On Mon, 2015-06-15 at 13:40 +0000, Madalin-Cristian Bucur wrote:

quoted

-----Original Message-----
From: Eric Dumazet [mailto:eric.dumazet@gmail.com]

On Wed, 2015-04-29 at 17:56 +0300, Madalin Bucur wrote:

quoted

This introduces the Freescale Data Path Acceleration Architecture
(DPAA) Ethernet driver (dpaa_eth) that builds upon the DPAA QMan,
BMan, PAMU and FMan drivers to deliver Ethernet connectivity on
the Freescale DPAA QorIQ platforms.

...

quoted

+	/* We're going to store the skb backpointer at the beginning
+	 * of the data buffer, so we need a privately owned skb
+	 */
+
+	/* Code borrowed from skb_unshare(). */
+	if (skb_cloned(skb)) {
+		struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
+
+		/* Finally, create a contig FD from this skb */
+		skb_to_contig_fd(priv, skb, &fd, countptr, &offset);
+
+		kfree_skb(skb);
+		skb = nskb;
+		/* skb_copy() has now linearized the skbuff. */
+	}
+

You know that TCP packets are clones, right ?
This code is killing performance.

TCP allows the headers part being modified by a driver if needed.

You should use skb_header_cloned() instead.

Thank you, I'll address this. I plan to do something like this:

+       if (!nonlinear) {
+               /* We're going to store the skb backpointer at the beginning
+                * of the data buffer, so we need a privately owned skb
+                */
+
+               /* make sure skb is not shared, skb_cow_head() assumes it's not */
+               skb = skb_share_check(skb, GFP_ATOMIC);
+               if (!skb)
+                       goto enomem;
+
+               /* verify the skb head is not cloned */
+               if (skb_cow_head(skb, priv->tx_headroom))
+                       goto enomem;
+
+               nonlinear = skb_is_nonlinear(skb);
+       }

but I'm not sure the skb_share_check() is required on tx.
I'm also a bit puzzled by the aliasing between shared and cloned terms
(i.e. skb_unshare() could be named something like skb_unclone();
 the skb_share_check() not only checks but also unshares an skb so
 the already taken skb_unshare() name would probably fit too).

A driver can ask tx skbs to be not shared, even if pktgen is used with
"clone=X" on it.

dev->priv_flags         &= ~IFF_TX_SKB_SHARING;

This way, you can avoid skb_share_check() completely. Only pktgen will
care.

Most skb_share_check() calls in drivers/net are related to input paths
on virtual drivers.