Excerpts from Yuya Nishihara's message of 2016-11-16 21:51:50 +0900: > On Tue, 15 Nov 2016 20:38:24 +0000, Jun Wu wrote: > > # HG changeset patch > > # User Jun Wu <qu...@fb.com> > > # Date 1479241551 0 > > # Tue Nov 15 20:25:51 2016 +0000 > > # Node ID f3d2f4ebc4006043684db52e4487756dd4e2d238 > > # Parent d1a0a64f6e16432333bea0476098c46a61222b9b > > # Available At https://bitbucket.org/quark-zju/hg-draft > > # hg pull https://bitbucket.org/quark-zju/hg-draft -r > > f3d2f4ebc400 > > util: improve iterfile so it chooses code path wisely > > > +if (pyplatform.python_implementation() == 'CPython' and > > + sys.version_info < (3, 0)): > > + # There is an issue in CPython that some IO methods do not handle EINTR > > + # correctly. The following table shows what CPython version (and > > functions) > > + # are affected (Y: has the EINTR bug, N: otherwise): > > + # > > + # | < 2.7.4 | 2.7.4 to 2.7.12 | >= 3.0 > > + # -------------------------------------------------- > > + # fp.__iter__ | Y | Y | N > > + # fp.read* | Y | N [1] | N > > + # > > + # [1]: fixed by hg changeset 67dc99a989cd. > > + # > > + # Here we workaround the EINTR issue for fileobj.__iter__. Other > > methods > > + # like "read*" are ignored for now, as Python < 2.7.4 is a minority. > > + # > > + # Although we can workaround the EINTR issue for fp.__iter__, it is > > slower: > > + # "for x in fp" is 4x faster than "for x in iter(fp.readline, '')" in > > + # CPython 2, because the latter maintains an internal readahead buffer. > > Do you mean "the former" ?
Good catch. Was trying to refer to "CPython" and saw it's at the right. > > + # On modern systems like Linux, the "read" syscall cannot be > > interrupted > > + # when reading "fast" files like on-disk files. So the EINTR issue only > > + # affects things like pipes, sockets, ttys etc. We treat "normal" > > (S_ISREG) > > + # files approximately as "fast" files and use the fast (unsafe) code > > path, > > + # to minimize the performance impact. > > + if sys.version_info >= (2, 7, 4): > > + # fp.readline deals with EINTR correctly, use it as a workaround > > + def _safeiterfile(fp): > > + return iter(fp.readline, '') > > + else: > > + # fp.read* are broken too, manually deal with EINTR in a stupid way > > + # note: this may block longer than necessary because of bufsize. > > + def _safeiterfile(fp, bufsize=4096): > > + fd = fp.fileno() > > + line = '' > > + while True: > > + try: > > + buf = os.read(fd, bufsize) > > + except OSError as ex: > > + if ex.errno == errno.EINTR: > > + continue > > + else: > > + raise > > + line += buf > > + if '\n' in buf: > > + splitted = line.splitlines(True) > > + line = '' > > + for l in splitted: > > + if l[-1] == '\n': > > + yield l > > + else: > > + line = l > > + if not buf: > > + break > > Missed the last line if not ends with '\n'. Aside from that, it seems the test cases introduced by 67dc99a989cd could be weak - I'll double check if "buf = os.read(fd, bufsize)" can "read some bytes, then interruptted" so "buf" gets modified, even on EINTR. If not, the only safe way seems to be the very stupid os.read(fd, 1) (an previous unsent version). It's 30x slower. I think we probably don't want a 30x slower workaround. As we still support Python 2.6, this should be solved before the SIGCHLD patch. > > + def iterfile(fp): > > + fastpath = True > > + if type(fp) is file: > > + fastpath = stat.S_ISREG(os.fstat(fp.fileno()).st_mode) > > + if fastpath: > > + return fp > > + else: > > + return _safeiterfile(fp) > > +else: > > + # PyPy and CPython 3 do not have the EINTR issue thus no workaround > > needed. > > + def iterfile(fp): > > + return fp _______________________________________________ Mercurial-devel mailing list Mercurial-devel@mercurial-scm.org https://www.mercurial-scm.org/mailman/listinfo/mercurial-devel
