igb: Add AF_XDP zero-copy Tx support

Add support for AF_XDP zero-copy transmit path.

A new TX buffer type IGB_TYPE_XSK is introduced to indicate that the Tx
frame was allocated from the xsk buff pool, so igb_clean_tx_ring() and
igb_clean_tx_irq() can clean the buffers correctly based on type.

igb_xmit_zc() performs the actual packet transmit when AF_XDP zero-copy is
enabled. We share the TX ring between slow path, XDP and AF_XDP
zero-copy, so we use the netdev queue lock to ensure mutual exclusion.

Signed-off-by: Sriram Yagnaraman <sriram.yagnaraman@est.tech>
[Kurt: Set olinfo_status in igb_xmit_zc() so that frames are transmitted,
       Use READ_ONCE() for xsk_pool and check Tx disabled and carrier in
       igb_xmit_zc(), Add FIXME for RS bit]
Signed-off-by: Kurt Kanzenbach <kurt@linutronix.de>
Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: George Kuruvinakunnel <george.kuruvinakunnel@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20250106221929.956999-7-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

drivers/net/ethernet/intel/igb/igb.h

@@ -258,6 +258,7 @@ enum igb_tx_flags {
 enum igb_tx_buf_type {
 	IGB_TYPE_SKB = 0,
 	IGB_TYPE_XDP,
+	IGB_TYPE_XSK
 };
 
 /* wrapper around a pointer to a socket buffer,
@@ -859,6 +860,7 @@ bool igb_alloc_rx_buffers_zc(struct igb_ring *rx_ring,
 void igb_clean_rx_ring_zc(struct igb_ring *rx_ring);
 int igb_clean_rx_irq_zc(struct igb_q_vector *q_vector,
 			struct xsk_buff_pool *xsk_pool, const int budget);
+bool igb_xmit_zc(struct igb_ring *tx_ring, struct xsk_buff_pool *xsk_pool);
 int igb_xsk_wakeup(struct net_device *dev, u32 qid, u32 flags);
 
 #endif /* _IGB_H_ */

drivers/net/ethernet/intel/igb/igb_main.c

@@ -2983,6 +2983,9 @@ static int igb_xdp_xmit(struct net_device *dev, int n,
 	if (unlikely(!tx_ring))
 		return -ENXIO;
 
+	if (unlikely(test_bit(IGB_RING_FLAG_TX_DISABLED, &tx_ring->flags)))
+		return -ENXIO;
+
 	nq = txring_txq(tx_ring);
 	__netif_tx_lock(nq, cpu);
 
@@ -3330,7 +3333,8 @@ static int igb_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 	netdev->priv_flags |= IFF_SUPP_NOFCS;
 
 	netdev->priv_flags |= IFF_UNICAST_FLT;
-	netdev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT;
+	netdev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT |
+			       NETDEV_XDP_ACT_XSK_ZEROCOPY;
 
 	/* MTU range: 68 - 9216 */
 	netdev->min_mtu = ETH_MIN_MTU;
@@ -4904,15 +4908,20 @@ void igb_clean_tx_ring(struct igb_ring *tx_ring)
 {
 	u16 i = tx_ring->next_to_clean;
 	struct igb_tx_buffer *tx_buffer = &tx_ring->tx_buffer_info[i];
+	u32 xsk_frames = 0;
 
 	while (i != tx_ring->next_to_use) {
 		union e1000_adv_tx_desc *eop_desc, *tx_desc;
 
 		/* Free all the Tx ring sk_buffs or xdp frames */
-		if (tx_buffer->type == IGB_TYPE_SKB)
+		if (tx_buffer->type == IGB_TYPE_SKB) {
 			dev_kfree_skb_any(tx_buffer->skb);
-		else
+		} else if (tx_buffer->type == IGB_TYPE_XDP) {
 			xdp_return_frame(tx_buffer->xdpf);
+		} else if (tx_buffer->type == IGB_TYPE_XSK) {
+			xsk_frames++;
+			goto skip_for_xsk;
+		}
 
 		/* unmap skb header data */
 		dma_unmap_single(tx_ring->dev,
@@ -4943,6 +4952,7 @@ void igb_clean_tx_ring(struct igb_ring *tx_ring)
 					       DMA_TO_DEVICE);
 		}
 
+skip_for_xsk:
 		tx_buffer->next_to_watch = NULL;
 
 		/* move us one more past the eop_desc for start of next pkt */
@@ -4957,6 +4967,9 @@ void igb_clean_tx_ring(struct igb_ring *tx_ring)
 	/* reset BQL for queue */
 	netdev_tx_reset_queue(txring_txq(tx_ring));
 
+	if (tx_ring->xsk_pool && xsk_frames)
+		xsk_tx_completed(tx_ring->xsk_pool, xsk_frames);
+
 	/* reset next_to_use and next_to_clean */
 	tx_ring->next_to_use = 0;
 	tx_ring->next_to_clean = 0;
@@ -6490,6 +6503,9 @@ netdev_tx_t igb_xmit_frame_ring(struct sk_buff *skb,
 		return NETDEV_TX_BUSY;
 	}
 
+	if (unlikely(test_bit(IGB_RING_FLAG_TX_DISABLED, &tx_ring->flags)))
+		return NETDEV_TX_BUSY;
+
 	/* record the location of the first descriptor for this packet */
 	first = &tx_ring->tx_buffer_info[tx_ring->next_to_use];
 	first->type = IGB_TYPE_SKB;
@@ -8264,13 +8280,18 @@ static int igb_poll(struct napi_struct *napi, int budget)
  **/
 static bool igb_clean_tx_irq(struct igb_q_vector *q_vector, int napi_budget)
 {
-	struct igb_adapter *adapter = q_vector->adapter;
-	struct igb_ring *tx_ring = q_vector->tx.ring;
-	struct igb_tx_buffer *tx_buffer;
-	union e1000_adv_tx_desc *tx_desc;
 	unsigned int total_bytes = 0, total_packets = 0;
+	struct igb_adapter *adapter = q_vector->adapter;
 	unsigned int budget = q_vector->tx.work_limit;
+	struct igb_ring *tx_ring = q_vector->tx.ring;
 	unsigned int i = tx_ring->next_to_clean;
+	union e1000_adv_tx_desc *tx_desc;
+	struct igb_tx_buffer *tx_buffer;
+	struct xsk_buff_pool *xsk_pool;
+	int cpu = smp_processor_id();
+	bool xsk_xmit_done = true;
+	struct netdev_queue *nq;
+	u32 xsk_frames = 0;
 
 	if (test_bit(__IGB_DOWN, &adapter->state))
 		return true;
@@ -8301,10 +8322,14 @@ static bool igb_clean_tx_irq(struct igb_q_vector *q_vector, int napi_budget)
 		total_packets += tx_buffer->gso_segs;
 
 		/* free the skb */
-		if (tx_buffer->type == IGB_TYPE_SKB)
+		if (tx_buffer->type == IGB_TYPE_SKB) {
 			napi_consume_skb(tx_buffer->skb, napi_budget);
-		else
+		} else if (tx_buffer->type == IGB_TYPE_XDP) {
 			xdp_return_frame(tx_buffer->xdpf);
+		} else if (tx_buffer->type == IGB_TYPE_XSK) {
+			xsk_frames++;
+			goto skip_for_xsk;
+		}
 
 		/* unmap skb header data */
 		dma_unmap_single(tx_ring->dev,
@@ -8336,6 +8361,7 @@ static bool igb_clean_tx_irq(struct igb_q_vector *q_vector, int napi_budget)
 			}
 		}
 
+skip_for_xsk:
 		/* move us one more past the eop_desc for start of next pkt */
 		tx_buffer++;
 		tx_desc++;
@@ -8364,6 +8390,21 @@ static bool igb_clean_tx_irq(struct igb_q_vector *q_vector, int napi_budget)
 	q_vector->tx.total_bytes += total_bytes;
 	q_vector->tx.total_packets += total_packets;
 
+	xsk_pool = READ_ONCE(tx_ring->xsk_pool);
+	if (xsk_pool) {
+		if (xsk_frames)
+			xsk_tx_completed(xsk_pool, xsk_frames);
+		if (xsk_uses_need_wakeup(xsk_pool))
+			xsk_set_tx_need_wakeup(xsk_pool);
+
+		nq = txring_txq(tx_ring);
+		__netif_tx_lock(nq, cpu);
+		/* Avoid transmit queue timeout since we share it with the slow path */
+		txq_trans_cond_update(nq);
+		xsk_xmit_done = igb_xmit_zc(tx_ring, xsk_pool);
+		__netif_tx_unlock(nq);
+	}
+
 	if (test_bit(IGB_RING_FLAG_TX_DETECT_HANG, &tx_ring->flags)) {
 		struct e1000_hw *hw = &adapter->hw;
 
@@ -8426,7 +8467,7 @@ static bool igb_clean_tx_irq(struct igb_q_vector *q_vector, int napi_budget)
 		}
 	}
 
-	return !!budget;
+	return !!budget && xsk_xmit_done;
 }
 
 /**

drivers/net/ethernet/intel/igb/igb_xsk.c

@@ -461,6 +461,69 @@ int igb_clean_rx_irq_zc(struct igb_q_vector *q_vector,
 	return failure ? budget : (int)total_packets;
 }
 
+bool igb_xmit_zc(struct igb_ring *tx_ring, struct xsk_buff_pool *xsk_pool)
+{
+	unsigned int budget = igb_desc_unused(tx_ring);
+	u32 cmd_type, olinfo_status, nb_pkts, i = 0;
+	struct xdp_desc *descs = xsk_pool->tx_descs;
+	union e1000_adv_tx_desc *tx_desc = NULL;
+	struct igb_tx_buffer *tx_buffer_info;
+	unsigned int total_bytes = 0;
+	dma_addr_t dma;
+
+	if (!netif_carrier_ok(tx_ring->netdev))
+		return true;
+
+	if (test_bit(IGB_RING_FLAG_TX_DISABLED, &tx_ring->flags))
+		return true;
+
+	nb_pkts = xsk_tx_peek_release_desc_batch(xsk_pool, budget);
+	if (!nb_pkts)
+		return true;
+
+	while (nb_pkts-- > 0) {
+		dma = xsk_buff_raw_get_dma(xsk_pool, descs[i].addr);
+		xsk_buff_raw_dma_sync_for_device(xsk_pool, dma, descs[i].len);
+
+		tx_buffer_info = &tx_ring->tx_buffer_info[tx_ring->next_to_use];
+		tx_buffer_info->bytecount = descs[i].len;
+		tx_buffer_info->type = IGB_TYPE_XSK;
+		tx_buffer_info->xdpf = NULL;
+		tx_buffer_info->gso_segs = 1;
+		tx_buffer_info->time_stamp = jiffies;
+
+		tx_desc = IGB_TX_DESC(tx_ring, tx_ring->next_to_use);
+		tx_desc->read.buffer_addr = cpu_to_le64(dma);
+
+		/* put descriptor type bits */
+		cmd_type = E1000_ADVTXD_DTYP_DATA | E1000_ADVTXD_DCMD_DEXT |
+			   E1000_ADVTXD_DCMD_IFCS;
+		olinfo_status = descs[i].len << E1000_ADVTXD_PAYLEN_SHIFT;
+
+		/* FIXME: This sets the Report Status (RS) bit for every
+		 * descriptor. One nice to have optimization would be to set it
+		 * only for the last descriptor in the whole batch. See Intel
+		 * ice driver for an example on how to do it.
+		 */
+		cmd_type |= descs[i].len | IGB_TXD_DCMD;
+		tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type);
+		tx_desc->read.olinfo_status = cpu_to_le32(olinfo_status);
+
+		total_bytes += descs[i].len;
+
+		i++;
+		tx_ring->next_to_use++;
+		tx_buffer_info->next_to_watch = tx_desc;
+		if (tx_ring->next_to_use == tx_ring->count)
+			tx_ring->next_to_use = 0;
+	}
+
+	netdev_tx_sent_queue(txring_txq(tx_ring), total_bytes);
+	igb_xdp_ring_update_tail(tx_ring);
+
+	return nb_pkts < budget;
+}
+
 int igb_xsk_wakeup(struct net_device *dev, u32 qid, u32 flags)
 {
 	struct igb_adapter *adapter = netdev_priv(dev);