Message217596
It looks like a bug in the subprocess module e.g., if child process does:
sys.stdout.write(sys.stdin.readline())
sys.stdout.flush()
and the parent:
p.stdin.write(line) #NOTE: no flush
line = p.stdout.readline()
then a deadlock may happen with bufsize=1 (because it is equivalent to bufsize=-1 in the current code)
Surprisingly, it works if universal_newlines=True but only for C implementation of io i.e., if C extension is disabled:
import io, _pyio
io.TextIOWrapper = _pyio.TextIOWrapper
then it blocks forever even if universal_newlines=True with bufsize=1 that is clearly wrong (<-- bug) e.g., `test_universal_newlines` deadlocks with _pyio.TextIOWrapper
C implementation works because it sets `needflush` flag even if only `write_through` is provided [1]:
if (self->write_through)
needflush = 1;
else if (self->line_buffering &&
(haslf ||
PyUnicode_FindChar(text, '\r', 0, PyUnicode_GET_LENGTH(text), 1) != -1))
needflush = 1;
[1]: http://hg.python.org/cpython/file/0b2e199ad088/Modules/_io/textio.c#l1333
Python io implementation doesn't flush with only `write_through` flag.
It doesn't deadlock if bufsize=0 whether universal_newlines=True or not.
Note: Python 2.7 doesn't deadlock with bufsize=0 and bufsize=1 in this case as expected.
What is not clear is whether it should work with universal_newline=False and bufsize=1: both current docs and Python 2.7 behaviour say that there should not be a deadlock.
I've updated the docs to mention that bufsize=1 works only with
universal_newlines=True and added corresponding tests. I've also updated the subprocess' code to pass line_buffering explicitly.
Patch is uploaded. |
|
Date |
User |
Action |
Args |
2014-04-30 10:09:15 | akira | set | recipients:
+ akira, jwilk, raylu |
2014-04-30 10:09:15 | akira | set | messageid: <1398852555.77.0.122778986188.issue21332@psf.upfronthosting.co.za> |
2014-04-30 10:09:15 | akira | link | issue21332 messages |
2014-04-30 10:09:15 | akira | create | |
|