Commit 76afdd0586
src/link/MachO/zld.zig
@@ -4172,6 +4172,7 @@ pub fn linkWithZld(macho_file: *MachO, comp: *Compilation, prog_node: *std.Progr
if (codesig) |*csig| {
try zld.writeCodeSignature(comp, csig); // code signing always comes last
+ try MachO.invalidateKernelCache(directory.handle, zld.options.emit.?.sub_path);
}
}
src/link/MachO.zig
@@ -662,6 +662,8 @@ pub fn flushModule(self: *MachO, comp: *Compilation, prog_node: *std.Progress.No
if (codesig) |*csig| {
try self.writeCodeSignature(comp, csig); // code signing always comes last
+ const emit = self.base.options.emit.?;
+ try invalidateKernelCache(emit.directory.handle, emit.sub_path);
}
if (self.d_sym) |*d_sym| {
@@ -691,6 +693,21 @@ pub fn flushModule(self: *MachO, comp: *Compilation, prog_node: *std.Progress.No
self.cold_start = false;
}
+
+/// XNU starting with Big Sur running on arm64 is caching inodes of running binaries.
+/// Any change to the binary will effectively invalidate the kernel's cache
+/// resulting in a SIGKILL on each subsequent run. Since when doing incremental
+/// linking we're modifying a binary in-place, this will end up with the kernel
+/// killing it on every subsequent run. To circumvent it, we will copy the file
+/// into a new inode, remove the original file, and rename the copy to match
+/// the original file. This is super messy, but there doesn't seem any other
+/// way to please the XNU.
+pub fn invalidateKernelCache(dir: std.fs.Dir, sub_path: []const u8) !void {
+ if (comptime builtin.target.isDarwin() and builtin.target.cpu.arch == .aarch64) {
+ try dir.copyFile(sub_path, dir, sub_path, .{});
+ }
+}
+
inline fn conformUuid(out: *[Md5.digest_length]u8) void {
// LC_UUID uuids should conform to RFC 4122 UUID version 4 & UUID version 5 formats
out[6] = (out[6] & 0x0F) | (3 << 4);
src/link.zig
@@ -418,26 +418,7 @@ pub const File = struct {
.Exe => {},
}
switch (base.tag) {
- .macho => if (base.file) |f| {
- if (build_options.only_c) unreachable;
- if (comptime builtin.target.isDarwin() and builtin.target.cpu.arch == .aarch64) {
- if (base.options.target.cpu.arch == .aarch64) {
- // XNU starting with Big Sur running on arm64 is caching inodes of running binaries.
- // Any change to the binary will effectively invalidate the kernel's cache
- // resulting in a SIGKILL on each subsequent run. Since when doing incremental
- // linking we're modifying a binary in-place, this will end up with the kernel
- // killing it on every subsequent run. To circumvent it, we will copy the file
- // into a new inode, remove the original file, and rename the copy to match
- // the original file. This is super messy, but there doesn't seem any other
- // way to please the XNU.
- const emit = base.options.emit orelse return;
- try emit.directory.handle.copyFile(emit.sub_path, emit.directory.handle, emit.sub_path, .{});
- }
- }
- f.close();
- base.file = null;
- },
- .coff, .elf, .plan9, .wasm => if (base.file) |f| {
+ .coff, .elf, .macho, .plan9, .wasm => if (base.file) |f| {
if (build_options.only_c) unreachable;
if (base.intermediary_basename != null) {
// The file we have open is not the final file that we want to