mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
From: Julius Bairaktaris <julius@bairaktaris.de>
To: Johannes Berg <johannes@sipsolutions.net>
Cc: linux-wireless@vger.kernel.org,
	"Jeff Johnson" <jjohnson@kernel.org>,
	"Felix Fietkau" <nbd@nbd.name>,
	"Toke Høiland-Jørgensen" <toke@toke.dk>,
	ath11k@lists.infradead.org, ath12k@lists.infradead.org,
	linux-kernel@vger.kernel.org
Subject: [PATCH wireless-next 1/4] wifi: mac80211: advertise TSO and SG for encap offloaded frames
Date: Sun, 22 Nov 2026 14:42:25 +0100	[thread overview]
Message-ID: <20261122134228.956044-2-julius@bairaktaris.de> (raw)
In-Reply-To: <20261122134228.956044-1-julius@bairaktaris.de>

A driver whose hardware segments TCP frames and takes frames with
fragments on its 802.3 encap path cannot say so. hw->netdev_features
applies to every netdev of the hw, so a feature set there also reaches
the frames that go out through the 802.11 path.

Add hw->encap_netdev_features and hw->encap_gso_max_segs. mac80211 sets
the features on the AP, AP_VLAN and station netdevs when the hw supports
encap offload, and ieee80211_8023_xmit() passes them to
ieee80211_tx_skb_fixup(), so a GSO frame reaches the driver whole. The
802.11 path keeps using the features of the vif, which do not include
them, and segments such a frame in software as before.

AQL charges a frame by tx_time_est, which ieee80211_info_set_tx_time_est()
caps at 4095 us. A whole GSO frame to a slow station can take longer and
would be charged less than its airtime. Segment such a frame in software
instead.

The tx statistics count a TCP GSO frame as gso_segs packets, with the
headers of every segment in its length.

Signed-off-by: Julius Bairaktaris <julius@bairaktaris.de>
---
 include/net/mac80211.h | 15 ++++++++++++
 net/mac80211/iface.c   | 14 +++++++++++
 net/mac80211/main.c    |  3 ++-
 net/mac80211/tx.c      | 53 ++++++++++++++++++++++++++++++++++++++++--
 4 files changed, 82 insertions(+), 3 deletions(-)

diff --git a/include/net/mac80211.h b/include/net/mac80211.h
index f768ef9184eb..8eac4d50d06c 100644
--- a/include/net/mac80211.h
+++ b/include/net/mac80211.h
@@ -3271,6 +3271,19 @@ enum ieee80211_hw_flags {
  * @netdev_features: netdev features to be set in each netdev created
  *	from this HW. Note that not all features are usable with mac80211,
  *	other features will be rejected during HW registration.
+ * @encap_netdev_features: netdev features, in addition to @netdev_features,
+ *	set in each AP, AP VLAN and station netdev created from this HW, for
+ *	the frames that are sent with 802.3 encapsulation offload (see
+ *	%IEEE80211_OFFLOAD_ENCAP_ENABLED). A driver sets scatter-gather and
+ *	segmentation offload here when the hardware takes such frames whole.
+ *	mac80211 hands a driver a frame with these features only with
+ *	%IEEE80211_TX_CTL_HW_80211_ENCAP, and segments or linearizes it in
+ *	software on every other path, and for a station when the airtime
+ *	estimate of the whole frame is more than AQL can account to one frame.
+ *	Such a frame starts with its 802.3 header; a GSO frame has at most
+ *	@encap_gso_max_segs segments.
+ * @encap_gso_max_segs: the most segments of a GSO frame the driver takes
+ *	whole, when @encap_netdev_features has a segmentation offload.
  *
  * @uapsd_queues: This bitmap is included in (re)association frame to indicate
  *	for each access category if it is uAPSD trigger-enabled and delivery-
@@ -3330,6 +3343,8 @@ struct ieee80211_hw {
 		s16 accuracy;
 	} radiotap_timestamp;
 	netdev_features_t netdev_features;
+	netdev_features_t encap_netdev_features;
+	u16 encap_gso_max_segs;
 	u8 uapsd_queues;
 	u8 uapsd_max_sp_len;
 	u8 max_nan_de_entries;
diff --git a/net/mac80211/iface.c b/net/mac80211/iface.c
index b6ed6541bc42..c3869dbfdab7 100644
--- a/net/mac80211/iface.c
+++ b/net/mac80211/iface.c
@@ -2407,6 +2407,20 @@ int ieee80211_if_add(struct ieee80211_local *local, const char *name,
 			sdata->u.mgd.use_4addr = params->use_4addr;
 
 		ndev->features |= local->hw.netdev_features;
+		/* Only the netdevs that can use the encap offload offer its
+		 * features: every other path segments in software, with the
+		 * features of the vif.
+		 */
+		if (ieee80211_hw_check(&local->hw, SUPPORTS_TX_ENCAP_OFFLOAD) &&
+		    local->hw.encap_netdev_features &&
+		    (type == NL80211_IFTYPE_AP ||
+		     type == NL80211_IFTYPE_AP_VLAN ||
+		     type == NL80211_IFTYPE_STATION)) {
+			ndev->features |= local->hw.encap_netdev_features;
+			if (local->hw.encap_gso_max_segs)
+				netif_set_tso_max_segs(ndev,
+						       local->hw.encap_gso_max_segs);
+		}
 		ndev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
 		ndev->hw_features |= ndev->features &
 					MAC80211_SUPPORTED_FEATURES_TX;
diff --git a/net/mac80211/main.c b/net/mac80211/main.c
index e03c2dbf4fb8..2a49fe7ed367 100644
--- a/net/mac80211/main.c
+++ b/net/mac80211/main.c
@@ -1243,7 +1243,8 @@ int ieee80211_register_hw(struct ieee80211_hw *hw)
 	}
 
 	/* Only HW csum features are currently compatible with mac80211 */
-	if (WARN_ON(hw->netdev_features & ~MAC80211_SUPPORTED_FEATURES))
+	if (WARN_ON((hw->netdev_features | hw->encap_netdev_features) &
+		    ~MAC80211_SUPPORTED_FEATURES))
 		return -EINVAL;
 
 	if (hw->max_report_rates == 0)
diff --git a/net/mac80211/tx.c b/net/mac80211/tx.c
index cb9ecac25a9c..faf3e6b9e4f8 100644
--- a/net/mac80211/tx.c
+++ b/net/mac80211/tx.c
@@ -14,6 +14,7 @@
 #include <linux/slab.h>
 #include <linux/skbuff.h>
 #include <linux/if_vlan.h>
+#include <linux/tcp.h>
 #include <linux/etherdevice.h>
 #include <linux/bitmap.h>
 #include <linux/rcupdate.h>
@@ -4797,6 +4798,35 @@ static bool ieee80211_tx_8023(struct ieee80211_sub_if_data *sdata,
 	return ret;
 }
 
+/* A frame's airtime estimate has 10 bits, in units of 4 us. A GSO frame whose
+ * estimate would not fit is charged less than it takes and stays in the AQL
+ * limits for as long as it takes, so it is segmented in software instead.
+ */
+#define IEEE80211_TX_GSO_AIRTIME_MAX	4092
+
+static bool ieee80211_8023_gso_fits(struct ieee80211_sub_if_data *sdata,
+				    struct sta_info *sta, struct sk_buff *skb,
+				    u16 queue)
+{
+	struct ieee80211_local *local = sdata->local;
+	u32 airtime;
+
+	if (!sta || !wiphy_ext_feature_isset(local->hw.wiphy,
+					     NL80211_EXT_FEATURE_AQL))
+		return true;
+
+	/* AQL charges the vif of the AP, as ieee80211_tx_dequeue() does */
+	if (sdata->vif.type == NL80211_IFTYPE_AP_VLAN)
+		sdata = container_of(sdata->bss, struct ieee80211_sub_if_data,
+				     u.ap);
+
+	airtime = ieee80211_calc_expected_tx_airtime(&local->hw, &sdata->vif,
+						     &sta->sta, skb->len,
+						     queue != IEEE80211_AC_VO);
+
+	return airtime <= IEEE80211_TX_GSO_AIRTIME_MAX;
+}
+
 static void ieee80211_8023_xmit(struct ieee80211_sub_if_data *sdata,
 				struct net_device *dev, struct sta_info *sta,
 				struct ieee80211_key *key, struct sk_buff *skb)
@@ -4807,6 +4837,7 @@ static void ieee80211_8023_xmit(struct ieee80211_sub_if_data *sdata,
 	struct tid_ampdu_tx *tid_tx = NULL;
 	struct sk_buff *seg, *next;
 	unsigned int skbs = 0, len = 0;
+	netdev_features_t features;
 	u16 queue;
 	u8 tid;
 
@@ -4835,7 +4866,12 @@ static void ieee80211_8023_xmit(struct ieee80211_sub_if_data *sdata,
 			tid_tx->last_tx = jiffies;
 	}
 
-	skb = ieee80211_tx_skb_fixup(skb, ieee80211_sdata_netdev_features(sdata));
+	features = ieee80211_sdata_netdev_features(sdata) |
+		   local->hw.encap_netdev_features;
+	if (skb_is_gso(skb) && !ieee80211_8023_gso_fits(sdata, sta, skb, queue))
+		features &= ~(NETIF_F_ALL_TSO | NETIF_F_SG);
+
+	skb = ieee80211_tx_skb_fixup(skb, features);
 	if (!skb)
 		return;
 
@@ -4865,8 +4901,21 @@ static void ieee80211_8023_xmit(struct ieee80211_sub_if_data *sdata,
 		info->control.hw_key = &key->conf;
 
 	skb_list_walk_safe(skb, seg, next) {
-		skbs++;
+		unsigned int segs = 1;
+
 		len += seg->len;
+		/* a GSO frame is one frame of the queues and of the airtime
+		 * estimate, and as many packets of the counters as it has
+		 * segments
+		 */
+		if (skb_is_gso(seg) &&
+		    (skb_shinfo(seg)->gso_type &
+		     (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
+			segs = max_t(unsigned int, skb_shinfo(seg)->gso_segs, 1);
+			len += (segs - 1) * (skb_transport_offset(seg) +
+					     tcp_hdrlen(seg));
+		}
+		skbs += segs;
 		if (seg != skb)
 			memcpy(IEEE80211_SKB_CB(seg), info, sizeof(*info));
 	}
-- 
2.53.0


  reply	other threads:[~2026-11-22 13:42 UTC|newest]

Thread overview: 5+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-22 13:42 [PATCH wireless-next 0/4] wifi: mac80211: TSO and SG for encap offload, and the TID of 802.3 RX frames Julius Bairaktaris
2026-11-22 13:42 ` Julius Bairaktaris [this message]
2026-11-22 13:42 ` [PATCH wireless-next 2/4] wifi: mac80211: take the TID of an 802.3 frame from the driver Julius Bairaktaris
2026-11-22 13:42 ` [PATCH wireless-next 3/4] wifi: ath11k: pass the TID of a received 802.3 frame to mac80211 Julius Bairaktaris
2026-11-22 13:42 ` [PATCH wireless-next 4/4] wifi: ath12k: " Julius Bairaktaris

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20261122134228.956044-2-julius@bairaktaris.de \
    --to=julius@bairaktaris.de \
    --cc=ath11k@lists.infradead.org \
    --cc=ath12k@lists.infradead.org \
    --cc=jjohnson@kernel.org \
    --cc=johannes@sipsolutions.net \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-wireless@vger.kernel.org \
    --cc=nbd@nbd.name \
    --cc=toke@toke.dk \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®