bnx2x: replace mechanism to check for next available packet

Check next packet availability by validating that HW has finished CQE
placement. This saves latency of another dma transaction performed to update
SB indexes.

Signed-off-by: Dmitry Kravkov <dmitry@broadcom.com>
Signed-off-by: Eilon Greenstein <eilong@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c

@@ -804,40 +804,32 @@ int bnx2x_rx_int(struct bnx2x_fastpath *fp, int budget)
 {
 	struct bnx2x *bp = fp->bp;
 	u16 bd_cons, bd_prod, bd_prod_fw, comp_ring_cons;
-	u16 hw_comp_cons, sw_comp_cons, sw_comp_prod;
+	u16 sw_comp_cons, sw_comp_prod;
 	int rx_pkt = 0;
+	union eth_rx_cqe *cqe;
+	struct eth_fast_path_rx_cqe *cqe_fp;
 
 #ifdef BNX2X_STOP_ON_ERROR
 	if (unlikely(bp->panic))
 		return 0;
 #endif
 
-	/* CQ "next element" is of the size of the regular element,
-	   that's why it's ok here */
-	hw_comp_cons = le16_to_cpu(*fp->rx_cons_sb);
-	if ((hw_comp_cons & MAX_RCQ_DESC_CNT) == MAX_RCQ_DESC_CNT)
-		hw_comp_cons++;
-
 	bd_cons = fp->rx_bd_cons;
 	bd_prod = fp->rx_bd_prod;
 	bd_prod_fw = bd_prod;
 	sw_comp_cons = fp->rx_comp_cons;
 	sw_comp_prod = fp->rx_comp_prod;
 
-	/* Memory barrier necessary as speculative reads of the rx
-	 * buffer can be ahead of the index in the status block
-	 */
-	rmb();
+	comp_ring_cons = RCQ_BD(sw_comp_cons);
+	cqe = &fp->rx_comp_ring[comp_ring_cons];
+	cqe_fp = &cqe->fast_path_cqe;
 
 	DP(NETIF_MSG_RX_STATUS,
-	   "queue[%d]:  hw_comp_cons %u  sw_comp_cons %u\n",
-	   fp->index, hw_comp_cons, sw_comp_cons);
+	   "queue[%d]: sw_comp_cons %u\n", fp->index, sw_comp_cons);
 
-	while (sw_comp_cons != hw_comp_cons) {
+	while (BNX2X_IS_CQE_COMPLETED(cqe_fp)) {
 		struct sw_rx_bd *rx_buf = NULL;
 		struct sk_buff *skb;
-		union eth_rx_cqe *cqe;
-		struct eth_fast_path_rx_cqe *cqe_fp;
 		u8 cqe_fp_flags;
 		enum eth_rx_cqe_type cqe_fp_type;
 		u16 len, pad, queue;
@@ -849,12 +841,9 @@ int bnx2x_rx_int(struct bnx2x_fastpath *fp, int budget)
 			return 0;
 #endif
 
-		comp_ring_cons = RCQ_BD(sw_comp_cons);
 		bd_prod = RX_BD(bd_prod);
 		bd_cons = RX_BD(bd_cons);
 
-		cqe = &fp->rx_comp_ring[comp_ring_cons];
-		cqe_fp = &cqe->fast_path_cqe;
 		cqe_fp_flags = cqe_fp->type_error_flags;
 		cqe_fp_type = cqe_fp_flags & ETH_FAST_PATH_RX_CQE_TYPE;
 
@@ -1018,8 +1007,15 @@ next_cqe:
 		sw_comp_prod = NEXT_RCQ_IDX(sw_comp_prod);
 		sw_comp_cons = NEXT_RCQ_IDX(sw_comp_cons);
 
+		/* mark CQE as free */
+		BNX2X_SEED_CQE(cqe_fp);
+
 		if (rx_pkt == budget)
 			break;
+
+		comp_ring_cons = RCQ_BD(sw_comp_cons);
+		cqe = &fp->rx_comp_ring[comp_ring_cons];
+		cqe_fp = &cqe->fast_path_cqe;
 	} /* while */
 
 	fp->rx_bd_cons = bd_cons;
@@ -1055,8 +1051,6 @@ static irqreturn_t bnx2x_msix_fp_int(int irq, void *fp_cookie)
 #endif
 
 	/* Handle Rx and Tx according to MSI-X vector */
-	prefetch(fp->rx_cons_sb);
-
 	for_each_cos_in_tx_queue(fp, cos)
 		prefetch(fp->txdata_ptr[cos]->tx_cons_sb);
 
@@ -3137,10 +3131,8 @@ int bnx2x_low_latency_recv(struct napi_struct *napi)
 	if (!bnx2x_fp_lock_poll(fp))
 		return LL_FLUSH_BUSY;
 
-	if (bnx2x_has_rx_work(fp)) {
-		bnx2x_update_fpsb_idx(fp);
+	if (bnx2x_has_rx_work(fp))
 		found = bnx2x_rx_int(fp, 4);
-	}
 
 	bnx2x_fp_unlock_poll(fp);
 
@@ -4339,10 +4331,11 @@ static int bnx2x_alloc_fp_mem_at(struct bnx2x *bp, int index)
 				&bnx2x_fp(bp, index, rx_desc_mapping),
 				sizeof(struct eth_rx_bd) * NUM_RX_BD);
 
-		BNX2X_PCI_ALLOC(bnx2x_fp(bp, index, rx_comp_ring),
-				&bnx2x_fp(bp, index, rx_comp_mapping),
-				sizeof(struct eth_fast_path_rx_cqe) *
-				NUM_RCQ_BD);
+		/* Seed all CQEs by 1s */
+		BNX2X_PCI_FALLOC(bnx2x_fp(bp, index, rx_comp_ring),
+				 &bnx2x_fp(bp, index, rx_comp_mapping),
+				 sizeof(struct eth_fast_path_rx_cqe) *
+				 NUM_RCQ_BD);
 
 		/* SGE ring */
 		BNX2X_ALLOC(bnx2x_fp(bp, index, rx_page_ring),

drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.h

@@ -59,6 +59,16 @@ extern int int_mode;
 		   (unsigned long long)(*y), x); \
 	} while (0)
 
+#define BNX2X_PCI_FALLOC(x, y, size) \
+	do { \
+		x = dma_alloc_coherent(&bp->pdev->dev, size, y, GFP_KERNEL); \
+		if (x == NULL) \
+			goto alloc_mem_err; \
+		memset((void *)x, 0xFFFFFFFF, size); \
+		DP(NETIF_MSG_HW, "BNX2X_PCI_FALLOC: Physical %Lx Virtual %p\n",\
+		   (unsigned long long)(*y), x); \
+	} while (0)
+
 #define BNX2X_ALLOC(x, size) \
 	do { \
 		x = kzalloc(size, GFP_KERNEL); \
@@ -805,16 +815,18 @@ static inline bool bnx2x_has_tx_work(struct bnx2x_fastpath *fp)
 	return false;
 }
 
+#define BNX2X_IS_CQE_COMPLETED(cqe_fp) (cqe_fp->marker == 0x0)
+#define BNX2X_SEED_CQE(cqe_fp) (cqe_fp->marker = 0xFFFFFFFF)
 static inline int bnx2x_has_rx_work(struct bnx2x_fastpath *fp)
 {
-	u16 rx_cons_sb;
+	u16 cons;
+	union eth_rx_cqe *cqe;
+	struct eth_fast_path_rx_cqe *cqe_fp;
 
-	/* Tell compiler that status block fields can change */
-	barrier();
-	rx_cons_sb = le16_to_cpu(*fp->rx_cons_sb);
-	if ((rx_cons_sb & MAX_RCQ_DESC_CNT) == MAX_RCQ_DESC_CNT)
-		rx_cons_sb++;
-	return (fp->rx_comp_cons != rx_cons_sb);
+	cons = RCQ_BD(fp->rx_comp_cons);
+	cqe = &fp->rx_comp_ring[cons];
+	cqe_fp = &cqe->fast_path_cqe;
+	return BNX2X_IS_CQE_COMPLETED(cqe_fp);
 }
 
 /**

drivers/net/ethernet/broadcom/bnx2x/bnx2x_hsi.h

@@ -3818,7 +3818,8 @@ struct eth_fast_path_rx_cqe {
 	__le16 len_on_bd;
 	struct parsing_flags pars_flags;
 	union eth_sgl_or_raw_data sgl_or_raw_data;
-	__le32 reserved1[8];
+	__le32 reserved1[7];
+	u32 marker;
 };
 
 

drivers/net/ethernet/broadcom/bnx2x/bnx2x_main.c

@@ -1866,7 +1866,6 @@ irqreturn_t bnx2x_interrupt(int irq, void *dev_instance)
 		mask = 0x2 << (fp->index + CNIC_SUPPORT(bp));
 		if (status & mask) {
 			/* Handle Rx or Tx according to SB id */
-			prefetch(fp->rx_cons_sb);
 			for_each_cos_in_tx_queue(fp, cos)
 				prefetch(fp->txdata_ptr[cos]->tx_cons_sb);
 			prefetch(&fp->sb_running_index[SM_RX_ID]);