cpu_patches: Patch stack canary accesses (#3857)

* Patch stack checks done using fs:[0x28]

Additionally adds support for multiple patches per instruction, since this makes two separate patches we need to conditionally perform for mov instructions.

* Missing include

* Disable patches for Apple

Mac can use their native FS segment directly, so these patches aren't needed

* Oops
This commit is contained in:
Stephen Miller
2025-12-07 00:47:39 -06:00
committed by GitHub
parent d3ad728ac0
commit 391d30cbb1

View File

@@ -5,6 +5,7 @@
#include <memory>
#include <mutex>
#include <set>
#include <vector>
#include <Zydis/Zydis.h>
#include <xbyak/xbyak.h>
#include <xbyak/xbyak_util.h>
@@ -122,6 +123,30 @@ static void GenerateTcbAccess(void* /* address */, const ZydisDecodedOperand* op
#endif
}
static bool FilterStackCheck(const ZydisDecodedOperand* operands) {
const auto& dst_op = operands[0];
const auto& src_op = operands[1];
// Some compilers emit stack checks by starting a function with
// 'mov (64-bit register), fs:[0x28]', then checking with `xor (64-bit register), fs:[0x28]`
return src_op.type == ZYDIS_OPERAND_TYPE_MEMORY && src_op.mem.segment == ZYDIS_REGISTER_FS &&
src_op.mem.base == ZYDIS_REGISTER_NONE && src_op.mem.index == ZYDIS_REGISTER_NONE &&
src_op.mem.disp.value == 0x28 && dst_op.reg.value >= ZYDIS_REGISTER_RAX &&
dst_op.reg.value <= ZYDIS_REGISTER_R15;
}
static void GenerateStackCheck(void* /* address */, const ZydisDecodedOperand* operands,
Xbyak::CodeGenerator& c) {
const auto dst = ZydisToXbyakRegisterOperand(operands[0]);
c.xor_(dst, 0);
}
static void GenerateStackCanary(void* /* address */, const ZydisDecodedOperand* operands,
Xbyak::CodeGenerator& c) {
const auto dst = ZydisToXbyakRegisterOperand(operands[0]);
c.mov(dst, 0);
}
static bool FilterNoSSE4a(const ZydisDecodedOperand*) {
Cpu cpu;
return !cpu.has(Cpu::tSSE4a);
@@ -440,18 +465,26 @@ struct PatchInfo {
bool trampoline;
};
static const std::unordered_map<ZydisMnemonic, PatchInfo> Patches = {
static const std::unordered_map<ZydisMnemonic, std::vector<PatchInfo>> Patches = {
// SSE4a
{ZYDIS_MNEMONIC_EXTRQ, {FilterNoSSE4a, GenerateEXTRQ, true}},
{ZYDIS_MNEMONIC_INSERTQ, {FilterNoSSE4a, GenerateINSERTQ, true}},
{ZYDIS_MNEMONIC_MOVNTSS, {FilterNoSSE4a, ReplaceMOVNTSS, false}},
{ZYDIS_MNEMONIC_MOVNTSD, {FilterNoSSE4a, ReplaceMOVNTSD, false}},
{ZYDIS_MNEMONIC_EXTRQ, {{FilterNoSSE4a, GenerateEXTRQ, true}}},
{ZYDIS_MNEMONIC_INSERTQ, {{FilterNoSSE4a, GenerateINSERTQ, true}}},
{ZYDIS_MNEMONIC_MOVNTSS, {{FilterNoSSE4a, ReplaceMOVNTSS, false}}},
{ZYDIS_MNEMONIC_MOVNTSD, {{FilterNoSSE4a, ReplaceMOVNTSD, false}}},
#if !defined(__APPLE__)
// FS segment patches
// These first two patches are for accesses to the stack canary, fs:[0x28]
{ZYDIS_MNEMONIC_XOR, {{FilterStackCheck, GenerateStackCheck, false}}},
{ZYDIS_MNEMONIC_MOV,
{{FilterStackCheck, GenerateStackCanary, false},
#if defined(_WIN32)
// Windows needs a trampoline.
{ZYDIS_MNEMONIC_MOV, {FilterTcbAccess, GenerateTcbAccess, true}},
#elif !defined(__APPLE__)
{ZYDIS_MNEMONIC_MOV, {FilterTcbAccess, GenerateTcbAccess, false}},
// Windows needs a trampoline for Tcb accesses.
{FilterTcbAccess, GenerateTcbAccess, true}
#else
{FilterTcbAccess, GenerateTcbAccess, false}
#endif
}},
#endif
};
@@ -503,7 +536,8 @@ static std::pair<bool, u64> TryPatch(u8* code, PatchModule* module) {
}
if (Patches.contains(instruction.mnemonic)) {
const auto& patch_info = Patches.at(instruction.mnemonic);
const auto& patches = Patches.at(instruction.mnemonic);
for (const auto& patch_info : patches) {
bool needs_trampoline = patch_info.trampoline;
if (patch_info.filter(operands)) {
auto& patch_gen = module->patch_gen;
@@ -551,6 +585,7 @@ static std::pair<bool, u64> TryPatch(u8* code, PatchModule* module) {
}
}
}
}
return std::make_pair(false, instruction.length);
}