Re: oom_killer crash linux system

From: KAMEZAWA Hiroyuki
Date: Tue Oct 19 2010 - 01:32:12 EST


On Tue, 19 Oct 2010 14:23:29 +0900
Minchan Kim <minchan.kim@xxxxxxxxx> wrote:

> On Tue, Oct 19, 2010 at 11:59 AM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > On Tue, 19 Oct 2010 10:07:38 +0800
> > "Figo.zhang" <zhangtianfei@xxxxxxxxxxxxxxxx> wrote:
> >
> >>
> >> >
> >> > very lots of change ;)
> >> > can you please send us your crash log?
> >>
> >> i add some prink in select_bad_process() and oom_badness() to see
> >> pid/totalpages/points/memoryuseage/and finally process to selet to kill.
> >>
> >> i found it the oom-killer select: syslog-ng,mysqld,nautilus,VirtualBox
> >> to kill, so my question is:
> >>
> >> 1. the syslog-ng,mysqld,nautilus is the system foundamental process, so
> >> if oom-killer kill those process, the system will be damaged, such as
> >> lose some important data.
> >>
> >> 2. the new oom-killer just use percentage of used memory as score to
> >> select the candidate to kill, but how to know this process to very
> >> important for system?
> >>
> >
> > The kernel can never know it. Just an admin (a man or management software) knows.
> > Old kernel tries to guess it, but it tend to be wrong and many many report comes
> > "why my ....is killed..." All guesswork the kernel does is not enough, I think.
> >
> >> oom_score_adj, it is anyone commercial linux distributions to use this
> >> to protect the critical process.
> >>
> > oom_adj may be used in some system. All my customers select panic_at_oom=1
> > and cause cluster fail over rather than half-broken.
> >
> > <Off topic>
> > Your another choice is memory cgroup, I think.
> > please see documentation/cgroup/memory.txt or libcgroup.
> > http://sourceforge.net/projects/libcg/
> > You can use some fancy controls with it.
> > </Off topic>
> >
> >
> > BTW, there seems to be some strange things.
> > (CC'ed to linux-mm)
> > Brief Summary:
> > Â an oom-killer happens on swapless environment with 2.6.36-rc8.
> > Â It has 2G memory.
> > a reporter says
> > ==
> >> i want to test the oom-killer. My desktop (Dell optiplex 780, i686
> >> kernel)have 2GB ram, i turn off the swap partition, and open a huge pdf
> >> files and applications, and let the system eat huge ram.
> >>
> >> in 2.6.35, i can use ram up to 1.75GB,
> >>
> >> but in 2.6.36-rc8, i just use to 1.53GB ram , the system come very slow
> >> and crashed after some minutes , the DiskIO is very busy. i see the
> >> DiskIO read is up to 8MB/s, write just only 400KB/s, (see by conky).
> > ==
> >
> > The trigger of oom-kill is order=0 allocation. (see original mail for full log)
> >
> >
> > Oct 19 09:44:08 myhost kernel: [ Â618.441470] httpd invoked oom-killer:
> > gfp_mask=0x201da, order=0, oom_adj=0, oom_score_adj=0
> >
> > Zone's stat is.
> >
> > Oct 19 09:44:08 myhost kernel: [ Â618.441551]
> > DMA free:7968kB min:64kB low:80kB high:96kB active_anon:3700kB inactive_anon:3752kB
> > Â Âactive_file:12kB inactive_file:252kB unevictable:0kB isolated(anon):0kB
> > Â Âisolated(file):0kB present:15788kB mlocked:0kB dirty:0kB writeback:4kB
> > Â Âmapped:52kB shmem:348kB slab_reclaimable:0kB slab_unreclaimable:16kB
> > Â Âkernel_stack:0kB pagetables:0kB unstable:0kB bounce:0kB
> > Â Âwriteback_tmp:0kB pages_scanned:421 all_unreclaimable? yes
> > Â Âlowmem_reserve[]: 0 865 1980 1980
> >
> > Oct 19 09:44:08 myhost kernel: [ Â618.441560]
> > Normal free:39348kB min:3728kB low:4660kB high:5592kB active_anon:176740kB
> > Â Â Â inactive_anon:25640kB active_file:84kB inactive_file:308kB
> > Â Â Â unevictable:0kB isolated(anon):0kB isolated(file):0kB present:885944kB
> > Â Â Â mlocked:0kB dirty:0kB writeback:4kB mapped:576992kB shmem:5024kB
> > Â Â Â slab_reclaimable:7612kB slab_unreclaimable:15512kB kernel_stack:2792kB
> > Â Â Â pagetables:6884kB unstable:0kB bounce:0kB writeback_tmp:0kB
> > Â Â Â pages_scanned:741 all_unreclaimable? yes
> > Â Â Â lowmem_reserve[]: 0 0 8921 8921
> >
> > Oct 19 09:44:08 myhost kernel: [ Â618.441569]
> > HighMem free:392kB min:512kB low:1712kB high:2912kB active_anon:492208kB
> > Â Â Â Âinactive_anon:166404kB active_file:180kB inactive_file:840kB
> > Â Â Â Âunevictable:40kB isolated(anon):0kB isolated(file):0kB present:1141984kB
> > Â Â Â Âmlocked:40kB dirty:0kB writeback:12kB mapped:493648kB shmem:72216kB
> > Â Â Â Âslab_reclaimable:0kB slab_unreclaimable:0kB kernel_stack:0kB
> > Â Â Â Âpagetables:0kB unstable:0kB bounce:0kB writeback_tmp:0kB
> > Â Â Â Âpages_scanned:1552 all_unreclaimable? yes
> >
> > Highmem seems a bit strange.
> > Âpresent(1141984) - active_anon - inactive_anon - inactive_file - active_file
> > Â= 482352kB but free is 392kB.
> >
> > ÂHighmem is used for some other purpose than usual user's page.(pagetable is 0.)
> > ÂAnd, Hmm, mapped:493648kB seems too large for me.
> > Â(active/inactive-file + shmem is not enough.)
> > ÂAnd "mapped" in NORMAL zone is large, too.
> >
> > ÂDoes anyone have idea about file-mapped-but-not-on-LRU pages ?
>
> Isn't it possible some file pages are much sharable?
> Please see the page_add_file_rmap.
>

page_add_file_rmap() just counts an event where mapcount goes 0->1.
Even if thousands process shares a page, it's just counted into file_mapped as 1.

Then, there are 480MB of mapped file caches. Do I miss something ?

Anyway, sum-of-all-lru-of-highmem is 480MB smaller than present pages.
and isolated(anon/file) is 0kB.
(NORMAL has similar problem)


Thanks,
-Kame




--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/