ppp: enable TX scatter-gather

PPP channels using chan->direct_xmit prepend the PPP header to a skb and call dev_queue_xmit() directly. In this mode the skb does not need to be linear, but the PPP netdevice currently does not advertise scatter-gather features, causing unnecessary linearization and preventing GSO.

Enable NETIF_F_SG and NETIF_F_FRAGLIST on PPP devices. In case a linear buffer is required (PPP compression, multilink, and channels without direct_xmit), call skb_linearize() explicitly.

Signed-off-by Qingfang Deng <[email protected]>

---
v3 -> v4:
Always expose SG and FRAGLIST, and linearize the skb if necessary.
- https://lore.kernel.org/linux-ppp/[email protected]/

 drivers/net/ppp/ppp_generic.c | 30 +++++++++++++++++++++++++-----
 1 file changed, 25 insertions(+), 5 deletions(-)

drivers/net/ppp/ppp_generic.c +11 -3

index f9f0f16c41d1..f8814d7be6f1 100644
@@ -1641,6 +1641,8 @@ static void ppp_setup(struct net_device *dev)
 	dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST;
 	dev->priv_destructor = ppp_dev_priv_destructor;
 	dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS;
+	dev->features = NETIF_F_SG | NETIF_F_FRAGLIST;
+	dev->hw_features = dev->features;
 	netif_keep_dst(dev);
 }
 

@@ -1710,6 +1712,10 @@ pad_compress_skb(struct ppp *ppp, struct sk_buff *skb)
 		ppp->xcomp->comp_extra + ppp->dev->hard_header_len;
 	int compressor_skb_size = ppp->dev->mtu +
 		ppp->xcomp->comp_extra + PPP_HDRLEN;
+
+	if (skb_linearize(skb))
+		return NULL;
+
 	new_skb = alloc_skb(new_skb_size, GFP_ATOMIC);
 	if (!new_skb) {
 		if (net_ratelimit())

@@ -1797,6 +1803,10 @@ ppp_send_frame(struct ppp *ppp, struct sk_buff *skb)
 	case PPP_IP:
 		if (!ppp->vj || (ppp->flags & SC_COMP_TCP) == 0)
 			break;
+
+		if (skb_linearize(skb))
+			goto drop;
+
 		/* try to do VJ TCP header compression */
 		new_skb = alloc_skb(skb->len + ppp->dev->hard_header_len - 2,
 				    GFP_ATOMIC);

@@ -1894,19 +1904,26 @@ ppp_push(struct ppp *ppp)
 	}
 
 	if ((ppp->flags & SC_MULTILINK) == 0) {
+		struct ppp_channel *chan;
 		/* not doing multilink: send it down the first channel */
 		list = list->next;
 		pch = list_entry(list, struct channel, clist);
 
 		spin_lock(&pch->downl);
-		if (pch->chan) {
-			if (pch->chan->ops->start_xmit(pch->chan, skb))
-				ppp->xmit_pending = NULL;

drivers/net/ppp/ppp_generic.c +8 -0

 		}
+
+		if (chan->ops->start_xmit(chan, skb))
+			ppp->xmit_pending = NULL;
+
+out:
 		spin_unlock(&pch->downl);
 		return;
 	}

@@ -1991,6 +2008,8 @@ static int ppp_mp_explode(struct ppp *ppp, struct sk_buff *skb)
 		return 0; /* can't take now, leave it in xmit_pending */
 
 	/* Do protocol field compression */
+	if (skb_linearize(skb))
+		goto err_linearize;
 	p = skb->data;
 	len = skb->len;
 	if (*p == 0 && mp_protocol_compress) {

@@ -2149,6 +2168,7 @@ static int ppp_mp_explode(struct ppp *ppp, struct sk_buff *skb)
 
  noskb:
 	spin_unlock(&pch->downl);
+ err_linearize:
 	if (ppp->debug & 1)
 		netdev_err(ppp->dev, "PPP: no memory (fragment)\n");
 	++ppp->dev->stats.tx_errors;

I'm sorry for missing this point before, but AFAICS channels with chan->direct_xmit == 0 do not support GSO packets at all, and here such packets will be transmitted after linearization (but still as GSO).

I think that while transmitting over !chan->direct_xmit you additionally need to check for GSO packet. If ppp features are recomputed depending on the channels configuration, you could chose to drop such packets. Otherwise you will need to segment them before transmission, and that looks error prone in case of multilink.

/P

Quoted 9 lines · Paolo Abeni

I'm sorry for missing this point before, but AFAICS channels with chan->direct_xmit == 0 do not support GSO packets at all, and here such packets will be transmitted after linearization (but still as GSO).

I think that while transmitting over !chan->direct_xmit you additionally need to check for GSO packet. If ppp features are recomputed depending on the channels configuration, you could chose to drop such packets. Otherwise you will need to segment them before transmission, and that looks error prone in case of multilink.

I don't think that's true. Using the current features set (NETIF_F_SG | NETIF_F_FRAGLIST), the network core will only enable the generic software GSO (NETIF_F_GSO), and a GSO skb will be segmented in validate_xmit_skb before being passed to the PPP driver.