std.process.Child: use clone3 on x86 and x86_64

Author: ruihe774

lib/std/os/linux.zig

@@ -87,6 +87,40 @@ pub fn clone(
     ) callconv(.C) usize, @ptrCast(&syscall_bits.clone))(func, stack, flags, arg, ptid, tp, ctid);
 }
 
+pub const clone_args = extern struct {
+    flags: u64,
+    pidfd: u64,
+    child_tid: u64,
+    parent_tid: u64,
+    exit_signal: u64,
+    stack: u64,
+    stack_size: u64,
+    tls: u64,
+    set_tid: u64,
+    set_tid_size: u64,
+    cgroup: u64,
+};
+
+pub fn clone3(
+    cl_args: *const clone_args,
+    size: usize,
+    func: *const fn (arg: usize) callconv(.C) u8,
+    arg: usize,
+) usize {
+    // TODO: write asm for other arch.
+    if (@hasDecl(syscall_bits, "clone3")) {
+        // Can't directly call a naked function; cast to C calling convention first.
+        return @as(*const fn (
+            cl_args: *const clone_args,
+            size: usize,
+            func: *const fn (arg: usize) callconv(.C) u8,
+            arg: usize,
+        ) callconv(.C) usize, @ptrCast(&syscall_bits.clone3))(cl_args, size, func, arg);
+    } else {
+        return @bitCast(-@as(isize, @intFromEnum(E.NOSYS)));
+    }
+}
+
 pub const ARCH = arch_bits.ARCH;
 pub const Elf_Symndx = arch_bits.Elf_Symndx;
 pub const F = arch_bits.F;

lib/std/os/linux/x86.zig

@@ -167,6 +167,37 @@ pub fn clone() callconv(.Naked) usize {
     );
 }
 
+pub fn clone3() callconv(.Naked) usize {
+    asm volatile (
+        \\  movl 4(%%esp),%%ecx
+        \\  movl 12(%esp),%%edx
+        \\  pushl %%ebx
+        \\  pushl %%esi
+        \\  movl 24(%%esp),%%esi
+        \\  movl %%ecx,%%ebx
+        \\  movl 16(%%esp),%%ecx
+        \\  movl $435,%%eax // SYS_clone3
+        \\  int $128
+        \\  testl %%eax,%%eax
+        \\  jz 1f
+        \\  popl %%esi
+        \\  popl %%ebx
+        \\  retl
+        \\
+        \\1:
+        \\  .cfi_undefined %%eip
+        \\  xorl %%ebp,%%ebp
+        \\
+        \\  andl $-16,%%esp
+        \\  subl $12,%%esp
+        \\  pushl %%esi
+        \\  calll *%%edx
+        \\  movl %%eax,%%ebx
+        \\  movl $1,%%eax // SYS_exit
+        \\  int $128
+    );
+}
+
 pub fn restore() callconv(.Naked) noreturn {
     switch (@import("builtin").zig_backend) {
         .stage2_c => asm volatile (

lib/std/os/linux/x86_64.zig

@@ -129,6 +129,27 @@ pub fn clone() callconv(.Naked) usize {
     );
 }
 
+pub fn clone3() callconv(.Naked) usize {
+    asm volatile (
+        \\      movl $435,%%eax // SYS_clone3
+        \\      movq %%rcx,%%r8
+        \\      syscall
+        \\      testq %%rax,%%rax
+        \\      jz 1f
+        \\      retq
+        \\
+        \\1:    .cfi_undefined %%rip
+        \\      xorl %%ebp,%%ebp
+        \\
+        \\      movq %%r8,%%rdi
+        \\      callq *%%rdx
+        \\      movl %%eax,%%edi
+        \\      movl $60,%%eax // SYS_exit
+        \\      syscall
+        \\
+    );
+}
+
 pub const restore = restore_rt;
 
 pub fn restore_rt() callconv(.Naked) noreturn {

lib/std/process/Child.zig

@@ -724,10 +724,6 @@ fn spawnPosix(self: *ChildProcess) SpawnError!void {
             immediateExit(spawnPosixChildHelper(@intFromPtr(&child_arg)));
         }
     } else {
-        var old_mask: posix.sigset_t = undefined;
-        posix.sigprocmask(posix.SIG.SETMASK, &linux.all_mask, &old_mask);
-        defer posix.sigprocmask(posix.SIG.SETMASK, &old_mask, null);
-        child_arg.sigmask = &old_mask;
         child_arg.ret_err = null;
         // Although the stack is fixed sized, we alloc it here,
         // because stack-smashing protection may have higher overhead than allocation.
@@ -737,13 +733,34 @@ fn spawnPosix(self: *ChildProcess) SpawnError!void {
         // For simplicity, we just align it to page boundary here.
         const stack = try self.allocator.alignedAlloc(u8, mem.page_size, stack_size);
         defer self.allocator.free(stack);
-        const rc = linux.clone(spawnPosixChildHelper, @intFromPtr(stack.ptr) + stack_size, linux.CLONE.VM | linux.CLONE.VFORK | linux.SIG.CHLD, @intFromPtr(&child_arg), null, 0, null);
-        pid_result = switch (posix.errno(rc)) {
-            .SUCCESS => @intCast(rc),
-            .AGAIN => return error.SystemResources,
-            .NOMEM => return error.SystemResources,
+
+        var clone_args = mem.zeroes(linux.clone_args);
+        clone_args.flags = linux.CLONE.VM | linux.CLONE.VFORK | linux.CLONE.CLEAR_SIGHAND;
+        clone_args.exit_signal = linux.SIG.CHLD;
+        clone_args.stack = @intFromPtr(stack.ptr);
+        clone_args.stack_size = stack_size;
+        var rc = linux.clone3(&clone_args, @sizeOf(linux.clone_args), spawnPosixChildHelper, @intFromPtr(&child_arg));
+        switch (posix.errno(rc)) {
+            .SUCCESS => {},
+            .AGAIN, .NOMEM => return error.SystemResources,
+            .INVAL, .NOSYS => {
+                // Fallback to use clone().
+                // We need to block signals here because we share VM with child before exec.
+                // Signal handlers may mess up our memory.
+                var old_mask: posix.sigset_t = undefined;
+                posix.sigprocmask(posix.SIG.SETMASK, &linux.all_mask, &old_mask);
+                defer posix.sigprocmask(posix.SIG.SETMASK, &old_mask, null);
+                child_arg.sigmask = &old_mask;
+                rc = linux.clone(spawnPosixChildHelper, @intFromPtr(stack.ptr) + stack_size, linux.CLONE.VM | linux.CLONE.VFORK | linux.SIG.CHLD, @intFromPtr(&child_arg), null, 0, null);
+                switch (posix.errno(rc)) {
+                    .SUCCESS => {},
+                    .AGAIN, .NOMEM => return error.SystemResources,
+                    else => |err| return posix.unexpectedErrno(err),
+                }
+            },
             else => |err| return posix.unexpectedErrno(err),
-        };
+        }
+        pid_result = @intCast(rc);
         if (child_arg.ret_err) |err| {
             return err;
         }