Re: [Bugme-new] [Bug 38102] New: BUG kmalloc-2048: Poison overwritten

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mardi 05 juillet 2011 à 14:06 -0400, Neil Horman a écrit :
> On Tue, Jul 05, 2011 at 06:47:21PM +0200, Eric Dumazet wrote:
> > Le mardi 05 juillet 2011 à 12:42 -0400, Neil Horman a écrit :
> > > On Tue, Jul 05, 2011 at 06:12:32PM +0200, Eric Dumazet wrote:
> > 
> > > > So all descriptors before prod are guaranteed to be ready for host
> > > > consume... Fact that a dma access is running on 'next descriptor' should
> > > > be irrelevant.
> > > > 
> > > But we handle more than one descriptor per b44_rx call - theres a while loop in
> > > there where we do advance to the next descriptor.
> > 
> > Yes, but we advance up to 'prod', which is the very last safe
> > descriptor.
> > 
> > If hardware advertises descriptor X being ready to be handled by host,
> > while DMA on this X descriptor is not yet finished, this would be a
> > really useless hardware ;)
> > 
> > 
> > 
> > --
> > To unsubscribe from this list: send the line "unsubscribe netdev" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > 
> 
> 
> Something else just jumped out at me.  During b44_open, we call b44_init_rings.
> This function allocates bp->rx_pending skb's and iteratively puts them in the rx
> dma ring. bp->rx_pending is initalized to B44_DEF_RX_RING_PENDING, which is
> defined as 200 (just about half of the 512 entries that the dma ring actually
> supports in the hardware.  This is normally ok, as subsequent calls to
> b44_alloc_rx_skb will fill in entries in the ring as those skbs are consumed.
> The problem with this however is that b44_alloc_rx_skb only sets the
> DESC_CTRL_EOT bit in the descriptor of the 512th entry, indicating that the
> hardware should wrap around and reset the index counter.  If a large volume of
> traffic is pushed through the adapter early on after initalization, or if the
> cpu is busy during init, it would be possible that the ring buffer would fill up
> prior to having additional entries added to the ring, the result being that the
> dma engine would reach the end of the allocated descriptors, not see an EOT bit
> set, and continue on using unallocated descriptors.
> 
> Just a theory, but it would be interesting to see if the problem subsided if you
> ensured that you allocated  a full descriptor ring on b44_open
> Neil
>  
> diff --git a/drivers/net/b44.c b/drivers/net/b44.c
> index 3d247f3..1b58a7c 100644
> --- a/drivers/net/b44.c
> +++ b/drivers/net/b44.c
> @@ -57,7 +57,7 @@
>  #define B44_MAX_MTU			1500
>  
>  #define B44_RX_RING_SIZE		512
> -#define B44_DEF_RX_RING_PENDING		200
> +#define B44_DEF_RX_RING_PENDING		512
>  #define B44_RX_RING_BYTES	(sizeof(struct dma_desc) * \
>  				 B44_RX_RING_SIZE)
>  #define B44_TX_RING_SIZE		512

No

Please take time to read the driver again.

200 desc are setup, and NIC is not allowed to use more than 200 descs.

( B44_DMARX_PTR )

We carefuly advance this pointer after a new desc(s) is(are) setup