[PATCH] oom, memcg: handle sysctl oom_kill_allocating_task while memcg oom happening

From: Sha Zhengju <redacted>

Sysctl oom_kill_allocating_task enables or disables killing the OOM-triggering
task in out-of-memory situations, but it only works on overall system-wide oom.
But it's also a useful indication in memcg so we take it into consideration
while oom happening in memcg. Other sysctl such as panic_on_oom has already
been memcg-ware.


Signed-off-by: Sha Zhengju <redacted>
---
 mm/memcontrol.c |    9 +++++++++
 1 files changed, 9 insertions(+), 0 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index e4e9b18..c329940 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c

@@ -1486,6 +1486,15 @@ static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
 
 	check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
 	totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
+	if (sysctl_oom_kill_allocating_task && current->mm &&
+	    !oom_unkillable_task(current, memcg, NULL) &&
+	    current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
+		get_task_struct(current);
+		oom_kill_process(current, gfp_mask, order, 0, totalpages, memcg, NULL,
+				 "Memory cgroup out of memory (oom_kill_allocating_task)");
+		return;
+	}
+
 	for_each_mem_cgroup_tree(iter, memcg) {
 		struct cgroup *cgroup = iter->css.cgroup;
 		struct cgroup_iter it;

-- 
1.7.6.1

On 10/16/2012 06:12 PM, Sha Zhengju wrote:

From: Sha Zhengju <redacted>

Sysctl oom_kill_allocating_task enables or disables killing the OOM-triggering
task in out-of-memory situations, but it only works on overall system-wide oom.
But it's also a useful indication in memcg so we take it into consideration
while oom happening in memcg. Other sysctl such as panic_on_oom has already
been memcg-ware.

Is it the resend one or new version, could you add changelog if it is 
the last case?

--
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>

On 10/16/2012 06:20 PM, Ni zhan Chen wrote:

On 10/16/2012 06:12 PM, Sha Zhengju wrote:

quoted

From: Sha Zhengju <redacted>

Sysctl oom_kill_allocating_task enables or disables killing the 
OOM-triggering
task in out-of-memory situations, but it only works on overall 
system-wide oom.
But it's also a useful indication in memcg so we take it into 
consideration
while oom happening in memcg. Other sysctl such as panic_on_oom has 
already
been memcg-ware.

Is it the resend one or new version, could you add changelog if it is 
the last case?

Sorry, forget to mention that this patch is an updated one rebased on 
mhocko mm tree, since-3.6 branch.
The first one is on old kernel, please ignore it. :-)


Thanks,
Sha

On Tue 16-10-12 18:12:08, Sha Zhengju wrote:

From: Sha Zhengju <redacted>

Sysctl oom_kill_allocating_task enables or disables killing the OOM-triggering
task in out-of-memory situations, but it only works on overall system-wide oom.
But it's also a useful indication in memcg so we take it into consideration
while oom happening in memcg. Other sysctl such as panic_on_oom has already
been memcg-ware.

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.

The primary motivation for oom_kill_allocating_task AFAIU was to reduce
search over huge tasklists and reduce task_lock holding times. I am not
sure whether the original concern is still valid since 6b0c81b (mm,
oom: reduce dependency on tasklist_lock) as the tasklist_lock usage has
been reduced conciderably in favor of RCU read locks is taken but maybe
even that can be too disruptive?
David?

Moreover memcg oom killer doesn't iterate over tasklist (it uses
cgroup_iter*) so this shouldn't cause the performance problem like
for the global case.
On the other hand we are taking css_set_lock for reading for the whole
iteration which might cause some issues as well but those should better
be described in the changelog.

-- 
Michal Hocko
SUSE Labs

On Tuesday, October 16, 2012, Michal Hocko [off-list ref] wrote:

On Tue 16-10-12 18:12:08, Sha Zhengju wrote:

quoted

From: Sha Zhengju <redacted>

Sysctl oom_kill_allocating_task enables or disables killing the

OOM-triggering

quoted

task in out-of-memory situations, but it only works on overall

system-wide oom.

quoted

But it's also a useful indication in memcg so we take it into

consideration

quoted

while oom happening in memcg. Other sysctl such as panic_on_oom has

already

quoted

been memcg-ware.

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).
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.

The primary motivation for oom_kill_allocating_tas AFAIU was to reduce
search over huge tasklists and reduce task_lock holding times. I am not
sure whether the original concern is still valid since 6b0c81b (mm,
oom: reduce dependency on tasklist_lock) as the tasklist_lock usage has
been reduced conciderably in favor of RCU read locks is taken but maybe
even that can be too disruptive?
David?

On the other hand, from the semantic meaning of oom_kill_allocating_task,
it implies to allow
suicide-like oom, which has no obvious relationship with performance
problems(such as huge task lists
or task_lock holding time). So make the sysctl be consistent with global
oom will be better or set an
individual option for memcg oom just as panic_on_oom does.

Moreover memcg oom killer doesn't iterate over tasklist (it uses
cgroup_iter*) so this shouldn't cause the performance problem like
for the global case.
On the other hand we are taking css_set_lock for reading for the whole
iteration which might cause some issues as well but those should better
be described in the changelog.

quoted

Signed-off-by: Sha Zhengju <redacted>
---
 mm/memcontrol.c |    9 +++++++++
 1 files changed, 9 insertions(+), 0 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index e4e9b18..c329940 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c

@@ -1486,6 +1486,15 @@ static void mem_cgroup_out_of_memory(struct

mem_cgroup *memcg, gfp_t gfp_mask,

quoted

      check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
      totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
+     if (sysctl_oom_kill_allocating_task && current->mm &&
+         !oom_unkillable_task(current, memcg, NULL) &&
+         current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
+             get_task_struct(current);
+             oom_kill_process(current, gfp_mask, order, 0, totalpages,

memcg, NULL,

quoted

+                              "Memory cgroup out of memory

(oom_kill_allocating_task)");

quoted

+             return;
+     }
+
      for_each_mem_cgroup_tree(iter, memcg) {
              struct cgroup *cgroup = iter->css.cgroup;
              struct cgroup_iter it;
--
1.7.6.1

--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

--
Michal Hocko
SUSE Labs

On Tue, 16 Oct 2012, Michal Hocko wrote:

The primary motivation for oom_kill_allocating_task AFAIU was to reduce
search over huge tasklists and reduce task_lock holding times. I am not
sure whether the original concern is still valid since 6b0c81b (mm,
oom: reduce dependency on tasklist_lock) as the tasklist_lock usage has
been reduced conciderably in favor of RCU read locks is taken but maybe
even that can be too disruptive?
David?

When the oom killer became serialized, the folks from SGI requested this 
tunable to be able to avoid the expensive tasklist scan on their systems 
and to be able to avoid killing threads that aren't allocating memory at 
all in a steady state.  It wasn't necessarily about tasklist_lock holding 
time but rather the expensive iteration over such a large number of 
processes.

Moreover memcg oom killer doesn't iterate over tasklist (it uses
cgroup_iter*) so this shouldn't cause the performance problem like
for the global case.

Depends on how many threads are attached to a memcg.

On Tue, 16 Oct 2012, Sha Zhengju wrote:

Please try to compile your patches and run scripts/checkpatch.pl on them 
before proposing them.

You'll also need to update Documentation/sysctl/vm.txt.

On Wed 17-10-12 01:14:48, Sha Zhengju wrote:

On Tuesday, October 16, 2012, Michal Hocko [off-list ref] wrote:

[...]

quoted

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)?

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.

quoted

The primary motivation for oom_kill_allocating_tas AFAIU was to reduce
search over huge tasklists and reduce task_lock holding times. I am not
sure whether the original concern is still valid since 6b0c81b (mm,
oom: reduce dependency on tasklist_lock) as the tasklist_lock usage has
been reduced conciderably in favor of RCU read locks is taken but maybe
even that can be too disruptive?
David?

On the other hand, from the semantic meaning of oom_kill_allocating_task,
it implies to allow suicide-like oom, which has no obvious relationship
with performance problems(such as huge task lists or task_lock holding
time). 

I guess that suicide-like oom in fact means "kill the poor soul that
happened to charge the last". I do not see any use case for this from
top of my head (appart from the performance benefits of course).

So make the sysctl be consistent with global oom will be better or set
an individual option for memcg oom just as panic_on_oom does.

-- 
Michal Hocko
SUSE Labs

On 10/18/2012 07:56 PM, Michal Hocko wrote:

On Wed 17-10-12 01:14:48, Sha Zhengju wrote:

quoted

On Tuesday, October 16, 2012, Michal Hocko[off-list ref]  wrote:

[...]

quoted

quoted

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)
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 backported 
to --stable tree
to deliver RHEL users. ;-) )

quoted

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.
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.


Thanks,
Sha

quoted

quoted

The primary motivation for oom_kill_allocating_tas AFAIU was to reduce
search over huge tasklists and reduce task_lock holding times. I am not
sure whether the original concern is still valid since 6b0c81b (mm,
oom: reduce dependency on tasklist_lock) as the tasklist_lock usage has
been reduced conciderably in favor of RCU read locks is taken but maybe
even that can be too disruptive?
David?

On the other hand, from the semantic meaning of oom_kill_allocating_task,
it implies to allow suicide-like oom, which has no obvious relationship
with performance problems(such as huge task lists or task_lock holding
time).

I guess that suicide-like oom in fact means "kill the poor soul that
happened to charge the last". I do not see any use case for this from
top of my head (appart from the performance benefits of course).

quoted

So make the sysctl be consistent with global oom will be better or set
an individual option for memcg oom just as panic_on_oom does.

On Thu 18-10-12 21:51:57, Sha Zhengju wrote:

On 10/18/2012 07:56 PM, Michal Hocko wrote:

quoted

On Wed 17-10-12 01:14:48, Sha Zhengju wrote:

quoted

On Tuesday, October 16, 2012, Michal Hocko[off-list ref]  wrote:

[...]

quoted

quoted

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.
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.

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.

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.

quoted

quoted

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.

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.
-- 
Michal Hocko
SUSE Labs

On 10/18/2012 11:32 PM, Michal Hocko wrote:

On Thu 18-10-12 21:51:57, Sha Zhengju wrote:

quoted

On 10/18/2012 07:56 PM, Michal Hocko wrote:

quoted

On Wed 17-10-12 01:14:48, Sha Zhengju wrote:

quoted

On Tuesday, October 16, 2012, Michal Hocko[off-list ref]   wrote:

[...]

quoted

quoted

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.

quoted

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.

quoted

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.

quoted

quoted

quoted

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

quoted

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.

On Fri 19-10-12 12:11:52, Sha Zhengju wrote:

On 10/18/2012 11:32 PM, Michal Hocko wrote:

quoted

On Thu 18-10-12 21:51:57, Sha Zhengju wrote:

quoted

On 10/18/2012 07:56 PM, Michal Hocko wrote:

quoted

On Wed 17-10-12 01:14:48, Sha Zhengju wrote:

quoted

On Tuesday, October 16, 2012, Michal Hocko[off-list ref]   wrote:

[...]

quoted

quoted

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.

And I was probably blind yesterday because if I have looked inside
add_to_page_cache_lru then I would have found out that we lock the page
before charging it. /me stupid. Sorry about the confusion.
That one is OK, though, because the page is fresh new and not visible
when we charge it. This is not related to your problem, more on that
below.

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.

OK I think I am seeing what you are trying to say, finally. But you
are wrong here. Previously locked&charged pages were already submitted
(every do_mpage_readpage submits the given page) so the IO will finish
eventually so those pages get unlocked.
 
[...]

quoted

quoted

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.

Yes the trashing on oom is certainly possible without that patch and it
seems that this is what the culprit of the problem you are describing.

[...]
-- 
Michal Hocko
SUSE Labs