sfc: support passing a representor to the EF100 TX path

A non-null efv in __ef100_enqueue_skb() indicates that the packet is
 from that representor, should be transmitted with a suitable option
 descriptor (to instruct the switch to deliver it to the representee),
 and should not be accounted to the parent PF's stats or BQL.
Signed-off-by: Edward Cree <ecree.xilinx@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/sfc/ef100_rep.h

@@ -15,6 +15,12 @@
 
 #include "net_driver.h"
 
+struct efx_rep_sw_stats {
+	atomic64_t rx_packets, tx_packets;
+	atomic64_t rx_bytes, tx_bytes;
+	atomic64_t rx_dropped, tx_errors;
+};
+
 /**
  * struct efx_rep - Private data for an Efx representor
  *
@@ -24,6 +30,7 @@
  * @mport: m-port ID of corresponding VF
  * @idx: VF index
  * @list: entry on efx->vf_reps
+ * @stats: software traffic counters for netdev stats
  */
 struct efx_rep {
 	struct efx_nic *parent;
@@ -32,6 +39,7 @@ struct efx_rep {
 	u32 mport;
 	unsigned int idx;
 	struct list_head list;
+	struct efx_rep_sw_stats stats;
 };
 
 int efx_ef100_vfrep_create(struct efx_nic *efx, unsigned int i);

drivers/net/ethernet/sfc/ef100_tx.c

@@ -254,7 +254,8 @@ static void ef100_make_tso_desc(struct efx_nic *efx,
 
 static void ef100_tx_make_descriptors(struct efx_tx_queue *tx_queue,
 				      const struct sk_buff *skb,
-				      unsigned int segment_count)
+				      unsigned int segment_count,
+				      struct efx_rep *efv)
 {
 	unsigned int old_write_count = tx_queue->write_count;
 	unsigned int new_write_count = old_write_count;
@@ -272,6 +273,20 @@ static void ef100_tx_make_descriptors(struct efx_tx_queue *tx_queue,
 	else
 		next_desc_type = ESE_GZ_TX_DESC_TYPE_SEND;
 
+	if (unlikely(efv)) {
+		/* Create TX override descriptor */
+		write_ptr = new_write_count & tx_queue->ptr_mask;
+		txd = ef100_tx_desc(tx_queue, write_ptr);
+		++new_write_count;
+
+		tx_queue->packet_write_count = new_write_count;
+		EFX_POPULATE_OWORD_3(*txd,
+				     ESF_GZ_TX_DESC_TYPE, ESE_GZ_TX_DESC_TYPE_PREFIX,
+				     ESF_GZ_TX_PREFIX_EGRESS_MPORT, efv->mport,
+				     ESF_GZ_TX_PREFIX_EGRESS_MPORT_EN, 1);
+		nr_descs--;
+	}
+
 	/* if it's a raw write (such as XDP) then always SEND single frames */
 	if (!skb)
 		nr_descs = 1;
@@ -306,6 +321,9 @@ static void ef100_tx_make_descriptors(struct efx_tx_queue *tx_queue,
 		/* if it's a raw write (such as XDP) then always SEND */
 		next_desc_type = skb ? ESE_GZ_TX_DESC_TYPE_SEG :
 				       ESE_GZ_TX_DESC_TYPE_SEND;
+		/* mark as an EFV buffer if applicable */
+		if (unlikely(efv))
+			buffer->flags |= EFX_TX_BUF_EFV;
 
 	} while (new_write_count != tx_queue->insert_count);
 
@@ -324,7 +342,7 @@ static void ef100_tx_make_descriptors(struct efx_tx_queue *tx_queue,
 
 void ef100_tx_write(struct efx_tx_queue *tx_queue)
 {
-	ef100_tx_make_descriptors(tx_queue, NULL, 0);
+	ef100_tx_make_descriptors(tx_queue, NULL, 0, NULL);
 	ef100_tx_push_buffers(tx_queue);
 }
 
@@ -350,6 +368,12 @@ void ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event)
  * function will free the SKB.
  */
 int ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
+{
+	return __ef100_enqueue_skb(tx_queue, skb, NULL);
+}
+
+int __ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb,
+			struct efx_rep *efv)
 {
 	unsigned int old_insert_count = tx_queue->insert_count;
 	struct efx_nic *efx = tx_queue->efx;
@@ -376,16 +400,64 @@ int ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 			return 0;
 	}
 
+	if (unlikely(efv)) {
+		struct efx_tx_buffer *buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
+
+		/* Drop representor packets if the queue is stopped.
+		 * We currently don't assert backoff to representors so this is
+		 * to make sure representor traffic can't starve the main
+		 * net device.
+		 * And, of course, if there are no TX descriptors left.
+		 */
+		if (netif_tx_queue_stopped(tx_queue->core_txq) ||
+		    unlikely(efx_tx_buffer_in_use(buffer))) {
+			atomic64_inc(&efv->stats.tx_errors);
+			rc = -ENOSPC;
+			goto err;
+		}
+
+		/* Also drop representor traffic if it could cause us to
+		 * stop the queue. If we assert backoff and we haven't
+		 * received traffic on the main net device recently then the
+		 * TX watchdog can go off erroneously.
+		 */
+		fill_level = efx_channel_tx_old_fill_level(tx_queue->channel);
+		fill_level += efx_tx_max_skb_descs(efx);
+		if (fill_level > efx->txq_stop_thresh) {
+			struct efx_tx_queue *txq2;
+
+			/* Refresh cached fill level and re-check */
+			efx_for_each_channel_tx_queue(txq2, tx_queue->channel)
+				txq2->old_read_count = READ_ONCE(txq2->read_count);
+
+			fill_level = efx_channel_tx_old_fill_level(tx_queue->channel);
+			fill_level += efx_tx_max_skb_descs(efx);
+			if (fill_level > efx->txq_stop_thresh) {
+				atomic64_inc(&efv->stats.tx_errors);
+				rc = -ENOSPC;
+				goto err;
+			}
+		}
+
+		buffer->flags = EFX_TX_BUF_OPTION | EFX_TX_BUF_EFV;
+		tx_queue->insert_count++;
+	}
+
 	/* Map for DMA and create descriptors */
 	rc = efx_tx_map_data(tx_queue, skb, segments);
 	if (rc)
 		goto err;
-	ef100_tx_make_descriptors(tx_queue, skb, segments);
+	ef100_tx_make_descriptors(tx_queue, skb, segments, efv);
 
 	fill_level = efx_channel_tx_old_fill_level(tx_queue->channel);
 	if (fill_level > efx->txq_stop_thresh) {
 		struct efx_tx_queue *txq2;
 
+		/* Because of checks above, representor traffic should
+		 * not be able to stop the queue.
+		 */
+		WARN_ON(efv);
+
 		netif_tx_stop_queue(tx_queue->core_txq);
 		/* Re-read after a memory barrier in case we've raced with
 		 * the completion path. Otherwise there's a danger we'll never
@@ -404,8 +476,12 @@ int ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 	/* If xmit_more then we don't need to push the doorbell, unless there
 	 * are 256 descriptors already queued in which case we have to push to
 	 * ensure we never push more than 256 at once.
+	 *
+	 * Always push for representor traffic, and don't account it to parent
+	 * PF netdevice's BQL.
 	 */
-	if (__netdev_tx_sent_queue(tx_queue->core_txq, skb->len, xmit_more) ||
+	if (unlikely(efv) ||
+	    __netdev_tx_sent_queue(tx_queue->core_txq, skb->len, xmit_more) ||
 	    tx_queue->write_count - tx_queue->notify_count > 255)
 		ef100_tx_push_buffers(tx_queue);
 

drivers/net/ethernet/sfc/ef100_tx.h

@@ -13,6 +13,7 @@
 #define EFX_EF100_TX_H
 
 #include "net_driver.h"
+#include "ef100_rep.h"
 
 int ef100_tx_probe(struct efx_tx_queue *tx_queue);
 void ef100_tx_init(struct efx_tx_queue *tx_queue);
@@ -22,4 +23,6 @@ unsigned int ef100_tx_max_skb_descs(struct efx_nic *efx);
 void ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event);
 
 netdev_tx_t ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb);
+int __ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb,
+			struct efx_rep *efv);
 #endif

drivers/net/ethernet/sfc/net_driver.h

@@ -178,6 +178,7 @@ struct efx_tx_buffer {
 #define EFX_TX_BUF_OPTION	0x10	/* empty buffer for option descriptor */
 #define EFX_TX_BUF_XDP		0x20	/* buffer was sent with XDP */
 #define EFX_TX_BUF_TSO_V3	0x40	/* empty buffer for a TSO_V3 descriptor */
+#define EFX_TX_BUF_EFV		0x100	/* buffer was sent from representor */
 
 /**
  * struct efx_tx_queue - An Efx TX queue

drivers/net/ethernet/sfc/tx.c

@@ -559,6 +559,7 @@ netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
 void efx_xmit_done_single(struct efx_tx_queue *tx_queue)
 {
 	unsigned int pkts_compl = 0, bytes_compl = 0;
+	unsigned int efv_pkts_compl = 0;
 	unsigned int read_ptr;
 	bool finished = false;
 
@@ -580,7 +581,8 @@ void efx_xmit_done_single(struct efx_tx_queue *tx_queue)
 		/* Need to check the flag before dequeueing. */
 		if (buffer->flags & EFX_TX_BUF_SKB)
 			finished = true;
-		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
+		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl,
+				   &efv_pkts_compl);
 
 		++tx_queue->read_count;
 		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
@@ -589,7 +591,7 @@ void efx_xmit_done_single(struct efx_tx_queue *tx_queue)
 	tx_queue->pkts_compl += pkts_compl;
 	tx_queue->bytes_compl += bytes_compl;
 
-	EFX_WARN_ON_PARANOID(pkts_compl != 1);
+	EFX_WARN_ON_PARANOID(pkts_compl + efv_pkts_compl != 1);
 
 	efx_xmit_done_check_empty(tx_queue);
 }

drivers/net/ethernet/sfc/tx_common.c

@@ -109,9 +109,11 @@ void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
 	/* Free any buffers left in the ring */
 	while (tx_queue->read_count != tx_queue->write_count) {
 		unsigned int pkts_compl = 0, bytes_compl = 0;
+		unsigned int efv_pkts_compl = 0;
 
 		buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
-		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
+		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl,
+				   &efv_pkts_compl);
 
 		++tx_queue->read_count;
 	}
@@ -146,7 +148,8 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
 void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 			struct efx_tx_buffer *buffer,
 			unsigned int *pkts_compl,
-			unsigned int *bytes_compl)
+			unsigned int *bytes_compl,
+			unsigned int *efv_pkts_compl)
 {
 	if (buffer->unmap_len) {
 		struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
@@ -164,9 +167,15 @@ void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 	if (buffer->flags & EFX_TX_BUF_SKB) {
 		struct sk_buff *skb = (struct sk_buff *)buffer->skb;
 
+		if (unlikely(buffer->flags & EFX_TX_BUF_EFV)) {
+			EFX_WARN_ON_PARANOID(!efv_pkts_compl);
+			(*efv_pkts_compl)++;
+		} else {
 			EFX_WARN_ON_PARANOID(!pkts_compl || !bytes_compl);
 			(*pkts_compl)++;
 			(*bytes_compl) += skb->len;
+		}
+
 		if (tx_queue->timestamping &&
 		    (tx_queue->completed_timestamp_major ||
 		     tx_queue->completed_timestamp_minor)) {
@@ -199,7 +208,8 @@ void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
 				unsigned int index,
 				unsigned int *pkts_compl,
-				unsigned int *bytes_compl)
+				unsigned int *bytes_compl,
+				unsigned int *efv_pkts_compl)
 {
 	struct efx_nic *efx = tx_queue->efx;
 	unsigned int stop_index, read_ptr;
@@ -218,7 +228,8 @@ static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
 			return;
 		}
 
-		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
+		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl,
+				   efv_pkts_compl);
 
 		++tx_queue->read_count;
 		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
@@ -241,15 +252,17 @@ void efx_xmit_done_check_empty(struct efx_tx_queue *tx_queue)
 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 {
 	unsigned int fill_level, pkts_compl = 0, bytes_compl = 0;
+	unsigned int efv_pkts_compl = 0;
 	struct efx_nic *efx = tx_queue->efx;
 
 	EFX_WARN_ON_ONCE_PARANOID(index > tx_queue->ptr_mask);
 
-	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
+	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl,
+			    &efv_pkts_compl);
 	tx_queue->pkts_compl += pkts_compl;
 	tx_queue->bytes_compl += bytes_compl;
 
-	if (pkts_compl > 1)
+	if (pkts_compl + efv_pkts_compl > 1)
 		++tx_queue->merge_events;
 
 	/* See if we need to restart the netif queue.  This memory
@@ -274,6 +287,7 @@ void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
 			unsigned int insert_count)
 {
+	unsigned int efv_pkts_compl = 0;
 	struct efx_tx_buffer *buffer;
 	unsigned int bytes_compl = 0;
 	unsigned int pkts_compl = 0;
@@ -282,7 +296,8 @@ void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
 	while (tx_queue->insert_count != insert_count) {
 		--tx_queue->insert_count;
 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
-		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
+		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl,
+				   &efv_pkts_compl);
 	}
 }
 

drivers/net/ethernet/sfc/tx_common.h

@@ -19,7 +19,8 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue);
 void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 			struct efx_tx_buffer *buffer,
 			unsigned int *pkts_compl,
-			unsigned int *bytes_compl);
+			unsigned int *bytes_compl,
+			unsigned int *efv_pkts_compl);
 
 static inline bool efx_tx_buffer_in_use(struct efx_tx_buffer *buffer)
 {