[Python-checkins] r86992 - in python/branches/release27-maint: Doc/library/io.rst Lib/test/test_io.py Misc/NEWS Modules/_io/bufferedio.c

antoine.pitrou python-checkins at python.org
Fri Dec 3 20:31:52 CET 2010


Author: antoine.pitrou
Date: Fri Dec  3 20:31:52 2010
New Revision: 86992

Log:
Merged revisions 86981,86984 via svnmerge from 
svn+ssh://pythondev@svn.python.org/python/branches/py3k

........
  r86981 | antoine.pitrou | 2010-12-03 19:41:39 +0100 (ven., 03 déc. 2010) | 5 lines
  
  Issue #10478: Reentrant calls inside buffered IO objects (for example by
  way of a signal handler) now raise a RuntimeError instead of freezing the
  current process.
........
  r86984 | antoine.pitrou | 2010-12-03 20:14:17 +0100 (ven., 03 déc. 2010) | 3 lines
  
  Add an "advanced topics" section to the io doc.
........


Modified:
   python/branches/release27-maint/   (props changed)
   python/branches/release27-maint/Doc/library/io.rst
   python/branches/release27-maint/Lib/test/test_io.py
   python/branches/release27-maint/Misc/NEWS
   python/branches/release27-maint/Modules/_io/bufferedio.c

Modified: python/branches/release27-maint/Doc/library/io.rst
==============================================================================
--- python/branches/release27-maint/Doc/library/io.rst	(original)
+++ python/branches/release27-maint/Doc/library/io.rst	Fri Dec  3 20:31:52 2010
@@ -515,8 +515,8 @@
 Buffered Streams
 ----------------
 
-In many situations, buffered I/O streams will provide higher performance
-(bandwidth and latency) than raw I/O streams.  Their API is also more usable.
+Buffered I/O streams provide a higher-level interface to an I/O device
+than raw I/O does.
 
 .. class:: BytesIO([initial_bytes])
 
@@ -776,8 +776,72 @@
       # .getvalue() will now raise an exception.
       output.close()
 
+
 .. class:: IncrementalNewlineDecoder
 
    A helper codec that decodes newlines for universal newlines mode.  It
    inherits :class:`codecs.IncrementalDecoder`.
 
+
+Advanced topics
+---------------
+
+Here we will discuss several advanced topics pertaining to the concrete
+I/O implementations described above.
+
+Performance
+^^^^^^^^^^^
+
+Binary I/O
+""""""""""
+
+By reading and writing only large chunks of data even when the user asks
+for a single byte, buffered I/O is designed to hide any inefficiency in
+calling and executing the operating system's unbuffered I/O routines.  The
+gain will vary very much depending on the OS and the kind of I/O which is
+performed (for example, on some contemporary OSes such as Linux, unbuffered
+disk I/O can be as fast as buffered I/O).  The bottom line, however, is
+that buffered I/O will offer you predictable performance regardless of the
+platform and the backing device.  Therefore, it is most always preferable to
+use buffered I/O rather than unbuffered I/O.
+
+Text I/O
+""""""""
+
+Text I/O over a binary storage (such as a file) is significantly slower than
+binary I/O over the same storage, because it implies conversions from
+unicode to binary data using a character codec.  This can become noticeable
+if you handle huge amounts of text data (for example very large log files).
+
+:class:`StringIO`, however, is a native in-memory unicode container and will
+exhibit similar speed to :class:`BytesIO`.
+
+Multi-threading
+^^^^^^^^^^^^^^^
+
+:class:`FileIO` objects are thread-safe to the extent that the operating
+system calls (such as ``read(2)`` under Unix) they are wrapping are thread-safe
+too.
+
+Binary buffered objects (instances of :class:`BufferedReader`,
+:class:`BufferedWriter`, :class:`BufferedRandom` and :class:`BufferedRWPair`)
+protect their internal structures using a lock; it is therefore safe to call
+them from multiple threads at once.
+
+:class:`TextIOWrapper` objects are not thread-safe.
+
+Reentrancy
+^^^^^^^^^^
+
+Binary buffered objects (instances of :class:`BufferedReader`,
+:class:`BufferedWriter`, :class:`BufferedRandom` and :class:`BufferedRWPair`)
+are not reentrant.  While reentrant calls will not happen in normal situations,
+they can arise if you are doing I/O in a :mod:`signal` handler.  If it is
+attempted to enter a buffered object again while already being accessed
+*from the same thread*, then a :exc:`RuntimeError` is raised.
+
+The above implicitly extends to text files, since the :func:`open()`
+function will wrap a buffered object inside a :class:`TextIOWrapper`.  This
+includes standard streams and therefore affects the built-in function
+:func:`print()` as well.
+

Modified: python/branches/release27-maint/Lib/test/test_io.py
==============================================================================
--- python/branches/release27-maint/Lib/test/test_io.py	(original)
+++ python/branches/release27-maint/Lib/test/test_io.py	Fri Dec  3 20:31:52 2010
@@ -2561,12 +2561,50 @@
     def test_interrupted_write_text(self):
         self.check_interrupted_write("xy", b"xy", mode="w", encoding="ascii")
 
+    def check_reentrant_write(self, data, **fdopen_kwargs):
+        def on_alarm(*args):
+            # Will be called reentrantly from the same thread
+            wio.write(data)
+            1/0
+        signal.signal(signal.SIGALRM, on_alarm)
+        r, w = os.pipe()
+        wio = self.io.open(w, **fdopen_kwargs)
+        try:
+            signal.alarm(1)
+            # Either the reentrant call to wio.write() fails with RuntimeError,
+            # or the signal handler raises ZeroDivisionError.
+            with self.assertRaises((ZeroDivisionError, RuntimeError)) as cm:
+                while 1:
+                    for i in range(100):
+                        wio.write(data)
+                        wio.flush()
+                    # Make sure the buffer doesn't fill up and block further writes
+                    os.read(r, len(data) * 100)
+            exc = cm.exception
+            if isinstance(exc, RuntimeError):
+                self.assertTrue(str(exc).startswith("reentrant call"), str(exc))
+        finally:
+            wio.close()
+            os.close(r)
+
+    def test_reentrant_write_buffered(self):
+        self.check_reentrant_write(b"xy", mode="wb")
+
+    def test_reentrant_write_text(self):
+        self.check_reentrant_write("xy", mode="w", encoding="ascii")
+
+
 class CSignalsTest(SignalsTest):
     io = io
 
 class PySignalsTest(SignalsTest):
     io = pyio
 
+    # Handling reentrancy issues would slow down _pyio even more, so the
+    # tests are disabled.
+    test_reentrant_write_buffered = None
+    test_reentrant_write_text = None
+
 
 def test_main():
     tests = (CIOTest, PyIOTest,

Modified: python/branches/release27-maint/Misc/NEWS
==============================================================================
--- python/branches/release27-maint/Misc/NEWS	(original)
+++ python/branches/release27-maint/Misc/NEWS	Fri Dec  3 20:31:52 2010
@@ -1,4 +1,3 @@
-:xq
 Python News
 +++++++++++
 
@@ -13,6 +12,10 @@
 Library
 -------
 
+- Issue #10478: Reentrant calls inside buffered IO objects (for example by
+  way of a signal handler) now raise a RuntimeError instead of freezing the
+  current process.
+
 - Issue #10497: Fix incorrect use of gettext in argparse.
 
 - Issue #10464: netrc now correctly handles lines with embedded '#' characters.

Modified: python/branches/release27-maint/Modules/_io/bufferedio.c
==============================================================================
--- python/branches/release27-maint/Modules/_io/bufferedio.c	(original)
+++ python/branches/release27-maint/Modules/_io/bufferedio.c	Fri Dec  3 20:31:52 2010
@@ -224,6 +224,7 @@
 
 #ifdef WITH_THREAD
     PyThread_type_lock lock;
+    volatile long owner;
 #endif
 
     Py_ssize_t buffer_size;
@@ -259,17 +260,39 @@
 /* These macros protect the buffered object against concurrent operations. */
 
 #ifdef WITH_THREAD
-#define ENTER_BUFFERED(self) \
-    if (!PyThread_acquire_lock(self->lock, 0)) { \
-        Py_BEGIN_ALLOW_THREADS \
-        PyThread_acquire_lock(self->lock, 1); \
-        Py_END_ALLOW_THREADS \
+
+static int
+_enter_buffered_busy(buffered *self)
+{
+    if (self->owner == PyThread_get_thread_ident()) {
+        PyObject *r = PyObject_Repr((PyObject *) self);
+        if (r != NULL) {
+            PyErr_Format(PyExc_RuntimeError,
+                         "reentrant call inside %s",
+                         PyString_AS_STRING(r));
+            Py_DECREF(r);
+        }
+        return 0;
     }
+    Py_BEGIN_ALLOW_THREADS
+    PyThread_acquire_lock(self->lock, 1);
+    Py_END_ALLOW_THREADS
+    return 1;
+}
+
+#define ENTER_BUFFERED(self) \
+    ( (PyThread_acquire_lock(self->lock, 0) ? \
+       1 : _enter_buffered_busy(self)) \
+     && (self->owner = PyThread_get_thread_ident(), 1) )
 
 #define LEAVE_BUFFERED(self) \
-    PyThread_release_lock(self->lock);
+    do { \
+        self->owner = 0; \
+        PyThread_release_lock(self->lock); \
+    } while(0);
+
 #else
-#define ENTER_BUFFERED(self)
+#define ENTER_BUFFERED(self) 1
 #define LEAVE_BUFFERED(self)
 #endif
 
@@ -425,7 +448,8 @@
     int r;
 
     CHECK_INITIALIZED(self)
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self))
+        return NULL;
 
     r = buffered_closed(self);
     if (r < 0)
@@ -438,7 +462,8 @@
     /* flush() will most probably re-take the lock, so drop it first */
     LEAVE_BUFFERED(self)
     res = PyObject_CallMethodObjArgs((PyObject *)self, _PyIO_str_flush, NULL);
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self))
+        return NULL;
     if (res == NULL) {
         goto end;
     }
@@ -643,6 +668,7 @@
         PyErr_SetString(PyExc_RuntimeError, "can't allocate read lock");
         return -1;
     }
+    self->owner = 0;
 #endif
     /* Find out whether buffer_size is a power of 2 */
     /* XXX is this optimization useful? */
@@ -669,7 +695,8 @@
     CHECK_INITIALIZED(self)
     CHECK_CLOSED(self, "flush of closed file")
 
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self))
+        return NULL;
     res = _bufferedwriter_flush_unlocked(self, 0);
     if (res != NULL && self->readable) {
         /* Rewind the raw stream so that its position corresponds to
@@ -696,7 +723,8 @@
         return NULL;
     }
 
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self))
+        return NULL;
 
     if (self->writable) {
         res = _bufferedwriter_flush_unlocked(self, 1);
@@ -731,7 +759,8 @@
 
     if (n == -1) {
         /* The number of bytes is unspecified, read until the end of stream */
-        ENTER_BUFFERED(self)
+        if (!ENTER_BUFFERED(self))
+            return NULL;
         res = _bufferedreader_read_all(self);
         LEAVE_BUFFERED(self)
     }
@@ -739,7 +768,8 @@
         res = _bufferedreader_read_fast(self, n);
         if (res == Py_None) {
             Py_DECREF(res);
-            ENTER_BUFFERED(self)
+            if (!ENTER_BUFFERED(self))
+                return NULL;
             res = _bufferedreader_read_generic(self, n);
             LEAVE_BUFFERED(self)
         }
@@ -767,7 +797,8 @@
     if (n == 0)
         return PyBytes_FromStringAndSize(NULL, 0);
 
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self))
+        return NULL;
     
     if (self->writable) {
         res = _bufferedwriter_flush_unlocked(self, 1);
@@ -823,7 +854,8 @@
     
     /* TODO: use raw.readinto() instead! */
     if (self->writable) {
-        ENTER_BUFFERED(self)
+        if (!ENTER_BUFFERED(self))
+            return NULL;
         res = _bufferedwriter_flush_unlocked(self, 0);
         LEAVE_BUFFERED(self)
         if (res == NULL)
@@ -867,7 +899,8 @@
         goto end_unlocked;
     }
 
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self))
+        goto end_unlocked;
 
     /* Now we try to get some more from the raw stream */
     if (self->writable) {
@@ -1017,7 +1050,8 @@
         }
     }
 
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self))
+        return NULL;
 
     /* Fallback: invoke raw seek() method and clear buffer */
     if (self->writable) {
@@ -1055,7 +1089,8 @@
         return NULL;
     }
 
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self))
+        return NULL;
 
     if (self->writable) {
         res = _bufferedwriter_flush_unlocked(self, 0);
@@ -1715,7 +1750,10 @@
         return NULL;
     }
 
-    ENTER_BUFFERED(self)
+    if (!ENTER_BUFFERED(self)) {
+        PyBuffer_Release(&buf);
+        return NULL;
+    }
 
     /* Fast path: the data to write can be fully buffered. */
     if (!VALID_READ_BUFFER(self) && !VALID_WRITE_BUFFER(self)) {


More information about the Python-checkins mailing list