From 610042a8e927106cb2d852fba2a90faa0d0e050f Mon Sep 17 00:00:00 2001
From: Rui Ueyama <ruiu@bluewhale.systems>
Date: Fri, 21 Oct 2022 07:08:09 +0800
Subject: [PATCH] [ELF] Set .ARM.exidx's sh_link to .text

GNU binutil's strip command doesn't like an .ARM.exidx section with
sh_link set to 0. arm-linux-gnueabihf-strip actually crashes on my
machine when the command is run on such a file.

https://github.com/rui314/mold/issues/801
---
 elf/arch-arm32.cc     | 56 +++++++++++++++++++++++++++++--------------
 elf/mold.h            |  2 +-
 elf/output-chunks.cc  |  4 ++--
 elf/passes.cc         |  2 +-
 test/elf/exception.sh |  8 +++++++
 5 files changed, 50 insertions(+), 22 deletions(-)

diff --git a/elf/arch-arm32.cc b/elf/arch-arm32.cc
index 46e45d43d..bea510763 100644
--- a/elf/arch-arm32.cc
+++ b/elf/arch-arm32.cc
@@ -552,14 +552,6 @@ void RangeExtensionThunk<E>::copy_buf(Context<E> &ctx) {
   }
 }
 
-template <typename E>
-static OutputSection<E> *find_exidx_section(Context<E> &ctx) {
-  for (std::unique_ptr<OutputSection<E>> &osec : ctx.output_sections)
-    if (osec->shdr.sh_type == SHT_ARM_EXIDX)
-      return osec.get();
-  return nullptr;
-}
-
 // ARM executables use an .ARM.exidx section to look up an exception
 // handling record for the current instruction pointer. The table needs
 // to be sorted by their addresses.
@@ -569,13 +561,7 @@ static OutputSection<E> *find_exidx_section(Context<E> &ctx) {
 // likely that it's due to some historical reason.
 //
 // This function sorts .ARM.exidx records.
-void sort_arm_exidx(Context<E> &ctx) {
-  Timer t(ctx, "sort_arm_exidx");
-
-  OutputSection<E> *osec = find_exidx_section(ctx);
-  if (!osec)
-    return;
-
+static void sort_exidx(Context<E> &ctx, OutputSection<E> &osec) {
   // .ARM.exidx records consists of a signed 31-bit relative address
   // and a 32-bit value. The relative address indicates the start
   // address of a function that the record covers. The value is one of
@@ -595,11 +581,11 @@ void sort_arm_exidx(Context<E> &ctx) {
     ul32 val;
   };
 
-  if (osec->shdr.sh_size % sizeof(Entry))
+  if (osec.shdr.sh_size % sizeof(Entry))
     Fatal(ctx) << "invalid .ARM.exidx section size";
 
-  Entry *ent = (Entry *)(ctx.buf + osec->shdr.sh_offset);
-  i64 num_entries = osec->shdr.sh_size / sizeof(Entry);
+  Entry *ent = (Entry *)(ctx.buf + osec.shdr.sh_offset);
+  i64 num_entries = osec.shdr.sh_size / sizeof(Entry);
 
   // Entry's addresses are relative to themselves. In order to sort
   // records by addresses, we first translate them so that the addresses
@@ -628,4 +614,38 @@ void sort_arm_exidx(Context<E> &ctx) {
   });
 }
 
+void fixup_arm_exidx_section(Context<E> &ctx) {
+  Timer t(ctx, "fixup_arm_exidx_section");
+
+  auto find_exidx = [&]() -> OutputSection<E> * {
+    for (std::unique_ptr<OutputSection<E>> &osec : ctx.output_sections)
+      if (osec->shdr.sh_type == SHT_ARM_EXIDX)
+        return osec.get();
+    return nullptr;
+  };
+
+  OutputSection<E> *exidx = find_exidx();
+  if (!exidx)
+    return;
+
+  // Sort .ARM.exidx contents
+  sort_exidx(ctx, *exidx);
+
+  // .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) {
+    auto find_text = [&]() -> OutputSection<E> * {
+      for (std::unique_ptr<OutputSection<E>> &osec : ctx.output_sections)
+        if (osec->name == ".text")
+          return osec.get();
+      return nullptr;
+    };
+
+    if (OutputSection<E> *text = find_exidx()) {
+      exidx->shdr.sh_link = text->shndx;
+      ctx.shdr->copy_buf(ctx);
+    }
+  }
+}
+
 } // namespace mold::elf
diff --git a/elf/mold.h b/elf/mold.h
index 11938270e..c8bac2b4e 100644
--- a/elf/mold.h
+++ b/elf/mold.h
@@ -1434,7 +1434,7 @@ template <typename E> void write_dependency_file(Context<E> &);
 // arch-arm32.cc
 //
 
-void sort_arm_exidx(Context<ARM32> &ctx);
+void fixup_arm_exidx_section(Context<ARM32> &ctx);
 
 //
 // arch-riscv64.cc
diff --git a/elf/output-chunks.cc b/elf/output-chunks.cc
index 54562feb2..0d1d0dcfd 100644
--- a/elf/output-chunks.cc
+++ b/elf/output-chunks.cc
@@ -847,9 +847,9 @@ get_output_name(Context<E> &ctx, std::string_view name, u64 flags) {
 
   if ((name == ".rodata" || name.starts_with(".rodata.")) && (flags & SHF_MERGE))
     return (flags & SHF_STRINGS) ? ".rodata.str" : ".rodata.cst";
-  if (name == ".ARM.exidx" || name.starts_with(".ARM.exidx."))
+  if (name.starts_with(".ARM.exidx"))
     return ".ARM.exidx";
-  if (name == ".ARM.extab" || name.starts_with(".ARM.extab."))
+  if (name.starts_with(".ARM.extab"))
     return ".ARM.extab";
 
   if (ctx.arg.z_keep_text_section_prefix) {
diff --git a/elf/passes.cc b/elf/passes.cc
index 713a06a84..4bffc3dbd 100644
--- a/elf/passes.cc
+++ b/elf/passes.cc
@@ -1102,7 +1102,7 @@ void copy_chunks(Context<E> &ctx) {
   report_undef_errors(ctx);
 
   if constexpr (std::is_same_v<E, ARM32>)
-    sort_arm_exidx(ctx);
+    fixup_arm_exidx_section(ctx);
 }
 
 template <typename E>
diff --git a/test/elf/exception.sh b/test/elf/exception.sh
index 6832966f0..c584c99fa 100755
--- a/test/elf/exception.sh
+++ b/test/elf/exception.sh
@@ -65,3 +65,11 @@ if [ $MACHINE = x86_64 -o $MACHINE = aarch64 ]; then
   $CXX -B. -o $t/exe10 $t/e.o -no-pie
   $QEMU $t/exe10
 fi
+
+$CXX -B. -o $t/exe11 $t/b.o -pie
+$STRIP $t/exe11
+$QEMU $t/exe11
+
+$CXX -B. -o $t/exe12 $t/c.o -no-pie
+$STRIP $t/exe12
+$QEMU $t/exe12