Swap read on timerfd/eventfd to poll to avoid Kernel 6.15.4 bug (#171)

This commit changes our operations on timerfd and eventfd from using
read to using poll. This is a workaround for a bug in Kernel 6.15.4 that
causes read operations on these file descriptors to hang indefinitely.

A poll is equivalent to what we're trying to achieve here and is
equivalent to epoll because epoll is already polling and `read` was just
a wrapper that subsequently called `read` anyways.

Prior to this commit, tests would hang on 6.15.4, now they pass.

Author: mitchellh

src/backend/io_uring.zig

@@ -1197,9 +1197,12 @@ test "io_uring: timerfd" {
     var called = false;
     var c: Completion = .{
         .op = .{
-            .read = .{
+            // Note: we should be able to use `read` here but on
+            // Kernel 6.15.4 there is a bug that prevents the read
+            // from ever firing with io_uring. I don't know why. I changed
+            // this to a poll so tests pass, which should also be fine!
+            .poll = .{
                 .fd = t.fd,
-                .buffer = .{ .array = undefined },
             },
         },
 

src/watcher/async.zig

@@ -52,13 +52,14 @@ fn AsyncEventFd(comptime xev: type) type {
                     // TODO: error handling
                     .freebsd => eventfd(
                         0,
-                        0x100000, // EFD_CLOEXEC
+                        0x100000 | 0x4, // EFD_CLOEXEC | EFD_NONBLOCK
                     ),
 
                     // Use std.posix if we can.
                     else => try std.posix.eventfd(
                         0,
-                        std.os.linux.EFD.CLOEXEC,
+                        std.os.linux.EFD.CLOEXEC |
+                            std.os.linux.EFD.NONBLOCK,
                     ),
                 },
             };
@@ -80,7 +81,13 @@ fn AsyncEventFd(comptime xev: type) type {
         /// You should NOT register an async with multiple loops (the same loop
         /// is fine -- but unnecessary). The behavior when waiting on multiple
         /// loops is undefined.
-        pub fn wait(
+        pub const wait = switch (xev.backend) {
+            .io_uring, .epoll => waitPoll,
+            .kqueue => waitRead,
+            .iocp, .wasi_poll => @compileError("AsyncEventFd does not support wait for this backend"),
+        };
+
+        fn waitRead(
             self: Self,
             loop: *xev.Loop,
             c: *xev.Completion,
@@ -128,6 +135,63 @@ fn AsyncEventFd(comptime xev: type) type {
             loop.add(c);
         }
 
+        fn waitPoll(
+            self: Self,
+            loop: *xev.Loop,
+            c: *xev.Completion,
+            comptime Userdata: type,
+            userdata: ?*Userdata,
+            comptime cb: *const fn (
+                ud: ?*Userdata,
+                l: *xev.Loop,
+                c: *xev.Completion,
+                r: WaitError!void,
+            ) xev.CallbackAction,
+        ) void {
+            c.* = .{
+                .op = .{
+                    // We use a poll operation instead of a read operation
+                    // because in Kernel 6.15.4, read was regressed for
+                    // io_uring on eventfd/timerfd and would block forever.
+                    // However, poll works fine.
+                    .poll = .{
+                        .fd = self.fd,
+                        .events = posix.POLL.IN,
+                    },
+                },
+
+                .userdata = userdata,
+                .callback = (struct {
+                    fn callback(
+                        ud: ?*anyopaque,
+                        l_inner: *xev.Loop,
+                        c_inner: *xev.Completion,
+                        r: xev.Result,
+                    ) xev.CallbackAction {
+                        if (r.poll) |_| {
+                            // We need to read so that we can consume the
+                            // eventfd value. We only read 8 bytes because
+                            // we only write up to 8 bytes and we own the fd.
+                            // We ignore errors here because we expect the
+                            // read to succeed given we just polled it.
+                            var buf: [8]u8 = undefined;
+                            _ = posix.read(c_inner.op.poll.fd, &buf) catch {};
+                        } else |_| {
+                            // We'll call the callback with the error later.
+                        }
+
+                        return @call(.always_inline, cb, .{
+                            common.userdataValue(Userdata, ud),
+                            l_inner,
+                            c_inner,
+                            if (r.poll) |_| {} else |err| err,
+                        });
+                    }
+                }).callback,
+            };
+            loop.add(c);
+        }
+
         /// Notify a loop to wake up synchronously. This should never block forever
         /// (it will always EVENTUALLY succeed regardless of if the loop is currently
         /// ticking or not).
@@ -676,16 +740,21 @@ fn AsyncTests(comptime xev: type, comptime Impl: type) type {
                 ) xev.CallbackAction {
                     _ = r catch unreachable;
                     ud.?.* = true;
-                    return .disarm;
+                    return .rearm;
                 }
             }).callback);
 
             // Send a notification
             try notifier.notify();
 
             // Wait for wake
-            try loop.run(.until_done);
+            try loop.run(.once);
             try testing.expect(wake);
+
+            // Make sure it only triggers once
+            wake = false;
+            try loop.run(.no_wait);
+            try testing.expect(!wake);
         }
 
         test "async: notify first" {