Re: Why kmem_cache_free occupy CPU for more than 10 seconds?

[Andrew Morton <akpm@linux-foundation.org>]

On Thu, 12 Apr 2007 09:39:25 +0200 Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:

> On Wed, 2007-04-11 at 15:30 -0700, Andrew Morton wrote:
> 
> > There used to be a cond_resched() in invalidate_mapping_pages() which would
> > have prevented this, but I rudely removed it to support
> > /proc/sys/vm/drop_caches (which needs to call invalidate_inode_pages()
> > under spinlock).
> > 
> > We could resurrect that cond_resched() by passing in some flag, I guess. 
> > Or change the code to poke the softlockup detector.  The former would be
> > better.
> 
> cond_resched() is conditional on __resched_legal(0)

What's __resched_legal()?

> which should take
> care of being called under a spinlock.

We only increment preempt_count() in spin_lock() if CONFIG_PREEMPT.

> so I guess we can just reinstate the call in invalidate_mapping_pages()
> 
> (still waiting on the compile to finish...)
> ---
> invalidate_mapping_pages() is called under locks (usually preemptable)
> but can do a _lot_ of work, stick in a voluntary preemption point to
> avoid excessive latencies (over 10 seconds was reported by softlockup).
> 
> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> ---
>  mm/truncate.c |    2 ++
>  1 file changed, 2 insertions(+)
> 
> Index: linux-2.6-mm/mm/truncate.c
> ===================================================================
> --- linux-2.6-mm.orig/mm/truncate.c
> +++ linux-2.6-mm/mm/truncate.c
> @@ -292,6 +292,8 @@ unsigned long invalidate_mapping_pages(s
>  			pgoff_t index;
>  			int lock_failed;
>  
> +			cond_resched();
> +
>  			lock_failed = TestSetPageLocked(page);

Is deadlocky on the drop_caches path and if CONFIG_PREEMPT we'll get
scheduling-in-spinlock warnings.

For the blkdev_close() path the change is unneeded if CONFIG_PREEMPT and
will fix things if !CONFIG_PREEMPT.

We can presumably just remove the invalidate_mapping_pages() call from the
kill_bdev() path (at least) - kill_bdev()'s truncate_inode_pages() will do
the same thing.

It might be time to remove that unused-for-six-years destroy_dirty_buffers
too.