[3/4] OOM, PM: OOM killed task shouldn't escape PM suspend

From: Michal Hocko <mhocko@suse.cz>

On Thu 06-11-14 10:01:21, Tejun Heo wrote:
> On Thu, Nov 06, 2014 at 01:49:53PM +0100, Michal Hocko wrote:
> > On Wed 05-11-14 12:55:27, Tejun Heo wrote:
> > > On Wed, Nov 05, 2014 at 06:46:09PM +0100, Michal Hocko wrote:
> > > > Because out_of_memory can be called from mutliple paths. And
> > > > the only interesting one should be the page allocation path.
> > > > pagefault_out_of_memory is not interesting because it cannot happen for
> > > > the frozen task.
> > > 
> > > Hmmm.... wouldn't that be broken by definition tho?  So, if the oom
> > > killer is invoked from somewhere else than page allocation path, it
> > > would proceed ignoring the disabled setting and would race against PM
> > > freeze path all the same. 
> > 
> > Not really because try_to_freeze_tasks doesn't finish until _all_ tasks
> > are frozen and a task in the page fault path cannot be frozen, can it?
> 
> We used to have freezing points deep in file system code which may be
> reacheable from page fault.

If that is really the case then there is no way around and use
out_of_memory from the page fault path as well. I cannot say I would be
happy about that though. There should be ideally only single freezing
place. But that is another story.

> Please take a step back and look at the paragraph above.  Doesn't
> it sound extremely contrived and brittle even if it's not outright
> broken?  What if somebody adds another oom killing site somewhere
> else?

The only way to add an oom killing site is out_of_memory and that does
all the magic with my patch.

> How can this possibly be a solution that we intentionally implement?
>
> > I mean there shouldn't be any problem to not invoke OOM killer under
> > from the page fault path as well but that might lead to looping in the
> > page fault path without any progress until freezer enables OOM killer on
> > the failure path because the said task cannot be frozen.
> > 
> > Is this preferable?
> 
> Why would PM freezing make OOM killing fail?  That doesn't make much
> sense.  Sure, it can block it for a finite duration for sync purposes
> but making OOM killing fail seems the wrong way around.  

We cannot block in the allocation path because the request might come
from the freezer path itself (e.g. when suspending devices etc.).
At least this is my understanding why the original oom disable approach
was implemented.

> We're doing one thing for non-PM freezing and the other way around for
> PM freezing, which indicates one of the two directions is wrong.

Because those two paths are quite different in their requirements. The
cgroup freezer only cares about freezing tasks and it doesn't have to
care about tasks accessing a possibly half suspended device on their way
out.

> Shouldn't it be that OOM killing happening while PM freezing is in
> progress cancels PM freezing rather than the other way around?  Find a
> point in PM suspend/hibernation operation where everything must be
> stable, disable OOM killing there and check whether OOM killing
> happened inbetween and if so back out. 

This is freeze_processes AFAIU. I might be wrong of course but this is
the time since when nobody should be waking processes up because they
could access half suspended devices.

> It seems rather obvious to me that OOM killing has to have precedence
> over PM freezing.
> 
> Sure, once the system reaches a point where the whole system must be
> in a stable state for snapshotting or whatever, disabling OOM killing
> is fine but at that point the system is in a very limited execution
> mode and sure won't be processing page faults from userland for
> example and we can actually disable OOM killing knowing that anything
> afterwards is ready to handle memory allocation failures.

I am really confused now. This is basically what the final patch does
actually.  Here is the what I have currently just to make the further
discussion easier.
---
From 337e772eaf636a96409e84bcd33d77ebc2950549 Mon Sep 17 00:00:00 2001
From: Michal Hocko <mhocko@suse.cz>
Date: Wed, 5 Nov 2014 15:09:56 +0100
Subject: [PATCH 1/2] OOM, PM: make OOM detection in the freezer path raceless

5695be142e20 (OOM, PM: OOM killed task shouldn't escape PM suspend)
has left a race window when OOM killer manages to note_oom_kill after
freeze_processes checks the counter. The race window is quite small and
really unlikely and partial solution deemed sufficient at the time of
submission.

Tejun wasn't happy about this partial solution though and insisted on
a full solution. That requires the full OOM and freezer exclusion,
though. This is done by this patch which introduces oom_sem RW lock
and gets rid of oom_killer_disabled global flag.

The PM code uses oom_killer_{disable,enable} which takes the lock
for write and excludes all the OOM killer invocation from any
out_of_memory users which newly returns a success status. It fails
only if the oom_sem cannot be taken for read which indicates that
OOM has been disabled. This is done by read trylock so we can never
deadlock.

The caller has to take an appropriate action when the out_of_memory
fails.

The allocation path simply fails the allocation request the same
way as previously. Sysrq path notes that the OOM didn't happen due to
OOM disable.

The page fault path ignored oom disabled flag previously with an
assumption that the page fault path cannot enter the fridge. As per
Tejun the freezing point used to be deep in the fs code. Therefore it
is safer and more robust to include pagefault_out_of_memory as well.
The task will be refaulting until there is some memory freed or PM
freezer fails because the said task cannot be frozen and re-enable OOM
killer when the OOM eventually happens if the memory still short.

There is no need to recheck all the processes with the full
synchronization anymore.

Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Michal Hocko <mhocko@suse.cz>

fold me
---
 drivers/tty/sysrq.c    |  6 ++++--
 include/linux/oom.h    | 25 +++++++++++++----------
 kernel/power/process.c | 50 ++++++++--------------------------------------
 mm/oom_kill.c          | 54 ++++++++++++++++++++++++++++++++------------------
 mm/page_alloc.c        | 32 +++++++++++++++---------------
 5 files changed, 77 insertions(+), 90 deletions(-)

[3/4] OOM, PM: OOM killed task shouldn't escape PM suspend

Commit Message

Comments

Patch