diff -r 53fa2c9523d4 Lib/test/test_subprocess.py
--- a/Lib/test/test_subprocess.py	Fri May 30 23:46:19 2014 -0700
+++ b/Lib/test/test_subprocess.py	Sun Jun 01 00:16:07 2014 -0700
@@ -1926,6 +1926,59 @@
                          "Some fds not in pass_fds were left open")
         self.assertIn(1, remaining_fds, "Subprocess failed")
 
+
+    def test_close_fds_when_max_fd_is_lowered(self):
+        """Confirm that issue21618 is fixed (mail fail under valgrind)."""
+        fd_status = support.findfile("fd_status.py", subdir="subprocessdata")
+
+        open_fds = set()
+        # Add a bunch more fds to pass down.
+        for _ in range(10):
+            fd = os.open("/dev/null", os.O_RDONLY)
+            open_fds.add(fd)
+
+        # Leave a two pairs of low ones available for use by the
+        # internal child error pipe and the stdout pipe.
+        for fd in sorted(open_fds)[:4]:
+            os.close(fd)
+            open_fds.remove(fd)
+
+        for fd in open_fds:
+            self.addCleanup(os.close, fd)
+            os.set_inheritable(fd, True)
+
+        max_fd_open = max(open_fds)
+
+        import resource
+        rlim_cur, rlim_max = resource.getrlimit(resource.RLIMIT_NOFILE)
+        try:
+            # 9 is lower than the highest fds we are leaving open.
+            resource.setrlimit(resource.RLIMIT_NOFILE, (9, rlim_max))
+            # Launch a new Python interpreter with our low fd rlim_cur that
+            # inherits open fds above that limit.  It then uses subprocess
+            # with close_fds=True to get a report of open fds in the child.
+            # An explicit list of fds to check is passed to fd_status.py as
+            # letting fd_status rely on its default logic would miss the
+            # fds above rlim_cur as it normally only checks up to that limit.
+            p = subprocess.Popen(
+                [sys.executable, '-c',
+                 textwrap.dedent("""
+                     import subprocess, sys
+                     subprocess.Popen([sys.executable, {fd_status!r}] +
+                                      [str(x) for x in range({max_fd})],
+                                      close_fds=True)
+                     """.format(fd_status=fd_status, max_fd=max_fd_open+1))],
+                stdout=subprocess.PIPE, close_fds=False)
+        finally:
+            resource.setrlimit(resource.RLIMIT_NOFILE, (rlim_cur, rlim_max))
+
+        output, unused_stderr = p.communicate()
+        remaining_fds = set(map(int, output.strip().split(b',')))
+
+        self.assertFalse(remaining_fds & open_fds,
+                         msg="Some fds were left open.")
+
+
     # Mac OS X Tiger (10.4) has a kernel bug: sometimes, the file
     # descriptor of a pipe closed in the parent process is valid in the
     # child process according to fstat(), but the mode of the file
diff -r 53fa2c9523d4 Modules/_posixsubprocess.c
--- a/Modules/_posixsubprocess.c	Fri May 30 23:46:19 2014 -0700
+++ b/Modules/_posixsubprocess.c	Sun Jun 01 00:16:07 2014 -0700
@@ -44,10 +44,6 @@
 #define POSIX_CALL(call)   do { if ((call) == -1) goto error; } while (0)
 
 
-/* Maximum file descriptor, initialized on module load. */
-static long max_fd;
-
-
 /* Given the gc module call gc.enable() and return 0 on success. */
 static int
 _enable_gc(PyObject *gc_module)
@@ -166,14 +162,39 @@
 }
 
 
-/* Close all file descriptors in the range start_fd inclusive to
- * end_fd exclusive except for those in py_fds_to_keep.  If the
- * range defined by [start_fd, end_fd) is large this will take a
- * long time as it calls close() on EVERY possible fd.
+/* Get the maximum file descriptor that could be opened by this process.
+ * This function is async signal safe for use between fork() and exec().
+ */
+static long
+safe_get_max_fd(void)
+{
+    long local_max_fd;
+#if defined(__NetBSD__)
+    local_max_fd = fcntl(0, F_MAXFD);
+    if (local_max_fd >= 0)
+        return local_max_fd;
+#endif
+#ifdef _SC_OPEN_MAX
+    local_max_fd = sysconf(_SC_OPEN_MAX);
+    if (local_max_fd == -1)
+#endif
+        local_max_fd = 256;  /* Matches Lib/subprocess.py */
+    return local_max_fd;
+}
+
+
+/* Close all file descriptors in the range from start_fd and higher
+ * except for those in py_fds_to_keep.  If the range defined by
+ * [start_fd, safe_get_max_fd()) is large this will take a long
+ * time as it calls close() on EVERY possible fd.
+ *
+ * It isn't possible to know for sure what the max fd to go up to
+ * is for processes with the capability of raising their maximum.
  */
 static void
-_close_fds_by_brute_force(int start_fd, int end_fd, PyObject *py_fds_to_keep)
+_close_fds_by_brute_force(long start_fd, PyObject *py_fds_to_keep)
 {
+    long end_fd = safe_get_max_fd();
     Py_ssize_t num_fds_to_keep = PySequence_Length(py_fds_to_keep);
     Py_ssize_t keep_seq_idx;
     int fd_num;
@@ -229,16 +250,14 @@
  * it with some cpp #define magic to work on other OSes as well if you want.
  */
 static void
-_close_open_fd_range_safe(int start_fd, int end_fd, PyObject* py_fds_to_keep)
+_close_open_fds_safe(int start_fd, PyObject* py_fds_to_keep)
 {
     int fd_dir_fd;
-    if (start_fd >= end_fd)
-        return;
 
     fd_dir_fd = _Py_open(FD_DIR, O_RDONLY);
     if (fd_dir_fd == -1) {
         /* No way to get a list of open fds. */
-        _close_fds_by_brute_force(start_fd, end_fd, py_fds_to_keep);
+        _close_fds_by_brute_force(start_fd, py_fds_to_keep);
         return;
     } else {
         char buffer[sizeof(struct linux_dirent64)];
@@ -253,23 +272,23 @@
                 entry = (struct linux_dirent64 *)(buffer + offset);
                 if ((fd = _pos_int_from_ascii(entry->d_name)) < 0)
                     continue;  /* Not a number. */
-                if (fd != fd_dir_fd && fd >= start_fd && fd < end_fd &&
+                if (fd != fd_dir_fd && fd >= start_fd &&
                     !_is_fd_in_sorted_fd_sequence(fd, py_fds_to_keep)) {
                     while (close(fd) < 0 && errno == EINTR);
                 }
             }
         }
-        close(fd_dir_fd);
+        while (close(fd_dir_fd) < 0 && errno == EINTR);
     }
 }
 
-#define _close_open_fd_range _close_open_fd_range_safe
+#define _close_open_fds _close_open_fds_safe
 
 #else  /* NOT (defined(__linux__) && defined(HAVE_SYS_SYSCALL_H)) */
 
 
-/* Close all open file descriptors in the range start_fd inclusive to end_fd
- * exclusive. Do not close any in the sorted py_fds_to_keep list.
+/* Close all open file descriptors from start_fd and higher.
+ * Do not close any in the sorted py_fds_to_keep list.
  *
  * This function violates the strict use of async signal safe functions. :(
  * It calls opendir(), readdir() and closedir().  Of these, the one most
@@ -282,17 +301,13 @@
  *   http://womble.decadent.org.uk/readdir_r-advisory.html
  */
 static void
-_close_open_fd_range_maybe_unsafe(int start_fd, int end_fd,
-                                  PyObject* py_fds_to_keep)
+_close_open_fds_maybe_unsafe(long start_fd, PyObject* py_fds_to_keep)
 {
     DIR *proc_fd_dir;
 #ifndef HAVE_DIRFD
-    while (_is_fd_in_sorted_fd_sequence(start_fd, py_fds_to_keep) &&
-           (start_fd < end_fd)) {
+    while (_is_fd_in_sorted_fd_sequence(start_fd, py_fds_to_keep)) {
         ++start_fd;
     }
-    if (start_fd >= end_fd)
-        return;
     /* Close our lowest fd before we call opendir so that it is likely to
      * reuse that fd otherwise we might close opendir's file descriptor in
      * our loop.  This trick assumes that fd's are allocated on a lowest
@@ -300,8 +315,6 @@
     while (close(start_fd) < 0 && errno == EINTR);
     ++start_fd;
 #endif
-    if (start_fd >= end_fd)
-        return;
 
 #if defined(__FreeBSD__)
     if (!_is_fdescfs_mounted_on_dev_fd())
@@ -311,7 +324,7 @@
         proc_fd_dir = opendir(FD_DIR);
     if (!proc_fd_dir) {
         /* No way to get a list of open fds. */
-        _close_fds_by_brute_force(start_fd, end_fd, py_fds_to_keep);
+        _close_fds_by_brute_force(start_fd, py_fds_to_keep);
     } else {
         struct dirent *dir_entry;
 #ifdef HAVE_DIRFD
@@ -324,7 +337,7 @@
             int fd;
             if ((fd = _pos_int_from_ascii(dir_entry->d_name)) < 0)
                 continue;  /* Not a number. */
-            if (fd != fd_used_by_opendir && fd >= start_fd && fd < end_fd &&
+            if (fd != fd_used_by_opendir && fd >= start_fd &&
                 !_is_fd_in_sorted_fd_sequence(fd, py_fds_to_keep)) {
                 while (close(fd) < 0 && errno == EINTR);
             }
@@ -332,13 +345,13 @@
         }
         if (errno) {
             /* readdir error, revert behavior. Highly Unlikely. */
-            _close_fds_by_brute_force(start_fd, end_fd, py_fds_to_keep);
+            _close_fds_by_brute_force(start_fd, py_fds_to_keep);
         }
         closedir(proc_fd_dir);
     }
 }
 
-#define _close_open_fd_range _close_open_fd_range_maybe_unsafe
+#define _close_open_fds _close_open_fds_maybe_unsafe
 
 #endif  /* else NOT (defined(__linux__) && defined(HAVE_SYS_SYSCALL_H)) */
 
@@ -457,14 +470,8 @@
 
     /* close FDs after executing preexec_fn, which might open FDs */
     if (close_fds) {
-        int local_max_fd = max_fd;
-#if defined(__NetBSD__)
-        local_max_fd = fcntl(0, F_MAXFD);
-        if (local_max_fd < 0)
-            local_max_fd = max_fd;
-#endif
         /* TODO HP-UX could use pstat_getproc() if anyone cares about it. */
-        _close_open_fd_range(3, local_max_fd, py_fds_to_keep);
+        _close_open_fds(3, py_fds_to_keep);
     }
 
     /* This loop matches the Lib/os.py _execvpe()'s PATH search when */
@@ -759,11 +766,5 @@
 PyMODINIT_FUNC
 PyInit__posixsubprocess(void)
 {
-#ifdef _SC_OPEN_MAX
-    max_fd = sysconf(_SC_OPEN_MAX);
-    if (max_fd == -1)
-#endif
-        max_fd = 256;  /* Matches Lib/subprocess.py */
-
     return PyModule_Create(&_posixsubprocessmodule);
 }