On Tue, Jun 30, 2009 at 10:16:57PM +0200, Paweł Staszewski wrote:

Jarek Poplawski pisze:

quoted

On Mon, Jun 29, 2009 at 10:47:03AM +0000, Jarek Poplawski wrote:
  

quoted

On Mon, Jun 29, 2009 at 11:51:52AM +0200, Paweł Staszewski wrote:
    

quoted

I apply this patch

fib_triestats in attached file :)
      

Great! But it would be nice to check if this (accidentally ;-) might
fix the previous problem, so I attach below the patch with "manual
RCU", which btw. (or even more important) should verify RCU use here.

It should be applied on top of this last "Fix..., part3". And
again: it's quite probable it can fail, so with caution, no hurry
(it can wait for quiet time)...
    

Pawel, here is another try to check what's going on here, so just
like before, but this one on top of these 2 last working patches,
plus quite time... (Stats aren't necessary; if these are some doubts
let me know.)

Thanks,
Jarek P.
--------------------> (synchronize_rcu take 5)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

--- a/net/ipv4/fib_trie.c	2009-06-29 10:00:14.000000000 +0000
+++ b/net/ipv4/fib_trie.c	2009-06-30 06:50:35.000000000 +0000

@@ -1036,6 +1036,7 @@ static void trie_rebalance(struct trie *
  	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
+	synchronize_rcu();
  	return;
 }
  

Apply and tested

Traffic is not forwarded after apply this patch.:)

A little comment: these last 2 patches weren't exactly to fix the
problem you reported, which should be mostly fixed by the earlier
patch.

There is some other bug, which you omit with CONFIG_PREEMPT_NONE
(but it's not for sure there is no by effects). So, I'd like to be
sure you are willing and can (without too much risk) to do more such
tests. Alas I've no way to generate similar conditions so it would
simply have to wait for somebody else.

Many thanks again,
Jarek P.