From 80972f1e0e6b6d409cc4d86202608c22e5ee3e5a Mon Sep 17 00:00:00 2001
From: Dario Nieuwenhuis <dirbaio@dirbaio.net>
Date: Thu, 30 Mar 2023 17:55:55 +0200
Subject: executor,sync: add support for turbo-wakers.

This is a `core` patch to make wakers 1 word (the task pointer) instead of 2 (task pointer + vtable). It allows having the "waker optimization" we had a while back on `WakerRegistration/AtomicWaker`, but EVERYWHERE, without patching all crates.

Advantages:
- Less memory usage.
- Faster.
- `AtomicWaker` can actually use atomics to load/store the waker, No critical section needed.
- No `dyn` call, which means `cargo-call-stack` can now see through wakes.

Disadvantages:
- You have to patch `core`...
- Breaks all executors and other things that create wakers, unless they opt in to using the new `from_ptr` API.

How to use:

- Run this shell script to patch `core`. https://gist.github.com/Dirbaio/c67da7cf318515181539122c9d32b395
- Enable `build-std`
- Enable `build-std-features = core/turbowakers`
- Enable feature `turbowakers` in `embassy-executor`, `embassy-sync`.
- Make sure you have no other crate creating wakers other than `embassy-executor`. These will panic at runtime.

Note that the patched `core` is equivalent to the unpached one when the `turbowakers` feature is not enabled, so it should be fine to leave it there.
---
 embassy-sync/src/waitqueue/waker_registration.rs | 52 ++++++++++++++++++++++++
 1 file changed, 52 insertions(+)
 create mode 100644 embassy-sync/src/waitqueue/waker_registration.rs

(limited to 'embassy-sync/src/waitqueue/waker_registration.rs')

diff --git a/embassy-sync/src/waitqueue/waker_registration.rs b/embassy-sync/src/waitqueue/waker_registration.rs
new file mode 100644
index 000000000..9b666e7c4
--- /dev/null
+++ b/embassy-sync/src/waitqueue/waker_registration.rs
@@ -0,0 +1,52 @@
+use core::mem;
+use core::task::Waker;
+
+/// Utility struct to register and wake a waker.
+#[derive(Debug, Default)]
+pub struct WakerRegistration {
+    waker: Option<Waker>,
+}
+
+impl WakerRegistration {
+    /// Create a new `WakerRegistration`.
+    pub const fn new() -> Self {
+        Self { waker: None }
+    }
+
+    /// Register a waker. Overwrites the previous waker, if any.
+    pub fn register(&mut self, w: &Waker) {
+        match self.waker {
+            // Optimization: If both the old and new Wakers wake the same task, we can simply
+            // keep the old waker, skipping the clone. (In most executor implementations,
+            // cloning a waker is somewhat expensive, comparable to cloning an Arc).
+            Some(ref w2) if (w2.will_wake(w)) => {}
+            _ => {
+                // clone the new waker and store it
+                if let Some(old_waker) = mem::replace(&mut self.waker, Some(w.clone())) {
+                    // We had a waker registered for another task. Wake it, so the other task can
+                    // reregister itself if it's still interested.
+                    //
+                    // If two tasks are waiting on the same thing concurrently, this will cause them
+                    // to wake each other in a loop fighting over this WakerRegistration. This wastes
+                    // CPU but things will still work.
+                    //
+                    // If the user wants to have two tasks waiting on the same thing they should use
+                    // a more appropriate primitive that can store multiple wakers.
+                    old_waker.wake()
+                }
+            }
+        }
+    }
+
+    /// Wake the registered waker, if any.
+    pub fn wake(&mut self) {
+        if let Some(w) = self.waker.take() {
+            w.wake()
+        }
+    }
+
+    /// Returns true if a waker is currently registered
+    pub fn occupied(&self) -> bool {
+        self.waker.is_some()
+    }
+}
-- 
cgit