Re: [PATCH v2] mm: correctly synchronize rss-counters at exit/exec

[Konstantin Khlebnikov]

Andrew Morton wrote:
> On Tue, 10 Apr 2012 21:07:32 +0400
> Konstantin Khlebnikov<khlebnikov@xxxxxxxxxx>  wrote:
>
> > mm->rss_stat counters have per-task delta: task->rss_stat, before changing
> > task->mm pointer kernel must flush this delta with help of sync_mm_rss().
> >
> > do_exit() already calls sync_mm_rss() to flush rss-counters before commiting
> > rss-statistics into task->signal->maxrss, taskstats, audit and other stuff.
> > Unfortunately kernel do this before calling mm_relese(), which can call put_user()
> > for processing task->clear_child_tid. So at this point we can trigger page-faults
> > and task->rss_stat becomes non-zero again, as result mm->rss_stat becomes
> > inconsistent and check_mm() will print something like this:
> >
> > | BUG: Bad rss-counter state mm:ffff88020813c380 idx:1 val:-1
> > | BUG: Bad rss-counter state mm:ffff88020813c380 idx:2 val:1
> >
> > This patch moves sync_mm_rss() into mm_release(), and moves mm_release() out of
> > do_exit() and calls it earlier. After mm_release() there should be no page-faults.
> >
> > ...
> >
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -751,6 +751,14 @@ void mm_release(struct task_struct *tsk, struct mm_struct *mm)
> >   		}
> >   		tsk->clear_child_tid = NULL;
> >   	}
> > +
> > +	/*
> > +	 * Final rss-counter synchronization. After this point must be
> > +	 * no page-faults into this mm from current context, otherwise
> > +	 * mm->rss_stat will be inconsistent.
> > +	 */
> > +	if (mm)
> > +		sync_mm_rss(mm);
> >   }
>
> Well that's scary.  AFACIT `mm' can indeed be NULL here, when a kernel
> thread calls do_exit().  No implementation of deactivate_mm() actually
> uses its `mm' arg and I guess that kernel threads never set
> tsk->clear_child_tid.  Whee.

But it works as designed =)

> Do we think we should backport this into -stable kernels?  How hard is
> it to make that warning come out?

Warning was introduced in v3.3-3720-gc3f0327, so no releases with them.
And I think older kernels can live fine with slightly racy rss counters.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/