Re: lockdep complaints in slab allocator

[Paul E. McKenney]

On Fri, Nov 20, 2009 at 04:17:40PM +0100, Peter Zijlstra wrote:
> On Fri, 2009-11-20 at 06:48 -0800, Paul E. McKenney wrote:
> > On Fri, Nov 20, 2009 at 01:05:58PM +0200, Pekka Enberg wrote:
> > > Peter Zijlstra kirjoitti:
> > >> On Fri, 2009-11-20 at 12:38 +0200, Pekka Enberg wrote:
> > >>>
> > >>> On Fri, Nov 20, 2009 at 11:25 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> 
> > >>> wrote:
> > >>>>  2) propagate the nesting information and user spin_lock_nested(), given
> > >>>> that slab is already a rat's nest, this won't make it any less obvious.
> > >>> spin_lock_nested() doesn't really help us here because there's a
> > >>> _real_ possibility of a recursive spin lock here, right? 
> > >> Well, I was working under the assumption that your analysis of it being
> > >> a false positive was right ;-)
> > >> I briefly tried to verify that, but got lost and gave up, at which point
> > >> I started looking for ways to annotate.
> > >
> > > Uh, ok, so apparently I was right after all. There's a comment in 
> > > free_block() above the slab_destroy() call that refers to the comment above 
> > > alloc_slabmgmt() function definition which explains it all.
> > >
> > > Long story short: ->slab_cachep never points to the same kmalloc cache 
> > > we're allocating or freeing from. Where do we need to put the 
> > > spin_lock_nested() annotation? Would it be enough to just use it in 
> > > cache_free_alien() for alien->lock or do we need it in cache_flusharray() 
> > > as well?
> > 
> > Hmmm...  If the nc->lock spinlocks are always from different slabs
> > (as alloc_slabmgmt()'s block comment claims), why not just give each
> > array_cache structure's lock its own struct lock_class_key?  They
> > are zero size unless you have lockdep enabled.
> 
> Because more classes:
> 
>  - takes more (static/limited) lockdep resources
> 
>  - make more chains, weakening lock dependency tracking
>    because it can no longer use the state observed in one branch
>    on state observed in another branch.
> 
> Suppose you have 3 locks and 2 classes, lock 1 and 2 part of class A and
> lock 3 of class B
> 
> Then if we observe 1 -> 3, and 3 -> 2, we'd see A->B and B->A, and go
> yell. Now if we split class A into two classes and these locks get into
> separate classes we loose that cycle.
> 
> Now in this case we want to break a cycle, so the above will be correct,
> but all resulting chains will be equivalent for 99% (with the one
> exception of this funny recursion case) wasting lots of resources and
> state matching opportunity.
> 
> Therefore it would be much better to use the _nested annotation if
> possible.

Got it, thank you for the explanation!!!

I will keep this in mind when reconsidering the RCU lockdep interactions.

							Thanx, Paul
--
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/