Patchwork mtd: m25p80: make command buffer DMA-safe

login
register
mail settings
Submitter Johannes Stezenbach
Date Oct. 28, 2009, 1:21 p.m.
Message ID <20091028132137.GA17813@sig21.net>
Download mbox | patch
Permalink /patch/37104/
State Accepted
Commit 61c3506c2cabe58bcdfe438d1e57b62994db1616
Headers show

Comments

Johannes Stezenbach - Oct. 28, 2009, 1:21 p.m.
spi_write() requires the buffer to be DMA-safe, kmalloc()
it seperately to ensure this.

Signed-off-by: Johannes Stezenbach <js@sig21.net>
Artem Bityutskiy - Nov. 3, 2009, 6:36 a.m.
On Wed, 2009-10-28 at 14:21 +0100, Johannes Stezenbach wrote:
> spi_write() requires the buffer to be DMA-safe, kmalloc()
> it seperately to ensure this.
> 
> Signed-off-by: Johannes Stezenbach <js@sig21.net>
> 
> 
> diff --git a/drivers/mtd/devices/m25p80.c b/drivers/mtd/devices/m25p80.c
> index 4c19269..21dd4d9 100644
> --- a/drivers/mtd/devices/m25p80.c
> +++ b/drivers/mtd/devices/m25p80.c
> @@ -79,7 +79,7 @@ struct m25p {
>  	struct mtd_info		mtd;
>  	unsigned		partitioned:1;
>  	u8			erase_opcode;
> -	u8			command[CMD_SIZE + FAST_READ_DUMMY_BYTE];
> +	u8			*command;
>  };
>  
>  static inline struct m25p *mtd_to_m25p(struct mtd_info *mtd)
> @@ -770,6 +770,11 @@ static int __devinit m25p_probe(struct spi_device *spi)
>  	flash = kzalloc(sizeof *flash, GFP_KERNEL);
>  	if (!flash)
>  		return -ENOMEM;
> +	flash->command = kmalloc(CMD_SIZE + FAST_READ_DUMMY_BYTE, GFP_KERNEL);
> +	if (!flash->command) {
> +		kfree(flash);
> +		return -ENOMEM;
> +	}

Even though it is just 4 or 5 bytes it can do DMA? Does not sound too
sane to use DMA in that case. Does this patch fix a real error?

I do not know much about SPI, but for me it sounds like there should be
a method to ask SPI to avoid using DMA, and you should use that method.
Johannes Stezenbach - Nov. 3, 2009, 10 a.m.
On Tue, Nov 03, 2009 at 08:36:05AM +0200, Artem Bityutskiy wrote:
> On Wed, 2009-10-28 at 14:21 +0100, Johannes Stezenbach wrote:
> > spi_write() requires the buffer to be DMA-safe, kmalloc()
> > it seperately to ensure this.
> 
> Even though it is just 4 or 5 bytes it can do DMA? Does not sound too
> sane to use DMA in that case. Does this patch fix a real error?
> 
> I do not know much about SPI, but for me it sounds like there should be
> a method to ask SPI to avoid using DMA, and you should use that method.

It fixes a real error -- with an out-of-tree driver and ancient kernel.
For the flash read/write case it is used with list-DMA (e.g. write 4 bytes,
read 64KB).  There are extensive comments in include/linux/spi/spi.h
which document the DMA-safe requirement of the buffers.

Thanks
Johannes
Artem Bityutskiy - Nov. 10, 2009, 2:54 p.m.
On Tue, 2009-11-03 at 11:00 +0100, Johannes Stezenbach wrote:
> On Tue, Nov 03, 2009 at 08:36:05AM +0200, Artem Bityutskiy wrote:
> > On Wed, 2009-10-28 at 14:21 +0100, Johannes Stezenbach wrote:
> > > spi_write() requires the buffer to be DMA-safe, kmalloc()
> > > it seperately to ensure this.
> > 
> > Even though it is just 4 or 5 bytes it can do DMA? Does not sound too
> > sane to use DMA in that case. Does this patch fix a real error?
> > 
> > I do not know much about SPI, but for me it sounds like there should be
> > a method to ask SPI to avoid using DMA, and you should use that method.
> 
> It fixes a real error -- with an out-of-tree driver and ancient kernel.
> For the flash read/write case it is used with list-DMA (e.g. write 4 bytes,
> read 64KB).  There are extensive comments in include/linux/spi/spi.h
> which document the DMA-safe requirement of the buffers.

Pushed to my l2-mtd-2.6 tree, thanks.

Patch

diff --git a/drivers/mtd/devices/m25p80.c b/drivers/mtd/devices/m25p80.c
index 4c19269..21dd4d9 100644
--- a/drivers/mtd/devices/m25p80.c
+++ b/drivers/mtd/devices/m25p80.c
@@ -79,7 +79,7 @@  struct m25p {
 	struct mtd_info		mtd;
 	unsigned		partitioned:1;
 	u8			erase_opcode;
-	u8			command[CMD_SIZE + FAST_READ_DUMMY_BYTE];
+	u8			*command;
 };
 
 static inline struct m25p *mtd_to_m25p(struct mtd_info *mtd)
@@ -770,6 +770,11 @@  static int __devinit m25p_probe(struct spi_device *spi)
 	flash = kzalloc(sizeof *flash, GFP_KERNEL);
 	if (!flash)
 		return -ENOMEM;
+	flash->command = kmalloc(CMD_SIZE + FAST_READ_DUMMY_BYTE, GFP_KERNEL);
+	if (!flash->command) {
+		kfree(flash);
+		return -ENOMEM;
+	}
 
 	flash->spi = spi;
 	mutex_init(&flash->lock);
@@ -888,8 +893,10 @@  static int __devexit m25p_remove(struct spi_device *spi)
 		status = del_mtd_partitions(&flash->mtd);
 	else
 		status = del_mtd_device(&flash->mtd);
-	if (status == 0)
+	if (status == 0) {
+		kfree(flash->command);
 		kfree(flash);
+	}
 	return 0;
 }