OK to set PF_MEMDIE on cleanup tasks?

OK to set PF_MEMDIE on cleanup tasks?

[Paul E. McKenney]

Hello!

We have tasks that actively return memory to the system, which we
would like to exempt from the OOM killer, as killing such tasks under
low-memory conditions would indeed be counterproductive.  It looks like
the "official" way to do this is to catch/ignore signal 15, which results
in PF_MEMDIE being set (in the 2.6 kernel), thus preventing the OOM killer
from killing the task again.  I don't see where PF_MEMDIE is cleared,
though there are a number of subtle ways one might do this that I would
have missed.

So...  Is it considered legit to simply set PF_MEMDIE when creating
the cleanup task?  Or is there some reason that one should deal with
signal 15?

All enlightenment much appreciated!

						Thanx, Paul

Re: OK to set PF_MEMDIE on cleanup tasks?

[Andrew Morton]

"Paul E. McKenney" <paulmck@us.ibm.com> wrote:
>
> Hello!
> 
> We have tasks that actively return memory to the system, which we
> would like to exempt from the OOM killer, as killing such tasks under
> low-memory conditions would indeed be counterproductive.  It looks like
> the "official" way to do this is to catch/ignore signal 15, which results
> in PF_MEMDIE being set (in the 2.6 kernel), thus preventing the OOM killer
> from killing the task again.  I don't see where PF_MEMDIE is cleared,
> though there are a number of subtle ways one might do this that I would
> have missed.

The PF_MEMDIE flag is there so the oom killer doesn't just sit there
hitting the same task over and over again.

We leave PF_MEMDIE set because we expect the task to exit, or to not want
any more oomkiller attention.

The SIGTERM behaviour is there because the CAP_SYS_RAWIO process may need
to release critical resources.

So as long as your process has CAP_SYS_RAWIO, everything happens to work as
you want it.  I don't think it was really designed that way though.

> So...  Is it considered legit to simply set PF_MEMDIE when creating
> the cleanup task?  Or is there some reason that one should deal with
> signal 15?

Well it's all very unconventional.  Catching SIGTERM seems like a suitable
way to do what you want to do.

Possibly your special process should also run as PF_MEMALLOC.  I've seen
that done before, with success.  There is no existing API with which this
can be set.

Re: OK to set PF_MEMDIE on cleanup tasks?

[Gerrit Huizenga]

Re: OK to set PF_MEMDIE on cleanup tasks?

[Paul E. McKenney]

Hello, Andrew,

On Tue, Oct 14, 2003 at 05:12:27PM -0700, Andrew Morton wrote:
> > So...  Is it considered legit to simply set PF_MEMDIE when creating
> > the cleanup task?  Or is there some reason that one should deal with
> > signal 15?
> 
> Well it's all very unconventional.  Catching SIGTERM seems like a suitable
> way to do what you want to do.

OK, since this particular case is a strictly in-kernel task, SIGTERM
should be a no-op anyway.  Unless I am missing something in the signal
delivery code, which is quite probable.  ;-)

So the magic code would then be:

	cap_raise(current->cap_effective, CAP_SYS_RAWIO);

perhaps with:

	cap_raise(current->cap_effective, CAP_SYS_ADMIN);

thrown in for good measure.

> Possibly your special process should also run as PF_MEMALLOC.  I've seen
> that done before, with success.  There is no existing API with which this
> can be set.

This would certainly head off at least some OOM deadlock situations.
On the (perhaps unlikely) chance that this was an invitation, here
is a patch to create an API.

					Thanx, Paul

diff -urN -X /home/linux/2.5/dontdiff linux-2.6.0-test7-mm1/include/linux/sched.h linux-2.6.0-test7-mm1-PF_MEMALLOC/include/linux/sched.h
--- linux-2.6.0-test7-mm1/include/linux/sched.h	2003-10-16 07:16:05.000000000 -0700
+++ linux-2.6.0-test7-mm1-PF_MEMALLOC/include/linux/sched.h	2003-10-16 09:20:34.000000000 -0700
@@ -508,6 +508,27 @@
 #define PF_LESS_THROTTLE 0x00100000	/* Throttle me less: I clean memory */
 #define PF_SYNCWRITE	0x00200000	/* I am doing a sync write */
 
+/**
+ * mark_task_memalloc - mark the specified task as deserving of preferential
+ * access to free memory.  Note that with great power comes great
+ * responsibility.
+ * @p: the task structure to be granted preferential access.
+ */
+static inline void mark_task_memalloc(task_t *p)
+{
+	p->flags |= PF_MEMALLOC;
+}
+
+/**
+ * unmark_task_memalloc - mark the specified task as no longer deserving
+ * of preferential access to free memory.
+ * @p: the task structure to have its preferential access revoked.
+ */
+static inline void unmark_task_memalloc(task_t *p)
+{
+	p->flags &= ~PF_MEMALLOC;
+}
+
 #ifdef CONFIG_SMP
 extern int set_cpus_allowed(task_t *p, cpumask_t new_mask);
 #else