ice: Add support for AF_XDP

Add zero copy AF_XDP support.  This patch adds zero copy support for
Tx and Rx; code for zero copy is added to ice_xsk.h and ice_xsk.c.

For Tx, implement ndo_xsk_wakeup. As with other drivers, reuse
existing XDP Tx queues for this task, since XDP_REDIRECT guarantees
mutual exclusion between different NAPI contexts based on CPU ID. In
turn, a netdev can XDP_REDIRECT to another netdev with a different
NAPI context, since the operation is bound to a specific core and each
core has its own hardware ring.

For Rx, allocate frames as MEM_TYPE_ZERO_COPY on queues that AF_XDP is
enabled.

Signed-off-by: Krzysztof Kazimierczak <krzysztof.kazimierczak@intel.com>
Co-developed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Author: kkazimie

drivers/net/ethernet/intel/ice/Makefile

@@ -21,3 +21,4 @@ ice-y := ice_main.o	\
 	 ice_ethtool.o
 ice-$(CONFIG_PCI_IOV) += ice_virtchnl_pf.o ice_sriov.o
 ice-$(CONFIG_DCB) += ice_dcb.o ice_dcb_lib.o
+ice-$(CONFIG_XDP_SOCKETS) += ice_xsk.o

drivers/net/ethernet/intel/ice/ice.h

@@ -35,6 +35,7 @@
 #include <linux/bpf.h>
 #include <linux/avf/virtchnl.h>
 #include <net/ipv6.h>
+#include <net/xdp_sock.h>
 #include "ice_devids.h"
 #include "ice_type.h"
 #include "ice_txrx.h"
@@ -44,6 +45,7 @@
 #include "ice_sched.h"
 #include "ice_virtchnl_pf.h"
 #include "ice_sriov.h"
+#include "ice_xsk.h"
 
 extern const char ice_drv_ver[];
 #define ICE_BAR0		0
@@ -287,6 +289,9 @@ struct ice_vsi {
 	struct ice_ring **xdp_rings;	 /* XDP ring array */
 	u16 num_xdp_txq;		 /* Used XDP queues */
 	u8 xdp_mapping_mode;		 /* ICE_MAP_MODE_[CONTIG|SCATTER] */
+	struct xdp_umem **xsk_umems;
+	u16 num_xsk_umems_used;
+	u16 num_xsk_umems;
 } ____cacheline_internodealigned_in_smp;
 
 /* struct that defines an interrupt vector */
@@ -440,6 +445,27 @@ static inline void ice_set_ring_xdp(struct ice_ring *ring)
 	ring->flags |= ICE_TX_FLAGS_RING_XDP;
 }
 
+/**
+ * ice_xsk_umem - get XDP UMEM bound to a ring
+ * @ring - ring to use
+ *
+ * Returns a pointer to xdp_umem structure if there is an UMEM present,
+ * NULL otherwise.
+ */
+static inline struct xdp_umem *ice_xsk_umem(struct ice_ring *ring)
+{
+	struct xdp_umem **umems = ring->vsi->xsk_umems;
+	int qid = ring->q_index;
+
+	if (ice_ring_is_xdp(ring))
+		qid -= ring->vsi->num_xdp_txq;
+
+	if (!umems || !umems[qid] || !ice_is_xdp_ena_vsi(ring->vsi))
+		return NULL;
+
+	return umems[qid];
+}
+
 /**
  * ice_get_main_vsi - Get the PF VSI
  * @pf: PF instance

drivers/net/ethernet/intel/ice/ice_base.c

@@ -276,14 +276,17 @@ ice_setup_tx_ctx(struct ice_ring *ring, struct ice_tlan_ctx *tlan_ctx, u16 pf_q)
  */
 int ice_setup_rx_ctx(struct ice_ring *ring)
 {
+	int chain_len = ICE_MAX_CHAINED_RX_BUFS;
 	struct ice_vsi *vsi = ring->vsi;
-	struct ice_hw *hw = &vsi->back->hw;
 	u32 rxdid = ICE_RXDID_FLEX_NIC;
 	struct ice_rlan_ctx rlan_ctx;
+	struct ice_hw *hw;
 	u32 regval;
 	u16 pf_q;
 	int err;
 
+	hw = &vsi->back->hw;
+
 	/* what is Rx queue number in global space of 2K Rx queues */
 	pf_q = vsi->rxq_map[ring->q_index];
 
@@ -297,10 +300,38 @@ int ice_setup_rx_ctx(struct ice_ring *ring)
 			xdp_rxq_info_reg(&ring->xdp_rxq, ring->netdev,
 					 ring->q_index);
 
-		err = xdp_rxq_info_reg_mem_model(&ring->xdp_rxq,
-						 MEM_TYPE_PAGE_SHARED, NULL);
-		if (err)
-			return err;
+		ring->xsk_umem = ice_xsk_umem(ring);
+		if (ring->xsk_umem) {
+			xdp_rxq_info_unreg_mem_model(&ring->xdp_rxq);
+
+			ring->rx_buf_len = ring->xsk_umem->chunk_size_nohr -
+					   XDP_PACKET_HEADROOM;
+			/* For AF_XDP ZC, we disallow packets to span on
+			 * multiple buffers, thus letting us skip that
+			 * handling in the fast-path.
+			 */
+			chain_len = 1;
+			ring->zca.free = ice_zca_free;
+			err = xdp_rxq_info_reg_mem_model(&ring->xdp_rxq,
+							 MEM_TYPE_ZERO_COPY,
+							 &ring->zca);
+			if (err)
+				return err;
+
+			dev_info(&vsi->back->pdev->dev, "Registered XDP mem model MEM_TYPE_ZERO_COPY on Rx ring %d\n",
+				 ring->q_index);
+		} else {
+			if (!xdp_rxq_info_is_reg(&ring->xdp_rxq))
+				xdp_rxq_info_reg(&ring->xdp_rxq,
+						 ring->netdev,
+						 ring->q_index);
+
+			err = xdp_rxq_info_reg_mem_model(&ring->xdp_rxq,
+							 MEM_TYPE_PAGE_SHARED,
+							 NULL);
+			if (err)
+				return err;
+		}
 	}
 	/* Receive Queue Base Address.
 	 * Indicates the starting address of the descriptor queue defined in
@@ -340,7 +371,7 @@ int ice_setup_rx_ctx(struct ice_ring *ring)
 	 * than 5 x DBUF
 	 */
 	rlan_ctx.rxmax = min_t(u16, vsi->max_frame,
-			       ICE_MAX_CHAINED_RX_BUFS * vsi->rx_buf_len);
+			       chain_len * ring->rx_buf_len);
 
 	/* Rx queue threshold in units of 64 */
 	rlan_ctx.lrxqthresh = 1;
@@ -378,7 +409,15 @@ int ice_setup_rx_ctx(struct ice_ring *ring)
 	/* init queue specific tail register */
 	ring->tail = hw->hw_addr + QRX_TAIL(pf_q);
 	writel(0, ring->tail);
-	ice_alloc_rx_bufs(ring, ICE_DESC_UNUSED(ring));
+
+	err = ring->xsk_umem ?
+	      ice_alloc_rx_bufs_slow_zc(ring, ICE_DESC_UNUSED(ring)) :
+	      ice_alloc_rx_bufs(ring, ICE_DESC_UNUSED(ring));
+	if (err)
+		dev_info(&vsi->back->pdev->dev,
+			 "Failed allocate some buffers on %sRx ring %d (pf_q %d)\n",
+			 ring->xsk_umem ? "UMEM enabled " : "",
+			 ring->q_index, pf_q);
 
 	return 0;
 }

drivers/net/ethernet/intel/ice/ice_ethtool.c

@@ -2612,6 +2612,13 @@ ice_set_ringparam(struct net_device *netdev, struct ethtool_ringparam *ring)
 		return 0;
 	}
 
+	/* If there is a AF_XDP UMEM attached to any of Rx rings,
+	 * disallow changing the number of descriptors -- regardless
+	 * if the netdev is running or not.
+	 */
+	if (ice_xsk_any_rx_ring_ena(vsi))
+		return -EBUSY;
+
 	while (test_and_set_bit(__ICE_CFG_BUSY, pf->state)) {
 		timeout--;
 		if (!timeout)

drivers/net/ethernet/intel/ice/ice_lib.c

@@ -1283,7 +1283,17 @@ int ice_vsi_cfg_lan_txqs(struct ice_vsi *vsi)
  */
 int ice_vsi_cfg_xdp_txqs(struct ice_vsi *vsi)
 {
-	return ice_vsi_cfg_txqs(vsi, vsi->xdp_rings);
+	int ret;
+	int i;
+
+	ret = ice_vsi_cfg_txqs(vsi, vsi->xdp_rings);
+	if (ret)
+		return ret;
+
+	for (i = 0; i < vsi->num_xdp_txq; i++)
+		vsi->xdp_rings[i]->xsk_umem = ice_xsk_umem(vsi->xdp_rings[i]);
+
+	return ret;
 }
 
 /**
@@ -2514,6 +2524,51 @@ char *ice_nvm_version_str(struct ice_hw *hw)
 	return buf;
 }
 
+/**
+ * ice_update_ring_stats - Update ring statistics
+ * @ring: ring to update
+ * @cont: used to increment per-vector counters
+ * @pkts: number of processed packets
+ * @bytes: number of processed bytes
+ *
+ * This function assumes that caller has acquired a u64_stats_sync lock.
+ */
+static void
+ice_update_ring_stats(struct ice_ring *ring, struct ice_ring_container *cont,
+		      u64 pkts, u64 bytes)
+{
+	ring->stats.bytes += bytes;
+	ring->stats.pkts += pkts;
+	cont->total_bytes += bytes;
+	cont->total_pkts += pkts;
+}
+
+/**
+ * ice_update_tx_ring_stats - Update Tx ring specific counters
+ * @tx_ring: ring to update
+ * @pkts: number of processed packets
+ * @bytes: number of processed bytes
+ */
+void ice_update_tx_ring_stats(struct ice_ring *tx_ring, u64 pkts, u64 bytes)
+{
+	u64_stats_update_begin(&tx_ring->syncp);
+	ice_update_ring_stats(tx_ring, &tx_ring->q_vector->tx, pkts, bytes);
+	u64_stats_update_end(&tx_ring->syncp);
+}
+
+/**
+ * ice_update_rx_ring_stats - Update Rx ring specific counters
+ * @rx_ring: ring to update
+ * @pkts: number of processed packets
+ * @bytes: number of processed bytes
+ */
+void ice_update_rx_ring_stats(struct ice_ring *rx_ring, u64 pkts, u64 bytes)
+{
+	u64_stats_update_begin(&rx_ring->syncp);
+	ice_update_ring_stats(rx_ring, &rx_ring->q_vector->rx, pkts, bytes);
+	u64_stats_update_end(&rx_ring->syncp);
+}
+
 /**
  * ice_vsi_cfg_mac_fltr - Add or remove a MAC address filter for a VSI
  * @vsi: the VSI being configured MAC filter

drivers/net/ethernet/intel/ice/ice_lib.h

@@ -83,6 +83,10 @@ void ice_vsi_free_tx_rings(struct ice_vsi *vsi);
 
 int ice_vsi_manage_rss_lut(struct ice_vsi *vsi, bool ena);
 
+void ice_update_tx_ring_stats(struct ice_ring *ring, u64 pkts, u64 bytes);
+
+void ice_update_rx_ring_stats(struct ice_ring *ring, u64 pkts, u64 bytes);
+
 void ice_vsi_cfg_frame_size(struct ice_vsi *vsi);
 
 u32 ice_intrl_usec_to_reg(u8 intrl, u8 gran);

drivers/net/ethernet/intel/ice/ice_main.c

@@ -1692,6 +1692,7 @@ static int ice_xdp_alloc_setup_rings(struct ice_vsi *vsi)
 		if (ice_setup_tx_ring(xdp_ring))
 			goto free_xdp_rings;
 		ice_set_ring_xdp(xdp_ring);
+		xdp_ring->xsk_umem = ice_xsk_umem(xdp_ring);
 	}
 
 	return 0;
@@ -1934,6 +1935,17 @@ ice_xdp_setup_prog(struct ice_vsi *vsi, struct bpf_prog *prog,
 	if (if_running)
 		ret = ice_up(vsi);
 
+	if (!ret && prog && vsi->xsk_umems) {
+		int i;
+
+		ice_for_each_rxq(vsi, i) {
+			struct ice_ring *rx_ring = vsi->rx_rings[i];
+
+			if (rx_ring->xsk_umem)
+				napi_schedule(&rx_ring->q_vector->napi);
+		}
+	}
+
 	return (ret || xdp_ring_err) ? -ENOMEM : 0;
 }
 
@@ -1959,6 +1971,9 @@ static int ice_xdp(struct net_device *dev, struct netdev_bpf *xdp)
 	case XDP_QUERY_PROG:
 		xdp->prog_id = vsi->xdp_prog ? vsi->xdp_prog->aux->id : 0;
 		return 0;
+	case XDP_SETUP_XSK_UMEM:
+		return ice_xsk_umem_setup(vsi, xdp->xsk.umem,
+					  xdp->xsk.queue_id);
 	default:
 		return -EINVAL;
 	}
@@ -5205,4 +5220,5 @@ static const struct net_device_ops ice_netdev_ops = {
 	.ndo_tx_timeout = ice_tx_timeout,
 	.ndo_bpf = ice_xdp,
 	.ndo_xdp_xmit = ice_xdp_xmit,
+	.ndo_xsk_wakeup = ice_xsk_wakeup,
 };

drivers/net/ethernet/intel/ice/ice_txrx.c

@@ -11,6 +11,7 @@
 #include "ice_lib.h"
 #include "ice.h"
 #include "ice_dcb_lib.h"
+#include "ice_xsk.h"
 
 #define ICE_RX_HDR_SIZE		256
 
@@ -58,6 +59,11 @@ void ice_clean_tx_ring(struct ice_ring *tx_ring)
 {
 	u16 i;
 
+	if (ice_ring_is_xdp(tx_ring) && tx_ring->xsk_umem) {
+		ice_xsk_clean_xdp_ring(tx_ring);
+		goto tx_skip_free;
+	}
+
 	/* ring already cleared, nothing to do */
 	if (!tx_ring->tx_buf)
 		return;
@@ -66,6 +72,7 @@ void ice_clean_tx_ring(struct ice_ring *tx_ring)
 	for (i = 0; i < tx_ring->count; i++)
 		ice_unmap_and_free_tx_buf(tx_ring, &tx_ring->tx_buf[i]);
 
+tx_skip_free:
 	memset(tx_ring->tx_buf, 0, sizeof(*tx_ring->tx_buf) * tx_ring->count);
 
 	/* Zero out the descriptor ring */
@@ -198,12 +205,8 @@ static bool ice_clean_tx_irq(struct ice_ring *tx_ring, int napi_budget)
 
 	i += tx_ring->count;
 	tx_ring->next_to_clean = i;
-	u64_stats_update_begin(&tx_ring->syncp);
-	tx_ring->stats.bytes += total_bytes;
-	tx_ring->stats.pkts += total_pkts;
-	u64_stats_update_end(&tx_ring->syncp);
-	tx_ring->q_vector->tx.total_bytes += total_bytes;
-	tx_ring->q_vector->tx.total_pkts += total_pkts;
+
+	ice_update_tx_ring_stats(tx_ring, total_pkts, total_bytes);
 
 	if (ice_ring_is_xdp(tx_ring))
 		return !!budget;
@@ -286,6 +289,11 @@ void ice_clean_rx_ring(struct ice_ring *rx_ring)
 	if (!rx_ring->rx_buf)
 		return;
 
+	if (rx_ring->xsk_umem) {
+		ice_xsk_clean_rx_ring(rx_ring);
+		goto rx_skip_free;
+	}
+
 	/* Free all the Rx ring sk_buffs */
 	for (i = 0; i < rx_ring->count; i++) {
 		struct ice_rx_buf *rx_buf = &rx_ring->rx_buf[i];
@@ -313,6 +321,7 @@ void ice_clean_rx_ring(struct ice_ring *rx_ring)
 		rx_buf->page_offset = 0;
 	}
 
+rx_skip_free:
 	memset(rx_ring->rx_buf, 0, sizeof(*rx_ring->rx_buf) * rx_ring->count);
 
 	/* Zero out the descriptor ring */
@@ -1073,13 +1082,7 @@ static int ice_clean_rx_irq(struct ice_ring *rx_ring, int budget)
 	if (xdp_prog)
 		ice_finalize_xdp_rx(rx_ring, xdp_xmit);
 
-	/* update queue and vector specific stats */
-	u64_stats_update_begin(&rx_ring->syncp);
-	rx_ring->stats.pkts += total_rx_pkts;
-	rx_ring->stats.bytes += total_rx_bytes;
-	u64_stats_update_end(&rx_ring->syncp);
-	rx_ring->q_vector->rx.total_pkts += total_rx_pkts;
-	rx_ring->q_vector->rx.total_bytes += total_rx_bytes;
+	ice_update_rx_ring_stats(rx_ring, total_rx_pkts, total_rx_bytes);
 
 	/* guarantee a trip back through this routine if there was a failure */
 	return failure ? budget : (int)total_rx_pkts;
@@ -1457,9 +1460,14 @@ int ice_napi_poll(struct napi_struct *napi, int budget)
 	/* Since the actual Tx work is minimal, we can give the Tx a larger
 	 * budget and be more aggressive about cleaning up the Tx descriptors.
 	 */
-	ice_for_each_ring(ring, q_vector->tx)
-		if (!ice_clean_tx_irq(ring, budget))
+	ice_for_each_ring(ring, q_vector->tx) {
+		bool wd = ring->xsk_umem ?
+			  ice_clean_tx_irq_zc(ring, budget) :
+			  ice_clean_tx_irq(ring, budget);
+
+		if (!wd)
 			clean_complete = false;
+	}
 
 	/* Handle case where we are called by netpoll with a budget of 0 */
 	if (unlikely(budget <= 0))
@@ -1479,7 +1487,13 @@ int ice_napi_poll(struct napi_struct *napi, int budget)
 	ice_for_each_ring(ring, q_vector->rx) {
 		int cleaned;
 
-		cleaned = ice_clean_rx_irq(ring, budget_per_ring);
+		/* A dedicated path for zero-copy allows making a single
+		 * comparison in the irq context instead of many inside the
+		 * ice_clean_rx_irq function and makes the codebase cleaner.
+		 */
+		cleaned = ring->xsk_umem ?
+			  ice_clean_rx_irq_zc(ring, budget_per_ring) :
+			  ice_clean_rx_irq(ring, budget_per_ring);
 		work_done += cleaned;
 		/* if we clean as many as budgeted, we must not be done */
 		if (cleaned >= budget_per_ring)