bnx2x: fix BRB thresholds for dropless_fc mode

Fix the thresholds according to 5778x HW and increase rx_ring size
to suit new thresholds in dropless_fc mode.
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/bnx2x/bnx2x.h

@@ -315,6 +315,14 @@ union db_prod {
 	u32		raw;
 };
 
+/* dropless fc FW/HW related params */
+#define BRB_SIZE(bp)		(CHIP_IS_E3(bp) ? 1024 : 512)
+#define MAX_AGG_QS(bp)		(CHIP_IS_E1(bp) ? \
+					ETH_MAX_AGGREGATION_QUEUES_E1 :\
+					ETH_MAX_AGGREGATION_QUEUES_E1H_E2)
+#define FW_DROP_LEVEL(bp)	(3 + MAX_SPQ_PENDING + MAX_AGG_QS(bp))
+#define FW_PREFETCH_CNT		16
+#define DROPLESS_FC_HEADROOM	100
 
 /* MC hsi */
 #define BCM_PAGE_SHIFT		12
@@ -331,15 +339,35 @@ union db_prod {
 /* SGE ring related macros */
 #define NUM_RX_SGE_PAGES	2
 #define RX_SGE_CNT		(BCM_PAGE_SIZE / sizeof(struct eth_rx_sge))
-#define MAX_RX_SGE_CNT		(RX_SGE_CNT - 2)
+#define NEXT_PAGE_SGE_DESC_CNT	2
+#define MAX_RX_SGE_CNT		(RX_SGE_CNT - NEXT_PAGE_SGE_DESC_CNT)
 /* RX_SGE_CNT is promised to be a power of 2 */
 #define RX_SGE_MASK		(RX_SGE_CNT - 1)
 #define NUM_RX_SGE		(RX_SGE_CNT * NUM_RX_SGE_PAGES)
 #define MAX_RX_SGE		(NUM_RX_SGE - 1)
 #define NEXT_SGE_IDX(x)		((((x) & RX_SGE_MASK) == \
-				  (MAX_RX_SGE_CNT - 1)) ? (x) + 3 : (x) + 1)
+				  (MAX_RX_SGE_CNT - 1)) ? \
+					(x) + 1 + NEXT_PAGE_SGE_DESC_CNT : \
+					(x) + 1)
 #define RX_SGE(x)		((x) & MAX_RX_SGE)
 
+/*
+ * Number of required  SGEs is the sum of two:
+ * 1. Number of possible opened aggregations (next packet for
+ *    these aggregations will probably consume SGE immidiatelly)
+ * 2. Rest of BRB blocks divided by 2 (block will consume new SGE only
+ *    after placement on BD for new TPA aggregation)
+ *
+ * Takes into account NEXT_PAGE_SGE_DESC_CNT "next" elements on each page
+ */
+#define NUM_SGE_REQ		(MAX_AGG_QS(bp) + \
+					(BRB_SIZE(bp) - MAX_AGG_QS(bp)) / 2)
+#define NUM_SGE_PG_REQ		((NUM_SGE_REQ + MAX_RX_SGE_CNT - 1) / \
+						MAX_RX_SGE_CNT)
+#define SGE_TH_LO(bp)		(NUM_SGE_REQ + \
+				 NUM_SGE_PG_REQ * NEXT_PAGE_SGE_DESC_CNT)
+#define SGE_TH_HI(bp)		(SGE_TH_LO(bp) + DROPLESS_FC_HEADROOM)
+
 /* Manipulate a bit vector defined as an array of u64 */
 
 /* Number of bits in one sge_mask array element */
@@ -551,24 +579,43 @@ struct bnx2x_fastpath {
 
 #define NUM_TX_RINGS		16
 #define TX_DESC_CNT		(BCM_PAGE_SIZE / sizeof(union eth_tx_bd_types))
-#define MAX_TX_DESC_CNT		(TX_DESC_CNT - 1)
+#define NEXT_PAGE_TX_DESC_CNT	1
+#define MAX_TX_DESC_CNT		(TX_DESC_CNT - NEXT_PAGE_TX_DESC_CNT)
 #define NUM_TX_BD		(TX_DESC_CNT * NUM_TX_RINGS)
 #define MAX_TX_BD		(NUM_TX_BD - 1)
 #define MAX_TX_AVAIL		(MAX_TX_DESC_CNT * NUM_TX_RINGS - 2)
 #define NEXT_TX_IDX(x)		((((x) & MAX_TX_DESC_CNT) == \
-				  (MAX_TX_DESC_CNT - 1)) ? (x) + 2 : (x) + 1)
+				  (MAX_TX_DESC_CNT - 1)) ? \
+					(x) + 1 + NEXT_PAGE_TX_DESC_CNT : \
+					(x) + 1)
 #define TX_BD(x)		((x) & MAX_TX_BD)
 #define TX_BD_POFF(x)		((x) & MAX_TX_DESC_CNT)
 
 /* The RX BD ring is special, each bd is 8 bytes but the last one is 16 */
 #define NUM_RX_RINGS		8
 #define RX_DESC_CNT		(BCM_PAGE_SIZE / sizeof(struct eth_rx_bd))
-#define MAX_RX_DESC_CNT		(RX_DESC_CNT - 2)
+#define NEXT_PAGE_RX_DESC_CNT	2
+#define MAX_RX_DESC_CNT		(RX_DESC_CNT - NEXT_PAGE_RX_DESC_CNT)
 #define RX_DESC_MASK		(RX_DESC_CNT - 1)
 #define NUM_RX_BD		(RX_DESC_CNT * NUM_RX_RINGS)
 #define MAX_RX_BD		(NUM_RX_BD - 1)
 #define MAX_RX_AVAIL		(MAX_RX_DESC_CNT * NUM_RX_RINGS - 2)
-#define MIN_RX_AVAIL		128
+
+/* dropless fc calculations for BDs
+ *
+ * Number of BDs should as number of buffers in BRB:
+ * Low threshold takes into account NEXT_PAGE_RX_DESC_CNT
+ * "next" elements on each page
+ */
+#define NUM_BD_REQ		BRB_SIZE(bp)
+#define NUM_BD_PG_REQ		((NUM_BD_REQ + MAX_RX_DESC_CNT - 1) / \
+					      MAX_RX_DESC_CNT)
+#define BD_TH_LO(bp)		(NUM_BD_REQ + \
+				 NUM_BD_PG_REQ * NEXT_PAGE_RX_DESC_CNT + \
+				 FW_DROP_LEVEL(bp))
+#define BD_TH_HI(bp)		(BD_TH_LO(bp) + DROPLESS_FC_HEADROOM)
+
+#define MIN_RX_AVAIL		((bp)->dropless_fc ? BD_TH_HI(bp) + 128 : 128)
 
 #define MIN_RX_SIZE_TPA_HW	(CHIP_IS_E1(bp) ? \
 					ETH_MIN_RX_CQES_WITH_TPA_E1 : \
@@ -579,7 +626,9 @@ struct bnx2x_fastpath {
 								MIN_RX_AVAIL))
 
 #define NEXT_RX_IDX(x)		((((x) & RX_DESC_MASK) == \
-				  (MAX_RX_DESC_CNT - 1)) ? (x) + 3 : (x) + 1)
+				  (MAX_RX_DESC_CNT - 1)) ? \
+					(x) + 1 + NEXT_PAGE_RX_DESC_CNT : \
+					(x) + 1)
 #define RX_BD(x)		((x) & MAX_RX_BD)
 
 /*
@@ -589,14 +638,31 @@ struct bnx2x_fastpath {
 #define CQE_BD_REL	(sizeof(union eth_rx_cqe) / sizeof(struct eth_rx_bd))
 #define NUM_RCQ_RINGS		(NUM_RX_RINGS * CQE_BD_REL)
 #define RCQ_DESC_CNT		(BCM_PAGE_SIZE / sizeof(union eth_rx_cqe))
-#define MAX_RCQ_DESC_CNT	(RCQ_DESC_CNT - 1)
+#define NEXT_PAGE_RCQ_DESC_CNT	1
+#define MAX_RCQ_DESC_CNT	(RCQ_DESC_CNT - NEXT_PAGE_RCQ_DESC_CNT)
 #define NUM_RCQ_BD		(RCQ_DESC_CNT * NUM_RCQ_RINGS)
 #define MAX_RCQ_BD		(NUM_RCQ_BD - 1)
 #define MAX_RCQ_AVAIL		(MAX_RCQ_DESC_CNT * NUM_RCQ_RINGS - 2)
 #define NEXT_RCQ_IDX(x)		((((x) & MAX_RCQ_DESC_CNT) == \
-				  (MAX_RCQ_DESC_CNT - 1)) ? (x) + 2 : (x) + 1)
+				  (MAX_RCQ_DESC_CNT - 1)) ? \
+					(x) + 1 + NEXT_PAGE_RCQ_DESC_CNT : \
+					(x) + 1)
 #define RCQ_BD(x)		((x) & MAX_RCQ_BD)
 
+/* dropless fc calculations for RCQs
+ *
+ * Number of RCQs should be as number of buffers in BRB:
+ * Low threshold takes into account NEXT_PAGE_RCQ_DESC_CNT
+ * "next" elements on each page
+ */
+#define NUM_RCQ_REQ		BRB_SIZE(bp)
+#define NUM_RCQ_PG_REQ		((NUM_BD_REQ + MAX_RCQ_DESC_CNT - 1) / \
+					      MAX_RCQ_DESC_CNT)
+#define RCQ_TH_LO(bp)		(NUM_RCQ_REQ + \
+				 NUM_RCQ_PG_REQ * NEXT_PAGE_RCQ_DESC_CNT + \
+				 FW_DROP_LEVEL(bp))
+#define RCQ_TH_HI(bp)		(RCQ_TH_LO(bp) + DROPLESS_FC_HEADROOM)
+
 
 /* This is needed for determining of last_max */
 #define SUB_S16(a, b)		(s16)((s16)(a) - (s16)(b))

drivers/net/bnx2x/bnx2x_cmn.c

@@ -987,8 +987,6 @@ void __bnx2x_link_report(struct bnx2x *bp)
 void bnx2x_init_rx_rings(struct bnx2x *bp)
 {
 	int func = BP_FUNC(bp);
-	int max_agg_queues = CHIP_IS_E1(bp) ? ETH_MAX_AGGREGATION_QUEUES_E1 :
-					      ETH_MAX_AGGREGATION_QUEUES_E1H_E2;
 	u16 ring_prod;
 	int i, j;
 
@@ -1001,7 +999,7 @@ void bnx2x_init_rx_rings(struct bnx2x *bp)
 
 		if (!fp->disable_tpa) {
 			/* Fill the per-aggregtion pool */
-			for (i = 0; i < max_agg_queues; i++) {
+			for (i = 0; i < MAX_AGG_QS(bp); i++) {
 				struct bnx2x_agg_info *tpa_info =
 					&fp->tpa_info[i];
 				struct sw_rx_bd *first_buf =
@@ -1041,7 +1039,7 @@ void bnx2x_init_rx_rings(struct bnx2x *bp)
 					bnx2x_free_rx_sge_range(bp, fp,
 								ring_prod);
 					bnx2x_free_tpa_pool(bp, fp,
-							    max_agg_queues);
+							    MAX_AGG_QS(bp));
 					fp->disable_tpa = 1;
 					ring_prod = 0;
 					break;
@@ -1137,9 +1135,7 @@ static void bnx2x_free_rx_skbs(struct bnx2x *bp)
 		bnx2x_free_rx_bds(fp);
 
 		if (!fp->disable_tpa)
-			bnx2x_free_tpa_pool(bp, fp, CHIP_IS_E1(bp) ?
-					    ETH_MAX_AGGREGATION_QUEUES_E1 :
-					    ETH_MAX_AGGREGATION_QUEUES_E1H_E2);
+			bnx2x_free_tpa_pool(bp, fp, MAX_AGG_QS(bp));
 	}
 }
 

drivers/net/bnx2x/bnx2x_main.c

@@ -2756,8 +2756,14 @@ static void bnx2x_pf_rx_q_prep(struct bnx2x *bp,
 	u16 tpa_agg_size = 0;
 
 	if (!fp->disable_tpa) {
-		pause->sge_th_hi = 250;
-		pause->sge_th_lo = 150;
+		pause->sge_th_lo = SGE_TH_LO(bp);
+		pause->sge_th_hi = SGE_TH_HI(bp);
+
+		/* validate SGE ring has enough to cross high threshold */
+		WARN_ON(bp->dropless_fc &&
+				pause->sge_th_hi + FW_PREFETCH_CNT >
+				MAX_RX_SGE_CNT * NUM_RX_SGE_PAGES);
+
 		tpa_agg_size = min_t(u32,
 			(min_t(u32, 8, MAX_SKB_FRAGS) *
 			SGE_PAGE_SIZE * PAGES_PER_SGE), 0xffff);
@@ -2771,10 +2777,21 @@ static void bnx2x_pf_rx_q_prep(struct bnx2x *bp,
 
 	/* pause - not for e1 */
 	if (!CHIP_IS_E1(bp)) {
-		pause->bd_th_hi = 350;
-		pause->bd_th_lo = 250;
-		pause->rcq_th_hi = 350;
-		pause->rcq_th_lo = 250;
+		pause->bd_th_lo = BD_TH_LO(bp);
+		pause->bd_th_hi = BD_TH_HI(bp);
+
+		pause->rcq_th_lo = RCQ_TH_LO(bp);
+		pause->rcq_th_hi = RCQ_TH_HI(bp);
+		/*
+		 * validate that rings have enough entries to cross
+		 * high thresholds
+		 */
+		WARN_ON(bp->dropless_fc &&
+				pause->bd_th_hi + FW_PREFETCH_CNT >
+				bp->rx_ring_size);
+		WARN_ON(bp->dropless_fc &&
+				pause->rcq_th_hi + FW_PREFETCH_CNT >
+				NUM_RCQ_RINGS * MAX_RCQ_DESC_CNT);
 
 		pause->pri_map = 1;
 	}
@@ -2802,9 +2819,7 @@ static void bnx2x_pf_rx_q_prep(struct bnx2x *bp,
 	 * For PF Clients it should be the maximum avaliable number.
 	 * VF driver(s) may want to define it to a smaller value.
 	 */
-	rxq_init->max_tpa_queues =
-		(CHIP_IS_E1(bp) ? ETH_MAX_AGGREGATION_QUEUES_E1 :
-		ETH_MAX_AGGREGATION_QUEUES_E1H_E2);
+	rxq_init->max_tpa_queues = MAX_AGG_QS(bp);
 
 	rxq_init->cache_line_log = BNX2X_RX_ALIGN_SHIFT;
 	rxq_init->fw_sb_id = fp->fw_sb_id;