[PATCH] ibmveth: Fix index increment calculation
authorDavid Gibson <dwg@au1.ibm.com>
Fri, 13 Oct 2006 04:20:59 +0000 (14:20 +1000)
committerJeff Garzik <jeff@garzik.org>
Sat, 21 Oct 2006 18:34:21 +0000 (14:34 -0400)
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.

The following change corrects this:

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>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
drivers/net/ibmveth.c

index 2802db2..44c9f99 100644 (file)
@@ -212,8 +212,8 @@ static void ibmveth_replenish_buffer_pool(struct ibmveth_adapter *adapter, struc
                        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_pool(struct ibmveth_adapter *adapter, u64
                         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();