[ofa-general] Re: [PATCH 08 of 11] anon-vma-rwsem

[Linus Torvalds]

On Wed, 7 May 2008, Andrew Morton wrote:
> 
> Now, if we need to take both anon_vma->lock AND i_mmap_lock in the newly
> added mm_lock() thing and we also take both those locks at the same time in
> regular code, we're probably screwed.

No, just use the normal static ordering for that case: one type of lock 
goes before the other kind. If those locks nest in regular code, you have 
to do that *anyway*.

The code that can take many locks, will have to get the global lock *and* 
order the types, but that's still trivial. It's something like

	spin_lock(&global_lock);
	for (vma = mm->mmap; vma; vma = vma->vm_next) {
		if (vma->anon_vma)
			spin_lock(&vma->anon_vma->lock);
	}
	for (vma = mm->mmap; vma; vma = vma->vm_next) {
		if (!vma->anon_vma && vma->vm_file && vma->vm_file->f_mapping)
			spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
	}
	spin_unlock(&global_lock);

and now everybody follows the rule that "anon_vma->lock" precedes 
"i_mmap_lock". So there can be no ABBA deadlock between the normal users 
and the many-locks version, and there can be no ABBA deadlock between 
many-locks-takers because they use the global_lock to serialize.

This really isn't rocket science, guys.

(I really hope and believe that they don't nest anyway, and that you can 
just use a single for-loop for the many-lock case)

		Linus