[RFC PATCH net-next] ppp: enable TX scatter-gather

[RFC PATCH net-next] ppp: enable TX scatter-gather

[Qingfang Deng]

When chan->direct_xmit is true, and no compressors are in use, PPP
prepends its header to a skb, and calls dev_queue_xmit directly. In this
mode the skb does not need to be linearized.
Enable NETIF_F_SG and NETIF_F_FRAGLIST if chan->direct_xmit is true, so
the networking core can transmit non-linear skbs directly. The
compressors still require a linear buffer so call skb_linearize() before
passing skb->data to them.
This is required to support PPPoE GSO.

Signed-off-by: Qingfang Deng <dqfext@gmail.com>
---
RFC:
 This depends on a pending fix:
  https://lore.kernel.org/netdev/20250903100726.269839-1-dqfext@gmail.com/
 There are also alternative approaches:
 - set SG and FRAGLIST unconditionally, and use skb_linearize()
   on !chan->direct_xmit paths.
 - don't use skb_linearize(), instead fix the compressors to handle
   non-linear sk_buffs.
 - conditionally set SG and FRAGLIST based on whether compressors are
   in use.

 drivers/net/ppp/ppp_generic.c | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ppp/ppp_generic.c b/drivers/net/ppp/ppp_generic.c
index f9f0f16c41d1..3bf37871a1aa 100644
--- a/drivers/net/ppp/ppp_generic.c
+++ b/drivers/net/ppp/ppp_generic.c
@@ -1710,6 +1710,12 @@ 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;
+	/* Until we fix the compressor need to make sure data portion is
+	 * linear.
+	 */
+	if (skb_linearize(skb))
+		return NULL;
+
 	new_skb = alloc_skb(new_skb_size, GFP_ATOMIC);
 	if (!new_skb) {
 		if (net_ratelimit())
@@ -1797,6 +1803,12 @@ ppp_send_frame(struct ppp *ppp, struct sk_buff *skb)
 	case PPP_IP:
 		if (!ppp->vj || (ppp->flags & SC_COMP_TCP) == 0)
 			break;
+		/* Until we fix the compressor need to make sure data portion
+		 * is linear.
+		 */
+		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);
@@ -3516,10 +3528,13 @@ ppp_connect_channel(struct channel *pch, int unit)
 		ret = -ENOTCONN;
 		goto outl;
 	}
-	if (pch->chan->direct_xmit)
+	if (pch->chan->direct_xmit) {
 		ppp->dev->priv_flags |= IFF_NO_QUEUE;
-	else
+		ppp->dev->features |= NETIF_F_SG | NETIF_F_FRAGLIST;
+	} else {
 		ppp->dev->priv_flags &= ~IFF_NO_QUEUE;
+		ppp->dev->features &= ~(NETIF_F_SG | NETIF_F_FRAGLIST);
+	}
 	spin_unlock_bh(&pch->downl);
 	if (pch->file.hdrlen > ppp->file.hdrlen)
 		ppp->file.hdrlen = pch->file.hdrlen;
-- 
2.43.0

Re: [RFC PATCH net-next] ppp: enable TX scatter-gather

[Andrew Lunn]

> diff --git a/drivers/net/ppp/ppp_generic.c b/drivers/net/ppp/ppp_generic.c
> index f9f0f16c41d1..3bf37871a1aa 100644
> --- a/drivers/net/ppp/ppp_generic.c
> +++ b/drivers/net/ppp/ppp_generic.c
> @@ -1710,6 +1710,12 @@ 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;
> +	/* Until we fix the compressor need to make sure data portion is
> +	 * linear.
> +	 */
> +	if (skb_linearize(skb))
> +		return NULL;

The word 'fix' suggest something is broken. I don't think that is
true, its just a limitation of the compressor. Please avoid 'fix'.

	Andrew