Merge branch 'xdp_xmit-bulking'

Jesper Dangaard Brouer says:

====================
This patchset change ndo_xdp_xmit API to take a bulk of xdp frames.

When kernel is compiled with CONFIG_RETPOLINE, every indirect function
pointer (branch) call hurts performance. For XDP this have a huge
negative performance impact.

This patchset reduce the needed (indirect) calls to ndo_xdp_xmit, but
also prepares for further optimizations.  The DMA APIs use of indirect
function pointer calls is the primary source the regression.  It is
left for a followup patchset, to use bulking calls towards the DMA API
(via the scatter-gatter calls).

The other advantage of this API change is that drivers can easier
amortize the cost of any sync/locking scheme, over the bulk of
packets.  The assumption of the current API is that the driver
implemementing the NDO will also allocate a dedicated XDP TX queue for
every CPU in the system.  Which is not always possible or practical to
configure. E.g. ixgbe cannot load an XDP program on a machine with
more than 96 CPUs, due to limited hardware TX queues.  E.g. virtio_net
is hard to configure as it requires manually increasing the
queues. E.g. tun driver chooses to use a per XDP frame producer lock
modulo smp_processor_id over avail queues.

I'm considered adding 'flags' to ndo_xdp_xmit, but it's not part of
this patchset.  This will be a followup patchset, once we know if this
will be needed (e.g. for non-map xdp_redirect flush-flag, and if
AF_XDP chooses to use ndo_xdp_xmit for TX).

---
V5: Fixed up issues spotted by Daniel and John

V4: Splitout the patches from 4 to 8 patches.  I cannot split the
driver changes from the NDO change, but I've tried to isolated the NDO
change together with the driver change as much as possible.
====================

Signed-off-by: Alexei Starovoitov <[email protected]>

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

@@ -3664,26 +3664,38 @@ netdev_tx_t i40e_lan_xmit_frame(struct sk_buff *skb, struct net_device *netdev)
  * @dev: netdev
  * @xdp: XDP buffer
  *
- * Returns Zero if sent, else an error code
+ * Returns number of frames successfully sent. Frames that fail are
+ * free'ed via XDP return API.
+ *
+ * For error cases, a negative errno code is returned and no-frames
+ * are transmitted (caller must handle freeing frames).
  **/
-int i40e_xdp_xmit(struct net_device *dev, struct xdp_frame *xdpf)
+int i40e_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames)
 {
 	struct i40e_netdev_priv *np = netdev_priv(dev);
 	unsigned int queue_index = smp_processor_id();
 	struct i40e_vsi *vsi = np->vsi;
-	int err;
+	int drops = 0;
+	int i;
 
 	if (test_bit(__I40E_VSI_DOWN, vsi->state))
 		return -ENETDOWN;
 
 	if (!i40e_enabled_xdp_vsi(vsi) || queue_index >= vsi->num_queue_pairs)
 		return -ENXIO;
 
-	err = i40e_xmit_xdp_ring(xdpf, vsi->xdp_rings[queue_index]);
-	if (err != I40E_XDP_TX)
-		return -ENOSPC;
+	for (i = 0; i < n; i++) {
+		struct xdp_frame *xdpf = frames[i];
+		int err;
 
-	return 0;
+		err = i40e_xmit_xdp_ring(xdpf, vsi->xdp_rings[queue_index]);
+		if (err != I40E_XDP_TX) {
+			xdp_return_frame_rx_napi(xdpf);
+			drops++;
+		}
+	}
+
+	return n - drops;
 }
 
 /**

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

@@ -487,7 +487,7 @@ u32 i40e_get_tx_pending(struct i40e_ring *ring, bool in_sw);
 void i40e_detect_recover_hung(struct i40e_vsi *vsi);
 int __i40e_maybe_stop_tx(struct i40e_ring *tx_ring, int size);
 bool __i40e_chk_linearize(struct sk_buff *skb);
-int i40e_xdp_xmit(struct net_device *dev, struct xdp_frame *xdpf);
+int i40e_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames);
 void i40e_xdp_flush(struct net_device *dev);
 
 /**

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

@@ -10017,11 +10017,13 @@ static int ixgbe_xdp(struct net_device *dev, struct netdev_bpf *xdp)
 	}
 }
 
-static int ixgbe_xdp_xmit(struct net_device *dev, struct xdp_frame *xdpf)
+static int ixgbe_xdp_xmit(struct net_device *dev, int n,
+			  struct xdp_frame **frames)
 {
 	struct ixgbe_adapter *adapter = netdev_priv(dev);
 	struct ixgbe_ring *ring;
-	int err;
+	int drops = 0;
+	int i;
 
 	if (unlikely(test_bit(__IXGBE_DOWN, &adapter->state)))
 		return -ENETDOWN;
@@ -10033,11 +10035,18 @@ static int ixgbe_xdp_xmit(struct net_device *dev, struct xdp_frame *xdpf)
 	if (unlikely(!ring))
 		return -ENXIO;
 
-	err = ixgbe_xmit_xdp_ring(adapter, xdpf);
-	if (err != IXGBE_XDP_TX)
-		return -ENOSPC;
+	for (i = 0; i < n; i++) {
+		struct xdp_frame *xdpf = frames[i];
+		int err;
 
-	return 0;
+		err = ixgbe_xmit_xdp_ring(adapter, xdpf);
+		if (err != IXGBE_XDP_TX) {
+			xdp_return_frame_rx_napi(xdpf);
+			drops++;
+		}
+	}
+
+	return n - drops;
 }
 
 static void ixgbe_xdp_flush(struct net_device *dev)

drivers/net/tun.c

@@ -70,6 +70,7 @@
 #include <net/netns/generic.h>
 #include <net/rtnetlink.h>
 #include <net/sock.h>
+#include <net/xdp.h>
 #include <linux/seq_file.h>
 #include <linux/uio.h>
 #include <linux/skb_array.h>
@@ -1290,34 +1291,44 @@ static const struct net_device_ops tun_netdev_ops = {
 	.ndo_get_stats64	= tun_net_get_stats64,
 };
 
-static int tun_xdp_xmit(struct net_device *dev, struct xdp_frame *frame)
+static int tun_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames)
 {
 	struct tun_struct *tun = netdev_priv(dev);
 	struct tun_file *tfile;
 	u32 numqueues;
-	int ret = 0;
+	int drops = 0;
+	int cnt = n;
+	int i;
 
 	rcu_read_lock();
 
 	numqueues = READ_ONCE(tun->numqueues);
 	if (!numqueues) {
-		ret = -ENOSPC;
-		goto out;
+		rcu_read_unlock();
+		return -ENXIO; /* Caller will free/return all frames */
 	}
 
 	tfile = rcu_dereference(tun->tfiles[smp_processor_id() %
 					    numqueues]);
-	/* Encode the XDP flag into lowest bit for consumer to differ
-	 * XDP buffer from sk_buff.
-	 */
-	if (ptr_ring_produce(&tfile->tx_ring, tun_xdp_to_ptr(frame))) {
-		this_cpu_inc(tun->pcpu_stats->tx_dropped);
-		ret = -ENOSPC;
+
+	spin_lock(&tfile->tx_ring.producer_lock);
+	for (i = 0; i < n; i++) {
+		struct xdp_frame *xdp = frames[i];
+		/* Encode the XDP flag into lowest bit for consumer to differ
+		 * XDP buffer from sk_buff.
+		 */
+		void *frame = tun_xdp_to_ptr(xdp);
+
+		if (__ptr_ring_produce(&tfile->tx_ring, frame)) {
+			this_cpu_inc(tun->pcpu_stats->tx_dropped);
+			xdp_return_frame_rx_napi(xdp);
+			drops++;
+		}
 	}
+	spin_unlock(&tfile->tx_ring.producer_lock);
 
-out:
 	rcu_read_unlock();
-	return ret;
+	return cnt - drops;
 }
 
 static int tun_xdp_tx(struct net_device *dev, struct xdp_buff *xdp)
@@ -1327,7 +1338,7 @@ static int tun_xdp_tx(struct net_device *dev, struct xdp_buff *xdp)
 	if (unlikely(!frame))
 		return -EOVERFLOW;
 
-	return tun_xdp_xmit(dev, frame);
+	return tun_xdp_xmit(dev, 1, &frame);
 }
 
 static void tun_xdp_flush(struct net_device *dev)

drivers/net/virtio_net.c

@@ -419,23 +419,13 @@ static void virtnet_xdp_flush(struct net_device *dev)
 	virtqueue_kick(sq->vq);
 }
 
-static int __virtnet_xdp_xmit(struct virtnet_info *vi,
-			       struct xdp_frame *xdpf)
+static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
+				   struct send_queue *sq,
+				   struct xdp_frame *xdpf)
 {
 	struct virtio_net_hdr_mrg_rxbuf *hdr;
-	struct xdp_frame *xdpf_sent;
-	struct send_queue *sq;
-	unsigned int len;
-	unsigned int qp;
 	int err;
 
-	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
-	sq = &vi->sq[qp];
-
-	/* Free up any pending old buffers before queueing new ones. */
-	while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
-		xdp_return_frame(xdpf_sent);
-
 	/* virtqueue want to use data area in-front of packet */
 	if (unlikely(xdpf->metasize > 0))
 		return -EOPNOTSUPP;
@@ -459,11 +449,40 @@ static int __virtnet_xdp_xmit(struct virtnet_info *vi,
 	return 0;
 }
 
-static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_frame *xdpf)
+static int __virtnet_xdp_tx_xmit(struct virtnet_info *vi,
+				   struct xdp_frame *xdpf)
+{
+	struct xdp_frame *xdpf_sent;
+	struct send_queue *sq;
+	unsigned int len;
+	unsigned int qp;
+
+	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
+	sq = &vi->sq[qp];
+
+	/* Free up any pending old buffers before queueing new ones. */
+	while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
+		xdp_return_frame(xdpf_sent);
+
+	return __virtnet_xdp_xmit_one(vi, sq, xdpf);
+}
+
+static int virtnet_xdp_xmit(struct net_device *dev,
+			    int n, struct xdp_frame **frames)
 {
 	struct virtnet_info *vi = netdev_priv(dev);
 	struct receive_queue *rq = vi->rq;
+	struct xdp_frame *xdpf_sent;
 	struct bpf_prog *xdp_prog;
+	struct send_queue *sq;
+	unsigned int len;
+	unsigned int qp;
+	int drops = 0;
+	int err;
+	int i;
+
+	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
+	sq = &vi->sq[qp];
 
 	/* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 	 * indicate XDP resources have been successfully allocated.
@@ -472,7 +491,20 @@ static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_frame *xdpf)
 	if (!xdp_prog)
 		return -ENXIO;
 
-	return __virtnet_xdp_xmit(vi, xdpf);
+	/* Free up any pending old buffers before queueing new ones. */
+	while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
+		xdp_return_frame(xdpf_sent);
+
+	for (i = 0; i < n; i++) {
+		struct xdp_frame *xdpf = frames[i];
+
+		err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
+		if (err) {
+			xdp_return_frame_rx_napi(xdpf);
+			drops++;
+		}
+	}
+	return n - drops;
 }
 
 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
@@ -616,7 +648,7 @@ static struct sk_buff *receive_small(struct net_device *dev,
 			xdpf = convert_to_xdp_frame(&xdp);
 			if (unlikely(!xdpf))
 				goto err_xdp;
-			err = __virtnet_xdp_xmit(vi, xdpf);
+			err = __virtnet_xdp_tx_xmit(vi, xdpf);
 			if (unlikely(err)) {
 				trace_xdp_exception(vi->dev, xdp_prog, act);
 				goto err_xdp;
@@ -779,7 +811,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			xdpf = convert_to_xdp_frame(&xdp);
 			if (unlikely(!xdpf))
 				goto err_xdp;
-			err = __virtnet_xdp_xmit(vi, xdpf);
+			err = __virtnet_xdp_tx_xmit(vi, xdpf);
 			if (unlikely(err)) {
 				trace_xdp_exception(vi->dev, xdp_prog, act);
 				if (unlikely(xdp_page != page))

include/linux/bpf.h

@@ -487,14 +487,17 @@ int bpf_check(struct bpf_prog **fp, union bpf_attr *attr);
 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth);
 
 /* Map specifics */
-struct net_device  *__dev_map_lookup_elem(struct bpf_map *map, u32 key);
+struct xdp_buff;
+
+struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key);
 void __dev_map_insert_ctx(struct bpf_map *map, u32 index);
 void __dev_map_flush(struct bpf_map *map);
+int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
+		    struct net_device *dev_rx);
 
 struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key);
 void __cpu_map_insert_ctx(struct bpf_map *map, u32 index);
 void __cpu_map_flush(struct bpf_map *map);
-struct xdp_buff;
 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
 		    struct net_device *dev_rx);
 
@@ -573,6 +576,16 @@ static inline void __dev_map_flush(struct bpf_map *map)
 {
 }
 
+struct xdp_buff;
+struct bpf_dtab_netdev;
+
+static inline
+int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
+		    struct net_device *dev_rx)
+{
+	return 0;
+}
+
 static inline
 struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
 {
@@ -587,7 +600,6 @@ static inline void __cpu_map_flush(struct bpf_map *map)
 {
 }
 
-struct xdp_buff;
 static inline int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu,
 				  struct xdp_buff *xdp,
 				  struct net_device *dev_rx)

include/linux/netdevice.h

@@ -1185,9 +1185,13 @@ struct dev_ifalias {
  *	This function is used to set or query state related to XDP on the
  *	netdevice and manage BPF offload. See definition of
  *	enum bpf_netdev_command for details.
- * int (*ndo_xdp_xmit)(struct net_device *dev, struct xdp_frame *xdp);
- *	This function is used to submit a XDP packet for transmit on a
- *	netdevice.
+ * int (*ndo_xdp_xmit)(struct net_device *dev, int n, struct xdp_frame **xdp);
+ *	This function is used to submit @n XDP packets for transmit on a
+ *	netdevice. Returns number of frames successfully transmitted, frames
+ *	that got dropped are freed/returned via xdp_return_frame().
+ *	Returns negative number, means general error invoking ndo, meaning
+ *	no frames were xmit'ed and core-caller will free all frames.
+ *	TODO: Consider add flag to allow sending flush operation.
  * void (*ndo_xdp_flush)(struct net_device *dev);
  *	This function is used to inform the driver to flush a particular
  *	xdp tx queue. Must be called on same CPU as xdp_xmit.
@@ -1375,8 +1379,8 @@ struct net_device_ops {
 						       int needed_headroom);
 	int			(*ndo_bpf)(struct net_device *dev,
 					   struct netdev_bpf *bpf);
-	int			(*ndo_xdp_xmit)(struct net_device *dev,
-						struct xdp_frame *xdp);
+	int			(*ndo_xdp_xmit)(struct net_device *dev, int n,
+						struct xdp_frame **xdp);
 	void			(*ndo_xdp_flush)(struct net_device *dev);
 };
 

include/net/page_pool.h

@@ -115,13 +115,14 @@ void page_pool_destroy(struct page_pool *pool);
 void __page_pool_put_page(struct page_pool *pool,
 			  struct page *page, bool allow_direct);
 
-static inline void page_pool_put_page(struct page_pool *pool, struct page *page)
+static inline void page_pool_put_page(struct page_pool *pool,
+				      struct page *page, bool allow_direct)
 {
 	/* When page_pool isn't compiled-in, net/core/xdp.c doesn't
 	 * allow registering MEM_TYPE_PAGE_POOL, but shield linker.
 	 */
 #ifdef CONFIG_PAGE_POOL
-	__page_pool_put_page(pool, page, false);
+	__page_pool_put_page(pool, page, allow_direct);
 #endif
 }
 /* Very limited use-cases allow recycle direct */

include/net/xdp.h

@@ -104,6 +104,7 @@ struct xdp_frame *convert_to_xdp_frame(struct xdp_buff *xdp)
 }
 
 void xdp_return_frame(struct xdp_frame *xdpf);
+void xdp_return_frame_rx_napi(struct xdp_frame *xdpf);
 void xdp_return_buff(struct xdp_buff *xdp);
 
 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,