Patchwork [3/6] OMAP NAND: configurable fifo threshold to gain the throughput

login
register
mail settings
Submitter Sukumar Ghorai
Date April 16, 2010, 11:35 a.m.
Message ID <1271417714-4408-1-git-send-email-s-ghorai@ti.com>
Download mbox | patch
Permalink /patch/50319/
State New
Headers show

Comments

Sukumar Ghorai - April 16, 2010, 11:35 a.m.
Configure the FIFO THREASHOLD value to 50% (32 bytes) to keep busy both
  filling and to drain out of FIFO at reading and writing.

Signed-off-by: Sukumar Ghorai <s-ghorai@ti.com>
Signed-off-by: Vimal Singh <vimalsingh@ti.com>
---
 arch/arm/mach-omap2/gpmc.c             |    9 ++++++---
 arch/arm/plat-omap/include/plat/gpmc.h |    7 +++++--
 drivers/mtd/nand/omap2.c               |   25 +++++++++++++------------
 3 files changed, 24 insertions(+), 17 deletions(-)
vimal singh - April 16, 2010, 12:45 p.m.
Hi Ghorai,

On Fri, Apr 16, 2010 at 5:05 PM, Sukumar Ghorai <s-ghorai@ti.com> wrote:
[...]
> -       ret = gpmc_prefetch_enable(info->gpmc_cs, 0x0, len, 0x0);
> +       ret = gpmc_prefetch_enable(info->gpmc_cs, 0x40, 0x0, len, 0x0);

Use macros here too then.

>        if (ret) {
>                /* PFPW engine is busy, use cpu copy method */
>                if (info->nand.options & NAND_BUSWIDTH_16)
> @@ -354,7 +354,7 @@ static void omap_write_buf_pref(struct mtd_info *mtd,
>        }
>
>        /*  configure and start prefetch transfer */
> -       ret = gpmc_prefetch_enable(info->gpmc_cs, 0x0, len, 0x1);
> +       ret = gpmc_prefetch_enable(info->gpmc_cs, 0x40, 0x0, len, 0x1);

here too

>        if (ret) {
>                /* PFPW engine is busy, use cpu copy method */
>                if (info->nand.options & NAND_BUSWIDTH_16)
> @@ -405,10 +405,11 @@ static inline int omap_nand_dma_transfer(struct mtd_info *mtd, void *addr,
>        dma_addr_t dma_addr;
>        int ret;
>
> -       /* The fifo depth is 64 bytes. We have a sync at each frame and frame
> -        * length is 64 bytes.
> +       /* The fifo depth is 64 bytes max.
> +        * But configure the FIFO-threahold to 32 to get a sync at each frame
> +        * and frame length is 32 bytes.
>         */
> -       int buf_len = len >> 6;
> +       int buf_len = len >> 5;
>
>        if (addr >= high_memory) {
>                struct page *p1;
> @@ -447,7 +448,7 @@ static inline int omap_nand_dma_transfer(struct mtd_info *mtd, void *addr,
>                                        OMAP24XX_DMA_GPMC, OMAP_DMA_SRC_SYNC);
>        }
>        /*  configure and start prefetch transfer */
> -       ret = gpmc_prefetch_enable(info->gpmc_cs, 0x1, len, is_write);
> +       ret = gpmc_prefetch_enable(info->gpmc_cs, 0x20, 0x1, len, is_write);
>        if (ret)
>                /* PFPW engine is busy, use cpu copy methode */
>                goto out_copy;
> @@ -524,6 +525,7 @@ static void omap_write_buf_dma_pref(struct mtd_info *mtd,
>  static irqreturn_t omap_nand_irq(int this_irq, void *dev)
>  {
>        struct omap_nand_info *info = (struct omap_nand_info *) dev;
> +       u32 *p = (u32 *) info->buf;
>        u32 irq_enb = 0, pref_status = 0, bytes = 0;
>        u32 irq_stats = __raw_readl(info->gpmc_baseaddr + GPMC_IRQSTATUS);
>        u32 pref_config = __raw_readl(info->gpmc_baseaddr +
> @@ -533,14 +535,11 @@ static irqreturn_t omap_nand_irq(int this_irq, void *dev)
>                if (irq_stats & 0x2)
>                        goto done;
>
> -               u32 *p = (u32 *) info->buf;
> -               pref_status = gpmc_prefetch_status();
> -               bytes = ((pref_status >> 24) & 0x7F);
> +               bytes = PREFETCH_FIFOTHRESHOLD_WRITE;

By this, you are not really keeping prefetch the busyest.
you are filling only 'PREFETCH_FIFOTHRESHOLD_WRITE' bytes, while there
could be more free spaces.
Previous way of doing it was more efficent.

>                iowrite32_rep(info->nand_pref_fifo_add, p, bytes >> 2);
>                info->buf = info->buf + bytes;
>
>        } else {
> -               u32 *p = (u32 *) info->buf;
>                pref_status = gpmc_prefetch_status();
>                bytes = ((pref_status >> 24) & 0x7F);
>                ioread32_rep(info->nand_pref_fifo_add, p, bytes >> 2);
> @@ -586,7 +585,8 @@ static void omap_read_buf_irq_pref(struct mtd_info *mtd, u_char *buf, int len)
>        init_completion(&info->comp);
>
>        /*  configure and start prefetch transfer */
> -       ret = gpmc_prefetch_enable(info->gpmc_cs, 0x0, len, 0x0);
> +       ret = gpmc_prefetch_enable(info->gpmc_cs,
> +                               PREFETCH_FIFOTHRESHOLD_READ, 0x0, len, 0x0);
>        if (ret)
>                /* PFPW engine is busy, use cpu copy methode */
>                goto out_copy;
> @@ -630,7 +630,8 @@ static void omap_write_buf_irq_pref(struct mtd_info *mtd,
>        init_completion(&info->comp);
>
>        /*  configure and start prefetch transfer */
> -       ret = gpmc_prefetch_enable(info->gpmc_cs, 0x0, len, 0x1);
> +       ret = gpmc_prefetch_enable(info->gpmc_cs,
> +                               PREFETCH_FIFOTHRESHOLD_WRITE, 0x0, len, 0x1);

In case of write, in my experiments, fifo thresholed '24' was the best
compromise for throughput and cpu load.


Regards,
Vimal

Patch

diff --git a/arch/arm/mach-omap2/gpmc.c b/arch/arm/mach-omap2/gpmc.c
index 9c77af0..1380886 100644
--- a/arch/arm/mach-omap2/gpmc.c
+++ b/arch/arm/mach-omap2/gpmc.c
@@ -391,12 +391,15 @@  EXPORT_SYMBOL(gpmc_cs_free);
  * @u32_count: number of bytes to be transferred
  * @is_write: prefetch read(0) or write post(1) mode
  */
-int gpmc_prefetch_enable(int cs, int dma_mode,
+int gpmc_prefetch_enable(int cs, int fifo_th, int dma_mode,
 				unsigned int u32_count, int is_write)
 {
 	uint32_t prefetch_config1;
 
-	if (!(gpmc_read_reg(GPMC_PREFETCH_CONTROL))) {
+	if (fifo_th > PREFETCH_FIFOTHRESHOLD_MAX) {
+		printk(KERN_ERR "PREFETCH Fifo Threshold is not supported\n");
+		return -1;
+	} else if (!(gpmc_read_reg(GPMC_PREFETCH_CONTROL))) {
 		/* Set the amount of bytes to be prefetched */
 		gpmc_write_reg(GPMC_PREFETCH_CONFIG2, u32_count);
 
@@ -404,7 +407,7 @@  int gpmc_prefetch_enable(int cs, int dma_mode,
 		 * enable the engine. Set which cs is has requested for.
 		 */
 		prefetch_config1 = ((cs << CS_NUM_SHIFT) |
-					PREFETCH_FIFOTHRESHOLD |
+					PREFETCH_FIFOTHRESHOLD(fifo_th) |
 					ENABLE_PREFETCH |
 					(dma_mode << DMA_MPU_MODE) |
 					(0x1 & is_write));
diff --git a/arch/arm/plat-omap/include/plat/gpmc.h b/arch/arm/plat-omap/include/plat/gpmc.h
index 347d212..a36a046 100644
--- a/arch/arm/plat-omap/include/plat/gpmc.h
+++ b/arch/arm/plat-omap/include/plat/gpmc.h
@@ -55,10 +55,13 @@ 
 #define GPMC_CHUNK_SHIFT        24              /* 16 MB */
 #define GPMC_SECTION_SHIFT      28              /* 128 MB */
 
-#define PREFETCH_FIFOTHRESHOLD  (0x40 << 8)
+#define PREFETCH_FIFOTHRESHOLD_MAX	0x40
+#define PREFETCH_FIFOTHRESHOLD(val)	(val << 8)
 #define CS_NUM_SHIFT            24
 #define ENABLE_PREFETCH         (0x1 << 7)
 #define DMA_MPU_MODE            2
+#define PREFETCH_FIFOTHRESHOLD_READ	32 /* threashold size for read */
+#define PREFETCH_FIFOTHRESHOLD_WRITE	32 /* threashold size for write */
 
 #define GPMC_CONFIG1_WRAPBURST_SUPP     (1 << 31)
 #define GPMC_CONFIG1_READMULTIPLE_SUPP  (1 << 30)
@@ -136,7 +139,7 @@  extern int gpmc_cs_request(int cs, unsigned long size, unsigned long *base);
 extern void gpmc_cs_free(int cs);
 extern int gpmc_cs_set_reserved(int cs, int reserved);
 extern int gpmc_cs_reserved(int cs);
-extern int gpmc_prefetch_enable(int cs, int dma_mode,
+extern int gpmc_prefetch_enable(int cs, int fifo_th, int dma_mode,
 					unsigned int u32_count, int is_write);
 extern void gpmc_prefetch_reset(void);
 extern int gpmc_prefetch_status(void);
diff --git a/drivers/mtd/nand/omap2.c b/drivers/mtd/nand/omap2.c
index de9b058..61c0c01
--- a/drivers/mtd/nand/omap2.c
+++ b/drivers/mtd/nand/omap2.c
@@ -310,7 +310,7 @@  static void omap_read_buf_pref(struct mtd_info *mtd, u_char *buf, int len)
 	}
 
 	/* configure and start prefetch transfer */
-	ret = gpmc_prefetch_enable(info->gpmc_cs, 0x0, len, 0x0);
+	ret = gpmc_prefetch_enable(info->gpmc_cs, 0x40, 0x0, len, 0x0);
 	if (ret) {
 		/* PFPW engine is busy, use cpu copy method */
 		if (info->nand.options & NAND_BUSWIDTH_16)
@@ -354,7 +354,7 @@  static void omap_write_buf_pref(struct mtd_info *mtd,
 	}
 
 	/*  configure and start prefetch transfer */
-	ret = gpmc_prefetch_enable(info->gpmc_cs, 0x0, len, 0x1);
+	ret = gpmc_prefetch_enable(info->gpmc_cs, 0x40, 0x0, len, 0x1);
 	if (ret) {
 		/* PFPW engine is busy, use cpu copy method */
 		if (info->nand.options & NAND_BUSWIDTH_16)
@@ -405,10 +405,11 @@  static inline int omap_nand_dma_transfer(struct mtd_info *mtd, void *addr,
 	dma_addr_t dma_addr;
 	int ret;
 
-	/* The fifo depth is 64 bytes. We have a sync at each frame and frame
-	 * length is 64 bytes.
+	/* The fifo depth is 64 bytes max.
+	 * But configure the FIFO-threahold to 32 to get a sync at each frame
+	 * and frame length is 32 bytes.
 	 */
-	int buf_len = len >> 6;
+	int buf_len = len >> 5;
 
 	if (addr >= high_memory) {
 		struct page *p1;
@@ -447,7 +448,7 @@  static inline int omap_nand_dma_transfer(struct mtd_info *mtd, void *addr,
 					OMAP24XX_DMA_GPMC, OMAP_DMA_SRC_SYNC);
 	}
 	/*  configure and start prefetch transfer */
-	ret = gpmc_prefetch_enable(info->gpmc_cs, 0x1, len, is_write);
+	ret = gpmc_prefetch_enable(info->gpmc_cs, 0x20, 0x1, len, is_write);
 	if (ret)
 		/* PFPW engine is busy, use cpu copy methode */
 		goto out_copy;
@@ -524,6 +525,7 @@  static void omap_write_buf_dma_pref(struct mtd_info *mtd,
 static irqreturn_t omap_nand_irq(int this_irq, void *dev)
 {
 	struct omap_nand_info *info = (struct omap_nand_info *) dev;
+	u32 *p = (u32 *) info->buf;
 	u32 irq_enb = 0, pref_status = 0, bytes = 0;
 	u32 irq_stats = __raw_readl(info->gpmc_baseaddr + GPMC_IRQSTATUS);
 	u32 pref_config = __raw_readl(info->gpmc_baseaddr +
@@ -533,14 +535,11 @@  static irqreturn_t omap_nand_irq(int this_irq, void *dev)
 		if (irq_stats & 0x2)
 			goto done;
 
-		u32 *p = (u32 *) info->buf;
-		pref_status = gpmc_prefetch_status();
-		bytes = ((pref_status >> 24) & 0x7F);
+		bytes = PREFETCH_FIFOTHRESHOLD_WRITE;
 		iowrite32_rep(info->nand_pref_fifo_add, p, bytes >> 2);
 		info->buf = info->buf + bytes;
 
 	} else {
-		u32 *p = (u32 *) info->buf;
 		pref_status = gpmc_prefetch_status();
 		bytes = ((pref_status >> 24) & 0x7F);
 		ioread32_rep(info->nand_pref_fifo_add, p, bytes >> 2);
@@ -586,7 +585,8 @@  static void omap_read_buf_irq_pref(struct mtd_info *mtd, u_char *buf, int len)
 	init_completion(&info->comp);
 
 	/*  configure and start prefetch transfer */
-	ret = gpmc_prefetch_enable(info->gpmc_cs, 0x0, len, 0x0);
+	ret = gpmc_prefetch_enable(info->gpmc_cs,
+				PREFETCH_FIFOTHRESHOLD_READ, 0x0, len, 0x0);
 	if (ret)
 		/* PFPW engine is busy, use cpu copy methode */
 		goto out_copy;
@@ -630,7 +630,8 @@  static void omap_write_buf_irq_pref(struct mtd_info *mtd,
 	init_completion(&info->comp);
 
 	/*  configure and start prefetch transfer */
-	ret = gpmc_prefetch_enable(info->gpmc_cs, 0x0, len, 0x1);
+	ret = gpmc_prefetch_enable(info->gpmc_cs,
+				PREFETCH_FIFOTHRESHOLD_WRITE, 0x0, len, 0x1);
 	if (ret)
 		/* PFPW engine is busy, use cpu copy methode */
 		goto out_copy;