Source code
Revision control
Copy as Markdown
Other Tools
/* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
// This program acts as a linker wrapper. Its executable name is meant
// to be that of a linker, and it will find the next linker with the same
// name in $PATH. However, if for some reason the next linker cannot be
// found this way, the caller may pass its path via the --real-linker
// option.
//
#include "relrhack.h"
#include <algorithm>
#include <array>
#include <cstdio>
#include <cstring>
#include <filesystem>
#include <fstream>
#include <iostream>
#include <optional>
#include <spawn.h>
#include <sstream>
#include <stdexcept>
#include <sys/wait.h>
#include <unistd.h>
#include <unordered_map>
#include <utility>
#include <vector>
#include "mozilla/ScopeExit.h"
namespace fs = std::filesystem;
class CantSwapSections : public std::runtime_error {
public:
CantSwapSections(const char* what) : std::runtime_error(what) {}
};
template <int bits>
struct Elf {};
#define ELF(bits) \
template <> \
struct Elf<bits> { \
using Ehdr = Elf##bits##_Ehdr; \
using Phdr = Elf##bits##_Phdr; \
using Shdr = Elf##bits##_Shdr; \
using Dyn = Elf##bits##_Dyn; \
using Addr = Elf##bits##_Addr; \
using Word = Elf##bits##_Word; \
using Off = Elf##bits##_Off; \
using Verneed = Elf##bits##_Verneed; \
using Vernaux = Elf##bits##_Vernaux; \
}
ELF(32);
ELF(64);
template <int bits>
struct RelR : public Elf<bits> {
using Elf_Ehdr = typename Elf<bits>::Ehdr;
using Elf_Phdr = typename Elf<bits>::Phdr;
using Elf_Shdr = typename Elf<bits>::Shdr;
using Elf_Dyn = typename Elf<bits>::Dyn;
using Elf_Addr = typename Elf<bits>::Addr;
using Elf_Word = typename Elf<bits>::Word;
using Elf_Off = typename Elf<bits>::Off;
using Elf_Verneed = typename Elf<bits>::Verneed;
using Elf_Vernaux = typename Elf<bits>::Vernaux;
#define TAG_NAME(t) \
{ t, #t }
class DynInfo {
public:
using Tag = decltype(Elf_Dyn::d_tag);
using Value = decltype(Elf_Dyn::d_un.d_val);
bool is_wanted(Tag tag) const { return tag_names.count(tag); }
void insert(off_t offset, Tag tag, Value val) {
data[tag] = std::make_pair(offset, val);
}
off_t offset(Tag tag) const { return data.at(tag).first; }
bool contains(Tag tag) const { return data.count(tag); }
Value& operator[](Tag tag) {
if (!is_wanted(tag)) {
std::stringstream msg;
msg << "Tag 0x" << std::hex << tag << " is not in DynInfo::tag_names";
throw std::runtime_error(msg.str());
}
return data[tag].second;
}
const char* name(Tag tag) const { return tag_names.at(tag); }
private:
std::unordered_map<Tag, std::pair<off_t, Value>> data;
const std::unordered_map<Tag, const char*> tag_names = {
TAG_NAME(DT_JMPREL), TAG_NAME(DT_PLTRELSZ), TAG_NAME(DT_RELR),
TAG_NAME(DT_RELRENT), TAG_NAME(DT_RELRSZ), TAG_NAME(DT_RELA),
TAG_NAME(DT_RELASZ), TAG_NAME(DT_RELAENT), TAG_NAME(DT_REL),
TAG_NAME(DT_RELSZ), TAG_NAME(DT_RELENT), TAG_NAME(DT_STRTAB),
TAG_NAME(DT_STRSZ), TAG_NAME(DT_VERNEED), TAG_NAME(DT_VERNEEDNUM),
};
};
// Translate a virtual address into an offset in the file based on the program
// headers' PT_LOAD.
static Elf_Addr get_offset(const std::vector<Elf_Phdr>& phdr, Elf_Addr addr) {
for (const auto& p : phdr) {
if (p.p_type == PT_LOAD && addr >= p.p_vaddr &&
addr < p.p_vaddr + p.p_filesz) {
return addr - (p.p_vaddr - p.p_paddr);
}
}
return 0;
}
static bool hack(std::fstream& f, bool set_relrhack_bit);
};
template <typename T>
T read_one_at(std::istream& in, off_t pos) {
T result;
in.seekg(pos, std::ios::beg);
in.read(reinterpret_cast<char*>(&result), sizeof(T));
return result;
}
template <typename T>
std::vector<T> read_vector_at(std::istream& in, off_t pos, size_t num) {
std::vector<T> result(num);
in.seekg(pos, std::ios::beg);
in.read(reinterpret_cast<char*>(result.data()), num * sizeof(T));
return result;
}
void write_at(std::ostream& out, off_t pos, const char* buf, size_t len) {
out.seekp(pos, std::ios::beg);
out.write(buf, len);
}
template <typename T>
void write_one_at(std::ostream& out, off_t pos, const T& data) {
write_at(out, pos, reinterpret_cast<const char*>(&data), sizeof(T));
}
template <typename T>
void write_vector_at(std::ostream& out, off_t pos, const std::vector<T>& vec) {
write_at(out, pos, reinterpret_cast<const char*>(&vec.front()),
vec.size() * sizeof(T));
}
template <int bits>
bool RelR<bits>::hack(std::fstream& f, bool set_relrhack_bit) {
auto ehdr = read_one_at<Elf_Ehdr>(f, 0);
if (ehdr.e_phentsize != sizeof(Elf_Phdr)) {
throw std::runtime_error("Invalid ELF?");
}
auto phdr = read_vector_at<Elf_Phdr>(f, ehdr.e_phoff, ehdr.e_phnum);
const auto& dyn_phdr =
std::find_if(phdr.begin(), phdr.end(),
[](const auto& p) { return p.p_type == PT_DYNAMIC; });
if (dyn_phdr == phdr.end()) {
return false;
}
if (dyn_phdr->p_filesz % sizeof(Elf_Dyn)) {
throw std::runtime_error("Invalid ELF?");
}
auto dyn = read_vector_at<Elf_Dyn>(f, dyn_phdr->p_offset,
dyn_phdr->p_filesz / sizeof(Elf_Dyn));
off_t dyn_offset = dyn_phdr->p_offset;
DynInfo dyn_info;
for (const auto& d : dyn) {
if (d.d_tag == DT_NULL) {
break;
}
if (dyn_info.is_wanted(d.d_tag)) {
if (dyn_info.contains(d.d_tag)) {
std::stringstream msg;
msg << dyn_info.name(d.d_tag) << " appears twice?";
throw std::runtime_error(msg.str());
}
dyn_info.insert(dyn_offset, d.d_tag, d.d_un.d_val);
}
dyn_offset += sizeof(Elf_Dyn);
}
// Find the location and size of the SHT_RELR section, which contains the
// packed-relative-relocs.
Elf_Addr relr_off =
dyn_info.contains(DT_RELR) ? get_offset(phdr, dyn_info[DT_RELR]) : 0;
Elf_Off relrsz = dyn_info[DT_RELRSZ];
const decltype(Elf_Dyn::d_tag) rel_tags[3][2] = {
{DT_REL, DT_RELA}, {DT_RELSZ, DT_RELASZ}, {DT_RELENT, DT_RELAENT}};
for (const auto& [rel_tag, rela_tag] : rel_tags) {
if (dyn_info.contains(rel_tag) && dyn_info.contains(rela_tag)) {
std::stringstream msg;
msg << "Both " << dyn_info.name(rel_tag) << " and "
<< dyn_info.name(rela_tag) << " appear?";
throw std::runtime_error(msg.str());
}
}
Elf_Off relent =
dyn_info.contains(DT_RELENT) ? dyn_info[DT_RELENT] : dyn_info[DT_RELAENT];
// Estimate the size of the unpacked relative relocations corresponding
// to the SHT_RELR section.
auto relr = read_vector_at<Elf_Addr>(f, relr_off, relrsz / sizeof(Elf_Addr));
size_t relocs = 0;
for (const auto& entry : relr) {
if ((entry & 1) == 0) {
// LSB is 0, this is a pointer for a single relocation.
relocs++;
} else {
// LSB is 1, remaining bits are a bitmap. Each bit represents a
// relocation.
relocs += __builtin_popcount(entry) - 1;
}
}
// If the packed relocations + some overhead (we pick 4K arbitrarily, the
// real size would require digging into the section sizes of the injected
// .o file, which is not worth the error) is larger than the estimated
// unpacked relocations, we'll just relink without packed relocations.
if (relocs * relent < relrsz + 4096) {
return false;
}
if (set_relrhack_bit) {
// Change DT_RELR* tags to add DT_RELRHACK_BIT.
for (const auto tag : {DT_RELR, DT_RELRSZ, DT_RELRENT}) {
write_one_at(f, dyn_info.offset(tag), tag | DT_RELRHACK_BIT);
}
}
bool is_glibc = false;
if (dyn_info.contains(DT_VERNEEDNUM) && dyn_info.contains(DT_VERNEED) &&
dyn_info.contains(DT_STRSZ) && dyn_info.contains(DT_STRTAB)) {
// Scan SHT_VERNEED for the GLIBC_ABI_DT_RELR version on the libc
// library.
Elf_Addr verneed_off = get_offset(phdr, dyn_info[DT_VERNEED]);
Elf_Off verneednum = dyn_info[DT_VERNEEDNUM];
// SHT_STRTAB section, which contains the string table for, among other
// things, the symbol versions in the SHT_VERNEED section.
auto strtab = read_vector_at<char>(f, get_offset(phdr, dyn_info[DT_STRTAB]),
dyn_info[DT_STRSZ]);
// Guarantee a nul character at the end of the string table.
strtab.push_back(0);
while (verneednum--) {
auto verneed = read_one_at<Elf_Verneed>(f, verneed_off);
if (std::string_view{"libc.so.6"} == &strtab.at(verneed.vn_file)) {
is_glibc = true;
Elf_Addr vernaux_off = verneed_off + verneed.vn_aux;
Elf_Addr relr = 0;
Elf_Vernaux reuse;
for (auto n = 0; n < verneed.vn_cnt; n++) {
auto vernaux = read_one_at<Elf_Vernaux>(f, vernaux_off);
if (std::string_view{"GLIBC_ABI_DT_RELR"} ==
&strtab.at(vernaux.vna_name)) {
relr = vernaux_off;
} else {
reuse = vernaux;
}
vernaux_off += vernaux.vna_next;
}
// In the case where we do have the GLIBC_ABI_DT_RELR version, we
// need to edit the binary to make the following changes:
// - Remove the GLIBC_ABI_DT_RELR version, we replace it with an
// arbitrary other version entry, which is simpler than completely
// removing it. We need to remove it because older versions of glibc
// don't have the version (after all, that's why the symbol version
// is there in the first place, to avoid running against older versions
// of glibc that don't support packed relocations).
// - Alter the DT_RELR* tags in the dynamic section, so that they
// are not recognized by ld.so, because, while all versions of ld.so
// ignore tags they don't know, glibc's ld.so versions that support
// packed relocations don't want to load a binary that has DT_RELR*
// tags but *not* a dependency on the GLIBC_ABI_DT_RELR version.
if (relr) {
// Don't overwrite vn_aux.
write_at(f, relr, reinterpret_cast<char*>(&reuse),
sizeof(reuse) - sizeof(Elf_Word));
}
}
verneed_off += verneed.vn_next;
}
}
// Location of the .rel.plt section.
Elf_Addr jmprel = dyn_info.contains(DT_JMPREL) ? dyn_info[DT_JMPREL] : 0;
if (is_glibc) {
#ifndef MOZ_STDCXX_COMPAT
try {
#endif
// ld.so in glibc 2.16 to 2.23 expects .rel.plt to strictly follow
// BFD ld places .relr.dyn after .rel.plt, so this works fine, but lld
// places it between both sections, which doesn't work out for us. In that
// case, we want to swap .relr.dyn and .rel.plt.
Elf_Addr rel_end = dyn_info.contains(DT_REL)
? (dyn_info[DT_REL] + dyn_info[DT_RELSZ])
: (dyn_info[DT_RELA] + dyn_info[DT_RELASZ]);
if (dyn_info.contains(DT_JMPREL) && dyn_info[DT_PLTRELSZ] &&
dyn_info[DT_JMPREL] != rel_end) {
if (dyn_info[DT_RELR] != rel_end) {
throw CantSwapSections("RELR section doesn't follow REL/RELA?");
}
if (dyn_info[DT_JMPREL] != dyn_info[DT_RELR] + dyn_info[DT_RELRSZ]) {
throw CantSwapSections("PLT REL/RELA doesn't follow RELR?");
}
auto plt_rel = read_vector_at<char>(
f, get_offset(phdr, dyn_info[DT_JMPREL]), dyn_info[DT_PLTRELSZ]);
// Write the content of both sections swapped, and adjust the
// corresponding PT_DYNAMIC entries.
write_vector_at(f, relr_off, plt_rel);
write_vector_at(f, relr_off + plt_rel.size(), relr);
dyn_info[DT_JMPREL] = rel_end;
dyn_info[DT_RELR] = rel_end + plt_rel.size();
for (const auto tag : {DT_JMPREL, DT_RELR}) {
write_one_at(f, dyn_info.offset(tag) + sizeof(typename DynInfo::Tag),
dyn_info[tag]);
}
}
#ifndef MOZ_STDCXX_COMPAT
} catch (const CantSwapSections& err) {
// When binary compatibility with older libstdc++/glibc is not enabled, we
// only emit a warning about why swapping the sections is not happening.
std::cerr << "WARNING: " << err.what() << std::endl;
}
#endif
}
off_t shdr_offset = ehdr.e_shoff;
auto shdr = read_vector_at<Elf_Shdr>(f, ehdr.e_shoff, ehdr.e_shnum);
for (auto& s : shdr) {
// Some tools don't like sections of types they don't know, so change
// SHT_RELR, which might be unknown on older systems, to SHT_PROGBITS.
if (s.sh_type == SHT_RELR) {
s.sh_type = SHT_PROGBITS;
// If DT_RELR has been adjusted to swap with DT_JMPREL, also adjust
// the corresponding SHT_RELR section header.
if (s.sh_addr != dyn_info[DT_RELR]) {
s.sh_offset += dyn_info[DT_RELR] - s.sh_addr;
s.sh_addr = dyn_info[DT_RELR];
}
write_one_at(f, shdr_offset, s);
} else if (jmprel && (s.sh_addr == jmprel) &&
(s.sh_addr != dyn_info[DT_JMPREL])) {
// If DT_JMPREL has been adjusted to swap with DT_RELR, also adjust
// the corresponding section header.
s.sh_offset -= s.sh_addr - dyn_info[DT_JMPREL];
s.sh_addr = dyn_info[DT_JMPREL];
write_one_at(f, shdr_offset, s);
}
shdr_offset += sizeof(Elf_Shdr);
}
return true;
}
std::vector<std::string> get_path() {
std::vector<std::string> result;
std::stringstream stream{std::getenv("PATH")};
std::string item;
while (std::getline(stream, item, ':')) {
result.push_back(std::move(item));
}
return result;
}
std::optional<fs::path> next_program(fs::path& this_program,
std::optional<fs::path>& program) {
auto program_name = program ? *program : this_program.filename();
for (const auto& dir : get_path()) {
auto path = fs::path(dir) / program_name;
auto status = fs::status(path);
if ((status.type() == fs::file_type::regular) &&
((status.permissions() & fs::perms::owner_exec) ==
fs::perms::owner_exec) &&
!fs::equivalent(path, this_program))
return path;
}
return std::nullopt;
}
unsigned char get_elf_class(unsigned char (&e_ident)[EI_NIDENT]) {
if (std::string_view{reinterpret_cast<char*>(e_ident), SELFMAG} !=
std::string_view{ELFMAG, SELFMAG}) {
throw std::runtime_error("Not ELF?");
}
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
if (e_ident[EI_DATA] != ELFDATA2LSB) {
throw std::runtime_error("Not Little Endian ELF?");
}
#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
if (e_ident[EI_DATA] != ELFDATA2MSB) {
throw std::runtime_error("Not Big Endian ELF?");
}
#else
# error Unknown byte order.
#endif
if (e_ident[EI_VERSION] != 1) {
throw std::runtime_error("Not ELF version 1?");
}
auto elf_class = e_ident[EI_CLASS];
if (elf_class != ELFCLASS32 && elf_class != ELFCLASS64) {
throw std::runtime_error("Not 32 or 64-bits ELF?");
}
return elf_class;
}
unsigned char get_elf_class(std::istream& in) {
unsigned char e_ident[EI_NIDENT];
in.read(reinterpret_cast<char*>(e_ident), sizeof(e_ident));
return get_elf_class(e_ident);
}
uint16_t get_elf_machine(std::istream& in) {
// As far as e_machine is concerned, both Elf32_Ehdr and Elf64_Ehdr are equal.
Elf32_Ehdr ehdr;
in.read(reinterpret_cast<char*>(&ehdr), sizeof(ehdr));
// get_elf_class will throw exceptions for the cases we don't handle.
get_elf_class(ehdr.e_ident);
return ehdr.e_machine;
}
int run_command(std::vector<const char*>& args, bool use_response_file) {
std::string at_file;
const char** argv = args.data();
std::array<const char*, 3> args_with_atfile{};
if (use_response_file) {
const char* tmpdir = getenv("TMPDIR");
if (!tmpdir) {
tmpdir = "/tmp";
}
std::string tmpfile = tmpdir;
tmpfile += "/relrhackXXXXXX";
int fd = mkstemp(tmpfile.data());
if (fd < 0) {
std::cerr << "Failed to create temporary file." << std::endl;
return 1;
}
close(fd);
std::ofstream f{tmpfile, f.binary};
for (auto arg = std::next(args.begin()); arg != args.end(); ++arg) {
f << *arg << "\n";
}
at_file = "@";
at_file += tmpfile;
args_with_atfile = {args.front(), at_file.c_str(), nullptr};
argv = args_with_atfile.data();
}
auto guard = mozilla::MakeScopeExit([&] {
if (!at_file.empty()) {
unlink(at_file.c_str() + 1);
}
});
pid_t child_pid;
if (posix_spawn(&child_pid, args[0], nullptr, nullptr,
const_cast<char* const*>(argv), environ) != 0) {
throw std::runtime_error("posix_spawn failed");
}
int status;
waitpid(child_pid, &status, 0);
return WEXITSTATUS(status);
}
int main(int argc, char* argv[]) {
auto this_program = fs::absolute(argv[0]);
std::vector<const char*> args;
int i, crti = 0;
std::optional<fs::path> output = std::nullopt;
std::optional<fs::path> real_linker = std::nullopt;
bool shared = false;
bool is_android = false;
bool use_response_file = false;
std::vector<char> response_file;
std::vector<const char*> response_file_args;
uint16_t elf_machine = EM_NONE;
// Scan argv in order to prepare the following:
// - get the output file. That's the file we may need to adjust.
// - get the --real-linker if one was passed.
// - detect whether we're linking a shared library or something else. As of
// now, only shared libraries are handled. Technically speaking, programs
// could be handled as well, but for the purpose of Firefox, that actually
// doesn't work because programs contain a memory allocator that ends up
// being called before the injected code has any chance to apply relocations,
// and the allocator itself needs the relocations to have been applied.
// - detect the position of crti.o so that we can inject our own object
// right after it, and also to detect the machine type to pick the right
// object to inject.
//
// At the same time, we also construct a new list of arguments, with
// --real-linker filtered out. We'll later inject arguments in that list.
if (argc == 2 && argv[1] && argv[1][0] == '@') {
// When GCC is given a response file, it wraps all arguments into a
// new response file with all arguments, even if originally there were
// arguments and a response file.
// In that case, we can't scan for arguments, so we need to read the
// response file. And as we change the arguments, we'll need to write
// a new one.
std::ifstream f{argv[1] + 1, f.binary | f.ate};
if (!f) {
std::cerr << "Failed to read " << argv[1] + 1 << std::endl;
return 1;
}
size_t len = f.tellg();
response_file = read_vector_at<char>(f, 0, len);
std::replace(response_file.begin(), response_file.end(), '\n', '\0');
if (len && response_file[len - 1] != '\0') {
response_file.push_back('\0');
}
response_file_args.push_back(argv[0]);
for (const char* a = response_file.data();
a < response_file.data() + response_file.size(); a += strlen(a) + 1) {
response_file_args.push_back(a);
}
argv = const_cast<char**>(response_file_args.data());
argc = response_file_args.size();
use_response_file = true;
}
for (i = 1, argv++; i < argc && *argv; argv++, i++) {
std::string_view arg{*argv};
if (arg == "-shared") {
shared = true;
} else if (arg == "-o") {
args.push_back(*(argv++));
++i;
output = *argv;
} else if (arg == "--real-linker") {
++i;
real_linker = *(++argv);
continue;
} else if (elf_machine == EM_NONE) {
auto filename = fs::path(arg).filename();
if (filename == "crti.o" || filename == "crtbegin_so.o") {
is_android = (filename == "crtbegin_so.o");
crti = i;
std::fstream f{std::string(arg), f.binary | f.in};
f.exceptions(f.failbit);
elf_machine = get_elf_machine(f);
}
}
args.push_back(*argv);
}
if (!output) {
std::cerr << "Could not determine output file." << std::endl;
return 1;
}
if (!crti) {
std::cerr << "Could not find CRT object on the command line." << std::endl;
return 1;
}
if (!real_linker || !real_linker->has_parent_path()) {
auto linker = next_program(this_program, real_linker);
if (!linker) {
std::cerr << "Could not find next "
<< (real_linker ? real_linker->filename()
: this_program.filename())
<< std::endl;
return 1;
}
real_linker = linker;
}
args.insert(args.begin(), real_linker->c_str());
args.push_back(nullptr);
std::string stem;
switch (elf_machine) {
case EM_NONE:
std::cerr << "Could not determine target machine type." << std::endl;
return 1;
case EM_386:
stem = "x86";
break;
case EM_X86_64:
stem = "x86_64";
break;
case EM_ARM:
stem = "arm";
break;
case EM_AARCH64:
stem = "aarch64";
break;
default:
std::cerr << "Unsupported target machine type." << std::endl;
return 1;
}
if (is_android) {
stem += "-android";
}
if (shared) {
std::vector<const char*> hacked_args(args);
auto inject = this_program.parent_path() / "inject" / (stem + ".o");
hacked_args.insert(hacked_args.begin() + crti + 1, inject.c_str());
hacked_args.insert(hacked_args.end() - 1, {"-z", "pack-relative-relocs",
"-init=_relrhack_wrap_init"});
int status = run_command(hacked_args, use_response_file);
if (status) {
return status;
}
bool hacked = false;
try {
std::fstream f{*output, f.binary | f.in | f.out};
f.exceptions(f.failbit);
auto elf_class = get_elf_class(f);
f.seekg(0, std::ios::beg);
// On Android, we don't set the relrhack bit so that the system linker
// can handle the RELR relocations when supported. On desktop, the
// glibc unfortunately rejects the binaries without the bit set.
// (see comment about GLIBC_ABI_DT_RELR)
if (elf_class == ELFCLASS32) {
hacked = RelR<32>::hack(f, /* set_relrhack_bit = */ !is_android);
} else if (elf_class == ELFCLASS64) {
hacked = RelR<64>::hack(f, /* set_relrhack_bit = */ !is_android);
}
} catch (const std::runtime_error& err) {
std::cerr << "Failed to hack " << output->string() << ": " << err.what()
<< std::endl;
return 1;
}
if (hacked) {
return 0;
}
}
return run_command(args, use_response_file);
}