[ofa-general] Re: [patch 01/10] emm: mm_lock: Lock a process against reclaim

[Andrea Arcangeli <andrea@qumranet.com>]

On Fri, Apr 04, 2008 at 04:12:42PM -0700, Jeremy Fitzhardinge wrote:
> I think you can break this if() down a bit:
>
> 			if (!(vma->vm_file && vma->vm_file->f_mapping))
> 				continue;

It makes no difference at runtime, coding style preferences are quite
subjective.

> So this is an O(n^2) algorithm to take the i_mmap_locks from low to high 
> order?  A comment would be nice.  And O(n^2)?  Ouch.  How often is it 
> called?

It's called a single time when the mmu notifier is registered. It's a
very slow path of course. Any other approach to reduce the complexity
would require memory allocations and it would require
mmu_notifier_register to return -ENOMEM failure. It didn't seem worth
it.

> And is it necessary to mush lock and unlock together?  Unlock ordering 
> doesn't matter, so you should just be able to have a much simpler loop, no?

That avoids duplicating .text. Originally they were separated. unlock
can't be a simpler loop because I didn't reserve vm_flags bitflags to
do a single O(N) loop for unlock. If you do malloc+fork+munmap two
vmas will point to the same anon-vma lock, that's why the unlock isn't
simpler unless I mark what I locked with a vm_flags bitflag.