On 11/30/23 10:11, Lorenzo Bianconi wrote:

I like this is split out into a check function.

  {
  	struct sk_buff *skb = *pskb;
-	u32 act = XDP_DROP;
-
-	/* Reinjected packets coming from act_mirred or similar should
-	 * not get XDP generic processing.
-	 */
-	if (skb_is_redirected(skb))
-		return XDP_PASS;

(For other reviewers)
This reinjected check is moved further down.

+	int err;
  
-	/* XDP packets must be linear and must have sufficient headroom
-	 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
-	 * native XDP provides, thus we need to do it here as well.
+	/* XDP does not support fraglist so we need to linearize
+	 * the skb.
  	 */
-	if (skb_cloned(skb) || skb_is_nonlinear(skb) ||
-	    skb_headroom(skb) < XDP_PACKET_HEADROOM) {
+	if (skb_has_frag_list(skb) || !prog->aux->xdp_has_frags) {
  		int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
  		int troom = skb->tail + skb->data_len - skb->end;
  
  		/* In case we have to go down the path and also linearize,
  		 * then lets do the pskb_expand_head() work just once here.
  		 */
-		if (pskb_expand_head(skb,
-				     hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
-				     troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
-			goto do_drop;
-		if (skb_linearize(skb))
-			goto do_drop;
+		err = pskb_expand_head(skb,
+				       hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
+				       troom > 0 ? troom + 128 : 0, GFP_ATOMIC);
+		if (err)
+			return err;
+
+		err = skb_linearize(skb);
+		if (err)
+			return err;
+
+		return 0;
+	}
+
+	/* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM
+	 * bytes. This is the guarantee that also native XDP provides,
+	 * thus we need to do it here as well.
+	 */
+	if (skb_cloned(skb) || skb_shinfo(skb)->nr_frags ||

I though we could allow a SKB with skb_shinfo(skb)->nr_frags (that isn't
cloned or shared) to be processed by generic XDP without any reallocation?

So check would be: (skb_cloned(skb) || skb_shared(skb) ||)

+	    skb_headroom(skb) < XDP_PACKET_HEADROOM) {

[Headroom trick]
For layered devices the netstack could be the process that created the
SKB.  If you noticed my veth patchset[4/4], when I detect an XDP-prog
attach, I'm increasing the net_device headroom (.ndo_set_rx_headroom)
such that netstack will allocated enough headroom to satisfy
XDP_PACKET_HEADROOM.

[4/4] 
https://lore.kernel.org/netdev/169272716651.1975370.10514711233878278884.stgit@firesoul/ (local)

+		u32 mac_len = skb->data - skb_mac_header(skb);
+		u32 size, len, max_head_size, off;
+		struct sk_buff *nskb;
+		int i, head_off;
+
+		__skb_push(skb, mac_len);
+		max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE -
+						  XDP_PACKET_HEADROOM);
+		if (skb->len > max_head_size + MAX_SKB_FRAGS * PAGE_SIZE)
+			return -ENOMEM;
+
+		size = min_t(u32, skb->len, max_head_size);
+		nskb = netdev_alloc_skb(skb->dev, size + XDP_PACKET_HEADROOM);

Would is be possible to use napi_alloc_skb() here?

The napi_alloc_skb() is faster than netdev_alloc_skb(), but it as name
suggest assumes this is called under NAPI protection/context.  It
used-to-be the case for generic XDP, but code got moved around to
support layered devices, so I not 100% sure if this is always true (NAPI
context).

+		if (!nskb)
+			return -ENOMEM;
+
+		skb_reserve(nskb, XDP_PACKET_HEADROOM);
+		skb_copy_header(nskb, skb);
+
+		err = skb_copy_bits(skb, 0, nskb->data, size);
+		if (err) {
+			consume_skb(nskb);
+			return err;
+		}
+		skb_put(nskb, size);
+
+		head_off = skb_headroom(nskb) - skb_headroom(skb);
+		skb_headers_offset_update(nskb, head_off);
+
+		off = size;
+		len = skb->len - off;
+		for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) {
+			struct page *page;
+			void *frag;
+
+			size = min_t(u32, len, PAGE_SIZE);
+			frag = netdev_alloc_frag(size);

Again the slower variant.

+			if (!frag) {
+				consume_skb(nskb);
+				return -ENOMEM;
+			}
+
+			page = virt_to_head_page(frag);
+			skb_add_rx_frag(nskb, i, page,
+					frag - page_address(page), size, size);
+			err = skb_copy_bits(skb, off, frag, size);
+			if (err) {
+				consume_skb(nskb);
+				return err;
+			}
+
+			len -= size;
+			off += size;
+		}
+
+		consume_skb(skb);
+		*pskb = nskb;
+		__skb_pull(nskb, mac_len);
  	}
  
-	act = bpf_prog_run_generic_xdp(skb, xdp, xdp_prog);
+	return 0;
+}
+
+static u32 netif_receive_generic_xdp(struct sk_buff **pskb,
+				     struct xdp_buff *xdp,
+				     struct bpf_prog *xdp_prog)
+{
+	u32 act = XDP_DROP;
+
+	/* Reinjected packets coming from act_mirred or similar should
+	 * not get XDP generic processing.
+	 */
+	if (skb_is_redirected(*pskb))
+		return XDP_PASS;
+
+	if (netif_skb_check_for_generic_xdp(pskb, xdp_prog))
+		goto do_drop;
+
+	act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog);
  	switch (act) {
  	case XDP_REDIRECT:
  	case XDP_TX:
  	case XDP_PASS:
  		break;
  	default:
-		bpf_warn_invalid_xdp_action(skb->dev, xdp_prog, act);
+		bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act);
  		fallthrough;
  	case XDP_ABORTED:
-		trace_xdp_exception(skb->dev, xdp_prog, act);
+		trace_xdp_exception((*pskb)->dev, xdp_prog, act);
  		fallthrough;
  	case XDP_DROP:
  	do_drop:
-		kfree_skb(skb);
+		kfree_skb(*pskb);
  		break;
  	}
  

Overall I like the patch :-)

Are we missing more things to allow GRO packets getting processed by 
generic XDP?

--Jesper