Re: getting warn once around skb_try_coalesce

On Tue, 2012-07-10 at 12:18 +0200, Eric Dumazet wrote:

On Tue, 2012-07-10 at 12:54 +0300, Or Gerlitz wrote:

quoted

Hi Dave, Eric,

Another trace that I see here with net-next is this one-time warning. I 
get it always
on the passive side of TCP, something that seems related to GRO, it 
happens only with
IPoIB, not with mlx4_en and igb (when igb get to work on net-next...)

The latest commit in this area is bad43ca8325f493dcaa0896c2f036276af059c7e
"net: introduce skb_try_coalesce()" from Eric.

Or.

-----------[ cut here ]------------
WARNING: at net/core/skbuff.c:3413 skb_try_coalesce+0x1f8/0x31d()

This warning catch skb truesize offenders, most probably its a driver
issue.

By the way, this driver allocates not enough tailroom in skbs, so IP/TCP
stacks need to reallocate skb head to pull IP/TCP headers. Thats not
efficient.

I suggest using following patch :

diff --git a/drivers/infiniband/ulp/ipoib/ipoib_ib.c b/drivers/infiniband/ulp/ipoib/ipoib_ib.c
index 5c1bc99..9939869 100644
--- a/drivers/infiniband/ulp/ipoib/ipoib_ib.c
+++ b/drivers/infiniband/ulp/ipoib/ipoib_ib.c

@@ -159,7 +159,7 @@ static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
 	u64 *mapping;
 
 	if (ipoib_ud_need_sg(priv->max_ib_mtu))
-		buf_size = IPOIB_UD_HEAD_SIZE;
+		buf_size = IPOIB_UD_HEAD_SIZE + 128; /* reserve some tailroom for IP/TCP headers */
 	else
 		buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);