fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[CPU/XThread] Add DWARF .eh_frame unwind info for JIT code on Linux

Browse Source 
Generate per-function CIE+FDE records and register them via
__register_frame so the C++ exception unwinder can propagate through
JIT frames. Replace setjmp/longjmp fiber reentry with throw/catch
on Linux to ensure destructors and RAII guards run during fiber
stack switches.
This commit is contained in:
Herman S.
2026-03-10 22:48:15 +09:00
parent f45a254774
commit b24ea8ef99
3 changed files with 358 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

293
src/xenia/cpu/backend/x64/x64_code_cache_posix.cc
View File

@@ -9,11 +9,74 @@
#include "xenia/cpu/backend/x64/x64_code_cache.h"
#include <cstring>
#include <vector>
#include "xenia/base/assert.h"
#include "xenia/base/logging.h"
#include "xenia/base/math.h"
#include "xenia/cpu/backend/x64/x64_stack_layout.h"
// libgcc/libunwind APIs for registering DWARF .eh_frame unwind info.
extern "C" void __register_frame(void*);
extern "C" void __deregister_frame(void*);
namespace xe {
namespace cpu {
namespace backend {
namespace x64 {
// Maximum size of DWARF .eh_frame data per function (CIE + FDE + terminator).
static constexpr uint32_t kMaxUnwindInfoSize = 96;
// DWARF register numbers for x86-64.
static constexpr uint8_t kDwarfRegRBX = 3;
static constexpr uint8_t kDwarfRegRBP = 6;
static constexpr uint8_t kDwarfRegRSP = 7;
static constexpr uint8_t kDwarfRegR12 = 12;
static constexpr uint8_t kDwarfRegR13 = 13;
static constexpr uint8_t kDwarfRegR14 = 14;
static constexpr uint8_t kDwarfRegR15 = 15;
static constexpr uint8_t kDwarfRegRA = 16;
// DWARF CFA opcodes.
static constexpr uint8_t kDW_CFA_advance_loc1 = 0x02;
static constexpr uint8_t kDW_CFA_advance_loc2 = 0x03;
static constexpr uint8_t kDW_CFA_def_cfa = 0x0c;
static constexpr uint8_t kDW_CFA_def_cfa_offset = 0x0e;
static constexpr uint8_t kDW_CFA_nop = 0x00;
// DWARF pointer encoding constants.
static constexpr uint8_t kDW_EH_PE_pcrel = 0x10;
static constexpr uint8_t kDW_EH_PE_sdata4 = 0x0b;
static size_t WriteULEB128(uint8_t* p, uint64_t value) {
size_t count = 0;
do {
uint8_t byte = value & 0x7F;
value >>= 7;
if (value) byte |= 0x80;
p[count++] = byte;
} while (value);
return count;
}
static size_t WriteSLEB128(uint8_t* p, int64_t value) {
size_t count = 0;
bool more = true;
while (more) {
uint8_t byte = value & 0x7F;
value >>= 7;
if ((value == 0 && !(byte & 0x40)) || (value == -1 && (byte & 0x40))) {
more = false;
} else {
byte |= 0x80;
}
p[count++] = byte;
}
return count;
}
class PosixX64CodeCache : public X64CodeCache {
public:
PosixX64CodeCache();
@@ -24,16 +87,19 @@ class PosixX64CodeCache : public X64CodeCache {
void* LookupUnwindInfo(uint64_t host_pc) override { return nullptr; }
private:
/*
UnwindReservation RequestUnwindReservation(uint8_t* entry_address) override;
void PlaceCode(uint32_t guest_address, void* machine_code, size_t code_size,
size_t stack_size, void* code_execute_address,
void PlaceCode(uint32_t guest_address, void* machine_code,
const EmitFunctionInfo& func_info, void* code_execute_address,
UnwindReservation unwind_reservation) override;
void InitializeUnwindEntry(uint8_t* unwind_entry_address,
size_t unwind_table_slot, void* code_address,
size_t code_size, size_t stack_size);
*/
void* code_execute_address,
const EmitFunctionInfo& func_info);
// Pointers registered with __register_frame, for cleanup.
std::vector<void*> registered_frames_;
// Current number of unwind table entries.
uint32_t unwind_table_count_ = 0;
};
std::unique_ptr<X64CodeCache> X64CodeCache::Create() {
@@ -41,9 +107,220 @@ std::unique_ptr<X64CodeCache> X64CodeCache::Create() {
}
PosixX64CodeCache::PosixX64CodeCache() = default;
PosixX64CodeCache::~PosixX64CodeCache() = default;
bool PosixX64CodeCache::Initialize() { return X64CodeCache::Initialize(); }
PosixX64CodeCache::~PosixX64CodeCache() {
for (auto frame : registered_frames_) {
__deregister_frame(frame);
}
}
bool PosixX64CodeCache::Initialize() {
if (!X64CodeCache::Initialize()) {
return false;
}
registered_frames_.reserve(kMaximumFunctionCount);
return true;
}
X64CodeCache::UnwindReservation PosixX64CodeCache::RequestUnwindReservation(
uint8_t* entry_address) {
#if defined(NDEBUG)
if (unwind_table_count_ >= kMaximumFunctionCount) {
xe::FatalError(
"Unwind table count exceeded maximum! Please report this to "
"Xenia developers");
}
#else
assert_false(unwind_table_count_ >= kMaximumFunctionCount);
#endif
UnwindReservation unwind_reservation;
unwind_reservation.data_size = xe::round_up(kMaxUnwindInfoSize, 16);
unwind_reservation.table_slot = unwind_table_count_++;
unwind_reservation.entry_address = entry_address;
return unwind_reservation;
}
void PosixX64CodeCache::PlaceCode(uint32_t guest_address, void* machine_code,
const EmitFunctionInfo& func_info,
void* code_execute_address,
UnwindReservation unwind_reservation) {
// Write the DWARF .eh_frame data into the reserved unwind space.
InitializeUnwindEntry(unwind_reservation.entry_address, code_execute_address,
func_info);
// Register with the runtime unwinder using the execute-side address.
// The execute mapping is readable (kExecuteReadOnly = PROT_EXEC|PROT_READ),
// so the unwinder can read the .eh_frame data at runtime.
void* unwind_execute_address = unwind_reservation.entry_address -
generated_code_write_base_ +
generated_code_execute_base_;
__register_frame(unwind_execute_address);
registered_frames_.push_back(unwind_execute_address);
}
void PosixX64CodeCache::InitializeUnwindEntry(
uint8_t* unwind_entry_address, void* code_execute_address,
const EmitFunctionInfo& func_info) {
// Compute execute-side base address of the unwind buffer.
// We write via the write mapping but pc-relative offsets must be relative
// to the execute mapping (which is what __register_frame sees).
uint8_t* unwind_execute_base = unwind_entry_address -
generated_code_write_base_ +
generated_code_execute_base_;
uint8_t* p = unwind_entry_address;
uint8_t* cie_start = p;
// === CIE (Common Information Entry) ===
uint8_t* cie_length_ptr = p;
p += 4; // placeholder for length
uint8_t* cie_content_start = p;
// CIE ID = 0 (distinguishes CIE from FDE in .eh_frame format).
*reinterpret_cast<uint32_t*>(p) = 0;
p += 4;
// Version = 1.
*p++ = 1;
// Augmentation string "zR" - indicates augmentation data with FDE encoding.
*p++ = 'z';
*p++ = 'R';
*p++ = '\0';
// Code alignment factor = 1.
p += WriteULEB128(p, 1);
// Data alignment factor = -8.
p += WriteSLEB128(p, -8);
// Return address register column = 16 (x86-64 RA).
p += WriteULEB128(p, kDwarfRegRA);
// Augmentation data length = 1 (just the FDE encoding byte).
p += WriteULEB128(p, 1);
// FDE pointer encoding: pc-relative, signed 32-bit.
*p++ = kDW_EH_PE_pcrel | kDW_EH_PE_sdata4;
// Initial instructions:
// DW_CFA_def_cfa RSP, 8 — at function entry, CFA = RSP + 8.
*p++ = kDW_CFA_def_cfa;
p += WriteULEB128(p, kDwarfRegRSP);
p += WriteULEB128(p, 8);
// DW_CFA_offset RA, 1 — return address at CFA - 8 (factored: 1 * 8).
*p++ = 0x80 | kDwarfRegRA;
p += WriteULEB128(p, 1);
// Pad CIE to pointer-size (8-byte) alignment.
size_t cie_content_len = static_cast<size_t>(p - cie_content_start);
size_t cie_padded_len = xe::round_up(cie_content_len, sizeof(void*));
while (p < cie_content_start + cie_padded_len) {
*p++ = kDW_CFA_nop;
}
// Write CIE length (excludes the length field itself).
*reinterpret_cast<uint32_t*>(cie_length_ptr) =
static_cast<uint32_t>(p - cie_content_start);
// === FDE (Frame Description Entry) ===
uint8_t* fde_length_ptr = p;
p += 4; // placeholder for length
uint8_t* fde_content_start = p;
// CIE pointer: offset from this field back to the start of the CIE.
*reinterpret_cast<uint32_t*>(p) = static_cast<uint32_t>(p - cie_start);
p += 4;
// PC begin: pc-relative offset to the start of the function code.
// Computed relative to the execute-side address of this field.
uint8_t* pc_begin_execute_addr =
unwind_execute_base + (p - unwind_entry_address);
*reinterpret_cast<int32_t*>(p) =
static_cast<int32_t>(reinterpret_cast<intptr_t>(code_execute_address) -
reinterpret_cast<intptr_t>(pc_begin_execute_addr));
p += 4;
// PC range: size of the function code.
*reinterpret_cast<uint32_t*>(p) =
static_cast<uint32_t>(func_info.code_size.total);
p += 4;
// Augmentation data length = 0 (no LSDA pointer).
p += WriteULEB128(p, 0);
// FDE instructions: describe how the stack frame changes during the prolog.
if (func_info.stack_size > 0) {
// Advance location to the instruction after the stack allocation.
size_t alloc_offset = func_info.prolog_stack_alloc_offset;
assert_true(alloc_offset > 0);
if (alloc_offset < 64) {
*p++ = 0x40 | static_cast<uint8_t>(alloc_offset);
} else if (alloc_offset < 256) {
*p++ = kDW_CFA_advance_loc1;
*p++ = static_cast<uint8_t>(alloc_offset);
} else {
*p++ = kDW_CFA_advance_loc2;
*reinterpret_cast<uint16_t*>(p) = static_cast<uint16_t>(alloc_offset);
p += 2;
}
// DW_CFA_def_cfa_offset: CFA = RSP + 8 + stack_size after stack alloc.
*p++ = kDW_CFA_def_cfa_offset;
p += WriteULEB128(p, 8 + func_info.stack_size);
// For thunk functions, encode callee-saved register save locations.
// The thunk saves non-volatile registers at known offsets from RSP.
if (func_info.stack_size == StackLayout::THUNK_STACK_SIZE) {
size_t cfa = 8 + func_info.stack_size; // 272
// RBX at rsp+0x18 → CFA-248, factored offset = 31
*p++ = 0x80 | kDwarfRegRBX;
p += WriteULEB128(p, (cfa - 0x18) / 8);
// RBP at rsp+0x20 → CFA-240, factored offset = 30
*p++ = 0x80 | kDwarfRegRBP;
p += WriteULEB128(p, (cfa - 0x20) / 8);
// R12 at rsp+0x40 → CFA-208, factored offset = 26
*p++ = 0x80 | kDwarfRegR12;
p += WriteULEB128(p, (cfa - 0x40) / 8);
// R13 at rsp+0x48 → CFA-200, factored offset = 25
*p++ = 0x80 | kDwarfRegR13;
p += WriteULEB128(p, (cfa - 0x48) / 8);
// R14 at rsp+0x50 → CFA-192, factored offset = 24
*p++ = 0x80 | kDwarfRegR14;
p += WriteULEB128(p, (cfa - 0x50) / 8);
// R15 at rsp+0x58 → CFA-184, factored offset = 23
*p++ = 0x80 | kDwarfRegR15;
p += WriteULEB128(p, (cfa - 0x58) / 8);
}
}
// Pad FDE to pointer-size (8-byte) alignment.
size_t fde_content_len = static_cast<size_t>(p - fde_content_start);
size_t fde_padded_len = xe::round_up(fde_content_len, sizeof(void*));
while (p < fde_content_start + fde_padded_len) {
*p++ = kDW_CFA_nop;
}
// Write FDE length.
*reinterpret_cast<uint32_t*>(fde_length_ptr) =
static_cast<uint32_t>(p - fde_content_start);
// === Terminator (zero-length entry) ===
*reinterpret_cast<uint32_t*>(p) = 0;
p += 4;
assert_true(static_cast<size_t>(p - unwind_entry_address) <=
kMaxUnwindInfoSize);
}
} // namespace x64
} // namespace backend

67
src/xenia/kernel/xthread.cc
View File

@@ -9,6 +9,10 @@
#include "xenia/kernel/xthread.h"
#if XE_PLATFORM_LINUX
#include <signal.h>
#endif
#include "xenia/base/byte_stream.h"
#include "xenia/base/logging.h"
#include "xenia/base/platform.h"
@@ -541,29 +545,59 @@ void XThread::Execute() {
want_exit_code = true;
}
// Set up reentry jump buffer for fiber-based stack switching.
// When Reenter() is called (e.g., by KeSetCurrentStackPointers), it will
// longjmp back here instead of throwing an exception through JIT code.
// Set up reentry mechanism for fiber-based stack switching.
// When Reenter() is called (e.g., by KeSetCurrentStackPointers), it
// unwinds back here to re-enter at a new guest address.
//
// On Linux, C++ exceptions are used so that DWARF unwind info (registered
// for JIT code via __register_frame) allows proper destructor/RAII cleanup
// through both JIT and host C++ frames.
//
// On Windows, setjmp/longjmp is used because MSVC's longjmp performs SEH
// stack unwinding which already calls destructors.
uint32_t next_address;
#if XE_PLATFORM_LINUX
try {
exit_code = static_cast<int>(kernel_state()->processor()->Execute(
thread_state_, address, args.data(), args.size()));
next_address = 0;
} catch (const FiberReentryException& e) {
// Ensure SIGRTMIN (used for thread suspend) is not left blocked.
sigset_t set;
sigemptyset(&set);
sigaddset(&set, SIGRTMIN);
pthread_sigmask(SIG_UNBLOCK, &set, nullptr);
next_address = e.address;
}
while (next_address != 0) {
try {
kernel_state()->processor()->ExecuteRaw(thread_state_, next_address);
next_address = 0;
if (want_exit_code) {
exit_code = static_cast<int>(thread_state_->context()->r[3]);
}
} catch (const FiberReentryException& e) {
sigset_t set;
sigemptyset(&set);
sigaddset(&set, SIGRTMIN);
pthread_sigmask(SIG_UNBLOCK, &set, nullptr);
next_address = e.address;
}
}
#else
if (setjmp(reentry_jmp_buf_) != 0) {
// Longjmp returned here - reentry requested
next_address = reentry_address_;
} else {
// Initial execution
exit_code = static_cast<int>(kernel_state()->processor()->Execute(
thread_state_, address, args.data(), args.size()));
next_address = 0;
}
// Handle reentry loop for fiber switching.
// See XThread::Reenter comments.
while (next_address != 0) {
// Set up jump buffer for potential reentries during this execution
if (setjmp(reentry_jmp_buf_) != 0) {
// Nested reentry occurred
next_address = reentry_address_;
} else {
// Execute at the reentry address
kernel_state()->processor()->ExecuteRaw(thread_state_, next_address);
next_address = 0;
if (want_exit_code) {
@@ -571,6 +605,7 @@ void XThread::Execute() {
}
}
}
#endif
// If we got here it means the execute completed without an exit being called.
// Treat the return code as an implicit exit code (if desired).
@@ -578,12 +613,18 @@ void XThread::Execute() {
}
void XThread::Reenter(uint32_t address) {
// Use setjmp/longjmp instead of exceptions to avoid issues with unwinding
// through JIT-compiled code, which lacks exception handling metadata.
// This is called when the game switches fiber stacks (e.g., via
// Called when the game switches fiber stacks (e.g., via
// KeSetCurrentStackPointers in games like Forza Horizon 2).
// Must unwind through all frames between here and Execute().
#if XE_PLATFORM_LINUX
// Throw a C++ exception that unwinds through JIT frames (using DWARF
// .eh_frame info) and host frames (using compiler-generated DWARF),
// calling destructors properly along the way.
throw FiberReentryException{address};
#else
reentry_address_ = address;
std::longjmp(reentry_jmp_buf_, 1);
#endif
}
void XThread::EnterCriticalRegion() {

21
src/xenia/kernel/xthread.h
View File

@@ -11,14 +11,17 @@
#define XENIA_KERNEL_XTHREAD_H_
#include <atomic>
#include <csetjmp>
#include <string>
#include "xenia/base/mutex.h"
#if XE_PLATFORM_LINUX
#include <condition_variable>
#include <csignal>
#include <mutex>
#endif
#if XE_PLATFORM_WIN32
#include <csetjmp>
#endif
#include "xenia/base/threading.h"
#include "xenia/cpu/thread.h"
#include "xenia/cpu/thread_state.h"
@@ -340,6 +343,15 @@ struct X_KTHREAD {
};
static_assert_size(X_KTHREAD, 0xAB0);
#if XE_PLATFORM_LINUX
// Exception thrown by XThread::Reenter() to unwind through JIT frames.
// C++ exception unwinding uses DWARF .eh_frame info registered for JIT code,
// ensuring destructors and RAII guards in host C++ frames are properly called.
struct FiberReentryException {
uint32_t address;
};
#endif
class XThread : public XObject, public cpu::Thread {
public:
static const XObject::Type kObjectType = XObject::Type::Thread;
@@ -484,9 +496,14 @@ class XThread : public XObject, public cpu::Thread {
std::condition_variable suspend_cv_;
#endif
// Reentry context for setjmp/longjmp based stack unwinding
// Reentry mechanism for fiber-based stack switching.
// On Linux, C++ exceptions are used instead of setjmp/longjmp so that
// destructors and RAII guards in host C++ frames are properly unwound.
// JIT code has DWARF .eh_frame unwind info registered via __register_frame.
#if XE_PLATFORM_WIN32
std::jmp_buf reentry_jmp_buf_;
uint32_t reentry_address_ = 0;
#endif
std::mutex thread_lock_;
};