Skip to contentSkip to conversation

mtd: cqspi: Simplify indirect write code

PATCH New Marek Vasut 3 participants 6 messages May 3, 2016 – May 4, 2016 SPI

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/2 mtd: cqspi: Simplify indirect write code

The indirect write code is buggy pile of nastiness which fails horribly when the system runs fast enough to saturate the controller. The failure results in some pages (256B) not being written to the flash. This can be observed on systems which run with Dcache enabled and L2 cache enabled, like the Altera SoCFPGA.

This patch replaces the whole unmaintainable indirect write implementation with the one from upcoming Linux CQSPI driver, which went through multiple rounds of thorough review and testing. While this makes the patch look terrifying and violates all best-practices of software development, all the patch does is it plucks out duplicate ad-hoc code distributed across the driver and replaces it with more compact code doing exactly the same thing.

Cc Anatolij Gustschin <[email protected]>
Cc Chin Liang See <[email protected]>
Cc Dinh Nguyen <[email protected]>
Cc Jagan Teki <[email protected]>
Cc Pavel Machek <[email protected]>
Cc Stefan Roese <[email protected]>
Cc Vignesh R <[email protected]>

---
V2: Handle non-4-byte aligned transfers

---
 drivers/spi/cadence_qspi_apb.c | 126 ++++++++++-------------------------------
 1 file changed, 30 insertions(+), 96 deletions(-)
drivers/spi/cadence_qspi_apb.c +30 -96
index 7786dd6..5e84144 100644
@@ -28,6 +28,7 @@
 #include <common.h>
 #include <asm/io.h>
 #include <asm/errno.h>
+#include <wait_bit.h>
 #include "cadence_qspi.h"
 
 #define CQSPI_REG_POLL_US			(1) /* 1us */
@@ -214,32 +215,6 @@ static void cadence_qspi_apb_read_fifo_data(void *dest,
 	return;
 }
 
-static void cadence_qspi_apb_write_fifo_data(const void *dest_ahb_addr,
-	const void *src, unsigned int bytes)
-{
-	unsigned int temp = 0;
-	int i;
-	int remaining = bytes;
-	unsigned int *dest_ptr = (unsigned int *)dest_ahb_addr;
-	unsigned int *src_ptr = (unsigned int *)src;
-
-	while (remaining >= CQSPI_FIFO_WIDTH) {
-		for (i = CQSPI_FIFO_WIDTH/sizeof(src_ptr) - 1; i >= 0; i--)
-			writel(*(src_ptr+i), dest_ptr+i);
-		src_ptr += CQSPI_FIFO_WIDTH/sizeof(src_ptr);
-		remaining -= CQSPI_FIFO_WIDTH;
-	}
-	if (remaining) {
-		/* dangling bytes */
-		i = remaining/sizeof(dest_ptr);
-		memcpy(&temp, src_ptr+i, remaining % sizeof(dest_ptr));
-		writel(temp, dest_ptr+i);
-		for (--i; i >= 0; i--)
-			writel(*(src_ptr+i), dest_ptr+i);
-	}
-	return;
-}
-
 /* Read from SRAM FIFO with polling SRAM fill level. */
 static int qspi_read_sram_fifo_poll(const void *reg_base, void *dest_addr,
 			const void *src_addr,  unsigned int num_bytes)
@@ -276,44 +251,6 @@ static int qspi_read_sram_fifo_poll(const void *reg_base, void *dest_addr,
 	return 0;
 }
 
-/* Write to SRAM FIFO with polling SRAM fill level. */
-static int qpsi_write_sram_fifo_push(struct cadence_spi_platdata *plat,
-				const void *src_addr, unsigned int num_bytes)
-{
-	const void *reg_base = plat->regbase;
-	void *dest_addr = plat->ahbbase;
-	unsigned int retry = CQSPI_REG_RETRY;
-	unsigned int sram_level;
-	unsigned int wr_bytes;
-	unsigned char *src = (unsigned char *)src_addr;
-	int remaining = num_bytes;
-	unsigned int page_size = plat->page_size;
-	unsigned int sram_threshold_words = CQSPI_REG_SRAM_THRESHOLD_WORDS;
-
-	while (remaining > 0) {
-		retry = CQSPI_REG_RETRY;
-		while (retry--) {
-			sram_level = CQSPI_GET_WR_SRAM_LEVEL(reg_base);
-			if (sram_level <= sram_threshold_words)
-				break;
-		}
-		if (!retry) {
-			printf("QSPI: SRAM fill level (0x%08x) not hit lower expected level (0x%08x)",
-			       sram_level, sram_threshold_words);
-			return -1;
-		}
-		/* Write a page or remaining bytes. */
-		wr_bytes = (remaining > page_size) ?
-					page_size : remaining;
-
-		cadence_qspi_apb_write_fifo_data(dest_addr, src, wr_bytes);
-		src += wr_bytes;
-		remaining -= wr_bytes;
-	}
-
-	return 0;
-}
-
 void cadence_qspi_apb_controller_enable(void *reg_base)
 {
 	unsigned int reg;
@@ -810,48 +747,45 @@ int cadence_qspi_apb_indirect_write_setup(struct cadence_spi_platdata *plat,
 }
 
 int cadence_qspi_apb_indirect_write_execute(struct cadence_spi_platdata *plat,
-	unsigned int txlen, const u8 *txbuf)
+	unsigned int n_tx, const u8 *txbuf)
 {
-	unsigned int reg = 0;
-	unsigned int retry;
+	unsigned int page_size = plat->page_size;
+	unsigned int remaining = n_tx;
+	unsigned int write_bytes;
+	int ret;
 
 	/* Configure the indirect read transfer bytes */
-	writel(txlen, plat->regbase + CQSPI_REG_INDIRECTWRBYTES);
+	writel(n_tx, plat->regbase + CQSPI_REG_INDIRECTWRBYTES);
 
 	/* Start the indirect write transfer */
 	writel(CQSPI_REG_INDIRECTWR_START_MASK,
 	       plat->regbase + CQSPI_REG_INDIRECTWR);
 
-	if (qpsi_write_sram_fifo_push(plat, (const void *)txbuf, txlen))
-		goto failwr;
-
-	/* Wait until last write is completed (FIFO empty) */
-	retry = CQSPI_REG_RETRY;
-	while (retry--) {
-		reg = CQSPI_GET_WR_SRAM_LEVEL(plat->regbase);
-		if (reg == 0)
-			break;
-
-		udelay(1);
-	}
-
-	if (reg != 0) {
-		printf("QSPI: timeout for indirect write\n");
-		goto failwr;
-	}
+	while (remaining > 0) {
+		write_bytes = remaining > page_size ? page_size : remaining;
+		/* Handle non-4-byte aligned access to avoid data abort. */
+		if (((uintptr_t)txbuf % 4) || (write_bytes % 4))
+			writesb(plat->ahbbase, txbuf, write_bytes);
+		else
+			writesl(plat->ahbbase, txbuf, write_bytes >> 2);
+
+		ret = wait_for_bit("QSPI", plat->regbase + CQSPI_REG_SDRAMLEVEL,
+				   CQSPI_REG_SDRAMLEVEL_WR_MASK <<
+				   CQSPI_REG_SDRAMLEVEL_WR_LSB, 0, 10, 0);
+		if (ret) {
+			printf("Indirect write timed out (%i)\n", ret);
+			goto failwr;
+		}
 
-	/* Check flash indirect controller status */
-	retry = CQSPI_REG_RETRY;
-	while (retry--) {
-		reg = readl(plat->regbase + CQSPI_REG_INDIRECTWR);
-		if (reg & CQSPI_REG_INDIRECTWR_DONE_MASK)
-			break;
-		udelay(1);
+		txbuf += write_bytes;
+		remaining -= write_bytes;
 	}
 
-	if (!(reg & CQSPI_REG_INDIRECTWR_DONE_MASK)) {
-		printf("QSPI: indirect completion status error with reg 0x%08x\n",
-		       reg);
+	/* Check indirect done status */
+	ret = wait_for_bit("QSPI", plat->regbase + CQSPI_REG_INDIRECTWR,
+			   CQSPI_REG_INDIRECTWR_DONE_MASK, 1, 10, 0);
+	if (ret) {
+		printf("Indirect write completion error (%i)\n", ret);
 		goto failwr;
 	}
 
@@ -864,7 +798,7 @@ failwr:
 	/* Cancel the indirect write */
 	writel(CQSPI_REG_INDIRECTWR_CANCEL_MASK,
 	       plat->regbase + CQSPI_REG_INDIRECTWR);
-	return -1;
+	return ret;
 }
 
 void cadence_qspi_apb_enter_xip(void *reg_base, char xip_dummy)
Signature · 1 line
2.7.0
#1 lore raw

PATCH v2 2/2 mtd: cqspi: Simplify indirect read code

The indirect read code is a pile of nastiness. This patch replaces the whole unmaintainable indirect read implementation with the one from upcoming Linux CQSPI driver, which went through multiple rounds of thorough review and testing. All the patch does is it plucks out duplicate ad-hoc code distributed across the driver and replaces it with more compact code doing exactly the same thing. There is no speed change of the read operation.

Cc Anatolij Gustschin <[email protected]>
Cc Chin Liang See <[email protected]>
Cc Dinh Nguyen <[email protected]>
Cc Jagan Teki <[email protected]>
Cc Pavel Machek <[email protected]>
Cc Stefan Roese <[email protected]>
Cc Vignesh R <[email protected]>

---
V2: Handle non-4-byte aligned transfers

---
 drivers/spi/cadence_qspi_apb.c | 128 ++++++++++++++++++-----------------------
 1 file changed, 57 insertions(+), 71 deletions(-)
drivers/spi/cadence_qspi_apb.c +57 -71
index 5e84144..a71531d 100644
@@ -193,64 +193,6 @@ static unsigned int cadence_qspi_apb_cmd2addr(const unsigned char *addr_buf,
 	return addr;
 }
 
-static void cadence_qspi_apb_read_fifo_data(void *dest,
-	const void *src_ahb_addr, unsigned int bytes)
-{
-	unsigned int temp;
-	int remaining = bytes;
-	unsigned int *dest_ptr = (unsigned int *)dest;
-	unsigned int *src_ptr = (unsigned int *)src_ahb_addr;
-
-	while (remaining >= sizeof(dest_ptr)) {
-		*dest_ptr = readl(src_ptr);
-		remaining -= sizeof(src_ptr);
-		dest_ptr++;
-	}
-	if (remaining) {
-		/* dangling bytes */
-		temp = readl(src_ptr);
-		memcpy(dest_ptr, &temp, remaining);
-	}
-
-	return;
-}
-
-/* Read from SRAM FIFO with polling SRAM fill level. */
-static int qspi_read_sram_fifo_poll(const void *reg_base, void *dest_addr,
-			const void *src_addr,  unsigned int num_bytes)
-{
-	unsigned int remaining = num_bytes;
-	unsigned int retry;
-	unsigned int sram_level = 0;
-	unsigned char *dest = (unsigned char *)dest_addr;
-
-	while (remaining > 0) {
-		retry = CQSPI_REG_RETRY;
-		while (retry--) {
-			sram_level = CQSPI_GET_RD_SRAM_LEVEL(reg_base);
-			if (sram_level)
-				break;
-			udelay(1);
-		}
-
-		if (!retry) {
-			printf("QSPI: No receive data after polling for %d times\n",
-			       CQSPI_REG_RETRY);
-			return -1;
-		}
-
-		sram_level *= CQSPI_FIFO_WIDTH;
-		sram_level = sram_level > remaining ? remaining : sram_level;
-
-		/* Read data from FIFO. */
-		cadence_qspi_apb_read_fifo_data(dest, src_addr, sram_level);
-		dest += sram_level;
-		remaining -= sram_level;
-		udelay(1);
-	}
-	return 0;
-}
-
 void cadence_qspi_apb_controller_enable(void *reg_base)
 {
 	unsigned int reg;
@@ -679,40 +621,84 @@ int cadence_qspi_apb_indirect_read_setup(struct cadence_spi_platdata *plat,
 	return 0;
 }
 
+static u32 cadence_qspi_get_rd_sram_level(struct cadence_spi_platdata *plat)
+{
+	u32 reg = readl(plat->regbase + CQSPI_REG_SDRAMLEVEL);
+	reg >>= CQSPI_REG_SDRAMLEVEL_RD_LSB;
+	return reg & CQSPI_REG_SDRAMLEVEL_RD_MASK;
+}
+
+static int cadence_qspi_wait_for_data(struct cadence_spi_platdata *plat)
+{
+	unsigned int timeout = 10000;
+	u32 reg;
+
+	while (timeout--) {
+		reg = cadence_qspi_get_rd_sram_level(plat);
+		if (reg)
+			return reg;
+		udelay(1);
+	}
+
+	return -ETIMEDOUT;
+}
+
 int cadence_qspi_apb_indirect_read_execute(struct cadence_spi_platdata *plat,
-	unsigned int rxlen, u8 *rxbuf)
+	unsigned int n_rx, u8 *rxbuf)
 {
-	unsigned int reg;
+	unsigned int remaining = n_rx;
+	unsigned int bytes_to_read = 0;
+	int ret;
 
-	writel(rxlen, plat->regbase + CQSPI_REG_INDIRECTRDBYTES);
+	writel(n_rx, plat->regbase + CQSPI_REG_INDIRECTRDBYTES);
 
 	/* Start the indirect read transfer */
 	writel(CQSPI_REG_INDIRECTRD_START_MASK,
 	       plat->regbase + CQSPI_REG_INDIRECTRD);
 
-	if (qspi_read_sram_fifo_poll(plat->regbase, (void *)rxbuf,
-				     (const void *)plat->ahbbase, rxlen))
-		goto failrd;
+	while (remaining > 0) {
+		ret = cadence_qspi_wait_for_data(plat);
+		if (ret < 0) {
+			printf("Indirect write timed out (%i)\n", ret);
+			goto failrd;
+		}
+
+		bytes_to_read = ret;
+
+		while (bytes_to_read != 0) {
+			bytes_to_read *= CQSPI_FIFO_WIDTH;
+			bytes_to_read = bytes_to_read > remaining ?
+					remaining : bytes_to_read;
+			/* Handle non-4-byte aligned access to avoid data abort. */
+			if (((uintptr_t)rxbuf % 4) || (bytes_to_read % 4))
+				readsb(plat->ahbbase, rxbuf, bytes_to_read);
+			else
+				readsl(plat->ahbbase, rxbuf, bytes_to_read >> 2);
+			rxbuf += bytes_to_read;
+			remaining -= bytes_to_read;
+			bytes_to_read = cadence_qspi_get_rd_sram_level(plat);
+		}
+	}
 
-	/* Check flash indirect controller */
-	reg = readl(plat->regbase + CQSPI_REG_INDIRECTRD);
-	if (!(reg & CQSPI_REG_INDIRECTRD_DONE_MASK)) {
-		reg = readl(plat->regbase + CQSPI_REG_INDIRECTRD);
-		printf("QSPI: indirect completion status error with reg 0x%08x\n",
-		       reg);
+	/* Check indirect done status */
+	ret = wait_for_bit("QSPI", plat->regbase + CQSPI_REG_INDIRECTRD,
+			   CQSPI_REG_INDIRECTRD_DONE_MASK, 1, 10, 0);
+	if (ret) {
+		printf("Indirect read completion error (%i)\n", ret);
 		goto failrd;
 	}
 
 	/* Clear indirect completion status */
 	writel(CQSPI_REG_INDIRECTRD_DONE_MASK,
 	       plat->regbase + CQSPI_REG_INDIRECTRD);
+
 	return 0;
 
 failrd:
 	/* Cancel the indirect read */
 	writel(CQSPI_REG_INDIRECTRD_CANCEL_MASK,
 	       plat->regbase + CQSPI_REG_INDIRECTRD);
-	return -1;
+	return ret;
 }
 
 /* Opcode + Address (3/4 bytes) */
Signature · 1 line
2.7.0
#2 lore raw

PATCH v2 2/2 mtd: cqspi: Simplify indirect read code

Quoted 16 lines · Marek Vasut

The indirect read code is a pile of nastiness. This patch replaces the whole unmaintainable indirect read implementation with the one from upcoming Linux CQSPI driver, which went through multiple rounds of thorough review and testing. All the patch does is it plucks out duplicate ad-hoc code distributed across the driver and replaces it with more compact code doing exactly the same thing. There is no speed change of the read operation.

Cc Anatolij Gustschin <[email protected]>
Cc Chin Liang See <[email protected]>
Cc Dinh Nguyen <[email protected]>
Cc Jagan Teki <[email protected]>
Cc Pavel Machek <[email protected]>
Cc Stefan Roese <[email protected]>
Cc Vignesh R <[email protected]>
Quoted 2 lines

---
V2: Handle non-4-byte aligned transfers

Thanks,
Stefan

#3 lore raw

PATCH v2 1/2 mtd: cqspi: Simplify indirect write code

In reply to Marek Vasut
Quoted 22 lines · Marek Vasut

The indirect write code is buggy pile of nastiness which fails horribly when the system runs fast enough to saturate the controller. The failure results in some pages (256B) not being written to the flash. This can be observed on systems which run with Dcache enabled and L2 cache enabled, like the Altera SoCFPGA.

This patch replaces the whole unmaintainable indirect write implementation with the one from upcoming Linux CQSPI driver, which went through multiple rounds of thorough review and testing. While this makes the patch look terrifying and violates all best-practices of software development, all the patch does is it plucks out duplicate ad-hoc code distributed across the driver and replaces it with more compact code doing exactly the same thing.

Cc Anatolij Gustschin <[email protected]>
Cc Chin Liang See <[email protected]>
Cc Dinh Nguyen <[email protected]>
Cc Jagan Teki <[email protected]>
Cc Pavel Machek <[email protected]>
Cc Stefan Roese <[email protected]>
Cc Vignesh R <[email protected]>
Quoted 2 lines

---
V2: Handle non-4-byte aligned transfers

Thanks,
Stefan

#4 lore raw

PATCH v2 1/2 mtd: cqspi: Simplify indirect write code

In reply to Marek Vasut

Hi!

Quoted 6 lines

The indirect write code is buggy pile of nastiness which fails horribly when the system runs fast enough to saturate the controller. The failure results in some pages (256B) not being written to the flash. This can be observed on systems which run with Dcache enabled and L2 cache enabled, like the Altera SoCFPGA.

Quoted 7 lines

This patch replaces the whole unmaintainable indirect write implementation with the one from upcoming Linux CQSPI driver, which went through multiple rounds of thorough review and testing. While this makes the patch look terrifying and violates all best-practices of software development, all the patch does is it plucks out duplicate ad-hoc code distributed across the driver and replaces it with more compact code doing exactly the same thing.

Ok, sorry, I still don't understand the changelog.

First, it describes the bug with L2 cache enabled, but then it says that "all the patch does .. doing exactly the same thing".

So I assume it does not do the same thing, but replaces duplicated code in u-boot with working code from Linux?

Thanks for doing this,
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#5 lore raw

PATCH v2 1/2 mtd: cqspi: Simplify indirect write code

Quoted 16 lines · Pavel Machek

Hi!

Quoted 6 lines

The indirect write code is buggy pile of nastiness which fails horribly when the system runs fast enough to saturate the controller. The failure results in some pages (256B) not being written to the flash. This can be observed on systems which run with Dcache enabled and L2 cache enabled, like the Altera SoCFPGA.

Quoted 7 lines

This patch replaces the whole unmaintainable indirect write implementation with the one from upcoming Linux CQSPI driver, which went through multiple rounds of thorough review and testing. While this makes the patch look terrifying and violates all best-practices of software development, all the patch does is it plucks out duplicate ad-hoc code distributed across the driver and replaces it with more compact code doing exactly the same thing.

Quoted 7 lines

Ok, sorry, I still don't understand the changelog.

First, it describes the bug with L2 cache enabled, but then it says that "all the patch does .. doing exactly the same thing".

So I assume it does not do the same thing, but replaces duplicated code in u-boot with working code from Linux?

Ah right, the linux code also does FIFO level checking, so it doesn't overflow during the writes.

Quoted 4 lines

Thanks for doing this,
Pavel

Signature · 2 lines
Best regards,
Marek Vasut