Merge branch 'sfc-Add-XDP-support'

Charles McLachlan says:

====================
sfc: Add XDP support

Supply the XDP callbacks in netdevice ops that enable lower level processing
of XDP frames.

Changes in v4:
- Handle the failure to send some frames in efx_xdp_tx_buffers() properly.

Changes in v3:
- Fix a BUG_ON when trying to allocate piobufs to xdp queues.
- Add a missed trace_xdp_exception.

Changes in v2:
- Use of xdp_return_frame_rx_napi() in tx.c
- Addition of xdp_rxq_info_valid and xdp_rxq_info_failed to track when
  xdp_rxq_info failures occur.
- Renaming of rc to err and more use of unlikely().
- Cut some duplicated code and fix an array overrun.
- Actually increment n_rx_xdp_tx when packets are transmitted.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/sfc/ef10.c

@@ -946,8 +946,10 @@ static int efx_ef10_link_piobufs(struct efx_nic *efx)
 		/* Extra channels, even those with TXQs (PTP), do not require
 		 * PIO resources.
 		 */
-		if (!channel->type->want_pio)
+		if (!channel->type->want_pio ||
+		    channel->channel >= efx->xdp_channel_offset)
 			continue;
+
 		efx_for_each_channel_tx_queue(tx_queue, channel) {
 			/* We assign the PIO buffers to queues in
 			 * reverse order to allow for the following
@@ -1296,8 +1298,9 @@ static int efx_ef10_dimension_resources(struct efx_nic *efx)
 	int rc;
 
 	channel_vis = max(efx->n_channels,
-			  (efx->n_tx_channels + efx->n_extra_tx_channels) *
-			  EFX_TXQ_TYPES);
+			  ((efx->n_tx_channels + efx->n_extra_tx_channels) *
+			   EFX_TXQ_TYPES) +
+			   efx->n_xdp_channels * efx->xdp_tx_per_channel);
 
 #ifdef EFX_USE_PIO
 	/* Try to allocate PIO buffers if wanted and if the full
@@ -2434,11 +2437,12 @@ static void efx_ef10_tx_init(struct efx_tx_queue *tx_queue)
 	/* TSOv2 is a limited resource that can only be configured on a limited
 	 * number of queues. TSO without checksum offload is not really a thing,
 	 * so we only enable it for those queues.
-	 * TSOv2 cannot be used with Hardware timestamping.
+	 * TSOv2 cannot be used with Hardware timestamping, and is never needed
+	 * for XDP tx.
 	 */
 	if (csum_offload && (nic_data->datapath_caps2 &
 			(1 << MC_CMD_GET_CAPABILITIES_V2_OUT_TX_TSO_V2_LBN)) &&
-	    !tx_queue->timestamping) {
+	    !tx_queue->timestamping && !tx_queue->xdp_tx) {
 		tso_v2 = true;
 		netif_dbg(efx, hw, efx->net_dev, "Using TSOv2 for channel %u\n",
 				channel->channel);

drivers/net/ethernet/sfc/efx.c

@@ -226,6 +226,10 @@ static void efx_fini_napi_channel(struct efx_channel *channel);
 static void efx_fini_struct(struct efx_nic *efx);
 static void efx_start_all(struct efx_nic *efx);
 static void efx_stop_all(struct efx_nic *efx);
+static int efx_xdp_setup_prog(struct efx_nic *efx, struct bpf_prog *prog);
+static int efx_xdp(struct net_device *dev, struct netdev_bpf *xdp);
+static int efx_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **xdpfs,
+			u32 flags);
 
 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
 	do {						\
@@ -340,6 +344,8 @@ static int efx_poll(struct napi_struct *napi, int budget)
 
 	spent = efx_process_channel(channel, budget);
 
+	xdp_do_flush_map();
+
 	if (spent < budget) {
 		if (efx_channel_has_rx_queue(channel) &&
 		    efx->irq_rx_adaptive &&
@@ -579,9 +585,14 @@ efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
 	int number;
 
 	number = channel->channel;
-	if (efx->tx_channel_offset == 0) {
+
+	if (number >= efx->xdp_channel_offset &&
+	    !WARN_ON_ONCE(!efx->n_xdp_channels)) {
+		type = "-xdp";
+		number -= efx->xdp_channel_offset;
+	} else if (efx->tx_channel_offset == 0) {
 		type = "";
-	} else if (channel->channel < efx->tx_channel_offset) {
+	} else if (number < efx->tx_channel_offset) {
 		type = "-rx";
 	} else {
 		type = "-tx";
@@ -651,7 +662,7 @@ static void efx_start_datapath(struct efx_nic *efx)
 	efx->rx_dma_len = (efx->rx_prefix_size +
 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
 			   efx->type->rx_buffer_padding);
-	rx_buf_len = (sizeof(struct efx_rx_page_state) +
+	rx_buf_len = (sizeof(struct efx_rx_page_state) + XDP_PACKET_HEADROOM +
 		      efx->rx_ip_align + efx->rx_dma_len);
 	if (rx_buf_len <= PAGE_SIZE) {
 		efx->rx_scatter = efx->type->always_rx_scatter;
@@ -774,6 +785,7 @@ static void efx_stop_datapath(struct efx_nic *efx)
 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
 			efx_fini_tx_queue(tx_queue);
 	}
+	efx->xdp_rxq_info_failed = false;
 }
 
 static void efx_remove_channel(struct efx_channel *channel)
@@ -798,6 +810,8 @@ static void efx_remove_channels(struct efx_nic *efx)
 
 	efx_for_each_channel(channel, efx)
 		efx_remove_channel(channel);
+
+	kfree(efx->xdp_tx_queues);
 }
 
 int
@@ -1435,6 +1449,101 @@ static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
 	return count;
 }
 
+static int efx_allocate_msix_channels(struct efx_nic *efx,
+				      unsigned int max_channels,
+				      unsigned int extra_channels,
+				      unsigned int parallelism)
+{
+	unsigned int n_channels = parallelism;
+	int vec_count;
+	int n_xdp_tx;
+	int n_xdp_ev;
+
+	if (efx_separate_tx_channels)
+		n_channels *= 2;
+	n_channels += extra_channels;
+
+	/* To allow XDP transmit to happen from arbitrary NAPI contexts
+	 * we allocate a TX queue per CPU. We share event queues across
+	 * multiple tx queues, assuming tx and ev queues are both
+	 * maximum size.
+	 */
+
+	n_xdp_tx = num_possible_cpus();
+	n_xdp_ev = DIV_ROUND_UP(n_xdp_tx, EFX_TXQ_TYPES);
+
+	/* Check resources.
+	 * We need a channel per event queue, plus a VI per tx queue.
+	 * This may be more pessimistic than it needs to be.
+	 */
+	if (n_channels + n_xdp_ev > max_channels) {
+		netif_err(efx, drv, efx->net_dev,
+			  "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n",
+			  n_xdp_ev, n_channels, max_channels);
+		efx->n_xdp_channels = 0;
+		efx->xdp_tx_per_channel = 0;
+		efx->xdp_tx_queue_count = 0;
+	} else {
+		efx->n_xdp_channels = n_xdp_ev;
+		efx->xdp_tx_per_channel = EFX_TXQ_TYPES;
+		efx->xdp_tx_queue_count = n_xdp_tx;
+		n_channels += n_xdp_ev;
+		netif_dbg(efx, drv, efx->net_dev,
+			  "Allocating %d TX and %d event queues for XDP\n",
+			  n_xdp_tx, n_xdp_ev);
+	}
+
+	n_channels = min(n_channels, max_channels);
+
+	vec_count = pci_msix_vec_count(efx->pci_dev);
+	if (vec_count < 0)
+		return vec_count;
+	if (vec_count < n_channels) {
+		netif_err(efx, drv, efx->net_dev,
+			  "WARNING: Insufficient MSI-X vectors available (%d < %u).\n",
+			  vec_count, n_channels);
+		netif_err(efx, drv, efx->net_dev,
+			  "WARNING: Performance may be reduced.\n");
+		n_channels = vec_count;
+	}
+
+	efx->n_channels = n_channels;
+
+	/* Do not create the PTP TX queue(s) if PTP uses the MC directly. */
+	if (extra_channels && !efx_ptp_use_mac_tx_timestamps(efx))
+		n_channels--;
+
+	/* Ignore XDP tx channels when creating rx channels. */
+	n_channels -= efx->n_xdp_channels;
+
+	if (efx_separate_tx_channels) {
+		efx->n_tx_channels =
+			min(max(n_channels / 2, 1U),
+			    efx->max_tx_channels);
+		efx->tx_channel_offset =
+			n_channels - efx->n_tx_channels;
+		efx->n_rx_channels =
+			max(n_channels -
+			    efx->n_tx_channels, 1U);
+	} else {
+		efx->n_tx_channels = min(n_channels, efx->max_tx_channels);
+		efx->tx_channel_offset = 0;
+		efx->n_rx_channels = n_channels;
+	}
+
+	if (efx->n_xdp_channels)
+		efx->xdp_channel_offset = efx->tx_channel_offset +
+					  efx->n_tx_channels;
+	else
+		efx->xdp_channel_offset = efx->n_channels;
+
+	netif_dbg(efx, drv, efx->net_dev,
+		  "Allocating %u RX channels\n",
+		  efx->n_rx_channels);
+
+	return efx->n_channels;
+}
+
 /* Probe the number and type of interrupts we are able to obtain, and
  * the resulting numbers of channels and RX queues.
  */
@@ -1449,19 +1558,19 @@ static int efx_probe_interrupts(struct efx_nic *efx)
 			++extra_channels;
 
 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
+		unsigned int parallelism = efx_wanted_parallelism(efx);
 		struct msix_entry xentries[EFX_MAX_CHANNELS];
 		unsigned int n_channels;
 
-		n_channels = efx_wanted_parallelism(efx);
-		if (efx_separate_tx_channels)
-			n_channels *= 2;
-		n_channels += extra_channels;
-		n_channels = min(n_channels, efx->max_channels);
-
+		rc = efx_allocate_msix_channels(efx, efx->max_channels,
+						extra_channels, parallelism);
+		if (rc >= 0) {
+			n_channels = rc;
 			for (i = 0; i < n_channels; i++)
 				xentries[i].entry = i;
-		rc = pci_enable_msix_range(efx->pci_dev,
-					   xentries, 1, n_channels);
+			rc = pci_enable_msix_range(efx->pci_dev, xentries, 1,
+						   n_channels);
+		}
 		if (rc < 0) {
 			/* Fall back to single channel MSI */
 			netif_err(efx, drv, efx->net_dev,
@@ -1480,21 +1589,6 @@ static int efx_probe_interrupts(struct efx_nic *efx)
 		}
 
 		if (rc > 0) {
-			efx->n_channels = n_channels;
-			if (n_channels > extra_channels)
-				n_channels -= extra_channels;
-			if (efx_separate_tx_channels) {
-				efx->n_tx_channels = min(max(n_channels / 2,
-							     1U),
-							 efx->max_tx_channels);
-				efx->n_rx_channels = max(n_channels -
-							 efx->n_tx_channels,
-							 1U);
-			} else {
-				efx->n_tx_channels = min(n_channels,
-							 efx->max_tx_channels);
-				efx->n_rx_channels = n_channels;
-			}
 			for (i = 0; i < efx->n_channels; i++)
 				efx_get_channel(efx, i)->irq =
 					xentries[i].vector;
@@ -1506,6 +1600,8 @@ static int efx_probe_interrupts(struct efx_nic *efx)
 		efx->n_channels = 1;
 		efx->n_rx_channels = 1;
 		efx->n_tx_channels = 1;
+		efx->n_xdp_channels = 0;
+		efx->xdp_channel_offset = efx->n_channels;
 		rc = pci_enable_msi(efx->pci_dev);
 		if (rc == 0) {
 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
@@ -1524,12 +1620,14 @@ static int efx_probe_interrupts(struct efx_nic *efx)
 		efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
 		efx->n_rx_channels = 1;
 		efx->n_tx_channels = 1;
+		efx->n_xdp_channels = 0;
+		efx->xdp_channel_offset = efx->n_channels;
 		efx->legacy_irq = efx->pci_dev->irq;
 	}
 
-	/* Assign extra channels if possible */
+	/* Assign extra channels if possible, before XDP channels */
 	efx->n_extra_tx_channels = 0;
-	j = efx->n_channels;
+	j = efx->xdp_channel_offset;
 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
 		if (!efx->extra_channel_type[i])
 			continue;
@@ -1724,29 +1822,50 @@ static void efx_remove_interrupts(struct efx_nic *efx)
 	efx->legacy_irq = 0;
 }
 
-static void efx_set_channels(struct efx_nic *efx)
+static int efx_set_channels(struct efx_nic *efx)
 {
 	struct efx_channel *channel;
 	struct efx_tx_queue *tx_queue;
+	int xdp_queue_number;
 
 	efx->tx_channel_offset =
 		efx_separate_tx_channels ?
 		efx->n_channels - efx->n_tx_channels : 0;
 
+	if (efx->xdp_tx_queue_count) {
+		EFX_WARN_ON_PARANOID(efx->xdp_tx_queues);
+
+		/* Allocate array for XDP TX queue lookup. */
+		efx->xdp_tx_queues = kcalloc(efx->xdp_tx_queue_count,
+					     sizeof(*efx->xdp_tx_queues),
+					     GFP_KERNEL);
+		if (!efx->xdp_tx_queues)
+			return -ENOMEM;
+	}
+
 	/* We need to mark which channels really have RX and TX
 	 * queues, and adjust the TX queue numbers if we have separate
 	 * RX-only and TX-only channels.
 	 */
+	xdp_queue_number = 0;
 	efx_for_each_channel(channel, efx) {
 		if (channel->channel < efx->n_rx_channels)
 			channel->rx_queue.core_index = channel->channel;
 		else
 			channel->rx_queue.core_index = -1;
 
-		efx_for_each_channel_tx_queue(tx_queue, channel)
+		efx_for_each_channel_tx_queue(tx_queue, channel) {
 			tx_queue->queue -= (efx->tx_channel_offset *
 					    EFX_TXQ_TYPES);
+
+			if (efx_channel_is_xdp_tx(channel) &&
+			    xdp_queue_number < efx->xdp_tx_queue_count) {
+				efx->xdp_tx_queues[xdp_queue_number] = tx_queue;
+				xdp_queue_number++;
+			}
 		}
+	}
+	return 0;
 }
 
 static int efx_probe_nic(struct efx_nic *efx)
@@ -1776,7 +1895,9 @@ static int efx_probe_nic(struct efx_nic *efx)
 		if (rc)
 			goto fail1;
 
-		efx_set_channels(efx);
+		rc = efx_set_channels(efx);
+		if (rc)
+			goto fail1;
 
 		/* dimension_resources can fail with EAGAIN */
 		rc = efx->type->dimension_resources(efx);
@@ -2022,6 +2143,10 @@ static void efx_stop_all(struct efx_nic *efx)
 
 static void efx_remove_all(struct efx_nic *efx)
 {
+	rtnl_lock();
+	efx_xdp_setup_prog(efx, NULL);
+	rtnl_unlock();
+
 	efx_remove_channels(efx);
 	efx_remove_filters(efx);
 #ifdef CONFIG_SFC_SRIOV
@@ -2082,6 +2207,8 @@ int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
 			channel->irq_moderation_us = rx_usecs;
 		else if (efx_channel_has_tx_queues(channel))
 			channel->irq_moderation_us = tx_usecs;
+		else if (efx_channel_is_xdp_tx(channel))
+			channel->irq_moderation_us = tx_usecs;
 	}
 
 	return 0;
@@ -2277,6 +2404,17 @@ static void efx_watchdog(struct net_device *net_dev)
 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
 }
 
+static unsigned int efx_xdp_max_mtu(struct efx_nic *efx)
+{
+	/* The maximum MTU that we can fit in a single page, allowing for
+	 * framing, overhead and XDP headroom.
+	 */
+	int overhead = EFX_MAX_FRAME_LEN(0) + sizeof(struct efx_rx_page_state) +
+		       efx->rx_prefix_size + efx->type->rx_buffer_padding +
+		       efx->rx_ip_align + XDP_PACKET_HEADROOM;
+
+	return PAGE_SIZE - overhead;
+}
 
 /* Context: process, rtnl_lock() held. */
 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
@@ -2288,6 +2426,14 @@ static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
 	if (rc)
 		return rc;
 
+	if (rtnl_dereference(efx->xdp_prog) &&
+	    new_mtu > efx_xdp_max_mtu(efx)) {
+		netif_err(efx, drv, efx->net_dev,
+			  "Requested MTU of %d too big for XDP (max: %d)\n",
+			  new_mtu, efx_xdp_max_mtu(efx));
+		return -EINVAL;
+	}
+
 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
 
 	efx_device_detach_sync(efx);
@@ -2489,8 +2635,65 @@ static const struct net_device_ops efx_netdev_ops = {
 #endif
 	.ndo_udp_tunnel_add	= efx_udp_tunnel_add,
 	.ndo_udp_tunnel_del	= efx_udp_tunnel_del,
+	.ndo_xdp_xmit		= efx_xdp_xmit,
+	.ndo_bpf		= efx_xdp
 };
 
+static int efx_xdp_setup_prog(struct efx_nic *efx, struct bpf_prog *prog)
+{
+	struct bpf_prog *old_prog;
+
+	if (efx->xdp_rxq_info_failed) {
+		netif_err(efx, drv, efx->net_dev,
+			  "Unable to bind XDP program due to previous failure of rxq_info\n");
+		return -EINVAL;
+	}
+
+	if (prog && efx->net_dev->mtu > efx_xdp_max_mtu(efx)) {
+		netif_err(efx, drv, efx->net_dev,
+			  "Unable to configure XDP with MTU of %d (max: %d)\n",
+			  efx->net_dev->mtu, efx_xdp_max_mtu(efx));
+		return -EINVAL;
+	}
+
+	old_prog = rtnl_dereference(efx->xdp_prog);
+	rcu_assign_pointer(efx->xdp_prog, prog);
+	/* Release the reference that was originally passed by the caller. */
+	if (old_prog)
+		bpf_prog_put(old_prog);
+
+	return 0;
+}
+
+/* Context: process, rtnl_lock() held. */
+static int efx_xdp(struct net_device *dev, struct netdev_bpf *xdp)
+{
+	struct efx_nic *efx = netdev_priv(dev);
+	struct bpf_prog *xdp_prog;
+
+	switch (xdp->command) {
+	case XDP_SETUP_PROG:
+		return efx_xdp_setup_prog(efx, xdp->prog);
+	case XDP_QUERY_PROG:
+		xdp_prog = rtnl_dereference(efx->xdp_prog);
+		xdp->prog_id = xdp_prog ? xdp_prog->aux->id : 0;
+		return 0;
+	default:
+		return -EINVAL;
+	}
+}
+
+static int efx_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **xdpfs,
+			u32 flags)
+{
+	struct efx_nic *efx = netdev_priv(dev);
+
+	if (!netif_running(dev))
+		return -EINVAL;
+
+	return efx_xdp_tx_buffers(efx, n, xdpfs, flags & XDP_XMIT_FLUSH);
+}
+
 static void efx_update_name(struct efx_nic *efx)
 {
 	strcpy(efx->name, efx->net_dev->name);

drivers/net/ethernet/sfc/efx.h

@@ -322,4 +322,7 @@ static inline bool efx_rwsem_assert_write_locked(struct rw_semaphore *sem)
 	return true;
 }
 
+int efx_xdp_tx_buffers(struct efx_nic *efx, int n, struct xdp_frame **xdpfs,
+		       bool flush);
+
 #endif /* EFX_EFX_H */

drivers/net/ethernet/sfc/ethtool.c

@@ -83,6 +83,10 @@ static const struct efx_sw_stat_desc efx_sw_stat_desc[] = {
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_frm_trunc),
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_merge_events),
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_merge_packets),
+	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_xdp_drops),
+	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_xdp_bad_drops),
+	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_xdp_tx),
+	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_xdp_redirect),
 };
 
 #define EFX_ETHTOOL_SW_STAT_COUNT ARRAY_SIZE(efx_sw_stat_desc)
@@ -399,6 +403,19 @@ static size_t efx_describe_per_queue_stats(struct efx_nic *efx, u8 *strings)
 			}
 		}
 	}
+	if (efx->xdp_tx_queue_count && efx->xdp_tx_queues) {
+		unsigned short xdp;
+
+		for (xdp = 0; xdp < efx->xdp_tx_queue_count; xdp++) {
+			n_stats++;
+			if (strings) {
+				snprintf(strings, ETH_GSTRING_LEN,
+					 "tx-xdp-cpu-%hu.tx_packets", xdp);
+				strings += ETH_GSTRING_LEN;
+			}
+		}
+	}
+
 	return n_stats;
 }
 
@@ -509,6 +526,14 @@ static void efx_ethtool_get_stats(struct net_device *net_dev,
 			data++;
 		}
 	}
+	if (efx->xdp_tx_queue_count && efx->xdp_tx_queues) {
+		int xdp;
+
+		for (xdp = 0; xdp < efx->xdp_tx_queue_count; xdp++) {
+			data[0] = efx->xdp_tx_queues[xdp]->tx_packets;
+			data++;
+		}
+	}
 
 	efx_ptp_update_stats(efx, data);
 }

drivers/net/ethernet/sfc/net_driver.h

@@ -27,6 +27,7 @@
 #include <linux/i2c.h>
 #include <linux/mtd/mtd.h>
 #include <net/busy_poll.h>
+#include <net/xdp.h>
 
 #include "enum.h"
 #include "bitfield.h"
@@ -136,7 +137,8 @@ struct efx_special_buffer {
  * struct efx_tx_buffer - buffer state for a TX descriptor
  * @skb: When @flags & %EFX_TX_BUF_SKB, the associated socket buffer to be
  *	freed when descriptor completes
- * @option: When @flags & %EFX_TX_BUF_OPTION, a NIC-specific option descriptor.
+ * @xdpf: When @flags & %EFX_TX_BUF_XDP, the XDP frame information; its @data
+ *	member is the associated buffer to drop a page reference on.
  * @dma_addr: DMA address of the fragment.
  * @flags: Flags for allocation and DMA mapping type
  * @len: Length of this fragment.
@@ -146,7 +148,10 @@ struct efx_special_buffer {
  * Only valid if @unmap_len != 0.
  */
 struct efx_tx_buffer {
+	union {
 		const struct sk_buff *skb;
+		struct xdp_frame *xdpf;
+	};
 	union {
 		efx_qword_t option;
 		dma_addr_t dma_addr;
@@ -160,6 +165,7 @@ struct efx_tx_buffer {
 #define EFX_TX_BUF_SKB		2	/* buffer is last part of skb */
 #define EFX_TX_BUF_MAP_SINGLE	8	/* buffer was mapped with dma_map_single() */
 #define EFX_TX_BUF_OPTION	0x10	/* empty buffer for option descriptor */
+#define EFX_TX_BUF_XDP		0x20	/* buffer was sent with XDP */
 
 /**
  * struct efx_tx_queue - An Efx TX queue
@@ -189,6 +195,7 @@ struct efx_tx_buffer {
  * @piobuf_offset: Buffer offset to be specified in PIO descriptors
  * @initialised: Has hardware queue been initialised?
  * @timestamping: Is timestamping enabled for this channel?
+ * @xdp_tx: Is this an XDP tx queue?
  * @handle_tso: TSO xmit preparation handler.  Sets up the TSO metadata and
  *	may also map tx data, depending on the nature of the TSO implementation.
  * @read_count: Current read pointer.
@@ -250,6 +257,7 @@ struct efx_tx_queue {
 	unsigned int piobuf_offset;
 	bool initialised;
 	bool timestamping;
+	bool xdp_tx;
 
 	/* Function pointers used in the fast path. */
 	int (*handle_tso)(struct efx_tx_queue*, struct sk_buff*, bool *);
@@ -363,6 +371,8 @@ struct efx_rx_page_state {
  *	refill was triggered.
  * @recycle_count: RX buffer recycle counter.
  * @slow_fill: Timer used to defer efx_nic_generate_fill_event().
+ * @xdp_rxq_info: XDP specific RX queue information.
+ * @xdp_rxq_info_valid: Is xdp_rxq_info valid data?.
  */
 struct efx_rx_queue {
 	struct efx_nic *efx;
@@ -394,6 +404,8 @@ struct efx_rx_queue {
 	unsigned int slow_fill_count;
 	/* Statistics to supplement MAC stats */
 	unsigned long rx_packets;
+	struct xdp_rxq_info xdp_rxq_info;
+	bool xdp_rxq_info_valid;
 };
 
 enum efx_sync_events_state {
@@ -441,6 +453,10 @@ enum efx_sync_events_state {
  *	lack of descriptors
  * @n_rx_merge_events: Number of RX merged completion events
  * @n_rx_merge_packets: Number of RX packets completed by merged events
+ * @n_rx_xdp_drops: Count of RX packets intentionally dropped due to XDP
+ * @n_rx_xdp_bad_drops: Count of RX packets dropped due to XDP errors
+ * @n_rx_xdp_tx: Count of RX packets retransmitted due to XDP
+ * @n_rx_xdp_redirect: Count of RX packets redirected to a different NIC by XDP
  * @rx_pkt_n_frags: Number of fragments in next packet to be delivered by
  *	__efx_rx_packet(), or zero if there is none
  * @rx_pkt_index: Ring index of first buffer for next packet to be delivered
@@ -494,6 +510,10 @@ struct efx_channel {
 	unsigned int n_rx_nodesc_trunc;
 	unsigned int n_rx_merge_events;
 	unsigned int n_rx_merge_packets;
+	unsigned int n_rx_xdp_drops;
+	unsigned int n_rx_xdp_bad_drops;
+	unsigned int n_rx_xdp_tx;
+	unsigned int n_rx_xdp_redirect;
 
 	unsigned int rx_pkt_n_frags;
 	unsigned int rx_pkt_index;
@@ -818,6 +838,8 @@ struct efx_async_filter_insertion {
  * @msi_context: Context for each MSI
  * @extra_channel_types: Types of extra (non-traffic) channels that
  *	should be allocated for this NIC
+ * @xdp_tx_queue_count: Number of entries in %xdp_tx_queues.
+ * @xdp_tx_queues: Array of pointers to tx queues used for XDP transmit.
  * @rxq_entries: Size of receive queues requested by user.
  * @txq_entries: Size of transmit queues requested by user.
  * @txq_stop_thresh: TX queue fill level at or above which we stop it.
@@ -830,6 +852,9 @@ struct efx_async_filter_insertion {
  * @n_rx_channels: Number of channels used for RX (= number of RX queues)
  * @n_tx_channels: Number of channels used for TX
  * @n_extra_tx_channels: Number of extra channels with TX queues
+ * @n_xdp_channels: Number of channels used for XDP TX
+ * @xdp_channel_offset: Offset of zeroth channel used for XPD TX.
+ * @xdp_tx_per_channel: Max number of TX queues on an XDP TX channel.
  * @rx_ip_align: RX DMA address offset to have IP header aligned in
  *	in accordance with NET_IP_ALIGN
  * @rx_dma_len: Current maximum RX DMA length
@@ -894,6 +919,7 @@ struct efx_async_filter_insertion {
  * @loopback_mode: Loopback status
  * @loopback_modes: Supported loopback mode bitmask
  * @loopback_selftest: Offline self-test private state
+ * @xdp_prog: Current XDP programme for this interface
  * @filter_sem: Filter table rw_semaphore, protects existence of @filter_state
  * @filter_state: Architecture-dependent filter table state
  * @rps_mutex: Protects RPS state of all channels
@@ -919,6 +945,8 @@ struct efx_async_filter_insertion {
  * @ptp_data: PTP state data
  * @ptp_warned: has this NIC seen and warned about unexpected PTP events?
  * @vpd_sn: Serial number read from VPD
+ * @xdp_rxq_info_failed: Have any of the rx queues failed to initialise their
+ *      xdp_rxq_info structures?
  * @monitor_work: Hardware monitor workitem
  * @biu_lock: BIU (bus interface unit) lock
  * @last_irq_cpu: Last CPU to handle a possible test interrupt.  This
@@ -966,6 +994,9 @@ struct efx_nic {
 	const struct efx_channel_type *
 	extra_channel_type[EFX_MAX_EXTRA_CHANNELS];
 
+	unsigned int xdp_tx_queue_count;
+	struct efx_tx_queue **xdp_tx_queues;
+
 	unsigned rxq_entries;
 	unsigned txq_entries;
 	unsigned int txq_stop_thresh;
@@ -984,6 +1015,9 @@ struct efx_nic {
 	unsigned tx_channel_offset;
 	unsigned n_tx_channels;
 	unsigned n_extra_tx_channels;
+	unsigned int n_xdp_channels;
+	unsigned int xdp_channel_offset;
+	unsigned int xdp_tx_per_channel;
 	unsigned int rx_ip_align;
 	unsigned int rx_dma_len;
 	unsigned int rx_buffer_order;
@@ -1053,6 +1087,10 @@ struct efx_nic {
 	u64 loopback_modes;
 
 	void *loopback_selftest;
+	/* We access loopback_selftest immediately before running XDP,
+	 * so we want them next to each other.
+	 */
+	struct bpf_prog __rcu *xdp_prog;
 
 	struct rw_semaphore filter_sem;
 	void *filter_state;
@@ -1082,6 +1120,7 @@ struct efx_nic {
 	bool ptp_warned;
 
 	char *vpd_sn;
+	bool xdp_rxq_info_failed;
 
 	/* The following fields may be written more often */
 
@@ -1473,10 +1512,24 @@ efx_get_tx_queue(struct efx_nic *efx, unsigned index, unsigned type)
 	return &efx->channel[efx->tx_channel_offset + index]->tx_queue[type];
 }
 
+static inline struct efx_channel *
+efx_get_xdp_channel(struct efx_nic *efx, unsigned int index)
+{
+	EFX_WARN_ON_ONCE_PARANOID(index >= efx->n_xdp_channels);
+	return efx->channel[efx->xdp_channel_offset + index];
+}
+
+static inline bool efx_channel_is_xdp_tx(struct efx_channel *channel)
+{
+	return channel->channel - channel->efx->xdp_channel_offset <
+	       channel->efx->n_xdp_channels;
+}
+
 static inline bool efx_channel_has_tx_queues(struct efx_channel *channel)
 {
-	return channel->type && channel->type->want_txqs &&
-				channel->type->want_txqs(channel);
+	return efx_channel_is_xdp_tx(channel) ||
+	       (channel->type && channel->type->want_txqs &&
+		channel->type->want_txqs(channel));
 }
 
 static inline struct efx_tx_queue *
@@ -1500,7 +1553,8 @@ static inline bool efx_tx_queue_used(struct efx_tx_queue *tx_queue)
 	else								\
 		for (_tx_queue = (_channel)->tx_queue;			\
 		     _tx_queue < (_channel)->tx_queue + EFX_TXQ_TYPES && \
-			     efx_tx_queue_used(_tx_queue);		\
+			     (efx_tx_queue_used(_tx_queue) ||            \
+			      efx_channel_is_xdp_tx(_channel));		\
 		     _tx_queue++)
 
 /* Iterate over all possible TX queues belonging to a channel */

drivers/net/ethernet/sfc/rx.c

@@ -17,6 +17,8 @@
 #include <linux/iommu.h>
 #include <net/ip.h>
 #include <net/checksum.h>
+#include <net/xdp.h>
+#include <linux/bpf_trace.h>
 #include "net_driver.h"
 #include "efx.h"
 #include "filter.h"
@@ -27,6 +29,9 @@
 /* Preferred number of descriptors to fill at once */
 #define EFX_RX_PREFERRED_BATCH 8U
 
+/* Maximum rx prefix used by any architecture. */
+#define EFX_MAX_RX_PREFIX_SIZE 16
+
 /* Number of RX buffers to recycle pages for.  When creating the RX page recycle
  * ring, this number is divided by the number of buffers per page to calculate
  * the number of pages to store in the RX page recycle ring.
@@ -95,7 +100,7 @@ void efx_rx_config_page_split(struct efx_nic *efx)
 				      EFX_RX_BUF_ALIGNMENT);
 	efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
 		((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
-		 efx->rx_page_buf_step);
+		(efx->rx_page_buf_step + XDP_PACKET_HEADROOM));
 	efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
 		efx->rx_bufs_per_page;
 	efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
@@ -185,6 +190,9 @@ static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
 		page_offset = sizeof(struct efx_rx_page_state);
 
 		do {
+			page_offset += XDP_PACKET_HEADROOM;
+			dma_addr += XDP_PACKET_HEADROOM;
+
 			index = rx_queue->added_count & rx_queue->ptr_mask;
 			rx_buf = efx_rx_buffer(rx_queue, index);
 			rx_buf->dma_addr = dma_addr + efx->rx_ip_align;
@@ -635,6 +643,123 @@ static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
 		netif_receive_skb(skb);
 }
 
+/** efx_do_xdp: perform XDP processing on a received packet
+ *
+ * Returns true if packet should still be delivered.
+ */
+static bool efx_do_xdp(struct efx_nic *efx, struct efx_channel *channel,
+		       struct efx_rx_buffer *rx_buf, u8 **ehp)
+{
+	u8 rx_prefix[EFX_MAX_RX_PREFIX_SIZE];
+	struct efx_rx_queue *rx_queue;
+	struct bpf_prog *xdp_prog;
+	struct xdp_frame *xdpf;
+	struct xdp_buff xdp;
+	u32 xdp_act;
+	s16 offset;
+	int err;
+
+	rcu_read_lock();
+	xdp_prog = rcu_dereference(efx->xdp_prog);
+	if (!xdp_prog) {
+		rcu_read_unlock();
+		return true;
+	}
+
+	rx_queue = efx_channel_get_rx_queue(channel);
+
+	if (unlikely(channel->rx_pkt_n_frags > 1)) {
+		/* We can't do XDP on fragmented packets - drop. */
+		rcu_read_unlock();
+		efx_free_rx_buffers(rx_queue, rx_buf,
+				    channel->rx_pkt_n_frags);
+		if (net_ratelimit())
+			netif_err(efx, rx_err, efx->net_dev,
+				  "XDP is not possible with multiple receive fragments (%d)\n",
+				  channel->rx_pkt_n_frags);
+		channel->n_rx_xdp_bad_drops++;
+		return false;
+	}
+
+	dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr,
+				rx_buf->len, DMA_FROM_DEVICE);
+
+	/* Save the rx prefix. */
+	EFX_WARN_ON_PARANOID(efx->rx_prefix_size > EFX_MAX_RX_PREFIX_SIZE);
+	memcpy(rx_prefix, *ehp - efx->rx_prefix_size,
+	       efx->rx_prefix_size);
+
+	xdp.data = *ehp;
+	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
+
+	/* No support yet for XDP metadata */
+	xdp_set_data_meta_invalid(&xdp);
+	xdp.data_end = xdp.data + rx_buf->len;
+	xdp.rxq = &rx_queue->xdp_rxq_info;
+
+	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
+	rcu_read_unlock();
+
+	offset = (u8 *)xdp.data - *ehp;
+
+	switch (xdp_act) {
+	case XDP_PASS:
+		/* Fix up rx prefix. */
+		if (offset) {
+			*ehp += offset;
+			rx_buf->page_offset += offset;
+			rx_buf->len -= offset;
+			memcpy(*ehp - efx->rx_prefix_size, rx_prefix,
+			       efx->rx_prefix_size);
+		}
+		break;
+
+	case XDP_TX:
+		/* Buffer ownership passes to tx on success. */
+		xdpf = convert_to_xdp_frame(&xdp);
+		err = efx_xdp_tx_buffers(efx, 1, &xdpf, true);
+		if (unlikely(err != 1)) {
+			efx_free_rx_buffers(rx_queue, rx_buf, 1);
+			if (net_ratelimit())
+				netif_err(efx, rx_err, efx->net_dev,
+					  "XDP TX failed (%d)\n", err);
+			channel->n_rx_xdp_bad_drops++;
+		} else {
+			channel->n_rx_xdp_tx++;
+		}
+		break;
+
+	case XDP_REDIRECT:
+		err = xdp_do_redirect(efx->net_dev, &xdp, xdp_prog);
+		if (unlikely(err)) {
+			efx_free_rx_buffers(rx_queue, rx_buf, 1);
+			if (net_ratelimit())
+				netif_err(efx, rx_err, efx->net_dev,
+					  "XDP redirect failed (%d)\n", err);
+			channel->n_rx_xdp_bad_drops++;
+		} else {
+			channel->n_rx_xdp_redirect++;
+		}
+		break;
+
+	default:
+		bpf_warn_invalid_xdp_action(xdp_act);
+		efx_free_rx_buffers(rx_queue, rx_buf, 1);
+		channel->n_rx_xdp_bad_drops++;
+		break;
+
+	case XDP_ABORTED:
+		trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
+		/* Fall through */
+	case XDP_DROP:
+		efx_free_rx_buffers(rx_queue, rx_buf, 1);
+		channel->n_rx_xdp_drops++;
+		break;
+	}
+
+	return xdp_act == XDP_PASS;
+}
+
 /* Handle a received packet.  Second half: Touches packet payload. */
 void __efx_rx_packet(struct efx_channel *channel)
 {
@@ -663,6 +788,9 @@ void __efx_rx_packet(struct efx_channel *channel)
 		goto out;
 	}
 
+	if (!efx_do_xdp(efx, channel, rx_buf, &eh))
+		goto out;
+
 	if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 		rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 
@@ -731,6 +859,7 @@ void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 {
 	struct efx_nic *efx = rx_queue->efx;
 	unsigned int max_fill, trigger, max_trigger;
+	int rc = 0;
 
 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 		  "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
@@ -764,6 +893,19 @@ void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 	rx_queue->fast_fill_trigger = trigger;
 	rx_queue->refill_enabled = true;
 
+	/* Initialise XDP queue information */
+	rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev,
+			      rx_queue->core_index);
+
+	if (rc) {
+		netif_err(efx, rx_err, efx->net_dev,
+			  "Failure to initialise XDP queue information rc=%d\n",
+			  rc);
+		efx->xdp_rxq_info_failed = true;
+	} else {
+		rx_queue->xdp_rxq_info_valid = true;
+	}
+
 	/* Set up RX descriptor ring */
 	efx_nic_init_rx(rx_queue);
 }
@@ -805,6 +947,11 @@ void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 	}
 	kfree(rx_queue->page_ring);
 	rx_queue->page_ring = NULL;
+
+	if (rx_queue->xdp_rxq_info_valid)
+		xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info);
+
+	rx_queue->xdp_rxq_info_valid = false;
 }
 
 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)

drivers/net/ethernet/sfc/tx.c

@@ -95,6 +95,8 @@ static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 		netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
 			   "TX queue %d transmission id %x complete\n",
 			   tx_queue->queue, tx_queue->read_count);
+	} else if (buffer->flags & EFX_TX_BUF_XDP) {
+		xdp_return_frame_rx_napi(buffer->xdpf);
 	}
 
 	buffer->len = 0;
@@ -597,6 +599,94 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 	return NETDEV_TX_OK;
 }
 
+static void efx_xdp_return_frames(int n,  struct xdp_frame **xdpfs)
+{
+	int i;
+
+	for (i = 0; i < n; i++)
+		xdp_return_frame_rx_napi(xdpfs[i]);
+}
+
+/* Transmit a packet from an XDP buffer
+ *
+ * Returns number of packets sent on success, error code otherwise.
+ * Runs in NAPI context, either in our poll (for XDP TX) or a different NIC
+ * (for XDP redirect).
+ */
+int efx_xdp_tx_buffers(struct efx_nic *efx, int n, struct xdp_frame **xdpfs,
+		       bool flush)
+{
+	struct efx_tx_buffer *tx_buffer;
+	struct efx_tx_queue *tx_queue;
+	struct xdp_frame *xdpf;
+	dma_addr_t dma_addr;
+	unsigned int len;
+	int space;
+	int cpu;
+	int i;
+
+	cpu = raw_smp_processor_id();
+
+	if (!efx->xdp_tx_queue_count ||
+	    unlikely(cpu >= efx->xdp_tx_queue_count))
+		return -EINVAL;
+
+	tx_queue = efx->xdp_tx_queues[cpu];
+	if (unlikely(!tx_queue))
+		return -EINVAL;
+
+	if (unlikely(n && !xdpfs))
+		return -EINVAL;
+
+	if (!n)
+		return 0;
+
+	/* Check for available space. We should never need multiple
+	 * descriptors per frame.
+	 */
+	space = efx->txq_entries +
+		tx_queue->read_count - tx_queue->insert_count;
+
+	for (i = 0; i < n; i++) {
+		xdpf = xdpfs[i];
+
+		if (i >= space)
+			break;
+
+		/* We'll want a descriptor for this tx. */
+		prefetchw(__efx_tx_queue_get_insert_buffer(tx_queue));
+
+		len = xdpf->len;
+
+		/* Map for DMA. */
+		dma_addr = dma_map_single(&efx->pci_dev->dev,
+					  xdpf->data, len,
+					  DMA_TO_DEVICE);
+		if (dma_mapping_error(&efx->pci_dev->dev, dma_addr))
+			break;
+
+		/*  Create descriptor and set up for unmapping DMA. */
+		tx_buffer = efx_tx_map_chunk(tx_queue, dma_addr, len);
+		tx_buffer->xdpf = xdpf;
+		tx_buffer->flags = EFX_TX_BUF_XDP |
+				   EFX_TX_BUF_MAP_SINGLE;
+		tx_buffer->dma_offset = 0;
+		tx_buffer->unmap_len = len;
+		tx_queue->tx_packets++;
+	}
+
+	/* Pass mapped frames to hardware. */
+	if (flush && i > 0)
+		efx_nic_push_buffers(tx_queue);
+
+	if (i == 0)
+		return -EIO;
+
+	efx_xdp_return_frames(n - i, xdpfs + i);
+
+	return i;
+}
+
 /* Remove packets from the TX queue
  *
  * This removes packets from the TX queue, up to and including the
@@ -857,6 +947,8 @@ void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
 	tx_queue->completed_timestamp_major = 0;
 	tx_queue->completed_timestamp_minor = 0;
 
+	tx_queue->xdp_tx = efx_channel_is_xdp_tx(tx_queue->channel);
+
 	/* Set up default function pointers. These may get replaced by
 	 * efx_nic_init_tx() based off NIC/queue capabilities.
 	 */