fileutils.c

#include "Python.h"
#include "osdefs.h"
#include <locale.h>

#ifdef MS_WINDOWS
#  include <malloc.h>
#  include <windows.h>
extern int winerror_to_errno(int);
#endif

#ifdef HAVE_LANGINFO_H
#include <langinfo.h>
#endif

#ifdef HAVE_SYS_IOCTL_H
#include <sys/ioctl.h>
#endif

#ifdef HAVE_FCNTL_H
#include <fcntl.h>
#endif /* HAVE_FCNTL_H */

#ifdef __APPLE__
extern wchar_t* _Py_DecodeUTF8_surrogateescape(const char *s, Py_ssize_t size);
#endif

#ifdef O_CLOEXEC
/* Does open() support the O_CLOEXEC flag? Possible values:

   -1: unknown
    0: open() ignores O_CLOEXEC flag, ex: Linux kernel older than 2.6.23
    1: open() supports O_CLOEXEC flag, close-on-exec is set

   The flag is used by _Py_open(), _Py_open_noraise(), io.FileIO
   and os.open(). */
int _Py_open_cloexec_works = -1;
#endif

PyObject *
_Py_device_encoding(int fd)
{
#if defined(MS_WINDOWS)
    UINT cp;
#endif
    int valid;
    _Py_BEGIN_SUPPRESS_IPH
    valid = _PyVerify_fd(fd) && isatty(fd);
    _Py_END_SUPPRESS_IPH
    if (!valid)
        Py_RETURN_NONE;

#if defined(MS_WINDOWS)
    if (fd == 0)
        cp = GetConsoleCP();
    else if (fd == 1 || fd == 2)
        cp = GetConsoleOutputCP();
    else
        cp = 0;
    /* GetConsoleCP() and GetConsoleOutputCP() return 0 if the application
       has no console */
    if (cp != 0)
        return PyUnicode_FromFormat("cp%u", (unsigned int)cp);
#elif defined(CODESET)
    {
        char *codeset = nl_langinfo(CODESET);
        if (codeset != NULL && codeset[0] != 0)
            return PyUnicode_FromString(codeset);
    }
#endif
    Py_RETURN_NONE;
}

#if !defined(__APPLE__) && !defined(MS_WINDOWS)
extern int _Py_normalize_encoding(const char *, char *, size_t);

/* Workaround FreeBSD and OpenIndiana locale encoding issue with the C locale.
   On these operating systems, nl_langinfo(CODESET) announces an alias of the
   ASCII encoding, whereas mbstowcs() and wcstombs() functions use the
   ISO-8859-1 encoding. The problem is that os.fsencode() and os.fsdecode() use
   locale.getpreferredencoding() codec. For example, if command line arguments
   are decoded by mbstowcs() and encoded back by os.fsencode(), we get a
   UnicodeEncodeError instead of retrieving the original byte string.

   The workaround is enabled if setlocale(LC_CTYPE, NULL) returns "C",
   nl_langinfo(CODESET) announces "ascii" (or an alias to ASCII), and at least
   one byte in range 0x80-0xff can be decoded from the locale encoding. The
   workaround is also enabled on error, for example if getting the locale
   failed.

   Values of force_ascii:

       1: the workaround is used: Py_EncodeLocale() uses
          encode_ascii_surrogateescape() and Py_DecodeLocale() uses
          decode_ascii_surrogateescape()
       0: the workaround is not used: Py_EncodeLocale() uses wcstombs() and
          Py_DecodeLocale() uses mbstowcs()
      -1: unknown, need to call check_force_ascii() to get the value
*/
static int force_ascii = -1;

static int
check_force_ascii(void)
{
    char *loc;
#if defined(HAVE_LANGINFO_H) && defined(CODESET)
    char *codeset, **alias;
    char encoding[100];
    int is_ascii;
    unsigned int i;
    char* ascii_aliases[] = {
        "ascii",
        "646",
        "ansi-x3.4-1968",
        "ansi-x3-4-1968",
        "ansi-x3.4-1986",
        "cp367",
        "csascii",
        "ibm367",
        "iso646-us",
        "iso-646.irv-1991",
        "iso-ir-6",
        "us",
        "us-ascii",
        NULL
    };
#endif

    loc = setlocale(LC_CTYPE, NULL);
    if (loc == NULL)
        goto error;
    if (strcmp(loc, "C") != 0) {
        /* the LC_CTYPE locale is different than C */
        return 0;
    }

#if defined(HAVE_LANGINFO_H) && defined(CODESET)
    codeset = nl_langinfo(CODESET);
    if (!codeset || codeset[0] == '\0') {
        /* CODESET is not set or empty */
        goto error;
    }
    if (!_Py_normalize_encoding(codeset, encoding, sizeof(encoding)))
        goto error;

    is_ascii = 0;
    for (alias=ascii_aliases; *alias != NULL; alias++) {
        if (strcmp(encoding, *alias) == 0) {
            is_ascii = 1;
            break;
        }
    }
    if (!is_ascii) {
        /* nl_langinfo(CODESET) is not "ascii" or an alias of ASCII */
        return 0;
    }

    for (i=0x80; i<0xff; i++) {
        unsigned char ch;
        wchar_t wch;
        size_t res;

        ch = (unsigned char)i;
        res = mbstowcs(&wch, (char*)&ch, 1);
        if (res != (size_t)-1) {
            /* decoding a non-ASCII character from the locale encoding succeed:
               the locale encoding is not ASCII, force ASCII */
            return 1;
        }
    }
    /* None of the bytes in the range 0x80-0xff can be decoded from the locale
       encoding: the locale encoding is really ASCII */
    return 0;
#else
    /* nl_langinfo(CODESET) is not available: always force ASCII */
    return 1;
#endif

error:
    /* if an error occurred, force the ASCII encoding */
    return 1;
}

static char*
encode_ascii_surrogateescape(const wchar_t *text, size_t *error_pos)
{
    char *result = NULL, *out;
    size_t len, i;
    wchar_t ch;

    if (error_pos != NULL)
        *error_pos = (size_t)-1;

    len = wcslen(text);

    result = PyMem_Malloc(len + 1);  /* +1 for NUL byte */
    if (result == NULL)
        return NULL;

    out = result;
    for (i=0; i<len; i++) {
        ch = text[i];

        if (ch <= 0x7f) {
            /* ASCII character */
            *out++ = (char)ch;
        }
        else if (0xdc80 <= ch && ch <= 0xdcff) {
            /* UTF-8b surrogate */
            *out++ = (char)(ch - 0xdc00);
        }
        else {
            if (error_pos != NULL)
                *error_pos = i;
            PyMem_Free(result);
            return NULL;
        }
    }
    *out = '\0';
    return result;
}
#endif   /* !defined(__APPLE__) && !defined(MS_WINDOWS) */

#if !defined(__APPLE__) && (!defined(MS_WINDOWS) || !defined(HAVE_MBRTOWC))
static wchar_t*
decode_ascii_surrogateescape(const char *arg, size_t *size)
{
    wchar_t *res;
    unsigned char *in;
    wchar_t *out;
    size_t argsize = strlen(arg) + 1;

    if (argsize > PY_SSIZE_T_MAX/sizeof(wchar_t))
        return NULL;
    res = PyMem_RawMalloc(argsize*sizeof(wchar_t));
    if (!res)
        return NULL;

    in = (unsigned char*)arg;
    out = res;
    while(*in)
        if(*in < 128)
            *out++ = *in++;
        else
            *out++ = 0xdc00 + *in++;
    *out = 0;
    if (size != NULL)
        *size = out - res;
    return res;
}
#endif


/* Decode a byte string from the locale encoding with the
   surrogateescape error handler: undecodable bytes are decoded as characters
   in range U+DC80..U+DCFF. If a byte sequence can be decoded as a surrogate
   character, escape the bytes using the surrogateescape error handler instead
   of decoding them.

   Return a pointer to a newly allocated wide character string, use
   PyMem_RawFree() to free the memory. If size is not NULL, write the number of
   wide characters excluding the null character into *size

   Return NULL on decoding error or memory allocation error. If *size* is not
   NULL, *size is set to (size_t)-1 on memory error or set to (size_t)-2 on
   decoding error.

   Decoding errors should never happen, unless there is a bug in the C
   library.

   Use the Py_EncodeLocale() function to encode the character string back to a
   byte string. */
wchar_t*
Py_DecodeLocale(const char* arg, size_t *size)
{
#ifdef __APPLE__
    wchar_t *wstr;
    wstr = _Py_DecodeUTF8_surrogateescape(arg, strlen(arg));
    if (size != NULL) {
        if (wstr != NULL)
            *size = wcslen(wstr);
        else
            *size = (size_t)-1;
    }
    return wstr;
#else
    wchar_t *res;
    size_t argsize;
    size_t count;
#ifdef HAVE_MBRTOWC
    unsigned char *in;
    wchar_t *out;
    mbstate_t mbs;
#endif

#ifndef MS_WINDOWS
    if (force_ascii == -1)
        force_ascii = check_force_ascii();

    if (force_ascii) {
        /* force ASCII encoding to workaround mbstowcs() issue */
        res = decode_ascii_surrogateescape(arg, size);
        if (res == NULL)
            goto oom;
        return res;
    }
#endif

#ifdef HAVE_BROKEN_MBSTOWCS
    /* Some platforms have a broken implementation of
     * mbstowcs which does not count the characters that
     * would result from conversion.  Use an upper bound.
     */
    argsize = strlen(arg);
#else
    argsize = mbstowcs(NULL, arg, 0);
#endif
    if (argsize != (size_t)-1) {
        if (argsize == PY_SSIZE_T_MAX)
            goto oom;
        argsize += 1;
        if (argsize > PY_SSIZE_T_MAX/sizeof(wchar_t))
            goto oom;
        res = (wchar_t *)PyMem_RawMalloc(argsize*sizeof(wchar_t));
        if (!res)
            goto oom;
        count = mbstowcs(res, arg, argsize);
        if (count != (size_t)-1) {
            wchar_t *tmp;
            /* Only use the result if it contains no
               surrogate characters. */
            for (tmp = res; *tmp != 0 &&
                         !Py_UNICODE_IS_SURROGATE(*tmp); tmp++)
                ;
            if (*tmp == 0) {
                if (size != NULL)
                    *size = count;
                return res;
            }
        }
        PyMem_RawFree(res);
    }
    /* Conversion failed. Fall back to escaping with surrogateescape. */
#ifdef HAVE_MBRTOWC
    /* Try conversion with mbrtwoc (C99), and escape non-decodable bytes. */

    /* Overallocate; as multi-byte characters are in the argument, the
       actual output could use less memory. */
    argsize = strlen(arg) + 1;
    if (argsize > PY_SSIZE_T_MAX/sizeof(wchar_t))
        goto oom;
    res = (wchar_t*)PyMem_RawMalloc(argsize*sizeof(wchar_t));
    if (!res)
        goto oom;
    in = (unsigned char*)arg;
    out = res;
    memset(&mbs, 0, sizeof mbs);
    while (argsize) {
        size_t converted = mbrtowc(out, (char*)in, argsize, &mbs);
        if (converted == 0)
            /* Reached end of string; null char stored. */
            break;
        if (converted == (size_t)-2) {
            /* Incomplete character. This should never happen,
               since we provide everything that we have -
               unless there is a bug in the C library, or I
               misunderstood how mbrtowc works. */
            PyMem_RawFree(res);
            if (size != NULL)
                *size = (size_t)-2;
            return NULL;
        }
        if (converted == (size_t)-1) {
            /* Conversion error. Escape as UTF-8b, and start over
               in the initial shift state. */
            *out++ = 0xdc00 + *in++;
            argsize--;
            memset(&mbs, 0, sizeof mbs);
            continue;
        }
        if (Py_UNICODE_IS_SURROGATE(*out)) {
            /* Surrogate character.  Escape the original
               byte sequence with surrogateescape. */
            argsize -= converted;
            while (converted--)
                *out++ = 0xdc00 + *in++;
            continue;
        }
        /* successfully converted some bytes */
        in += converted;
        argsize -= converted;
        out++;
    }
    if (size != NULL)
        *size = out - res;
#else   /* HAVE_MBRTOWC */
    /* Cannot use C locale for escaping; manually escape as if charset
       is ASCII (i.e. escape all bytes > 128. This will still roundtrip
       correctly in the locale's charset, which must be an ASCII superset. */
    res = decode_ascii_surrogateescape(arg, size);
    if (res == NULL)
        goto oom;
#endif   /* HAVE_MBRTOWC */
    return res;
oom:
    if (size != NULL)
        *size = (size_t)-1;
    return NULL;
#endif   /* __APPLE__ */
}

/* Encode a wide character string to the locale encoding with the
   surrogateescape error handler: surrogate characters in the range
   U+DC80..U+DCFF are converted to bytes 0x80..0xFF.

   Return a pointer to a newly allocated byte string, use PyMem_Free() to free
   the memory. Return NULL on encoding or memory allocation error.

   If error_pos is not NULL, *error_pos is set to the index of the invalid
   character on encoding error, or set to (size_t)-1 otherwise.

   Use the Py_DecodeLocale() function to decode the bytes string back to a wide
   character string. */
char*
Py_EncodeLocale(const wchar_t *text, size_t *error_pos)
{
#ifdef __APPLE__
    Py_ssize_t len;
    PyObject *unicode, *bytes = NULL;
    char *cpath;

    unicode = PyUnicode_FromWideChar(text, wcslen(text));
    if (unicode == NULL)
        return NULL;

    bytes = _PyUnicode_AsUTF8String(unicode, "surrogateescape");
    Py_DECREF(unicode);
    if (bytes == NULL) {
        PyErr_Clear();
        if (error_pos != NULL)
            *error_pos = (size_t)-1;
        return NULL;
    }

    len = PyBytes_GET_SIZE(bytes);
    cpath = PyMem_Malloc(len+1);
    if (cpath == NULL) {
        PyErr_Clear();
        Py_DECREF(bytes);
        if (error_pos != NULL)
            *error_pos = (size_t)-1;
        return NULL;
    }
    memcpy(cpath, PyBytes_AsString(bytes), len + 1);
    Py_DECREF(bytes);
    return cpath;
#else   /* __APPLE__ */
    const size_t len = wcslen(text);
    char *result = NULL, *bytes = NULL;
    size_t i, size, converted;
    wchar_t c, buf[2];

#ifndef MS_WINDOWS
    if (force_ascii == -1)
        force_ascii = check_force_ascii();

    if (force_ascii)
        return encode_ascii_surrogateescape(text, error_pos);
#endif

    /* The function works in two steps:
       1. compute the length of the output buffer in bytes (size)
       2. outputs the bytes */
    size = 0;
    buf[1] = 0;
    while (1) {
        for (i=0; i < len; i++) {
            c = text[i];
            if (c >= 0xdc80 && c <= 0xdcff) {
                /* UTF-8b surrogate */
                if (bytes != NULL) {
                    *bytes++ = c - 0xdc00;
                    size--;
                }
                else
                    size++;
                continue;
            }
            else {
                buf[0] = c;
                if (bytes != NULL)
                    converted = wcstombs(bytes, buf, size);
                else
                    converted = wcstombs(NULL, buf, 0);
                if (converted == (size_t)-1) {
                    if (result != NULL)
                        PyMem_Free(result);
                    if (error_pos != NULL)
                        *error_pos = i;
                    return NULL;
                }
                if (bytes != NULL) {
                    bytes += converted;
                    size -= converted;
                }
                else
                    size += converted;
            }
        }
        if (result != NULL) {
            *bytes = '\0';
            break;
        }

        size += 1; /* nul byte at the end */
        result = PyMem_Malloc(size);
        if (result == NULL) {
            if (error_pos != NULL)
                *error_pos = (size_t)-1;
            return NULL;
        }
        bytes = result;
    }
    return result;
#endif   /* __APPLE__ */
}


#ifdef MS_WINDOWS
static __int64 secs_between_epochs = 11644473600; /* Seconds between 1.1.1601 and 1.1.1970 */

static void
FILE_TIME_to_time_t_nsec(FILETIME *in_ptr, time_t *time_out, int* nsec_out)
{
    /* XXX endianness. Shouldn't matter, as all Windows implementations are little-endian */
    /* Cannot simply cast and dereference in_ptr,
       since it might not be aligned properly */
    __int64 in;
    memcpy(&in, in_ptr, sizeof(in));
    *nsec_out = (int)(in % 10000000) * 100; /* FILETIME is in units of 100 nsec. */
    *time_out = Py_SAFE_DOWNCAST((in / 10000000) - secs_between_epochs, __int64, time_t);
}

void
_Py_time_t_to_FILE_TIME(time_t time_in, int nsec_in, FILETIME *out_ptr)
{
    /* XXX endianness */
    __int64 out;
    out = time_in + secs_between_epochs;
    out = out * 10000000 + nsec_in / 100;
    memcpy(out_ptr, &out, sizeof(out));
}

/* Below, we *know* that ugo+r is 0444 */
#if _S_IREAD != 0400
#error Unsupported C library
#endif
static int
attributes_to_mode(DWORD attr)
{
    int m = 0;
    if (attr & FILE_ATTRIBUTE_DIRECTORY)
        m |= _S_IFDIR | 0111; /* IFEXEC for user,group,other */
    else
        m |= _S_IFREG;
    if (attr & FILE_ATTRIBUTE_READONLY)
        m |= 0444;
    else
        m |= 0666;
    return m;
}

void
_Py_attribute_data_to_stat(BY_HANDLE_FILE_INFORMATION *info, ULONG reparse_tag,
                           struct _Py_stat_struct *result)
{
    memset(result, 0, sizeof(*result));
    result->st_mode = attributes_to_mode(info->dwFileAttributes);
    result->st_size = (((__int64)info->nFileSizeHigh)<<32) + info->nFileSizeLow;
    result->st_dev = info->dwVolumeSerialNumber;
    result->st_rdev = result->st_dev;
    FILE_TIME_to_time_t_nsec(&info->ftCreationTime, &result->st_ctime, &result->st_ctime_nsec);
    FILE_TIME_to_time_t_nsec(&info->ftLastWriteTime, &result->st_mtime, &result->st_mtime_nsec);
    FILE_TIME_to_time_t_nsec(&info->ftLastAccessTime, &result->st_atime, &result->st_atime_nsec);
    result->st_nlink = info->nNumberOfLinks;
    result->st_ino = (((__int64)info->nFileIndexHigh)<<32) + info->nFileIndexLow;
    if (reparse_tag == IO_REPARSE_TAG_SYMLINK) {
        /* first clear the S_IFMT bits */
        result->st_mode ^= (result->st_mode & S_IFMT);
        /* now set the bits that make this a symlink */
        result->st_mode |= S_IFLNK;
    }
    result->st_file_attributes = info->dwFileAttributes;
}
#endif

/* Return information about a file.

   On POSIX, use fstat().

   On Windows, use GetFileType() and GetFileInformationByHandle() which support
   files larger than 2 GB.  fstat() may fail with EOVERFLOW on files larger
   than 2 GB because the file size type is a signed 32-bit integer: see issue
   #23152.

   On Windows, set the last Windows error and return nonzero on error. On
   POSIX, set errno and return nonzero on error. Fill status and return 0 on
   success. */
int
_Py_fstat_noraise(int fd, struct _Py_stat_struct *status)
{
#ifdef MS_WINDOWS
    BY_HANDLE_FILE_INFORMATION info;
    HANDLE h;
    int type;

    if (!_PyVerify_fd(fd))
        h = INVALID_HANDLE_VALUE;
    else {
        _Py_BEGIN_SUPPRESS_IPH
        h = (HANDLE)_get_osfhandle(fd);
        _Py_END_SUPPRESS_IPH
    }

    if (h == INVALID_HANDLE_VALUE) {
        /* errno is already set by _get_osfhandle, but we also set
           the Win32 error for callers who expect that */
        SetLastError(ERROR_INVALID_HANDLE);
        return -1;
    }
    memset(status, 0, sizeof(*status));

    type = GetFileType(h);
    if (type == FILE_TYPE_UNKNOWN) {
        DWORD error = GetLastError();
        if (error != 0) {
            errno = winerror_to_errno(error);
            return -1;
        }
        /* else: valid but unknown file */
    }

    if (type != FILE_TYPE_DISK) {
        if (type == FILE_TYPE_CHAR)
            status->st_mode = _S_IFCHR;
        else if (type == FILE_TYPE_PIPE)
            status->st_mode = _S_IFIFO;
        return 0;
    }

    if (!GetFileInformationByHandle(h, &info)) {
        /* The Win32 error is already set, but we also set errno for
           callers who expect it */
        errno = winerror_to_errno(GetLastError());
        return -1;
    }

    _Py_attribute_data_to_stat(&info, 0, status);
    /* specific to fstat() */
    status->st_ino = (((__int64)info.nFileIndexHigh)<<32) + info.nFileIndexLow;
    return 0;
#else
    return fstat(fd, status);
#endif
}

/* Return information about a file.

   On POSIX, use fstat().

   On Windows, use GetFileType() and GetFileInformationByHandle() which support
   files larger than 2 GB.  fstat() may fail with EOVERFLOW on files larger
   than 2 GB because the file size type is a signed 32-bit integer: see issue
   #23152.

   Raise an exception and return -1 on error. On Windows, set the last Windows
   error on error. On POSIX, set errno on error. Fill status and return 0 on
   success.

   Release the GIL to call GetFileType() and GetFileInformationByHandle(), or
   to call fstat(). The caller must hold the GIL. */
int
_Py_fstat(int fd, struct _Py_stat_struct *status)
{
    int res;

    Py_BEGIN_ALLOW_THREADS
    res = _Py_fstat_noraise(fd, status);
    Py_END_ALLOW_THREADS

    if (res != 0) {
#ifdef MS_WINDOWS
        PyErr_SetFromWindowsErr(0);
#else
        PyErr_SetFromErrno(PyExc_OSError);
#endif
        return -1;
    }
    return 0;
}

/* Call _wstat() on Windows, or encode the path to the filesystem encoding and
   call stat() otherwise. Only fill st_mode attribute on Windows.

   Return 0 on success, -1 on _wstat() / stat() error, -2 if an exception was
   raised. */

int
_Py_stat(PyObject *path, struct stat *statbuf)
{
#ifdef MS_WINDOWS
    int err;
    struct _stat wstatbuf;
    const wchar_t *wpath;

    wpath = _PyUnicode_AsUnicode(path);
    if (wpath == NULL)
        return -2;

    err = _wstat(wpath, &wstatbuf);
    if (!err)
        statbuf->st_mode = wstatbuf.st_mode;
    return err;
#else
    int ret;
    PyObject *bytes;
    char *cpath;

    bytes = PyUnicode_EncodeFSDefault(path);
    if (bytes == NULL)
        return -2;

    /* check for embedded null bytes */
    if (PyBytes_AsStringAndSize(bytes, &cpath, NULL) == -1) {
        Py_DECREF(bytes);
        return -2;
    }

    ret = stat(cpath, statbuf);
    Py_DECREF(bytes);
    return ret;
#endif
}


static int
get_inheritable(int fd, int raise)
{
#ifdef MS_WINDOWS
    HANDLE handle;
    DWORD flags;

    if (!_PyVerify_fd(fd)) {
        if (raise)
            PyErr_SetFromErrno(PyExc_OSError);
        return -1;
    }

    _Py_BEGIN_SUPPRESS_IPH
    handle = (HANDLE)_get_osfhandle(fd);
    _Py_END_SUPPRESS_IPH
    if (handle == INVALID_HANDLE_VALUE) {
        if (raise)
            PyErr_SetFromErrno(PyExc_OSError);
        return -1;
    }

    if (!GetHandleInformation(handle, &flags)) {
        if (raise)
            PyErr_SetFromWindowsErr(0);
        return -1;
    }

    return (flags & HANDLE_FLAG_INHERIT);
#else
    int flags;

    flags = fcntl(fd, F_GETFD, 0);
    if (flags == -1) {
        if (raise)
            PyErr_SetFromErrno(PyExc_OSError);
        return -1;
    }
    return !(flags & FD_CLOEXEC);
#endif
}

/* Get the inheritable flag of the specified file descriptor.
   Return 1 if the file descriptor can be inherited, 0 if it cannot,
   raise an exception and return -1 on error. */
int
_Py_get_inheritable(int fd)
{
    return get_inheritable(fd, 1);
}

static int
set_inheritable(int fd, int inheritable, int raise, int *atomic_flag_works)
{
#ifdef MS_WINDOWS
    HANDLE handle;
    DWORD flags;
#else
#if defined(HAVE_SYS_IOCTL_H) && defined(FIOCLEX) && defined(FIONCLEX)
    static int ioctl_works = -1;
    int request;
    int err;
#endif
    int flags;
    int res;
#endif

    /* atomic_flag_works can only be used to make the file descriptor
       non-inheritable */
    assert(!(atomic_flag_works != NULL && inheritable));

    if (atomic_flag_works != NULL && !inheritable) {
        if (*atomic_flag_works == -1) {
            int isInheritable = get_inheritable(fd, raise);
            if (isInheritable == -1)
                return -1;
            *atomic_flag_works = !isInheritable;
        }

        if (*atomic_flag_works)
            return 0;
    }

#ifdef MS_WINDOWS
    if (!_PyVerify_fd(fd)) {
        if (raise)
            PyErr_SetFromErrno(PyExc_OSError);
        return -1;
    }

    _Py_BEGIN_SUPPRESS_IPH
    handle = (HANDLE)_get_osfhandle(fd);
    _Py_END_SUPPRESS_IPH
    if (handle == INVALID_HANDLE_VALUE) {
        if (raise)
            PyErr_SetFromErrno(PyExc_OSError);
        return -1;
    }

    if (inheritable)
        flags = HANDLE_FLAG_INHERIT;
    else
        flags = 0;
    if (!SetHandleInformation(handle, HANDLE_FLAG_INHERIT, flags)) {
        if (raise)
            PyErr_SetFromWindowsErr(0);
        return -1;
    }
    return 0;

#else

#if defined(HAVE_SYS_IOCTL_H) && defined(FIOCLEX) && defined(FIONCLEX)
    if (ioctl_works != 0) {
        /* fast-path: ioctl() only requires one syscall */
        if (inheritable)
            request = FIONCLEX;
        else
            request = FIOCLEX;
        err = ioctl(fd, request, NULL);
        if (!err) {
            ioctl_works = 1;
            return 0;
        }

        if (errno != ENOTTY && errno != EACCES) {
            if (raise)
                PyErr_SetFromErrno(PyExc_OSError);
            return -1;
        }
        else {
            /* Issue #22258: Here, ENOTTY means "Inappropriate ioctl for
               device". The ioctl is declared but not supported by the kernel.
               Remember that ioctl() doesn't work. It is the case on
               Illumos-based OS for example.

               Issue #27057: When SELinux policy disallows ioctl it will fail
               with EACCES. While FIOCLEX is safe operation it may be
               unavailable because ioctl was denied altogether.
               This can be the case on Android. */
            ioctl_works = 0;
        }
        /* fallback to fcntl() if ioctl() does not work */
    }
#endif

    /* slow-path: fcntl() requires two syscalls */
    flags = fcntl(fd, F_GETFD);
    if (flags < 0) {
        if (raise)
            PyErr_SetFromErrno(PyExc_OSError);
        return -1;
    }

    if (inheritable)
        flags &= ~FD_CLOEXEC;
    else
        flags |= FD_CLOEXEC;
    res = fcntl(fd, F_SETFD, flags);
    if (res < 0) {
        if (raise)
            PyErr_SetFromErrno(PyExc_OSError);
        return -1;
    }
    return 0;
#endif
}

/* Make the file descriptor non-inheritable.
   Return 0 on success, set errno and return -1 on error. */
static int
make_non_inheritable(int fd)
{
    return set_inheritable(fd, 0, 0, NULL);
}

/* Set the inheritable flag of the specified file descriptor.
   On success: return 0, on error: raise an exception if raise is nonzero
   and return -1.

   If atomic_flag_works is not NULL:

    * if *atomic_flag_works==-1, check if the inheritable is set on the file
      descriptor: if yes, set *atomic_flag_works to 1, otherwise set to 0 and
      set the inheritable flag
    * if *atomic_flag_works==1: do nothing
    * if *atomic_flag_works==0: set inheritable flag to False

   Set atomic_flag_works to NULL if no atomic flag was used to create the
   file descriptor.

   atomic_flag_works can only be used to make a file descriptor
   non-inheritable: atomic_flag_works must be NULL if inheritable=1. */
int
_Py_set_inheritable(int fd, int inheritable, int *atomic_flag_works)
{
    return set_inheritable(fd, inheritable, 1, atomic_flag_works);
}

static int
_Py_open_impl(const char *pathname, int flags, int gil_held)
{
    int fd;
    int async_err = 0;
#ifndef MS_WINDOWS
    int *atomic_flag_works;
#endif

#ifdef MS_WINDOWS
    flags |= O_NOINHERIT;
#elif defined(O_CLOEXEC)
    atomic_flag_works = &_Py_open_cloexec_works;
    flags |= O_CLOEXEC;
#else
    atomic_flag_works = NULL;
#endif

    if (gil_held) {
        do {
            Py_BEGIN_ALLOW_THREADS
            fd = open(pathname, flags);
            Py_END_ALLOW_THREADS
        } while (fd < 0
                 && errno == EINTR && !(async_err = PyErr_CheckSignals()));
        if (async_err)
            return -1;
        if (fd < 0) {
            PyErr_SetFromErrnoWithFilename(PyExc_OSError, pathname);
            return -1;
        }
    }
    else {
        fd = open(pathname, flags);
        if (fd < 0)
            return -1;
    }

#ifndef MS_WINDOWS
    if (set_inheritable(fd, 0, gil_held, atomic_flag_works) < 0) {
        close(fd);
        return -1;
    }
#endif

    return fd;
}

/* Open a file with the specified flags (wrapper to open() function).
   Return a file descriptor on success. Raise an exception and return -1 on
   error.

   The file descriptor is created non-inheritable.

   When interrupted by a signal (open() fails with EINTR), retry the syscall,
   except if the Python signal handler raises an exception.