Try setting up io_uring with IORING_SETUP_SUBMIT_ALL

This decreases the likelihood of a short submit.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: 2 people

ext/io/event/selector/uring.c

@@ -283,9 +283,23 @@ VALUE IO_Event_Selector_URing_initialize(VALUE self, VALUE loop) {
 #ifdef IORING_SETUP_TASKRUN_FLAG
 	flags |= IORING_SETUP_TASKRUN_FLAG;
 #endif
+	// IORING_SETUP_SUBMIT_ALL (kernel 5.18+): keep processing the rest of the SQE
+	// batch even when one fails, reducing the frequency of short submits.
+#ifdef IORING_SETUP_SUBMIT_ALL
+	flags |= IORING_SETUP_SUBMIT_ALL;
+#endif
 
 	int result = io_uring_queue_init(URING_ENTRIES, &selector->ring, flags);
 
+#ifdef IORING_SETUP_SUBMIT_ALL
+	if (result == -EINVAL) {
+		// IORING_SETUP_SUBMIT_ALL was added in Linux 5.18; retry without it.
+		if (DEBUG) fprintf(stderr, "IO_Event_Selector_URing_initialize: no IORING_SETUP_SUBMIT_ALL\n");
+		flags &= ~IORING_SETUP_SUBMIT_ALL;
+		result = io_uring_queue_init(URING_ENTRIES, &selector->ring, flags);
+	}
+#endif
+	
 	if (result < 0) {
 		rb_syserr_fail(-result, "IO_Event_Selector_URing_initialize:io_uring_queue_init");
 	}
@@ -1306,9 +1320,19 @@ static int IO_Event_Selector_URing_supported_p(void) {
 #endif
 #ifdef IORING_SETUP_TASKRUN_FLAG
 	flags |= IORING_SETUP_TASKRUN_FLAG;
+#endif
+#ifdef IORING_SETUP_SUBMIT_ALL
+	flags |= IORING_SETUP_SUBMIT_ALL;
 #endif
 	int result = io_uring_queue_init(32, &ring, flags);
 
+#ifdef IORING_SETUP_SUBMIT_ALL
+	if (result == -EINVAL) {
+		flags &= ~IORING_SETUP_SUBMIT_ALL;
+		result = io_uring_queue_init(32, &ring, flags);
+	}
+#endif
+	
 	if (result < 0) {
 		rb_warn("io_uring_queue_init() was available at compile time but failed at run time: %s\n", strerror(-result));
 

releases.md

@@ -7,6 +7,7 @@
   - Improve `WorkerPool` GC compaction support and add proper write barriers, fixing potential use-after-free under compacting GC.
   - Keep blocked scheduler fibers alive during GC by registering them as roots in `TestScheduler#block`, preventing premature collection and the resulting use-after-free crash on resume.
   - Correctly handle short `io_uring_submit()` results in the `URing` selector. `io_uring_submit()` returns the number of SQEs actually accepted by the kernel and can be short (SQE prep errors, `ENOMEM`, transient `EAGAIN`); the old accounting reset `pending = 0` on any success and silently lost track of unsubmitted SQEs.
+  - Enable `IORING_SETUP_SUBMIT_ALL` (kernel 5.18+) on the `URing` selector so the kernel keeps processing the rest of an SQE batch past individual errors, reducing the frequency of short submits in practice.
 
 ## v1.15.1