On Tue, Aug 26, 2025 at 1:34 AM Stanislav Fomichev [off-list ref] wrote:

On 08/25, Jason Xing wrote:

quoted

From: Jason Xing <kernelxing@tencent.com>

Add batch xmit logic.

Only grabbing the lock and disable bottom half once and sent all
the aggregated packets in one loop.

Since previous patch puts descriptors in xs->skb_cache in a reversed
order, this patch sends each skb out from start to end when 'start' is
not smaller than 'end'.

Signed-off-by: Jason Xing <kernelxing@tencent.com>
---
 include/linux/netdevice.h |  3 +++
 net/core/dev.c            | 19 +++++++++++++++++++
 2 files changed, 22 insertions(+)

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 5e5de4b0a433..8e2688e3f2e4 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h

@@ -3352,6 +3352,9 @@ u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,

 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev);
 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id);
+int xsk_direct_xmit_batch(struct sk_buff **skbs, struct net_device *dev,
+                       struct netdev_queue *txq, int *cur,
+                       int start, int end);

 static inline int dev_queue_xmit(struct sk_buff *skb)
 {

diff --git a/net/core/dev.c b/net/core/dev.c
index 68dc47d7e700..a5a6b9a199e9 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -4742,6 +4742,25 @@ int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
 }
 EXPORT_SYMBOL(__dev_queue_xmit);

+int xsk_direct_xmit_batch(struct sk_buff **skbs, struct net_device *dev,
+                       struct netdev_queue *txq, int *cur,
+                       int start, int end)
+{
+     int ret = NETDEV_TX_BUSY;
+
+     local_bh_disable();
+     HARD_TX_LOCK(dev, txq, smp_processor_id());
+     for (*cur = start; *cur >= end; (*cur)--) {

skbs support chaining (via list member), any reason not to use that for
batching purposes?

Good point, let me dig into it :)

Thanks,
Jason