[PATCH] dw_mmc: add support for pre_req and post_req

[PATCH] dw_mmc: add support for pre_req and post_req

[Jaehoon Chung]

This patch is based on Per Forlin's patches.
([PATCH v2 00/12]mmc: use nonblock mmc requests to minimize latency)
After applied Per Forlin's patches, this patch must be apply.
In dw_mmc.c, support for pre/post_reqeust().

Signed-off-by: Jaehoon Chung <jh80.chung@samsung.com>
Signed-off-by: kyungmin Park <kyungmin.park@samsung.com>
---
 drivers/mmc/host/dw_mmc.c  |   84 ++++++++++++++++++++++++++++++++++++++++----
 include/linux/mmc/dw_mmc.h |    7 ++++
 2 files changed, 84 insertions(+), 7 deletions(-)

diff --git a/drivers/mmc/host/dw_mmc.c b/drivers/mmc/host/dw_mmc.c
index 87e1f57..67eeb85 100644
--- a/drivers/mmc/host/dw_mmc.c
+++ b/drivers/mmc/host/dw_mmc.c
@@ -417,6 +417,42 @@ static int dw_mci_idmac_init(struct dw_mci *host)
 	return 0;
 }
 
+static unsigned int dw_mci_pre_dma_transfer(struct dw_mci *host,
+			struct mmc_data *data, struct dw_mci_next *next)
+{
+	unsigned int sg_len;
+
+	BUG_ON(next && data->host_cookie);
+	BUG_ON(!next && data->host_cookie &&
+			data->host_cookie != host->next_data.cookie);
+
+	if (!next && data->host_cookie &&
+			data->host_cookie != host->next_data.cookie) {
+		data->host_cookie = 0;
+	}
+
+	if (next ||
+		(!next && data->host_cookie != host->next_data.cookie)) {
+		sg_len = dma_map_sg(&host->pdev->dev, data->sg,
+				data->sg_len, ((data->flags & MMC_DATA_WRITE)
+			      ? DMA_TO_DEVICE : DMA_FROM_DEVICE));
+	} else {
+		sg_len = host->next_data.sg_len;
+		host->next_data.sg_len = 0;
+	}
+
+	if (sg_len == 0)
+		return -EINVAL;
+
+	if (next) {
+		next->sg_len = sg_len;
+		data->host_cookie = ++next->cookie < 0 ? 1 : next->cookie;
+	} else
+		data->sg_len = sg_len;
+
+	return sg_len;
+}
+
 static struct dw_mci_dma_ops dw_mci_idmac_ops = {
 	.init = dw_mci_idmac_init,
 	.start = dw_mci_idmac_start_dma,
@@ -451,13 +487,9 @@ static int dw_mci_submit_data_dma(struct dw_mci *host, struct mmc_data *data)
 			return -EINVAL;
 	}
 
-	if (data->flags & MMC_DATA_READ)
-		direction = DMA_FROM_DEVICE;
-	else
-		direction = DMA_TO_DEVICE;
-
-	sg_len = dma_map_sg(&host->pdev->dev, data->sg, data->sg_len,
-			    direction);
+	sg_len = dw_mci_pre_dma_transfer(host, data, NULL);
+	if (sg_len < 0)
+		return sg_len;
 
 	dev_vdbg(&host->pdev->dev,
 		 "sd sg_cpu: %#lx sg_dma: %#lx sg_len: %d\n",
@@ -643,6 +675,42 @@ static void dw_mci_queue_request(struct dw_mci *host, struct dw_mci_slot *slot,
 	spin_unlock_bh(&host->lock);
 }
 
+static void dw_mci_post_request(struct mmc_host *mmc, struct mmc_request *mrq,
+		int err)
+{
+	struct dw_mci_slot *slot = mmc_priv(mmc);
+	struct mmc_data *data = mrq->data;
+
+	if (!data)
+		return;
+
+	if (slot->host->use_dma) {
+		dma_unmap_sg(&slot->host->pdev->dev, data->sg, data->sg_len,
+			     ((data->flags & MMC_DATA_WRITE)
+			      ? DMA_TO_DEVICE : DMA_FROM_DEVICE));
+
+		data->host_cookie = 0;
+	}
+}
+
+static void dw_mci_pre_request(struct mmc_host *mmc, struct mmc_request *mrq,
+		bool is_first_req)
+{
+	struct dw_mci_slot *slot = mmc_priv(mmc);
+	struct mmc_data *data = mrq->data;
+
+	if (!data)
+		return;
+
+	BUG_ON(mrq->data->host_cookie);
+
+	if (slot->host->use_dma) {
+		if (dw_mci_pre_dma_transfer(slot->host, mrq->data,
+					&slot->host->next_data))
+			mrq->data->host_cookie = 0;
+	}
+}
+
 static void dw_mci_request(struct mmc_host *mmc, struct mmc_request *mrq)
 {
 	struct dw_mci_slot *slot = mmc_priv(mmc);
@@ -748,6 +816,8 @@ static int dw_mci_get_cd(struct mmc_host *mmc)
 
 static const struct mmc_host_ops dw_mci_ops = {
 	.request	= dw_mci_request,
+	.pre_req	= dw_mci_pre_request,
+	.post_req	= dw_mci_post_request,
 	.set_ios	= dw_mci_set_ios,
 	.get_ro		= dw_mci_get_ro,
 	.get_cd		= dw_mci_get_cd,
diff --git a/include/linux/mmc/dw_mmc.h b/include/linux/mmc/dw_mmc.h
index c0207a7..dca82ee 100644
--- a/include/linux/mmc/dw_mmc.h
+++ b/include/linux/mmc/dw_mmc.h
@@ -35,6 +35,11 @@ enum {
 
 struct mmc_data;
 
+struct dw_mci_next {
+	unsigned int	sg_len;
+	s32		cookie;
+};
+
 /**
  * struct dw_mci - MMC controller state shared between all slots
  * @lock: Spinlock protecting the queue and associated data.
@@ -154,6 +159,8 @@ struct dw_mci {
 	u32			quirks;
 
 	struct regulator	*vmmc;	/* Power regulator */
+
+	struct dw_mci_next	next_data;
 };
 
 /* DMA ops for Internal/External DMAC interface */

Re: [PATCH] dw_mmc: add support for pre_req and post_req

[Shawn Guo]

Hi Jaehoon,

On Thu, Apr 07, 2011 at 05:04:52PM +0900, Jaehoon Chung wrote:
[...]
> +static unsigned int dw_mci_pre_dma_transfer(struct dw_mci *host,
> +			struct mmc_data *data, struct dw_mci_next *next)
> +{
> +	unsigned int sg_len;
> +
> +	BUG_ON(next && data->host_cookie);
> +	BUG_ON(!next && data->host_cookie &&
> +			data->host_cookie != host->next_data.cookie);
> +
> +	if (!next && data->host_cookie &&
> +			data->host_cookie != host->next_data.cookie) {
> +		data->host_cookie = 0;
> +	}
> +
I'm unsure if the 'if' statement makes any sense here, since the
exactly same conditions have been caught by the BUG_ON just above
it.

> +	if (next ||
> +		(!next && data->host_cookie != host->next_data.cookie)) {
> +		sg_len = dma_map_sg(&host->pdev->dev, data->sg,
> +				data->sg_len, ((data->flags & MMC_DATA_WRITE)
> +			      ? DMA_TO_DEVICE : DMA_FROM_DEVICE));
> +	} else {
> +		sg_len = host->next_data.sg_len;
> +		host->next_data.sg_len = 0;
> +	}
> +
> +	if (sg_len == 0)
> +		return -EINVAL;
> +
> +	if (next) {
> +		next->sg_len = sg_len;
> +		data->host_cookie = ++next->cookie < 0 ? 1 : next->cookie;
> +	} else
> +		data->sg_len = sg_len;
> +
> +	return sg_len;
> +}
> +
Function dw_mci_pre_dma_transfer() returns non-zero value anyway,
either -EINVAL or sg_len ...

[...]
> +static void dw_mci_pre_request(struct mmc_host *mmc, struct mmc_request *mrq,
> +		bool is_first_req)
> +{
> +	struct dw_mci_slot *slot = mmc_priv(mmc);
> +	struct mmc_data *data = mrq->data;
> +
> +	if (!data)
> +		return;
> +
> +	BUG_ON(mrq->data->host_cookie);
> +
> +	if (slot->host->use_dma) {
> +		if (dw_mci_pre_dma_transfer(slot->host, mrq->data,
> +					&slot->host->next_data))
> +			mrq->data->host_cookie = 0;
... while it steps back to old blocking way by setting
data->host_cookie 0 when dw_mci_pre_dma_transfer returns non-zero.

Per my understanding, it means the non-blocking optimization will
always get bypassed anyway, so I doubt the patch can really gain
performance improvement.  Did you get the chance to measure?

-- 
Regards,
Shawn

Re: [PATCH] dw_mmc: add support for pre_req and post_req

[Jaehoon Chung]

Hi Shawn..

Shawn Guo wrote:
> Hi Jaehoon,
> 
> On Thu, Apr 07, 2011 at 05:04:52PM +0900, Jaehoon Chung wrote:
> [...]
>> +static unsigned int dw_mci_pre_dma_transfer(struct dw_mci *host,
>> +			struct mmc_data *data, struct dw_mci_next *next)
>> +{
>> +	unsigned int sg_len;
>> +
>> +	BUG_ON(next && data->host_cookie);
>> +	BUG_ON(!next && data->host_cookie &&
>> +			data->host_cookie != host->next_data.cookie);
>> +
>> +	if (!next && data->host_cookie &&
>> +			data->host_cookie != host->next_data.cookie) {
>> +		data->host_cookie = 0;
>> +	}
>> +
> I'm unsure if the 'if' statement makes any sense here, since the
> exactly same conditions have been caught by the BUG_ON just above
> it.
> 
You're right..i'll modify this..

>> +	if (next ||
>> +		(!next && data->host_cookie != host->next_data.cookie)) {
>> +		sg_len = dma_map_sg(&host->pdev->dev, data->sg,
>> +				data->sg_len, ((data->flags & MMC_DATA_WRITE)
>> +			      ? DMA_TO_DEVICE : DMA_FROM_DEVICE));
>> +	} else {
>> +		sg_len = host->next_data.sg_len;
>> +		host->next_data.sg_len = 0;
>> +	}
>> +
>> +	if (sg_len == 0)
>> +		return -EINVAL;
>> +
>> +	if (next) {
>> +		next->sg_len = sg_len;
>> +		data->host_cookie = ++next->cookie < 0 ? 1 : next->cookie;
>> +	} else
>> +		data->sg_len = sg_len;
>> +
>> +	return sg_len;
>> +}
>> +
> Function dw_mci_pre_dma_transfer() returns non-zero value anyway,
> either -EINVAL or sg_len ...
> 
Sorry,, i didn't understand this your comments..

> [...]
>> +static void dw_mci_pre_request(struct mmc_host *mmc, struct mmc_request *mrq,
>> +		bool is_first_req)
>> +{
>> +	struct dw_mci_slot *slot = mmc_priv(mmc);
>> +	struct mmc_data *data = mrq->data;
>> +
>> +	if (!data)
>> +		return;
>> +
>> +	BUG_ON(mrq->data->host_cookie);
>> +
>> +	if (slot->host->use_dma) {
>> +		if (dw_mci_pre_dma_transfer(slot->host, mrq->data,
>> +					&slot->host->next_data))
>> +			mrq->data->host_cookie = 0;
> ... while it steps back to old blocking way by setting
> data->host_cookie 0 when dw_mci_pre_dma_transfer returns non-zero.
> 
> Per my understanding, it means the non-blocking optimization will
> always get bypassed anyway, so I doubt the patch can really gain
> performance improvement.  Did you get the chance to measure?
> 
Actually, i didn't get performance improvement, but didn't fully affect by CPU_FREQ.
Somebody get performance improvement?

Regards,
Jaehoon Chung

Re: [PATCH] dw_mmc: add support for pre_req and post_req

[Shawn Guo]

On Mon, Apr 18, 2011 at 01:55:11PM +0900, Jaehoon Chung wrote:
> Hi Shawn..
> 
> Shawn Guo wrote:
> > Hi Jaehoon,
> > 
> > On Thu, Apr 07, 2011 at 05:04:52PM +0900, Jaehoon Chung wrote:
> > [...]
> >> +static unsigned int dw_mci_pre_dma_transfer(struct dw_mci *host,
> >> +			struct mmc_data *data, struct dw_mci_next *next)
> >> +{
> >> +	unsigned int sg_len;
> >> +
> >> +	BUG_ON(next && data->host_cookie);
> >> +	BUG_ON(!next && data->host_cookie &&
> >> +			data->host_cookie != host->next_data.cookie);
> >> +
> >> +	if (!next && data->host_cookie &&
> >> +			data->host_cookie != host->next_data.cookie) {
> >> +		data->host_cookie = 0;
> >> +	}
> >> +
> > I'm unsure if the 'if' statement makes any sense here, since the
> > exactly same conditions have been caught by the BUG_ON just above
> > it.
> > 
> You're right..i'll modify this..
> 
> >> +	if (next ||
> >> +		(!next && data->host_cookie != host->next_data.cookie)) {
> >> +		sg_len = dma_map_sg(&host->pdev->dev, data->sg,
> >> +				data->sg_len, ((data->flags & MMC_DATA_WRITE)
> >> +			      ? DMA_TO_DEVICE : DMA_FROM_DEVICE));
> >> +	} else {
> >> +		sg_len = host->next_data.sg_len;
> >> +		host->next_data.sg_len = 0;
> >> +	}
> >> +
> >> +	if (sg_len == 0)
> >> +		return -EINVAL;
> >> +
> >> +	if (next) {
> >> +		next->sg_len = sg_len;
> >> +		data->host_cookie = ++next->cookie < 0 ? 1 : next->cookie;
> >> +	} else
> >> +		data->sg_len = sg_len;
> >> +
> >> +	return sg_len;
> >> +}
> >> +
> > Function dw_mci_pre_dma_transfer() returns non-zero value anyway,
> > either -EINVAL or sg_len ...
> > 
> Sorry,, i didn't understand this your comments..
> 
> > [...]
> >> +static void dw_mci_pre_request(struct mmc_host *mmc, struct mmc_request *mrq,
> >> +		bool is_first_req)
> >> +{
> >> +	struct dw_mci_slot *slot = mmc_priv(mmc);
> >> +	struct mmc_data *data = mrq->data;
> >> +
> >> +	if (!data)
> >> +		return;
> >> +
> >> +	BUG_ON(mrq->data->host_cookie);
> >> +
> >> +	if (slot->host->use_dma) {
> >> +		if (dw_mci_pre_dma_transfer(slot->host, mrq->data,
> >> +					&slot->host->next_data))
> >> +			mrq->data->host_cookie = 0;
> > ... while it steps back to old blocking way by setting
> > data->host_cookie 0 when dw_mci_pre_dma_transfer returns non-zero.
> > 
> > Per my understanding, it means the non-blocking optimization will
> > always get bypassed anyway, so I doubt the patch can really gain
> > performance improvement.  Did you get the chance to measure?
> > 
> Actually, i didn't get performance improvement, but didn't fully affect by CPU_FREQ.
> Somebody get performance improvement?
> 
Though the performance improvement is limited, I did see nearly 4%
improvement with best case on mxs-mmc.


Try the following change in dw_mci_pre_request to see if you can see
any performance difference.

	if (slot->host->use_dma) {
		if (dw_mci_pre_dma_transfer(slot->host, mrq->data,
					&slot->host->next_data) < 0)
			mrq->data->host_cookie = 0;

-- 
Regards,
Shawn

Re: [PATCH] dw_mmc: add support for pre_req and post_req

[Jaehoon Chung]

Hi Shawn,

Shawn Guo wrote:
> 
> 	if (slot->host->use_dma) {
> 		if (dw_mci_pre_dma_transfer(slot->host, mrq->data,
> 					&slot->host->next_data) < 0)
> 			mrq->data->host_cookie = 0;
> 

Ok..I understood yours..:) i'll check and test..
Thanks for your comment.

Regards,
Jaehoon Chung