wake

    fn wake(ptr: *const u32, max_waiters: u32) void {
        if (max_waiters == 0) return;

        const bucket = bucketForAddress(@intFromPtr(ptr));

        // To ensure the store to `ptr` is ordered before this check, we effectively want a `.release`
        // load, but that doesn't exist in the C11 memory model, so emulate it with a non-mutating rmw.
        if (bucket.num_waiters.fetchAdd(0, .release) == 0) {
            @branchHint(.likely);
            return; // no waiters
        }

        // Waiters removed from the linked list under the mutex so we can unpark their threads outside
        // of the critical section. This forms a singly-linked list of waiters using `Waiter.node.next`.
        var waking_head: ?*std.DoublyLinkedList.Node = null;
        {
            bucket.mutex.lock();
            defer bucket.mutex.unlock();

            var num_removed: u32 = 0;
            var it = bucket.waiters.first;
            while (num_removed < max_waiters) {
                const waiter: *Waiter = @fieldParentPtr("node", it orelse break);
                it = waiter.node.next;
                if (waiter.address != @intFromPtr(ptr)) continue;
                const old_status = waiter.thread_status.fetchAnd(
                    .{ .cancelation = @fromBackingInt(@intCast(0b110)), .awaitable = .all_ones },
                    .monotonic,
                );
                switch (old_status.cancelation) {
                    .parked => {}, // state updated to `.none`
                    .none => continue, // race with timeout; they are about to lock `bucket.mutex` and remove themselves from the bucket
                    .canceling => continue, // race with a canceler who hasn't called `removeCanceledWaiter` yet
                    .canceled => unreachable,
                    .blocked => unreachable,
                    .blocked_alertable => unreachable,
                    .blocked_alertable_canceling => unreachable,
                    .blocked_canceling => unreachable,
                }
                // We're waking this waiter. Remove them from the bucket and add them to our local list.
                bucket.waiters.remove(&waiter.node);
                waiter.node.next = waking_head;
                waking_head = &waiter.node;
                num_removed += 1;
            }
            _ = bucket.num_waiters.fetchSub(num_removed, .monotonic);
        }

        var unpark_buf: [128]UnparkTid = undefined;
        var unpark_len: usize = 0;

        // Finally, unpark the threads.
        while (waking_head) |node| {
            waking_head = node.next;
            const waiter: *Waiter = @fieldParentPtr("node", node);
            unpark_buf[unpark_len] = waiter.tid;
            if (need_unpark_flag) setUnparkFlag(waiter.unpark_flag);
            unpark_len += 1;
            if (unpark_len == unpark_buf.len) {
                unpark(&unpark_buf, ptr);
                unpark_len = 0;
            }
        }
        if (unpark_len > 0) {
            unpark(unpark_buf[0..unpark_len], ptr);
        }
    }