event 0.1.2 → 0.4.0

data/ext/event/backend/kqueue.c

@@ -19,6 +19,7 @@
 // THE SOFTWARE.
 
 #include "kqueue.h"
+#include "backend.h"
 
 #include <sys/event.h>
 #include <sys/ioctl.h>
@@ -27,9 +28,7 @@
 static VALUE Event_Backend_KQueue = Qnil;
 static ID id_fileno, id_transfer;
 
-static const int READABLE = 1, PRIORITY = 2, WRITABLE = 4;
-
-static const unsigned KQUEUE_MAX_EVENTS = 1024;
+enum {KQUEUE_MAX_EVENTS = 64};
 
 struct Event_Backend_KQueue {
 	VALUE loop;
@@ -42,13 +41,19 @@ void Event_Backend_KQueue_Type_mark(void *_data)
 	rb_gc_mark(data->loop);
 }
 
+static
+void close_internal(struct Event_Backend_KQueue *data) {
+	if (data->descriptor >= 0) {
+		close(data->descriptor);
+		data->descriptor = -1;
+	}
+}
+
 void Event_Backend_KQueue_Type_free(void *_data)
 {
 	struct Event_Backend_KQueue *data = _data;
 	
-	if (data->descriptor >= 0) {
-		close(data->descriptor);
-	}
+	close_internal(data);
 	
 	free(data);
 }
@@ -98,41 +103,121 @@ VALUE Event_Backend_KQueue_initialize(VALUE self, VALUE loop) {
 	return self;
 }
 
-VALUE Event_Backend_KQueue_io_wait(VALUE self, VALUE fiber, VALUE io, VALUE events) {
+VALUE Event_Backend_KQueue_close(VALUE self) {
 	struct Event_Backend_KQueue *data = NULL;
 	TypedData_Get_Struct(self, struct Event_Backend_KQueue, &Event_Backend_KQueue_Type, data);
 	
-	struct kevent event;
-	u_short flags = 0;
-	
-	int descriptor = NUM2INT(rb_funcall(io, id_fileno, 0));
+	close_internal(data);
 	
-	int mask = NUM2INT(events);
+	return Qnil;
+}
+
+static
+int io_add_filters(int descriptor, int ident, int events, VALUE fiber) {
+	int count = 0;
+	struct kevent kevents[2] = {0};
 	
-	if (mask & READABLE) {
-		flags |= EVFILT_READ;
+	if (events & READABLE) {
+		kevents[count].ident = ident;
+		kevents[count].filter = EVFILT_READ;
+		kevents[count].flags = EV_ADD | EV_ENABLE | EV_ONESHOT;
+		kevents[count].udata = (void*)fiber;
+		
+// #ifdef EV_OOBAND
+// 		if (events & PRIORITY) {
+// 			kevents[count].flags |= EV_OOBAND;
+// 		}
+// #endif
+		
+		count++;
 	}
 	
-	if (mask & PRIORITY) {
-		flags |= EV_OOBAND;
+	if (events & WRITABLE) {
+		kevents[count].ident = ident;
+		kevents[count].filter = EVFILT_WRITE;
+		kevents[count].flags = EV_ADD | EV_ENABLE | EV_ONESHOT;
+		kevents[count].udata = (void*)fiber;
+		count++;
 	}
 	
-	if (mask & WRITABLE) {
-		flags |= EVFILT_WRITE;
+	int result = kevent(descriptor, kevents, count, NULL, 0, NULL);
+	
+	if (result == -1) {
+		rb_sys_fail("kevent(register)");
 	}
+	
+	return events;
+}
 
-	EV_SET(&event, descriptor, flags, EV_ADD|EV_ENABLE|EV_ONESHOT, 0, 0, (void*)fiber);
+static
+void io_remove_filters(int descriptor, int ident, int events) {
+	int count = 0;
+	struct kevent kevents[2] = {0};
 	
-	// A better approach is to batch all changes:
-	int result = kevent(data->descriptor, &event, 1, NULL, 0, NULL);
+	if (events & READABLE) {
+		kevents[count].ident = ident;
+		kevents[count].filter = EVFILT_READ;
+		kevents[count].flags = EV_DELETE;
+		
+		count++;
+	}
 	
-	if (result == -1) {
-		rb_sys_fail("kevent");
+	if (events & WRITABLE) {
+		kevents[count].ident = ident;
+		kevents[count].filter = EVFILT_WRITE;
+		kevents[count].flags = EV_DELETE;
+		count++;
 	}
 	
-	rb_funcall(data->loop, id_transfer, 0);
+	// Ignore the result.
+	kevent(descriptor, kevents, count, NULL, 0, NULL);
+}
+
+struct io_wait_arguments {
+	struct Event_Backend_KQueue *data;
+	int events;
+	int descriptor;
+};
+
+static
+VALUE io_wait_rescue(VALUE _arguments, VALUE exception) {
+	struct io_wait_arguments *arguments = (struct io_wait_arguments *)_arguments;
+	
+	io_remove_filters(arguments->data->descriptor, arguments->descriptor, arguments->events);
 	
-	return Qnil;
+	rb_exc_raise(exception);
+};
+
+static inline
+int events_from_kqueue_filter(int filter) {
+	if (filter == EVFILT_READ) return READABLE;
+	if (filter == EVFILT_WRITE) return WRITABLE;
+	
+	return 0;
+}
+
+static
+VALUE io_wait_transfer(VALUE _arguments) {
+	struct io_wait_arguments *arguments = (struct io_wait_arguments *)_arguments;
+	
+	VALUE result = rb_funcall(arguments->data->loop, id_transfer, 0);
+	
+	return INT2NUM(events_from_kqueue_filter(NUM2INT(result)));
+};
+
+VALUE Event_Backend_KQueue_io_wait(VALUE self, VALUE fiber, VALUE io, VALUE events) {
+	struct Event_Backend_KQueue *data = NULL;
+	TypedData_Get_Struct(self, struct Event_Backend_KQueue, &Event_Backend_KQueue_Type, data);
+	
+	int descriptor = NUM2INT(rb_funcall(io, id_fileno, 0));
+	
+	struct io_wait_arguments io_wait_arguments = {
+		.events = io_add_filters(data->descriptor, descriptor, NUM2INT(events), fiber),
+		.data = data,
+		.descriptor = descriptor,
+	};
+	
+	return rb_rescue(io_wait_transfer, (VALUE)&io_wait_arguments, io_wait_rescue, (VALUE)&io_wait_arguments);
 }
 
 static
@@ -150,7 +235,7 @@ struct timespec * make_timeout(VALUE duration, struct timespec * storage) {
 	
 	else if (RB_FLOAT_TYPE_P(duration)) {
 		double value = RFLOAT_VALUE(duration);
-		time_t seconds = duration;
+		time_t seconds = value;
 		
 		storage->tv_sec = seconds;
 		storage->tv_nsec = (value - seconds) * 1000000000L;
@@ -161,25 +246,89 @@ struct timespec * make_timeout(VALUE duration, struct timespec * storage) {
 	rb_raise(rb_eRuntimeError, "unable to convert timeout");
 }
 
+static
+int timeout_nonblocking(struct timespec * timespec) {
+	return timespec && timespec->tv_sec == 0 && timespec->tv_nsec == 0;
+}
+
+struct select_arguments {
+	struct Event_Backend_KQueue *data;
+	
+	int count;
+	struct kevent events[KQUEUE_MAX_EVENTS];
+	
+	struct timespec storage;
+	struct timespec *timeout;
+};
+
+static
+void * select_internal(void *_arguments) {
+	struct select_arguments * arguments = (struct select_arguments *)_arguments;
+	
+	arguments->count = kevent(arguments->data->descriptor, NULL, 0, arguments->events, arguments->count, arguments->timeout);
+	
+	return NULL;
+}
+
+static
+void select_internal_without_gvl(struct select_arguments *arguments) {
+	rb_thread_call_without_gvl(select_internal, (void *)arguments, RUBY_UBF_IO, 0);
+	
+	if (arguments->count == -1) {
+		rb_sys_fail("select_internal_without_gvl:kevent");
+	}
+}
+
+static
+void select_internal_with_gvl(struct select_arguments *arguments) {
+	select_internal((void *)arguments);
+	
+	if (arguments->count == -1) {
+		rb_sys_fail("select_internal_with_gvl:kevent");
+	}
+}
+
 VALUE Event_Backend_KQueue_select(VALUE self, VALUE duration) {
 	struct Event_Backend_KQueue *data = NULL;
 	TypedData_Get_Struct(self, struct Event_Backend_KQueue, &Event_Backend_KQueue_Type, data);
 	
-	struct kevent events[KQUEUE_MAX_EVENTS];
-	struct timespec storage;
+	struct select_arguments arguments = {
+		.data = data,
+		.count = KQUEUE_MAX_EVENTS,
+		.storage = {
+			.tv_sec = 0,
+			.tv_nsec = 0
+		}
+	};
 	
-	int count = kevent(data->descriptor, NULL, 0, events, KQUEUE_MAX_EVENTS, make_timeout(duration, &storage));
+	// We break this implementation into two parts.
+	// (1) count = kevent(..., timeout = 0)
+	// (2) without gvl: kevent(..., timeout = 0) if count == 0 and timeout != 0
+	// This allows us to avoid releasing and reacquiring the GVL.
+	// Non-comprehensive testing shows this gives a 1.5x speedup.
+	arguments.timeout = &arguments.storage;
 	
-	if (count == -1) {
-		rb_sys_fail("kevent");
+	// First do the syscall with no timeout to get any immediately available events:
+	select_internal_with_gvl(&arguments);
+	
+	// If there were no pending events, if we have a timeout, wait for more events:
+	if (arguments.count == 0) {
+		arguments.timeout = make_timeout(duration, &arguments.storage);
+		
+		if (!timeout_nonblocking(arguments.timeout)) {
+			arguments.count = KQUEUE_MAX_EVENTS;
+			
+			select_internal_without_gvl(&arguments);
+		}
 	}
 	
-	for (int i = 0; i < count; i += 1) {
-		VALUE fiber = (VALUE)events[i].udata;
-		rb_funcall(fiber, id_transfer, 0);
+	for (int i = 0; i < arguments.count; i += 1) {
+		VALUE fiber = (VALUE)arguments.events[i].udata;
+		VALUE result = INT2NUM(arguments.events[i].filter);
+		rb_funcall(fiber, id_transfer, 1, result);
 	}
 	
-	return INT2NUM(count);
+	return INT2NUM(arguments.count);
 }
 
 void Init_Event_Backend_KQueue(VALUE Event_Backend) {
@@ -190,6 +339,7 @@ void Init_Event_Backend_KQueue(VALUE Event_Backend) {
 	
 	rb_define_alloc_func(Event_Backend_KQueue, Event_Backend_KQueue_allocate);
 	rb_define_method(Event_Backend_KQueue, "initialize", Event_Backend_KQueue_initialize, 1);
+	rb_define_method(Event_Backend_KQueue, "initialize", Event_Backend_KQueue_close, 0);
 	
 	rb_define_method(Event_Backend_KQueue, "io_wait", Event_Backend_KQueue_io_wait, 3);
 	rb_define_method(Event_Backend_KQueue, "select", Event_Backend_KQueue_select, 1);

data/ext/event/backend/uring.c

@@ -19,21 +19,21 @@
 // THE SOFTWARE.
 
 #include "uring.h"
+#include "backend.h"
 
 #include <liburing.h>
+#include <poll.h>
 #include <time.h>
 
 static VALUE Event_Backend_URing = Qnil;
 static ID id_fileno, id_transfer;
 
-static const int READABLE = 1, PRIORITY = 2, WRITABLE = 4;
-
-static const int URING_ENTRIES = 1024;
-static const int URING_MAX_EVENTS = 1024;
+enum {URING_ENTRIES = 128};
+enum {URING_MAX_EVENTS = 128};
 
 struct Event_Backend_URing {
 	VALUE loop;
-	struct io_uring* ring;
+	struct io_uring ring;
 };
 
 void Event_Backend_URing_Type_mark(void *_data)
@@ -42,14 +42,19 @@ void Event_Backend_URing_Type_mark(void *_data)
 	rb_gc_mark(data->loop);
 }
 
+static
+void close_internal(struct Event_Backend_URing *data) {
+	if (data->ring.ring_fd >= 0) {
+		io_uring_queue_exit(&data->ring);
+		data->ring.ring_fd = -1;
+	}
+}
+
 void Event_Backend_URing_Type_free(void *_data)
 {
 	struct Event_Backend_URing *data = _data;
 	
-	if (data->ring) {
-		io_uring_queue_exit(data->ring);
-		xfree(data->ring);
-	}
+	close_internal(data);
 	
 	free(data);
 }
@@ -75,7 +80,7 @@ VALUE Event_Backend_URing_allocate(VALUE self) {
 	VALUE instance = TypedData_Make_Struct(self, struct Event_Backend_URing, &Event_Backend_URing_Type, data);
 	
 	data->loop = Qnil;
-	data->ring = NULL;
+	data->ring.ring_fd = -1;
 	
 	return instance;
 }
@@ -85,85 +90,130 @@ VALUE Event_Backend_URing_initialize(VALUE self, VALUE loop) {
 	TypedData_Get_Struct(self, struct Event_Backend_URing, &Event_Backend_URing_Type, data);
 	
 	data->loop = loop;
-	data->ring = xmalloc(sizeof(struct io_uring));
 	
-	int result = io_uring_queue_init(URING_ENTRIES, data->ring, 0);
+	int result = io_uring_queue_init(URING_ENTRIES, &data->ring, 0);
 	
-	if (result == -1) {
-		rb_sys_fail("io_uring_queue_init");
+	if (result < 0) {
+		rb_syserr_fail(-result, "io_uring_queue_init");
 	}
 	
+	rb_update_max_fd(data->ring.ring_fd);
+	
 	return self;
 }
 
-VALUE Event_Backend_URing_io_wait(VALUE self, VALUE fiber, VALUE io, VALUE events) {
+VALUE Event_Backend_URing_close(VALUE self) {
 	struct Event_Backend_URing *data = NULL;
 	TypedData_Get_Struct(self, struct Event_Backend_URing, &Event_Backend_URing_Type, data);
 	
-	int descriptor = NUM2INT(rb_funcall(io, id_fileno, 0));
-	struct io_uring_sqe *sqe = io_uring_get_sqe(data->ring);
+	close_internal(data);
 	
-	int mask = NUM2INT(events);
+	return Qnil;
+}
+
+static inline
+short poll_flags_from_events(int events) {
 	short flags = 0;
 	
-	if (mask & READABLE) {
-		flags |= POLL_IN;
-	}
+	if (events & READABLE) flags |= POLLIN;
+	if (events & PRIORITY) flags |= POLLPRI;
+	if (events & WRITABLE) flags |= POLLOUT;
 	
-	if (mask & PRIORITY) {
-		flags |= POLL_PRI;
-	}
+	flags |= POLLERR;
+	flags |= POLLHUP;
 	
-	if (mask & WRITABLE) {
-		flags |= POLL_OUT;
-	}
-
-	// fprintf(stderr, "poll_add(%p, %d, %d)\n", sqe, descriptor, flags);
+	return flags;
+}
 
-	io_uring_prep_poll_add(sqe, descriptor, flags);
-	io_uring_sqe_set_data(sqe, (void*)fiber);
-	io_uring_submit(data->ring);
+static inline
+int events_from_poll_flags(short flags) {
+	int events = 0;
+	
+	if (flags & POLLIN) events |= READABLE;
+	if (flags & POLLPRI) events |= PRIORITY;
+	if (flags & POLLOUT) events |= WRITABLE;
 	
-	// fprintf(stderr, "count = %d, errno = %d\n", count, errno);
+	return events;
+}
+
+struct io_wait_arguments {
+	struct Event_Backend_URing *data;
+	VALUE fiber;
+	short flags;
+};
 
-	rb_funcall(data->loop, id_transfer, 0);
+struct io_uring_sqe * io_get_sqe(struct Event_Backend_URing *data) {
+	struct io_uring_sqe *sqe = io_uring_get_sqe(&data->ring);
 	
-	return Qnil;
+	while (sqe == NULL) {
+		sqe = io_uring_get_sqe(&data->ring);
+	}
+	
+	return sqe;
 }
 
 static
-struct __kernel_timespec * make_timeout(VALUE duration, struct __kernel_timespec *storage) {
-	if (duration == Qnil) {
-		return NULL;
-	}
+VALUE io_wait_rescue(VALUE _arguments, VALUE exception) {
+	struct io_wait_arguments *arguments = (struct io_wait_arguments *)_arguments;
+	struct Event_Backend_URing *data = arguments->data;
 	
-	if (FIXNUM_P(duration)) {
-		storage->tv_sec = NUM2TIMET(duration);
-		storage->tv_nsec = 0;
-		
-		return storage;
-	}
+	struct io_uring_sqe *sqe = io_get_sqe(data);
 	
-	else if (RB_FLOAT_TYPE_P(duration)) {
-		double value = RFLOAT_VALUE(duration);
-		time_t seconds = duration;
-		
-		storage->tv_sec = seconds;
-		storage->tv_nsec = (value - seconds) * 1000000000L;
-		
-		return storage;
-	}
+	// fprintf(stderr, "poll_remove(%p, %p)\n", sqe, (void*)arguments->fiber);
 	
-	rb_raise(rb_eRuntimeError, "unable to convert timeout");
+	io_uring_prep_poll_remove(sqe, (void*)arguments->fiber);
+	io_uring_submit(&data->ring);
+	
+	rb_exc_raise(exception);
+};
+
+static
+VALUE io_wait_transfer(VALUE _arguments) {
+	struct io_wait_arguments *arguments = (struct io_wait_arguments *)_arguments;
+	struct Event_Backend_URing *data = arguments->data;
+	
+	VALUE result = rb_funcall(data->loop, id_transfer, 0);
+	
+	// We explicitly filter the resulting events based on the requested events.
+	// In some cases, poll will report events we didn't ask for.
+	short flags = arguments->flags & NUM2INT(result);
+	
+	return INT2NUM(events_from_poll_flags(flags));
+};
+
+VALUE Event_Backend_URing_io_wait(VALUE self, VALUE fiber, VALUE io, VALUE events) {
+	struct Event_Backend_URing *data = NULL;
+	TypedData_Get_Struct(self, struct Event_Backend_URing, &Event_Backend_URing_Type, data);
+	
+	int descriptor = NUM2INT(rb_funcall(io, id_fileno, 0));
+	struct io_uring_sqe *sqe = io_get_sqe(data);
+	
+	if (!sqe) return INT2NUM(0);
+	
+	short flags = poll_flags_from_events(NUM2INT(events));
+	
+	// fprintf(stderr, "poll_add(%p, %d, %d, %p)\n", sqe, descriptor, flags, (void*)fiber);
+	
+	io_uring_prep_poll_add(sqe, descriptor, flags);
+	io_uring_sqe_set_data(sqe, (void*)fiber);
+	io_uring_submit(&data->ring);
+	
+	struct io_wait_arguments io_wait_arguments = {
+		.data = data,
+		.fiber = fiber,
+		.flags = flags
+	};
+	
+	return rb_rescue(io_wait_transfer, (VALUE)&io_wait_arguments, io_wait_rescue, (VALUE)&io_wait_arguments);
 }
 
 inline static
 void resize_to_capacity(VALUE string, size_t offset, size_t length) {
 	size_t current_length = RSTRING_LEN(string);
 	long difference = (long)(offset + length) - (long)current_length;
-
+	
 	difference += 1;
-
+	
 	if (difference > 0) {
 		rb_str_modify_expand(string, difference);
 	} else {
@@ -174,7 +224,7 @@ void resize_to_capacity(VALUE string, size_t offset, size_t length) {
 inline static
 void resize_to_fit(VALUE string, size_t offset, size_t length) {
 	size_t current_length = RSTRING_LEN(string);
-
+	
 	if (current_length < (offset + length)) {
 		rb_str_set_len(string, offset + length);
 	}
@@ -183,92 +233,174 @@ void resize_to_fit(VALUE string, size_t offset, size_t length) {
 VALUE Event_Backend_URing_io_read(VALUE self, VALUE fiber, VALUE io, VALUE buffer, VALUE offset, VALUE length) {
 	struct Event_Backend_URing *data = NULL;
 	TypedData_Get_Struct(self, struct Event_Backend_URing, &Event_Backend_URing_Type, data);
-
+	
 	resize_to_capacity(buffer, NUM2SIZET(offset), NUM2SIZET(length));
-
+	
 	int descriptor = NUM2INT(rb_funcall(io, id_fileno, 0));
-	struct io_uring_sqe *sqe = io_uring_get_sqe(data->ring);
-
+	struct io_uring_sqe *sqe = io_get_sqe(data);
+	
 	struct iovec iovecs[1];
 	iovecs[0].iov_base = RSTRING_PTR(buffer) + NUM2SIZET(offset);
 	iovecs[0].iov_len = NUM2SIZET(length);
-
+	
 	io_uring_prep_readv(sqe, descriptor, iovecs, 1, 0);
 	io_uring_sqe_set_data(sqe, (void*)fiber);
-	io_uring_submit(data->ring);
+	io_uring_submit(&data->ring);
+	
+	// fprintf(stderr, "prep_readv(%p, %d, %ld)\n", sqe, descriptor, iovecs[0].iov_len);
 	
 	int result = NUM2INT(rb_funcall(data->loop, id_transfer, 0));
-
+	
 	if (result < 0) {
 		rb_syserr_fail(-result, strerror(-result));
 	}
-
+	
 	resize_to_fit(buffer, NUM2SIZET(offset), (size_t)result);
-
+	
 	return INT2NUM(result);
 }
 
 VALUE Event_Backend_URing_io_write(VALUE self, VALUE fiber, VALUE io, VALUE buffer, VALUE offset, VALUE length) {
 	struct Event_Backend_URing *data = NULL;
 	TypedData_Get_Struct(self, struct Event_Backend_URing, &Event_Backend_URing_Type, data);
-
+	
 	if ((size_t)RSTRING_LEN(buffer) < NUM2SIZET(offset) + NUM2SIZET(length)) {
 		rb_raise(rb_eRuntimeError, "invalid offset/length exceeds bounds of buffer");
 	}
-
+	
 	int descriptor = NUM2INT(rb_funcall(io, id_fileno, 0));
-	struct io_uring_sqe *sqe = io_uring_get_sqe(data->ring);
-
+	struct io_uring_sqe *sqe = io_get_sqe(data);
+	
 	struct iovec iovecs[1];
 	iovecs[0].iov_base = RSTRING_PTR(buffer) + NUM2SIZET(offset);
 	iovecs[0].iov_len = NUM2SIZET(length);
-
+	
 	io_uring_prep_writev(sqe, descriptor, iovecs, 1, 0);
 	io_uring_sqe_set_data(sqe, (void*)fiber);
-	io_uring_submit(data->ring);
+	io_uring_submit(&data->ring);
+	
+	// fprintf(stderr, "prep_writev(%p, %d, %ld)\n", sqe, descriptor, iovecs[0].iov_len);
 	
 	int result = NUM2INT(rb_funcall(data->loop, id_transfer, 0));
-
+	
 	if (result < 0) {
 		rb_syserr_fail(-result, strerror(-result));
 	}
-
+	
 	return INT2NUM(result);
 }
 
+static
+struct __kernel_timespec * make_timeout(VALUE duration, struct __kernel_timespec *storage) {
+	if (duration == Qnil) {
+		return NULL;
+	}
+	
+	if (FIXNUM_P(duration)) {
+		storage->tv_sec = NUM2TIMET(duration);
+		storage->tv_nsec = 0;
+		
+		return storage;
+	}
+	
+	else if (RB_FLOAT_TYPE_P(duration)) {
+		double value = RFLOAT_VALUE(duration);
+		time_t seconds = value;
+		
+		storage->tv_sec = seconds;
+		storage->tv_nsec = (value - seconds) * 1000000000L;
+		
+		return storage;
+	}
+	
+	rb_raise(rb_eRuntimeError, "unable to convert timeout");
+}
+
+static
+int timeout_nonblocking(struct __kernel_timespec *timespec) {
+	return timespec && timespec->tv_sec == 0 && timespec->tv_nsec == 0;
+}
+
+struct select_arguments {
+	struct Event_Backend_URing *data;
+	
+	int count;
+	struct io_uring_cqe **cqes;
+	
+	struct __kernel_timespec storage;
+	struct __kernel_timespec *timeout;
+};
+
+static
+void * select_internal(void *_arguments) {
+	struct select_arguments * arguments = (struct select_arguments *)_arguments;
+	
+	arguments->count = io_uring_wait_cqes(&arguments->data->ring, arguments->cqes, 1, arguments->timeout, NULL);
+	
+	// If waiting resulted in a timeout, there are 0 events.
+	if (arguments->count == -ETIME) {
+		arguments->count = 0;
+	}
+	
+	return NULL;
+}
+
+static
+int select_internal_without_gvl(struct select_arguments *arguments) {
+	rb_thread_call_without_gvl(select_internal, (void *)arguments, RUBY_UBF_IO, 0);
+	
+	if (arguments->count < 0) {
+		rb_syserr_fail(-arguments->count, "select_internal_without_gvl:io_uring_wait_cqes");
+	}
+	
+	return arguments->count;
+}
+
 VALUE Event_Backend_URing_select(VALUE self, VALUE duration) {
 	struct Event_Backend_URing *data = NULL;
 	TypedData_Get_Struct(self, struct Event_Backend_URing, &Event_Backend_URing_Type, data);
 	
 	struct io_uring_cqe *cqes[URING_MAX_EVENTS];
-	struct __kernel_timespec storage;
-
-	if (duration != Qnil) {
-		int result = io_uring_wait_cqe_timeout(data->ring, cqes, make_timeout(duration, &storage));
+	
+	// This is a non-blocking operation:
+	int result = io_uring_peek_batch_cqe(&data->ring, cqes, URING_MAX_EVENTS);
+	
+	if (result < 0) {
+		rb_syserr_fail(-result, strerror(-result));
+	} else if (result == 0) {
+		// We might need to wait for events:
+		struct select_arguments arguments = {
+			.data = data,
+			.cqes = cqes,
+			.timeout = NULL,
+		};
+		
+		arguments.timeout = make_timeout(duration, &arguments.storage);
 		
-		if (result == -ETIME) {
-			// Timeout.
-		} else if (result < 0) {
-			rb_syserr_fail(-result, strerror(-result));
+		if (!timeout_nonblocking(arguments.timeout)) {
+			result = select_internal_without_gvl(&arguments);
 		}
 	}
 	
-	int count = io_uring_peek_batch_cqe(data->ring, cqes, URING_MAX_EVENTS);
+	// fprintf(stderr, "cqes count=%d\n", result);
 	
-	if (count == -1) {
-		rb_sys_fail("io_uring_peek_batch_cqe");
-	}
-	
-	for (int i = 0; i < count; i += 1) {
+	for (int i = 0; i < result; i += 1) {
+		// If the operation was cancelled, or the operation has no user data (fiber):
+		if (cqes[i]->res == -ECANCELED || cqes[i]->user_data == 0) {
+			continue;
+		}
+		
 		VALUE fiber = (VALUE)io_uring_cqe_get_data(cqes[i]);
 		VALUE result = INT2NUM(cqes[i]->res);
-
-		io_uring_cqe_seen(data->ring, cqes[i]);
+		
+		// fprintf(stderr, "cqes[i] res=%d user_data=%p\n", cqes[i]->res, (void*)cqes[i]->user_data);
+		
+		io_uring_cqe_seen(&data->ring, cqes[i]);
 		
 		rb_funcall(fiber, id_transfer, 1, result);
 	}
 	
-	return INT2NUM(count);
+	return INT2NUM(result);
 }
 
 void Init_Event_Backend_URing(VALUE Event_Backend) {
@@ -279,10 +411,11 @@ void Init_Event_Backend_URing(VALUE Event_Backend) {
 	
 	rb_define_alloc_func(Event_Backend_URing, Event_Backend_URing_allocate);
 	rb_define_method(Event_Backend_URing, "initialize", Event_Backend_URing_initialize, 1);
+	rb_define_method(Event_Backend_URing, "close", Event_Backend_URing_close, 0);
 	
 	rb_define_method(Event_Backend_URing, "io_wait", Event_Backend_URing_io_wait, 3);
 	rb_define_method(Event_Backend_URing, "select", Event_Backend_URing_select, 1);
-
+	
 	rb_define_method(Event_Backend_URing, "io_read", Event_Backend_URing_io_read, 5);
 	rb_define_method(Event_Backend_URing, "io_write", Event_Backend_URing_io_write, 5);
 }