Hi Willy, Op 9-11-2017 om 5:45 schreef Willy Tarreau:
Hi Pieter,We had something similar on Linux in relation with TCP splicing and the fd cache, for which a fix was emitted. But yesterday Christopher explained me that the fix has an impact on the way applets are scheduled in 1.8, so actually it could mean that the initial bug might possibly cover a larger scope than splicing only, and that recv+send might also be affected. If you're interested in testing, the commit in 1.7 is c040c1f ("BUG/MAJOR: stream-int: don't re-arm recv if send fails") and is present in the latest snapshot (we really need to emit 1.7.10 BTW). I'd be curious to know if it fixes it or not. At least it will tell us if that's related to this fd cache thing or to something completely different such as Lua. I also need to check with Thierry if we could find a way to add some stats about the time spent in Lua to "show info" to help debugging such cases where Lua is involved. By the way, thanks for your dump, we'll check the sessions' statuses. There are not that many, and maybe it will give us a useful indication! Cheers, Willy
Okay have been running with haproxy-ss-20171017 for a day now. Sofar it sticks to <1% cpu usage.
Will report if anything changes, cant tell for sure as don't have a clear reproduction of the issue, but my issue 'seems' fixed.
Regards, PiBa-NL / Pieter
