Re: [GIT PULL] core kernel fixes

[Thomas Gleixner <tglx@linutronix.de>]

On Sat, 20 Jun 2009, Linus Torvalds wrote:
> On Sat, 20 Jun 2009, Linus Torvalds wrote:
> > 
> > On x86, the natural way to do what you want done is ONE SINGLE 
> > INSTRUCTION! As far as I can tell, the above crazy function is 100% 
> > equivalent to this:
> > 
> > 	asm __inline__("lock ; addl $0,%0":"+m" (*uaddr): :"memory", "cc");
> > 
> > which really makes me think that using "get_user_pages_fast()" for it is 
> > some truly crazy crap.
> 
> We could also take the opposite approach - knowing that this is called 
> only when the page doesn't exist, and just doing
> 
> 	down_read(mmap_sem)
> 	vma = find_vma(..)
> 	ret = VM_FAULT_ERROR;
> 	if (vma && vma->vm_start <= address)
> 		ret = handle_mm_fault(mm, vma, address, 1);
> 	up_read(mmap_sem);
> 	return (ret & VM_FAULT_ERROR) ? -EFAULT : 0;
> 
> or something like that. Again, that looks saner than using 
> get_user_pages() for this and then dropping the page.

Hmm. The main reason why we switched to get_user_pages_fast() in the
futex code is to avoid mmap_sem contention which was observed as a
real big performance problem especially with those horrible JavaVM
applications.

As a fallout of this we got rid of the private find_vma /
handle_mm_fault magic (as above) in the futex code which mm folks
frowned upon for quite a while. Unfortunately we got it wrong :(

I agree that in the fault path we might go back to the mmap_sem
version, but I want to avoid it when possible. At least we need to run
some of those horrible JavaVM apps to see whether it matters or not.

OTOH, I really wonder whether we can simplify the whole logic when we
keep the page reference for the atomic access to the user space
address, but I still have doubts about the life time rules of all
this. Also I need to find out how this would affect the fast path
optimization of private futexes where we avoid the lookup in the first
place.

Thanks,

	tglx