Hi Willy! I have 2 more haproxy-servers with exactly the same configuration and load. Both has threads compiled in but not enabled in config (no nbthreads). And there're no segfaults at all. So I'm sure everything is fine without threads. Haproxy's config file itself is way too large to find out exact proxy-section where the fault occurred to tell you about configuration :( But if it could help: we have few heavy loaded proxy-sections with hundred of servers inside, round-roubin algo and maxconn=2. Stick tables are enabled but only to calculate average characteristics of traffic and use it in ACL so no real stickness on backends. I still have the core file available, so I can extract some more detailed traces if you need'em.
Thank you, Maksim 2018-03-05 18:56 GMT+05:00 Willy Tarreau <w...@1wt.eu>: > Hi Maxsim, > > On Mon, Mar 05, 2018 at 03:08:11PM +0300, ?????? ????????? wrote: > > Hi! > > > > I have a backtrace for segfault in haproxy=1.8.4 with 4 threads. It > happens > > usually under heavy load. Can you take a look? > > > > Using host libthread_db library "/lib/x86_64-linux-gnu/ > libthread_db.so.1". > > Core was generated by `/usr/sbin/haproxy -f /etc/haproxy/haproxy-market. > cfg > > -p /var/run/haproxy-market'. > > Program terminated with signal SIGSEGV, Segmentation fault. > > #0 __pendconn_free (p=0x55d61e970cd8) at src/queue.c:292 > > 292 HA_ATOMIC_SUB(&p->srv->nbpend, 1); > > (gdb) bt > > #0 __pendconn_free (p=0x55d61e970cd8) at src/queue.c:292 > > #1 0x000055d61be491de in pendconn_get_next_strm (px=0x55d61e96fea0, > > srv=0x55d61ea3d950) at src/queue.c:122 > > #2 process_srv_queue (s=0x55d61ea3d950) at src/queue.c:153 > > It looks very likely that this could be directly related to the threads > indeed, and pendconns can be recycled very quickly, so maybe we're reusing > one a bit too fast (or maybe we have a very old use-after-free there that > is magnified by the threads). > > Did you have the opportunity to get this one without threads, or conversely > do you know if it works fine without threads ? > > I guess you have maxconn enabled on your servers, could you tell us a bit > more about the way the farm is set up (LB algo, #servers, average maxconn, > stickiness or not). > > Thanks, > Willy >