On Sun, May 31, 2020 at 11:46:49PM +0200, Lorenzo Bianconi wrote:

+
+		prog = READ_ONCE(rcpu->prog);
 		for (i = 0; i < n; i++) {
-			void *f = frames[i];
+			void *f = xdp_frames[i];
 			struct page *page = virt_to_page(f);
+			struct xdp_frame *xdpf;
+			struct xdp_buff xdp;
+			u32 act;
+			int err;
 
 			/* Bring struct page memory area to curr CPU. Read by
 			 * build_skb_around via page_is_pfmemalloc(), and when
 			 * freed written by page_frag_free call.
 			 */
 			prefetchw(page);
+			if (!prog) {
+				frames[nframes++] = xdp_frames[i];
+				continue;
+			}

I'm not sure compiler will be smart enough to hoist !prog check out of the loop.
Otherwise default cpumap case will be a bit slower.
I'd like to see performance numbers before/after and acks from folks
who are using cpumap before applying.
Also please add selftest for it. samples/bpf/ in patch 6 is not enough.

Other than the above the feature looks good to me. It nicely complements devmap.