[PATCH] qtnfmac: pcie datapath optimizations and cleanups

[PATCH] qtnfmac: pcie datapath optimizations and cleanups

[Sergey Matyukevich]

Hello Kalle and all,

This patchset implements several optimizations and cleanups for
pcie datapath in qtnfmac driver. Major changes include:
- switch to kernel circ_buf implementation
- modify tx reclaim locking
- introduce counter for rx underflow events

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>

 bus.h                   |    1 
 pearl/pcie.c            |  285 +++++++++++++++++++++++++++++-------------------
 pearl/pcie_bus_priv.h   |   15 +-
 pearl/pcie_ipc.h        |    1 
 pearl/pcie_regs_pearl.h |    1 
 5 files changed, 183 insertions(+), 120 deletions(-)

[PATCH 1/9] qtnfmac: remove unused qtnf_rx_frame declaration

[Sergey Matyukevich]

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
---
 drivers/net/wireless/quantenna/qtnfmac/bus.h | 1 -
 1 file changed, 1 deletion(-)

diff --git a/drivers/net/wireless/quantenna/qtnfmac/bus.h b/drivers/net/wireless/quantenna/qtnfmac/bus.h
index dda05003d522..56e5fed92a2a 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/bus.h
+++ b/drivers/net/wireless/quantenna/qtnfmac/bus.h
@@ -130,7 +130,6 @@ static __always_inline void qtnf_bus_unlock(struct qtnf_bus *bus)
 
 /* interface functions from common layer */
 
-void qtnf_rx_frame(struct device *dev, struct sk_buff *rxp);
 int qtnf_core_attach(struct qtnf_bus *bus);
 void qtnf_core_detach(struct qtnf_bus *bus);
 void qtnf_txflowblock(struct device *dev, bool state);
-- 
2.11.0

Re: [1/9] qtnfmac: remove unused qtnf_rx_frame declaration

[Kalle Valo]

Sergey Matyukevich <sergey.matyukevich.os@quantenna.com> wrote:

> Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>

8 patches applied to wireless-drivers-next.git, thanks.

0db63e37992c qtnfmac: remove unused qtnf_rx_frame declaration
7376947dfb80 qtnfmac: switch to napi_gro_receive
c58730cab8ea qtnfmac: use __netdev_alloc_skb_ip_align
867ba964fa69 qtnfmac: skb2rbd_attach cleanup
dfb13db68f3e qtnfmac: decrease default Tx queue size
3cbc3a0f19ac qtnfmac: switch to kernel circ_buf implementation
cc75f9e5bc66 qtnfmac: introduce counter for Rx underflow events
0593da274d4d qtnfmac: modify tx reclaim locking

-- 
https://patchwork.kernel.org/patch/9901845/

https://wireless.wiki.kernel.org/en/developers/documentation/submittingpatches

[PATCH 2/9] qtnfmac: switch to napi_gro_receive

[Sergey Matyukevich]

Use napi_gro_receive() rather than netif_receive_skb() in qtnfmac driver.

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
---
 drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
index ae8acc1bf291..08b35dc30bc8 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
@@ -762,7 +762,7 @@ static int qtnf_rx_poll(struct napi_struct *napi, int budget)
 				ndev->stats.rx_bytes += skb->len;
 
 				skb->protocol = eth_type_trans(skb, ndev);
-				netif_receive_skb(skb);
+				napi_gro_receive(napi, skb);
 			} else {
 				pr_debug("drop untagged skb\n");
 				bus->mux_dev.stats.rx_dropped++;
-- 
2.11.0

Re: [PATCH 2/9] qtnfmac: switch to napi_gro_receive

[Kalle Valo]

Sergey Matyukevich <sergey.matyukevich.os@quantenna.com> writes:

> Use napi_gro_receive() rather than netif_receive_skb() in qtnfmac driver.

Yes, I can easily see that from the diff so no need to document that :)
But the commit log should tell _why_ you did it.

-- 
Kalle Valo

Re: [PATCH 2/9] qtnfmac: switch to napi_gro_receive

[Kalle Valo]

Kalle Valo <kvalo@codeaurora.org> writes:

> Sergey Matyukevich <sergey.matyukevich.os@quantenna.com> writes:
>
>> Use napi_gro_receive() rather than netif_receive_skb() in qtnfmac driver.
>
> Yes, I can easily see that from the diff so no need to document that :)
> But the commit log should tell _why_ you did it.

And no need to resend because of this, I can improve the commit log
during commit. Just let me know what I should add.

-- 
Kalle Valo

Re: [PATCH 2/9] qtnfmac: switch to napi_gro_receive

[Sergey Matyukevich]

> >> Use napi_gro_receive() rather than netif_receive_skb() in qtnfmac driver.
> >
> > Yes, I can easily see that from the diff so no need to document that :)
> > But the commit log should tell _why_ you did it.
> 
> And no need to resend because of this, I can improve the commit log
> during commit. Just let me know what I should add.

Sure. The motivation is to improve performance when GRO is enabled, e.g. by
reducing the number of TCP ACKs. Updated message would be:

Use napi_gro_receive() rather than netif_receive_skb() to improve
performance when GRO is enabled.

Regards,
Sergey

[PATCH 3/9] qtnfmac: use __netdev_alloc_skb_ip_align

[Sergey Matyukevich]

Replace __dev_alloc_skb and explicit NET_IP_ALIGN alignment by built-in
__netdev_alloc_skb_ip_align function.

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
---
 drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
index 08b35dc30bc8..079aa1693ff5 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
@@ -429,8 +429,7 @@ static int skb2rbd_attach(struct qtnf_pcie_bus_priv *priv, u16 rx_bd_index)
 	struct sk_buff *skb;
 	dma_addr_t paddr;
 
-	skb = __dev_alloc_skb(SKB_BUF_SIZE + NET_IP_ALIGN,
-			      GFP_ATOMIC);
+	skb = __netdev_alloc_skb_ip_align(NULL, SKB_BUF_SIZE, GFP_ATOMIC);
 	if (!skb) {
 		priv->rx_skb[rx_bd_index] = NULL;
 		return -ENOMEM;
@@ -438,8 +437,6 @@ static int skb2rbd_attach(struct qtnf_pcie_bus_priv *priv, u16 rx_bd_index)
 
 	priv->rx_skb[rx_bd_index] = skb;
 
-	skb_reserve(skb, NET_IP_ALIGN);
-
 	rxbd = &priv->rx_bd_vbase[rx_bd_index];
 
 	paddr = pci_map_single(priv->pdev, skb->data,
-- 
2.11.0

[PATCH 4/9] qtnfmac: skb2rbd_attach cleanup

[Sergey Matyukevich]

Update PCIE_HDP_TX_HOST_Q_WR_PTR register in skb2rbd_attach as a part of
procedure of passing new Rx buffer to hardware. Sync up all the the
qtnf_rx_bd descriptor updates before passing Rx buffer to hardware.

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
---
 .../net/wireless/quantenna/qtnfmac/pearl/pcie.c    | 31 +++++++++-------------
 1 file changed, 13 insertions(+), 18 deletions(-)

diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
index 079aa1693ff5..a0b65d487ddb 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
@@ -411,11 +411,6 @@ static int alloc_bd_table(struct qtnf_pcie_bus_priv *priv)
 	writel(priv->rx_bd_num | (sizeof(struct qtnf_rx_bd)) << 16,
 	       PCIE_HDP_TX_HOST_Q_SZ_CTRL(priv->pcie_reg_base));
 
-	priv->hw_txproc_wr_ptr = priv->rx_bd_num - rx_bd_reserved_param;
-
-	writel(priv->hw_txproc_wr_ptr,
-	       PCIE_HDP_TX_HOST_Q_WR_PTR(priv->pcie_reg_base));
-
 	pr_debug("RX descriptor table: vaddr=0x%p paddr=%pad\n", vaddr, &paddr);
 
 	priv->rx_bd_index = 0;
@@ -423,7 +418,7 @@ static int alloc_bd_table(struct qtnf_pcie_bus_priv *priv)
 	return 0;
 }
 
-static int skb2rbd_attach(struct qtnf_pcie_bus_priv *priv, u16 rx_bd_index)
+static int skb2rbd_attach(struct qtnf_pcie_bus_priv *priv, u16 index)
 {
 	struct qtnf_rx_bd *rxbd;
 	struct sk_buff *skb;
@@ -431,13 +426,12 @@ static int skb2rbd_attach(struct qtnf_pcie_bus_priv *priv, u16 rx_bd_index)
 
 	skb = __netdev_alloc_skb_ip_align(NULL, SKB_BUF_SIZE, GFP_ATOMIC);
 	if (!skb) {
-		priv->rx_skb[rx_bd_index] = NULL;
+		priv->rx_skb[index] = NULL;
 		return -ENOMEM;
 	}
 
-	priv->rx_skb[rx_bd_index] = skb;
-
-	rxbd = &priv->rx_bd_vbase[rx_bd_index];
+	priv->rx_skb[index] = skb;
+	rxbd = &priv->rx_bd_vbase[index];
 
 	paddr = pci_map_single(priv->pdev, skb->data,
 			       SKB_BUF_SIZE, PCI_DMA_FROMDEVICE);
@@ -446,17 +440,20 @@ static int skb2rbd_attach(struct qtnf_pcie_bus_priv *priv, u16 rx_bd_index)
 		return -ENOMEM;
 	}
 
-	writel(QTN_HOST_LO32(paddr),
-	       PCIE_HDP_HHBM_BUF_PTR(priv->pcie_reg_base));
-	writel(QTN_HOST_HI32(paddr),
-	       PCIE_HDP_HHBM_BUF_PTR_H(priv->pcie_reg_base));
-
 	/* keep rx skb paddrs in rx buffer descriptors for cleanup purposes */
 	rxbd->addr = cpu_to_le32(QTN_HOST_LO32(paddr));
 	rxbd->addr_h = cpu_to_le32(QTN_HOST_HI32(paddr));
-
 	rxbd->info = 0x0;
 
+	/* sync up all descriptor updates */
+	wmb();
+
+	writel(QTN_HOST_HI32(paddr),
+	       PCIE_HDP_HHBM_BUF_PTR_H(priv->pcie_reg_base));
+	writel(QTN_HOST_LO32(paddr),
+	       PCIE_HDP_HHBM_BUF_PTR(priv->pcie_reg_base));
+
+	writel(index, PCIE_HDP_TX_HOST_Q_WR_PTR(priv->pcie_reg_base));
 	return 0;
 }
 
@@ -787,8 +784,6 @@ static int qtnf_rx_poll(struct napi_struct *napi, int budget)
 			break;
 		}
 
-		writel(priv->hw_txproc_wr_ptr,
-		       PCIE_HDP_TX_HOST_Q_WR_PTR(priv->pcie_reg_base));
 	}
 
 	if (processed < budget) {
-- 
2.11.0

[PATCH 5/9] qtnfmac: decrease default Tx queue size

[Sergey Matyukevich]

Avoid extra buffering in driver by default. Use max hardware Tx queue size.

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
---
 drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
index a0b65d487ddb..f18e8a724c68 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
@@ -36,7 +36,7 @@ static bool use_msi = true;
 module_param(use_msi, bool, 0644);
 MODULE_PARM_DESC(use_msi, "set 0 to use legacy interrupt");
 
-static unsigned int tx_bd_size_param = 256;
+static unsigned int tx_bd_size_param = 32;
 module_param(tx_bd_size_param, uint, 0644);
 MODULE_PARM_DESC(tx_bd_size_param, "Tx descriptors queue size");
 
-- 
2.11.0

[PATCH 6/9] qtnfmac: switch to kernel circ_buf implementation

[Sergey Matyukevich]

Current code for both Rx and Tx queue management is a custom and incomplete
circular buffer implementation. It makes a lot of sense to switch to kernel
built-in circ_buf implementation.

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
---
 .../net/wireless/quantenna/qtnfmac/pearl/pcie.c    | 206 +++++++++++++--------
 .../quantenna/qtnfmac/pearl/pcie_bus_priv.h        |  10 +-
 2 files changed, 136 insertions(+), 80 deletions(-)

diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
index f18e8a724c68..f8207ab25576 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
@@ -25,6 +25,7 @@
 #include <linux/completion.h>
 #include <linux/crc32.h>
 #include <linux/spinlock.h>
+#include <linux/circ_buf.h>
 
 #include "qtn_hw_ids.h"
 #include "pcie_bus_priv.h"
@@ -44,10 +45,6 @@ static unsigned int rx_bd_size_param = 256;
 module_param(rx_bd_size_param, uint, 0644);
 MODULE_PARM_DESC(rx_bd_size_param, "Rx descriptors queue size");
 
-static unsigned int rx_bd_reserved_param = 16;
-module_param(rx_bd_reserved_param, uint, 0644);
-MODULE_PARM_DESC(rx_bd_reserved_param, "Reserved RX descriptors");
-
 static u8 flashboot = 1;
 module_param(flashboot, byte, 0644);
 MODULE_PARM_DESC(flashboot, "set to 0 to use FW binary file on FS");
@@ -392,9 +389,8 @@ static int alloc_bd_table(struct qtnf_pcie_bus_priv *priv)
 
 	pr_debug("TX descriptor table: vaddr=0x%p paddr=%pad\n", vaddr, &paddr);
 
-	priv->tx_bd_reclaim_start = 0;
-	priv->tx_bd_index = 0;
-	priv->tx_queue_len = 0;
+	priv->tx_bd_r_index = 0;
+	priv->tx_bd_w_index = 0;
 
 	/* rx bd */
 
@@ -413,8 +409,6 @@ static int alloc_bd_table(struct qtnf_pcie_bus_priv *priv)
 
 	pr_debug("RX descriptor table: vaddr=0x%p paddr=%pad\n", vaddr, &paddr);
 
-	priv->rx_bd_index = 0;
-
 	return 0;
 }
 
@@ -445,6 +439,8 @@ static int skb2rbd_attach(struct qtnf_pcie_bus_priv *priv, u16 index)
 	rxbd->addr_h = cpu_to_le32(QTN_HOST_HI32(paddr));
 	rxbd->info = 0x0;
 
+	priv->rx_bd_w_index = index;
+
 	/* sync up all descriptor updates */
 	wmb();
 
@@ -510,6 +506,8 @@ static int qtnf_pcie_init_xfer(struct qtnf_pcie_bus_priv *priv)
 
 	priv->tx_bd_num = tx_bd_size_param;
 	priv->rx_bd_num = rx_bd_size_param;
+	priv->rx_bd_w_index = 0;
+	priv->rx_bd_r_index = 0;
 
 	ret = alloc_skb_array(priv);
 	if (ret) {
@@ -532,67 +530,69 @@ static int qtnf_pcie_init_xfer(struct qtnf_pcie_bus_priv *priv)
 	return ret;
 }
 
-static int qtnf_pcie_data_tx_reclaim(struct qtnf_pcie_bus_priv *priv)
+static void qtnf_pcie_data_tx_reclaim(struct qtnf_pcie_bus_priv *priv)
 {
 	struct qtnf_tx_bd *txbd;
 	struct sk_buff *skb;
 	dma_addr_t paddr;
-	int last_sent;
-	int count;
+	u32 tx_done_index;
+	int count = 0;
 	int i;
 
-	last_sent = readl(PCIE_HDP_RX0DMA_CNT(priv->pcie_reg_base))
-			% priv->tx_bd_num;
-	i = priv->tx_bd_reclaim_start;
-	count = 0;
 
-	while (i != last_sent) {
-		skb = priv->tx_skb[i];
-		if (!skb)
-			break;
+	tx_done_index = readl(PCIE_HDP_RX0DMA_CNT(priv->pcie_reg_base))
+			& (priv->tx_bd_num - 1);
 
-		txbd = &priv->tx_bd_vbase[i];
-		paddr = QTN_HOST_ADDR(le32_to_cpu(txbd->addr_h),
-				      le32_to_cpu(txbd->addr));
-		pci_unmap_single(priv->pdev, paddr, skb->len, PCI_DMA_TODEVICE);
+	i = priv->tx_bd_r_index;
 
-		if (skb->dev) {
-			skb->dev->stats.tx_packets++;
-			skb->dev->stats.tx_bytes += skb->len;
+	while (CIRC_CNT(tx_done_index, i, priv->tx_bd_num)) {
+		skb = priv->tx_skb[i];
+		if (likely(skb)) {
+			txbd = &priv->tx_bd_vbase[i];
+			paddr = QTN_HOST_ADDR(le32_to_cpu(txbd->addr_h),
+					      le32_to_cpu(txbd->addr));
+			pci_unmap_single(priv->pdev, paddr, skb->len,
+					 PCI_DMA_TODEVICE);
+
+			if (skb->dev) {
+				skb->dev->stats.tx_packets++;
+				skb->dev->stats.tx_bytes += skb->len;
+
+				if (netif_queue_stopped(skb->dev))
+					netif_wake_queue(skb->dev);
+			}
 
-			if (netif_queue_stopped(skb->dev))
-				netif_wake_queue(skb->dev);
+			dev_kfree_skb_any(skb);
 		}
 
-		dev_kfree_skb_any(skb);
 		priv->tx_skb[i] = NULL;
-		priv->tx_queue_len--;
 		count++;
 
 		if (++i >= priv->tx_bd_num)
 			i = 0;
 	}
 
-	priv->tx_bd_reclaim_start = i;
 	priv->tx_reclaim_done += count;
 	priv->tx_reclaim_req++;
+	priv->tx_bd_r_index = i;
 
-	return count;
 }
 
-static bool qtnf_tx_queue_ready(struct qtnf_pcie_bus_priv *priv)
+static int qtnf_tx_queue_ready(struct qtnf_pcie_bus_priv *priv)
 {
-	if (priv->tx_queue_len >= priv->tx_bd_num - 1) {
+	if (!CIRC_SPACE(priv->tx_bd_w_index, priv->tx_bd_r_index,
+			priv->tx_bd_num)) {
 		pr_err_ratelimited("reclaim full Tx queue\n");
 		qtnf_pcie_data_tx_reclaim(priv);
 
-		if (priv->tx_queue_len >= priv->tx_bd_num - 1) {
+		if (!CIRC_SPACE(priv->tx_bd_w_index, priv->tx_bd_r_index,
+				priv->tx_bd_num)) {
 			priv->tx_full_count++;
-			return false;
+			return 0;
 		}
 	}
 
-	return true;
+	return 1;
 }
 
 static int qtnf_pcie_data_tx(struct qtnf_bus *bus, struct sk_buff *skb)
@@ -617,7 +617,7 @@ static int qtnf_pcie_data_tx(struct qtnf_bus *bus, struct sk_buff *skb)
 		return NETDEV_TX_BUSY;
 	}
 
-	i = priv->tx_bd_index;
+	i = priv->tx_bd_w_index;
 	priv->tx_skb[i] = skb;
 	len = skb->len;
 
@@ -649,8 +649,7 @@ static int qtnf_pcie_data_tx(struct qtnf_bus *bus, struct sk_buff *skb)
 	if (++i >= priv->tx_bd_num)
 		i = 0;
 
-	priv->tx_bd_index = i;
-	priv->tx_queue_len++;
+	priv->tx_bd_w_index = i;
 
 tx_done:
 	if (ret && skb) {
@@ -709,16 +708,19 @@ static irqreturn_t qtnf_interrupt(int irq, void *data)
 	return IRQ_HANDLED;
 }
 
-static inline void hw_txproc_wr_ptr_inc(struct qtnf_pcie_bus_priv *priv)
+static int qtnf_rx_data_ready(struct qtnf_pcie_bus_priv *priv)
 {
-	u32 index;
+	u16 index = priv->rx_bd_r_index;
+	struct qtnf_rx_bd *rxbd;
+	u32 descw;
 
-	index = priv->hw_txproc_wr_ptr;
+	rxbd = &priv->rx_bd_vbase[index];
+	descw = le32_to_cpu(rxbd->info);
 
-	if (++index >= priv->rx_bd_num)
-		index = 0;
+	if (descw & QTN_TXDONE_MASK)
+		return 1;
 
-	priv->hw_txproc_wr_ptr = index;
+	return 0;
 }
 
 static int qtnf_rx_poll(struct napi_struct *napi, int budget)
@@ -730,26 +732,52 @@ static int qtnf_rx_poll(struct napi_struct *napi, int budget)
 	int processed = 0;
 	struct qtnf_rx_bd *rxbd;
 	dma_addr_t skb_paddr;
+	int consume;
 	u32 descw;
-	u16 index;
+	u32 psize;
+	u16 r_idx;
+	u16 w_idx;
 	int ret;
 
-	index = priv->rx_bd_index;
-	rxbd = &priv->rx_bd_vbase[index];
+	while (processed < budget) {
 
-	descw = le32_to_cpu(rxbd->info);
 
-	while ((descw & QTN_TXDONE_MASK) && (processed < budget)) {
-		skb = priv->rx_skb[index];
+		if (!qtnf_rx_data_ready(priv))
+			goto rx_out;
 
-		if (likely(skb)) {
-			skb_put(skb, QTN_GET_LEN(descw));
+		r_idx = priv->rx_bd_r_index;
+		rxbd = &priv->rx_bd_vbase[r_idx];
+		descw = le32_to_cpu(rxbd->info);
+
+		skb = priv->rx_skb[r_idx];
+		psize = QTN_GET_LEN(descw);
+		consume = 1;
 
+		if (!(descw & QTN_TXDONE_MASK)) {
+			pr_warn("skip invalid rxbd[%d]\n", r_idx);
+			consume = 0;
+		}
+
+		if (!skb) {
+			pr_warn("skip missing rx_skb[%d]\n", r_idx);
+			consume = 0;
+		}
+
+		if (skb && (skb_tailroom(skb) <  psize)) {
+			pr_err("skip packet with invalid length: %u > %u\n",
+			       psize, skb_tailroom(skb));
+			consume = 0;
+		}
+
+		if (skb) {
 			skb_paddr = QTN_HOST_ADDR(le32_to_cpu(rxbd->addr_h),
 						  le32_to_cpu(rxbd->addr));
 			pci_unmap_single(priv->pdev, skb_paddr, SKB_BUF_SIZE,
 					 PCI_DMA_FROMDEVICE);
+		}
 
+		if (consume) {
+			skb_put(skb, psize);
 			ndev = qtnf_classify_skb(bus, skb);
 			if (likely(ndev)) {
 				ndev->stats.rx_packets++;
@@ -762,30 +790,38 @@ static int qtnf_rx_poll(struct napi_struct *napi, int budget)
 				bus->mux_dev.stats.rx_dropped++;
 				dev_kfree_skb_any(skb);
 			}
-
-			processed++;
 		} else {
-			pr_err("missing rx_skb[%d]\n", index);
+			if (skb) {
+				bus->mux_dev.stats.rx_dropped++;
+				dev_kfree_skb_any(skb);
+			}
 		}
 
-		/* attached rx buffer is passed upstream: map a new one */
-		ret = skb2rbd_attach(priv, index);
-		if (likely(!ret)) {
-			if (++index >= priv->rx_bd_num)
-				index = 0;
+		priv->rx_skb[r_idx] = NULL;
+		if (++r_idx >= priv->rx_bd_num)
+			r_idx = 0;
 
-			priv->rx_bd_index = index;
-			hw_txproc_wr_ptr_inc(priv);
+		priv->rx_bd_r_index = r_idx;
 
-			rxbd = &priv->rx_bd_vbase[index];
-			descw = le32_to_cpu(rxbd->info);
-		} else {
-			pr_err("failed to allocate new rx_skb[%d]\n", index);
-			break;
+		/* repalce processed buffer by a new one */
+		w_idx = priv->rx_bd_w_index;
+		while (CIRC_SPACE(priv->rx_bd_w_index, priv->rx_bd_r_index,
+				  priv->rx_bd_num) > 0) {
+			if (++w_idx >= priv->rx_bd_num)
+				w_idx = 0;
+
+			ret = skb2rbd_attach(priv, w_idx);
+			if (ret) {
+				pr_err("failed to allocate new rx_skb[%d]\n",
+				       w_idx);
+				break;
+			}
 		}
 
+		processed++;
 	}
 
+rx_out:
 	if (processed < budget) {
 		napi_complete(napi);
 		qtnf_en_rxdone_irq(priv);
@@ -1056,10 +1092,18 @@ static int qtnf_dbg_irq_stats(struct seq_file *s, void *data)
 {
 	struct qtnf_bus *bus = dev_get_drvdata(s->private);
 	struct qtnf_pcie_bus_priv *priv = get_bus_priv(bus);
+	u32 reg = readl(PCIE_HDP_INT_EN(priv->pcie_reg_base));
+	u32 status;
 
 	seq_printf(s, "pcie_irq_count(%u)\n", priv->pcie_irq_count);
 	seq_printf(s, "pcie_irq_tx_count(%u)\n", priv->pcie_irq_tx_count);
+	status = reg &  PCIE_HDP_INT_TX_BITS;
+	seq_printf(s, "pcie_irq_tx_status(%s)\n",
+		   (status == PCIE_HDP_INT_TX_BITS) ? "EN" : "DIS");
 	seq_printf(s, "pcie_irq_rx_count(%u)\n", priv->pcie_irq_rx_count);
+	status = reg &  PCIE_HDP_INT_RX_BITS;
+	seq_printf(s, "pcie_irq_rx_status(%s)\n",
+		   (status == PCIE_HDP_INT_RX_BITS) ? "EN" : "DIS");
 
 	return 0;
 }
@@ -1073,10 +1117,24 @@ static int qtnf_dbg_hdp_stats(struct seq_file *s, void *data)
 	seq_printf(s, "tx_done_count(%u)\n", priv->tx_done_count);
 	seq_printf(s, "tx_reclaim_done(%u)\n", priv->tx_reclaim_done);
 	seq_printf(s, "tx_reclaim_req(%u)\n", priv->tx_reclaim_req);
-	seq_printf(s, "tx_bd_reclaim_start(%u)\n", priv->tx_bd_reclaim_start);
-	seq_printf(s, "tx_bd_index(%u)\n", priv->tx_bd_index);
-	seq_printf(s, "rx_bd_index(%u)\n", priv->rx_bd_index);
-	seq_printf(s, "tx_queue_len(%u)\n", priv->tx_queue_len);
+
+	seq_printf(s, "tx_bd_r_index(%u)\n", priv->tx_bd_r_index);
+	seq_printf(s, "tx_bd_p_index(%u)\n",
+		   readl(PCIE_HDP_RX0DMA_CNT(priv->pcie_reg_base))
+			& (priv->tx_bd_num - 1));
+	seq_printf(s, "tx_bd_w_index(%u)\n", priv->tx_bd_w_index);
+	seq_printf(s, "tx queue len(%u)\n",
+		   CIRC_CNT(priv->tx_bd_w_index, priv->tx_bd_r_index,
+			    priv->tx_bd_num));
+
+	seq_printf(s, "rx_bd_r_index(%u)\n", priv->rx_bd_r_index);
+	seq_printf(s, "rx_bd_p_index(%u)\n",
+		   readl(PCIE_HDP_TX0DMA_CNT(priv->pcie_reg_base))
+			& (priv->rx_bd_num - 1));
+	seq_printf(s, "rx_bd_w_index(%u)\n", priv->rx_bd_w_index);
+	seq_printf(s, "rx alloc queue len(%u)\n",
+		   CIRC_SPACE(priv->rx_bd_w_index, priv->rx_bd_r_index,
+			      priv->rx_bd_num));
 
 	return 0;
 }
diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
index 2a897db2bd79..1b37914299e9 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
@@ -66,13 +66,11 @@ struct qtnf_pcie_bus_priv {
 	void *bd_table_vaddr;
 	u32 bd_table_len;
 
-	u32 hw_txproc_wr_ptr;
+	u32 rx_bd_w_index;
+	u32 rx_bd_r_index;
 
-	u16 tx_bd_reclaim_start;
-	u16 tx_bd_index;
-	u32 tx_queue_len;
-
-	u16 rx_bd_index;
+	u32 tx_bd_w_index;
+	u32 tx_bd_r_index;
 
 	u32 pcie_irq_mask;
 
-- 
2.11.0

[PATCH 7/9] qtnfmac: introduce counter for Rx underflow events

[Sergey Matyukevich]

Introduce counter for Rx underflow events. Export this counter via debugfs.

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
---
 drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c      | 16 ++++++++++++++--
 .../net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h |  1 +
 drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_ipc.h  |  1 +
 .../wireless/quantenna/qtnfmac/pearl/pcie_regs_pearl.h   |  1 +
 4 files changed, 17 insertions(+), 2 deletions(-)

diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
index f8207ab25576..72730aff2a41 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
@@ -686,14 +686,21 @@ static irqreturn_t qtnf_interrupt(int irq, void *data)
 	if (!(status & priv->pcie_irq_mask))
 		goto irq_done;
 
-	if (status & PCIE_HDP_INT_RX_BITS) {
+	if (status & PCIE_HDP_INT_RX_BITS)
 		priv->pcie_irq_rx_count++;
+
+	if (status & PCIE_HDP_INT_TX_BITS)
+		priv->pcie_irq_tx_count++;
+
+	if (status & PCIE_HDP_INT_HHBM_UF)
+		priv->pcie_irq_uf_count++;
+
+	if (status & PCIE_HDP_INT_RX_BITS) {
 		qtnf_dis_rxdone_irq(priv);
 		napi_schedule(&bus->mux_napi);
 	}
 
 	if (status & PCIE_HDP_INT_TX_BITS) {
-		priv->pcie_irq_tx_count++;
 		qtnf_dis_txdone_irq(priv);
 		tasklet_hi_schedule(&priv->reclaim_tq);
 	}
@@ -1104,6 +1111,10 @@ static int qtnf_dbg_irq_stats(struct seq_file *s, void *data)
 	status = reg &  PCIE_HDP_INT_RX_BITS;
 	seq_printf(s, "pcie_irq_rx_status(%s)\n",
 		   (status == PCIE_HDP_INT_RX_BITS) ? "EN" : "DIS");
+	seq_printf(s, "pcie_irq_uf_count(%u)\n", priv->pcie_irq_uf_count);
+	status = reg &  PCIE_HDP_INT_HHBM_UF;
+	seq_printf(s, "pcie_irq_hhbm_uf_status(%s)\n",
+		   (status == PCIE_HDP_INT_HHBM_UF) ? "EN" : "DIS");
 
 	return 0;
 }
@@ -1189,6 +1200,7 @@ static int qtnf_pcie_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 	pcie_priv->pcie_irq_count = 0;
 	pcie_priv->pcie_irq_rx_count = 0;
 	pcie_priv->pcie_irq_tx_count = 0;
+	pcie_priv->pcie_irq_uf_count = 0;
 	pcie_priv->tx_reclaim_done = 0;
 	pcie_priv->tx_reclaim_req = 0;
 
diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
index 1b37914299e9..698e42132ed4 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
@@ -78,6 +78,7 @@ struct qtnf_pcie_bus_priv {
 	u32 pcie_irq_count;
 	u32 pcie_irq_rx_count;
 	u32 pcie_irq_tx_count;
+	u32 pcie_irq_uf_count;
 	u32 tx_full_count;
 	u32 tx_done_count;
 	u32 tx_reclaim_done;
diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_ipc.h b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_ipc.h
index e00d508fbcf0..667f5ec457e3 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_ipc.h
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_ipc.h
@@ -50,6 +50,7 @@
 #define PCIE_HDP_INT_RX_BITS (0		\
 	| PCIE_HDP_INT_EP_TXDMA		\
 	| PCIE_HDP_INT_EP_TXEMPTY	\
+	| PCIE_HDP_INT_HHBM_UF		\
 	)
 
 #define PCIE_HDP_INT_TX_BITS (0		\
diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_regs_pearl.h b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_regs_pearl.h
index 78715b8a8ef9..69696f118769 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_regs_pearl.h
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_regs_pearl.h
@@ -333,6 +333,7 @@
 #define PCIE_HDP_INT_RX_LEN_ERR		(BIT(2))
 #define PCIE_HDP_INT_RX_HDR_LEN_ERR	(BIT(3))
 #define PCIE_HDP_INT_EP_TXDMA		(BIT(12))
+#define PCIE_HDP_INT_HHBM_UF		(BIT(13))
 #define PCIE_HDP_INT_EP_TXEMPTY		(BIT(15))
 #define PCIE_HDP_INT_IPC		(BIT(29))
 
-- 
2.11.0

[PATCH 8/9] qtnfmac: modify tx reclaim locking

[Sergey Matyukevich]

Perform additional reclaim from qtnf_pcie_data_tx. Lock tx_lock serves only
reclaim synchronization purposes. Rename it accordingly and improve
granularity moving this lock to qtnf_pcie_data_tx_reclaim.

Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
---
 drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c     | 17 ++++++-----------
 .../wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h    |  4 ++--
 2 files changed, 8 insertions(+), 13 deletions(-)

diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
index 72730aff2a41..cd2f2b667643 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie.c
@@ -534,11 +534,13 @@ static void qtnf_pcie_data_tx_reclaim(struct qtnf_pcie_bus_priv *priv)
 {
 	struct qtnf_tx_bd *txbd;
 	struct sk_buff *skb;
+	unsigned long flags;
 	dma_addr_t paddr;
 	u32 tx_done_index;
 	int count = 0;
 	int i;
 
+	spin_lock_irqsave(&priv->tx_reclaim_lock, flags);
 
 	tx_done_index = readl(PCIE_HDP_RX0DMA_CNT(priv->pcie_reg_base))
 			& (priv->tx_bd_num - 1);
@@ -576,6 +578,7 @@ static void qtnf_pcie_data_tx_reclaim(struct qtnf_pcie_bus_priv *priv)
 	priv->tx_reclaim_req++;
 	priv->tx_bd_r_index = i;
 
+	spin_unlock_irqrestore(&priv->tx_reclaim_lock, flags);
 }
 
 static int qtnf_tx_queue_ready(struct qtnf_pcie_bus_priv *priv)
@@ -600,20 +603,14 @@ static int qtnf_pcie_data_tx(struct qtnf_bus *bus, struct sk_buff *skb)
 	struct qtnf_pcie_bus_priv *priv = (void *)get_bus_priv(bus);
 	dma_addr_t txbd_paddr, skb_paddr;
 	struct qtnf_tx_bd *txbd;
-	unsigned long flags;
 	int len, i;
 	u32 info;
 	int ret = 0;
 
-	spin_lock_irqsave(&priv->tx_lock, flags);
-
-	priv->tx_done_count++;
-
 	if (!qtnf_tx_queue_ready(priv)) {
 		if (skb->dev)
 			netif_stop_queue(skb->dev);
 
-		spin_unlock_irqrestore(&priv->tx_lock, flags);
 		return NETDEV_TX_BUSY;
 	}
 
@@ -659,7 +656,8 @@ static int qtnf_pcie_data_tx(struct qtnf_bus *bus, struct sk_buff *skb)
 		dev_kfree_skb_any(skb);
 	}
 
-	spin_unlock_irqrestore(&priv->tx_lock, flags);
+	qtnf_pcie_data_tx_reclaim(priv);
+	priv->tx_done_count++;
 
 	return NETDEV_TX_OK;
 }
@@ -1067,11 +1065,8 @@ static int qtnf_bringup_fw(struct qtnf_bus *bus)
 static void qtnf_reclaim_tasklet_fn(unsigned long data)
 {
 	struct qtnf_pcie_bus_priv *priv = (void *)data;
-	unsigned long flags;
 
-	spin_lock_irqsave(&priv->tx_lock, flags);
 	qtnf_pcie_data_tx_reclaim(priv);
-	spin_unlock_irqrestore(&priv->tx_lock, flags);
 	qtnf_en_txdone_irq(priv);
 }
 
@@ -1192,7 +1187,7 @@ static int qtnf_pcie_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 	init_completion(&bus->request_firmware_complete);
 	mutex_init(&bus->bus_lock);
 	spin_lock_init(&pcie_priv->irq_lock);
-	spin_lock_init(&pcie_priv->tx_lock);
+	spin_lock_init(&pcie_priv->tx_reclaim_lock);
 
 	/* init stats */
 	pcie_priv->tx_full_count = 0;
diff --git a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
index 698e42132ed4..e76a23716ee0 100644
--- a/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
+++ b/drivers/net/wireless/quantenna/qtnfmac/pearl/pcie_bus_priv.h
@@ -32,8 +32,8 @@ struct qtnf_pcie_bus_priv {
 	/* lock for irq configuration changes */
 	spinlock_t irq_lock;
 
-	/* lock for tx operations */
-	spinlock_t tx_lock;
+	/* lock for tx reclaim operations */
+	spinlock_t tx_reclaim_lock;
 	u8 msi_enabled;
 	int mps;
 
-- 
2.11.0

Re: [PATCH] qtnfmac: pcie datapath optimizations and cleanups

[Kalle Valo]

Sergey Matyukevich <sergey.matyukevich.os@quantenna.com> writes:

> Hello Kalle and all,
>
> This patchset implements several optimizations and cleanups for
> pcie datapath in qtnfmac driver. Major changes include:
> - switch to kernel circ_buf implementation
> - modify tx reclaim locking
> - introduce counter for rx underflow events
>
> Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>
>
>  bus.h                   |    1 
>  pearl/pcie.c            |  285 +++++++++++++++++++++++++++++-------------------
>  pearl/pcie_bus_priv.h   |   15 +-
>  pearl/pcie_ipc.h        |    1 
>  pearl/pcie_regs_pearl.h |    1 
>  5 files changed, 183 insertions(+), 120 deletions(-)

I don't see patch 9 in the mailing list. Did it get lost somewhere?

-- 
Kalle Valo

Re: [PATCH] qtnfmac: pcie datapath optimizations and cleanups

[Sergey Matyukevich]

> I don't see patch 9 in the mailing list. Did it get lost somewhere?

No, this is my fault. The last patch was sort of experimental shuffling
dma memory barriers. I kept it for testing, but forgot to remove it
before sending. Let me know if it matters, then I will resend
a properly numbered patchset.

Regards,
Sergey

Re: [PATCH] qtnfmac: pcie datapath optimizations and cleanups

[Kalle Valo]

Sergey Matyukevich <sergey.matyukevich.os@quantenna.com> writes:

>> I don't see patch 9 in the mailing list. Did it get lost somewhere?
>
> No, this is my fault. The last patch was sort of experimental shuffling
> dma memory barriers. I kept it for testing, but forgot to remove it
> before sending. Let me know if it matters, then I will resend
> a properly numbered patchset.

Not a problem, I just always check if I see a patch missing.

-- 
Kalle Valo

Re: [PATCH] qtnfmac: pcie datapath optimizations and cleanups

[Igor Mitsyanko]

On 08/15/2017 06:06 AM, Sergey Matyukevich wrote:
> Hello Kalle and all,
> 
> This patchset implements several optimizations and cleanups for
> pcie datapath in qtnfmac driver. Major changes include:
> - switch to kernel circ_buf implementation
> - modify tx reclaim locking
> - introduce counter for rx underflow events
> 
> Signed-off-by: Sergey Matyukevich <sergey.matyukevich.os@quantenna.com>

Entire series:
Tested-by: Igor Mitsyanko <igor.mitsyanko.os@quantenna.com>

> 
>   bus.h                   |    1
>   pearl/pcie.c            |  285 +++++++++++++++++++++++++++++-------------------
>   pearl/pcie_bus_priv.h   |   15 +-
>   pearl/pcie_ipc.h        |    1
>   pearl/pcie_regs_pearl.h |    1
>   5 files changed, 183 insertions(+), 120 deletions(-)
>