with
e7a6c00dc77aedf27a601738ea509f1caea6d673.

Registering the io_uring's fd prevents having to look it up for each
io_uring_enter call reducing contention on the process file table.
Jens Axboe reports good results in his fio based benchmarks and
I see no reason for EMPER to not register the io_uring fds, especially
because we never pass or share rings.

Do not register the global io_uring since it is shared in the SINGLE_URING
case or it is passed by the main thread to the completer thread breaking the
assumption liburing has about the registered io_uring fd.

emper/io/IoContext.cpp

@@ -510,6 +510,7 @@ template auto IoContext::reapCompletionsLocked<CallerEnvironment::ANYWHERE,
 
 IoContext::IoContext(Runtime &runtime, size_t uring_entries, unsigned unboundedIowMax)
 		: runtime(runtime) {
+	const bool isWorkerIo = runtime.globalIo != nullptr;
 	struct io_uring_params params;
 	memset(&params, 0, sizeof(params));
 
@@ -528,9 +529,8 @@ IoContext::IoContext(Runtime &runtime, size_t uring_entries, unsigned unboundedI
 
 			// We use a single poller -> attach to the global io_uring
 			if constexpr (emper::IO_SQ_POLLER == IoSqPoller::one) {
-				auto *gio = runtime.globalIo;
-				if (gio) {
-					pollerRingFd = gio->ring.ring_fd;
+				if (isWorkerIo) {
+					pollerRingFd = runtime.globalIo->ring.ring_fd;
 				}
 				// One sq poller thread per numa node
 				// The CPU with the lowest id in a node is responsible to create a poller.
@@ -593,6 +593,25 @@ IoContext::IoContext(Runtime &runtime, size_t uring_entries, unsigned unboundedI
 		}
 	}
 
+	// Detect IORING_REGISTER_RING_FDS support
+	// Liburing hides the details of calling io_uring_enter with the index of the
+	// registered io_uring fd instead of the fd itself.
+	// We only have to register our io_uring's fd.
+	// Added with Linux: e7a6c00dc77aedf27a601738ea509f1caea6d673
+
+	// Do not register the ring fd on the global io_uring because it is created on
+	// the main thread and afterwards used on the completer thread.
+	// Registering on the main thread breaks the assumption done in
+	// liburing that the registered offset is valid on the using thread.
+	// Registering the global io_uring is also not possible when using SINGLE_URING.
+	if (EMPER_LINUX_GE("5.18") && isWorkerIo) {
+		ret = io_uring_register_ring_fd(&ring);
+		if (unlikely(ret < 0)) {
+			errno = -ret;
+			DIE_MSG_ERRNO("registering our io_uring fd " << ring.ring_fd << " failed");
+		}
+	}
+
 	// Limit the number of unbounded iow thread created by this io_uring
 	// By default the number of unbounded iow worker threads is RLIMIT_NPROC
 	// which may be excessive.