[PATCH linux-next v1 0/2] improve imx spi performance

[PATCH linux-next v1 0/2] improve imx spi performance

[Jiada Wang]

previously burst length (BURST_LENGTH) is always set to equal
to bits_per_word, causes a 10us gap between each word in transfer,
which significantly affects performance.

This patch set uses 32 bits tranfser to simulate lowers bits transfer,
and by set burst length to maximum possible value to reduce number of gaps
in both PIO & DMA transfers.

* This patch set is against linux-next tree

Jiada Wang (2):
  spi: imx: dynamic burst length adjust for PIO mode
  spi: imx: dynamic burst length adjust for DMA mode

 drivers/spi/spi-imx.c | 293 +++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 267 insertions(+), 26 deletions(-)

-- 
2.9.3

--
To unsubscribe from this list: send the line "unsubscribe linux-spi" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH linux-next v1 0/2] improve imx spi performance

[Jiada Wang]

previously burst length (BURST_LENGTH) is always set to equal
to bits_per_word, causes a 10us gap between each word in transfer,
which significantly affects performance.

This patch set uses 32 bits tranfser to simulate lowers bits transfer,
and by set burst length to maximum possible value to reduce number of gaps
in both PIO & DMA transfers.

* This patch set is against linux-next tree

Jiada Wang (2):
  spi: imx: dynamic burst length adjust for PIO mode
  spi: imx: dynamic burst length adjust for DMA mode

 drivers/spi/spi-imx.c | 293 +++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 267 insertions(+), 26 deletions(-)

-- 
2.9.3

[PATCH 1/2] spi: imx: dynamic burst length adjust for PIO mode

[Jiada Wang]

previously burst length (BURST_LENGTH) is always set to equal
to bits_per_word, causes a 10us gap between each word in
transfer, which significantly affects performance.

This patch uses 32 bits transfer to simulate lower bits transfer,
and adjusts burst length runtimely to use biggeest burst length
as possible to reduce the gaps in transfer for PIO mode.

Signed-off-by: Jiada Wang <jiada_wang@mentor.com>
---
 drivers/spi/spi-imx.c | 151 +++++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 143 insertions(+), 8 deletions(-)

diff --git a/drivers/spi/spi-imx.c b/drivers/spi/spi-imx.c
index 9a7c62f..04b4ea8 100644
--- a/drivers/spi/spi-imx.c
+++ b/drivers/spi/spi-imx.c
@@ -56,9 +56,11 @@
 
 /* The maximum  bytes that a sdma BD can transfer.*/
 #define MAX_SDMA_BD_BYTES  (1 << 15)
+#define MX51_ECSPI_CTRL_MAX_BURST	512
 struct spi_imx_config {
 	unsigned int speed_hz;
 	unsigned int bpw;
+	unsigned int len;
 };
 
 enum spi_imx_devtype {
@@ -96,12 +98,14 @@ struct spi_imx_data {
 
 	unsigned int bytes_per_word;
 
-	unsigned int count;
+	unsigned int count, count_index;
 	void (*tx)(struct spi_imx_data *);
 	void (*rx)(struct spi_imx_data *);
 	void *rx_buf;
 	const void *tx_buf;
 	unsigned int txfifo; /* number of words pushed in tx FIFO */
+	unsigned int dynamic_burst, bpw_rx;
+	unsigned int bpw_w;
 
 	/* DMA */
 	bool usedma;
@@ -250,6 +254,7 @@ static bool spi_imx_can_dma(struct spi_master *master, struct spi_device *spi,
 #define MX51_ECSPI_CTRL_PREDIV_OFFSET	12
 #define MX51_ECSPI_CTRL_CS(cs)		((cs) << 18)
 #define MX51_ECSPI_CTRL_BL_OFFSET	20
+#define MX51_ECSPI_CTRL_BL_MASK		(0xfff << 20)
 
 #define MX51_ECSPI_CONFIG	0x0c
 #define MX51_ECSPI_CONFIG_SCLKPHA(cs)	(1 << ((cs) +  0))
@@ -277,6 +282,79 @@ static bool spi_imx_can_dma(struct spi_master *master, struct spi_device *spi,
 #define MX51_ECSPI_TESTREG	0x20
 #define MX51_ECSPI_TESTREG_LBC	BIT(31)
 
+static void spi_imx_u32_swap_u8(struct spi_transfer *transfer, u8 *buf)
+{
+	int i;
+
+	for (i = 0; i < transfer->len / 4; i++) {
+		u8 temp;
+
+		temp = *(buf + i * 4);
+		*(buf + i * 4) = *(buf + i * 4 + 3);
+		*(buf + i * 4 + 3) = temp;
+
+		temp = *(buf + i * 4 + 1);
+		*(u8 *)(buf + i * 4 + 1) = *(buf + i * 4 + 2);
+		*(buf + i * 4 + 2) = temp;
+	}
+}
+
+static void spi_imx_u32_swap_u16(struct spi_transfer *transfer, u16 *buf)
+{
+	int i;
+
+	for (i = 0; i < transfer->len / 4; i++) {
+		u16 temp;
+
+		temp = *(buf + i * 2);
+		*(buf + i * 2) = *(buf + i * 2 + 1);
+		*(buf + i * 2 + 1) = temp;
+	}
+}
+
+static void spi_imx_buf_rx_swap(struct spi_imx_data *spi_imx)
+{
+	if (!spi_imx->bpw_rx) {
+		spi_imx_buf_rx_u32(spi_imx);
+		return;
+	}
+
+	if (spi_imx->bpw_w == 1)
+		spi_imx_buf_rx_u8(spi_imx);
+	else if (spi_imx->bpw_w == 2)
+		spi_imx_buf_rx_u16(spi_imx);
+}
+
+static void spi_imx_buf_tx_swap(struct spi_imx_data *spi_imx)
+{
+	u32 ctrl, val;
+
+	if (spi_imx->count == spi_imx->count_index) {
+		spi_imx->count_index = spi_imx->count > sizeof(u32) ?
+					spi_imx->count % sizeof(u32) : 0;
+		ctrl = readl(spi_imx->base + MX51_ECSPI_CTRL);
+		ctrl &= ~MX51_ECSPI_CTRL_BL_MASK;
+		if (spi_imx->count >= sizeof(u32))
+			val = spi_imx->count - spi_imx->count_index;
+		else {
+			val = spi_imx->bpw_w;
+			spi_imx->bpw_rx = 1;
+		}
+		ctrl |= ((val * 8 - 1) << MX51_ECSPI_CTRL_BL_OFFSET);
+		writel(ctrl, spi_imx->base + MX51_ECSPI_CTRL);
+	}
+
+	if (spi_imx->count >= sizeof(u32)) {
+		spi_imx_buf_tx_u32(spi_imx);
+		return;
+	}
+
+	if (spi_imx->bpw_w == 1)
+		spi_imx_buf_tx_u8(spi_imx);
+	else if (spi_imx->bpw_w == 2)
+		spi_imx_buf_tx_u16(spi_imx);
+}
+
 /* MX51 eCSPI */
 static unsigned int mx51_ecspi_clkdiv(struct spi_imx_data *spi_imx,
 				      unsigned int fspi, unsigned int *fres)
@@ -362,7 +440,14 @@ static int mx51_ecspi_config(struct spi_device *spi,
 	/* set chip select to use */
 	ctrl |= MX51_ECSPI_CTRL_CS(spi->chip_select);
 
-	ctrl |= (config->bpw - 1) << MX51_ECSPI_CTRL_BL_OFFSET;
+	if (spi_imx->dynamic_burst) {
+		if (config->len > MX51_ECSPI_CTRL_MAX_BURST)
+			ctrl |= MX51_ECSPI_CTRL_BL_MASK;
+		else
+			ctrl |= (((config->len - config->len % 4) * 8 - 1) <<
+				MX51_ECSPI_CTRL_BL_OFFSET);
+	} else
+		ctrl |= (config->bpw - 1) << MX51_ECSPI_CTRL_BL_OFFSET;
 
 	cfg |= MX51_ECSPI_CONFIG_SBBCTRL(spi->chip_select);
 
@@ -797,6 +882,8 @@ static void spi_imx_push(struct spi_imx_data *spi_imx)
 	while (spi_imx->txfifo < spi_imx_get_fifosize(spi_imx)) {
 		if (!spi_imx->count)
 			break;
+		if (spi_imx->txfifo && (spi_imx->count == spi_imx->count_index))
+			break;
 		spi_imx->tx(spi_imx);
 		spi_imx->txfifo++;
 	}
@@ -887,8 +974,12 @@ static int spi_imx_setupxfer(struct spi_device *spi,
 	struct spi_imx_config config;
 	int ret;
 
+	spi_imx->dynamic_burst = 0;
+	spi_imx->bpw_rx = 0;
+
 	config.bpw = t ? t->bits_per_word : spi->bits_per_word;
 	config.speed_hz  = t ? t->speed_hz : spi->max_speed_hz;
+	config.len = t->len;
 
 	if (!config.speed_hz)
 		config.speed_hz = spi->max_speed_hz;
@@ -897,14 +988,32 @@ static int spi_imx_setupxfer(struct spi_device *spi,
 
 	/* Initialize the functions for transfer */
 	if (config.bpw <= 8) {
-		spi_imx->rx = spi_imx_buf_rx_u8;
-		spi_imx->tx = spi_imx_buf_tx_u8;
+		if (t->len >= sizeof(u32) && is_imx51_ecspi(spi_imx)) {
+			spi_imx->dynamic_burst = 1;
+			spi_imx->rx = spi_imx_buf_rx_swap;
+			spi_imx->tx = spi_imx_buf_tx_swap;
+		} else {
+			spi_imx->rx = spi_imx_buf_rx_u8;
+			spi_imx->tx = spi_imx_buf_tx_u8;
+		}
 	} else if (config.bpw <= 16) {
-		spi_imx->rx = spi_imx_buf_rx_u16;
-		spi_imx->tx = spi_imx_buf_tx_u16;
+		if (t->len >= sizeof(u32) && is_imx51_ecspi(spi_imx)) {
+			spi_imx->dynamic_burst = 1;
+			spi_imx->rx = spi_imx_buf_rx_swap;
+			spi_imx->tx = spi_imx_buf_tx_swap;
+		} else {
+			spi_imx->rx = spi_imx_buf_rx_u16;
+			spi_imx->tx = spi_imx_buf_tx_u16;
+		}
 	} else {
-		spi_imx->rx = spi_imx_buf_rx_u32;
-		spi_imx->tx = spi_imx_buf_tx_u32;
+		if (is_imx51_ecspi(spi_imx)) {
+			spi_imx->dynamic_burst = 1;
+			spi_imx->rx = spi_imx_buf_rx_swap;
+			spi_imx->tx = spi_imx_buf_tx_swap;
+		} else {
+			spi_imx->rx = spi_imx_buf_rx_u32;
+			spi_imx->tx = spi_imx_buf_tx_u32;
+		}
 	}
 
 	if (spi_imx_can_dma(spi_imx->bitbang.master, spi, t))
@@ -912,6 +1021,8 @@ static int spi_imx_setupxfer(struct spi_device *spi,
 	else
 		spi_imx->usedma = 0;
 
+	spi_imx->bpw_w = DIV_ROUND_UP(config.bpw, 8);
+
 	if (spi_imx->usedma) {
 		ret = spi_imx_dma_configure(spi->master,
 					    spi_imx_bytes_per_word(config.bpw));
@@ -1086,6 +1197,20 @@ static int spi_imx_pio_transfer(struct spi_device *spi,
 	spi_imx->count = transfer->len;
 	spi_imx->txfifo = 0;
 
+	if (spi_imx->dynamic_burst) {
+		spi_imx->count_index =
+				spi_imx->count > MX51_ECSPI_CTRL_MAX_BURST ?
+				spi_imx->count % MX51_ECSPI_CTRL_MAX_BURST :
+				spi_imx->count % sizeof(u32);
+
+		if (spi_imx->bpw_w == 1)
+			spi_imx_u32_swap_u8(transfer,
+					(u8 *)transfer->tx_buf);
+		else if (spi_imx->bpw_w == 2)
+			spi_imx_u32_swap_u16(transfer,
+					(u16 *)transfer->tx_buf);
+	}
+
 	reinit_completion(&spi_imx->xfer_done);
 
 	spi_imx_push(spi_imx);
@@ -1102,6 +1227,16 @@ static int spi_imx_pio_transfer(struct spi_device *spi,
 		return -ETIMEDOUT;
 	}
 
+	if (spi_imx->dynamic_burst) {
+		if (spi_imx->bpw_w == 1)
+			spi_imx_u32_swap_u8(transfer,
+					(u8 *)transfer->rx_buf);
+		else if (spi_imx->bpw_w == 2)
+			spi_imx_u32_swap_u16(transfer,
+					(u16 *)transfer->rx_buf);
+		spi_imx->dynamic_burst = 0;
+	}
+
 	return transfer->len;
 }
 
-- 
2.9.3

Re: [PATCH 1/2] spi: imx: dynamic burst length adjust for PIO mode

[Mark Brown]

[-- Attachment #1: Type: text/plain, Size: 902 bytes --]

On Wed, Feb 08, 2017 at 03:20:27PM +0900, Jiada Wang wrote:

This looks basically fine, a couple of fairly minor things here:

> +	for (i = 0; i < transfer->len / 4; i++) {
> +		u8 temp;
> +
> +		temp = *(buf + i * 4);
> +		*(buf + i * 4) = *(buf + i * 4 + 3);
> +		*(buf + i * 4 + 3) = temp;

Should this be using one of the cpu_to_ functions?  It's a bit unclear
what the goal is here and if it'll work if the kernel is big endian
(which people do do with i.MX systems IIRC).

> +	if (spi_imx->bpw_w == 1)
> +		spi_imx_buf_rx_u8(spi_imx);
> +	else if (spi_imx->bpw_w == 2)
> +		spi_imx_buf_rx_u16(spi_imx);

switch statement please.

> +	if (spi_imx->dynamic_burst) {
> +		spi_imx->count_index =
> +				spi_imx->count > MX51_ECSPI_CTRL_MAX_BURST ?
> +				spi_imx->count % MX51_ECSPI_CTRL_MAX_BURST :
> +				spi_imx->count % sizeof(u32);

Just write a normal if statement please, it's easier to read.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH 1/2] spi: imx: dynamic burst length adjust for PIO mode

[Mark Brown]

[-- Attachment #1: Type: text/plain, Size: 902 bytes --]

On Wed, Feb 08, 2017 at 03:20:27PM +0900, Jiada Wang wrote:

This looks basically fine, a couple of fairly minor things here:

> +	for (i = 0; i < transfer->len / 4; i++) {
> +		u8 temp;
> +
> +		temp = *(buf + i * 4);
> +		*(buf + i * 4) = *(buf + i * 4 + 3);
> +		*(buf + i * 4 + 3) = temp;

Should this be using one of the cpu_to_ functions?  It's a bit unclear
what the goal is here and if it'll work if the kernel is big endian
(which people do do with i.MX systems IIRC).

> +	if (spi_imx->bpw_w == 1)
> +		spi_imx_buf_rx_u8(spi_imx);
> +	else if (spi_imx->bpw_w == 2)
> +		spi_imx_buf_rx_u16(spi_imx);

switch statement please.

> +	if (spi_imx->dynamic_burst) {
> +		spi_imx->count_index =
> +				spi_imx->count > MX51_ECSPI_CTRL_MAX_BURST ?
> +				spi_imx->count % MX51_ECSPI_CTRL_MAX_BURST :
> +				spi_imx->count % sizeof(u32);

Just write a normal if statement please, it's easier to read.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH 1/2] spi: imx: dynamic burst length adjust for PIO mode

[Jiada Wang]

Hello Mark

Sorry, somehow I missed your following comment

On 02/14/2017 10:20 AM, Mark Brown wrote:
> On Wed, Feb 08, 2017 at 03:20:27PM +0900, Jiada Wang wrote:
>
> This looks basically fine, a couple of fairly minor things here:
>
>> +	for (i = 0; i<  transfer->len / 4; i++) {
>> +		u8 temp;
>> +
>> +		temp = *(buf + i * 4);
>> +		*(buf + i * 4) = *(buf + i * 4 + 3);
>> +		*(buf + i * 4 + 3) = temp;
> Should this be using one of the cpu_to_ functions?  It's a bit unclear
> what the goal is here and if it'll work if the kernel is big endian
> (which people do do with i.MX systems IIRC).
indeed, with big endian kernel, this function won't work
I will replace the algo here with cpu_to_* in next version

Thanks,
Jiada

>> +	if (spi_imx->bpw_w == 1)
>> +		spi_imx_buf_rx_u8(spi_imx);
>> +	else if (spi_imx->bpw_w == 2)
>> +		spi_imx_buf_rx_u16(spi_imx);
> switch statement please.
>
>> +	if (spi_imx->dynamic_burst) {
>> +		spi_imx->count_index =
>> +				spi_imx->count>  MX51_ECSPI_CTRL_MAX_BURST ?
>> +				spi_imx->count % MX51_ECSPI_CTRL_MAX_BURST :
>> +				spi_imx->count % sizeof(u32);
> Just write a normal if statement please, it's easier to read.

--
To unsubscribe from this list: send the line "unsubscribe linux-spi" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/2] spi: imx: dynamic burst length adjust for PIO mode

[Jiada Wang]

Hello Mark

Sorry, somehow I missed your following comment

On 02/14/2017 10:20 AM, Mark Brown wrote:
> On Wed, Feb 08, 2017 at 03:20:27PM +0900, Jiada Wang wrote:
>
> This looks basically fine, a couple of fairly minor things here:
>
>> +	for (i = 0; i<  transfer->len / 4; i++) {
>> +		u8 temp;
>> +
>> +		temp = *(buf + i * 4);
>> +		*(buf + i * 4) = *(buf + i * 4 + 3);
>> +		*(buf + i * 4 + 3) = temp;
> Should this be using one of the cpu_to_ functions?  It's a bit unclear
> what the goal is here and if it'll work if the kernel is big endian
> (which people do do with i.MX systems IIRC).
indeed, with big endian kernel, this function won't work
I will replace the algo here with cpu_to_* in next version

Thanks,
Jiada

>> +	if (spi_imx->bpw_w == 1)
>> +		spi_imx_buf_rx_u8(spi_imx);
>> +	else if (spi_imx->bpw_w == 2)
>> +		spi_imx_buf_rx_u16(spi_imx);
> switch statement please.
>
>> +	if (spi_imx->dynamic_burst) {
>> +		spi_imx->count_index =
>> +				spi_imx->count>  MX51_ECSPI_CTRL_MAX_BURST ?
>> +				spi_imx->count % MX51_ECSPI_CTRL_MAX_BURST :
>> +				spi_imx->count % sizeof(u32);
> Just write a normal if statement please, it's easier to read.

[PATCH 2/2] spi: imx: dynamic burst length adjust for DMA mode

[Jiada Wang]

previously burst length (BURST_LENGTH) is always set to equal
to bits_per_word, causes a 10us gap between each word in
transfer, which significantly affects performance.

This patch uses 32 bits transfer to simulate lower bits transfer,
and adjusts burst length to reduce the number of gaps in DMA
transfer.

Signed-off-by: Jiada Wang <jiada_wang@mentor.com>
---
 drivers/spi/spi-imx.c | 154 ++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 130 insertions(+), 24 deletions(-)

diff --git a/drivers/spi/spi-imx.c b/drivers/spi/spi-imx.c
index 04b4ea8..68ff781 100644
--- a/drivers/spi/spi-imx.c
+++ b/drivers/spi/spi-imx.c
@@ -39,6 +39,8 @@
 #include <linux/of_device.h>
 #include <linux/of_gpio.h>
 
+#include <asm/cacheflush.h>
+
 #include <linux/platform_data/dma-imx.h>
 #include <linux/platform_data/spi-imx.h>
 
@@ -216,6 +218,7 @@ static bool spi_imx_can_dma(struct spi_master *master, struct spi_device *spi,
 {
 	struct spi_imx_data *spi_imx = spi_master_get_devdata(master);
 	unsigned int bpw, i;
+	u32 length, div;
 
 	if (!master->dma_rx)
 		return false;
@@ -232,8 +235,18 @@ static bool spi_imx_can_dma(struct spi_master *master, struct spi_device *spi,
 	if (bpw != 1 && bpw != 2 && bpw != 4)
 		return false;
 
+	length = transfer->len;
+
+	if (spi_imx->dynamic_burst) {
+		bpw = sizeof(u32);
+		length = transfer->len - transfer->len % sizeof(u32);
+		div = length / MX51_ECSPI_CTRL_MAX_BURST  + 1;
+		length = (length / div) - (length / div) % sizeof(u32);
+		spi_imx->count_index = transfer->len - length * div;
+	}
+
 	for (i = spi_imx_get_fifosize(spi_imx) / 2; i > 0; i--) {
-		if (!(transfer->len % (i * bpw)))
+		if (!(length % (i * bpw)))
 			break;
 	}
 
@@ -423,6 +436,7 @@ static int mx51_ecspi_config(struct spi_device *spi,
 	u32 ctrl = MX51_ECSPI_CTRL_ENABLE;
 	u32 clk = config->speed_hz, delay, reg;
 	u32 cfg = readl(spi_imx->base + MX51_ECSPI_CONFIG);
+	u32 div, length;
 
 	/*
 	 * The hardware seems to have a race condition when changing modes. The
@@ -441,9 +455,18 @@ static int mx51_ecspi_config(struct spi_device *spi,
 	ctrl |= MX51_ECSPI_CTRL_CS(spi->chip_select);
 
 	if (spi_imx->dynamic_burst) {
-		if (config->len > MX51_ECSPI_CTRL_MAX_BURST)
-			ctrl |= MX51_ECSPI_CTRL_BL_MASK;
-		else
+		if (config->len > MX51_ECSPI_CTRL_MAX_BURST) {
+			if (spi_imx->usedma) {
+				length = config->len -
+					 config->len % sizeof(u32);
+				div = length / MX51_ECSPI_CTRL_MAX_BURST  + 1;
+				length = (length / div) -
+					 (length / div) % sizeof(u32);
+				ctrl |= ((length * 8 - 1) <<
+					MX51_ECSPI_CTRL_BL_OFFSET);
+			} else
+				ctrl |= MX51_ECSPI_CTRL_BL_MASK;
+		} else
 			ctrl |= (((config->len - config->len % 4) * 8 - 1) <<
 				MX51_ECSPI_CTRL_BL_OFFSET);
 	} else
@@ -933,10 +956,16 @@ static int spi_imx_dma_configure(struct spi_master *master,
 		buswidth = DMA_SLAVE_BUSWIDTH_4_BYTES;
 		break;
 	case 2:
-		buswidth = DMA_SLAVE_BUSWIDTH_2_BYTES;
+		if (spi_imx->dynamic_burst)
+			buswidth = DMA_SLAVE_BUSWIDTH_4_BYTES;
+		else
+			buswidth = DMA_SLAVE_BUSWIDTH_2_BYTES;
 		break;
 	case 1:
-		buswidth = DMA_SLAVE_BUSWIDTH_1_BYTE;
+		if (spi_imx->dynamic_burst)
+			buswidth = DMA_SLAVE_BUSWIDTH_4_BYTES;
+		else
+			buswidth = DMA_SLAVE_BUSWIDTH_1_BYTE;
 		break;
 	default:
 		return -EINVAL;
@@ -1122,6 +1151,32 @@ static int spi_imx_calculate_timeout(struct spi_imx_data *spi_imx, int size)
 	return msecs_to_jiffies(2 * timeout * MSEC_PER_SEC);
 }
 
+static int spi_imx_pio_txrx(struct spi_imx_data *spi_imx)
+{
+	unsigned long transfer_timeout;
+	unsigned long timeout;
+
+	spi_imx->txfifo = 0;
+
+	reinit_completion(&spi_imx->xfer_done);
+
+	spi_imx_push(spi_imx);
+
+	spi_imx->devtype_data->intctrl(spi_imx, MXC_INT_TE);
+
+	transfer_timeout = spi_imx_calculate_timeout(spi_imx, spi_imx->count);
+
+	timeout = wait_for_completion_timeout(&spi_imx->xfer_done,
+					      transfer_timeout);
+	if (!timeout) {
+		dev_err(spi_imx->dev, "I/O Error in PIO\n");
+		spi_imx->devtype_data->reset(spi_imx);
+		return -ETIMEDOUT;
+	}
+
+	return 0;
+}
+
 static int spi_imx_dma_transfer(struct spi_imx_data *spi_imx,
 				struct spi_transfer *transfer)
 {
@@ -1130,6 +1185,20 @@ static int spi_imx_dma_transfer(struct spi_imx_data *spi_imx,
 	unsigned long timeout;
 	struct spi_master *master = spi_imx->bitbang.master;
 	struct sg_table *tx = &transfer->tx_sg, *rx = &transfer->rx_sg;
+	unsigned int old_nents = 0;
+	int ret;
+
+	spi_imx->count = transfer->len - spi_imx->count_index;
+	if (spi_imx->dynamic_burst && spi_imx->count_index) {
+		/* Cut RX data tail */
+		old_nents = rx->nents;
+		WARN_ON(sg_dma_len(&rx->sgl[rx->nents - 1]) <
+			spi_imx->count_index);
+		sg_dma_len(&rx->sgl[rx->nents - 1]) -=
+			spi_imx->count_index;
+		if (sg_dma_len(&rx->sgl[rx->nents - 1]) == 0)
+			--rx->nents;
+	}
 
 	/*
 	 * The TX DMA setup starts the transfer, so make sure RX is configured
@@ -1147,6 +1216,30 @@ static int spi_imx_dma_transfer(struct spi_imx_data *spi_imx,
 	reinit_completion(&spi_imx->dma_rx_completion);
 	dma_async_issue_pending(master->dma_rx);
 
+	if (spi_imx->dynamic_burst) {
+		dma_sync_sg_for_cpu(master->dma_tx->device->dev,
+				    tx->sgl, tx->nents, DMA_TO_DEVICE);
+		if (spi_imx->bpw_w == 1)
+			spi_imx_u32_swap_u8(transfer, (u8 *)transfer->tx_buf);
+		if (spi_imx->bpw_w == 2)
+			spi_imx_u32_swap_u16(transfer,
+				(u16 *)transfer->tx_buf);
+
+		if (spi_imx->count_index) {
+			/* Cut TX data tail */
+			old_nents = tx->nents;
+			WARN_ON(sg_dma_len(&tx->sgl[tx->nents - 1]) <
+				spi_imx->count_index);
+			sg_dma_len(&tx->sgl[tx->nents - 1]) -=
+				spi_imx->count_index;
+			if (sg_dma_len(&tx->sgl[tx->nents - 1]) == 0)
+				--tx->nents;
+		}
+
+		dma_sync_sg_for_device(master->dma_tx->device->dev,
+				       tx->sgl, tx->nents, DMA_TO_DEVICE);
+	}
+
 	desc_tx = dmaengine_prep_slave_sg(master->dma_tx,
 				tx->sgl, tx->nents, DMA_MEM_TO_DEV,
 				DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
@@ -1161,6 +1254,12 @@ static int spi_imx_dma_transfer(struct spi_imx_data *spi_imx,
 	reinit_completion(&spi_imx->dma_tx_completion);
 	dma_async_issue_pending(master->dma_tx);
 
+	if (spi_imx->dynamic_burst && spi_imx->count_index) {
+		spi_imx->tx_buf = transfer->tx_buf + spi_imx->count;
+		spi_imx->rx_buf = transfer->rx_buf + spi_imx->count;
+		spi_imx->count = spi_imx->count_index;
+	}
+
 	transfer_timeout = spi_imx_calculate_timeout(spi_imx, transfer->len);
 
 	/* Wait SDMA to finish the data transfer.*/
@@ -1182,6 +1281,27 @@ static int spi_imx_dma_transfer(struct spi_imx_data *spi_imx,
 		return -ETIMEDOUT;
 	}
 
+	if (spi_imx->dynamic_burst) {
+		spi_imx->dynamic_burst = 0;
+
+		if (spi_imx->count_index) {
+			ret = spi_imx_pio_txrx(spi_imx);
+			if (ret < 0)
+				return ret;
+		}
+
+		if (spi_imx->bpw_w == 1)
+			spi_imx_u32_swap_u8(transfer, (u8 *)transfer->rx_buf);
+		if (spi_imx->bpw_w == 2)
+			spi_imx_u32_swap_u16(transfer,
+					(u16 *)transfer->rx_buf);
+		dmac_flush_range(transfer->rx_buf,
+				transfer->rx_buf + transfer->len);
+		outer_flush_range(virt_to_phys(transfer->rx_buf),
+				virt_to_phys(transfer->rx_buf) +
+				transfer->len);
+	}
+
 	return transfer->len;
 }
 
@@ -1189,13 +1309,11 @@ static int spi_imx_pio_transfer(struct spi_device *spi,
 				struct spi_transfer *transfer)
 {
 	struct spi_imx_data *spi_imx = spi_master_get_devdata(spi->master);
-	unsigned long transfer_timeout;
-	unsigned long timeout;
+	int ret;
 
 	spi_imx->tx_buf = transfer->tx_buf;
 	spi_imx->rx_buf = transfer->rx_buf;
 	spi_imx->count = transfer->len;
-	spi_imx->txfifo = 0;
 
 	if (spi_imx->dynamic_burst) {
 		spi_imx->count_index =
@@ -1211,21 +1329,9 @@ static int spi_imx_pio_transfer(struct spi_device *spi,
 					(u16 *)transfer->tx_buf);
 	}
 
-	reinit_completion(&spi_imx->xfer_done);
-
-	spi_imx_push(spi_imx);
-
-	spi_imx->devtype_data->intctrl(spi_imx, MXC_INT_TE);
-
-	transfer_timeout = spi_imx_calculate_timeout(spi_imx, transfer->len);
-
-	timeout = wait_for_completion_timeout(&spi_imx->xfer_done,
-					      transfer_timeout);
-	if (!timeout) {
-		dev_err(&spi->dev, "I/O Error in PIO\n");
-		spi_imx->devtype_data->reset(spi_imx);
-		return -ETIMEDOUT;
-	}
+	ret = spi_imx_pio_txrx(spi_imx);
+	if (ret < 0)
+		return ret;
 
 	if (spi_imx->dynamic_burst) {
 		if (spi_imx->bpw_w == 1)
-- 
2.9.3

Re: [PATCH 2/2] spi: imx: dynamic burst length adjust for DMA mode

[kbuild test robot]

[-- Attachment #1: Type: text/plain, Size: 1565 bytes --]

Hi Jiada,

[auto build test ERROR on next-20170207]

url:    https://github.com/0day-ci/linux/commits/Jiada-Wang/improve-imx-spi-performance/20170208-142524
config: i386-allmodconfig (attached as .config)
compiler: gcc-6 (Debian 6.2.0-3) 6.2.0 20160901
reproduce:
        # save the attached .config to linux build tree
        make ARCH=i386 

All errors (new ones prefixed by >>):

   drivers/spi/spi-imx.c: In function 'spi_imx_dma_transfer':
>> drivers/spi/spi-imx.c:1298:3: error: implicit declaration of function 'dmac_flush_range' [-Werror=implicit-function-declaration]
      dmac_flush_range(transfer->rx_buf,
      ^~~~~~~~~~~~~~~~
>> drivers/spi/spi-imx.c:1300:3: error: implicit declaration of function 'outer_flush_range' [-Werror=implicit-function-declaration]
      outer_flush_range(virt_to_phys(transfer->rx_buf),
      ^~~~~~~~~~~~~~~~~
   cc1: some warnings being treated as errors

vim +/dmac_flush_range +1298 drivers/spi/spi-imx.c

  1292	
  1293			if (spi_imx->bpw_w == 1)
  1294				spi_imx_u32_swap_u8(transfer, (u8 *)transfer->rx_buf);
  1295			if (spi_imx->bpw_w == 2)
  1296				spi_imx_u32_swap_u16(transfer,
  1297						(u16 *)transfer->rx_buf);
> 1298			dmac_flush_range(transfer->rx_buf,
  1299					transfer->rx_buf + transfer->len);
> 1300			outer_flush_range(virt_to_phys(transfer->rx_buf),
  1301					virt_to_phys(transfer->rx_buf) +
  1302					transfer->len);
  1303		}

---
0-DAY kernel test infrastructure                Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all                   Intel Corporation

[-- Attachment #2: .config.gz --]
[-- Type: application/gzip, Size: 58520 bytes --]

Re: [PATCH 2/2] spi: imx: dynamic burst length adjust for DMA mode

[kbuild test robot]

[-- Attachment #1: Type: text/plain, Size: 780 bytes --]

Hi Jiada,

[auto build test ERROR on next-20170207]

url:    https://github.com/0day-ci/linux/commits/Jiada-Wang/improve-imx-spi-performance/20170208-142524
config: arm-arm5 (attached as .config)
compiler: arm-linux-gnueabi-gcc (Debian 6.1.1-9) 6.1.1 20160705
reproduce:
        wget https://git.kernel.org/cgit/linux/kernel/git/wfg/lkp-tests.git/plain/sbin/make.cross -O ~/bin/make.cross
        chmod +x ~/bin/make.cross
        # save the attached .config to linux build tree
        make.cross ARCH=arm 

All errors (new ones prefixed by >>):

>> ERROR: "arm926_dma_flush_range" [drivers/spi/spi-imx.ko] undefined!

---
0-DAY kernel test infrastructure                Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all                   Intel Corporation

[-- Attachment #2: .config.gz --]
[-- Type: application/gzip, Size: 28175 bytes --]