Improve crash unwinding on Fuchsia

[aarongreen]

Summary

Fuchsia doesn't have signals; instead it expects processes to have a dedicated exception thread that binds to the process' exception port and waits for exception packets to be delivered. On the other hand, libFuzzer and sanitizer_common use expect to collect crash information via libunwind from the same thread that caused the exception.

The long term fix is to improve support for remote unwinding in libunbwind, plumb this through sanitizer_common and libFuzzer, and handle the exception exclusively on the exception thread. In the meantime, this revision has the exception thread "resurrect" the crashing thread by:

• saving its general purpose register state onto the crashing thread's stack,
• setting the crashing thread's program counter to an assembly trampoline with the CFI information needed by libunwind, and
• resuming the crashed thread.

[phosek]

lib/fuzzer/FuzzerUtilFuchsia.cpp

80–81

Do you care about atexit hooks being executed? I think this should probably use _exit instead.

183

I'd use an explicit #elif defined(__aarch64__) and #error in the #else case. You never know when someone tries to compile this on some other architecture.

[aarongreen]

Addressed Petr's comment, and modified the visibility of CrashTrampolineAsm and MakeTrampoline to suppress their warnings.

[mcgrathr]

Almost there, but some nits.

lib/fuzzer/FuzzerUtilFuchsia.cpp

44

Clarify in the comment that this appears to have external linkage to C++ and that's why it's not in the anonymous namespace, but the assembly definition inside MakeTrampoline(), below, actually defines the symbol with internal linkage only.

59

You don't want this printf for the success case in production code.

82

typo: s/and/an/

84

typo: trailing slash should be period

89

Alternatively, Fuchsia may in future actually implement basic signal handling for the machine trap signals.

142

lr is not special, it's just an alias for x30 so I'd use OP_NUM(30) here.
Only sp (and pc) are really special.

157

This can use offsetof(..., r[num]) instead of local arithmetic. It's not standard but it works in Clang and GCC and always will.

163

Make this static (no extern "C") and pass it into the asm.

171

Clarify that this function is never really used as such and that's why the attribute is necessary.
It's just a container around the asm so it can use operands for compile-time computed constants.

174

No need for a variable here.

179

This can be ".cfi_startproc simple" because you supply all the details explicitly.

185

This can be "call %c[StaticCrashHandler]" with the operand being "i' (StaticCrashHandler).

192

This can be "bl %[StaticCrashHandler]" with the operand being "i' (StaticCrashHandler).
Make this the last operand and it solves to comma-eating problem too.

205

Just use "i" (0) here.

212

Can use initialized member decls and default constructor.
I'd make a ScopedHandle instead and just have two of them.

261

This should use __unsanitized_memcpy from <zircon/sanitizer.h>.
Otherwise it's theoretically possible for this thread to reenter the sanitizer runtime and that could go badly.

267

Also needs rounded down to modulo 16.

307

typo: s/its/it's/

[aarongreen]

Addressed Roland's comments.

[mcgrathr]

Looks great!

lib/fuzzer/FuzzerUtilFuchsia.cpp

178

superfluous ;

[aarongreen]

Scrubbed superfluous semicolon.

lib/fuzzer/FuzzerUtilFuchsia.cpp

 #include <string>
 #include <sys/select.h>
 #include <thread>
 #include <unistd.h>
 #include <zircon/errors.h>
 #include <zircon/process.h>
 #include <zircon/status.h>
 #include <zircon/syscalls.h>
+#include <zircon/syscalls/debug.h>
+#include <zircon/syscalls/exception.h>
 #include <zircon/syscalls/port.h>
 #include <zircon/types.h>
 
 namespace fuzzer {
 
+// Given that Fuchsia doesn't have the POSIX signals that libFuzzer was written
+// around, the general approach is to spin up dedicated threads to watch for
+// each requested condition (alarm, interrupt, crash).  Of these, the crash
+// handler is the most involved, as it requires resuming the crashed thread in
+// order to invoke the sanitizers to get the needed state.
+
+// Forward declaration of assembly trampoline needed to resume crashed threads.
+void CrashTrampolineAsm() __asm__("CrashTrampolineAsm");
+
 namespace {
 
 // A magic value for the Zircon exception port, chosen to spell 'FUZZING'
 // when interpreted as a byte sequence on little-endian platforms.
 const uint64_t kFuzzingCrash = 0x474e495a5a5546;
 
+// Helper function to handle Zircon syscall failures.
+void ExitOnErr(zx_status_t Status, const char *Syscall) {
+  if (Status != ZX_OK) {
+    Printf("libFuzzer: %s failed: %s\n", Syscall,
+           _zx_status_get_string(Status));
+    exit(1);
+  } else {
+    Printf("libFuzzer: %s succeeded\n", Syscall);
+  }
+}
+
 void AlarmHandler(int Seconds) {
   while (true) {
     SleepSeconds(Seconds);
     Fuzzer::StaticAlarmCallback();
   }
 }
 
 void InterruptHandler() {
   fd_set readfds;
   // Ctrl-C sends ETX in Zircon.
   do {
     FD_ZERO(&readfds);
     FD_SET(STDIN_FILENO, &readfds);
     select(STDIN_FILENO + 1, &readfds, nullptr, nullptr, nullptr);
   } while(!FD_ISSET(STDIN_FILENO, &readfds) || getchar() != 0x03);
   Fuzzer::StaticInterruptCallback();
 }
 
-void CrashHandler(zx_handle_t *Port) {
-  std::unique_ptr<zx_handle_t> ExceptionPort(Port);
-  zx_port_packet_t Packet;
-  _zx_port_wait(*ExceptionPort, ZX_TIME_INFINITE, &Packet);
-  // Unbind as soon as possible so we don't receive exceptions from this thread.
-  if (_zx_task_bind_exception_port(ZX_HANDLE_INVALID, ZX_HANDLE_INVALID,
-                                   kFuzzingCrash, 0) != ZX_OK) {
-    // Shouldn't happen; if it does the safest option is to just exit.
-    Printf("libFuzzer: unable to unbind exception port; aborting!\n");
-    exit(1);
+// For the crash handler, we need to call Fuzzer::StaticCrashSignalCallback
+// without POSIX signal handlers.  To achieve this, we use and assembly function
+// to add the necessary CFI unwinding information and a C function to bridge
+// from that back into C++/
+
+// FIXME: This works as a short-term solution, but this code really shouldn't be
+// architecture dependent. A better long term solution is to implement remote
+// unwinding and expose the necessary APIs through sanitizer_common and/or ASAN
+// to allow the exception handling thread to gather the crash state directly.
+#if defined(__x86_64__)
+#define FOREACH_REGISTER(OP_REG, OP_NUM) \
+  OP_REG(rax)                            \
+  OP_REG(rbx)                            \
+  OP_REG(rcx)                            \
+  OP_REG(rdx)                            \
+  OP_REG(rsi)                            \
+  OP_REG(rdi)                            \
+  OP_REG(rbp)                            \
+  OP_REG(rsp)                            \
+  OP_REG(r8)                             \
+  OP_REG(r9)                             \
+  OP_REG(r10)                            \
+  OP_REG(r11)                            \
+  OP_REG(r12)                            \
+  OP_REG(r13)                            \
+  OP_REG(r14)                            \
+  OP_REG(r15)                            \
+  OP_REG(rip)
+
+#elif defined(__aarch64__)
+#define FOREACH_REGISTER(OP_REG, OP_NUM) \
+  OP_NUM(0)                              \
+  OP_NUM(1)                              \
+  OP_NUM(2)                              \
+  OP_NUM(3)                              \
+  OP_NUM(4)                              \
+  OP_NUM(5)                              \
+  OP_NUM(6)                              \
+  OP_NUM(7)                              \
+  OP_NUM(8)                              \
+  OP_NUM(9)                              \
+  OP_NUM(10)                             \
+  OP_NUM(11)                             \
+  OP_NUM(12)                             \
+  OP_NUM(13)                             \
+  OP_NUM(14)                             \
+  OP_NUM(15)                             \
+  OP_NUM(16)                             \
+  OP_NUM(17)                             \
+  OP_NUM(18)                             \
+  OP_NUM(19)                             \
+  OP_NUM(20)                             \
+  OP_NUM(21)                             \
+  OP_NUM(22)                             \
+  OP_NUM(23)                             \
+  OP_NUM(24)                             \
+  OP_NUM(25)                             \
+  OP_NUM(26)                             \
+  OP_NUM(27)                             \
+  OP_NUM(28)                             \
+  OP_NUM(29)                             \
+  OP_REG(lr)                             \
+  OP_REG(sp)
+
+#else
+#error "Unsupported architecture for fuzzing on Fuchsia"
+#endif
+
+// Produces a CFI directive for the named or numbered register.
+#define CFI_OFFSET_REG(reg) ".cfi_offset " #reg ", %c[" #reg "]\n"
+#define CFI_OFFSET_NUM(num) CFI_OFFSET_REG(r##num)
+
+// Produces an assembler input operand for the named or numbered register.
+#define ASM_OPERAND_REG(reg) \
+  [reg] "i"(offsetof(zx_thread_state_general_regs_t, reg)),
+#define ASM_OPERAND_NUM(num)                                 \
+  [r##num] "i"(offsetof(zx_thread_state_general_regs_t, r) + \
+               (num * sizeof(uint64_t))),
+
+// Trampoline to bridge from the assembly below to the static C++ crash
+// callback.
+__attribute__((noreturn))
+extern "C" void StaticCrashHandler() {
+  Fuzzer::StaticCrashSignalCallback();
+  for (;;) {
+    _Exit(1);
   }
-  if (Packet.key != kFuzzingCrash) {
-    Printf("libFuzzer: invalid crash key: %" PRIx64 "; aborting!\n",
-           Packet.key);
-    exit(1);
-  }
-  // CrashCallback should not return from this call
-  Fuzzer::StaticCrashSignalCallback();
+
+// Creates the trampoline with the necessary CFI information to unwind through
+// to the crashing call stack.
+__attribute__((used))
+void MakeTrampoline() {
+  uint64_t dummy = 0;  // ignored; used to swallow the trailing comma.
+  __asm__(".cfi_endproc\n"
+    ".pushsection .text.CrashTrampolineAsm\n"
+    ".type CrashTrampolineAsm,STT_FUNC;\n"
+"CrashTrampolineAsm:\n"
+    ".cfi_startproc\n"
+    ".cfi_signal_frame\n"
+#if defined(__x86_64__)
+    ".cfi_return_column rip\n"
+    ".cfi_def_cfa rsp, 0\n"
+    FOREACH_REGISTER(CFI_OFFSET_REG, CFI_OFFSET_NUM)
+    "call StaticCrashHandler\n"
+    "ud2\n"
+#elif defined(__aarch64__)
+    ".cfi_return_column 33\n"
+    ".cfi_def_cfa sp, 0\n"
+    ".cfi_offset 33, %c[pc]\n"
+    FOREACH_REGISTER(CFI_OFFSET_REG, CFI_OFFSET_NUM)
+    "bl StaticCrashHandler\n"
+#else
+#error "Unsupported architecture for fuzzing on Fuchsia"
+#endif
+    ".cfi_endproc\n"
+    ".size CrashTrampolineAsm, . - CrashTrampolineAsm\n"
+    ".popsection\n"
+    ".cfi_startproc\n"
+    : // No outputs
+    : FOREACH_REGISTER(ASM_OPERAND_REG, ASM_OPERAND_NUM)
+#if defined(__aarch64__)
+      ASM_OPERAND_REG(pc)
+#endif
+      "i" (dummy));
+}
+
+void CrashHandler(zx_handle_t *Event) {
+  // This structure is used to ensure we close handles to objects we create in
+  // this handler.
+  struct ScopedCrashContext {
+    ScopedCrashContext() : Port(ZX_HANDLE_INVALID), Thread(ZX_HANDLE_INVALID) {}
+    ~ScopedCrashContext() {
+      _zx_handle_close(Port);
+      _zx_handle_close(Thread);
+    }
+    zx_handle_t Port;
+    zx_handle_t Thread;
+  } Scoped;
+
+  // Create and bind the exception port.  We need to claim to be a "debugger" so
+  // the kernel will allow us to modify and resume dying threads (see below).
+  // Once the port is set, we can signal the main thread to continue and wait
+  // for the exception to arrive.
+  ExitOnErr(_zx_port_create(0, &Scoped.Port), "_zx_port_create");
+  zx_handle_t Self = _zx_process_self();
+
+  ExitOnErr(_zx_task_bind_exception_port(Self, Scoped.Port, kFuzzingCrash,
+                                         ZX_EXCEPTION_PORT_DEBUGGER),
+            "_zx_task_bind_exception_port");
+
+  ExitOnErr(_zx_object_signal(*Event, 0, ZX_USER_SIGNAL_0),
+            "_zx_object_signal");
+
+  zx_port_packet_t Packet;
+  ExitOnErr(_zx_port_wait(Scoped.Port, ZX_TIME_INFINITE, &Packet),
+            "_zx_port_wait");
+
+  // At this point, we want to get the state of the crashing thread, but
+  // libFuzzer and the sanitizers assume this will happen from that same thread
+  // via a POSIX signal handler. "Resurrecting" the thread in the middle of the
+  // appropriate callback is as simple as forcibly setting the instruction
+  // pointer/program counter, provided we NEVER EVER return from that function
+  // (since otherwise our stack will not be valid).
+  ExitOnErr(_zx_object_get_child(Self, Packet.exception.tid,
+                                 ZX_RIGHT_SAME_RIGHTS, &Scoped.Thread),
+            "_zx_object_get_child");
+
+  zx_thread_state_general_regs_t GeneralRegisters;
+  ExitOnErr(_zx_thread_read_state(Scoped.Thread, ZX_THREAD_STATE_GENERAL_REGS,
+                                  &GeneralRegisters, sizeof(GeneralRegisters)),
+            "_zx_thread_read_state");
+
+  // To unwind properly, we need to push the crashing thread's register state
+  // onto the stack and jump into a trampoline with CFI instructions on how
+  // to restore it.
+#if defined(__x86_64__)
+  uintptr_t StackPtr =
+      (GeneralRegisters.rsp - (128 + sizeof(GeneralRegisters))) &
+      -(uintptr_t)16;
+  memcpy(reinterpret_cast<void *>(StackPtr), &GeneralRegisters,
+         sizeof(GeneralRegisters));
+  GeneralRegisters.rsp = StackPtr;
+  GeneralRegisters.rip = reinterpret_cast<zx_vaddr_t>(CrashTrampolineAsm);
+
+#elif defined(__aarch64__)
+  uintptr_t StackPtr = GeneralRegisters.sp - sizeof(GeneralRegisters);
+  memcpy(reinterpret_cast<void *>(StackPtr), &GeneralRegisters,
+         sizeof(GeneralRegisters));
+  GeneralRegisters.sp = StackPtr;
+  GeneralRegisters.pc = reinterpret_cast<zx_vaddr_t>(CrashTrampolineAsm);
+
+#else
+#error "Unsupported architecture for fuzzing on Fuchsia"
+#endif
+
+  // Now force the crashing thread's state.
+  ExitOnErr(_zx_thread_write_state(Scoped.Thread, ZX_THREAD_STATE_GENERAL_REGS,
+                                   &GeneralRegisters, sizeof(GeneralRegisters)),
+            "_zx_thread_write_state");
+
+  ExitOnErr(_zx_task_resume_from_exception(Scoped.Thread, Scoped.Port, 0),
+            "_zx_task_resume_from_exception");
 }
 
 } // namespace
 
 // Platform specific functions.
 void SetSignalHandler(const FuzzingOptions &Options) {
-  zx_status_t rc;
-
   // Set up alarm handler if needed.
   if (Options.UnitTimeoutSec > 0) {
     std::thread T(AlarmHandler, Options.UnitTimeoutSec / 2 + 1);
     T.detach();
   }
 
   // Set up interrupt handler if needed.
   if (Options.HandleInt || Options.HandleTerm) {
     std::thread T(InterruptHandler);
     T.detach();
   }
 
   // Early exit if no crash handler needed.
   if (!Options.HandleSegv && !Options.HandleBus && !Options.HandleIll &&
       !Options.HandleFpe && !Options.HandleAbrt)
     return;
 
-  // Create an exception port
-  zx_handle_t *ExceptionPort = new zx_handle_t;
-  if ((rc = _zx_port_create(0, ExceptionPort)) != ZX_OK) {
-    Printf("libFuzzer: zx_port_create failed: %s\n", _zx_status_get_string(rc));
-    exit(1);
-  }
-
-  // Bind the port to receive exceptions from our process
-  if ((rc = _zx_task_bind_exception_port(_zx_process_self(), *ExceptionPort,
-                                         kFuzzingCrash, 0)) != ZX_OK) {
-    Printf("libFuzzer: unable to bind exception port: %s\n",
-           _zx_status_get_string(rc));
-    exit(1);
-  }
+  // Set up the crash handler and wait until its ready before proceeding.
+  zx_handle_t Event;
+  ExitOnErr(_zx_event_create(0, &Event), "_zx_event_create");
+
+  std::thread T(CrashHandler, &Event);
+  zx_status_t Status =
+      _zx_object_wait_one(Event, ZX_USER_SIGNAL_0, ZX_TIME_INFINITE, nullptr);
+  _zx_handle_close(Event);
+  ExitOnErr(Status, "_zx_object_wait_one");
 
-  // Set up the crash handler.
-  std::thread T(CrashHandler, ExceptionPort);
   T.detach();
 }
 
 void SleepSeconds(int Seconds) {
   _zx_nanosleep(_zx_deadline_after(ZX_SEC(Seconds)));
 }
 
 unsigned long GetPid() {
   zx_status_t rc;
   zx_info_handle_basic_t Info;
   if ((rc = _zx_object_get_info(_zx_process_self(), ZX_INFO_HANDLE_BASIC, &Info,
                                 sizeof(Info), NULL, NULL)) != ZX_OK) {
     Printf("libFuzzer: unable to get info about self: %s\n",
            _zx_status_get_string(rc));
     exit(1);
   }
   return Info.koid;
 }
 
 size_t GetPeakRSSMb() {
   zx_status_t rc;
   zx_info_task_stats_t Info;
   if ((rc = _zx_object_get_info(_zx_process_self(), ZX_INFO_TASK_STATS, &Info,
                                 sizeof(Info), NULL, NULL)) != ZX_OK) {
     Printf("libFuzzer: unable to get info about self: %s\n",
            _zx_status_get_string(rc));
     exit(1);
   }
   return (Info.mem_private_bytes + Info.mem_shared_bytes) >> 20;
 }
 
 template <typename Fn>