diff --git a/Lib/test/test_io.py b/Lib/test/test_io.py index 76ce347aa75..edb593b23d5 100644 --- a/Lib/test/test_io.py +++ b/Lib/test/test_io.py @@ -1336,6 +1336,26 @@ class BufferedRandomTest(BufferedReaderTest, BufferedWriterTest): bufio.readinto(bytearray(1)) self.check_writes(_read) + def test_write_after_readahead(self): + # Issue #6629: writing after the buffer was filled by readahead should + # first rewind the raw stream. + for overwrite_size in [1, 5]: + raw = self.BytesIO(b"A" * 10) + bufio = self.tp(raw, 4) + # Trigger readahead + self.assertEqual(bufio.read(1), b"A") + self.assertEqual(bufio.tell(), 1) + # Overwriting should rewind the raw stream if it needs so + bufio.write(b"B" * overwrite_size) + self.assertEqual(bufio.tell(), overwrite_size + 1) + # If the write size was smaller than the buffer size, flush() and + # check that rewind happens. + bufio.flush() + self.assertEqual(bufio.tell(), overwrite_size + 1) + s = raw.getvalue() + self.assertEqual(s, + b"A" + b"B" * overwrite_size + b"A" * (9 - overwrite_size)) + def test_misbehaved_io(self): BufferedReaderTest.test_misbehaved_io(self) BufferedWriterTest.test_misbehaved_io(self) diff --git a/Misc/NEWS b/Misc/NEWS index 9fd2a513c0b..48c58fbfef2 100644 --- a/Misc/NEWS +++ b/Misc/NEWS @@ -354,6 +354,11 @@ Core and Builtins Library ------- +- Issue #6629: Fix a data corruption issue in the new `io` package, which could + occur when writing to a BufferedRandom object (e.g. a file opened in "rb+" or + "wb+" mode) after having buffered a certain amount of data for reading. This + bug was not present in the pure Python implementation. + - Issue #4660: If a multiprocessing.JoinableQueue.put() was preempted, it was possible to get a spurious 'task_done() called too many times' error. diff --git a/Modules/_io/bufferedio.c b/Modules/_io/bufferedio.c index 554e113daa3..6e5e0fafc7b 100644 --- a/Modules/_io/bufferedio.c +++ b/Modules/_io/bufferedio.c @@ -1774,6 +1774,19 @@ bufferedwriter_write(buffered *self, PyObject *args) } Py_CLEAR(res); + /* Adjust the raw stream position if it is away from the logical stream + position. This happens if the read buffer has been filled but not + modified (and therefore _bufferedwriter_flush_unlocked() didn't rewind + the raw stream by itself). + Fixes issue #6629. + */ + n = RAW_OFFSET(self); + if (n != 0) { + if (_buffered_raw_seek(self, -n, 1) < 0) + goto error; + self->raw_pos -= n; + } + /* Then write buf itself. At this point the buffer has been emptied. */ remaining = buf.len; written = 0;