On 10/18/2012 11:32 PM, Michal Hocko wrote:
On Thu 18-10-12 21:51:57, Sha Zhengju wrote:
On 10/18/2012 07:56 PM, Michal Hocko wrote:
On Wed 17-10-12 01:14:48, Sha Zhengju wrote:
On Tuesday, October 16, 2012, Michal Hocko<mhocko@xxxxxxx> wrote:
[...]
Could you be more specific about the motivation for this patch? Is it
"let's be consistent with the global oom" or you have a real use case
for this knob.
In our environment(rhel6), we encounter a memcg oom 'deadlock'
problem. Simply speaking, suppose process A is selected to be killed
by memcg oom killer, but A is uninterruptible sleeping on a page
lock. What's worse, the exact page lock is holding by another memcg
process B which is trapped in mem_croup_oom_lock(proves to be a
livelock).
Hmm, this is strange. How can you get down that road with the page lock
held? Is it possible this is related to the issue fixed by: 1d65f86d
(mm: preallocate page before lock_page() at filemap COW)?
No, it has nothing with the cow page. By checking stack of the process A
selected to be killed(uninterruptible sleeping), it was stuck at:
__do_fault->filemap_fault->__lock_page_or_retry->wait_on_page_bit--(D
state).
The person B holding the exactly page lock is on the following path:
__do_fault->filemap_fault->__do_page_cache_readahead->..->mpage_readpages
->add_to_page_cache_locked ---->(in memcg oom and cannot exit)
Hmm filemap_fault locks the page after the read ahead is triggered
already so it doesn't call mpage_readpages with any page locked - the
add_to_page_cache_lru is called without any page locked.
It's not the page being fault in filemap_fault that causing the problem,
but those
pages handling by readhead. To clarify the point, the more detailed call
stack is:
filemap_fault->do_async/sync_mmap_readahead->ondemand_readahead->
__do_page_cache_readahead->read_pages->ext3/4_readpages->*mpage_readpages*
It is because mpage_readpages that bring the risk:
for each of readahead pages
(1)add_to_page_cache_lru (--> *will lock page and go through memcg
charging*)
add the page to a big bio
submit_bio (So those locked pages will be unlocked in end_bio after swapin)
So if a page is being charged and cannot exit from memcg oom successfully
(following I'll explain the reason) in step (1), it will cause the
submit_bio indefinitely
postponed while holding the PageLock of previous pages.
This is at least the current code. It might be different in rhel6 but
calling memcg charging with a page lock is definitely a bug.
The current code (mm repo since-3.6) here remains unchanged. Through we
may need
to take care of page lock and memcg charging in mpage_readpages, it
dives to fs level.
Besides 37b23e05 have already fixed the deadlock from the other side:
process still can be
killed even waiting for pagelock. But considering other potential
problem, we may as well do
something in mpage_readpages to avoid calling add_to_page_cache_lru with
any page locked.
In mpage_readpages, B tends to read a dozen of pages in: for each of
page will do
locking, charging, and then send out a big bio. And A is waiting for
one of the pages
and stuck.
As I said, 37b23e05 has made pagefault killable by changing
uninterruptible sleeping to killable sleeping. So A can be woke up to
exit successfully and free the memory which can in turn help B pass
memcg charging period.
(By the way, it seems commit 37b23e05 and 7d9fdac need to be
79dfdaccd1d5 you mean, right? That one just helps when there are too
many tasks trashing oom killer so it is not related to what you are
trying to achieve. Besides that make sure you take 23751be0 if you
take it.
Here is the reason why I said a process may go though memcg oom and cannot
exit. It's just the phenomenon described in the commit log of 79dfdaccd:
the old version of memcg oom lock can lead to serious starvation and make
many tasks trash oom killer but nothing useful can be done.
It is for these two reasons that cause the bug and can make the memcg
unusable
(sys up to almost 100%)for hours even days... Once we give some extra memory
to the memcg(such as increase hardlimit a little), the processes tending
into oom killer
will pass the charging and send bio out eventually, which will unlock
those pages and
wake up the D sleeper.
backported to --stable tree to deliver RHEL users. ;-) )
I am not sure the first one qualifies the stable tree inclusion as it is
a feature.
When debugging the problem, we indeed found 37b23e05 is the key
enemy of the deadlock bug.
Then A can not exit successfully to free the memory and both of them
can not moving on.
Indeed, we should dig into these locks to find the solution and
in fact the 37b23e05 (x86, mm: make pagefault killable) and
7d9fdac(Memcg: make oom_lock 0 and 1 based other than counter) have
already solved the problem, but if oom_killing_allocating_task is
memcg aware, enabling this suicide oom behavior will be a simpler
workaround. What's more, enabling the sysctl can avoid other potential
oom problems to some extent.
As I said, I am not against this but I really want to see a valid use
case first. So far I haven't seen any because what you mention above is
a clear bug which should be fixed. I can imagine the huge number of
tasks in the group could be a problem as well but I would like to see
what are those problems first.
In view of consistent with global oom and performance benefit, I suggest
we may as well open it in memcg oom as there's no obvious harm.
I am not sure about "no obvious harm" part. The policy could be
different in different groups e.g. and the global knob could be really
misleading. But the question is. Is it worth having this per group? To
be honest, I do not like the global knob either and I am not entirely
keen on spreading it out into memcg unless there is a real use case for
it.
Okay...then let's lie it on the table. We may use it as a in-house
patch. :-)
Thanks,
Sha
As refer to the bug I mentioned, obviously the key solution is the above two
patchset, but considing other *potential* memcg oom bugs, the sysctl may
be a role of temporary workaround to some extent... but it's just a
workaround.
We shouldn't add something like that just to workaround obvious bugs.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>