On 28/05/2013, at 4:30 PM, Andrew Beekhof <and...@beekhof.net> wrote:
> > On 28/05/2013, at 10:12 AM, Andrew Beekhof <and...@beekhof.net> wrote: > >> >> On 27/05/2013, at 5:08 PM, Vladislav Bogdanov <bub...@hoster-ok.com> wrote: >> >>> 27.05.2013 04:20, Yuichi SEINO wrote: >>>> Hi, >>>> >>>> 2013/5/24 Vladislav Bogdanov <bub...@hoster-ok.com>: >>>>> 24.05.2013 06:34, Andrew Beekhof wrote: >>>>>> Any help figuring out where the leaks might be would be very much >>>>>> appreciated :) >>>>> >>>>> One (and the only) suspect is unfortunately crmd itself. >>>>> It has private heap grown from 2708 to 3680 kB. >>>>> >>>>> All other relevant differences are in qb shm buffers, which are >>>>> controlled and may grow until they reach configured size. >>>>> >>>>> @Yuichi >>>>> I would recommend to try running under valgrind on a testing cluster to >>>>> figure out is that a memleak (lost memory) or some history data >>>>> (referenced memory). Latter may be a logical memleak though. You may >>>>> look in /etc/sysconfig/pacemaker for details. >>>> >>>> I got valgrind for about 2 days. And, I attached valgrind in ACT node >>>> and SBY node. >>> >>> >>> I do not see any "direct" memory leaks (repeating 'definitely-lost' >>> allocations) there. >>> >>> So what we see is probably one of: >>> * Cache/history/etc, which grows up to some limit (or expired at the >>> some point in time). >>> * Unlimited/not-expirable lists/hashes of data structures, which are >>> correctly freed at exit >> >> There is still plenty of memory chunks not free'd at exit, I'm slowly >> working through those. > > I've pushed the following to my repo: > > + Andrew Beekhof (2 hours ago) d070092: Test: More glib suppressions > + Andrew Beekhof (2 hours ago) ec74bf0: Fix: Fencing: Ensure API object is > consistently free'd > + Andrew Beekhof (2 hours ago) 6130d23: Fix: Free additional memory at exit > + Andrew Beekhof (2 hours ago) b76d6be: Refactor: crmd: Allocate a mainloop > before doing anything to help valgrind > + Andrew Beekhof (3 hours ago) d4041de: Log: init: Remove unnecessary detail > from shutdown message > + Andrew Beekhof (3 hours ago) 282032b: Fix: Clean up internal mainloop > structures at exit > + Andrew Beekhof (4 hours ago) 0947721: Fix: Core: Correctly unreference > GSource inputs > + Andrew Beekhof (25 hours ago) d94140d: Fix: crmd: Clean up more memory > before exit > + Andrew Beekhof (25 hours ago) b44257c: Test: cman: Ignore additional > valgrind errors > > If someone would like to run the cluster (no valgrind needed) for a while with > > export > PCMK_trace_functions=mainloop_gio_destroy,mainloop_add_fd,mainloop_del_fd,crmd_exit,crm_peer_destroy,empty_uuid_cache,lrm_state_destroy_all,internal_lrm_state_destroy,do_stop,mainloop_destroy_trigger,mainloop_setup_trigger,do_startup,stonith_api_delete > > and then (after grabbing smaps) shut it down, we should have some information > about any lists/hashes that are growing too large. > > Also, be sure to run with: > > export G_SLICE=always-malloc > > which will prevent glib from accumulating pools of memory and distorting any > results. I did this today with 2747e25 and it looks to me like there is no leak (anymore?) For context, between smaps.5 and smaps.6, the 4 node cluster ran over 120 "standby" tests (lots of PE runs and resource activity). So unless someone can show me otherwise, I'm going to move on :) Note that the [heap] changes are actually the memory usage going _backwards_. Raw results below. [root@corosync-host-1 ~]# cat /proc/`pidof crmd`/smaps > smaps.6 ; diff -u smaps.5 smaps.6; --- smaps.5 2013-05-29 02:39:25.032940230 -0400 +++ smaps.6 2013-05-29 03:48:51.278940819 -0400 @@ -40,16 +40,16 @@ Swap: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB -0226b000-02517000 rw-p 00000000 00:00 0 [heap] -Size: 2736 kB -Rss: 2268 kB -Pss: 2268 kB +0226b000-02509000 rw-p 00000000 00:00 0 [heap] +Size: 2680 kB +Rss: 2212 kB +Pss: 2212 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB -Private_Dirty: 2268 kB -Referenced: 2268 kB -Anonymous: 2268 kB +Private_Dirty: 2212 kB +Referenced: 2212 kB +Anonymous: 2212 kB AnonHugePages: 0 kB Swap: 0 kB KernelPageSize: 4 kB @@ -112,13 +112,13 @@ MMUPageSize: 4 kB 7f0c6e918000-7f0c6ee18000 rw-s 00000000 00:10 522579 /dev/shm/qb-pengine-event-27411-27412-6-data Size: 5120 kB -Rss: 3572 kB -Pss: 1785 kB +Rss: 4936 kB +Pss: 2467 kB Shared_Clean: 0 kB -Shared_Dirty: 3572 kB +Shared_Dirty: 4936 kB Private_Clean: 0 kB Private_Dirty: 0 kB -Referenced: 3572 kB +Referenced: 4936 kB Anonymous: 0 kB AnonHugePages: 0 kB Swap: 0 kB @@ -841,7 +841,7 @@ 7f0c72b00000-7f0c72b1d000 r-xp 00000000 fd:00 119 /lib64/libselinux.so.1 Size: 116 kB Rss: 36 kB -Pss: 5 kB +Pss: 4 kB Shared_Clean: 36 kB Shared_Dirty: 0 kB Private_Clean: 0 kB @@ -1401,7 +1401,7 @@ 7f0c740c6000-7f0c74250000 r-xp 00000000 fd:00 45 /lib64/libc-2.12.so Size: 1576 kB Rss: 588 kB -Pss: 20 kB +Pss: 19 kB Shared_Clean: 588 kB Shared_Dirty: 0 kB Private_Clean: 0 kB > > >> Once we know all memory is being cleaned up, the next step is to check the >> size of things beforehand. >> >> I'm hoping one or more of them show up as unnaturally large, indicating >> things are being added but not removed. >> >>> (f.e like dlm_controld has(had???) for a >>> debugging buffer or like glibc resolver had in EL3). This cannot be >>> caught with valgrind if you use it in a standard way. >>> >>> I believe we have former one. To prove that, it would be very >>> interesting to run under valgrind *debugger* (--vgdb=yes|full) for some >>> long enough (2-3 weeks) period of time and periodically get memory >>> allocation state from there (with 'monitor leak_check full reachable >>> any' gdb command). I wanted to do that a long time ago, but >>> unfortunately did not have enough spare time to even try that (although >>> I tried to valgrind other programs that way). >>> >>> This is described in valgrind documentation: >>> http://valgrind.org/docs/manual/manual-core-adv.html#manual-core-adv.gdbserver >>> >>> We probably do not need to specify '--vgdb-error=0' because we do not >>> need to install watchpoints at the start (and we do not need/want to >>> immediately connect to crmd with gdb to tell it to continue), we just >>> need to periodically get status of memory allocations >>> (stop-leak_check-cont sequence). Probably that should be done in a >>> 'fast' manner, so crmd does not stop for a long time, and the rest of >>> pacemaker does not see it 'hanged'. Again, I did not try that, and I do >>> not know if it's even possible to do that with crmd. >>> >>> And, as pacemaker heavily utilizes glib, which has own memory allocator >>> (slices), it is better to switch it to a 'standard' malloc/free for >>> debugging with G_SLICE=always-malloc env var. >>> >>> Last, I did memleak checks for a 'static' (i.e. no operations except >>> monitors are performed) cluster for ~1.1.8, and did not find any. It >>> would be interesting to see if that is true for an 'active' one, which >>> starts/stops resources, handles failures, etc. >>> >>>> >>>> Sincerely, >>>> Yuichi >>>> >>>>> >>>>>> >>>>>> Also, the measurements are in pages... could you run "getconf PAGESIZE" >>>>>> and let us know the result? >>>>>> I'm guessing 4096 bytes. >>>>>> >>>>>> On 23/05/2013, at 5:47 PM, Yuichi SEINO <seino.clust...@gmail.com> wrote: >>>>>> >>>>>>> Hi, >>>>>>> >>>>>>> I retry the test after we updated packages to the latest tag and OS. >>>>>>> glue and booth is latest. >>>>>>> >>>>>>> * Environment >>>>>>> OS:RHEL 6.4 >>>>>>> cluster-glue:latest(commit:2755:8347e8c9b94f) + >>>>>>> patch[detail:http://www.gossamer-threads.com/lists/linuxha/dev/85787] >>>>>>> resource-agent:v3.9.5 >>>>>>> libqb:v0.14.4 >>>>>>> corosync:v2.3.0 >>>>>>> pacemaker:v1.1.10-rc2 >>>>>>> crmsh:v1.2.5 >>>>>>> booth:latest(commit:67e1208973de728958432aaba165766eac1ce3a0) >>>>>>> >>>>>>> * Test procedure >>>>>>> we regularly switch a ticket. The previous test also used the same way. >>>>>>> And, There was no a memory leak when we tested pacemaker-1.1 before >>>>>>> pacemaker use libqb. >>>>>>> >>>>>>> * Result >>>>>>> As a result, I think that crmd may cause the memory leak. >>>>>>> >>>>>>> crmd smaps(a total of each addresses) >>>>>>> In detail, we attached smaps of start and end. And, I recorded smaps >>>>>>> every 1 minutes. >>>>>>> >>>>>>> Start >>>>>>> RSS: 7396 >>>>>>> SHR(Shared_Clean+Shared_Dirty):3560 >>>>>>> Private(Private_Clean+Private_Dirty):3836 >>>>>>> >>>>>>> Interbal(about 30h later) >>>>>>> RSS:18464 >>>>>>> SHR:14276 >>>>>>> Private:4188 >>>>>>> >>>>>>> End(about 70h later) >>>>>>> RSS:19104 >>>>>>> SHR:14336 >>>>>>> Private:4768 >>>>>>> >>>>>>> Sincerely, >>>>>>> Yuichi >>>>>>> >>>>>>> 2013/5/15 Yuichi SEINO <seino.clust...@gmail.com>: >>>>>>>> Hi, >>>>>>>> >>>>>>>> I ran the test for about two days. >>>>>>>> >>>>>>>> Environment >>>>>>>> >>>>>>>> OS:RHEL 6.3 >>>>>>>> pacemaker-1.1.9-devel (commit 138556cb0b375a490a96f35e7fbeccc576a22011) >>>>>>>> corosync-2.3.0 >>>>>>>> cluster-glue >>>>>>>> latest+patch(detail:http://www.gossamer-threads.com/lists/linuxha/dev/85787) >>>>>>>> libqb- 0.14.4 >>>>>>>> >>>>>>>> There may be a memory leak in crmd and lrmd. I regularly got rss of ps. >>>>>>>> >>>>>>>> start-up >>>>>>>> crmd:5332 >>>>>>>> lrmd:3625 >>>>>>>> >>>>>>>> interval(about 30h later) >>>>>>>> crmd:7716 >>>>>>>> lrmd:3744 >>>>>>>> >>>>>>>> ending(about 60h later) >>>>>>>> crmd:8336 >>>>>>>> lrmd:3780 >>>>>>>> >>>>>>>> I still don't run a test that pacemaker-1.1.10-rc2 use. So, I will run >>>>>>>> its test. >>>>>>>> >>>>>>>> Sincerely, >>>>>>>> Yuichi >>>>>>>> >>>>>>>> -- >>>>>>>> Yuichi SEINO >>>>>>>> METROSYSTEMS CORPORATION >>>>>>>> E-mail:seino.clust...@gmail.com >>>>>>> >>>>>>> >>>>>>> >>>>>>> -- >>>>>>> Yuichi SEINO >>>>>>> METROSYSTEMS CORPORATION >>>>>>> E-mail:seino.clust...@gmail.com >>>>>>> <smaps_log.tar.gz>_______________________________________________ >>>>>>> Pacemaker mailing list: Pacemaker@oss.clusterlabs.org >>>>>>> http://oss.clusterlabs.org/mailman/listinfo/pacemaker >>>>>>> >>>>>>> Project Home: http://www.clusterlabs.org >>>>>>> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf >>>>>>> Bugs: http://bugs.clusterlabs.org >>>>>> >>>>>> >>>>>> _______________________________________________ >>>>>> Pacemaker mailing list: Pacemaker@oss.clusterlabs.org >>>>>> http://oss.clusterlabs.org/mailman/listinfo/pacemaker >>>>>> >>>>>> Project Home: http://www.clusterlabs.org >>>>>> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf >>>>>> Bugs: http://bugs.clusterlabs.org >>>>>> >>>>> >>>>> >>>>> _______________________________________________ >>>>> Pacemaker mailing list: Pacemaker@oss.clusterlabs.org >>>>> http://oss.clusterlabs.org/mailman/listinfo/pacemaker >>>>> >>>>> Project Home: http://www.clusterlabs.org >>>>> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf >>>>> Bugs: http://bugs.clusterlabs.org >>>> >>>> >>>> >>>> -- >>>> Yuichi SEINO >>>> METROSYSTEMS CORPORATION >>>> E-mail:seino.clust...@gmail.com >>>> >>>> >>>> >>>> _______________________________________________ >>>> Pacemaker mailing list: Pacemaker@oss.clusterlabs.org >>>> http://oss.clusterlabs.org/mailman/listinfo/pacemaker >>>> >>>> Project Home: http://www.clusterlabs.org >>>> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf >>>> Bugs: http://bugs.clusterlabs.org >>>> >>> >>> >>> _______________________________________________ >>> Pacemaker mailing list: Pacemaker@oss.clusterlabs.org >>> http://oss.clusterlabs.org/mailman/listinfo/pacemaker >>> >>> Project Home: http://www.clusterlabs.org >>> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf >>> Bugs: http://bugs.clusterlabs.org >> > _______________________________________________ Pacemaker mailing list: Pacemaker@oss.clusterlabs.org http://oss.clusterlabs.org/mailman/listinfo/pacemaker Project Home: http://www.clusterlabs.org Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf Bugs: http://bugs.clusterlabs.org
