Bug 1838

Summary:

/proc/self/oom_adj needs to use /proc/self/oom_score_adj

Product:

Portable OpenSSH

Reporter:

David Rientjes <rientjes>

Component:

sshd

Assignee:

Assigned to nobody <unassigned-bugs>

Status:

CLOSED FIXED

Severity:

minor

CC:

djm, donald.r.gray, dtucker, pipping, rientjes, vapier

Priority:

Version:

5.6p1

Hardware:

All

OS:

Linux

Bug Depends on:

Bug Blocks:

1803

Attachments:

Description	Flags
openssh-linux-new-oomkill.patch	djm: ok+
Switched to end-of-array sentinel.	djm: ok+

Description David Rientjes 2010-11-16 11:31:09 AEDT

/proc/self/oom_adj is deprecated in Linux 2.6.36 and later kernels; it is scheduled for removal in August 2012.

sshd uses this interface to disable oom killing.

The new interface to use is /proc/self/oom_score_adj.  Instead of writing -17 to /proc/self/oom_adj, sshd should now write -1000 to /proc/self/oom_score_adj for the same behavior.

The filename and OOM_ADJ_NOKILL needs to be changed in openbsd-compat/port-linux.c for 2.6.36 and later kernels.

Comment 1 Darren Tucker 2010-11-16 16:24:22 AEDT

Created attachment 1957 [details]
openssh-linux-new-oomkill.patch

Try new OOM interface first then fall back if necessary.  Tested only on 2.6.32.

Comment 2 Darren Tucker 2010-11-24 10:22:56 AEDT

Comment on attachment 1957 [details]
openssh-linux-new-oomkill.patch

Works for me (2.6.32) but it would be nice if someone with a 2.6.36+ kernel could confirm it behaves as expected on that too.

Comment 3 Damien Miller 2010-12-18 09:31:12 AEDT

ping: Darren's patch can't go in until someone with a 2.6.36 system tests his patch.

Comment 4 Don Gray 2011-01-05 12:13:31 AEDT

(In reply to comment #3)
> ping: Darren's patch can't go in until someone with a 2.6.36 system
> tests his patch.

Tested on 2.6.36-ck5-r5 x86_64 openssh-5.6_p1

Error is gone from dmesg when restarting sshd.

value of /proc/$PID/oom_score_adj is -1000
value of /proc/$PID/oom_adj is -17

Comment 5 Mike Frysinger 2011-01-09 09:08:10 AEDT

your oom_adjust should be static and constant, and you should be using
ARRAY_SIZE() instead of "2"

Comment 6 Damien Miller 2011-01-10 06:41:32 AEDT

The idiom we use is to have a sentinel array entry with an obvious NULL or -1 somewhere we can test for:

for (i = 0; whatever[i].fnord != NULL; i++) ...

Comment 7 Darren Tucker 2011-01-17 11:04:18 AEDT

Created attachment 1980 [details]
Switched to end-of-array sentinel.

fair comments, I've changed it to use the sentinel since that's consistent with what we do elsewhere.

Comment 8 Darren Tucker 2011-01-17 11:56:21 AEDT

Thanks, this has been committed and will be in 5.7p1.

Comment 9 Damien Miller 2011-01-24 12:33:46 AEDT

Move resolved bugs to CLOSED after 5.7 release