Re: [PATCH net-next v2 5/9] xsk: add xsk_alloc_batch_skb() to build skbs in batch

From: Jason Xing <redacted>
Date: Thu, 28 Aug 2025 08:38:42 +0800

On Wed, Aug 27, 2025 at 10:33 PM Alexander Lobakin
[off-list ref] wrote:

quoted

From: Jason Xing <redacted>
Date: Mon, 25 Aug 2025 21:53:38 +0800

quoted

From: Jason Xing <kernelxing@tencent.com>

Support allocating and building skbs in batch.

[...]

quoted

+     base_len = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
+     if (!(dev->priv_flags & IFF_TX_SKB_NO_LINEAR))
+             base_len += dev->needed_tailroom;
+
+     if (xs->skb_count >= nb_pkts)
+             goto build;
+
+     if (xs->skb) {
+             i = 1;
+             xs->skb_count++;
+     }
+
+     xs->skb_count += kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
+                                            gfp_mask, nb_pkts - xs->skb_count,
+                                            (void **)&skbs[xs->skb_count]);

Have you tried napi_skb_cache_get_bulk()? Depending on the workload, it
may give better perf numbers.

Sure, my initial try is using this interface. But later I want to see
a standalone cache belonging to xsk. The whole xsk_alloc_batch_skb
function I added is quite similar to napi_skb_cache_get_bulk(), to
some extent.

And if using napi_xxx(), we need a lock to avoid the race between this
context and softirq context on the same core.

Are you saying this particular function is not run in the softirq
context? I thought all Tx is done in BH.
If it's not BH, then ignore my suggestion -- napi_skb_cache_get_bulk()
requires BH, that's true.

Thanks,
Jason

Thanks,
Olek