On Tue 23-08-16 11:06:55, Michal Hocko wrote:
> On Tue 23-08-16 09:55:55, Michal Hocko wrote:
> > On Tue 23-08-16 00:01:23, Michael S. Tsirkin wrote:
> > [...]
> > > Actually, vhost net calls out to tun which does regular copy_from_iter.
> > > Returning 0 there will cause corrupted packets in the network: not a
> > > huge deal, but ugly.  And I don't think we want to annotate run and
> > > macvtap as well.
> > 
> > Hmm, OK, I wasn't aware of that path and being consistent here matters.
> > If the vhost driver can interact with other subsystems then there is
> > really no other option than hooking into the page fault path. Ohh well.
> 
> Here is a completely untested patch just for sanity check.

OK, so I've tested the patch and it seems to work as expected. I didn't
know how to configure vhost so I've just hacked up a quick kernel thread
which picks on a task (I am always selecting a first task which is the
OOM_SCORE_ADJ_MAX because that would be the selected victim - see the
code attached) and then read through its address space in a loop. The
oom victim then just mmaps and poppulates a private anon mapping which
causes the oom killer. It properly notices that the memor could have
been reaped.
[  628.559374] Out of memory: Kill process 3193 (oom_victim) score 1868 or sacrifice child
[  628.561713] Killed process 3193 (oom_victim) total-vm:1052540kB, anon-rss:782964kB, file-rss:12kB, shmem-rss:0kB
[  628.568120] Found a dead vma: ret:-14vma ffff88003c697000 start 00007f9227b33000 end 00007f9267b33000
next ffff88003d80d8a0 prev ffff88003d80d228 mm ffff88003d97b200
prot 8000000000000025 anon_vma ffff88003dbbc000 vm_ops           (null)
pgoff 7f9227b33 file           (null) private_data           (null)
flags: 0x100073(read|write|mayread|maywrite|mayexec|account)
[  628.595684] oom_reaper: reaped process 3193 (oom_victim), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[  673.366282] Waiting for kthread to stop
[  673.367487] Done

Are there any objections or suggestions to the apporach?
-- 
Michal Hocko
SUSE Labs