Re: [PATCH 1/2] mm,migration: Prevent rmap_walk_[anon|ksm] seeing the wrong VMA information

[Minchan Kim]

On Thu, May 6, 2010 at 9:22 AM, Mel Gorman <mel@xxxxxxxxx> wrote:
> On Wed, May 05, 2010 at 11:02:25AM -0700, Linus Torvalds wrote:
>>
>>
>> On Wed, 5 May 2010, Mel Gorman wrote:
>> >
>> > If the same_vma list is properly ordered then maybe something like the
>> > following is allowed?
>>
>> Heh. This is the same logic I just sent out. However:
>>
>> > + Â anon_vma = page_rmapping(page);
>> > + Â if (!anon_vma)
>> > + Â Â Â Â Â return NULL;
>> > +
>> > + Â spin_lock(&anon_vma->lock);
>>
>> RCU should guarantee that this spin_lock() is valid, but:
>>
>> > + Â /*
>> > + Â Â* Get the oldest anon_vma on the list by depending on the ordering
>> > + Â Â* of the same_vma list setup by __page_set_anon_rmap
>> > + Â Â*/
>> > + Â avc = list_entry(&anon_vma->head, struct anon_vma_chain, same_anon_vma);
>>
>> We're not guaranteed that the 'anon_vma->head' list is non-empty.
>>
>> Somebody could have freed the list and the anon_vma and we have a stale
>> 'page->anon_vma' (that has just not been _released_ yet).
>>
>> And shouldn't that be 'list_first_entry'? Or &anon_vma->head.next?
>>
>> How did that line actually work for you? Or was it just a "it boots", but
>> no actual testing of the rmap walk?
>>
>
> This is what I just started testing on a 4-core machine. Lockdep didn't
> complain but there are two potential sources of badness in anon_vma_lock_root
> marked with XXX. The second is the most important because I can't see how the
> local and root anon_vma locks can be safely swapped - i.e. release local and
> get the root without the root disappearing. I haven't considered the other
> possibilities yet such as always locking the root anon_vma. Going to
> sleep on it.
>
> Any comments?

<snip>
> +/* Given an anon_vma, find the root of the chain, lock it and return the root */
> +struct anon_vma *anon_vma_lock_root(struct anon_vma *anon_vma)
> +{
> + Â Â Â struct anon_vma *root_anon_vma;
> + Â Â Â struct anon_vma_chain *avc, *root_avc;
> + Â Â Â struct vm_area_struct *vma;
> +
> + Â Â Â /* Lock the same_anon_vma list and make sure we are on a chain */
> + Â Â Â spin_lock(&anon_vma->lock);
> + Â Â Â if (list_empty(&anon_vma->head)) {
> + Â Â Â Â Â Â Â spin_unlock(&anon_vma->lock);
> + Â Â Â Â Â Â Â return NULL;
> + Â Â Â }
> +
> + Â Â Â /*
> + Â Â Â Â* Get the root anon_vma on the list by depending on the ordering
> + Â Â Â Â* of the same_vma list setup by __page_set_anon_rmap. Basically
> + Â Â Â Â* we are doing
> + Â Â Â Â*
> + Â Â Â Â* local anon_vma -> local vma -> deepest vma -> anon_vma
> + Â Â Â Â*/
> + Â Â Â avc = list_first_entry(&anon_vma->head, struct anon_vma_chain, same_anon_vma);

Dumb question.

I can't understand why we should use list_first_entry.

I looked over the code.
anon_vma_chain_link uses list_add_tail so I think that's right.
But anon_vma_prepare uses list_add. So it's not consistent.
How do we make sure list_first_entry returns deepest vma?

Sorry if I am missing.


-- 
Kind regards,
Minchan Kim
--
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/