Date: Sun, 14 Apr 2019 09:25:54 +0200 From: Jan Beich <jbeich@FreeBSD.org> To: Kubilay Kocak <koobs@FreeBSD.org> Cc: FreeBSD Ports <freebsd-ports@freebsd.org> Subject: Re: python 3 subprocess performance Message-ID: <36ml-oxod-wny@FreeBSD.org> In-Reply-To: <800de545-0310-fcda-ebae-a4dffabe47e9@FreeBSD.org> (Kubilay Kocak's message of "Sun, 14 Apr 2019 14:59:53 %2B1000") References: <20190411161649.1b740d21@vm2.home.zagrebin.ru> <8f3f8413-60f2-bb03-a6b4-4f6364cdc3df@rlwinm.de> <20190411143926.5rg4jskmodt4shhi@laparbeit> <9729db47-12c4-caf4-cdcf-1913dab73c8e@rlwinm.de> <20190412101012.4142854f@vm2.home.zagrebin.ru> <CAAWYfq2P5QsSbc2bzQauApF6=ijQYpKCb2HQ1jy-4dNZCOB_=g@mail.gmail.com> <20190412104531.7b492a3c@vm2.home.zagrebin.ru> <CAAWYfq16Y90-nXeV3TkT10uh=_tK=Vgjk_QEJo=DM-P8KGFNQg@mail.gmail.com> <800de545-0310-fcda-ebae-a4dffabe47e9@FreeBSD.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Kubilay Kocak <koobs@FreeBSD.org> writes: > On 12/04/2019 8:41 pm, Dima Pasechnik wrote: > >> On Fri, Apr 12, 2019 at 9:46 AM Alexander Zagrebin <alex@zagrebin.ru> wr= ote: >>> >>> =D0=92 Fri, 12 Apr 2019 09:36:13 +0200 >>> Dima Pasechnik <dimpase+freebsd@gmail.com> =D0=BF=D0=B8=D1=88=D0=B5=D1= =82: >>> >>>> On Fri, Apr 12, 2019 at 9:11 AM Alexander Zagrebin <alex@zagrebin.ru> >>>> wrote: >>>>> >>>>> =D0=92 Thu, 11 Apr 2019 17:32:42 +0200 >>>>> Jan Bramkamp <crest@rlwinm.de> =D0=BF=D0=B8=D1=88=D0=B5=D1=82: >>>>> >>>>>> The reason is that that python does something stupid (tm). It >>>>>> tries to close all file descriptors (except a few whitelisted >>>>>> ones) up to the maximum file descriptor number. It does this by >>>>>> asking the kernel for the maximum possible number and closing >>>>>> everything it doesn't want to keep. Some time later someone came >>>>>> up with an optimization (read the open file descriptors >>>>>> from /dev/fd). All of this pain and suffering is caused by good >>>>>> old Ulrich Drepper braindamage: >>>>>> https://sourceware.org/bugzilla/show_bug.cgi?id=3D10353. >>>>>> >>>>>> Most Linux distros have lower default file descriptor limits than >>>>>> FreeBSD making this workaround less painful. The correct solution >>>>>> would be to teach python3 about closefrom(2). >>>>> >>>>> Thank you for hint and testing! >>>>> >>>>> Indeed the problem is in closing more than 400,000 file descriptors >>>>> in loop. It seems that all current versions of Python are affected. >>>>> Python2 uses False as default value for the close_fds parameter of >>>>> the Popen constructor, so this issue is mostly not visible. >>>>> Python3 has changed this default to True. >>>>> >>>>> As Jan Bramkamp suggested, I've wrote simple patch to fix an issue >>>>> (see attached file). It seems the problem has gone. >>>> >>>> The attachment has been stripped out. Could you paste the diff into >>>> the message? >>> >>> Yes, sure. >>> >>> --- Modules/_posixsubprocess.c.orig 2018-12-24 00:37:14.000000000 >>> +0300 +++ Modules/_posixsubprocess.c 2019-04-12 >>> 09:25:21.549389000 +0300 @@ -235,11 +235,15 @@ >>> _close_fds_by_brute_force(long start_fd, } >>> start_fd =3D keep_fd + 1; >>> } >>> +#if defined(__FreeBSD__) >>> + closefrom(start_fd); >>> +#else >>> if (start_fd <=3D end_fd) { >>> for (fd_num =3D start_fd; fd_num < end_fd; ++fd_num) { >>> close(fd_num); >>> } >>> } >>> +#endif >>> } >>> >>>> If this is a Python issue, shouldn't this be reported upstream, on >>>> https://bugs.python.org ? >>> >>> May be. Rather, it is a FreeBSD-specific optimization. >> >> Well, closefrom() is also available in Darwin (a.k.a. MacOSX :-)), >> OpenBSD and NetBSD. (It's not documented in current MacOSX, but it is >> there, I just checked) FreeBSD was late to the party. Even DragonFly got closefrom() before. ;) Note, closefrom() looks non-atomic on Solaris. http://src.illumos.org/source/xref/illumos-gate/usr/src/lib/libc/port/gen/c= losefrom.c > Issue exists for this: > > https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=3D221700 Oh, nice. I've recently it hit via www/firefox build. https://bugzilla.mozilla.org/show_bug.cgi?id=3D1507655
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?36ml-oxod-wny>