Emit PT_GNU_PROPERTY program header if .note.gnu.property is present

It looks like Linux kernel reads this program header to enable BTI on ARM64.

CMakeLists.txt

@@ -362,7 +362,6 @@ target_sources(mold PRIVATE
   lib/compress.cc
   lib/crc32.cc
   lib/demangle.cc
-  lib/elf.cc
   lib/filepath.cc
   lib/glob.cc
   lib/hyperloglog.cc
@@ -384,6 +383,7 @@ target_sources(mold PRIVATE
   src/arch-sparc64.cc
   src/arch-x86-64.cc
   src/config.cc
+  src/elf.cc
   third-party/rust-demangle/rust-demangle.c
   )
 

lib/archive-file.h

@@ -26,7 +26,6 @@
 #pragma once
 
 #include "common.h"
-#include "filetype.h"
 
 namespace mold {
 
@@ -164,14 +163,11 @@ std::vector<MappedFile *> read_fat_archive_members(Context &ctx, MappedFile *mf)
 
 template <typename Context, typename MappedFile>
 std::vector<MappedFile *> read_archive_members(Context &ctx, MappedFile *mf) {
-  switch (get_file_type(ctx, mf)) {
-  case FileType::AR:
+  std::string_view str = mf->get_contents();
+  if (str.starts_with("!<arch>\n"))
     return read_fat_archive_members(ctx, mf);
-  case FileType::THIN_AR:
-    return read_thin_archive_members(ctx, mf);
-  default:
-    unreachable();
-  }
+  assert(str.starts_with("!<thin>\n"));
+  return read_thin_archive_members(ctx, mf);
 }
 
 } // namespace mold

lib/integers.h

@@ -8,12 +8,12 @@
 //    big-endian SPARC machine to create a little-endian RV64 binary.
 //
 // 2. Even though data members in all ELF data strucutres are naturally
-//    aligned, they are not guaranteed to be aligned on memory. Because
-//    archive file (.a file) aligns each member only to a 2 byte boundary,
-//    anything larger than 2 bytes may be unaligned in an mmap'ed memory.
-//    Unaligned access is an undefined behavior in C/C++, so we shouldn't
-//    cast an arbitrary pointer to a uint32_t, for example, to read a
-//    32 bit value.
+//    aligned, they are not guaranteed to be aligned on memory because of
+//    archive files. Archive files (.a files) align each member only to a
+//    2 byte boundary, so anything larger than 2 bytes may be misaligned
+//    in an mmap'ed memory. Misaligned access is an undefined behavior in
+//    C/C++, so we shouldn't cast an arbitrary pointer to a uint32_t, for
+//    example, to read a 32 bit value.
 //
 // The data types defined in this file don't depend on host byte order and
 // don't do unaligned access.

lib/tar.cc

@@ -1,3 +1,5 @@
+// This file contains functions to create a tar file.
+
 #include "common.h"
 
 #ifdef _WIN32

src/arch-arm32.cc

@@ -722,7 +722,11 @@ u64 get_eflags(Context<E> &ctx) {
 void fixup_arm_exidx_section(Context<E> &ctx) {
   Timer t(ctx, "fixup_arm_exidx_section");
 
-  OutputSection<E> *osec = find_section(ctx, SHT_ARM_EXIDX);
+  Chunk<E> *chunk = find_chunk(ctx, SHT_ARM_EXIDX);
+  if (!chunk)
+    return;
+
+  OutputSection<E> *osec = chunk->to_osec();
   if (!osec)
     return;
 
@@ -780,7 +784,7 @@ void fixup_arm_exidx_section(Context<E> &ctx) {
   // .ARM.exidx's sh_link should be set to the .text section index.
   // Runtime doesn't care about it, but the binutils's strip command does.
   if (ctx.shdr) {
-    if (Chunk<E> *text = find_section(ctx, ".text")) {
+    if (Chunk<E> *text = find_chunk(ctx, ".text")) {
       osec->shdr.sh_link = text->shndx;
       ctx.shdr->copy_buf(ctx);
     }

src/elf.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "integers.h"
+#include "../lib/integers.h"
 
 #include <concepts>
 #include <ostream>
@@ -191,6 +191,7 @@ enum : u32 {
   PT_GNU_EH_FRAME = 0x6474e550,
   PT_GNU_STACK = 0x6474e551,
   PT_GNU_RELRO = 0x6474e552,
+  PT_GNU_PROPERTY = 0x6474e553,
   PT_OPENBSD_RANDOMIZE = 0x65a3dbe6,
   PT_ARM_EXIDX = 0x70000001,
   PT_RISCV_ATTRIBUTES = 0x70000003,

src/filetype.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "common.h"
+#include "../lib/common.h"
 #include "elf.h"
 
 namespace mold {
@@ -140,23 +140,23 @@ FileType get_file_type(Context &ctx, MappedFile *mf) {
   return FileType::UNKNOWN;
 }
 
-inline std::string filetype_to_string(FileType type) {
-  switch (type) {
-  case FileType::UNKNOWN: return "UNKNOWN";
-  case FileType::EMPTY: return "EMPTY";
-  case FileType::ELF_OBJ: return "ELF_OBJ";
-  case FileType::ELF_DSO: return "ELF_DSO";
-  case FileType::AR: return "AR";
-  case FileType::THIN_AR: return "THIN_AR";
-  case FileType::TEXT: return "TEXT";
-  case FileType::GCC_LTO_OBJ: return "GCC_LTO_OBJ";
-  case FileType::LLVM_BITCODE: return "LLVM_BITCODE";
-  }
-  return "UNKNOWN";
-}
-
 inline std::ostream &operator<<(std::ostream &out, FileType type) {
-  out << filetype_to_string(type);
+  auto to_string = [&] {
+    switch (type) {
+    case FileType::UNKNOWN: return "UNKNOWN";
+    case FileType::EMPTY: return "EMPTY";
+    case FileType::ELF_OBJ: return "ELF_OBJ";
+    case FileType::ELF_DSO: return "ELF_DSO";
+    case FileType::AR: return "AR";
+    case FileType::THIN_AR: return "THIN_AR";
+    case FileType::TEXT: return "TEXT";
+    case FileType::GCC_LTO_OBJ: return "GCC_LTO_OBJ";
+    case FileType::LLVM_BITCODE: return "LLVM_BITCODE";
+    default: return "UNKNOWN";
+    }
+  };
+
+  out << to_string();
   return out;
 }
 

src/main.cc

@@ -1,4 +1,5 @@
 #include "mold.h"
+#include "filetype.h"
 #include "../lib/archive-file.h"
 #include "../lib/output-file.h"
 

src/mold.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include "../lib/common.h"
-#include "../lib/elf.h"
+#include "elf.h"
 
 #include <atomic>
 #include <bitset>
@@ -341,10 +341,10 @@ template <typename E> u64 get_dtp_addr(const ElfPhdr<E> &);
 //
 
 template <typename E>
-OutputSection<E> *find_section(Context<E> &ctx, u32 sh_type);
+Chunk<E> *find_chunk(Context<E> &ctx, u32 sh_type);
 
 template <typename E>
-OutputSection<E> *find_section(Context<E> &ctx, std::string_view name);
+Chunk<E> *find_chunk(Context<E> &ctx, std::string_view name);
 
 template <typename E>
 u64 get_eflags(Context<E> &ctx) {

src/output-chunks.cc

@@ -25,20 +25,18 @@ static u32 elf_hash(std::string_view name) {
 }
 
 template <typename E>
-OutputSection<E> *find_section(Context<E> &ctx, u32 sh_type) {
+Chunk<E> *find_chunk(Context<E> &ctx, u32 sh_type) {
   for (Chunk<E> *chunk : ctx.chunks)
-    if (OutputSection<E> *osec = chunk->to_osec())
-      if (osec->shdr.sh_type == sh_type)
-        return osec;
+    if (chunk->shdr.sh_type == sh_type)
+      return chunk;
   return nullptr;
 }
 
 template <typename E>
-OutputSection<E> *find_section(Context<E> &ctx, std::string_view name) {
+Chunk<E> *find_chunk(Context<E> &ctx, std::string_view name) {
   for (Chunk<E> *chunk : ctx.chunks)
-    if (OutputSection<E> *osec = chunk->to_osec())
-      if (osec->name == name)
-        return osec;
+    if (chunk->name == name)
+      return chunk;
   return nullptr;
 }
 
@@ -267,6 +265,10 @@ static std::vector<ElfPhdr<E>> create_phdr(Context<E> &ctx) {
   if (ctx.eh_frame_hdr)
     define(PT_GNU_EH_FRAME, PF_R, ctx.eh_frame_hdr);
 
+  // Add PT_GNU_PROPERTY
+  if (Chunk<E> *chunk = find_chunk(ctx, ".note.gnu.property"))
+    define(PT_GNU_PROPERTY, PF_R, chunk);
+
   // Add PT_GNU_STACK, which is a marker segment that doesn't really
   // contain any segments. It controls executable bit of stack area.
   {
@@ -293,8 +295,8 @@ static std::vector<ElfPhdr<E>> create_phdr(Context<E> &ctx) {
 
   // Create a PT_ARM_EDXIDX
   if constexpr (is_arm32<E>)
-    if (OutputSection<E> *osec = find_section(ctx, SHT_ARM_EXIDX))
-      define(PT_ARM_EXIDX, PF_R, osec);
+    if (Chunk<E> *chunk = find_chunk(ctx, SHT_ARM_EXIDX))
+      define(PT_ARM_EXIDX, PF_R, chunk);
 
   // Create a PT_RISCV_ATTRIBUTES
   if constexpr (is_riscv<E>)
@@ -734,19 +736,19 @@ static std::vector<Word<E>> create_dynamic_section(Context<E> &ctx) {
     define(DT_STRSZ, ctx.dynstr->shdr.sh_size);
   }
 
-  if (find_section(ctx, SHT_INIT_ARRAY)) {
+  if (find_chunk(ctx, SHT_INIT_ARRAY)) {
     define(DT_INIT_ARRAY, ctx.__init_array_start->value);
     define(DT_INIT_ARRAYSZ,
            ctx.__init_array_end->value - ctx.__init_array_start->value);
   }
 
-  if (find_section(ctx, SHT_PREINIT_ARRAY)) {
+  if (find_chunk(ctx, SHT_PREINIT_ARRAY)) {
     define(DT_PREINIT_ARRAY, ctx.__preinit_array_start->value);
     define(DT_PREINIT_ARRAYSZ,
            ctx.__preinit_array_end->value - ctx.__preinit_array_start->value);
   }
 
-  if (find_section(ctx, SHT_FINI_ARRAY)) {
+  if (find_chunk(ctx, SHT_FINI_ARRAY)) {
     define(DT_FINI_ARRAY, ctx.__fini_array_start->value);
     define(DT_FINI_ARRAYSZ,
            ctx.__fini_array_end->value - ctx.__fini_array_start->value);
@@ -2949,8 +2951,8 @@ template class RelocSection<E>;
 template class ComdatGroupSection<E>;
 template class GnuDebuglinkSection<E>;
 
-template OutputSection<E> *find_section(Context<E> &, u32);
-template OutputSection<E> *find_section(Context<E> &, std::string_view);
+template Chunk<E> *find_chunk(Context<E> &, u32);
+template Chunk<E> *find_chunk(Context<E> &, std::string_view);
 template i64 to_phdr_flags(Context<E> &ctx, Chunk<E> *chunk);
 template ElfSym<E> to_output_esym(Context<E> &, Symbol<E> &, u32, U32<E> *);
 

src/passes.cc

@@ -1271,10 +1271,13 @@ void fixup_ctors_in_init_array(Context<E> &ctx) {
     }
   };
 
-  if (OutputSection<E> *osec = find_section(ctx, ".init_array"))
-    fixup(*osec);
-  if (OutputSection<E> *osec = find_section(ctx, ".fini_array"))
-    fixup(*osec);
+  if (Chunk<E> *chunk = find_chunk(ctx, ".init_array"))
+    if (OutputSection<E> *osec = chunk->to_osec())
+      fixup(*osec);
+
+  if (Chunk<E> *chunk = find_chunk(ctx, ".fini_array"))
+    if (OutputSection<E> *osec = chunk->to_osec())
+      fixup(*osec);
 }
 
 template <typename T>

test/gnu-property.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -o $t/a.o -c -xc -
+int main() {}
+EOF
+
+$CC -B. -o $t/exe $t/a.o -no-pie
+readelf -W --sections $t/exe | grep -Fqw .note.gnu.property || skip
+readelf -W --segments $t/exe | grep -qw GNU_PROPERTY