On Wed, May 30, 2012 at 11:07:29AM +0200, Thibaut Girka wrote: > On Wed, May 30, 2012 at 02:22:00AM -0500, Jonathan Nieder wrote: > [...] > > If it reproduces the problem, then that would mean this is not Debian- > > specific, so please write to libc-help@sourceware.org in that case to > > request advice[1]. (If we are lucky, someone might recognize the bug or > > know of some command like "strace -f" or patch that could help track > > it down.) > > I'll try that. I couldn't reproduce this because of other failed tests. But... > In the meantime, I've came up with a simpler (only one spawner, no signal handling) > test (attached) that exhibits the same issue. > It appears that, for some reason, at some point in time, the threads aren't cleaned up anymore > (just monitor /proc/$pid/task when the test is running), even though the pthread_join calls succeed... > This isn't always reproductible, but it happens fairly often. This bug seems to be only present on SMP systems. I could run eglibc's tst-eintr1 with “ulimit -u 300” and TIMEOUTFACTOR=300 without any problem. Plus, the test I wrote runs without any issue (left it run for about 30 minutes) with a dedicated user and “ulimit -u $NB_THREADS + 5”, whereas it almost immediately fails on SMP systems. (I've tried on the same machine with a -486 kernel) Regards, Thibaut Girka.
Attachment:
signature.asc
Description: Digital signature