Messages in this thread | ![/](/images/icornerl.gif) | | Date | Wed, 28 Apr 2010 14:09:59 -0400 | From | Rik van Riel <> | Subject | Re: [RFC PATCH] take all anon_vma locks in anon_vma_lock |
| |
On 04/28/2010 02:03 PM, Andrea Arcangeli wrote: > On Wed, Apr 28, 2010 at 01:47:19PM -0400, Rik van Riel wrote: >> static inline void anon_vma_unlock(struct vm_area_struct *vma) > > never mind as this is RFC, lock is clear enough > >> @@ -1762,7 +1760,8 @@ static int expand_downwards(struct vm_area_struct *vma, >> if (error) >> return error; >> >> - anon_vma_lock(vma); >> + spin_lock(&mm->page_table_lock); >> + anon_vma_lock(vma,&mm->page_table_lock); > > This will cause a lock inversion (page_table_lock can only be taken > after the anon_vma lock). I don't immediately see why the > page_table_lock here though?
We need to safely walk the vma->anon_vma_chain / anon_vma_chain->same_vma list.
So much for using the mmap_sem for read + the page_table_lock to lock the anon_vma_chain list.
We'll need a new lock somewhere, probably in the mm_struct since one per process seems plenty.
I'll add that in the next version of the patch.
-- All rights reversed
| ![\](/images/icornerr.gif) |