1. Patchew
  2. linux
  3. View series

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

Qingfang Deng posted 1 patch 2 weeks, 6 days ago
There is a newer version of this series
drivers/net/ppp/ppp_generic.c | 32 ++++++++++++++++++++++++++++++++
1 file changed, 32 insertions(+)
[PATCH net-next] ppp: enable TX scatter-gather
Posted by Qingfang Deng 2 weeks, 6 days ago 
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, and add .ndo_fix_features()
callback to conditionally disable them if a linear skb is required.
This is required to support PPPoE GSO.

Signed-off-by: Qingfang Deng <dqfext@gmail.com>
---
v1:
 Remove the test for SC_CCP_OPEN and instead test for xc_state changes.
 Link to RFC v2: https://lore.kernel.org/netdev/20250909012742.424771-1-dqfext@gmail.com/
RFC v2:
 Dynamically update netdev features with ndo_fix_features() callback.
 Link to RFC v1: https://lore.kernel.org/netdev/20250904021328.24329-1-dqfext@gmail.com/

 drivers/net/ppp/ppp_generic.c | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/drivers/net/ppp/ppp_generic.c b/drivers/net/ppp/ppp_generic.c
index f9f0f16c41d1..1132159a8b92 100644
--- a/drivers/net/ppp/ppp_generic.c
+++ b/drivers/net/ppp/ppp_generic.c
@@ -835,6 +835,10 @@ static long ppp_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 		ppp_unlock(ppp);
 		if (cflags & SC_CCP_OPEN)
 			ppp_ccp_closed(ppp);
+
+		rtnl_lock();
+		netdev_update_features(ppp->dev);
+		rtnl_unlock();
 		err = 0;
 		break;
 
@@ -1545,6 +1549,22 @@ ppp_get_stats64(struct net_device *dev, struct rtnl_link_stats64 *stats64)
 	dev_fetch_sw_netstats(stats64, dev->tstats);
 }
 
+static netdev_features_t
+ppp_fix_features(struct net_device *dev, netdev_features_t features)
+{
+	struct ppp *ppp = netdev_priv(dev);
+
+	ppp_xmit_lock(ppp);
+	/* Allow SG/FRAGLIST only when we have direct-xmit, and no compression
+	 * path that wants a linear skb.
+	 */
+	if (!(dev->priv_flags & IFF_NO_QUEUE) || ppp->xc_state ||
+	    ppp->flags & (SC_COMP_TCP | SC_CCP_UP))
+		features &= ~(NETIF_F_SG | NETIF_F_FRAGLIST);
+	ppp_xmit_unlock(ppp);
+	return features;
+}
+
 static int ppp_dev_init(struct net_device *dev)
 {
 	struct ppp *ppp;
@@ -1619,6 +1639,7 @@ static const struct net_device_ops ppp_netdev_ops = {
 	.ndo_start_xmit  = ppp_start_xmit,
 	.ndo_siocdevprivate = ppp_net_siocdevprivate,
 	.ndo_get_stats64 = ppp_get_stats64,
+	.ndo_fix_features = ppp_fix_features,
 	.ndo_fill_forward_path = ppp_fill_forward_path,
 };
 
@@ -1641,6 +1662,8 @@ 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);
 }
 
@@ -3081,6 +3104,9 @@ ppp_set_compress(struct ppp *ppp, struct ppp_option_data *data)
 				ocomp->comp_free(ostate);
 				module_put(ocomp->owner);
 			}
+			rtnl_lock();
+			netdev_update_features(ppp->dev);
+			rtnl_unlock();
 			err = 0;
 		} else
 			module_put(cp->owner);
@@ -3537,6 +3563,12 @@ ppp_connect_channel(struct channel *pch, int unit)
 	spin_unlock(&pch->upl);
  out:
 	mutex_unlock(&pn->all_ppp_mutex);
+	if (ret == 0) {
+		rtnl_lock();
+		netdev_update_features(ppp->dev);
+		rtnl_unlock();
+	}
+
 	return ret;
 }
 
-- 
2.43.0
Re: [PATCH net-next] ppp: enable TX scatter-gather
Posted by Jakub Kicinski 2 weeks, 2 days ago 
On Fri, 12 Sep 2025 17:59:27 +0800 Qingfang Deng wrote:
> 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, and add .ndo_fix_features()
> callback to conditionally disable them if a linear skb is required.
> This is required to support PPPoE GSO.

Seems a bit racy. We can't netdev_update_features() under the spin lock
so there's going to be a window of time where datapath will see new
state but netdev flags won't be cleared, yet?

We either need to add a explicit linearization check in the xmit path,
or always reset the flags to disabled before we start tweaking the
config and re-enable after config (tho the latter feels like a bit of 
a hack).
-- 
pw-bot: cr
Re: [PATCH net-next] ppp: enable TX scatter-gather
Posted by Qingfang Deng 2 weeks, 2 days ago 
Hi Jakub,

On Tue, Sep 16, 2025 at 9:10 AM Jakub Kicinski <kuba@kernel.org> wrote:
> Seems a bit racy. We can't netdev_update_features() under the spin lock
> so there's going to be a window of time where datapath will see new
> state but netdev flags won't be cleared, yet?
>
> We either need to add a explicit linearization check in the xmit path,
> or always reset the flags to disabled before we start tweaking the
> config and re-enable after config (tho the latter feels like a bit of
> a hack).

Can I modify dev->features directly under the spin lock (without
.ndo_fix_features) ?

> --
> pw-bot: cr
Re: [PATCH net-next] ppp: enable TX scatter-gather
Posted by Jakub Kicinski 2 weeks, 2 days ago 
On Tue, 16 Sep 2025 10:57:49 +0800 Qingfang Deng wrote:
> Hi Jakub,
> 
> On Tue, Sep 16, 2025 at 9:10 AM Jakub Kicinski <kuba@kernel.org> wrote:
> > Seems a bit racy. We can't netdev_update_features() under the spin lock
> > so there's going to be a window of time where datapath will see new
> > state but netdev flags won't be cleared, yet?
> >
> > We either need to add a explicit linearization check in the xmit path,
> > or always reset the flags to disabled before we start tweaking the
> > config and re-enable after config (tho the latter feels like a bit of
> > a hack).  
> 
> Can I modify dev->features directly under the spin lock (without
> .ndo_fix_features) ?

Hm, I'm not aware of a reason not to. You definitely need to hold
rtnl_lock, and call netdev_update_features() after.
Re: [PATCH net-next] ppp: enable TX scatter-gather
Posted by Qingfang Deng 2 weeks, 1 day ago 
On Tue, Sep 16, 2025 at 10:57 PM Jakub Kicinski <kuba@kernel.org> wrote:
>
> On Tue, 16 Sep 2025 10:57:49 +0800 Qingfang Deng wrote:
> > Can I modify dev->features directly under the spin lock (without
> > .ndo_fix_features) ?
>
> Hm, I'm not aware of a reason not to. You definitely need to hold
> rtnl_lock, and call netdev_update_features() after.

Will the modification race against __netdev_update_features(), where
dev->features is assigned a new value?
Re: [PATCH net-next] ppp: enable TX scatter-gather
Posted by Jakub Kicinski 2 weeks ago 
On Wed, 17 Sep 2025 19:00:16 +0800 Qingfang Deng wrote:
> On Tue, Sep 16, 2025 at 10:57 PM Jakub Kicinski <kuba@kernel.org> wrote:
> >
> > On Tue, 16 Sep 2025 10:57:49 +0800 Qingfang Deng wrote:  
> > > Can I modify dev->features directly under the spin lock (without
> > > .ndo_fix_features) ?  
> >
> > Hm, I'm not aware of a reason not to. You definitely need to hold
> > rtnl_lock, and call netdev_update_features() after.  
> 
> Will the modification race against __netdev_update_features(), where
> dev->features is assigned a new value?

Shouldn't race if we're holding rtnl_lock when we make the modification
and until we call netdev_update_features()? I'm just spitballing tho,
haven't studied the code.

Patchew 2.1-devel-b67e4c2

© 2016 - 2025 Red Hat, Inc.