On Thu, 2011-01-20 at 11:57 -0800, Hugh Dickins wrote:

quoted

quoted

quoted

21/21 mm-optimize_page_lock_anon_vma_fast-path.patch
      I certainly see the call for this patch, I want to eliminate those
      doubled atomics too.  This appears correct to me, and I've not dreamt
      up an alternative; but I do dislike it, and I suspect you don't like
      it much either.  I'm ambivalent about it, would love a better patch.

Like said, I fully agree with that sentiment, just haven't been able to
come up with anything saner :/ Although I can optimize the
__put_anon_vma() path a bit by doing something like:

  if (mutex_is_locked()) { anon_vma_lock(); anon_vma_unlock(); }

But I bet that wants a barrier someplace and my head hurts.. 

Without daring to hurt my head very much, yes, I'd say those kind
of "optimizations" have a habit of turning out to be racily wrong.

But you put your finger on it: if you hadn't had to add that lock-
unlock pair into __put_anon_vma(), I wouldn't have minded the
contortions added to page_lock_anon_vma(). 

I think there's just about enough implied barriers there that the
'simple' code just works ;-)

But given that I'm trying to think with snot for brains thanks to some
cold, I don't trust myself at all to have gotten this right.

[ for Oleg and Paul: https://lkml.org/lkml/2010/11/26/213 contains the
full patch this is against ]

---
Index: linux-2.6/mm/rmap.c
===================================================================

--- linux-2.6.orig/mm/rmap.c
+++ linux-2.6/mm/rmap.c

@@ -1559,9 +1559,20 @@ void __put_anon_vma(struct anon_vma *ano
 	 * Synchronize against page_lock_anon_vma() such that
 	 * we can safely hold the lock without the anon_vma getting
 	 * freed.
+	 *
+	 * Relies on the full mb implied by the atomic_dec_and_test() from
+	 * put_anon_vma() against the full mb implied by mutex_trylock() from
+	 * page_lock_anon_vma(). This orders:
+	 *
+	 * page_lock_anon_vma()		VS	put_anon_vma()
+	 *   mutex_trylock()			  atomic_dec_and_test()
+	 *   smp_mb()				  smp_mb()
+	 *   atomic_read()			  mutex_is_locked()
 	 */
-	anon_vma_lock(anon_vma);
-	anon_vma_unlock(anon_vma);
+	if (mutex_is_locked(&anon_vma->root->mutex)) {
+		anon_vma_lock(anon_vma);
+		anon_vma_unlock(anon_vma);
+	}
 
 	if (anon_vma->root != anon_vma)
 		put_anon_vma(anon_vma->root);