gve: Do lazy cleanup in TX path

When TX queue is full, attemt to process enough TX completions
to avoid stalling the queue.

Fixes: f5cedc84 ("gve: Add transmit and receive support")
Signed-off-by: Tao Liu <xliutaox@google.com>
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/google/gve/gve.h

@@ -341,8 +341,8 @@ struct gve_tx_ring {
 	union {
 		/* GQI fields */
 		struct {
-			/* NIC tail pointer */
-			__be32 last_nic_done;
+			/* Spinlock for when cleanup in progress */
+			spinlock_t clean_lock;
 		};
 
 		/* DQO fields. */
@@ -821,8 +821,9 @@ netdev_tx_t gve_tx(struct sk_buff *skb, struct net_device *dev);
 bool gve_tx_poll(struct gve_notify_block *block, int budget);
 int gve_tx_alloc_rings(struct gve_priv *priv);
 void gve_tx_free_rings_gqi(struct gve_priv *priv);
-__be32 gve_tx_load_event_counter(struct gve_priv *priv,
+u32 gve_tx_load_event_counter(struct gve_priv *priv,
 			      struct gve_tx_ring *tx);
+bool gve_tx_clean_pending(struct gve_priv *priv, struct gve_tx_ring *tx);
 /* rx handling */
 void gve_rx_write_doorbell(struct gve_priv *priv, struct gve_rx_ring *rx);
 int gve_rx_poll(struct gve_notify_block *block, int budget);

drivers/net/ethernet/google/gve/gve_ethtool.c

@@ -330,8 +330,7 @@ gve_get_ethtool_stats(struct net_device *netdev,
 			data[i++] = tmp_tx_bytes;
 			data[i++] = tx->wake_queue;
 			data[i++] = tx->stop_queue;
-			data[i++] = be32_to_cpu(gve_tx_load_event_counter(priv,
-									  tx));
+			data[i++] = gve_tx_load_event_counter(priv, tx);
 			data[i++] = tx->dma_mapping_error;
 			/* stats from NIC */
 			if (skip_nic_stats) {

drivers/net/ethernet/google/gve/gve_main.c

@@ -212,13 +212,13 @@ static int gve_napi_poll(struct napi_struct *napi, int budget)
 		irq_doorbell = gve_irq_doorbell(priv, block);
 		iowrite32be(GVE_IRQ_ACK | GVE_IRQ_EVENT, irq_doorbell);
 
-		/* Double check we have no extra work.
-		 * Ensure unmask synchronizes with checking for work.
+		/* Ensure IRQ ACK is visible before we check pending work.
+		 * If queue had issued updates, it would be truly visible.
 		 */
 		mb();
 
 		if (block->tx)
-			reschedule |= gve_tx_poll(block, -1);
+			reschedule |= gve_tx_clean_pending(priv, block->tx);
 		if (block->rx)
 			reschedule |= gve_rx_work_pending(block->rx);
 

drivers/net/ethernet/google/gve/gve_tx.c

@@ -144,7 +144,7 @@ static void gve_tx_free_ring(struct gve_priv *priv, int idx)
 
 	gve_tx_remove_from_block(priv, idx);
 	slots = tx->mask + 1;
-	gve_clean_tx_done(priv, tx, tx->req, false);
+	gve_clean_tx_done(priv, tx, priv->tx_desc_cnt, false);
 	netdev_tx_reset_queue(tx->netdev_txq);
 
 	dma_free_coherent(hdev, sizeof(*tx->q_resources),
@@ -176,6 +176,7 @@ static int gve_tx_alloc_ring(struct gve_priv *priv, int idx)
 
 	/* Make sure everything is zeroed to start */
 	memset(tx, 0, sizeof(*tx));
+	spin_lock_init(&tx->clean_lock);
 	tx->q_num = idx;
 
 	tx->mask = slots - 1;
@@ -328,10 +329,16 @@ static inline bool gve_can_tx(struct gve_tx_ring *tx, int bytes_required)
 	return (gve_tx_avail(tx) >= MAX_TX_DESC_NEEDED && can_alloc);
 }
 
+static_assert(NAPI_POLL_WEIGHT >= MAX_TX_DESC_NEEDED);
+
 /* Stops the queue if the skb cannot be transmitted. */
-static int gve_maybe_stop_tx(struct gve_tx_ring *tx, struct sk_buff *skb)
+static int gve_maybe_stop_tx(struct gve_priv *priv, struct gve_tx_ring *tx,
+			     struct sk_buff *skb)
 {
 	int bytes_required = 0;
+	u32 nic_done;
+	u32 to_do;
+	int ret;
 
 	if (!tx->raw_addressing)
 		bytes_required = gve_skb_fifo_bytes_required(tx, skb);
@@ -339,29 +346,28 @@ static int gve_maybe_stop_tx(struct gve_tx_ring *tx, struct sk_buff *skb)
 	if (likely(gve_can_tx(tx, bytes_required)))
 		return 0;
 
+	ret = -EBUSY;
+	spin_lock(&tx->clean_lock);
+	nic_done = gve_tx_load_event_counter(priv, tx);
+	to_do = nic_done - tx->done;
+
+	/* Only try to clean if there is hope for TX */
+	if (to_do + gve_tx_avail(tx) >= MAX_TX_DESC_NEEDED) {
+		if (to_do > 0) {
+			to_do = min_t(u32, to_do, NAPI_POLL_WEIGHT);
+			gve_clean_tx_done(priv, tx, to_do, false);
+		}
+		if (likely(gve_can_tx(tx, bytes_required)))
+			ret = 0;
+	}
+	if (ret) {
 		/* No space, so stop the queue */
 		tx->stop_queue++;
 		netif_tx_stop_queue(tx->netdev_txq);
-	smp_mb();	/* sync with restarting queue in gve_clean_tx_done() */
-
-	/* Now check for resources again, in case gve_clean_tx_done() freed
-	 * resources after we checked and we stopped the queue after
-	 * gve_clean_tx_done() checked.
-	 *
-	 * gve_maybe_stop_tx()			gve_clean_tx_done()
-	 *   nsegs/can_alloc test failed
-	 *					  gve_tx_free_fifo()
-	 *					  if (tx queue stopped)
-	 *					    netif_tx_queue_wake()
-	 *   netif_tx_stop_queue()
-	 *   Need to check again for space here!
-	 */
-	if (likely(!gve_can_tx(tx, bytes_required)))
-		return -EBUSY;
+	}
+	spin_unlock(&tx->clean_lock);
 
-	netif_tx_start_queue(tx->netdev_txq);
-	tx->wake_queue++;
-	return 0;
+	return ret;
 }
 
 static void gve_tx_fill_pkt_desc(union gve_tx_desc *pkt_desc,
@@ -576,7 +582,7 @@ netdev_tx_t gve_tx(struct sk_buff *skb, struct net_device *dev)
 	WARN(skb_get_queue_mapping(skb) >= priv->tx_cfg.num_queues,
 	     "skb queue index out of range");
 	tx = &priv->tx[skb_get_queue_mapping(skb)];
-	if (unlikely(gve_maybe_stop_tx(tx, skb))) {
+	if (unlikely(gve_maybe_stop_tx(priv, tx, skb))) {
 		/* We need to ring the txq doorbell -- we have stopped the Tx
 		 * queue for want of resources, but prior calls to gve_tx()
 		 * may have added descriptors without ringing the doorbell.
@@ -672,19 +678,19 @@ static int gve_clean_tx_done(struct gve_priv *priv, struct gve_tx_ring *tx,
 	return pkts;
 }
 
-__be32 gve_tx_load_event_counter(struct gve_priv *priv,
+u32 gve_tx_load_event_counter(struct gve_priv *priv,
 			      struct gve_tx_ring *tx)
 {
-	u32 counter_index = be32_to_cpu((tx->q_resources->counter_index));
+	u32 counter_index = be32_to_cpu(tx->q_resources->counter_index);
+	__be32 counter = READ_ONCE(priv->counter_array[counter_index]);
 
-	return READ_ONCE(priv->counter_array[counter_index]);
+	return be32_to_cpu(counter);
 }
 
 bool gve_tx_poll(struct gve_notify_block *block, int budget)
 {
 	struct gve_priv *priv = block->priv;
 	struct gve_tx_ring *tx = block->tx;
-	bool repoll = false;
 	u32 nic_done;
 	u32 to_do;
 
@@ -692,17 +698,23 @@ bool gve_tx_poll(struct gve_notify_block *block, int budget)
 	if (budget == 0)
 		budget = INT_MAX;
 
-	/* Find out how much work there is to be done */
-	tx->last_nic_done = gve_tx_load_event_counter(priv, tx);
-	nic_done = be32_to_cpu(tx->last_nic_done);
-	if (budget > 0) {
-		/* Do as much work as we have that the budget will
-		 * allow
+	/* In TX path, it may try to clean completed pkts in order to xmit,
+	 * to avoid cleaning conflict, use spin_lock(), it yields better
+	 * concurrency between xmit/clean than netif's lock.
 	 */
+	spin_lock(&tx->clean_lock);
+	/* Find out how much work there is to be done */
+	nic_done = gve_tx_load_event_counter(priv, tx);
 	to_do = min_t(u32, (nic_done - tx->done), budget);
 	gve_clean_tx_done(priv, tx, to_do, true);
-	}
+	spin_unlock(&tx->clean_lock);
 	/* If we still have work we want to repoll */
-	repoll |= (nic_done != tx->done);
-	return repoll;
+	return nic_done != tx->done;
+}
+
+bool gve_tx_clean_pending(struct gve_priv *priv, struct gve_tx_ring *tx)
+{
+	u32 nic_done = gve_tx_load_event_counter(priv, tx);
+
+	return nic_done != tx->done;
 }