all: make memory model saner

src/kern/log_ring.zig

@@ -220,7 +220,7 @@ pub fn RingBuffer(data_size_bits: usize, avg_msg_bits: usize) type {
             result.seq = info.sequence;
 
             // Ensure the copy is finished before the state load.
-            rtl.barrier.rmb();
+            rtl.barrier.fence(.acquire);
 
             state_var = desc.state.load(.acquire);
             result.state = get_desc_state(id, state_var);

src/kern/smr.zig

@@ -79,7 +79,7 @@ pub var full_scans: std.atomic.Value(usize) = .init(0);
 /// Scan all CPUs and return the minimum observed value.
 /// If `should_wait` is true, this will spinloop (or block) until all CPUs have reached the given goal.
 fn scan(dom: *Domain, goal: Sequence, clock: Clock, should_wait: bool) Sequence {
-    rtl.barrier.mb();
+    rtl.barrier.fence(.seq_cst);
 
     const clk_write = clock.write_seq.raw;
     const clk_read = clock.read_seq.raw;
@@ -125,7 +125,7 @@ fn scan(dom: *Domain, goal: Sequence, clock: Clock, should_wait: bool) Sequence
     }
 
     if (dom.preempt) {
-        rtl.barrier.mb();
+        rtl.barrier.fence(.seq_cst);
 
         // Preemption is enabled, check stalled readers.
         for (0..ke.ncpus) |i| {
@@ -181,7 +181,7 @@ fn scan(dom: *Domain, goal: Sequence, clock: Clock, should_wait: bool) Sequence
     // (1) is done by the acquire fence below, and (2) is done by the release
     // success ordering in the CAS. This pairs with the load acquire in poll().
     // It would also be valid to relax the CAS and have a full fence here.
-    rtl.barrier.acq();
+    rtl.barrier.fence(.acquire);
 
     // Advance the global read sequence.
     var dom_rd_seq = dom.clock.read_seq.load(.acquire);
@@ -259,7 +259,7 @@ pub fn advance(dom: *Domain) Sequence {
 /// the calling thread are visible. The global clock isn't advanced unlike `advance`,
 /// it is only when commit() is called.
 pub fn deferred_advance(dom: *Domain) Sequence {
-    rtl.barrier.mb();
+    rtl.barrier.fence(.seq_cst);
 
     return dom.clock.write_seq.load(.monotonic) + seq_incr;
 }
@@ -296,7 +296,7 @@ fn enter_internal(dom: *Domain) Sequence {
         _ = cpu.current_seq.fetchAdd(wr_seq, .seq_cst);
     } else {
         cpu.current_seq.store(wr_seq, .monotonic);
-        rtl.barrier.mb();
+        rtl.barrier.fence(.seq_cst);
     }
 
     return wr_seq;

src/kern/spinlock.zig

@@ -264,7 +264,8 @@ pub const QSpinLock = LockTemplate(struct {
         }
 
         // Store cpu + 1 so that we can distinguish from an empty tail field.
-        rtl.barrier.wmb();
+        rtl.barrier.fence(.release);
+
         const tail: u16 = @bitCast(Tail{ .cpu = @intCast(ke.cpu.current() + 1), .idx = @intCast(idx) });
         const old: Tail = @bitCast(self.data.split.tail.swap(tail, .acq_rel));
 

src/platform/pc/smp.zig

@@ -113,7 +113,7 @@ pub fn init() linksection(r.init) void {
 
         start_thread.remote(@intCast(cpu_id)).* = make_thread(@ptrCast(&ap_entry), stack_top - r.kib(32));
 
-        rtl.barrier.wmb();
+        rtl.barrier.fence(.release);
 
         // Send the INIT-SIPI-SIPI sequence to start the AP.
         apic.send_init(apic_id);

src/rtl/barrier.zig

@@ -4,51 +4,24 @@
 const std = @import("std");
 const builtin = @import("builtin");
 
-/// Read Memory Barrier (Load-Load)
-/// Ensures all loads before the barrier complete before any loads after it.
-pub inline fn rmb() void {
-    switch (builtin.cpu.arch) {
-        .x86, .x86_64 => {
-            asm volatile ("" ::: .{ .memory = true });
-        },
-        .aarch64 => {
-            asm volatile ("dmb ishld" ::: .{ .memory = true });
-        },
-        .arm => {
-            asm volatile ("dmb ish" ::: .{ .memory = true });
-        },
-        .riscv64, .riscv32 => {
-            asm volatile ("fence r,r" ::: .{ .memory = true });
-        },
-        else => @compileError("unsupported architecture"),
+pub inline fn fence(comptime ordering: std.builtin.AtomicOrder) void {
+    switch (ordering) {
+        .seq_cst, .acq_rel => mb(),
+        .acquire => acq(),
+        .release => rel(),
+        else => @compileError("Unsupported ordering"),
     }
 }
 
-/// Write Memory Barrier (Store-Store)
-/// Ensures all stores before the barrier complete before any stores after it.
-pub inline fn wmb() void {
+// Full memory barrier
+inline fn mb() void {
     switch (builtin.cpu.arch) {
-        .x86, .x86_64 => {
-            asm volatile ("" ::: .{ .memory = true });
+        .x86_64 => {
+            // Faster than mfence
+            asm volatile ("lock addl $0, (%%rsp)" ::: .{ .memory = true });
         },
-        .aarch64 => {
-            asm volatile ("dmb ishst" ::: .{ .memory = true });
-        },
-        .arm => {
-            asm volatile ("dmb ish" ::: .{ .memory = true });
-        },
-        .riscv64, .riscv32 => {
-            asm volatile ("fence w,w" ::: .{ .memory = true });
-        },
-        else => @compileError("unsupported architecture"),
-    }
-}
-
-/// Full Memory Barrier (Load/Store - Load/Store)
-pub inline fn mb() void {
-    switch (builtin.cpu.arch) {
-        .x86, .x86_64 => {
-            asm volatile ("mfence" ::: .{ .memory = true });
+        .x86 => {
+            asm volatile ("lock addl $0, (%%esp)" ::: .{ .memory = true });
         },
         .aarch64, .arm => {
             asm volatile ("dmb ish" ::: .{ .memory = true });
@@ -60,8 +33,8 @@ pub inline fn mb() void {
     }
 }
 
-/// Acquire Barrier (Load/Load - Load/Store)
-pub inline fn acq() void {
+// Acquire barrier
+inline fn acq() void {
     switch (builtin.cpu.arch) {
         .x86, .x86_64 => {
             asm volatile ("" ::: .{ .memory = true });
@@ -76,6 +49,22 @@ pub inline fn acq() void {
     }
 }
 
+// Release barrier
+inline fn rel() void {
+    switch (builtin.cpu.arch) {
+        .x86, .x86_64 => {
+            asm volatile ("" ::: .{ .memory = true });
+        },
+        .aarch64, .arm => {
+            asm volatile ("dmb ish" ::: .{ .memory = true });
+        },
+        .riscv64, .riscv32 => {
+            asm volatile ("fence rw,w" ::: .{ .memory = true });
+        },
+        else => @compileError("unsupported architecture"),
+    }
+}
+
 /// Copy to dest from src with atomic loads.
 /// Both addresses need to be aligned.
 pub fn atomic_load_memcpy(dest: anytype, src: anytype, comptime ordering: std.builtin.AtomicOrder) void {

src/rtl/seqlock.zig

@@ -26,7 +26,7 @@ pub fn SeqLock(comptime T: type) type {
                 rtl.barrier.atomic_load_memcpy(&ret, &self.raw, .monotonic);
 
                 // Ensure the sequence read happens *after* the data is fully loaded.
-                rtl.barrier.rmb();
+                rtl.barrier.fence(.acquire);
 
                 if (self.sequence.load(.monotonic) == seq)
                     break;
@@ -43,7 +43,7 @@ pub fn SeqLock(comptime T: type) type {
             self.sequence.store(seq + 1, .monotonic);
 
             // Ensure the data write happens *after* the sequence is incremented.
-            rtl.barrier.wmb();
+            rtl.barrier.fence(.release);
 
             // Relaxed store memcpy is needed here since data races are UB.
             rtl.barrier.atomic_store_memcpy(&self.raw, &data, .monotonic);