quoted

 static __always_inline void

@@ -116,7 +120,8 @@ struct xdp_frame {
 	u16 len;
 	u16 headroom;
 	u32 metasize:8;
-	u32 frame_sz:24;
+	u32 frame_sz:23;
+	u32 mb:1; /* xdp non-linear frame */
 	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
 	 * while mem info is valid on remote CPU.
 	 */

So, it seems that these bitfield's are the root-cause of the
performance regression.  Credit to Alexei whom wisely already point
this out[1] in V2 ;-)

[1] https://lore.kernel.org/netdev/20200904010705.jm6dnuyj3oq4cpjd@ast-mbp.dhcp.thefacebook.com/ (local)

yes, shame on me..yesterday I recalled email from Alexei debugging the issue
reported by Magnus.
In the current approach I am testing (not posted upstream yet) I reduced the
size of xdp_mem_info as proposed by Jesper in [0] and I added a flags field
in xdp_frame/xdp_buff we can use for multiple features (e.g. multi-buff or hw csum
hints). Doing so, running xdp_rxq_info sample on ixgbe 10Gbps NIC I do not have any
performance regressions for xdp_tx or xdp_drop. Same results have been reported by
Magnus off-list on i40e (we have a 1% regression on xdp_sock tests iiuc).
I will continue working on this.

Regards,
Lorenzo

[0] https://patchwork.kernel.org/project/netdevbpf/patch/20210409223801.104657-2-mcroce@linux.microsoft.com/

quoted

@@ -179,6 +184,7 @@ void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp)
 	xdp->data_end = frame->data + frame->len;
 	xdp->data_meta = frame->data - frame->metasize;
 	xdp->frame_sz = frame->frame_sz;
+	xdp->mb = frame->mb;
 }
 
 static inline

@@ -205,6 +211,7 @@ int xdp_update_frame_from_buff(struct xdp_buff *xdp,
 	xdp_frame->headroom = headroom - sizeof(*xdp_frame);
 	xdp_frame->metasize = metasize;
 	xdp_frame->frame_sz = xdp->frame_sz;
+	xdp_frame->mb = xdp->mb;
 
 	return 0;
 }

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer