Re: [PATCH] rwsem: reduce spinlock contention in wakeup code path

[Ingo Molnar <mingo@kernel.org>]

* Peter Zijlstra <peterz@infradead.org> wrote:

> On Mon, Sep 30, 2013 at 09:13:52AM -0700, Linus Torvalds wrote:
>  
> > So unlike a lot of other "let's try to make our locking fancy" that I 
> > dislike because it tends to hide the fundamental problem of 
> > contention, the rwlock patches make me go "those actually _fix_ a 
> > fundamental problem".
> 
> So here I'm slightly disagreeing; fixing a fundamental problem would be 
> coming up a better anon_vma management that doesn't create such immense 
> chains.

So, I think the fundamental problem seems to be that when rwsems are 
applied to this usecase, they still don't perform as well as a primitive 
rwlock.

That means that when rwsems cause a context switch it is a loss, while an 
rwlock_t burning CPU time by looping around is a win. I'm not sure it's 
even about 'immense chains' - if those were true then context-switching 
should actually improve performance by allowing other work to continue 
while the heavy chains are processed.

Alas that's not what happens!

Or is AIM7 essentially triggering a single large lock? I doubt that's the 
case though.

> Its still the same lock, spinlock or not.
> 
> And regardless of if we keep anon_vma lock a rwsem or not; I think we 
> should merge those rwsem patches as they do improve the lock 
> implementation and the hard work has already been done.

That I mostly agree with, except that without a serious usecase do we have 
a guarantee that bugs in fancies queueing in rwsems gets ironed out?

Thanks,

	Ingo