Re: veth crash (commit 751ae21c6cd1493e3d0a4935b08fb298b9d89773)

[David Gibson <dwg@au1.ibm.com>]

On Thu, Oct 12, 2006 at 06:22:14PM +1000, David Gibson wrote:
> Your recent ibmveth commit, 751ae21c6cd1493e3d0a4935b08fb298b9d89773
> ("fix int rollover panic"), causes a rapid oops on my test machine
> (POWER5 LPAR).
> 
> I've bisected it down to that commit, but am still investigating the
> cause of the crash itself.

Found the problem, I believe: an object lesson in the need for great
caution using ++.

[...]
@@ -213,6 +213,7 @@ static void ibmveth_replenish_buffer_poo
 		}
 
 		free_index = pool->consumer_index++ % pool->size;
+		pool->consumer_index = free_index;
 		index = pool->free_map[free_index];
 
 		ibmveth_assert(index != IBM_VETH_INVALID_MAP);

Since the ++ is used as post-increment, the increment is not included
in free_index, and so the added line effectively reverts the
increment.  The produced_index side has an analagous bug.

Jeff, Santiago, please apply the patch below to correct this:

ibmveth: Fix index increment calculation

The recent commit 751ae21c6cd1493e3d0a4935b08fb298b9d89773 introduced
a bug in the producer/consumer index calculation in the ibmveth driver
- incautious use of the post-increment ++ operator resulted in an
increment being immediately reverted.  This patch corrects the logic.

Without this patch, the driver oopses almost immediately after
activation on at least some machines.

Signed-off-by: David Gibson <dwg@au1.ibm.com>

Index: working-2.6/drivers/net/ibmveth.c
===================================================================
--- working-2.6.orig/drivers/net/ibmveth.c	2006-10-13 14:12:49.000000000 +1000
+++ working-2.6/drivers/net/ibmveth.c	2006-10-13 14:14:24.000000000 +1000
@@ -212,8 +212,8 @@ static void ibmveth_replenish_buffer_poo
 			break;
 		}
 
-		free_index = pool->consumer_index++ % pool->size;
-		pool->consumer_index = free_index;
+		free_index = pool->consumer_index;
+		pool->consumer_index = (pool->consumer_index + 1) % pool->size;
 		index = pool->free_map[free_index];
 
 		ibmveth_assert(index != IBM_VETH_INVALID_MAP);
@@ -329,8 +329,10 @@ static void ibmveth_remove_buffer_from_p
 			 adapter->rx_buff_pool[pool].buff_size,
 			 DMA_FROM_DEVICE);
 
-	free_index = adapter->rx_buff_pool[pool].producer_index++ % adapter->rx_buff_pool[pool].size;
-	adapter->rx_buff_pool[pool].producer_index = free_index;
+	free_index = adapter->rx_buff_pool[pool].producer_index;
+	adapter->rx_buff_pool[pool].producer_index
+		= (adapter->rx_buff_pool[pool].producer_index + 1)
+		% adapter->rx_buff_pool[pool].size;
 	adapter->rx_buff_pool[pool].free_map[free_index] = index;
 
 	mb();


-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson