Re: [PATCH] lowmemorykiller: prevent multiple instances of low memory killer

[Oskar Andero <oskar.andero@sonymobile.com>]

On 22:00 Tue 16 Apr     , David Rientjes wrote:
> On Tue, 16 Apr 2013, Oskar Andero wrote:
> 
> > > > The comment in shrinker.h is misleading, not the source code.
> > > > do_shrinker_shrink() will fail for anything negative and 0.
> > > 
> > > The comment is correct.  The only acceptable negative return is -1.
> > > Look at the second time do_shrinker_shrink() is called from
> > > shrink_slab().
> > > 
> > >    283                  while (total_scan >= batch_size) {
> > >    284                          int nr_before;
> > >    285  
> > >    286                          nr_before = do_shrinker_shrink(shrinker, shrink, 0);
> > >    287                          shrink_ret = do_shrinker_shrink(shrinker, shrink,
> > >    288                                                          batch_size);
> > >    289                          if (shrink_ret == -1)
> > >    290                                  break;
> > >    291                          if (shrink_ret < nr_before)
> > >    292                                  ret += nr_before - shrink_ret;
> > >    293                          count_vm_events(SLABS_SCANNED, batch_size);
> > 
> > Yes, the comment is correct with what is implemented in the code, but
> > that doesn't mean the code is right. IMHO, relaying on magical numbers is highly
> > questionable coding style.
> > 
> 
> No, it's not.  This is controlled higher in shrink_slab() by this:
> 
> 	max_pass = do_shrinker_shrink(shrinker, shrink, 0);
> 	if (max_pass <= 0)
> 		continue;
> 

Sure, that looks ok, but that doesn't change the fact that line 289
above has a magical number and I guess that explains the comment:
> > >    289                          if (shrink_ret == -1)
> > >    290                                  break;

Just to be clear - this is not about lowmemkiller, but rather a generic
clean-up of shrinkers that is needed IMO.

> and your patch is implemented incorrectly, i.e. it does not return 
> LMK_BUSY if the spinlock is contended which needlessly recalls the 
> shrinker later.
> 
> You have a couple of options:
> 
>  - return -1 when the spinlock is contended immediately when
>    !sc->nr_to_scan (although it should really be a cmpxchg since a
>    spinlock isn't needed), or

I leave it to Snild to comment on the patch, but could you elaborate on why
you think cmpxchg is a better alternative than a spin_trylock? I just had a
brief look at the implementation for ARM and it looks like cmpxchg means
two unconditional memory barriers, whereas spin_trylock has one
conditional memory barrier. See arch/arm/include/asm/spinlock.h:
	if (tmp == 0) {
		smp_mb();
		return 1;
	} else {
		return 0;
	}

...and arch/arm/include/asm/cmpxchg.h:
	smp_mb();
	ret = __cmpxchg(ptr, old, new, size);
	smp_mb();

AFAIK a memory barrier is pretty costly.

-Oskar