ar71xx: improve SPI flash read/write performance

mtd_speedtest results:

                        page write speed
                    old         new       delta
DB120            209 KiB/s   226 KiB/s     +8.13%
TL-WR1043ND v1   122 KiB/s   148 KiB/s    +21.31%
TL-WR703N v1     153 KiB/s   194 KiB/s    +26.80%
TL-MR3220 v1     130 KiB/s   156 KiB/s    +20.00%
TL-WR2543ND v1   158 KiB/s   202 KiB/s    +27.85%
TL-WR741ND v2    122 KiB/s   152 KiB/s    +24.59%
ALFA AP96        229 KiB/s   260 KiB/s    +13.54%
WNDR3700         202 KiB/s   223 KiB/s    +10.40%

                         page read speed
                    old        new       delta
DB120            691 KiB/s   929 KiB/s    +34.44%
TL-WR1043ND v1   372 KiB/s   754 KiB/s   +102.69%
TL-WR703N v1     375 KiB/s   745 KiB/s    +98.67%
TL-MR3220 v1     372 KiB/s   752 KiB/s   +102.15%
TL-WR2543ND v1   307 KiB/s   564 KiB/s    +83.71%
TL-WR741ND v2    315 KiB/s   525 KiB/s    +66.67%
ALFA AP96        515 KiB/s   702 KiB/s    +36.31%
WNDR3700         515 KiB/s   697 KiB/s    +35.34%

git-svn-id: svn://svn.openwrt.org/openwrt/trunk@31117 3c298f89-4303-0410-b956-a3cf2f4a3e73

1 files changed, 90 insertions, 10 deletions

diff --git a/target/linux/ar71xx/patches-3.2/201-spi-ath79-add-delay-between-SCK-changes.patch b/target/linux/ar71xx/patches-3.2/201-spi-ath79-add-delay-between-SCK-changes.patch
index c2681f9a5d..9294133d45 100644
--- a/target/linux/ar71xx/patches-3.2/201-spi-ath79-add-delay-between-SCK-changes.patch
+++ b/target/linux/ar71xx/patches-3.2/201-spi-ath79-add-delay-between-SCK-changes.patch
@@ -1,4 +1,4 @@
-From 48b7e765e6e097d20d809fadd17a4355d26ad6d5 Mon Sep 17 00:00:00 2001
+From 8e948c035dd7983eccc3a889f2497e64044f3a31 Mon Sep 17 00:00:00 2001
 From: Gabor Juhos <juhosg@openwrt.org>
 Date: Wed, 11 Jan 2012 20:06:35 +0100
 Subject: [PATCH 1/7] spi/ath79: add delay between SCK changes
@@ -12,31 +12,111 @@ well.
 
 Signed-off-by: Gabor Juhos <juhosg@openwrt.org>
 ---
- drivers/spi/spi-ath79.c |    8 ++++++++
- 1 files changed, 8 insertions(+), 0 deletions(-)
+ drivers/spi/spi-ath79.c |   44 +++++++++++++++++++++++++++++++++++++++++++-
+ 1 files changed, 43 insertions(+), 1 deletions(-)
 
 --- a/drivers/spi/spi-ath79.c
 +++ b/drivers/spi/spi-ath79.c
-@@ -52,6 +52,12 @@ static inline struct ath79_spi *ath79_sp
+@@ -24,17 +24,24 @@
+ #include <linux/spi/spi_bitbang.h>
+ #include <linux/bitops.h>
+ #include <linux/gpio.h>
++#include <linux/clk.h>
++#include <linux/err.h>
+ 
+ #include <asm/mach-ath79/ar71xx_regs.h>
+ #include <asm/mach-ath79/ath79_spi_platform.h>
+ 
+ #define DRV_NAME	"ath79-spi"
+ 
++#define ATH79_SPI_RRW_DELAY_FACTOR	12000
++#define MHZ				(1000 * 1000)
++
+ struct ath79_spi {
+ 	struct spi_bitbang	bitbang;
+ 	u32			ioc_base;
+ 	u32			reg_ctrl;
+ 	void __iomem		*base;
++	struct clk		*clk;
++	unsigned		rrw_delay;
+ };
+ 
+ static inline u32 ath79_spi_rr(struct ath79_spi *sp, unsigned reg)
+@@ -52,6 +59,12 @@ static inline struct ath79_spi *ath79_sp
  	return spi_master_get_devdata(spi->master);
  }
  
-+static inline void ath79_spi_delay(unsigned nsecs)
++static inline void ath79_spi_delay(struct ath79_spi *sp, unsigned nsecs)
 +{
-+	if (nsecs)
-+		ndelay(nsecs);
++	if (nsecs > sp->rrw_delay)
++		ndelay(nsecs - sp->rrw_delay);
 +}
 +
  static void ath79_spi_chipselect(struct spi_device *spi, int is_active)
  {
  	struct ath79_spi *sp = ath79_spidev_to_sp(spi);
-@@ -184,7 +190,9 @@ static u32 ath79_spi_txrx_mode0(struct s
+@@ -184,7 +197,9 @@ static u32 ath79_spi_txrx_mode0(struct s
  
  		/* setup MSB (to slave) on trailing edge */
  		ath79_spi_wr(sp, AR71XX_SPI_REG_IOC, out);
-+		ath79_spi_delay(nsecs);
++		ath79_spi_delay(sp, nsecs);
  		ath79_spi_wr(sp, AR71XX_SPI_REG_IOC, out | AR71XX_SPI_IOC_CLK);
-+		ath79_spi_delay(nsecs);
++		ath79_spi_delay(sp, nsecs);
  
  		word <<= 1;
  	}
+@@ -198,6 +213,7 @@ static __devinit int ath79_spi_probe(str
+ 	struct ath79_spi *sp;
+ 	struct ath79_spi_platform_data *pdata;
+ 	struct resource	*r;
++	unsigned long rate;
+ 	int ret;
+ 
+ 	master = spi_alloc_master(&pdev->dev, sizeof(*sp));
+@@ -239,12 +255,36 @@ static __devinit int ath79_spi_probe(str
+ 		goto err_put_master;
+ 	}
+ 
++	sp->clk = clk_get(&pdev->dev, "ahb");
++	if (IS_ERR(sp->clk)) {
++		ret = PTR_ERR(sp->clk);
++		goto err_unmap;
++	}
++
++	ret = clk_enable(sp->clk);
++	if (ret)
++		goto err_clk_put;
++
++	rate = DIV_ROUND_UP(clk_get_rate(sp->clk), MHZ);
++	if (!rate) {
++		ret = -EINVAL;
++		goto err_clk_disable;
++	}
++
++	sp->rrw_delay = ATH79_SPI_RRW_DELAY_FACTOR / rate;
++	dev_dbg(&pdev->dev, "register read/write delay is %u nsecs\n",
++		sp->rrw_delay);
++
+ 	ret = spi_bitbang_start(&sp->bitbang);
+ 	if (ret)
+-		goto err_unmap;
++		goto err_clk_disable;
+ 
+ 	return 0;
+ 
++err_clk_disable:
++	clk_disable(sp->clk);
++err_clk_put:
++	clk_put(sp->clk);
+ err_unmap:
+ 	iounmap(sp->base);
+ err_put_master:
+@@ -259,6 +299,8 @@ static __devexit int ath79_spi_remove(st
+ 	struct ath79_spi *sp = platform_get_drvdata(pdev);
+ 
+ 	spi_bitbang_stop(&sp->bitbang);
++	clk_disable(sp->clk);
++	clk_put(sp->clk);
+ 	iounmap(sp->base);
+ 	platform_set_drvdata(pdev, NULL);
+ 	spi_master_put(sp->bitbang.master);