faulthandler.c

#include "Python.h"
#include "pycore_ceval.h"         // _PyEval_IsGILEnabled()
#include "pycore_initconfig.h"    // _PyStatus_ERR()
#include "pycore_pyerrors.h"      // _Py_DumpExtensionModules()
#include "pycore_fileutils.h"     // _PyFile_Flush
#include "pycore_pystate.h"       // _PyThreadState_GET()
#include "pycore_runtime.h"       // _Py_ID()
#include "pycore_signal.h"        // Py_NSIG
#include "pycore_time.h"          // _PyTime_FromSecondsObject()
#include "pycore_traceback.h"     // _Py_DumpStack()
#ifdef HAVE_UNISTD_H
#  include <unistd.h>             // _exit()
#endif

#include <signal.h>               // sigaction()
#include <stdlib.h>               // abort()
#if defined(HAVE_PTHREAD_SIGMASK) && !defined(HAVE_BROKEN_PTHREAD_SIGMASK) && defined(HAVE_PTHREAD_H)
#  include <pthread.h>
#endif
#ifdef MS_WINDOWS
#  include <windows.h>
#endif
#ifdef HAVE_SYS_RESOURCE_H
#  include <sys/resource.h>       // setrlimit()
#endif

#if defined(FAULTHANDLER_USE_ALT_STACK) && defined(HAVE_LINUX_AUXVEC_H) && defined(HAVE_SYS_AUXV_H)
#  include <linux/auxvec.h>       // AT_MINSIGSTKSZ
#  include <sys/auxv.h>           // getauxval()
#endif


#include "clinic/faulthandler.c.h"


/* Sentinel to ignore all_threads on free-threading */
#define FT_IGNORE_ALL_THREADS 2

/* Allocate at maximum 100 MiB of the stack to raise the stack overflow */
#define STACK_OVERFLOW_MAX_SIZE (100 * 1024 * 1024)

#define PUTS(fd, str) (void)_Py_write_noraise(fd, str, strlen(str))


/*[clinic input]
module faulthandler
[clinic start generated code]*/
/*[clinic end generated code: output=da39a3ee5e6b4b0d input=c3d4f47c4f3d440f]*/


typedef struct {
    int signum;
    int enabled;
    const char* name;
    _Py_sighandler_t previous;
    int all_threads;
} fault_handler_t;

#define fatal_error _PyRuntime.faulthandler.fatal_error
#define thread _PyRuntime.faulthandler.thread

#ifdef FAULTHANDLER_USER
#define user_signals _PyRuntime.faulthandler.user_signals
typedef struct faulthandler_user_signal user_signal_t;
static void faulthandler_user(int signum);
#endif /* FAULTHANDLER_USER */


static fault_handler_t faulthandler_handlers[] = {
#ifdef SIGBUS
    {SIGBUS, 0, "Bus error", },
#endif
#ifdef SIGILL
    {SIGILL, 0, "Illegal instruction", },
#endif
    {SIGFPE, 0, "Floating-point exception", },
    {SIGABRT, 0, "Aborted", },
    /* define SIGSEGV at the end to make it the default choice if searching the
       handler fails in faulthandler_fatal_error() */
    {SIGSEGV, 0, "Segmentation fault", }
};
static const size_t faulthandler_nsignals = \
    Py_ARRAY_LENGTH(faulthandler_handlers);

#ifdef FAULTHANDLER_USE_ALT_STACK
#  define stack _PyRuntime.faulthandler.stack
#  define old_stack _PyRuntime.faulthandler.old_stack
#endif


/* Get the file descriptor of a file by calling its fileno() method and then
   call its flush() method.

   If file is NULL or Py_None, use sys.stderr as the new file.
   If file is an integer, it will be treated as file descriptor.

   On success, return the file descriptor and write the new file into *file_ptr.
   On error, return -1. */

static int
faulthandler_get_fileno(PyObject **file_ptr)
{
    PyObject *result;
    PyObject *file = *file_ptr;

    if (file == NULL || file == Py_None) {
        file = PySys_GetAttr(&_Py_ID(stderr));
        if (file == NULL) {
            return -1;
        }
        if (file == Py_None) {
            PyErr_SetString(PyExc_RuntimeError, "sys.stderr is None");
            Py_DECREF(file);
            return -1;
        }
    }
    else if (PyLong_Check(file)) {
        if (PyBool_Check(file)) {
            if (PyErr_WarnEx(PyExc_RuntimeWarning,
                    "bool is used as a file descriptor", 1))
            {
                return -1;
            }
        }
        int fd = PyLong_AsInt(file);
        if (fd == -1 && PyErr_Occurred())
            return -1;
        if (fd < 0) {
            PyErr_SetString(PyExc_ValueError,
                            "file is not a valid file descriptor");
            return -1;
        }
        *file_ptr = NULL;
        return fd;
    }
    else {
        Py_INCREF(file);
    }

    result = PyObject_CallMethodNoArgs(file, &_Py_ID(fileno));
    if (result == NULL) {
        Py_DECREF(file);
        return -1;
    }

    int fd;
    if (PyLong_Check(result)) {
        fd = PyLong_AsInt(result);
    }
    else {
        fd = -1;
    }
    Py_DECREF(result);

    if (fd < 0) {
        PyErr_SetString(PyExc_RuntimeError,
                        "file.fileno() is not a valid file descriptor");
        Py_DECREF(file);
        return -1;
    }

    if (_PyFile_Flush(file) < 0) {
        /* ignore flush() error */
        PyErr_Clear();
    }
    *file_ptr = file;
    return fd;
}

/* Get the state of the current thread: only call this function if the current
   thread holds the GIL. Raise an exception on error. */
static PyThreadState*
get_thread_state(void)
{
    PyThreadState *tstate = _PyThreadState_GET();
    if (tstate == NULL) {
        /* just in case but very unlikely... */
        PyErr_SetString(PyExc_RuntimeError,
                        "unable to get the current thread state");
        return NULL;
    }
    return tstate;
}

static void
faulthandler_dump_traceback(int fd, int all_threads,
                            PyInterpreterState *interp,
                            Py_ssize_t max_threads)
{
    static volatile int reentrant = 0;

    if (reentrant)
        return;

    reentrant = 1;

    /* SIGSEGV, SIGFPE, SIGABRT, SIGBUS and SIGILL are synchronous signals and
       are thus delivered to the thread that caused the fault. Get the Python
       thread state of the current thread.

       PyThreadState_Get() doesn't give the state of the thread that caused the
       fault if the thread released the GIL, and so this function cannot be
       used. Read the thread specific storage (TSS) instead: call
       PyGILState_GetThisThreadState(). */
    PyThreadState *tstate = PyGILState_GetThisThreadState();

    if (all_threads == 1) {
        (void)PyUnstable_DumpTracebackThreads(fd, NULL, tstate, max_threads);
    }
    else {
        if (all_threads == FT_IGNORE_ALL_THREADS) {
            PUTS(fd, "<Cannot show all threads while the GIL is disabled>\n");
        }
        if (tstate != NULL) {
            PyUnstable_DumpTraceback(fd, tstate);
        }
    }

    reentrant = 0;
}

static void
faulthandler_dump_c_stack(int fd)
{
    static volatile int reentrant = 0;

    if (reentrant) {
        return;
    }

    reentrant = 1;

    if (fatal_error.c_stack) {
        PUTS(fd, "\n");
        _Py_DumpStack(fd);
    }

    reentrant = 0;
}


/*[clinic input]
faulthandler.dump_traceback as faulthandler_dump_traceback_py

    file: object(py_default="sys.stderr") = NULL
    all_threads: bool = True
    *
    max_threads: Py_ssize_t = 100

Dump the traceback of the current thread into file.

Dump the traceback of all threads if all_threads is true. max_threads
caps the number of threads dumped.
[clinic start generated code]*/

static PyObject *
faulthandler_dump_traceback_py_impl(PyObject *module, PyObject *file,
                                    int all_threads, Py_ssize_t max_threads)
/*[clinic end generated code: output=ee1bbc2668e56e77 input=38630eb40e641de6]*/
{
    PyThreadState *tstate;
    const char *errmsg;
    int fd;

    fd = faulthandler_get_fileno(&file);
    if (fd < 0)
        return NULL;

    tstate = get_thread_state();
    if (tstate == NULL) {
        Py_XDECREF(file);
        return NULL;
    }

    if (all_threads) {
        PyInterpreterState *interp = _PyInterpreterState_GET();
        /* gh-128400: Accessing other thread states while they're running
         * isn't safe if those threads are running. */
        _PyEval_StopTheWorld(interp);
        errmsg = PyUnstable_DumpTracebackThreads(fd, NULL, tstate, max_threads);
        _PyEval_StartTheWorld(interp);
    }
    else {
        errmsg = PyUnstable_DumpTraceback(fd, tstate);
    }
    if (errmsg != NULL) {
        PyErr_SetString(PyExc_RuntimeError, errmsg);
        Py_XDECREF(file);
        return NULL;
    }

    Py_XDECREF(file);

    if (PyErr_CheckSignals())
        return NULL;

    Py_RETURN_NONE;
}


/*[clinic input]
faulthandler.dump_c_stack as faulthandler_dump_c_stack_py

    file: object(py_default="sys.stderr") = NULL

Dump the C stack of the current thread.
[clinic start generated code]*/

static PyObject *
faulthandler_dump_c_stack_py_impl(PyObject *module, PyObject *file)
/*[clinic end generated code: output=151d6c95e9f8c0f6 input=10f6b6f29b635109]*/
{
    int fd = faulthandler_get_fileno(&file);
    if (fd < 0) {
        return NULL;
    }

    _Py_DumpStack(fd);
    Py_XDECREF(file);

    if (PyErr_CheckSignals()) {
        return NULL;
    }

    Py_RETURN_NONE;
}

static void
faulthandler_disable_fatal_handler(fault_handler_t *handler)
{
    if (!handler->enabled)
        return;
    handler->enabled = 0;
#ifdef HAVE_SIGACTION
    (void)sigaction(handler->signum, &handler->previous, NULL);
#else
    (void)signal(handler->signum, handler->previous);
#endif
}

static int
deduce_all_threads(void)
{
#ifndef Py_GIL_DISABLED
    return fatal_error.all_threads;
#else
    if (fatal_error.all_threads == 0) {
        return 0;
    }
    // We can't use _PyThreadState_GET, so use the stored GILstate one
    PyThreadState *tstate = PyGILState_GetThisThreadState();
    if (tstate == NULL) {
        return 0;
    }

    /* In theory, it's safe to dump all threads if the GIL is enabled */
    return _PyEval_IsGILEnabled(tstate)
        ? fatal_error.all_threads
        : FT_IGNORE_ALL_THREADS;
#endif
}

/* Handler for SIGSEGV, SIGFPE, SIGABRT, SIGBUS and SIGILL signals.

   Display the current Python traceback, restore the previous handler and call
   the previous handler.

   On Windows, don't explicitly call the previous handler, because the Windows
   signal handler would not be called (for an unknown reason). The execution of
   the program continues at faulthandler_fatal_error() exit, but the same
   instruction will raise the same fault (signal), and so the previous handler
   will be called.

   This function is signal-safe and should only call signal-safe functions. */

static void
faulthandler_fatal_error(int signum)
{
    const int fd = fatal_error.fd;
    size_t i;
    fault_handler_t *handler = NULL;
    int save_errno = errno;
    int found = 0;

    if (!fatal_error.enabled)
        return;

    for (i=0; i < faulthandler_nsignals; i++) {
        handler = &faulthandler_handlers[i];
        if (handler->signum == signum) {
            found = 1;
            break;
        }
    }
    if (handler == NULL) {
        /* faulthandler_nsignals == 0 (unlikely) */
        return;
    }

    /* restore the previous handler */
    faulthandler_disable_fatal_handler(handler);

    if (found) {
        PUTS(fd, "Fatal Python error: ");
        PUTS(fd, handler->name);
        PUTS(fd, "\n\n");
    }
    else {
        PUTS(fd, "Fatal Python error from unexpected signum: ");
        _Py_DumpDecimal(fd, signum);
        PUTS(fd, "\n\n");
    }

    faulthandler_dump_traceback(fd, deduce_all_threads(),
                                fatal_error.interp,
                                fatal_error.max_threads);
    faulthandler_dump_c_stack(fd);

    _Py_DumpExtensionModules(fd, fatal_error.interp);

    errno = save_errno;
#ifdef MS_WINDOWS
    if (signum == SIGSEGV) {
        /* don't explicitly call the previous handler for SIGSEGV in this signal
           handler, because the Windows signal handler would not be called */
        return;
    }
#endif
    /* call the previous signal handler: it is called immediately if we use
       sigaction() thanks to SA_NODEFER flag, otherwise it is deferred */
    raise(signum);
}

#ifdef MS_WINDOWS
static int
faulthandler_ignore_exception(DWORD code)
{
    /* bpo-30557: ignore exceptions which are not errors */
    if (!(code & 0x80000000)) {
        return 1;
    }
    /* bpo-31701: ignore MSC and COM exceptions
       E0000000 + code */
    if (code == 0xE06D7363 /* MSC exception ("Emsc") */
        || code == 0xE0434352 /* COM Callable Runtime exception ("ECCR") */) {
        return 1;
    }
    /* Interesting exception: log it with the Python traceback */
    return 0;
}

static LONG WINAPI
faulthandler_exc_handler(struct _EXCEPTION_POINTERS *exc_info)
{
    const int fd = fatal_error.fd;
    DWORD code = exc_info->ExceptionRecord->ExceptionCode;

    if (faulthandler_ignore_exception(code)) {
        /* ignore the exception: call the next exception handler */
        return EXCEPTION_CONTINUE_SEARCH;
    }

    PUTS(fd, "Windows fatal exception: ");
    switch (code)
    {
    /* only format most common errors */
    case EXCEPTION_ACCESS_VIOLATION: PUTS(fd, "access violation"); break;
    case EXCEPTION_FLT_DIVIDE_BY_ZERO: PUTS(fd, "float divide by zero"); break;
    case EXCEPTION_FLT_OVERFLOW: PUTS(fd, "float overflow"); break;
    case EXCEPTION_INT_DIVIDE_BY_ZERO: PUTS(fd, "int divide by zero"); break;
    case EXCEPTION_INT_OVERFLOW: PUTS(fd, "integer overflow"); break;
    case EXCEPTION_IN_PAGE_ERROR: PUTS(fd, "page error"); break;
    case EXCEPTION_STACK_OVERFLOW: PUTS(fd, "stack overflow"); break;
    default:
        PUTS(fd, "code 0x");
        _Py_DumpHexadecimal(fd, code, 8);
    }
    PUTS(fd, "\n\n");

    if (code == EXCEPTION_ACCESS_VIOLATION) {
        /* disable signal handler for SIGSEGV */
        for (size_t i=0; i < faulthandler_nsignals; i++) {
            fault_handler_t *handler = &faulthandler_handlers[i];
            if (handler->signum == SIGSEGV) {
                faulthandler_disable_fatal_handler(handler);
                break;
            }
        }
    }

    faulthandler_dump_traceback(fd, deduce_all_threads(),
                                fatal_error.interp,
                                fatal_error.max_threads);
    faulthandler_dump_c_stack(fd);

    /* call the next exception handler */
    return EXCEPTION_CONTINUE_SEARCH;
}
#endif


#ifdef FAULTHANDLER_USE_ALT_STACK
static int
faulthandler_allocate_stack(void)
{
    if (stack.ss_sp != NULL) {
        return 0;
    }
    /* Allocate an alternate stack for faulthandler() signal handler
       to be able to execute a signal handler on a stack overflow error */
    stack.ss_sp = PyMem_Malloc(stack.ss_size);
    if (stack.ss_sp == NULL) {
        PyErr_NoMemory();
        return -1;
    }

    int err = sigaltstack(&stack, &old_stack);
    if (err) {
        PyErr_SetFromErrno(PyExc_OSError);
        /* Release the stack to retry sigaltstack() next time */
        PyMem_Free(stack.ss_sp);
        stack.ss_sp = NULL;
        return -1;
    }
    return 0;
}
#endif


/* Install the handler for fatal signals, faulthandler_fatal_error(). */

static int
faulthandler_enable(void)
{
    if (fatal_error.enabled) {
        return 0;
    }
    fatal_error.enabled = 1;

#ifdef FAULTHANDLER_USE_ALT_STACK
    if (faulthandler_allocate_stack() < 0) {
        return -1;
    }
#endif

    // gh-137185: Initialize C stack trace dumping outside of the signal
    // handler. Specifically, we call backtrace() to ensure that libgcc is
    // dynamically loaded outside of the signal handler.
    _Py_InitDumpStack();

    for (size_t i=0; i < faulthandler_nsignals; i++) {
        fault_handler_t *handler;
        int err;

        handler = &faulthandler_handlers[i];
        assert(!handler->enabled);
#ifdef HAVE_SIGACTION
        struct sigaction action;
        action.sa_handler = faulthandler_fatal_error;
        sigemptyset(&action.sa_mask);
        /* Do not prevent the signal from being received from within
           its own signal handler */
        action.sa_flags = SA_NODEFER;
#ifdef FAULTHANDLER_USE_ALT_STACK
        assert(stack.ss_sp != NULL);
        /* Call the signal handler on an alternate signal stack
           provided by sigaltstack() */
        action.sa_flags |= SA_ONSTACK;
#endif
        err = sigaction(handler->signum, &action, &handler->previous);
#else
        handler->previous = signal(handler->signum,
                                   faulthandler_fatal_error);
        err = (handler->previous == SIG_ERR);
#endif
        if (err) {
            PyErr_SetFromErrno(PyExc_RuntimeError);
            return -1;
        }

        handler->enabled = 1;
    }

#ifdef MS_WINDOWS_DESKTOP
    assert(fatal_error.exc_handler == NULL);
    fatal_error.exc_handler = AddVectoredExceptionHandler(1, faulthandler_exc_handler);
#endif
    return 0;
}


/*[clinic input]
faulthandler.enable as faulthandler_py_enable

    file: object(py_default="sys.stderr") = NULL
    all_threads: bool = True
    c_stack: bool = True
    *
    max_threads: Py_ssize_t = 100

Enable the fault handler.
[clinic start generated code]*/

static PyObject *
faulthandler_py_enable_impl(PyObject *module, PyObject *file,
                            int all_threads, int c_stack,
                            Py_ssize_t max_threads)
/*[clinic end generated code: output=7ee655332317c47a input=e64759714f27b466]*/
{
    int fd;
    PyThreadState *tstate;

    fd = faulthandler_get_fileno(&file);
    if (fd < 0)
        return NULL;

    tstate = get_thread_state();
    if (tstate == NULL) {
        Py_XDECREF(file);
        return NULL;
    }

    Py_XSETREF(fatal_error.file, file);
    fatal_error.fd = fd;
    fatal_error.all_threads = all_threads;
    fatal_error.interp = PyThreadState_GetInterpreter(tstate);
    fatal_error.c_stack = c_stack;
    fatal_error.max_threads = max_threads;

    if (faulthandler_enable() < 0) {
        return NULL;
    }

    Py_RETURN_NONE;
}

static void
faulthandler_disable(void)
{
    if (fatal_error.enabled) {
        fatal_error.enabled = 0;
        for (size_t i=0; i < faulthandler_nsignals; i++) {
            fault_handler_t *handler;
            handler = &faulthandler_handlers[i];
            faulthandler_disable_fatal_handler(handler);
        }
    }
#ifdef MS_WINDOWS_DESKTOP
    if (fatal_error.exc_handler != NULL) {
        RemoveVectoredExceptionHandler(fatal_error.exc_handler);
        fatal_error.exc_handler = NULL;
    }
#endif
    Py_CLEAR(fatal_error.file);
}


/*[clinic input]
faulthandler.disable as faulthandler_disable_py

Disable the fault handler.
[clinic start generated code]*/

static PyObject *
faulthandler_disable_py_impl(PyObject *module)
/*[clinic end generated code: output=e9087a04535af3cb input=6223eac6804550af]*/
{
    if (!fatal_error.enabled) {
        Py_RETURN_FALSE;
    }
    faulthandler_disable();
    Py_RETURN_TRUE;
}


/*[clinic input]
faulthandler.is_enabled -> bool

Check if the handler is enabled.
[clinic start generated code]*/

static int
faulthandler_is_enabled_impl(PyObject *module)
/*[clinic end generated code: output=b9f33a3e0f881a23 input=3d5532547eb14bf9]*/
{
    return fatal_error.enabled;
}

static void
faulthandler_thread(void *unused)
{
    PyLockStatus st;
    const char* errmsg;
    int ok;
#if defined(HAVE_PTHREAD_SIGMASK) && !defined(HAVE_BROKEN_PTHREAD_SIGMASK)
    sigset_t set;

    /* we don't want to receive any signal */
    sigfillset(&set);
    pthread_sigmask(SIG_SETMASK, &set, NULL);
#endif

    do {
        st = PyThread_acquire_lock_timed(thread.cancel_event,
                                         thread.timeout_us, 0);
        if (st == PY_LOCK_ACQUIRED) {
            PyThread_release_lock(thread.cancel_event);
            break;
        }
        /* Timeout => dump traceback */
        assert(st == PY_LOCK_FAILURE);

        (void)_Py_write_noraise(thread.fd, thread.header, thread.header_len);

        errmsg = PyUnstable_DumpTracebackThreads(thread.fd, thread.interp, NULL,
                                                 thread.max_threads);
        ok = (errmsg == NULL);

        if (thread.exit)
            _exit(1);
    } while (ok && thread.repeat);

    /* The only way out */
    PyThread_release_lock(thread.running);
}

static void
cancel_dump_traceback_later(void)
{
    /* If not scheduled, nothing to cancel */
    if (!thread.cancel_event) {
        return;
    }

    /* Notify cancellation */
    PyThread_release_lock(thread.cancel_event);

    /* Wait for thread to join */
    PyThread_acquire_lock(thread.running, 1);
    PyThread_release_lock(thread.running);

    /* The main thread should always hold the cancel_event lock */
    PyThread_acquire_lock(thread.cancel_event, 1);

    Py_CLEAR(thread.file);
    if (thread.header) {
        PyMem_Free(thread.header);
        thread.header = NULL;
    }
}

#define SEC_TO_US (1000 * 1000)

static char*
format_timeout(PyTime_t us)
{
    unsigned long sec, min, hour;
    char buffer[100];

    /* the downcast is safe: the caller check that 0 < us <= LONG_MAX */
    sec = (unsigned long)(us / SEC_TO_US);
    us %= SEC_TO_US;

    min = sec / 60;
    sec %= 60;
    hour = min / 60;
    min %= 60;

    if (us != 0) {
        PyOS_snprintf(buffer, sizeof(buffer),
                      "Timeout (%lu:%02lu:%02lu.%06u)!\n",
                      hour, min, sec, (unsigned int)us);
    }
    else {
        PyOS_snprintf(buffer, sizeof(buffer),
                      "Timeout (%lu:%02lu:%02lu)!\n",
                      hour, min, sec);
    }
    return _PyMem_Strdup(buffer);
}


/*[clinic input]
faulthandler.dump_traceback_later

    timeout as timeout_obj: object
    repeat: bool = False
    file: object(py_default="sys.stderr") = NULL
    exit: bool = False
    *
    max_threads: Py_ssize_t = 100

Dump the traceback of all threads in timeout seconds.

If repeat is true, the tracebacks of all threads are dumped every
timeout seconds.  If exit is true, call _exit(1) which is not safe.
max_threads caps the number of threads dumped.
[clinic start generated code]*/

static PyObject *
faulthandler_dump_traceback_later_impl(PyObject *module,
                                       PyObject *timeout_obj, int repeat,
                                       PyObject *file, int exit,
                                       Py_ssize_t max_threads)
/*[clinic end generated code: output=543a0f3807113394 input=32aaf7437d0928db]*/
{
    PyTime_t timeout, timeout_us;
    int fd;
    PyThreadState *tstate;
    char *header;
    size_t header_len;

    if (_PyTime_FromSecondsObject(&timeout, timeout_obj,
                                  _PyTime_ROUND_TIMEOUT) < 0) {
        return NULL;
    }
    timeout_us = _PyTime_AsMicroseconds(timeout, _PyTime_ROUND_TIMEOUT);
    if (timeout_us <= 0) {
        PyErr_SetString(PyExc_ValueError, "timeout must be greater than 0");
        return NULL;
    }
    /* Limit to LONG_MAX seconds for format_timeout() */
    if (timeout_us > PY_TIMEOUT_MAX || timeout_us / SEC_TO_US > LONG_MAX) {
        PyErr_SetString(PyExc_OverflowError,
                        "timeout value is too large");
        return NULL;
    }

    tstate = get_thread_state();
    if (tstate == NULL) {
        return NULL;
    }

    fd = faulthandler_get_fileno(&file);
    if (fd < 0) {
        return NULL;
    }

    if (!thread.running) {
        thread.running = PyThread_allocate_lock();
        if (!thread.running) {
            Py_XDECREF(file);
            return PyErr_NoMemory();
        }
    }
    if (!thread.cancel_event) {
        thread.cancel_event = PyThread_allocate_lock();
        if (!thread.cancel_event || !thread.running) {
            Py_XDECREF(file);
            return PyErr_NoMemory();
        }

        /* cancel_event starts to be acquired: it's only released to cancel
           the thread. */
        PyThread_acquire_lock(thread.cancel_event, 1);
    }

    /* format the timeout */
    header = format_timeout(timeout_us);
    if (header == NULL) {
        Py_XDECREF(file);
        return PyErr_NoMemory();
    }
    header_len = strlen(header);

    /* Cancel previous thread, if running */
    cancel_dump_traceback_later();

    Py_XSETREF(thread.file, file);
    thread.fd = fd;
    /* the downcast is safe: we check that 0 < timeout_us < PY_TIMEOUT_MAX */
    thread.timeout_us = (PY_TIMEOUT_T)timeout_us;
    thread.repeat = repeat;
    thread.interp = PyThreadState_GetInterpreter(tstate);
    thread.exit = exit;
    thread.header = header;
    thread.header_len = header_len;
    thread.max_threads = max_threads;

    /* Arm these locks to serve as events when released */
    PyThread_acquire_lock(thread.running, 1);

    if (PyThread_start_new_thread(faulthandler_thread, NULL) == PYTHREAD_INVALID_THREAD_ID) {
        PyThread_release_lock(thread.running);
        Py_CLEAR(thread.file);
        PyMem_Free(header);
        thread.header = NULL;
        PyErr_SetString(PyExc_RuntimeError,
                        "unable to start watchdog thread");
        return NULL;
    }

    Py_RETURN_NONE;
}


/*[clinic input]
faulthandler.cancel_dump_traceback_later as faulthandler_cancel_dump_traceback_later_py

Cancel the previous call to dump_traceback_later().
[clinic start generated code]*/

static PyObject *
faulthandler_cancel_dump_traceback_later_py_impl(PyObject *module)
/*[clinic end generated code: output=2cf303015d39c926 input=51ad64b6ca8412a4]*/
{
    cancel_dump_traceback_later();
    Py_RETURN_NONE;
}


#ifdef FAULTHANDLER_USER
static int
faulthandler_register(int signum, int chain, _Py_sighandler_t *previous_p)
{
#ifdef HAVE_SIGACTION
    struct sigaction action;
    action.sa_handler = faulthandler_user;
    sigemptyset(&action.sa_mask);
    /* if the signal is received while the kernel is executing a system
       call, try to restart the system call instead of interrupting it and
       return EINTR. */
    action.sa_flags = SA_RESTART;
    if (chain) {
        /* do not prevent the signal from being received from within its
           own signal handler */
        action.sa_flags = SA_NODEFER;
    }
#ifdef FAULTHANDLER_USE_ALT_STACK
    assert(stack.ss_sp != NULL);
    /* Call the signal handler on an alternate signal stack
       provided by sigaltstack() */
    action.sa_flags |= SA_ONSTACK;
#endif
    return sigaction(signum, &action, previous_p);
#else
    _Py_sighandler_t previous;
    previous = signal(signum, faulthandler_user);
    if (previous_p != NULL) {
        *previous_p = previous;
    }
    return (previous == SIG_ERR);
#endif
}

/* Handler of user signals (e.g. SIGUSR1).

   Dump the traceback of the current thread, or of all threads if
   thread.all_threads is true.

   This function is signal safe and should only call signal safe functions. */

static void
faulthandler_user(int signum)
{
    user_signal_t *user;
    int save_errno = errno;

    user = &user_signals[signum];
    if (!user->enabled)
        return;

    faulthandler_dump_traceback(user->fd, user->all_threads, user->interp,
                                user->max_threads);

#ifdef HAVE_SIGACTION
    if (user->chain) {
        (void)sigaction(signum, &user->previous, NULL);
        errno = save_errno;

        /* call the previous signal handler */
        raise(signum);

        save_errno = errno;
        (void)faulthandler_register(signum, user->chain, NULL);
        errno = save_errno;
    }
#else
    if (user->chain && user->previous != NULL) {
        errno = save_errno;
        /* call the previous signal handler */
        user->previous(signum);
    }
#endif
}

static int
check_signum(int signum)
{
    for (size_t i=0; i < faulthandler_nsignals; i++) {
        if (faulthandler_handlers[i].signum == signum) {
            PyErr_Format(PyExc_RuntimeError,
                         "signal %i cannot be registered, "
                         "use enable() instead",
                         signum);
            return 0;
        }
    }
    if (signum < 1 || Py_NSIG <= signum) {
        PyErr_SetString(PyExc_ValueError, "signal number out of range");
        return 0;
    }
    return 1;
}


/*[clinic input]
faulthandler.register as faulthandler_register_py

    signum: int
    file: object(py_default="sys.stderr") = NULL
    all_threads: bool = True
    chain: bool = False
    *
    max_threads: Py_ssize_t = 100

Register a handler for the signal 'signum'.

Dump the traceback of the current thread, or of all threads if
all_threads is True, into file. max_threads caps the number of threads
dumped.
[clinic start generated code]*/

static PyObject *
faulthandler_register_py_impl(PyObject *module, int signum, PyObject *file,
                              int all_threads, int chain,
                              Py_ssize_t max_threads)
/*[clinic end generated code: output=d63a5b4f388dee5f input=c75096a20de502fe]*/
{
    int fd;
    user_signal_t *user;
    _Py_sighandler_t previous;
    PyThreadState *tstate;
    int err;

    if (!check_signum(signum))
        return NULL;

    tstate = get_thread_state();
    if (tstate == NULL)
        return NULL;

    fd = faulthandler_get_fileno(&file);
    if (fd < 0)
        return NULL;

    if (user_signals == NULL) {
        user_signals = PyMem_Calloc(Py_NSIG, sizeof(user_signal_t));
        if (user_signals == NULL) {
            Py_XDECREF(file);
            return PyErr_NoMemory();
        }
    }
    user = &user_signals[signum];

    if (!user->enabled) {
#ifdef FAULTHANDLER_USE_ALT_STACK
        if (faulthandler_allocate_stack() < 0) {
            Py_XDECREF(file);
            return NULL;
        }
#endif

        err = faulthandler_register(signum, chain, &previous);
        if (err) {
            PyErr_SetFromErrno(PyExc_OSError);
            Py_XDECREF(file);
            return NULL;
        }

        user->previous = previous;
    }

    Py_XSETREF(user->file, file);
    user->fd = fd;
    user->all_threads = all_threads;
    user->chain = chain;
    user->interp = PyThreadState_GetInterpreter(tstate);
    user->max_threads = max_threads;
    user->enabled = 1;

    Py_RETURN_NONE;
}

static int
faulthandler_unregister(user_signal_t *user, int signum)
{
    if (!user->enabled)
        return 0;
    user->enabled = 0;
#ifdef HAVE_SIGACTION
    (void)sigaction(signum, &user->previous, NULL);
#else
    (void)signal(signum, user->previous);
#endif
    Py_CLEAR(user->file);
    user->fd = -1;
    return 1;
}


/*[clinic input]
faulthandler.unregister as faulthandler_unregister_py

     signum: int
     /

Unregister the handler of the signal 'signum' registered by register().
[clinic start generated code]*/

static PyObject *
faulthandler_unregister_py_impl(PyObject *module, int signum)
/*[clinic end generated code: output=01734423da1155ed input=c016de014495d384]*/
{
    user_signal_t *user;
    int change;

    if (!check_signum(signum))
        return NULL;

    if (user_signals == NULL)
        Py_RETURN_FALSE;

    user = &user_signals[signum];
    change = faulthandler_unregister(user, signum);
    return PyBool_FromLong(change);
}
#endif   /* FAULTHANDLER_USER */


static void
faulthandler_suppress_crash_report(void)
{
#ifdef MS_WINDOWS_DESKTOP
    UINT mode;

    /* Configure Windows to not display the Windows Error Reporting dialog */
    mode = SetErrorMode(SEM_NOGPFAULTERRORBOX);
    SetErrorMode(mode | SEM_NOGPFAULTERRORBOX);
#endif

#ifdef HAVE_SYS_RESOURCE_H
    struct rlimit rl;

    /* Disable creation of core dump */
    if (getrlimit(RLIMIT_CORE, &rl) == 0) {
        rl.rlim_cur = 0;
        setrlimit(RLIMIT_CORE, &rl);
    }
#endif

#ifdef _MSC_VER
    /* Visual Studio: configure abort() to not display an error message nor
       open a popup asking to report the fault. */
    _set_abort_behavior(0, _WRITE_ABORT_MSG | _CALL_REPORTFAULT);
#endif
}


static void
faulthandler_raise_sigsegv(void)
{
    faulthandler_suppress_crash_report();
#if defined(MS_WINDOWS)
    /* For SIGSEGV, faulthandler_fatal_error() restores the previous signal
       handler and then gives back the execution flow to the program (without
       explicitly calling the previous error handler). In a normal case, the
       SIGSEGV was raised by the kernel because of a fault, and so if the
       program retries to execute the same instruction, the fault will be
       raised again.

       Here the fault is simulated by a fake SIGSEGV signal raised by the
       application. We have to raise SIGSEGV at lease twice: once for
       faulthandler_fatal_error(), and one more time for the previous signal
       handler. */
    while(1)
        raise(SIGSEGV);
#else
    raise(SIGSEGV);
#endif
}


/*[clinic input]
faulthandler._sigsegv

     release_gil: bool = False
     /

Raise a SIGSEGV signal.
[clinic start generated code]*/

static PyObject *
faulthandler__sigsegv_impl(PyObject *module, int release_gil)
/*[clinic end generated code: output=96e5a2f215b01b76 input=c6ad893cf2ea2b41]*/
{
    if (release_gil) {
        Py_BEGIN_ALLOW_THREADS
        faulthandler_raise_sigsegv();
        Py_END_ALLOW_THREADS
    } else {
        faulthandler_raise_sigsegv();
    }
    Py_RETURN_NONE;
}

static void _Py_NO_RETURN
faulthandler_fatal_error_thread(void *plock)
{
    Py_FatalError("in new thread");
}


/*[clinic input]
faulthandler._fatal_error_c_thread

Call Py_FatalError() in a new C thread.
[clinic start generated code]*/

static PyObject *
faulthandler__fatal_error_c_thread_impl(PyObject *module)
/*[clinic end generated code: output=101bc8aaf4a5eec1 input=fbdca6fffd639a39]*/
{
    unsigned long tid;
    PyThread_type_lock lock;

    faulthandler_suppress_crash_report();

    lock = PyThread_allocate_lock();
    if (lock == NULL)
        return PyErr_NoMemory();

    PyThread_acquire_lock(lock, WAIT_LOCK);

    tid = PyThread_start_new_thread(faulthandler_fatal_error_thread, lock);
    if (tid == PYTHREAD_INVALID_THREAD_ID) {
        PyThread_free_lock(lock);
        PyErr_SetString(PyExc_RuntimeError, "unable to start the thread");
        return NULL;
    }

    /* wait until the thread completes: it will never occur, since Py_FatalError()
       exits the process immediately. */
    PyThread_acquire_lock(lock, WAIT_LOCK);
    PyThread_release_lock(lock);
    PyThread_free_lock(lock);

    Py_RETURN_NONE;
}


/*[clinic input]
faulthandler._sigfpe

Raise a SIGFPE signal.
[clinic start generated code]*/

static PyObject *
faulthandler__sigfpe_impl(PyObject *module)
/*[clinic end generated code: output=dec9c98100e986db input=fd608a92d4421d28]*/
{
    faulthandler_suppress_crash_report();
    raise(SIGFPE);
    Py_UNREACHABLE();
}


/*[clinic input]
faulthandler._sigabrt

Raise a SIGABRT signal.
[clinic start generated code]*/

static PyObject *
faulthandler__sigabrt_impl(PyObject *module)
/*[clinic end generated code: output=58c1378a0c166682 input=be3e0ecefb8676b8]*/
{
    faulthandler_suppress_crash_report();
    abort();
    Py_RETURN_NONE;
}

#if defined(FAULTHANDLER_USE_ALT_STACK)
#define FAULTHANDLER_STACK_OVERFLOW

static uintptr_t
stack_overflow(uintptr_t min_sp, uintptr_t max_sp, size_t *depth)
{
    /* Allocate (at least) 4096 bytes on the stack at each call.

       bpo-23654, bpo-38965: use volatile keyword to prevent tail call
       optimization. */
    volatile unsigned char buffer[4096];
    uintptr_t sp = (uintptr_t)&buffer;
    *depth += 1;
    if (sp < min_sp || max_sp < sp)
        return sp;
    buffer[0] = 1;
    buffer[4095] = 0;
    return stack_overflow(min_sp, max_sp, depth);
}


/*[clinic input]
faulthandler._stack_overflow

Recursive call to raise a stack overflow.
[clinic start generated code]*/

static PyObject *
faulthandler__stack_overflow_impl(PyObject *module)
/*[clinic end generated code: output=efffba4be522d8fb input=4291594a790b6c35]*/
{
    size_t depth, size;
    uintptr_t sp = (uintptr_t)&depth;
    uintptr_t stop, lower_limit, upper_limit;

    faulthandler_suppress_crash_report();
    depth = 0;

    if (STACK_OVERFLOW_MAX_SIZE <= sp) {
        lower_limit = sp - STACK_OVERFLOW_MAX_SIZE;
    }
    else {
        lower_limit = 0;
    }

    if (UINTPTR_MAX - STACK_OVERFLOW_MAX_SIZE >= sp) {
        upper_limit = sp + STACK_OVERFLOW_MAX_SIZE;
    }
    else {
        upper_limit = UINTPTR_MAX;
    }

    stop = stack_overflow(lower_limit, upper_limit, &depth);
    if (sp < stop)
        size = stop - sp;
    else
        size = sp - stop;
    PyErr_Format(PyExc_RuntimeError,
        "unable to raise a stack overflow (allocated %zu bytes "
        "on the stack, %zu recursive calls)",
        size, depth);
    return NULL;
}
#endif   /* defined(FAULTHANDLER_USE_ALT_STACK) && defined(HAVE_SIGACTION) */


#ifdef MS_WINDOWS
/*[clinic input]
faulthandler._raise_exception

    code: unsigned_int
    flags: unsigned_int = 0
    /

Call RaiseException(code, flags).
[clinic start generated code]*/

static PyObject *
faulthandler__raise_exception_impl(PyObject *module, unsigned int code,
                                   unsigned int flags)
/*[clinic end generated code: output=2346cf318eab10dc input=43a5ba0eb7794504]*/
{
    faulthandler_suppress_crash_report();
    RaiseException(code, flags, 0, NULL);
    Py_RETURN_NONE;
}
#endif

PyDoc_STRVAR(module_doc,
"faulthandler module.");

static PyMethodDef module_methods[] = {
    FAULTHANDLER_PY_ENABLE_METHODDEF
    FAULTHANDLER_DISABLE_PY_METHODDEF
    FAULTHANDLER_IS_ENABLED_METHODDEF
    FAULTHANDLER_DUMP_TRACEBACK_PY_METHODDEF
    FAULTHANDLER_DUMP_C_STACK_PY_METHODDEF
    FAULTHANDLER_DUMP_TRACEBACK_LATER_METHODDEF
    FAULTHANDLER_CANCEL_DUMP_TRACEBACK_LATER_PY_METHODDEF
#ifdef FAULTHANDLER_USER
    FAULTHANDLER_REGISTER_PY_METHODDEF
    FAULTHANDLER_UNREGISTER_PY_METHODDEF
#endif
    FAULTHANDLER__SIGSEGV_METHODDEF
    FAULTHANDLER__FATAL_ERROR_C_THREAD_METHODDEF
    FAULTHANDLER__SIGABRT_METHODDEF
    FAULTHANDLER__SIGFPE_METHODDEF
#ifdef FAULTHANDLER_STACK_OVERFLOW
    FAULTHANDLER__STACK_OVERFLOW_METHODDEF
#endif
#ifdef MS_WINDOWS
    FAULTHANDLER__RAISE_EXCEPTION_METHODDEF
#endif
    {NULL, NULL}  /* sentinel */
};

static int
PyExec_faulthandler(PyObject *module) {
    /* Add constants for unit tests */
#ifdef MS_WINDOWS
    /* RaiseException() codes (prefixed by an underscore) */
    if (PyModule_Add(module, "_EXCEPTION_ACCESS_VIOLATION",
                     PyLong_FromUnsignedLong(EXCEPTION_ACCESS_VIOLATION))) {
        return -1;
    }
    if (PyModule_Add(module, "_EXCEPTION_INT_DIVIDE_BY_ZERO",
                     PyLong_FromUnsignedLong(EXCEPTION_INT_DIVIDE_BY_ZERO))) {
        return -1;
    }
    if (PyModule_Add(module, "_EXCEPTION_STACK_OVERFLOW",
                     PyLong_FromUnsignedLong(EXCEPTION_STACK_OVERFLOW))) {
        return -1;
    }

    /* RaiseException() flags (prefixed by an underscore) */
    if (PyModule_Add(module, "_EXCEPTION_NONCONTINUABLE",
                     PyLong_FromUnsignedLong(EXCEPTION_NONCONTINUABLE))) {
        return -1;
    }
    if (PyModule_Add(module, "_EXCEPTION_NONCONTINUABLE_EXCEPTION",
                     PyLong_FromUnsignedLong(EXCEPTION_NONCONTINUABLE_EXCEPTION))) {
        return -1;
    }
#endif
    return 0;
}

static PyModuleDef_Slot faulthandler_slots[] = {
    _Py_ABI_SLOT,
    {Py_mod_exec, PyExec_faulthandler},
    // XXX gh-103092: fix isolation.
    //{Py_mod_multiple_interpreters, Py_MOD_PER_INTERPRETER_GIL_SUPPORTED},
    {Py_mod_gil, Py_MOD_GIL_NOT_USED},
    {0, NULL}
};

static struct PyModuleDef module_def = {
    PyModuleDef_HEAD_INIT,
    .m_name = "faulthandler",
    .m_doc = module_doc,
    .m_methods = module_methods,
    .m_slots = faulthandler_slots
};

PyMODINIT_FUNC
PyInit_faulthandler(void)
{
    return PyModuleDef_Init(&module_def);
}

static int
faulthandler_init_enable(void)
{
    PyObject *enable = PyImport_ImportModuleAttrString("faulthandler", "enable");
    if (enable == NULL) {
        return -1;
    }

    PyObject *res = PyObject_CallNoArgs(enable);
    Py_DECREF(enable);
    if (res == NULL) {
        return -1;
    }
    Py_DECREF(res);

    return 0;
}

PyStatus
_PyFaulthandler_Init(int enable)
{
#ifdef FAULTHANDLER_USE_ALT_STACK
    memset(&stack, 0, sizeof(stack));
    stack.ss_flags = 0;
    /* bpo-21131: allocate dedicated stack of SIGSTKSZ*2 bytes, instead of just
       SIGSTKSZ bytes. Calling the previous signal handler in faulthandler
       signal handler uses more than SIGSTKSZ bytes of stack memory on some
       platforms. */
    stack.ss_size = SIGSTKSZ * 2;
#ifdef AT_MINSIGSTKSZ
    /* bpo-46968: Query Linux for minimal stack size to ensure signal delivery
       for the hardware running CPython. This OS feature is available in
       Linux kernel version >= 5.14 */
    unsigned long at_minstack_size = getauxval(AT_MINSIGSTKSZ);
    if (at_minstack_size != 0) {
        stack.ss_size = SIGSTKSZ + at_minstack_size;
    }
#endif
#endif

    memset(&thread, 0, sizeof(thread));

    if (enable) {
        if (faulthandler_init_enable() < 0) {
            return _PyStatus_ERR("failed to enable faulthandler");
        }
    }
    return _PyStatus_OK();
}

void _PyFaulthandler_Fini(void)
{
    /* later */
    if (thread.cancel_event) {
        cancel_dump_traceback_later();
        PyThread_release_lock(thread.cancel_event);
        PyThread_free_lock(thread.cancel_event);
        thread.cancel_event = NULL;
    }
    if (thread.running) {
        PyThread_free_lock(thread.running);
        thread.running = NULL;
    }

#ifdef FAULTHANDLER_USER
    /* user */
    if (user_signals != NULL) {
        for (size_t signum=0; signum < Py_NSIG; signum++) {
            faulthandler_unregister(&user_signals[signum], signum);
        }
        PyMem_Free(user_signals);
        user_signals = NULL;
    }
#endif

    /* fatal */
    faulthandler_disable();

#ifdef FAULTHANDLER_USE_ALT_STACK
    if (stack.ss_sp != NULL) {
        /* Fetch the current alt stack */
        stack_t current_stack;
        memset(&current_stack, 0, sizeof(current_stack));
        if (sigaltstack(NULL, &current_stack) == 0) {
            if (current_stack.ss_sp == stack.ss_sp) {
                /* The current alt stack is the one that we installed.
                 It is safe to restore the old stack that we found when
                 we installed ours */
                sigaltstack(&old_stack, NULL);
            } else {
                /* Someone switched to a different alt stack and didn't
                   restore ours when they were done (if they're done).
                   There's not much we can do in this unlikely case */
            }
        }
        PyMem_Free(stack.ss_sp);
        stack.ss_sp = NULL;
    }
#endif
}