Created attachment 1122838 [details]
Proposed patch

totempg: Fix memory leak

Previously there were two free lists. One for operational and one for
transitional state. Because every node starts in transitional state and
always ends in the operational state, assembly was always put to normal
state free list and never in transitional free list, so new assembly
structure was always allocated after new node connected.

Solution is to have only one free list.

*** Bug 1309806 has been marked as a duplicate of this bug. ***

This bug was accidentally moved from POST to MODIFIED via an error in automation, please see mmccune@redhat.com with any questions

While looping:
systemctl start corosync.service; sleep 2; systemctl stop corosync.service; sleep 2

corosync on the non-looping node:
# cat /proc/`pgrep corosync`/status | grep RSS
Start: 
VmRSS:	    4268 kB
after 10 cycles:
VmRSS:	    5484 kB
after 50 cycles:
VmRSS:	    6980 kB
after next 50 (160 total):
VmRSS:      8784 kB

Unlike in RHEL6 this does not stop after couple of cycles. It goes like this, even after hundreds of cycles. Example of 50 run:

01 VmRSS:	    5388 kB
02 VmRSS:	    5388 kB
03 VmRSS:	    5388 kB
04 VmRSS:	    5388 kB
05 VmRSS:	    5388 kB
06 VmRSS:	    5648 kB
07 VmRSS:	    5716 kB
08 VmRSS:	    5716 kB
09 VmRSS:	    5716 kB
10 VmRSS:	    5716 kB
11 VmRSS:	    5716 kB
12 VmRSS:	    5716 kB
13 VmRSS:	    5716 kB
14 VmRSS:	    5976 kB
15 VmRSS:	    5976 kB
16 VmRSS:	    5976 kB
17 VmRSS:	    6108 kB
18 VmRSS:	    6108 kB
19 VmRSS:	    6108 kB
20 VmRSS:	    6108 kB
21 VmRSS:	    6108 kB
22 VmRSS:	    6108 kB
23 VmRSS:	    6108 kB
24 VmRSS:	    6368 kB
25 VmRSS:	    6368 kB
26 VmRSS:	    6368 kB
27 VmRSS:	    6504 kB
28 VmRSS:	    6504 kB
29 VmRSS:	    6504 kB
30 VmRSS:	    6504 kB
31 VmRSS:	    6504 kB
32 VmRSS:	    6504 kB
33 VmRSS:	    6504 kB
34 VmRSS:	    6764 kB
35 VmRSS:	    6764 kB
36 VmRSS:	    6764 kB
37 VmRSS:	    6904 kB
38 VmRSS:	    6904 kB
39 VmRSS:	    6904 kB
40 VmRSS:	    6904 kB
41 VmRSS:	    6904 kB
42 VmRSS:	    6904 kB
43 VmRSS:	    6904 kB
44 VmRSS:	    7164 kB
45 VmRSS:	    7164 kB
46 VmRSS:	    7164 kB
47 VmRSS:	    7280 kB
48 VmRSS:	    7280 kB
49 VmRSS:	    7280 kB
50 VmRSS:	    7280 kB

As you can see, it's steadily increasing, just not with every rejoin, but like every 5 or so. I had only corosync running on both nodes (no pacemaker), started via systemctl.

corosync.conf:

totem {
    version: 2
    secauth: off
    cluster_name: STSRHTS15446
    transport: udpu
}

nodelist {
    node {
        ring0_addr: virt-053
        nodeid: 1
    }

    node {
        ring0_addr: virt-054
        nodeid: 2
    }
}

quorum {
    provider: corosync_votequorum
    two_node: 1
}

logging {
    to_logfile: yes
    logfile: /var/log/cluster/corosync.log
    to_syslog: yes
}


corosynclib-2.4.0-4.el7.x86_64
libqb-1.0-1.el7.x86_64
corosync-2.4.0-4.el7.x86_64

@jkortus: Your numbers looks really interesting. I've tested exactly same versions of packages on RHEL 7.2. I've executed corosync on first node and it's memory usage right after startup was:

cat /proc/`pgrep corosync`/status | grep RSS
VmRSS:	   33276 kB

on the second node I've executed:

for i in `seq 1 200`;do systemctl start corosync.service; sleep 2; systemctl stop corosync.service; sleep 2;done

Memory usage on first node after all cycle finished:
cat /proc/`pgrep corosync`/status | grep RSS
VmRSS:	   33276 kB

Installed packages:
corosync-2.4.0-4.el7.x86_64
corosynclib-2.4.0-4.el7.x86_64
libqb-1.0-1.el7.x86_64

So conclusion:
- Your startup memory usage looks like way smaller (maybe some change in glibc memory allocation)
- I'm not able to reproduce the bug

Can you please try your test with:
- RHEL 7.2
- With RHEL 7 compose and retry more cycles (till memory grows to ~33Mb) and report what happens afterwards

?

I can reproduce this on my RHEL7.2 VMs, it behaves just as Jaroslav's cluster does, starting at 3988 kB and increasing.

libqb-1.0-1.el7.x86_64
glibc-2.17-106.el7_2.8.x86_64
corosync-2.4.0-4.el7.x86_64

@Chrissie, yep I'm now too. First run I was running corosync -f on first node. Without foreground option behavior is as described.

I've tried running test without -f and second node running:

for i in `seq 1 20000`;do systemctl start corosync.service; sleep 0.1; systemctl stop corosync.service; sleep 0.1;done

to speed things up.

Result is memory grow to 

VmRSS:	   12124 kB

after less than 1000 cycles and then memory stays on the same number for rest of the test.

So to conclude. I believe original bug is fixed (without fix memory would grow thru whole test) and described behavior is standard behavior of glibc not freeing all memory (this is expected). It's probably result of cmap (so libqb trie) or logging. Moving back to on_qa.

thanks to both of you for a quick response! I'll run it again and see if it stabilizes on ~12M.

Longer run:
00001 VmRSS:	    5044 kB
...
00215 VmRSS:	   13212 kB
00615 VmRSS:	   13212 kB
...
01570 VmRSS:	   13224 kB
...
02334 VmRSS:	   13224 kB


Looks good even after 2K+ rejoins. Marking as verified with corosync-2.4.0-4.el7.x86_64.

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://rhn.redhat.com/errata/RHBA-2016-2463.html