rtl: add racy_read

src/rtl/mem.zig

@@ -30,3 +30,8 @@ pub fn read_struct(
 
     return result;
 }
+
+/// Racily read a pointer on purpose.
+pub fn racy_read(comptime T: type, ptr: *T) T {
+    return @atomicLoad(T, ptr, .unordered);
+}

src/rtl/seqlock.zig

@@ -24,8 +24,8 @@ pub fn SeqLockType(comptime T: type) type {
                     continue;
                 }
 
-                // Relaxed load memcpy is needed here since data races are UB.
+                // Unordered load memcpy is needed here since data races are UB.
-                rtl.barrier.atomic_load_memcpy(&ret, &self.raw, .monotonic);
+                rtl.barrier.atomic_load_memcpy(&ret, &self.raw, .unordered);
 
                 // Ensure the sequence read happens *after* the data is fully
                 // loaded.
@@ -49,8 +49,8 @@ pub fn SeqLockType(comptime T: type) type {
             // is incremented.
             rtl.barrier.fence(.release);
 
-            // Relaxed store memcpy is needed here since data races are UB.
+            // Unordered store memcpy is needed here since data races are UB.
-            rtl.barrier.atomic_store_memcpy(&self.raw, &data, .monotonic);
+            rtl.barrier.atomic_store_memcpy(&self.raw, &data, .unordered);
 
             self.sequence.store(seq + 2, .release);
         }

src/sys/ex/work.zig

@@ -144,13 +144,12 @@ const Pool = struct {
 
     /// Return whether or not a new worker thread should be added to this pool.
     fn should_grow(self: *Pool) bool {
-        // NOTE: lockless racy check whether or not the queue is empty
+        // Lockless racy check whether or not the queue is empty
         // and its active count.
         // This is fine as this is best effort anyway and is a heuristic.
-        const is_empty = @atomicLoad(
+        const is_empty = rtl.mem.racy_read(
             *rtl.List.Entry,
             &self.queue.items.head.next,
-            .monotonic,
         ) == &self.queue.items.head;
 
         const active = self.queue.active.load(.monotonic);

src/sys/kern/log_ring.zig

@@ -293,11 +293,7 @@ pub fn RingBufferType(data_size_bits: usize, avg_msg_bits: usize) type {
                 // Loading the ID here is purposely racey because
                 // we check if the descriptor is valid anyway.
                 const id: Id = @truncate(
-                    @atomicLoad(
+                    rtl.mem.racy_read(usize, @ptrCast(&blk.id)),
-                        usize,
-                        @as(*usize, @ptrCast(&blk.id)),
-                        .monotonic,
-                    ),
                 );
                 const res = self.read_desc(id);
                 desc = res.desc;

src/sys/kern/mutex.zig

@@ -175,10 +175,9 @@ pub const Mutex = struct {
 
                 var gas: usize = 0;
 
-                while (@atomicLoad(
+                while (rtl.mem.racy_read(
                     ?*ke.Thread,
                     cpu_curtd,
-                    .unordered,
                 ) == owner and gas < optimistic_spins) : (gas += 1) {
                     if (m.state.load(.monotonic) == State.uncontended) {
                         // Got released, retry acquiring it.

src/sys/kern/sched.zig

@@ -1779,11 +1779,7 @@ pub fn idle(_: ?*anyopaque) noreturn {
 
     while (true) {
         // Racy but only a hint anyway
-        const next = @atomicLoad(
+        const next = rtl.mem.racy_read(?*ke.Thread, &cpu.next_thread);
-            ?*ke.Thread,
-            &cpu.next_thread,
-            .monotonic,
-        );
 
         if (cpu.load.load(.monotonic) != 0 or next != null) {
             std.atomic.spinLoopHint();

src/sys/mm/zone.zig

@@ -736,7 +736,7 @@ pub const Zone = struct {
                 // Local depot is empty, rebalance with the zone depot.
                 // Purposefully race against cpu_depot_size because locking is
                 // tricky here, we may get a stale value but it is harmless.
-                const n = @atomicLoad(usize, &self.cpu_depot_size, .monotonic);
+                const n = rtl.mem.racy_read(usize, &self.cpu_depot_size);
 
                 self.alloc_depot_rebalance(n, cpu);
 
@@ -924,7 +924,7 @@ pub const Zone = struct {
 
                 // Local depot is empty, rebalance with the zone depot.
                 // Race explained in alloc.
-                const n = @atomicLoad(usize, &self.cpu_depot_size, .monotonic);
+                const n = rtl.mem.racy_read(usize, &self.cpu_depot_size);
 
                 self.free_depot_rebalance(n, cpu);