[PATCH 0/2] dmaengine: hdmac: Add scatter-gathered memset support

[PATCH 0/2] dmaengine: hdmac: Add scatter-gathered memset support

[Maxime Ripard]

Hi Vinod,

Here is a patch serie that adds support in the Atmel HDMAC for memset
over a discontiguous memory area.

Just like for the memset, it's used on those SoCs to support
framebuffer related operations that cannot be performed either by the
display engine or the (non-existent) GPU.

Let me know what you think,
Maxime

Maxime Ripard (2):
  dmaengine: hdmac: factorise memset descriptor allocation
  dmaengine: hdmac: Add scatter-gathered memset support

 drivers/dma/at_hdmac.c      | 169 ++++++++++++++++++++++++++++++++++----------
 drivers/dma/at_hdmac_regs.h |   2 +-
 2 files changed, 134 insertions(+), 37 deletions(-)

-- 
2.5.3

[PATCH 1/2] dmaengine: hdmac: factorise memset descriptor allocation

[Maxime Ripard]

The memset and scatter gathered memset are going to use some common logic
to create their descriptors.

Move that logic into a function of its own so that we can share it with the
future memset_sg callback.

Signed-off-by: Maxime Ripard <maxime.ripard@free-electrons.com>
---
 drivers/dma/at_hdmac.c      | 98 +++++++++++++++++++++++++++------------------
 drivers/dma/at_hdmac_regs.h |  2 +-
 2 files changed, 59 insertions(+), 41 deletions(-)

diff --git a/drivers/dma/at_hdmac.c b/drivers/dma/at_hdmac.c
index 58d406230d89..009103fd6d9a 100644
--- a/drivers/dma/at_hdmac.c
+++ b/drivers/dma/at_hdmac.c
@@ -458,10 +458,10 @@ atc_chain_complete(struct at_dma_chan *atchan, struct at_desc *desc)
 		dma_cookie_complete(txd);
 
 	/* If the transfer was a memset, free our temporary buffer */
-	if (desc->memset) {
+	if (desc->memset_buffer) {
 		dma_pool_free(atdma->memset_pool, desc->memset_vaddr,
 			      desc->memset_paddr);
-		desc->memset = false;
+		desc->memset_buffer = false;
 	}
 
 	/* move children to free_list */
@@ -881,6 +881,47 @@ err_desc_get:
 	return NULL;
 }
 
+static struct at_desc *atc_create_memset_desc(struct dma_chan *chan,
+					      dma_addr_t psrc,
+					      dma_addr_t pdst,
+					      size_t len)
+{
+	struct at_dma_chan *atchan = to_at_dma_chan(chan);
+	struct at_desc *desc;
+	size_t xfer_count;
+
+	u32 ctrla = ATC_SRC_WIDTH(2) |
+		ATC_DST_WIDTH(2);
+	u32 ctrlb = ATC_DEFAULT_CTRLB | ATC_IEN |
+		ATC_SRC_ADDR_MODE_FIXED |
+		ATC_DST_ADDR_MODE_INCR |
+		ATC_FC_MEM2MEM;
+
+	xfer_count = len >> 2;
+	if (xfer_count > ATC_BTSIZE_MAX) {
+		dev_err(chan2dev(chan), "%s: buffer is too big\n",
+			__func__);
+		return NULL;
+	}
+
+	desc = atc_desc_get(atchan);
+	if (!desc) {
+		dev_err(chan2dev(chan), "%s: can't get a descriptor\n",
+			__func__);
+		return NULL;
+	}
+
+	desc->lli.saddr = psrc;
+	desc->lli.daddr = pdst;
+	desc->lli.ctrla = ctrla | xfer_count;
+	desc->lli.ctrlb = ctrlb;
+
+	desc->txd.cookie = 0;
+	desc->len = len;
+
+	return desc;
+}
+
 /**
  * atc_prep_dma_memset - prepare a memcpy operation
  * @chan: the channel to prepare operation on
@@ -893,12 +934,10 @@ static struct dma_async_tx_descriptor *
 atc_prep_dma_memset(struct dma_chan *chan, dma_addr_t dest, int value,
 		    size_t len, unsigned long flags)
 {
-	struct at_dma_chan	*atchan = to_at_dma_chan(chan);
 	struct at_dma		*atdma = to_at_dma(chan->device);
-	struct at_desc		*desc = NULL;
-	size_t			xfer_count;
-	u32			ctrla;
-	u32			ctrlb;
+	struct at_desc		*desc;
+	void __iomem		*vaddr;
+	dma_addr_t		paddr;
 
 	dev_vdbg(chan2dev(chan), "%s: d0x%x v0x%x l0x%zx f0x%lx\n", __func__,
 		dest, value, len, flags);
@@ -914,46 +953,26 @@ atc_prep_dma_memset(struct dma_chan *chan, dma_addr_t dest, int value,
 		return NULL;
 	}
 
-	xfer_count = len >> 2;
-	if (xfer_count > ATC_BTSIZE_MAX) {
-		dev_err(chan2dev(chan), "%s: buffer is too big\n",
+	vaddr = dma_pool_alloc(atdma->memset_pool, GFP_ATOMIC, &paddr);
+	if (!vaddr) {
+		dev_err(chan2dev(chan), "%s: couldn't allocate buffer\n",
 			__func__);
 		return NULL;
 	}
+	*(u32*)vaddr = value;
 
-	ctrlb =   ATC_DEFAULT_CTRLB | ATC_IEN
-		| ATC_SRC_ADDR_MODE_FIXED
-		| ATC_DST_ADDR_MODE_INCR
-		| ATC_FC_MEM2MEM;
-
-	ctrla = ATC_SRC_WIDTH(2) |
-		ATC_DST_WIDTH(2);
-
-	desc = atc_desc_get(atchan);
+	desc = atc_create_memset_desc(chan, paddr, dest, len);
 	if (!desc) {
-		dev_err(chan2dev(chan), "%s: can't get a descriptor\n",
+		dev_err(chan2dev(chan), "%s: couldn't get a descriptor\n",
 			__func__);
-		return NULL;
+		goto err_free_buffer;
 	}
 
-	desc->memset_vaddr = dma_pool_alloc(atdma->memset_pool, GFP_ATOMIC,
-					    &desc->memset_paddr);
-	if (!desc->memset_vaddr) {
-		dev_err(chan2dev(chan), "%s: couldn't allocate buffer\n",
-			__func__);
-		goto err_put_desc;
-	}
-
-	*desc->memset_vaddr = value;
-	desc->memset = true;
-
-	desc->lli.saddr = desc->memset_paddr;
-	desc->lli.daddr = dest;
-	desc->lli.ctrla = ctrla | xfer_count;
-	desc->lli.ctrlb = ctrlb;
+	desc->memset_paddr = paddr;
+	desc->memset_vaddr = vaddr;
+	desc->memset_buffer = true;
 
 	desc->txd.cookie = -EBUSY;
-	desc->len = len;
 	desc->total_len = len;
 
 	/* set end-of-link on the descriptor */
@@ -963,12 +982,11 @@ atc_prep_dma_memset(struct dma_chan *chan, dma_addr_t dest, int value,
 
 	return &desc->txd;
 
-err_put_desc:
-	atc_desc_put(atchan, desc);
+err_free_buffer:
+	dma_pool_free(atdma->memset_pool, vaddr, paddr);
 	return NULL;
 }
 
-
 /**
  * atc_prep_slave_sg - prepare descriptors for a DMA_SLAVE transaction
  * @chan: DMA channel
diff --git a/drivers/dma/at_hdmac_regs.h b/drivers/dma/at_hdmac_regs.h
index c3bebbe899ac..d1cfc8c876f9 100644
--- a/drivers/dma/at_hdmac_regs.h
+++ b/drivers/dma/at_hdmac_regs.h
@@ -202,7 +202,7 @@ struct at_desc {
 	size_t				src_hole;
 
 	/* Memset temporary buffer */
-	bool				memset;
+	bool				memset_buffer;
 	dma_addr_t			memset_paddr;
 	int				*memset_vaddr;
 };
-- 
2.5.3

[PATCH 2/2] dmaengine: hdmac: Add scatter-gathered memset support

[Maxime Ripard]

Just like memset support, the HDMAC might be used to do a memset over a
discontiguous memory area.

In such a case, we'll just build up a chain of memset descriptors over the
contiguous chunks of memory to set, in order to allow such a support.

Signed-off-by: Maxime Ripard <maxime.ripard@free-electrons.com>
---
 drivers/dma/at_hdmac.c | 79 ++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 79 insertions(+)

diff --git a/drivers/dma/at_hdmac.c b/drivers/dma/at_hdmac.c
index 009103fd6d9a..f19651645a33 100644
--- a/drivers/dma/at_hdmac.c
+++ b/drivers/dma/at_hdmac.c
@@ -987,6 +987,83 @@ err_free_buffer:
 	return NULL;
 }
 
+static struct dma_async_tx_descriptor *
+atc_prep_dma_memset_sg(struct dma_chan *chan,
+		       struct scatterlist *sgl,
+		       unsigned int sg_len, int value,
+		       unsigned long flags)
+{
+	struct at_dma_chan	*atchan = to_at_dma_chan(chan);
+	struct at_dma		*atdma = to_at_dma(chan->device);
+	struct at_desc		*desc = NULL, *first = NULL, *prev = NULL;
+	struct scatterlist	*sg;
+	void __iomem		*vaddr;
+	dma_addr_t		paddr;
+	size_t			total_len = 0;
+	int			i;
+
+	dev_vdbg(chan2dev(chan), "%s: v0x%x l0x%zx f0x%lx\n", __func__,
+		 value, sg_len, flags);
+
+	if (unlikely(!sgl || !sg_len)) {
+		dev_dbg(chan2dev(chan), "%s: scatterlist is empty!\n",
+			__func__);
+		return NULL;
+	}
+
+	vaddr = dma_pool_alloc(atdma->memset_pool, GFP_ATOMIC, &paddr);
+	if (!vaddr) {
+		dev_err(chan2dev(chan), "%s: couldn't allocate buffer\n",
+			__func__);
+		return NULL;
+	}
+	*(u32*)vaddr = value;
+
+	for_each_sg(sgl, sg, sg_len, i) {
+		dma_addr_t dest = sg_dma_address(sg);
+		size_t len = sg_dma_len(sg);
+
+		dev_vdbg(chan2dev(chan), "%s: d0x%08x, l0x%zx\n",
+			 __func__, dest, len);
+
+		if (!is_dma_fill_aligned(chan->device, dest, 0, len)) {
+			dev_err(chan2dev(chan), "%s: buffer is not aligned\n",
+				__func__);
+			goto err_put_desc;
+		}
+
+		desc = atc_create_memset_desc(chan, paddr, dest, len);
+		if (!desc)
+			goto err_put_desc;
+
+		atc_desc_chain(&first, &prev, desc);
+
+		total_len += len;
+	}
+
+	/*
+	 * Only set the buffer pointers on the last descriptor to
+	 * avoid free'ing while we have our transfer still going
+	 */
+	desc->memset_paddr = paddr;
+	desc->memset_vaddr = vaddr;
+	desc->memset_buffer = true;
+
+	first->txd.cookie = -EBUSY;
+	first->total_len = total_len;
+
+	/* set end-of-link on the descriptor */
+	set_desc_eol(desc);
+
+	first->txd.flags = flags;
+
+	return &first->txd;
+
+err_put_desc:
+	atc_desc_put(atchan, first);
+	return NULL;
+}
+
 /**
  * atc_prep_slave_sg - prepare descriptors for a DMA_SLAVE transaction
  * @chan: DMA channel
@@ -1869,6 +1946,7 @@ static int __init at_dma_probe(struct platform_device *pdev)
 	dma_cap_set(DMA_INTERLEAVE, at91sam9g45_config.cap_mask);
 	dma_cap_set(DMA_MEMCPY, at91sam9g45_config.cap_mask);
 	dma_cap_set(DMA_MEMSET, at91sam9g45_config.cap_mask);
+	dma_cap_set(DMA_MEMSET_SG, at91sam9g45_config.cap_mask);
 	dma_cap_set(DMA_PRIVATE, at91sam9g45_config.cap_mask);
 	dma_cap_set(DMA_SLAVE, at91sam9g45_config.cap_mask);
 	dma_cap_set(DMA_SG, at91sam9g45_config.cap_mask);
@@ -1990,6 +2068,7 @@ static int __init at_dma_probe(struct platform_device *pdev)
 
 	if (dma_has_cap(DMA_MEMSET, atdma->dma_common.cap_mask)) {
 		atdma->dma_common.device_prep_dma_memset = atc_prep_dma_memset;
+		atdma->dma_common.device_prep_dma_memset_sg = atc_prep_dma_memset_sg;
 		atdma->dma_common.fill_align = DMAENGINE_ALIGN_4_BYTES;
 	}
 
-- 
2.5.3

Re: [PATCH 1/2] dmaengine: hdmac: factorise memset descriptor allocation

[Vinod Koul]

On Thu, Oct 01, 2015 at 04:52:37PM +0200, Maxime Ripard wrote:

> +static struct at_desc *atc_create_memset_desc(struct dma_chan *chan,
> +					      dma_addr_t psrc,
> +					      dma_addr_t pdst,
> +					      size_t len)
> +{
> +	struct at_dma_chan *atchan = to_at_dma_chan(chan);
> +	struct at_desc *desc;
> +	size_t xfer_count;
> +
> +	u32 ctrla = ATC_SRC_WIDTH(2) |
> +		ATC_DST_WIDTH(2);

why is this over two lines :)

> +	u32 ctrlb = ATC_DEFAULT_CTRLB | ATC_IEN |
> +		ATC_SRC_ADDR_MODE_FIXED |
> +		ATC_DST_ADDR_MODE_INCR |
> +		ATC_FC_MEM2MEM;
> +
> +	xfer_count = len >> 2;
> +	if (xfer_count > ATC_BTSIZE_MAX) {
> +		dev_err(chan2dev(chan), "%s: buffer is too big\n",
> +			__func__);
> +		return NULL;
> +	}

This is fine, but this is driver limitation. We should really split the
txn to multiple descriptors here..

> @@ -914,46 +953,26 @@ atc_prep_dma_memset(struct dma_chan *chan, dma_addr_t dest, int value,
>  		return NULL;
>  	}
>  
> -	xfer_count = len >> 2;
> -	if (xfer_count > ATC_BTSIZE_MAX) {
> -		dev_err(chan2dev(chan), "%s: buffer is too big\n",
> +	vaddr = dma_pool_alloc(atdma->memset_pool, GFP_ATOMIC, &paddr);

Mostly people use _zalloc variant, any reason why you don't want that

-- 
~Vinod

Re: [PATCH 1/2] dmaengine: hdmac: factorise memset descriptor allocation

[Maxime Ripard]

[-- Attachment #1: Type: text/plain, Size: 1851 bytes --]

Hi Vinod,

On Wed, Oct 07, 2015 at 02:54:13PM +0100, Vinod Koul wrote:
> On Thu, Oct 01, 2015 at 04:52:37PM +0200, Maxime Ripard wrote:
> 
> > +static struct at_desc *atc_create_memset_desc(struct dma_chan *chan,
> > +					      dma_addr_t psrc,
> > +					      dma_addr_t pdst,
> > +					      size_t len)
> > +{
> > +	struct at_dma_chan *atchan = to_at_dma_chan(chan);
> > +	struct at_desc *desc;
> > +	size_t xfer_count;
> > +
> > +	u32 ctrla = ATC_SRC_WIDTH(2) |
> > +		ATC_DST_WIDTH(2);
> 
> why is this over two lines :)
> 
> > +	u32 ctrlb = ATC_DEFAULT_CTRLB | ATC_IEN |
> > +		ATC_SRC_ADDR_MODE_FIXED |
> > +		ATC_DST_ADDR_MODE_INCR |
> > +		ATC_FC_MEM2MEM;
> > +
> > +	xfer_count = len >> 2;
> > +	if (xfer_count > ATC_BTSIZE_MAX) {
> > +		dev_err(chan2dev(chan), "%s: buffer is too big\n",
> > +			__func__);
> > +		return NULL;
> > +	}
> 
> This is fine, but this is driver limitation. We should really split the
> txn to multiple descriptors here..

Both these things are actually just copy from the previous code.

I can probably fix the first style issue in this patch, but splitting
the memset descriptors should go in a separate patch.

> 
> > @@ -914,46 +953,26 @@ atc_prep_dma_memset(struct dma_chan *chan, dma_addr_t dest, int value,
> >  		return NULL;
> >  	}
> >  
> > -	xfer_count = len >> 2;
> > -	if (xfer_count > ATC_BTSIZE_MAX) {
> > -		dev_err(chan2dev(chan), "%s: buffer is too big\n",
> > +	vaddr = dma_pool_alloc(atdma->memset_pool, GFP_ATOMIC, &paddr);
> 
> Mostly people use _zalloc variant, any reason why you don't want that

Not necessarily, but I don't need zeroed memory either, since I
overwrite the content right after the allocation.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: [PATCH 1/2] dmaengine: hdmac: factorise memset descriptor allocation

[Vinod Koul]

On Wed, Oct 07, 2015 at 06:33:17PM +0100, Maxime Ripard wrote:
> > This is fine, but this is driver limitation. We should really split the
> > txn to multiple descriptors here..
> 
> Both these things are actually just copy from the previous code.
> 
> I can probably fix the first style issue in this patch, but splitting
> the memset descriptors should go in a separate patch.

That sounds fine

-- 
~Vinod