Merge branch 'bpf-metadata-direct-access'

Daniel Borkmann says:

====================
BPF metadata for direct access

This work enables generic transfer of metadata from XDP into skb,
meaning the packet has a flexible and programmable room for meta
data, which can later be used by BPF to set various skb members
when passing up the stack. For details, please see second patch.
Support has been implemented and tested with two drivers, and
should be straight forward to add to other drivers as well which
properly support head adjustment already.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnxt/bnxt_xdp.c

@@ -94,6 +94,7 @@ bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
 
 	xdp.data_hard_start = *data_ptr - offset;
 	xdp.data = *data_ptr;
+	xdp_set_data_meta_invalid(&xdp);
 	xdp.data_end = *data_ptr + *len;
 	orig_data = xdp.data;
 	mapping = rx_buf->mapping - bp->rx_dma_offset;

drivers/net/ethernet/cavium/thunder/nicvf_main.c

@@ -523,6 +523,7 @@ static inline bool nicvf_xdp_rx(struct nicvf *nic, struct bpf_prog *prog,
 
 	xdp.data_hard_start = page_address(page);
 	xdp.data = (void *)cpu_addr;
+	xdp_set_data_meta_invalid(&xdp);
 	xdp.data_end = xdp.data + len;
 	orig_data = xdp.data;
 

drivers/net/ethernet/intel/i40e/i40e_txrx.c

@@ -2107,6 +2107,7 @@ static int i40e_clean_rx_irq(struct i40e_ring *rx_ring, int budget)
 		if (!skb) {
 			xdp.data = page_address(rx_buffer->page) +
 				   rx_buffer->page_offset;
+			xdp_set_data_meta_invalid(&xdp);
 			xdp.data_hard_start = xdp.data -
 					      i40e_rx_offset(rx_ring);
 			xdp.data_end = xdp.data + size;

drivers/net/ethernet/intel/ixgbe/ixgbe_main.c

@@ -2133,6 +2133,21 @@ static struct sk_buff *ixgbe_construct_skb(struct ixgbe_ring *rx_ring,
 #if L1_CACHE_BYTES < 128
 	prefetch(xdp->data + L1_CACHE_BYTES);
 #endif
+	/* Note, we get here by enabling legacy-rx via:
+	 *
+	 *    ethtool --set-priv-flags <dev> legacy-rx on
+	 *
+	 * In this mode, we currently get 0 extra XDP headroom as
+	 * opposed to having legacy-rx off, where we process XDP
+	 * packets going to stack via ixgbe_build_skb(). The latter
+	 * provides us currently with 192 bytes of headroom.
+	 *
+	 * For ixgbe_construct_skb() mode it means that the
+	 * xdp->data_meta will always point to xdp->data, since
+	 * the helper cannot expand the head. Should this ever
+	 * change in future for legacy-rx mode on, then lets also
+	 * add xdp->data_meta handling here.
+	 */
 
 	/* allocate a skb to store the frags */
 	skb = napi_alloc_skb(&rx_ring->q_vector->napi, IXGBE_RX_HDR_SIZE);
@@ -2165,6 +2180,7 @@ static struct sk_buff *ixgbe_build_skb(struct ixgbe_ring *rx_ring,
 				       struct xdp_buff *xdp,
 				       union ixgbe_adv_rx_desc *rx_desc)
 {
+	unsigned int metasize = xdp->data - xdp->data_meta;
 #if (PAGE_SIZE < 8192)
 	unsigned int truesize = ixgbe_rx_pg_size(rx_ring) / 2;
 #else
@@ -2174,10 +2190,14 @@ static struct sk_buff *ixgbe_build_skb(struct ixgbe_ring *rx_ring,
 #endif
 	struct sk_buff *skb;
 
-	/* prefetch first cache line of first page */
-	prefetch(xdp->data);
+	/* Prefetch first cache line of first page. If xdp->data_meta
+	 * is unused, this points extactly as xdp->data, otherwise we
+	 * likely have a consumer accessing first few bytes of meta
+	 * data, and then actual data.
+	 */
+	prefetch(xdp->data_meta);
 #if L1_CACHE_BYTES < 128
-	prefetch(xdp->data + L1_CACHE_BYTES);
+	prefetch(xdp->data_meta + L1_CACHE_BYTES);
 #endif
 
 	/* build an skb to around the page buffer */
@@ -2188,6 +2208,8 @@ static struct sk_buff *ixgbe_build_skb(struct ixgbe_ring *rx_ring,
 	/* update pointers within the skb to store the data */
 	skb_reserve(skb, xdp->data - xdp->data_hard_start);
 	__skb_put(skb, xdp->data_end - xdp->data);
+	if (metasize)
+		skb_metadata_set(skb, metasize);
 
 	/* record DMA address if this is the start of a chain of buffers */
 	if (!ixgbe_test_staterr(rx_desc, IXGBE_RXD_STAT_EOP))
@@ -2326,6 +2348,7 @@ static int ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
 		if (!skb) {
 			xdp.data = page_address(rx_buffer->page) +
 				   rx_buffer->page_offset;
+			xdp.data_meta = xdp.data;
 			xdp.data_hard_start = xdp.data -
 					      ixgbe_rx_offset(rx_ring);
 			xdp.data_end = xdp.data + size;

drivers/net/ethernet/mellanox/mlx4/en_rx.c

@@ -762,6 +762,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 
 			xdp.data_hard_start = va - frags[0].page_offset;
 			xdp.data = va;
+			xdp_set_data_meta_invalid(&xdp);
 			xdp.data_end = xdp.data + length;
 			orig_data = xdp.data;
 

drivers/net/ethernet/mellanox/mlx5/core/en_rx.c

@@ -794,6 +794,7 @@ static inline int mlx5e_xdp_handle(struct mlx5e_rq *rq,
 		return false;
 
 	xdp.data = va + *rx_headroom;
+	xdp_set_data_meta_invalid(&xdp);
 	xdp.data_end = xdp.data + *len;
 	xdp.data_hard_start = va;
 

drivers/net/ethernet/netronome/nfp/nfp_net_common.c

@@ -1574,26 +1574,6 @@ nfp_net_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
 	return true;
 }
 
-static int nfp_net_run_xdp(struct bpf_prog *prog, void *data, void *hard_start,
-			   unsigned int *off, unsigned int *len)
-{
-	struct xdp_buff xdp;
-	void *orig_data;
-	int ret;
-
-	xdp.data_hard_start = hard_start;
-	xdp.data = data + *off;
-	xdp.data_end = data + *off + *len;
-
-	orig_data = xdp.data;
-	ret = bpf_prog_run_xdp(prog, &xdp);
-
-	*len -= xdp.data - orig_data;
-	*off += xdp.data - orig_data;
-
-	return ret;
-}
-
 /**
  * nfp_net_rx() - receive up to @budget packets on @rx_ring
  * @rx_ring:   RX ring to receive from
@@ -1629,6 +1609,7 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 		struct nfp_meta_parsed meta;
 		struct net_device *netdev;
 		dma_addr_t new_dma_addr;
+		u32 meta_len_xdp = 0;
 		void *new_frag;
 
 		idx = D_IDX(rx_ring, rx_ring->rd_p);
@@ -1707,16 +1688,24 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 
 		if (xdp_prog && !(rxd->rxd.flags & PCIE_DESC_RX_BPF &&
 				  dp->bpf_offload_xdp) && !meta.portid) {
+			void *orig_data = rxbuf->frag + pkt_off;
 			unsigned int dma_off;
-			void *hard_start;
+			struct xdp_buff xdp;
 			int act;
 
-			hard_start = rxbuf->frag + NFP_NET_RX_BUF_HEADROOM;
+			xdp.data_hard_start = rxbuf->frag + NFP_NET_RX_BUF_HEADROOM;
+			xdp.data = orig_data;
+			xdp.data_meta = orig_data;
+			xdp.data_end = orig_data + pkt_len;
+
+			act = bpf_prog_run_xdp(xdp_prog, &xdp);
+
+			pkt_len -= xdp.data - orig_data;
+			pkt_off += xdp.data - orig_data;
 
-			act = nfp_net_run_xdp(xdp_prog, rxbuf->frag, hard_start,
-					      &pkt_off, &pkt_len);
 			switch (act) {
 			case XDP_PASS:
+				meta_len_xdp = xdp.data - xdp.data_meta;
 				break;
 			case XDP_TX:
 				dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
@@ -1784,6 +1773,8 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 		if (rxd->rxd.flags & PCIE_DESC_RX_VLAN)
 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
 					       le16_to_cpu(rxd->rxd.vlan));
+		if (meta_len_xdp)
+			skb_metadata_set(skb, meta_len_xdp);
 
 		napi_gro_receive(&rx_ring->r_vec->napi, skb);
 	}

drivers/net/ethernet/qlogic/qede/qede_fp.c

@@ -1004,6 +1004,7 @@ static bool qede_rx_xdp(struct qede_dev *edev,
 
 	xdp.data_hard_start = page_address(bd->data);
 	xdp.data = xdp.data_hard_start + *data_offset;
+	xdp_set_data_meta_invalid(&xdp);
 	xdp.data_end = xdp.data + *len;
 
 	/* Queues always have a full reset currently, so for the time

drivers/net/tun.c

@@ -1468,6 +1468,7 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun,
 
 		xdp.data_hard_start = buf;
 		xdp.data = buf + pad;
+		xdp_set_data_meta_invalid(&xdp);
 		xdp.data_end = xdp.data + len;
 		orig_data = xdp.data;
 		act = bpf_prog_run_xdp(xdp_prog, &xdp);

drivers/net/virtio_net.c

@@ -554,6 +554,7 @@ static struct sk_buff *receive_small(struct net_device *dev,
 
 		xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
 		xdp.data = xdp.data_hard_start + xdp_headroom;
+		xdp_set_data_meta_invalid(&xdp);
 		xdp.data_end = xdp.data + len;
 		orig_data = xdp.data;
 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
@@ -686,6 +687,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 		data = page_address(xdp_page) + offset;
 		xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
 		xdp.data = data + vi->hdr_len;
+		xdp_set_data_meta_invalid(&xdp);
 		xdp.data_end = xdp.data + (len - vi->hdr_len);
 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
 

include/linux/bpf.h

@@ -137,6 +137,7 @@ enum bpf_reg_type {
 	PTR_TO_MAP_VALUE,	 /* reg points to map element value */
 	PTR_TO_MAP_VALUE_OR_NULL,/* points to map elem value or NULL */
 	PTR_TO_STACK,		 /* reg == frame_pointer + offset */
+	PTR_TO_PACKET_META,	 /* skb->data - meta_len */
 	PTR_TO_PACKET,		 /* reg points to skb->data */
 	PTR_TO_PACKET_END,	 /* skb->data + headlen */
 };

include/linux/filter.h

@@ -487,24 +487,30 @@ struct sk_filter {
 
 struct bpf_skb_data_end {
 	struct qdisc_skb_cb qdisc_cb;
+	void *data_meta;
 	void *data_end;
 };
 
 struct xdp_buff {
 	void *data;
 	void *data_end;
+	void *data_meta;
 	void *data_hard_start;
 };
 
-/* compute the linear packet data range [data, data_end) which
- * will be accessed by cls_bpf, act_bpf and lwt programs
+/* Compute the linear packet data range [data, data_end) which
+ * will be accessed by various program types (cls_bpf, act_bpf,
+ * lwt, ...). Subsystems allowing direct data access must (!)
+ * ensure that cb[] area can be written to when BPF program is
+ * invoked (otherwise cb[] save/restore is necessary).
  */
-static inline void bpf_compute_data_end(struct sk_buff *skb)
+static inline void bpf_compute_data_pointers(struct sk_buff *skb)
 {
 	struct bpf_skb_data_end *cb = (struct bpf_skb_data_end *)skb->cb;
 
 	BUILD_BUG_ON(sizeof(*cb) > FIELD_SIZEOF(struct sk_buff, cb));
-	cb->data_end = skb->data + skb_headlen(skb);
+	cb->data_meta = skb->data - skb_metadata_len(skb);
+	cb->data_end  = skb->data + skb_headlen(skb);
 }
 
 static inline u8 *bpf_skb_cb(struct sk_buff *skb)
@@ -725,8 +731,22 @@ int xdp_do_redirect(struct net_device *dev,
 		    struct bpf_prog *prog);
 void xdp_do_flush_map(void);
 
+/* Drivers not supporting XDP metadata can use this helper, which
+ * rejects any room expansion for metadata as a result.
+ */
+static __always_inline void
+xdp_set_data_meta_invalid(struct xdp_buff *xdp)
+{
+	xdp->data_meta = xdp->data + 1;
+}
+
+static __always_inline bool
+xdp_data_meta_unsupported(const struct xdp_buff *xdp)
+{
+	return unlikely(xdp->data_meta > xdp->data);
+}
+
 void bpf_warn_invalid_xdp_action(u32 act);
-void bpf_warn_invalid_xdp_redirect(u32 ifindex);
 
 struct sock *do_sk_redirect_map(void);
 

include/linux/skbuff.h

@@ -489,8 +489,9 @@ int skb_zerocopy_iter_stream(struct sock *sk, struct sk_buff *skb,
  * the end of the header data, ie. at skb->end.
  */
 struct skb_shared_info {
-	unsigned short	_unused;
-	unsigned char	nr_frags;
+	__u8		__unused;
+	__u8		meta_len;
+	__u8		nr_frags;
 	__u8		tx_flags;
 	unsigned short	gso_size;
 	/* Warning: this field is not always filled in (UFO)! */
@@ -3400,6 +3401,69 @@ static inline ktime_t net_invalid_timestamp(void)
 	return 0;
 }
 
+static inline u8 skb_metadata_len(const struct sk_buff *skb)
+{
+	return skb_shinfo(skb)->meta_len;
+}
+
+static inline void *skb_metadata_end(const struct sk_buff *skb)
+{
+	return skb_mac_header(skb);
+}
+
+static inline bool __skb_metadata_differs(const struct sk_buff *skb_a,
+					  const struct sk_buff *skb_b,
+					  u8 meta_len)
+{
+	const void *a = skb_metadata_end(skb_a);
+	const void *b = skb_metadata_end(skb_b);
+	/* Using more efficient varaiant than plain call to memcmp(). */
+#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && BITS_PER_LONG == 64
+	u64 diffs = 0;
+
+	switch (meta_len) {
+#define __it(x, op) (x -= sizeof(u##op))
+#define __it_diff(a, b, op) (*(u##op *)__it(a, op)) ^ (*(u##op *)__it(b, op))
+	case 32: diffs |= __it_diff(a, b, 64);
+	case 24: diffs |= __it_diff(a, b, 64);
+	case 16: diffs |= __it_diff(a, b, 64);
+	case  8: diffs |= __it_diff(a, b, 64);
+		break;
+	case 28: diffs |= __it_diff(a, b, 64);
+	case 20: diffs |= __it_diff(a, b, 64);
+	case 12: diffs |= __it_diff(a, b, 64);
+	case  4: diffs |= __it_diff(a, b, 32);
+		break;
+	}
+	return diffs;
+#else
+	return memcmp(a - meta_len, b - meta_len, meta_len);
+#endif
+}
+
+static inline bool skb_metadata_differs(const struct sk_buff *skb_a,
+					const struct sk_buff *skb_b)
+{
+	u8 len_a = skb_metadata_len(skb_a);
+	u8 len_b = skb_metadata_len(skb_b);
+
+	if (!(len_a | len_b))
+		return false;
+
+	return len_a != len_b ?
+	       true : __skb_metadata_differs(skb_a, skb_b, len_a);
+}
+
+static inline void skb_metadata_set(struct sk_buff *skb, u8 meta_len)
+{
+	skb_shinfo(skb)->meta_len = meta_len;
+}
+
+static inline void skb_metadata_clear(struct sk_buff *skb)
+{
+	skb_metadata_set(skb, 0);
+}
+
 struct sk_buff *skb_clone_sk(struct sk_buff *skb);
 
 #ifdef CONFIG_NETWORK_PHY_TIMESTAMPING