Re: Still OOM problems with 4.9er/4.10er kernels

On Fri 17-03-17 21:08:31, Gerhard Wiesinger wrote:

On 17.03.2017 18:13, Michal Hocko wrote:

quoted

On Fri 17-03-17 17:37:48, Gerhard Wiesinger wrote:
[...]

quoted

Why does the kernel prefer to swapin/out and not use

a.) the free memory?

It will use all the free memory up to min watermark which is set up
based on min_free_kbytes.

Makes sense, how is /proc/sys/vm/min_free_kbytes default value calculated?

See init_per_zone_wmark_min

quoted

quoted

b.) the buffer/cache?

the memory reclaim is strongly biased towards page cache and we try to
avoid swapout as much as possible (see get_scan_count).

If I understand it correctly, swapping is preferred over dropping the
cache, right. Can this behaviour be changed to prefer dropping the
cache to some minimum amount?  Is this also configurable in a way?

No, we enforce swapping if the amount of free + file pages are below the
cumulative high watermark.

(As far as I remember e.g. kernel 2.4 dropped the caches well).

quoted

quoted

There is ~100M memory available but kernel swaps all the time ...

Any ideas?

Kernel: 4.9.14-200.fc25.x86_64

top - 17:33:43 up 28 min,  3 users,  load average: 3.58, 1.67, 0.89
Tasks: 145 total,   4 running, 141 sleeping,   0 stopped,   0 zombie
%Cpu(s): 19.1 us, 56.2 sy,  0.0 ni,  4.3 id, 13.4 wa, 2.0 hi,  0.3 si,  4.7
st
KiB Mem :   230076 total,    61508 free,   123472 used,    45096 buff/cache

procs -----------memory---------- ---swap-- -----io---- -system--
------cpu-----
 r  b   swpd   free   buff  cache   si   so    bi    bo in   cs us sy id wa st
 3  5 303916  60372    328  43864 27828  200 41420   236 6984 11138 11 47  6 23 14

I am really surprised to see any reclaim at all. 26% of free memory
doesn't sound as if we should do a reclaim at all. Do you have an
unusual configuration of /proc/sys/vm/min_free_kbytes ? Or is there
anything running inside a memory cgroup with a small limit?

nothing special set regarding /proc/sys/vm/min_free_kbytes (default values),
detailed config below. Regarding cgroups, none of I know. How to check (I
guess nothing is set because cg* commands are not available)?

be careful because systemd started to use some controllers. You can
easily check cgroup mount points.

/proc/sys/vm/min_free_kbytes
45056

So at least 45M will be kept reserved for the system. Your data
indicated you had more memory. How does /proc/zoneinfo look like?
Btw. you seem to be using fc kernel, are there any patches applied on
top of Linus tree? Could you try to retest vanilla kernel?
-- 
Michal Hocko
SUSE Labs

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>