From 48242ba8d64992b7bcd4ea6c18611dffb5ccac6e Mon Sep 17 00:00:00 2001
From: Ayke van Laethem <aykevanlaethem@gmail.com>
Date: Mon, 23 May 2022 20:19:59 +0200
Subject: [PATCH] darwin: scan globals by reading MachO header

This replaces "precise" global scanning in LLVM with conservative
scanning of writable MachO segments. Eventually I'd like to get rid of
the AddGlobalsBitmap pass, and this is one step towards that goal.
---
 src/runtime/gc_globals_precise.go |  4 +-
 src/runtime/os_darwin.go          | 92 +++++++++++++++++++++++++++++++
 2 files changed, 94 insertions(+), 2 deletions(-)

diff --git a/src/runtime/gc_globals_precise.go b/src/runtime/gc_globals_precise.go
index 3cdba741..5fbaea37 100644
--- a/src/runtime/gc_globals_precise.go
+++ b/src/runtime/gc_globals_precise.go
@@ -1,5 +1,5 @@
-//go:build gc.conservative && !baremetal && !tinygo.wasm && !windows
-// +build gc.conservative,!baremetal,!tinygo.wasm,!windows
+//go:build gc.conservative && !baremetal && !darwin && !tinygo.wasm && !windows
+// +build gc.conservative,!baremetal,!darwin,!tinygo.wasm,!windows
 
 package runtime
 
diff --git a/src/runtime/os_darwin.go b/src/runtime/os_darwin.go
index d485f042..1d895722 100644
--- a/src/runtime/os_darwin.go
+++ b/src/runtime/os_darwin.go
@@ -3,6 +3,8 @@
 
 package runtime
 
+import "unsafe"
+
 const GOOS = "darwin"
 
 const (
@@ -18,3 +20,93 @@ const (
 	clock_REALTIME      = 0
 	clock_MONOTONIC_RAW = 4
 )
+
+// https://opensource.apple.com/source/xnu/xnu-7195.141.2/EXTERNAL_HEADERS/mach-o/loader.h.auto.html
+type machHeader struct {
+	magic      uint32
+	cputype    uint32
+	cpusubtype uint32
+	filetype   uint32
+	ncmds      uint32
+	sizeofcmds uint32
+	flags      uint32
+	reserved   uint32
+}
+
+// Struct for the LC_SEGMENT_64 load command.
+type segmentLoadCommand struct {
+	cmd      uint32 // LC_SEGMENT_64
+	cmdsize  uint32
+	segname  [16]byte
+	vmaddr   uintptr
+	vmsize   uintptr
+	fileoff  uintptr
+	filesize uintptr
+	maxprot  uint32
+	initprot uint32
+	nsects   uint32
+	flags    uint32
+}
+
+// MachO header of the currently running process.
+//go:extern _mh_execute_header
+var libc_mh_execute_header machHeader
+
+// Mark global variables.
+// The MachO linker doesn't seem to provide symbols for the start and end of the
+// data section. There is get_etext, get_edata, and get_end, but these are
+// undocumented and don't work with ASLR (which is enabled by default).
+// Therefore, read the MachO header directly.
+func markGlobals() {
+	// Here is a useful blog post to understand the MachO file format:
+	// https://h3adsh0tzz.com/2020/01/macho-file-format/
+
+	const (
+		MH_MAGIC_64   = 0xfeedfacf
+		LC_SEGMENT_64 = 0x19
+		VM_PROT_WRITE = 0x02
+	)
+
+	// Sanity check that we're actually looking at a MachO header.
+	if gcAsserts && libc_mh_execute_header.magic != MH_MAGIC_64 {
+		runtimePanic("gc: unexpected MachO header")
+	}
+
+	// Iterate through the load commands.
+	// Because we're only interested in LC_SEGMENT_64 load commands, cast the
+	// pointer to that struct in advance.
+	var offset uintptr
+	var hasOffset bool
+	cmd := (*segmentLoadCommand)(unsafe.Pointer(uintptr(unsafe.Pointer(&libc_mh_execute_header)) + unsafe.Sizeof(machHeader{})))
+	for i := libc_mh_execute_header.ncmds; i != 0; i-- {
+		if cmd.cmd == LC_SEGMENT_64 {
+			if cmd.fileoff == 0 && cmd.nsects != 0 {
+				// Detect ASLR offset by checking fileoff and nsects. This
+				// locates the __TEXT segment. This matches getsectiondata:
+				// https://opensource.apple.com/source/cctools/cctools-973.0.1/libmacho/getsecbyname.c.auto.html
+				offset = uintptr(unsafe.Pointer(&libc_mh_execute_header)) - cmd.vmaddr
+				hasOffset = true
+			}
+			if cmd.maxprot&VM_PROT_WRITE != 0 {
+				// Found a writable segment, which may contain Go globals.
+				if gcAsserts && !hasOffset {
+					// No ASLR offset detected. Did the __TEXT segment come
+					// after the __DATA segment?
+					// Note that when ASLR is disabled (for example, when
+					// running inside lldb), the offset is zero. That's why we
+					// need a separate hasOffset for this assert.
+					runtimePanic("gc: did not detect ASLR offset")
+				}
+				// Scan this segment for GC roots.
+				// This could be improved by only reading the memory areas
+				// covered by sections. That would reduce the amount of memory
+				// scanned a little bit (up to a single VM page).
+				markRoots(offset+cmd.vmaddr, offset+cmd.vmaddr+cmd.vmsize)
+			}
+		}
+
+		// Move on to the next load command (wich may or may not be a
+		// LC_SEGMENT_64).
+		cmd = (*segmentLoadCommand)(unsafe.Pointer(uintptr(unsafe.Pointer(cmd)) + uintptr(cmd.cmdsize)))
+	}
+}