Customer is using mrg rt kernel is his machines . He is getting a large number of system hangs. There is memory shortage in the server during the hang. But oom killer is not getting invoked.
DATE: Thu Apr 15 18:36:59 2010
UPTIME: 6 days, 04:37:46
LOAD AVERAGE: 1378.45, 1353.83, 1114.23
VERSION: #1 SMP PREEMPT RT Mon Nov 16 12:02:19 EST 2009
MACHINE: x86_64 (3333 Mhz)
MEMORY: 32 GB
PANIC: "SysRq : Trigger a crashdump"
I think there may be two issues here:
1) the reaper is not getting called. I check the oom flags in the sosreport and the oom-killer should be called.
2) the hang. I suspect that this may be a dup of bz 546428. There are 928 processes in shrink_zone which just doesn't seem right.
grep -c shrink_zone btall
Created attachment 413153 [details]
log from vmcore
Created attachment 413160 [details]
foreach bt dump
I found a patch from RHEL we have been carrying for a long time in the v1 kernel, linux-2.6-rt-oomkill.patch, that (among other things) includes a function called should_oom_kill() that attempts to avoid and OOM frenzy. As there are no comments in the patch, I am right now investigating why the patch was added in the first place and whether we need to remove the should_oom_kill() references or fix the function.
I have a test kernel running with the should_oom_kill() bits removed and it is behaving the expected way. I have several instances of three different memory hoggers running and they are being killed right away when an OOM situation happens.
As soon as the historical motivation for the inclusion of that patch are clear, I will have a fix added to our latest kernel and test it on bigger systems.
Created attachment 426656 [details]
reproducer for 158 crash
what is the status of this one?
A patch, described below, has been added to kernel -160 in order to enhance the behavior of the oom-killer (speed and accuracy). This patch also includes a new invocation path for the oom-killer (page fault handling).
oom-killer: several fixes and enhancements
Bugzilla: 589741 591137
This patch contains the backport of portions of upstream commits related
to the oom-killer, bringing mrg v1 oom-killer closer to upstream. The
following upstream commits have been backported:
1c0fe6e mm: invoke oom-killer from page fault
ff0ceb9 oom: serialize out of memory calls
28b83c5 oom: move oom_adj value from task_struct to signal_struct
4365a56 oom-kill: fix NUMA constraint check with nodemask
b95c35e oom: fix the unsafe usage of badness() in proc_oom_score()
d553ad8 param: fix NULL comparison on oom
1ac0cb5 mm: fix anonymous dirtying
5d863b8 oom: fix oom_adjust_write() input sanity check
6583bb6 mm: avoid endless looping for oom killed tasks
82553a9 oom: invoke oom killer for __GFP_NOFAIL
a12888f oom_kill: don't call for int_sqrt(0)
4779280 mm: make get_user_pages() interruptible (partially backported)
7a36a75 get_user_pages(): fix possible page leak on oom
a1e0961 relay: nopage
e91a810 oom_kill bug
7b1915a mm/oom_kill.c: Use list_for_each_entry instead of list_for_each
fyi: As per customer the kernel-rt supplied with MRG 1.3 beta works very good. OOM is done in under 2 seconds.
What kind of infos is needed?
I'm the customer with the problem.
That may be possible. We're not sure yet.
but if SELinux is all that is required, will we support 1.3 rt on rhel5.4+seLinux upgrage?
we have SELinux deactivated with selinux=0 in /proc/cmdline.
SELinux is not a showstopper.
aka the Customer ;-)
We updated to MRG 1.3 and the bug does not exsist there.
Service Ticket is already closed.