Skip to contentSkip to conversation

p54spi: get rid of busy-wait loops

PATCH In review Max Filippov 2 participants 4 messages Apr 5, 2009 – Apr 6, 2009 Networking drivers (wireless)

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 p54spi: get rid of busy-wait loops

p54spi_wakeup and p54spi_tx_frame used busy-waiting loop to poll for 'ready' bits in SPI_ADRS_HOST_INTERRUPTS register. With this change in place 'WR_READY timeout' messages do not show anymore.

---

I wish I knew for sure why. My guess is that busy-waiting somehow interferes with SPI transfer scheduling.

 drivers/net/wireless/p54/p54spi.c |   41 ++++++++++++------------------------
 1 files changed, 14 insertions(+), 27 deletions(-)
drivers/net/wireless/p54/p54spi.c +14 -27
index 2903672..e4e708e 100644
@@ -153,14 +153,13 @@ static const struct p54spi_spi_reg p54spi_registers_array[] =
 static int p54spi_wait_bit(struct p54s_priv *priv, u16 reg, __le32 bits)
 {
 	int i;
-	__le32 buffer;
 
 	for (i = 0; i < 2000; i++) {
-		p54spi_spi_read(priv, reg, &buffer, sizeof(buffer));
+		__le32 buffer = p54spi_read32(priv, reg);
 		if ((buffer & bits) == bits)
 			return 1;
 
-		msleep(1);
+		msleep(0);
 	}
 	return 0;
 }
@@ -171,10 +170,10 @@ static int p54spi_spi_write_dma(struct p54s_priv *priv, __le32 base,
 	p54spi_write16(priv, SPI_ADRS_DMA_WRITE_CTRL,
 		       cpu_to_le16(SPI_DMA_WRITE_CTRL_ENABLE));
 
-	if (p54spi_wait_bit(priv, SPI_ADRS_DMA_WRITE_CTRL,
-			    cpu_to_le32(HOST_ALLOWED)) == 0) {
+	if (!p54spi_wait_bit(priv, SPI_ADRS_DMA_WRITE_CTRL,
+			     cpu_to_le32(HOST_ALLOWED))) {
 		dev_err(&priv->spi->dev, "spi_write_dma not allowed "
-			"to DMA write.");
+			"to DMA write.\n");
 		return -EAGAIN;
 	}
 
@@ -316,21 +315,15 @@ static inline void p54spi_int_ack(struct p54s_priv *priv, u32 val)
 
 static void p54spi_wakeup(struct p54s_priv *priv)
 {
-	unsigned long timeout;
-	u32 ints;
-
 	/* wake the chip */
 	p54spi_write32(priv, SPI_ADRS_ARM_INTERRUPTS,
 		       cpu_to_le32(SPI_TARGET_INT_WAKEUP));
 
 	/* And wait for the READY interrupt */
-	timeout = jiffies + HZ;
-
-	ints =  p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
-	while (!(ints & SPI_HOST_INT_READY)) {
-		if (time_after(jiffies, timeout))
-				goto out;
-		ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
+	if (!p54spi_wait_bit(priv, SPI_ADRS_HOST_INTERRUPTS,
+			     cpu_to_le32(SPI_HOST_INT_READY))) {
+		dev_err(&priv->spi->dev, "INT_READY timeout\n");
+		goto out;
 	}
 
 	p54spi_int_ack(priv, SPI_HOST_INT_READY);
@@ -418,9 +411,7 @@ static irqreturn_t p54spi_interrupt(int irq, void *config)
 static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb)
 {
 	struct p54_hdr *hdr = (struct p54_hdr *) skb->data;
-	unsigned long timeout;
 	int ret = 0;
-	u32 ints;
 
 	p54spi_wakeup(priv);
 
@@ -428,15 +419,11 @@ static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb)
 	if (ret < 0)
 		goto out;
 
-	timeout = jiffies + 2 * HZ;
-	ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
-	while (!(ints & SPI_HOST_INT_WR_READY)) {
-		if (time_after(jiffies, timeout)) {
-			dev_err(&priv->spi->dev, "WR_READY timeout\n");
-			ret = -1;
-			goto out;
-		}
-		ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
+	if (!p54spi_wait_bit(priv, SPI_ADRS_HOST_INTERRUPTS,
+			     cpu_to_le32(SPI_HOST_INT_WR_READY))) {
+		dev_err(&priv->spi->dev, "WR_READY timeout\n");
+		ret = -1;
+		goto out;
 	}
 
 	p54spi_int_ack(priv, SPI_HOST_INT_WR_READY);
Signature · 1 line
1.5.4.3
Acked-by on v1 1/1
#1 lore raw
Quoted 7 lines · On Sunday 05 April 2009 23:03:09 Max Filippov wrote:

p54spi_wakeup and p54spi_tx_frame used busy-waiting loop to poll for 'ready' bits in SPI_ADRS_HOST_INTERRUPTS register. With this change in place 'WR_READY timeout' messages do not show anymore.

---

looks good to me.

Quoted 14 lines

I wish I knew for sure why. My guess is that busy-waiting somehow interferes with SPI transfer scheduling.

 drivers/net/wireless/p54/p54spi.c |   41 ++++++++++++------------------------
 1 files changed, 14 insertions(+), 27 deletions(-)
drivers/net/wireless/p54/p54spi.c +0 -0
index 2903672..e4e708e 100644
@@ -153,14 +153,13 @@ static const struct p54spi_spi_reg p54spi_registers_array[] =
 static int p54spi_wait_bit(struct p54s_priv *priv, u16 reg, __le32 bits)
 {
 	int i;

drivers/net/wireless/p54/p54spi.c static const struct p54spi_spi_reg p54spi_registers_array[] = · line 162

Quoted 10 lines
drivers/net/wireless/p54/p54spi.c +2 -3
-	__le32 buffer;
 
 	for (i = 0; i < 2000; i++) {
-		p54spi_spi_read(priv, reg, &buffer, sizeof(buffer));
+		__le32 buffer = p54spi_read32(priv, reg);
 		if ((buffer & bits) == bits)
 			return 1;
 
-		msleep(1);
+		msleep(0);

I never liked msleep in workqueues. So instead of msleep(0), what about replacing it with udelay or remove it entirely?

OR: maybe you can get the chip to generate IRQs when its ready, so we don't have to poll it at all?

Quoted 54 lines
drivers/net/wireless/p54/p54spi.c +7 -15
 	}
 	return 0;
 }
@@ -171,10 +170,10 @@ static int p54spi_spi_write_dma(struct p54s_priv *priv, __le32 base,
 	p54spi_write16(priv, SPI_ADRS_DMA_WRITE_CTRL,
 		       cpu_to_le16(SPI_DMA_WRITE_CTRL_ENABLE));
 
-	if (p54spi_wait_bit(priv, SPI_ADRS_DMA_WRITE_CTRL,
-			    cpu_to_le32(HOST_ALLOWED)) == 0) {
+	if (!p54spi_wait_bit(priv, SPI_ADRS_DMA_WRITE_CTRL,
+			     cpu_to_le32(HOST_ALLOWED))) {
 		dev_err(&priv->spi->dev, "spi_write_dma not allowed "
-			"to DMA write.");
+			"to DMA write.\n");
 		return -EAGAIN;
 	}
 
@@ -316,21 +315,15 @@ static inline void p54spi_int_ack(struct p54s_priv *priv, u32 val)
 
 static void p54spi_wakeup(struct p54s_priv *priv)
 {
-	unsigned long timeout;
-	u32 ints;
-
 	/* wake the chip */
 	p54spi_write32(priv, SPI_ADRS_ARM_INTERRUPTS,
 		       cpu_to_le32(SPI_TARGET_INT_WAKEUP));
 
 	/* And wait for the READY interrupt */
-	timeout = jiffies + HZ;
-
-	ints =  p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
-	while (!(ints & SPI_HOST_INT_READY)) {
-		if (time_after(jiffies, timeout))
-				goto out;
-		ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
+	if (!p54spi_wait_bit(priv, SPI_ADRS_HOST_INTERRUPTS,
+			     cpu_to_le32(SPI_HOST_INT_READY))) {
+		dev_err(&priv->spi->dev, "INT_READY timeout\n");
+		goto out;
 	}
 
 	p54spi_int_ack(priv, SPI_HOST_INT_READY);
@@ -418,9 +411,7 @@ static irqreturn_t p54spi_interrupt(int irq, void *config)
 static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb)
 {
 	struct p54_hdr *hdr = (struct p54_hdr *) skb->data;
-	unsigned long timeout;
 	int ret = 0;
-	u32 ints;
 
 	p54spi_wakeup(priv);
 

drivers/net/wireless/p54/p54spi.c static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb) · line 436

Quoted 9 lines
drivers/net/wireless/p54/p54spi.c +0 -6
@@ -428,15 +419,11 @@ static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb)
 	if (ret < 0)
 		goto out;
 
-	timeout = jiffies + 2 * HZ;
-	ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
-	while (!(ints & SPI_HOST_INT_WR_READY)) {
-		if (time_after(jiffies, timeout)) {
-			dev_err(&priv->spi->dev, "WR_READY timeout\n");
-			ret = -1;

what about replacing the generic code "-1" with EBUSY/EIO or ETIMEDOUT?

drivers/net/wireless/p54/p54spi.c static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb) · line 425

Quoted 7 lines
drivers/net/wireless/p54/p54spi.c +4 -3
-			goto out;
-		}
-		ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
+	if (!p54spi_wait_bit(priv, SPI_ADRS_HOST_INTERRUPTS,
+			     cpu_to_le32(SPI_HOST_INT_WR_READY))) {
+		dev_err(&priv->spi->dev, "WR_READY timeout\n");
+		ret = -1;

here too?!

drivers/net/wireless/p54/p54spi.c static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb) · line 429

Quoted 4 lines
drivers/net/wireless/p54/p54spi.c +1 -0
+		goto out;
 	}
 
 	p54spi_int_ack(priv, SPI_HOST_INT_WR_READY);

---

By the way a few days ago the libertas driver authors posted a patch which "improve the average throughput 13%."

http://marc.info/?l=linux-wireless&m=123871643203161

I wonder if this would also work with p54spi.
But to test this theory, one has to rewrite the driver to use kthread instead of hogging workqueues.... just in case you're looking for more TODOs ;-)

Regards,
Chr

#2 lore raw
Quoted 35 lines · On Sunday 05 April 2009 23:03:09 Max Filippov wrote:
Quoted 7 lines

p54spi_wakeup and p54spi_tx_frame used busy-waiting loop to poll for 'ready' bits in SPI_ADRS_HOST_INTERRUPTS register. With this change in place 'WR_READY timeout' messages do not show anymore.

---

looks good to me.

Quoted 23 lines

I wish I knew for sure why. My guess is that busy-waiting somehow interferes with SPI transfer scheduling.

 drivers/net/wireless/p54/p54spi.c |   41

++++++++++++------------------------ 1 files changed, 14 insertions(+), 27 deletions(-)

(quoted diff) +0 -0

b/drivers/net/wireless/p54/p54spi.c index 2903672..e4e708e 100644

drivers/net/wireless/p54/p54spi.c +0 -0
@@ -153,14 +153,13 @@ static const struct p54spi_spi_reg

p54spi_registers_array[] = static int p54spi_wait_bit(struct p54s_priv *priv, u16 reg, __le32 bits) {

drivers/net/wireless/p54/p54spi.c +1 -2
 	int i;
-	__le32 buffer;
 
 	for (i = 0; i < 2000; i++) {
-		p54spi_spi_read(priv, reg, &buffer, sizeof(buffer));
+		__le32 buffer = p54spi_read32(priv, reg);
 		if ((buffer & bits) == bits)
 			return 1;
Quoted 5 lines

drivers/net/wireless/p54/p54spi.c static const struct p54spi_spi_reg p54spi_registers_array[] = · line 162

Quoted 2 lines
drivers/net/wireless/p54/p54spi.c +1 -1
-		msleep(1);
+		msleep(0);

I never liked msleep in workqueues. So instead of msleep(0), what about replacing it with udelay or remove it entirely?

The whole point was to delay these repeated SPI reads (:
I'm not sure if kernel scheduling is important here. Will test and repost the patch.

Quoted 2 lines

OR: maybe you can get the chip to generate IRQs when its ready, so we don't have to poll it at all?

Will try it.

Quoted 67 lines
Quoted 55 lines
drivers/net/wireless/p54/p54spi.c +7 -13
 	}
 	return 0;
 }
@@ -171,10 +170,10 @@ static int p54spi_spi_write_dma(struct p54s_priv
 *priv, __le32 base, p54spi_write16(priv, SPI_ADRS_DMA_WRITE_CTRL,
 		       cpu_to_le16(SPI_DMA_WRITE_CTRL_ENABLE));
 
-	if (p54spi_wait_bit(priv, SPI_ADRS_DMA_WRITE_CTRL,
-			    cpu_to_le32(HOST_ALLOWED)) == 0) {
+	if (!p54spi_wait_bit(priv, SPI_ADRS_DMA_WRITE_CTRL,
+			     cpu_to_le32(HOST_ALLOWED))) {
 		dev_err(&priv->spi->dev, "spi_write_dma not allowed "
-			"to DMA write.");
+			"to DMA write.\n");
 		return -EAGAIN;
 	}
 
@@ -316,21 +315,15 @@ static inline void p54spi_int_ack(struct p54s_priv
 *priv, u32 val)
 
 static void p54spi_wakeup(struct p54s_priv *priv)
 {
-	unsigned long timeout;
-	u32 ints;
-
 	/* wake the chip */
 	p54spi_write32(priv, SPI_ADRS_ARM_INTERRUPTS,
 		       cpu_to_le32(SPI_TARGET_INT_WAKEUP));
 
 	/* And wait for the READY interrupt */
-	timeout = jiffies + HZ;
-
-	ints =  p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
-	while (!(ints & SPI_HOST_INT_READY)) {
-		if (time_after(jiffies, timeout))
-				goto out;
-		ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
+	if (!p54spi_wait_bit(priv, SPI_ADRS_HOST_INTERRUPTS,
+			     cpu_to_le32(SPI_HOST_INT_READY))) {
+		dev_err(&priv->spi->dev, "INT_READY timeout\n");
+		goto out;
 	}
 
drivers/net/wireless/p54/p54spi.c +0 -2
 	p54spi_int_ack(priv, SPI_HOST_INT_READY);
@@ -418,9 +411,7 @@ static irqreturn_t p54spi_interrupt(int irq, void
 *config) static int p54spi_tx_frame(struct p54s_priv *priv, struct
 sk_buff *skb) {
 	struct p54_hdr *hdr = (struct p54_hdr *) skb->data;
-	unsigned long timeout;
 	int ret = 0;
-	u32 ints;
 
 	p54spi_wakeup(priv);
 

drivers/net/wireless/p54/p54spi.c static int p54spi_tx_frame(struct p54s_priv *priv, · line 436

Quoted 9 lines
drivers/net/wireless/p54/p54spi.c +0 -6
@@ -428,15 +419,11 @@ static int p54spi_tx_frame(struct p54s_priv *priv,
 struct sk_buff *skb) if (ret < 0)
 		goto out;
 
-	timeout = jiffies + 2 * HZ;
-	ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
-	while (!(ints & SPI_HOST_INT_WR_READY)) {
-		if (time_after(jiffies, timeout)) {
-			dev_err(&priv->spi->dev, "WR_READY timeout\n");
-			ret = -1;

what about replacing the generic code "-1" with EBUSY/EIO or ETIMEDOUT?

Quoted 9 lines

drivers/net/wireless/p54/p54spi.c static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb) · line 425

Quoted 7 lines
drivers/net/wireless/p54/p54spi.c +4 -3
-			goto out;
-		}
-		ints = p54spi_read32(priv, SPI_ADRS_HOST_INTERRUPTS);
+	if (!p54spi_wait_bit(priv, SPI_ADRS_HOST_INTERRUPTS,
+			     cpu_to_le32(SPI_HOST_INT_WR_READY))) {
+		dev_err(&priv->spi->dev, "WR_READY timeout\n");
+		ret = -1;

here too?!

Will do so.

Quoted 10 lines

drivers/net/wireless/p54/p54spi.c static int p54spi_tx_frame(struct p54s_priv *priv, struct sk_buff *skb) · line 429

Quoted 4 lines
drivers/net/wireless/p54/p54spi.c +1 -0
+		goto out;
 	}
 
 	p54spi_int_ack(priv, SPI_HOST_INT_WR_READY);

---

By the way a few days ago the libertas driver authors posted a patch which "improve the average throughput 13%."

Quoted 6 lines

http://marc.info/?l=linux-wireless&m=123871643203161

I wonder if this would also work with p54spi.
But to test this theory, one has to rewrite the driver to use kthread instead of hogging workqueues.... just in case you're looking for more TODOs ;-)

I'm just getting it to work. And there are still problems in tx path, e.g. it cannot withstand dd if=/dev/zero | ssh ... "cat >/dev/null". Sure I will revisit performance issues once they remain the most important ones.

Quoted 3 lines

Regards,
Chr

Signature · 2 lines
Thanks.
-- Max
#3 lore raw
Quoted 4 lines · On Monday 06 April 2009 07:16:02 Max Filippov wrote:
Quoted 4 lines

By the way a few days ago the libertas driver authors posted a patch which "improve the average throughput 13%."

Quoted 8 lines
Quoted 6 lines

http://marc.info/?l=linux-wireless&m=123871643203161

I wonder if this would also work with p54spi.
But to test this theory, one has to rewrite the driver to use kthread instead of hogging workqueues.... just in case you're looking for more TODOs ;-)

I'm just getting it to work. And there are still problems in tx path, e.g. it cannot withstand dd if=/dev/zero | ssh ... "cat >/dev/null".

hmm, sounds familiar.
(" p54usb: fix random traffic stalls (LM87)" (43af18f06d5 and p54usb: fix packet loss with first generation devices ")

can you please run
for ((size=0; size < 9000;size++)); do ping $IP -s $size

and see if there's a connection between the stalls/hangs and the frame length?

Regards,
Chr