On Thu, Jun 06, 2024 at 04:05:03PM +0200, J. Hannken-Illjes wrote: > > On 6. Jun 2024, at 00:55, Greg Troxel <g...@lexort.com> wrote: > > > > Brian Marcotte <marco...@panix.com> writes: > > > >> Since upgrading to NetBSD-10, we've seen memory leaks in several > >> daemons which use libpthread: > <snip> > > > > > > I am seeing the leak with gpg-agent and mariadbd. I don't run the rest > > on 10. > > For me it is spamass-milter. On an amd64, 10.0, 56 days up: > > PID USERNAME PRI NICE SIZE RES STATE TIME WCPU CPU COMMAND > 387 root 85 0 26G 236M poll/0 0:50 0.00% 0.00% > spamass-mil > > On amd64 a vsize of 26G is only nasty, on my i386 mail server the > milter dies every few days at ~2G :-) > > I have a small test (p2.c, attached) that shows the leak. It is > always a combination of pthread_create()/pthread_detach(). Using > pthread_join() instead of pthread_detach() doesn't have the problem. > > All started with this commit: > > https://mail-index.netbsd.org/source-changes/2020/01/27/msg113341.html > > Undoing part of this commit (diff attached) makes the leak disappear > for me. I'm not sure if this is fix or it just plasters another bug, > I'm not very familiar with the lwp lifecycle. > > At least I'm sure pthread__deadqueue fills with dead pthreads and > the mapped stacks are the memory leak observed.
Ah-ha! Now I know why my toy code didn't replicate the issue - there's a required race between thread creates and exits. My code was effectively waiting for the child thread to exit before creating the next thread... Thanks for filling in the gap! -- Paul Ripke "Great minds discuss ideas, average minds discuss events, small minds discuss people." -- Disputed: Often attributed to Eleanor Roosevelt. 1948.
