os.notify: implement the kqueue backend for notify.FdNotifier (#19057)

vlib/os/notify/backend_darwin.c.v

@@ -0,0 +1,230 @@
+module notify
+
+import time
+import os
+
+#insert "@VEXEROOT/vlib/os/notify/kqueue.h"
+
+struct C.kevent {
+mut:
+	ident  u32
+	filter i16
+	flags  u16
+	fflags u32
+	data   int
+	udata  voidptr
+}
+
+fn C.kqueue() int
+fn C.__kevent__(int, voidptr, int, voidptr, int, voidptr) int
+fn C.EV_SET(voidptr, u32, i16, u16, u32, int, voidptr)
+
+// KqueueNotifier provides methods that implement FdNotifier using the
+// kqueue I/O event notification facility (macos, freeBSD, xxxBSD...unix only)
+[noinit]
+struct KqueueNotifier {
+	kqueue_fd int
+}
+
+// KqueueEvent describes an event that occurred for a file descriptor in
+// the watch list
+[noinit]
+struct KqueueEvent {
+pub:
+	fd   int
+	kind FdEventType
+}
+
+// new creates a new KqueueNotifier
+// The FdNotifier interface is returned to allow OS specific
+// implementations without exposing the concrete type
+pub fn new() !FdNotifier {
+	fd := C.kqueue()
+	if fd == -1 {
+		return error(os.posix_get_error_msg(C.errno))
+	}
+	// Needed to circumvent V limitations
+	x := &KqueueNotifier{
+		kqueue_fd: fd
+	}
+	return x
+}
+
+const (
+	// filter types
+	kqueue_read           = i16(C.EVFILT_READ)
+	kqueue_write          = i16(C.EVFILT_WRITE)
+	kqueue_aio            = i16(C.EVFILT_AIO)
+	kqueue_vnode          = i16(C.EVFILT_VNODE)
+	kqueue_proc           = i16(C.EVFILT_PROC)
+	kqueue_signal         = i16(C.EVFILT_SIGNAL)
+	kqueue_timer          = i16(C.EVFILT_TIMER)
+	kqueue_machport       = i16(C.EVFILT_MACHPORT)
+	kqueue_fs             = i16(C.EVFILT_FS)
+	kqueue_user           = i16(C.EVFILT_USER)
+	kqueue_vm             = i16(C.EVFILT_VM)
+	kqueue_exception      = i16(C.EVFILT_EXCEPT)
+	kqueue_syscount       = i16(C.EVFILT_SYSCOUNT)
+
+	// actions
+	kqueue_add            = u16(C.EV_ADD)
+	kqueue_delete         = u16(C.EV_DELETE)
+	kqueue_enable         = u16(C.EV_ENABLE)
+	kqueue_disable        = u16(C.EV_DISABLE)
+
+	// flags
+	kqueue_oneshot        = u16(C.EV_ONESHOT)
+	kqueue_edge_trigger   = u16(C.EV_CLEAR) // kqueue_clear
+	kqueue_receipt        = u16(C.EV_RECEIPT)
+	kqueue_dispatch       = u16(C.EV_DISPATCH)
+	kqueue_udata_specific = u16(C.EV_UDATA_SPECIFIC)
+	kqueue_dispatch2      = u16(C.EV_DISPATCH | C.EV_UDATA_SPECIFIC)
+	kqueue_vanished       = u16(C.EV_VANISHED)
+	kqueue_sysflags       = u16(C.EV_SYSFLAGS)
+	kqueue_flag0          = u16(C.EV_FLAG0)
+	kqueue_flag1          = u16(C.EV_FLAG1)
+
+	// returned values
+	kqueue_eof            = u16(C.EV_EOF)
+	kqueue_error          = u16(C.EV_ERROR)
+)
+
+// ctl is a helper method for add, modify, and remove
+fn (mut kn KqueueNotifier) ctl(fd int, filter i16, flags u16) ! {
+	event := [1]C.kevent{}
+	C.EV_SET(&event[0], fd, filter, flags, 0, 0, unsafe { nil })
+	if C.__kevent__(kn.kqueue_fd, &event[0], 1, unsafe { nil }, 0, unsafe { nil }) == -1 {
+		return error(os.posix_get_error_msg(C.errno))
+	}
+}
+
+// add adds a file descriptor to the watch list
+fn (mut kn KqueueNotifier) add(fd int, events FdEventType, conf ...FdConfigFlags) ! {
+	filter := filter_to_mask(events)
+	flags := flags_to_mask(...conf)
+	kn.ctl(fd, filter, flags)!
+}
+
+// modify sets an existing entry in the watch list to the provided events and configuration
+fn (mut kn KqueueNotifier) modify(fd int, events FdEventType, conf ...FdConfigFlags) ! {
+	kn.add(fd, events, ...conf)!
+}
+
+// remove removes a file descriptor from the watch list
+fn (mut kn KqueueNotifier) remove(fd int) ! {
+	filter := notify.kqueue_read | notify.kqueue_write | notify.kqueue_exception
+	flags := notify.kqueue_delete
+	kn.ctl(fd, filter, flags)!
+}
+
+// wait waits to be notified of events on the watch list,
+// returns at most 512 events
+fn (mut kn KqueueNotifier) wait(timeout time.Duration) []FdEvent {
+	// arbitrary 512 limit; events will round robin on successive
+	// waits if the number exceeds this
+	// NOTE: we use a fixed size array here for stack allocation; this has
+	//       the added bonus of making KqueueNotifier thread safe
+	events := [512]C.kevent{}
+	// populate events with the new events
+	to := &C.timespec{0, timeout.nanoseconds()}
+	count := C.__kevent__(kn.kqueue_fd, unsafe { nil }, 0, &events[0], events.len, to)
+
+	if count > 0 {
+		mut arr := []FdEvent{cap: count}
+		for i := 0; i < count; i++ {
+			fd := int(events[i].ident)
+			kind := event_mask_to_flag(events[i].filter, events[i].flags)
+			if kind.is_empty() {
+				// NOTE: tcc only reports the first event for some
+				// reason, leaving subsequent structs in the array as 0
+				// (or possibly garbage)
+				panic('encountered an empty event kind; this is most likely due to using tcc')
+			}
+			arr << &KqueueEvent{
+				fd: fd
+				kind: kind
+			}
+		}
+		return arr
+	}
+	return []
+}
+
+// close closes the KqueueNotifier,
+// any successive calls to add, modify, remove, and wait should fail
+fn (mut kn KqueueNotifier) close() ! {
+	if C.close(kn.kqueue_fd) == -1 {
+		return error(os.posix_get_error_msg(C.errno))
+	}
+}
+
+// event_mask_to_flag is a helper function that converts a bitmask
+// returned by kevent() wait to FdEventType
+fn event_mask_to_flag(filter i16, flags u16) FdEventType {
+	mut res := FdEventType.read
+
+	if filter & notify.kqueue_read != 0 {
+		res.set(.read)
+	}
+	if filter & notify.kqueue_write != 0 {
+		res.set(.write)
+	}
+	if filter & notify.kqueue_exception != 0 {
+		res.set(.exception)
+	}
+
+	if flags & notify.kqueue_eof != 0 {
+		res.set(.hangup)
+	}
+	if flags & notify.kqueue_error != 0 {
+		res.set(.error)
+	}
+
+	return res
+}
+
+// filter_to_mask is a helper function that converts FdEventType
+// to a bitmask used by the C functions
+fn filter_to_mask(events FdEventType) i16 {
+	mut mask := i16(0)
+	if events.has(.read) {
+		mask |= notify.kqueue_read
+	}
+	if events.has(.write) {
+		mask |= notify.kqueue_write
+	}
+	if events.has(.exception) {
+		mask |= notify.kqueue_exception
+	}
+	if events.has(.peer_hangup) {
+		panic("Kqueue does not support 'peer_hangup' event type.")
+	}
+	if events.has(.error) {
+		panic("Kqueue does not support 'error' event type.")
+	}
+	if events.has(.hangup) {
+		panic("Kqueue does not support 'hangup' event type.")
+	}
+	return mask
+}
+
+// flags_to_mask is a helper function that converts FdConfigFlags
+// to a bitmask used by the C functions
+fn flags_to_mask(confs ...FdConfigFlags) u16 {
+	mut mask := notify.kqueue_add | notify.kqueue_enable
+	for conf in confs {
+		if conf.has(.edge_trigger) {
+			mask |= notify.kqueue_edge_trigger
+		}
+		if conf.has(.one_shot) {
+			mask |= notify.kqueue_oneshot
+		}
+		if conf.has(.wake_up) {
+			panic("Kqueue does not support 'wake_up' flag.")
+		}
+		if conf.has(.exclusive) {
+			panic("Kqueue does not support 'exclusive' flag.")
+		}
+	}
+	return mask
+}

vlib/os/notify/kqueue.h

@@ -0,0 +1,12 @@
+#ifndef __KQUEUE_H
+#define __KQUEUE_H
+
+#include <sys/event.h>
+
+// Due to the renaming of 'struct kevent' and function 'kevent',
+// they are wrapped here to avoid conflicts.
+int __kevent__(int handle, const struct kevent* changelist, int nchanges, struct kevent* eventlist, int nevents, const struct timespec* timeout) {
+	return kevent(handle, changelist, nchanges, eventlist, nevents, timeout);
+}
+
+#endif

vlib/os/notify/notify_test.v

@@ -5,7 +5,7 @@ import os.notify
 
 // make a pipe and return the (read, write) file descriptors
 fn make_pipe() !(int, int) {
-	$if linux {
+	$if linux || macos {
 		pipefd := [2]int{}
 		if C.pipe(&pipefd[0]) != 0 {
 			return error('error ${C.errno}: ' + os.posix_get_error_msg(C.errno))
@@ -16,8 +16,8 @@ fn make_pipe() !(int, int) {
 }
 
 fn test_level_trigger() {
-	// currently only linux is supported
-	$if linux {
+	// currently only linux and macos are supported
+	$if linux || macos {
 		mut notifier := notify.new()!
 		reader, writer := make_pipe()!
 		defer {
@@ -37,8 +37,8 @@ fn test_level_trigger() {
 }
 
 fn test_edge_trigger() {
-	// currently only linux is supported
-	$if linux {
+	// currently only linux and macos are supported
+	$if linux || macos {
 		mut notifier := notify.new()!
 		reader, writer := make_pipe()!
 		defer {
@@ -53,7 +53,27 @@ fn test_edge_trigger() {
 		os.fd_write(writer, 'foobar')
 		check_read_event(mut n, reader, 'foo')
 
+		$if linux {
 			assert notifier.wait(0).len == 0
+		}
+		$if macos {
+			/*
+			In the kqueue of macos, EV_CLEAR flag represents a clear event,
+			which is mainly used for pipeline and socket class events. When this flag is set,
+			kqueue will trigger the corresponding event when the data is readable or writable,
+			but it is not guaranteed that the event will only be triggered once.
+			Compared to EPOLLET, EV_CLEAR's behavior varies. In epoll, the edge triggered mode only triggers
+			an event once when the state changes from unreadable/non writable to readable/writable,
+			that is, when the data changes from unreadable to readable,
+			or when the data changes from unreadable to writable. In the kqueue of macos,
+			EV_CLEAR does not possess this precise edge triggering behavior.
+			Therefore, in the kqueue of macos, even if the data is not completely read,
+			it is possible to continue triggering read events. This means that if you don't process all the data,
+			the next kqueue event notification may still be triggered
+			*/
+
+			// notifier.wait(0).len == 1 or 0
+		}
 
 		os.fd_write(writer, 'baz')
 		// we do not get an event because there is still data
@@ -65,7 +85,7 @@ fn test_edge_trigger() {
 }
 
 fn test_one_shot() {
-	$if linux {
+	$if linux || macos {
 		mut notifier := notify.new()!
 		reader, writer := make_pipe()!
 		defer {
@@ -89,6 +109,7 @@ fn test_one_shot() {
 	}
 }
 
+// Kqueue does not support 'hangup' event type.
 fn test_hangup() {
 	$if linux {
 		mut notifier := notify.new()!
@@ -112,7 +133,7 @@ fn test_hangup() {
 }
 
 fn test_write() {
-	$if linux {
+	$if linux || macos {
 		mut notifier := notify.new()!
 		reader, writer := make_pipe()!
 		defer {
@@ -133,7 +154,7 @@ fn test_write() {
 }
 
 fn test_remove() {
-	$if linux {
+	$if linux || macos {
 		mut notifier := notify.new()!
 		reader, writer := make_pipe()!
 		defer {