From e1ddb78f4185559bc6f9be1d1b302f9f52899f94 Mon Sep 17 00:00:00 2001
From: Peter Chiu <chui-hao.chiu@mediatek.com>
Date: Tue, 13 Jun 2023 09:04:43 +0800
Subject: [PATCH 10/98] wifi: mt76: mt7996: adjust wfdma setting to enhance
 throughput

1. Set band 1 traffic to pcie1.
2. Refactor dma prefetch and enlarge txd prefetch size.
3. Update pdma setting.

Signed-off-by: Peter Chiu <chui-hao.chiu@mediatek.com>
---
 mt7996/dma.c  | 70 +++++++++++++++++++++++++++++++++++----------------
 mt7996/regs.h |  9 +++++++
 2 files changed, 58 insertions(+), 21 deletions(-)

diff --git a/mt7996/dma.c b/mt7996/dma.c
index 3d04470..1ed91da 100644
--- a/mt7996/dma.c
+++ b/mt7996/dma.c
@@ -99,38 +99,49 @@ static void mt7996_dma_config(struct mt7996_dev *dev)
 	MCUQ_CONFIG(MT_MCUQ_FWDL, WFDMA0, MT_INT_TX_DONE_FWDL, MT7996_TXQ_FWDL);
 }
 
+static u32 __mt7996_dma_prefetch_base(u16 *base, u8 depth)
+{
+	u32 ret = *base << 16 | depth;
+
+	*base = *base + (depth << 4);
+
+	return ret;
+}
+
 static void __mt7996_dma_prefetch(struct mt7996_dev *dev, u32 ofs)
 {
-#define PREFETCH(_base, _depth)	((_base) << 16 | (_depth))
+	u16 base = 0;
+
+#define PREFETCH(_depth)	(__mt7996_dma_prefetch_base(&base, (_depth)))
 	/* prefetch SRAM wrapping boundary for tx/rx ring. */
-	mt76_wr(dev, MT_MCUQ_EXT_CTRL(MT_MCUQ_FWDL) + ofs, PREFETCH(0x0, 0x2));
-	mt76_wr(dev, MT_MCUQ_EXT_CTRL(MT_MCUQ_WM) + ofs, PREFETCH(0x20, 0x2));
-	mt76_wr(dev, MT_TXQ_EXT_CTRL(0) + ofs, PREFETCH(0x40, 0x4));
-	mt76_wr(dev, MT_TXQ_EXT_CTRL(1) + ofs, PREFETCH(0x80, 0x4));
-	mt76_wr(dev, MT_MCUQ_EXT_CTRL(MT_MCUQ_WA) + ofs, PREFETCH(0xc0, 0x2));
-	mt76_wr(dev, MT_TXQ_EXT_CTRL(2) + ofs, PREFETCH(0xe0, 0x4));
-	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MCU) + ofs, PREFETCH(0x120, 0x2));
-	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MCU_WA) + ofs, PREFETCH(0x140, 0x2));
-	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MAIN_WA) + ofs, PREFETCH(0x160, 0x2));
-	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_BAND2_WA) + ofs, PREFETCH(0x180, 0x2));
-	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MAIN) + ofs, PREFETCH(0x1a0, 0x10));
-	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_BAND2) + ofs, PREFETCH(0x2a0, 0x10));
+	mt76_wr(dev, MT_MCUQ_EXT_CTRL(MT_MCUQ_FWDL) + ofs, PREFETCH(0x2));
+	mt76_wr(dev, MT_MCUQ_EXT_CTRL(MT_MCUQ_WM) + ofs, PREFETCH(0x2));
+	mt76_wr(dev, MT_TXQ_EXT_CTRL(0) + ofs, PREFETCH(0x8));
+	mt76_wr(dev, MT_TXQ_EXT_CTRL(1) + ofs, PREFETCH(0x8));
+	mt76_wr(dev, MT_MCUQ_EXT_CTRL(MT_MCUQ_WA) + ofs, PREFETCH(0x2));
+	mt76_wr(dev, MT_TXQ_EXT_CTRL(2) + ofs, PREFETCH(0x8));
+	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MCU) + ofs, PREFETCH(0x2));
+	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MCU_WA) + ofs, PREFETCH(0x2));
+	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MAIN_WA) + ofs, PREFETCH(0x2));
+	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_BAND2_WA) + ofs, PREFETCH(0x2));
+	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MAIN) + ofs, PREFETCH(0x10));
+	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_BAND2) + ofs, PREFETCH(0x10));
 
 	if (dev->has_rro) {
 		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_RRO_BAND0) + ofs,
-			PREFETCH(0x3a0, 0x10));
+			PREFETCH(0x10));
 		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_RRO_BAND2) + ofs,
-			PREFETCH(0x4a0, 0x10));
+			PREFETCH(0x10));
 		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MSDU_PAGE_BAND0) + ofs,
-			PREFETCH(0x5a0, 0x4));
+			PREFETCH(0x4));
 		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MSDU_PAGE_BAND1) + ofs,
-			PREFETCH(0x5e0, 0x4));
+			PREFETCH(0x4));
 		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MSDU_PAGE_BAND2) + ofs,
-			PREFETCH(0x620, 0x4));
+			PREFETCH(0x4));
 		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_TXFREE_BAND0) + ofs,
-			PREFETCH(0x660, 0x4));
+			PREFETCH(0x4));
 		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_TXFREE_BAND2) + ofs,
-			PREFETCH(0x6a0, 0x4));
+			PREFETCH(0x4));
 	}
 #undef PREFETCH
 
@@ -295,6 +306,12 @@ static void mt7996_dma_enable(struct mt7996_dev *dev, bool reset)
 	mt76_set(dev, WF_WFDMA0_GLO_CFG_EXT1,
 		 WF_WFDMA0_GLO_CFG_EXT1_TX_FCTRL_MODE);
 
+	/* WFDMA rx threshold */
+	mt76_wr(dev, MT_WFDMA0_PAUSE_RX_Q_45_TH, 0xc000c);
+	mt76_wr(dev, MT_WFDMA0_PAUSE_RX_Q_67_TH, 0x10008);
+	mt76_wr(dev, MT_WFDMA0_PAUSE_RX_Q_89_TH, 0x10008);
+	mt76_wr(dev, MT_WFDMA0_PAUSE_RX_Q_RRO_TH, 0x20);
+
 	if (dev->hif2) {
 		/* GLO_CFG_EXT0 */
 		mt76_set(dev, WF_WFDMA0_GLO_CFG_EXT0 + hif1_ofs,
@@ -306,7 +323,18 @@ static void mt7996_dma_enable(struct mt7996_dev *dev, bool reset)
 			 WF_WFDMA0_GLO_CFG_EXT1_TX_FCTRL_MODE);
 
 		mt76_set(dev, MT_WFDMA_HOST_CONFIG,
-			 MT_WFDMA_HOST_CONFIG_PDMA_BAND);
+			 MT_WFDMA_HOST_CONFIG_PDMA_BAND |
+			 MT_WFDMA_HOST_CONFIG_BAND2_PCIE1);
+
+		/* AXI read outstanding number */
+		mt76_rmw(dev, MT_WFDMA_AXI_R2A_CTRL,
+			 MT_WFDMA_AXI_R2A_CTRL_OUTSTAND_MASK, 0x14);
+
+		/* WFDMA rx threshold */
+		mt76_wr(dev, MT_WFDMA0_PAUSE_RX_Q_45_TH + hif1_ofs, 0xc000c);
+		mt76_wr(dev, MT_WFDMA0_PAUSE_RX_Q_67_TH + hif1_ofs, 0x10008);
+		mt76_wr(dev, MT_WFDMA0_PAUSE_RX_Q_89_TH + hif1_ofs, 0x10008);
+		mt76_wr(dev, MT_WFDMA0_PAUSE_RX_Q_RRO_TH + hif1_ofs, 0x20);
 	}
 
 	if (dev->hif2) {
diff --git a/mt7996/regs.h b/mt7996/regs.h
index a4d5ad8..f7c99cd 100644
--- a/mt7996/regs.h
+++ b/mt7996/regs.h
@@ -366,6 +366,11 @@ enum base_rev {
 #define MT_WFDMA0_GLO_CFG_OMIT_RX_INFO		BIT(27)
 #define MT_WFDMA0_GLO_CFG_OMIT_RX_INFO_PFET2	BIT(21)
 
+#define MT_WFDMA0_PAUSE_RX_Q_45_TH		MT_WFDMA0(0x268)
+#define MT_WFDMA0_PAUSE_RX_Q_67_TH		MT_WFDMA0(0x26c)
+#define MT_WFDMA0_PAUSE_RX_Q_89_TH		MT_WFDMA0(0x270)
+#define MT_WFDMA0_PAUSE_RX_Q_RRO_TH		MT_WFDMA0(0x27c)
+
 #define WF_WFDMA0_GLO_CFG_EXT0			MT_WFDMA0(0x2b0)
 #define WF_WFDMA0_GLO_CFG_EXT0_RX_WB_RXD	BIT(18)
 #define WF_WFDMA0_GLO_CFG_EXT0_WED_MERGE_MODE	BIT(14)
@@ -388,10 +393,14 @@ enum base_rev {
 
 #define MT_WFDMA_HOST_CONFIG			MT_WFDMA_EXT_CSR(0x30)
 #define MT_WFDMA_HOST_CONFIG_PDMA_BAND		BIT(0)
+#define MT_WFDMA_HOST_CONFIG_BAND2_PCIE1	BIT(22)
 
 #define MT_WFDMA_EXT_CSR_HIF_MISC		MT_WFDMA_EXT_CSR(0x44)
 #define MT_WFDMA_EXT_CSR_HIF_MISC_BUSY		BIT(0)
 
+#define MT_WFDMA_AXI_R2A_CTRL			MT_WFDMA_EXT_CSR(0x500)
+#define MT_WFDMA_AXI_R2A_CTRL_OUTSTAND_MASK	GENMASK(4, 0)
+
 #define MT_PCIE_RECOG_ID			0xd7090
 #define MT_PCIE_RECOG_ID_MASK			GENMASK(30, 0)
 #define MT_PCIE_RECOG_ID_SEM			BIT(31)
-- 
2.18.0