From mboxrd@z Thu Jan 1 00:00:00 1970 From: Xiaoxi Chen Subject: Re: Failing OSDs (suicide timeout) due to flaky clients Date: Tue, 5 Jul 2016 09:38:11 +0800 Message-ID: References: <256042034.446.1467643268046@ox.pcextreme.nl> <410573451.453.1467650925047@ox.pcextreme.nl> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8BIT Return-path: Received: from mail-yw0-f180.google.com ([209.85.161.180]:32854 "EHLO mail-yw0-f180.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753223AbcGEBiM convert rfc822-to-8bit (ORCPT ); Mon, 4 Jul 2016 21:38:12 -0400 Received: by mail-yw0-f180.google.com with SMTP id v77so50654951ywg.0 for ; Mon, 04 Jul 2016 18:38:12 -0700 (PDT) In-Reply-To: <410573451.453.1467650925047@ox.pcextreme.nl> Sender: ceph-devel-owner@vger.kernel.org List-ID: To: Wido den Hollander Cc: Dan van der Ster , ceph-devel This is not heartbeat timeout between OSD, it's internal thread timeout, in this case is there is an OSD_OP_THREAD 7f8cb45a3700 timeout after 150s. The timeout value is counted by a alarm, which is reset every time the thread picked and about to start processing a request. So if something prevent the thread from finishing the request, it could timeout like this. But I dont understand as the send_message in ceph is async--just putting the message into buffer rather than send it, so theoretically it will not block the op thread? 2016-07-05 0:48 GMT+08:00 Wido den Hollander : > >> Op 4 juli 2016 om 17:54 schreef Dan van der Ster : >> >> >> Hi Wido, >> >> On Mon, Jul 4, 2016 at 4:41 PM, Wido den Hollander wrote: >> > Hi, >> > >> > On a Ceph Hammer (0.94.5) cluster I've seen OSDs fail multiple times when one or more clients are responding flaky (full netfilter conntrack table). >> > >> > The cluster is a 250 OSD cluster with roughly 30 clients all running KVM/Qemu with librbd. >> > >> > What happens is that the cluster is doing roughly 10k IOps and then a client starts to experience packet loss. Due to this responses sent back to the client by the OSD might be lost and the OSD (TCP) will re-transmit these. >> > >> > The clients in this case run into a full conntrack table on the hypervisor which also causes Ceph traffic to be dropped. >> > >> > After a few minutes OSDs will start to commit suicide throughout the cluster: >> > >> > -4> 2016-07-04 15:38:37.794230 7f8c9b63c700 10 monclient: renew_subs >> > -3> 2016-07-04 15:38:37.794235 7f8c9b63c700 10 monclient: _send_mon_message to mon.charlie at [2a00:f10:121:200::6789:3]:6789/0 >> > -2> 2016-07-04 15:38:39.944116 7f8cb45a3700 1 heartbeat_map is_healthy 'OSD::osd_op_tp thread 0x7f8c95630700' had timed out after 15 >> > -1> 2016-07-04 15:38:39.944146 7f8cb45a3700 1 heartbeat_map is_healthy 'OSD::osd_op_tp thread 0x7f8c95630700' had suicide timed out after 150 >> > 0> 2016-07-04 15:38:39.952054 7f8cb45a3700 -1 common/HeartbeatMap.cc: In function 'bool ceph::HeartbeatMap::_check(ceph::heartbeat_handle_d*, const char*, time_t)' thread 7f8cb45a3700 time 2016-07-04 15:38:39.944180 >> > common/HeartbeatMap.cc: 79: FAILED assert(0 == "hit suicide timeout") >> > ceph version 0.94.5 (9764da52395923e0b32908d83a9f7304401fee43) >> > 1: (ceph::__ceph_assert_fail(char const*, char const*, int, char const*)+0x8b) [0xbc60eb] >> > 2: (ceph::HeartbeatMap::_check(ceph::heartbeat_handle_d*, char const*, long)+0x2a9) [0xb02089] >> > 3: (ceph::HeartbeatMap::is_healthy()+0xd6) [0xb02916] >> > 4: (ceph::HeartbeatMap::check_touch_file()+0x17) [0xb02ff7] >> > 5: (CephContextServiceThread::entry()+0x154) [0xbd61d4] >> > 6: (()+0x8182) [0x7f8cb7a0d182] >> > 7: (clone()+0x6d) [0x7f8cb5f5747d] >> > NOTE: a copy of the executable, or `objdump -rdS ` is needed to interpret this. >> >> I don't understand why a client's full conntrack table would affect >> heartbeats between OSDs. > > Me neither, but it is happening. My theory for now is that somewhere in the network stack something becomes saturated because the OSDs can't send it's reply back properly. > > As soon as the client's network is working fine again the Ceph cluster is healthy and OSDs are no longer crashing. > >> Can you work backwards in the logs to see what 7f8c95630700 was doing? > > -1196> 2016-07-04 15:36:27.792105 7f8c9b63c700 10 log_client will send 2016-07-04 15:36:24.187675 osd.9 [XXXXX:2]:6820/3330 23763 : cluster [WRN] slow request 2.996474 seconds old, received at 2016-07-04 15:36:21.191081: osd_op(client.18130582.0:130421015 rbd_data.b7b662ae8944a.000000000000d339 [set-alloc-hint object_size 4194304 write_size 4194304,write 1339392~8192] 20.6b89ee5c snapc 37e=[] ack+ondisk+write e382057) currently no flag points reached > -1195> 2016-07-04 15:36:27.792121 7f8c9b63c700 10 log_client will send 2016-07-04 15:36:24.187680 osd.9 [XXXXX:2]:6820/3330 23764 : cluster [WRN] slow request 2.879790 seconds old, received at 2016-07-04 15:36:21.307765: osd_repop(client.12376090.0:17851029 26.1f2 4a7c1f2/rbd_data.270ca14cf1dfd2.0000000000000a3b/head//26 v 382057'4620873) currently no flag points reached > > That's what it was doing before it crashed. The 100 lines before are almost identical. > >> >> > >> > On other OSDs you might see messages like these come by: >> > >> > -96> 2016-07-04 15:33:23.645713 7fa9b7243700 0 -- [XX:YY:AA:BB::30]:6803/7451 >> [XX:YY:AA:BB::29]:6800/1019342 pipe(0x2c4e4000 sd=22 :56587 s=1 pgs=835 cs=1 l=0 c=0x318f1de0).connect got RESETSESSION >> > -95> 2016-07-04 15:33:23.645836 7fa9db92d700 0 -- [XX:YY:AA:BB::30]:6803/7451 >> [XX:YY:AA:BB::39]:6822/2687 pipe(0x4914000 sd=20 :38712 s=1 pgs=20 cs=1 l=0 c=0x3e4e1080).connect got RESETSESSION >> > -94> 2016-07-04 15:33:24.563595 7faa09f29700 10 monclient: renew_subs >> > -93> 2016-07-04 15:33:24.563621 7faa09f29700 10 monclient: _send_mon_message to mon.beta at [2a00:f10:121:200::6789:2]:6789/0 >> > -92> 2016-07-04 15:33:24.564221 7faa09f29700 10 monclient: renew_subs >> > -91> 2016-07-04 15:33:24.564226 7faa09f29700 10 monclient: _send_mon_message to mon.beta at [2a00:f10:121:200::6789:2]:6789/0 >> > -90> 2016-07-04 15:33:24.567635 7faa09f29700 10 monclient: renew_subs >> > -89> 2016-07-04 15:33:24.567642 7faa09f29700 10 monclient: _send_mon_message to mon.beta at [2a00:f10:121:200::6789:2]:6789/0 >> > -88> 2016-07-04 15:33:24.567667 7faa0cf2f700 0 log_channel(cluster) log [WRN] : map e382054 wrongly marked me down >> > -87> 2016-07-04 15:33:25.231536 7faa270cd700 1 heartbeat_map is_healthy 'OSD::osd_op_tp thread 0x7faa0271a700' had timed out after 15 >> > -86> 2016-07-04 15:33:28.800283 7fa9e65d9700 0 -- [XX:YY:AA:BB::30]:6813/1007451 >> [XX:YY:AA:BB::20]:6821/2641 pipe(0x3706c000 sd=25 :48471 s=2 pgs=25097 cs=1 l=0 c=0x120c5fa0).fault, initiating reconnect >> > -85> 2016-07-04 15:33:28.800795 7fa9ec437700 0 -- [XX:YY:AA:BB::30]:6813/1007451 >> [XX:YY:AA:BB::39]:6813/4497 pipe(0x341bc000 sd=26 :53469 s=2 pgs=24916 cs=1 l=0 c=0xdadf8c0).fault, initiating reconnect >> > -84> 2016-07-04 15:33:28.801374 7fa9a4f81700 0 -- [XX:YY:AA:BB::30]:6813/1007451 >> [XX:YY:AA:BB::20]:6821/2641 pipe(0x3706c000 sd=25 :48473 s=1 pgs=25097 cs=2 l=0 c=0x120c5fa0).connect got RESETSESSION >> > -83> 2016-07-04 15:33:28.801832 7fa9aced3700 0 -- [XX:YY:AA:BB::30]:6813/1007451 >> [XX:YY:AA:BB::39]:6813/4497 pipe(0x341bc000 sd=26 :53471 s=1 pgs=24916 cs=2 l=0 c=0xdadf8c0).connect got RESETSESSION >> > -82> 2016-07-04 15:33:30.231626 7faa270cd700 1 heartbeat_map is_healthy 'OSD::osd_op_tp thread 0x7faa0271a700' had timed out after 15 >> > -81> 2016-07-04 15:33:32.366023 7faa04f1f700 10 monclient: tick >> >> >> Again, what did thread 7faa0271a700 last do before timing out? >> >> (BTW, in the past we used >> >> ms tcp read timeout = 60 >> >> to work around strange el6 kernel networking bugs -- maybe it would help here.) >> >> -- Dan >> >> > >> > >> > Due to the flaky responses of the clients roughly 10 to 20 OSDs die or start to respond very, very slowly causing outage on RBD. >> > >> > As soon as the client is fixed it takes just a few minutes for the cluster to become healthy a proceed. >> > >> > To me it seems like there are some buffers inside the OSD which fill up and cause the OSD to respond slowly to other network traffic. >> > >> > Does that make any sense? >> > >> > Wido >> > -- >> > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in >> > the body of a message to majordomo@vger.kernel.org >> > More majordomo info at http://vger.kernel.org/majordomo-info.html > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html