Date: Thu, 22 Sep 2016 13:20:45 +0300 From: Slawa Olhovchenkov <slw@zxy.spb.ru> To: Julien Charbon <jch@freebsd.org> Cc: Konstantin Belousov <kostikbel@gmail.com>, freebsd-stable@FreeBSD.org, hiren panchasara <hiren@strugglingcoder.info> Subject: Re: 11.0 stuck on high network load Message-ID: <20160922102045.GC2840@zxy.spb.ru> In-Reply-To: <67862b33-63c0-2f23-d254-5ddc55dbb554@freebsd.org> References: <20160916190330.GG2840@zxy.spb.ru> <78cbcdc9-f565-1046-c157-2ddd8fcccc62@freebsd.org> <20160919204328.GN2840@zxy.spb.ru> <8ba75d6e-4f01-895e-0aed-53c6c6692cb9@freebsd.org> <20160920202633.GQ2840@zxy.spb.ru> <f644cd52-4377-aa90-123a-3a2887972bbc@freebsd.org> <20160921195155.GW2840@zxy.spb.ru> <e4e0188c-b22b-29af-ed15-b650c3ec4553@gmail.com> <20160922095331.GB2840@zxy.spb.ru> <67862b33-63c0-2f23-d254-5ddc55dbb554@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Thu, Sep 22, 2016 at 12:04:40PM +0200, Julien Charbon wrote: > >> These paths can indeed compete for the same INP lock, as both > >> tcp_tw_2msl_scan() calls always start with the first inp found in > >> twq_2msl list. But in both cases, this first inp should be quickly used > >> and its lock released anyway, thus that could explain your situation it > >> that the TCP stack is doing that all the time, for example: > >> > >> - Let say that you are running out completely and constantly of tcptw, > >> and then all connections transitioning to TIME_WAIT state are competing > >> with the TIME_WAIT timeout scan that tries to free all the expired > >> tcptw. If the stack is doing that all the time, it can appear like > >> "live" locked. > >> > >> This is just an hypothesis and as usual might be a red herring. > >> Anyway, could you run: > >> > >> $ vmstat -z | head -2; vmstat -z | grep -E 'tcp|sock' > > > > ITEM SIZE LIMIT USED FREE REQ FAIL SLEEP > > > > socket: 864, 4192664, 18604, 25348,49276158, 0, 0 > > tcp_inpcb: 464, 4192664, 34226, 18702,49250593, 0, 0 > > tcpcb: 1040, 4192665, 18424, 18953,49250593, 0, 0 > > tcptw: 88, 16425, 15802, 623,14526919, 8, 0 > > tcpreass: 40, 32800, 15, 2285, 632381, 0, 0 > > > > In normal case tcptw is about 16425/600/900 > > > > And after `sysctl -a | grep tcp` system stuck on serial console and I am reset it. > > > >> Ideally, once when everything is ok, and once when you have the issue > >> to see the differences (if any). > >> > >> If it appears your are quite low in tcptw, and if you have enough > >> memory, could you try increase the tcptw limit using sysctl > > > > I think this is not eliminate stuck, just may do it less frequency > > You are right, it would just be a big hint that the tcp_tw_2msl_scan() > contention hypothesis is the right one. As I see you have plenty of > memory on your server, thus could you try with: > > net.inet.tcp.maxtcptw=4192665 > > And see what happen. Just to validate this hypothesis. This is bad way for validate, with maxtcptw=16384 happened is random and can be waited for month. After maxtcptw=4192665 I am don't know how long need to wait for verification this hypothesis. More frequency (may be 3-5 times per day) happening less traffic drops (not to zero for minutes). May be this caused also by contention in tcp_tw_2msl_scan, but fast resolved (stochastic process). By eating CPU power nginx can't service connection and clients closed connections and need more TIME_WAIT and can trigered tcp_tw_2msl_scan(reuse=1). After this we can got live lock. May be after I learning to catch and dignostic this validation is more accurately.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20160922102045.GC2840>