octeon_ep: remove atomic variable usage in Tx data path

Replace atomic variable "instr_pending" which represents number of
posted tx instructions pending completion, with host_write_idx and
flush_index variables in the xmit and completion processing respectively.

Signed-off-by: Shinas Rasheed <srasheed@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Author: shinas-marvell

drivers/net/ethernet/marvell/octeon_ep/octep_config.h

@@ -13,6 +13,7 @@
 #define OCTEP_64BYTE_INSTR  64
 
 /* Tx Queue: maximum descriptors per ring */
+/* This needs to be a power of 2 */
 #define OCTEP_IQ_MAX_DESCRIPTORS    1024
 /* Minimum input (Tx) requests to be enqueued to ring doorbell */
 #define OCTEP_DB_MIN                8

drivers/net/ethernet/marvell/octeon_ep/octep_main.c

@@ -777,7 +777,7 @@ static int octep_stop(struct net_device *netdev)
  */
 static inline int octep_iq_full_check(struct octep_iq *iq)
 {
-	if (likely((iq->max_count - atomic_read(&iq->instr_pending)) >=
+	if (likely((IQ_INSTR_SPACE(iq)) >
 		   OCTEP_WAKE_QUEUE_THRESHOLD))
 		return 0;
 
@@ -794,7 +794,7 @@ static inline int octep_iq_full_check(struct octep_iq *iq)
 	/* check again and restart the queue, in case NAPI has just freed
 	 * enough Tx ring entries.
 	 */
-	if (unlikely((iq->max_count - atomic_read(&iq->instr_pending)) >=
+	if (unlikely(IQ_INSTR_SPACE(iq) >
 		     OCTEP_WAKE_QUEUE_THRESHOLD)) {
 		netif_start_subqueue(iq->netdev, iq->q_no);
 		iq->stats.restart_cnt++;
@@ -903,12 +903,9 @@ static netdev_tx_t octep_start_xmit(struct sk_buff *skb,
 	__netdev_tx_sent_queue(iq->netdev_q, skb->len, xmit_more);
 
 	skb_tx_timestamp(skb);
-	atomic_inc(&iq->instr_pending);
 	iq->fill_cnt++;
 	wi++;
-	if (wi == iq->max_count)
-		wi = 0;
-	iq->host_write_index = wi;
+	iq->host_write_index = wi & iq->ring_size_mask;
 
 	/* octep_iq_full_check stops the queue and returns
 	 * true if so, in case the queue has become full

drivers/net/ethernet/marvell/octeon_ep/octep_main.h

@@ -40,6 +40,15 @@
 #define  OCTEP_OQ_INTR_RESEND_BIT  59
 
 #define  OCTEP_MMIO_REGIONS     3
+
+#define  IQ_INSTR_PENDING(iq)  ({ typeof(iq) iq__ = (iq); \
+				  ((iq__)->host_write_index - (iq__)->flush_index) & \
+				  (iq__)->ring_size_mask; \
+				})
+#define  IQ_INSTR_SPACE(iq)    ({ typeof(iq) iq_ = (iq); \
+				  (iq_)->max_count - IQ_INSTR_PENDING(iq_); \
+				})
+
 /* PCI address space mapping information.
  * Each of the 3 address spaces given by BAR0, BAR2 and BAR4 of
  * Octeon gets mapped to different physical address spaces in

drivers/net/ethernet/marvell/octeon_ep/octep_tx.c

@@ -21,7 +21,6 @@ static void octep_iq_reset_indices(struct octep_iq *iq)
 	iq->flush_index = 0;
 	iq->pkts_processed = 0;
 	iq->pkt_in_done = 0;
-	atomic_set(&iq->instr_pending, 0);
 }
 
 /**
@@ -82,7 +81,6 @@ int octep_iq_process_completions(struct octep_iq *iq, u16 budget)
 	}
 
 	iq->pkts_processed += compl_pkts;
-	atomic_sub(compl_pkts, &iq->instr_pending);
 	iq->stats.instr_completed += compl_pkts;
 	iq->stats.bytes_sent += compl_bytes;
 	iq->stats.sgentry_sent += compl_sg;
@@ -91,7 +89,7 @@ int octep_iq_process_completions(struct octep_iq *iq, u16 budget)
 	netdev_tx_completed_queue(iq->netdev_q, compl_pkts, compl_bytes);
 
 	if (unlikely(__netif_subqueue_stopped(iq->netdev, iq->q_no)) &&
-	    ((iq->max_count - atomic_read(&iq->instr_pending)) >
+	    (IQ_INSTR_SPACE(iq) >
 	     OCTEP_WAKE_QUEUE_THRESHOLD))
 		netif_wake_subqueue(iq->netdev, iq->q_no);
 	return !budget;
@@ -144,7 +142,6 @@ static void octep_iq_free_pending(struct octep_iq *iq)
 		dev_kfree_skb_any(skb);
 	}
 
-	atomic_set(&iq->instr_pending, 0);
 	iq->flush_index = fi;
 	netdev_tx_reset_queue(netdev_get_tx_queue(iq->netdev, iq->q_no));
 }

drivers/net/ethernet/marvell/octeon_ep/octep_tx.h

@@ -172,9 +172,6 @@ struct octep_iq {
 	/* Statistics for this input queue. */
 	struct octep_iq_stats stats;
 
-	/* This field keeps track of the instructions pending in this queue. */
-	atomic_t instr_pending;
-
 	/* Pointer to the Virtual Base addr of the input ring. */
 	struct octep_tx_desc_hw *desc_ring;