[PATCH] mmc: sdio: To use multiple scatter gather list

[PATCH] mmc: sdio: To use multiple scatter gather list

[Kyoungil Kim]

We used only single sg entry for SDIO transfer.
This chagne is to use multiple sg entry.
In case of dwmci, it support only up to 4KB size per single sg entry.
So if we want to transfer more than 4KB, we should send more than 1 command.
This makes performance degradation for large data transfer.

Signed-off-by: Kyoungil Kim <ki0351.kim@samsung.com>
---
 drivers/mmc/core/sdio_io.c  |   10 +++-------
 drivers/mmc/core/sdio_ops.c |   33 +++++++++++++++++++++++++++++----
 2 files changed, 32 insertions(+), 11 deletions(-)

diff --git a/drivers/mmc/core/sdio_io.c b/drivers/mmc/core/sdio_io.c
index 8f6f5ac..78cb4d5 100644
--- a/drivers/mmc/core/sdio_io.c
+++ b/drivers/mmc/core/sdio_io.c
@@ -188,8 +188,7 @@ EXPORT_SYMBOL_GPL(sdio_set_block_size);
  */
 static inline unsigned int sdio_max_byte_size(struct sdio_func *func)
 {
-	unsigned mval =	min(func->card->host->max_seg_size,
-			    func->card->host->max_blk_size);
+	unsigned mval =	func->card->host->max_blk_size;
 
 	if (mmc_blksz_for_byte_mode(func->card))
 		mval = min(mval, func->cur_blksize);
@@ -311,11 +310,8 @@ static int sdio_io_rw_ext_helper(struct sdio_func *func, int write,
 	/* Do the bulk of the transfer using block mode (if supported). */
 	if (func->card->cccr.multi_block && (size > sdio_max_byte_size(func))) {
 		/* Blocks per command is limited by host count, host transfer
-		 * size (we only use a single sg entry) and the maximum for
-		 * IO_RW_EXTENDED of 511 blocks. */
-		max_blocks = min(func->card->host->max_blk_count,
-			func->card->host->max_seg_size / func->cur_blksize);
-		max_blocks = min(max_blocks, 511u);
+		 * size and the maximum for IO_RW_EXTENDED of 511 blocks. */
+		max_blocks = min(func->card->host->max_blk_count, 511u);
 
 		while (remainder >= func->cur_blksize) {
 			unsigned blocks;
diff --git a/drivers/mmc/core/sdio_ops.c b/drivers/mmc/core/sdio_ops.c
index d29e206..ac01352 100644
--- a/drivers/mmc/core/sdio_ops.c
+++ b/drivers/mmc/core/sdio_ops.c
@@ -124,7 +124,10 @@ int mmc_io_rw_extended(struct mmc_card *card, int write, unsigned fn,
 	struct mmc_request mrq = {NULL};
 	struct mmc_command cmd = {0};
 	struct mmc_data data = {0};
-	struct scatterlist sg;
+	struct scatterlist sg, *sg_ptr;
+	struct sg_table sgtable;
+	unsigned int nents, left_size, i;
+	unsigned int seg_size = card->host->max_seg_size;
 
 	BUG_ON(!card);
 	BUG_ON(fn > 7);
@@ -152,15 +155,37 @@ int mmc_io_rw_extended(struct mmc_card *card, int write, unsigned fn,
 	/* Code in host drivers/fwk assumes that "blocks" always is >=1 */
 	data.blocks = blocks ? blocks : 1;
 	data.flags = write ? MMC_DATA_WRITE : MMC_DATA_READ;
-	data.sg = &sg;
-	data.sg_len = 1;
 
-	sg_init_one(&sg, buf, data.blksz * data.blocks);
+	left_size = data.blksz * data.blocks;
+	nents = (left_size - 1) / seg_size + 1;
+	if (nents > 1) {
+		if (sg_alloc_table(&sgtable, nents, GFP_KERNEL))
+			return -ENOMEM;
+
+		data.sg = sgtable.sgl;
+		data.sg_len = nents;
+
+		for_each_sg(data.sg, sg_ptr, data.sg_len, i) {
+			sg_set_page(sg_ptr, virt_to_page(buf + (i * seg_size)),
+					min(seg_size, left_size),
+					offset_in_page(buf + (i * seg_size)));
+			left_size = left_size - seg_size;
+		}
+	} else {
+		data.sg = &sg;
+		data.sg_len = 1;
+
+		sg_init_one(&sg, buf, left_size);
+	}
 
 	mmc_set_data_timeout(&data, card);
 
 	mmc_wait_for_req(card->host, &mrq);
 
+	if (nents > 1)
+		sg_free_table(&sgtable);
+
+
 	if (cmd.error)
 		return cmd.error;
 	if (data.error)
-- 
1.7.4.1

Re: [PATCH] mmc: sdio: To use multiple scatter gather list

[Chris Ball]

Hi,

On Mon, Oct 22 2012, Kyoungil Kim wrote:
> We used only single sg entry for SDIO transfer.
> This chagne is to use multiple sg entry.
> In case of dwmci, it support only up to 4KB size per single sg entry.
> So if we want to transfer more than 4KB, we should send more than 1 command.
> This makes performance degradation for large data transfer.
>
> Signed-off-by: Kyoungil Kim <ki0351.kim@samsung.com>

Can you share the before/after benchmark on your controller, please?

Thanks,

- Chris.
-- 
Chris Ball   <cjb@laptop.org>   <http://printf.net/>
One Laptop Per Child

RE: [PATCH] mmc: sdio: To use multiple scatter gather list

[Kyoungil Kim]

On Tue, Oct 30 2012, Chris Ball wrote:
> Hi,
> 
> On Mon, Oct 22 2012, Kyoungil Kim wrote:
> > We used only single sg entry for SDIO transfer.
> > This chagne is to use multiple sg entry.
> > In case of dwmci, it support only up to 4KB size per single sg entry.
> > So if we want to transfer more than 4KB, we should send more than 1 command.
> > This makes performance degradation for large data transfer.
> >
> > Signed-off-by: Kyoungil Kim <ki0351.kim@samsung.com>
> 
> Can you share the before/after benchmark on your controller, please?

When we test before apply this patch, it takes around 335 us for 5K(5120) bytes transfer with dwmci controller.
But when we test after apply this patch, it takes 242 us for 5K bytes.
So this patch makes around 38% performance improvement for 5K bytes transfer.
If the transfer size is bigger, then the performance improvement ratio will be increased.

Thanks.

Re: [PATCH] mmc: sdio: To use multiple scatter gather list

[Chris Ball]

Hi,

On Wed, Nov 07 2012, Kyoungil Kim wrote:
> On Tue, Oct 30 2012, Chris Ball wrote:
>> Hi,
>> 
>> On Mon, Oct 22 2012, Kyoungil Kim wrote:
>> > We used only single sg entry for SDIO transfer.
>> > This chagne is to use multiple sg entry.
>> > In case of dwmci, it support only up to 4KB size per single sg entry.
>> > So if we want to transfer more than 4KB, we should send more than
>> > 1 command.
>> > This makes performance degradation for large data transfer.
>> >
>> > Signed-off-by: Kyoungil Kim <ki0351.kim@samsung.com>
>> 
>> Can you share the before/after benchmark on your controller, please?
>
> When we test before apply this patch, it takes around 335 us for
> 5K(5120) bytes transfer with dwmci controller.
> But when we test after apply this patch, it takes 242 us for 5K bytes.
> So this patch makes around 38% performance improvement for 5K bytes transfer.
> If the transfer size is bigger, then the performance improvement ratio
> will be increased.

Thanks, pushed to mmc-next for 3.8.

- Chris.
-- 
Chris Ball   <cjb@laptop.org>   <http://printf.net/>
One Laptop Per Child