netdev - [PATCH bpf-next RFCv3 4/6] veth: add zero-copy AF

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-Id: <1545856073-8680-5-git-send-email-u9012063@gmail.com>
Date:   Wed, 26 Dec 2018 12:27:51 -0800
From:   William Tu <u9012063@...il.com>
To:     bjorn.topel@...il.com, magnus.karlsson@...il.com, ast@...nel.org,
        daniel@...earbox.net, netdev@...r.kernel.org,
        makita.toshiaki@....ntt.co.jp, yihung.wei@...il.com,
        magnus.karlsson@...el.com
Subject: [PATCH bpf-next RFCv3 4/6] veth: add zero-copy AF_XDP TX support.

Remove the extra copy when doing AF_XDP TX.  The xdp frame comes
directly from the umem element and passes to the receiving logic.
Also, only depending on async_xmit to kick napi poll isn't fast
enough. So re-schedule the napi at the end of poll so the ksoftirqd
can keep processing the packets.  The performance increases from
1.1Mpps to 1.4Mpps, when running zero copy xdpsock as sender and
XDP_DROP at the receiver side.

Signed-off-by: William Tu <u9012063@...il.com>
---
 drivers/net/veth.c | 41 ++++++++++++++++++-----------------------
 1 file changed, 18 insertions(+), 23 deletions(-)

diff --git a/drivers/net/veth.c b/drivers/net/veth.c
index 10cf9ded59f1..551444195398 100644
--- a/drivers/net/veth.c
+++ b/drivers/net/veth.c
@@ -761,45 +761,34 @@ static int veth_xsk_poll(struct napi_struct *napi, int budget)
 	while (peer_rq->xsk_umem && budget--) {
 		unsigned int inner_xdp_xmit = 0;
 		unsigned int metasize = 0;
-		struct xdp_frame *xdpf;
+		struct xdp_frame xdpf;
 		bool dropped = false;
 		struct sk_buff *skb;
 		struct page *page;
 		void *vaddr;
-		void *addr;
 		u32 len;
 
 		if (!xsk_umem_consume_tx_virtual(peer_rq->xsk_umem, &vaddr, &len))
 			break;
 
-		page = dev_alloc_page();
-		if (!page) {
-			xsk_umem_complete_tx(peer_rq->xsk_umem, 1);
-			xsk_umem_consume_tx_done(peer_rq->xsk_umem);
-			return -ENOMEM;
-		}
-
-		addr = page_to_virt(page);
-		xdpf = addr;
-		memset(xdpf, 0, sizeof(*xdpf));
-
-		addr += sizeof(*xdpf);
-		memcpy(addr, vaddr, len);
+		xdpf.data = vaddr + metasize;
+		xdpf.len = len;
+		xdpf.headroom = 0;
+		xdpf.metasize = metasize;
+		xdpf.mem.type = MEM_TYPE_ZERO_COPY_VDEV;
 
-		xdpf->data = addr + metasize;
-		xdpf->len = len;
-		xdpf->headroom = 0;
-		xdpf->metasize = metasize;
-		xdpf->mem.type = MEM_TYPE_PAGE_SHARED;
+		page = virt_to_head_page(vaddr);
+		if (page->mem_cgroup)
+			page->mem_cgroup = NULL;
 
 		/* put into rq */
-		skb = veth_xdp_rcv_one(rq, xdpf, &inner_xdp_xmit);
+		skb = veth_xdp_rcv_one(rq, &xdpf, &inner_xdp_xmit);
 		if (!skb) {
 			/* Peer side has XDP program attached */
 			if (inner_xdp_xmit & VETH_XDP_TX) {
 				/* Not supported */
 				pr_warn("veth: peer XDP_TX not supported\n");
-				xdp_return_frame(xdpf);
+				xdp_return_frame(&xdpf);
 				dropped = true;
 				goto skip_tx;
 			} else if (inner_xdp_xmit & VETH_XDP_REDIR) {
@@ -808,7 +797,8 @@ static int veth_xsk_poll(struct napi_struct *napi, int budget)
 				dropped = true;
 			}
 		} else {
-			napi_gro_receive(&rq->xdp_napi, skb);
+			napi_gro_receive(&rq->xdp_napi, skb_copy(skb, GFP_KERNEL));
+			kfree(skb);
 		}
 skip_tx:
 		xsk_umem_complete_tx(peer_rq->xsk_umem, 1);
@@ -856,6 +846,11 @@ static int veth_poll(struct napi_struct *napi, int budget)
 		xdp_do_flush_map();
 	xdp_clear_return_frame_no_direct();
 
+	/* schedule again so the CPU can keep receiving
+	 * at higher rate
+	 */
+	napi_schedule(&rq->xdp_napi);
+
 	return done > budget ? budget : done;
 }
 
-- 
2.7.4