add std.os.cpuCount and have std.event.Loop use it for thread pool size

std/event.zig

@@ -150,8 +150,8 @@ pub const Loop = struct {
     /// TODO copy elision / named return values so that the threads referencing *Loop
     /// have the correct pointer value.
     fn initMultiThreaded(self: *Loop, allocator: *mem.Allocator) !void {
-        // TODO check the actual cpu core count
-        return self.initInternal(allocator, 4);
+        const core_count = try std.os.cpuCount(allocator);
+        return self.initInternal(allocator, core_count);
     }
 
     /// Thread count is the total thread count. The thread pool size will be

std/heap.zig

@@ -361,6 +361,73 @@ pub const ThreadSafeFixedBufferAllocator = struct {
     fn free(allocator: *Allocator, bytes: []u8) void {}
 };
 
+pub fn stackFallback(comptime size: usize, fallback_allocator: *Allocator) StackFallbackAllocator(size) {
+    return StackFallbackAllocator(size){
+        .buffer = undefined,
+        .fallback_allocator = fallback_allocator,
+        .fixed_buffer_allocator = undefined,
+        .allocator = Allocator{
+            .allocFn = StackFallbackAllocator(size).alloc,
+            .reallocFn = StackFallbackAllocator(size).realloc,
+            .freeFn = StackFallbackAllocator(size).free,
+        },
+    };
+}
+
+pub fn StackFallbackAllocator(comptime size: usize) type {
+    return struct {
+        const Self = this;
+
+        buffer: [size]u8,
+        allocator: Allocator,
+        fallback_allocator: *Allocator,
+        fixed_buffer_allocator: FixedBufferAllocator,
+
+        pub fn get(self: *Self) *Allocator {
+            self.fixed_buffer_allocator = FixedBufferAllocator.init(self.buffer[0..]);
+            return &self.allocator;
+        }
+
+        fn alloc(allocator: *Allocator, n: usize, alignment: u29) ![]u8 {
+            const self = @fieldParentPtr(Self, "allocator", allocator);
+            return FixedBufferAllocator.alloc(&self.fixed_buffer_allocator.allocator, n, alignment) catch
+                self.fallback_allocator.allocFn(self.fallback_allocator, n, alignment);
+        }
+
+        fn realloc(allocator: *Allocator, old_mem: []u8, new_size: usize, alignment: u29) ![]u8 {
+            const self = @fieldParentPtr(Self, "allocator", allocator);
+            const in_buffer = @ptrToInt(old_mem.ptr) >= @ptrToInt(&self.buffer) and
+                @ptrToInt(old_mem.ptr) < @ptrToInt(&self.buffer) + self.buffer.len;
+            if (in_buffer) {
+                return FixedBufferAllocator.realloc(
+                    &self.fixed_buffer_allocator.allocator,
+                    old_mem,
+                    new_size,
+                    alignment,
+                ) catch {
+                    const result = try self.fallback_allocator.allocFn(
+                        self.fallback_allocator,
+                        new_size,
+                        alignment,
+                    );
+                    mem.copy(u8, result, old_mem);
+                    return result;
+                };
+            }
+            return self.fallback_allocator.reallocFn(self.fallback_allocator, old_mem, new_size, alignment);
+        }
+
+        fn free(allocator: *Allocator, bytes: []u8) void {
+            const self = @fieldParentPtr(Self, "allocator", allocator);
+            const in_buffer = @ptrToInt(bytes.ptr) >= @ptrToInt(&self.buffer) and
+                @ptrToInt(bytes.ptr) < @ptrToInt(&self.buffer) + self.buffer.len;
+            if (!in_buffer) {
+                return self.fallback_allocator.freeFn(self.fallback_allocator, bytes);
+            }
+        }
+    };
+}
+
 test "c_allocator" {
     if (builtin.link_libc) {
         var slice = c_allocator.alloc(u8, 50) catch return;

std/os/index.zig

@@ -2748,3 +2748,42 @@ pub fn posixFStat(fd: i32) !posix.Stat {
 
     return stat;
 }
+
+pub const CpuCountError = error{
+    OutOfMemory,
+    PermissionDenied,
+    Unexpected,
+};
+
+pub fn cpuCount(fallback_allocator: *mem.Allocator) CpuCountError!usize {
+    const usize_count = 16;
+    const allocator = std.heap.stackFallback(usize_count * @sizeOf(usize), fallback_allocator).get();
+
+    var set = try allocator.alloc(usize, usize_count);
+    defer allocator.free(set);
+
+    while (true) {
+        const rc = posix.sched_getaffinity(0, set);
+        const err = posix.getErrno(rc);
+        switch (err) {
+            0 => {
+                if (rc < set.len * @sizeOf(usize)) {
+                    const result = set[0 .. rc / @sizeOf(usize)];
+                    var sum: usize = 0;
+                    for (result) |x| {
+                        sum += @popCount(x);
+                    }
+                    return sum;
+                } else {
+                    set = try allocator.realloc(usize, set, set.len * 2);
+                    continue;
+                }
+            },
+            posix.EFAULT => unreachable,
+            posix.EINVAL => unreachable,
+            posix.EPERM => return CpuCountError.PermissionDenied,
+            posix.ESRCH => unreachable,
+            else => return os.unexpectedErrorPosix(err),
+        }
+    }
+}

std/os/linux/index.zig

@@ -1197,6 +1197,10 @@ pub fn fremovexattr(fd: usize, name: [*]const u8) usize {
     return syscall2(SYS_fremovexattr, fd, @ptrToInt(name));
 }
 
+pub fn sched_getaffinity(pid: i32, set: []usize) usize {
+    return syscall3(SYS_sched_getaffinity, @bitCast(usize, isize(pid)), set.len * @sizeOf(usize), @ptrToInt(set.ptr));
+}
+
 pub const epoll_data = packed union {
     ptr: usize,
     fd: i32,

std/os/test.zig

@@ -58,3 +58,8 @@ fn start2(ctx: *i32) u8 {
     _ = @atomicRmw(i32, ctx, AtomicRmwOp.Add, 1, AtomicOrder.SeqCst);
     return 0;
 }
+
+test "cpu count" {
+    const cpu_count = try std.os.cpuCount(a);
+    assert(cpu_count >= 1);
+}