[PATCH 1/2] usb: cdns3: Use dma_pool_* api to alloc trb pool

[PATCH 1/2] usb: cdns3: Use dma_pool_* api to alloc trb pool

[Sanket Parmar]

Allocation of DMA coherent memory in atomic context using
dma_alloc_coherent() might fail on platforms with smaller
DMA region.

To fix it, dma_alloc_coherent() is replaced with dma_pool
API to allocate a smaller chunk of DMA coherent memory for
TRB rings.

Fixes: 7733f6c32e36 ("usb: cdns3: Add Cadence USB3 DRD Driver")
Reported-by: Aswath Govindraju <a-govindraju@ti.com>
Signed-off-by: Sanket Parmar <sparmar@cadence.com>
---
 drivers/usb/cdns3/cdns3-gadget.c |   42 +++++++++++++++++--------------------
 drivers/usb/cdns3/cdns3-gadget.h |    1 +
 2 files changed, 20 insertions(+), 23 deletions(-)

diff --git a/drivers/usb/cdns3/cdns3-gadget.c b/drivers/usb/cdns3/cdns3-gadget.c
index 582bfec..5f51215 100644
--- a/drivers/usb/cdns3/cdns3-gadget.c
+++ b/drivers/usb/cdns3/cdns3-gadget.c
@@ -59,6 +59,7 @@
 #include <linux/dma-mapping.h>
 #include <linux/usb/gadget.h>
 #include <linux/module.h>
+#include <linux/dmapool.h>
 #include <linux/iopoll.h>
 
 #include "core.h"
@@ -190,29 +191,13 @@ dma_addr_t cdns3_trb_virt_to_dma(struct cdns3_endpoint *priv_ep,
 	return priv_ep->trb_pool_dma + offset;
 }
 
-static int cdns3_ring_size(struct cdns3_endpoint *priv_ep)
-{
-	switch (priv_ep->type) {
-	case USB_ENDPOINT_XFER_ISOC:
-		return TRB_ISO_RING_SIZE;
-	case USB_ENDPOINT_XFER_CONTROL:
-		return TRB_CTRL_RING_SIZE;
-	default:
-		if (priv_ep->use_streams)
-			return TRB_STREAM_RING_SIZE;
-		else
-			return TRB_RING_SIZE;
-	}
-}
-
 static void cdns3_free_trb_pool(struct cdns3_endpoint *priv_ep)
 {
 	struct cdns3_device *priv_dev = priv_ep->cdns3_dev;
 
 	if (priv_ep->trb_pool) {
-		dma_free_coherent(priv_dev->sysdev,
-				  cdns3_ring_size(priv_ep),
-				  priv_ep->trb_pool, priv_ep->trb_pool_dma);
+		dma_pool_free(priv_dev->eps_dma_pool,
+			      priv_ep->trb_pool, priv_ep->trb_pool_dma);
 		priv_ep->trb_pool = NULL;
 	}
 }
@@ -226,7 +211,7 @@ static void cdns3_free_trb_pool(struct cdns3_endpoint *priv_ep)
 int cdns3_allocate_trb_pool(struct cdns3_endpoint *priv_ep)
 {
 	struct cdns3_device *priv_dev = priv_ep->cdns3_dev;
-	int ring_size = cdns3_ring_size(priv_ep);
+	int ring_size = TRB_RING_SIZE;
 	int num_trbs = ring_size / TRB_SIZE;
 	struct cdns3_trb *link_trb;
 
@@ -234,10 +219,10 @@ int cdns3_allocate_trb_pool(struct cdns3_endpoint *priv_ep)
 		cdns3_free_trb_pool(priv_ep);
 
 	if (!priv_ep->trb_pool) {
-		priv_ep->trb_pool = dma_alloc_coherent(priv_dev->sysdev,
-						       ring_size,
-						       &priv_ep->trb_pool_dma,
-						       GFP_DMA32 | GFP_ATOMIC);
+		priv_ep->trb_pool = dma_pool_alloc(priv_dev->eps_dma_pool,
+						   GFP_DMA32 | GFP_ATOMIC,
+						   &priv_ep->trb_pool_dma);
+
 		if (!priv_ep->trb_pool)
 			return -ENOMEM;
 
@@ -3113,6 +3098,7 @@ static void cdns3_gadget_exit(struct cdns *cdns)
 
 	dma_free_coherent(priv_dev->sysdev, 8, priv_dev->setup_buf,
 			  priv_dev->setup_dma);
+	dma_pool_destroy(priv_dev->eps_dma_pool);
 
 	kfree(priv_dev->zlp_buf);
 	usb_put_gadget(&priv_dev->gadget);
@@ -3185,6 +3171,14 @@ static int cdns3_gadget_start(struct cdns *cdns)
 	/* initialize endpoint container */
 	INIT_LIST_HEAD(&priv_dev->gadget.ep_list);
 	INIT_LIST_HEAD(&priv_dev->aligned_buf_list);
+	priv_dev->eps_dma_pool = dma_pool_create("cdns3_eps_dma_pool",
+						 priv_dev->sysdev,
+						 TRB_RING_SIZE, 8, 0);
+	if (!priv_dev->eps_dma_pool) {
+		dev_err(priv_dev->dev, "Failed to create TRB dma pool\n");
+		ret = -ENOMEM;
+		goto err1;
+	}
 
 	ret = cdns3_init_eps(priv_dev);
 	if (ret) {
@@ -3235,6 +3229,8 @@ static int cdns3_gadget_start(struct cdns *cdns)
 err2:
 	cdns3_free_all_eps(priv_dev);
 err1:
+	dma_pool_destroy(priv_dev->eps_dma_pool);
+
 	usb_put_gadget(&priv_dev->gadget);
 	cdns->gadget_dev = NULL;
 	return ret;
diff --git a/drivers/usb/cdns3/cdns3-gadget.h b/drivers/usb/cdns3/cdns3-gadget.h
index 21fa461..ecf9b91 100644
--- a/drivers/usb/cdns3/cdns3-gadget.h
+++ b/drivers/usb/cdns3/cdns3-gadget.h
@@ -1298,6 +1298,7 @@ struct cdns3_device {
 
 	struct cdns3_usb_regs		__iomem *regs;
 
+	struct dma_pool			*eps_dma_pool;
 	struct usb_ctrlrequest		*setup_buf;
 	dma_addr_t			setup_dma;
 	void				*zlp_buf;
-- 
1.7.1

[PATCH 2/2] usb: cdns3: Optimize DMA request buffer allocation

[Sanket Parmar]

dma_alloc_coherent() might fail on the platform with a small DMA region.

To avoid such failure in cdns3_prepare_aligned_request_buf(),
dma_alloc_coherent() is replaced with kmalloc and dma_map API to
allocate aligned request buffer of dynamic length.

Fixes: 7733f6c32e36 ("usb: cdns3: Add Cadence USB3 DRD Driver")
Reported-by: Aswath Govindraju <a-govindraju@ti.com>
Signed-off-by: Sanket Parmar <sparmar@cadence.com>
---
 drivers/usb/cdns3/cdns3-gadget.c |   73 +++++++++++++++++++++++++------------
 drivers/usb/cdns3/cdns3-gadget.h |    2 +
 2 files changed, 51 insertions(+), 24 deletions(-)

diff --git a/drivers/usb/cdns3/cdns3-gadget.c b/drivers/usb/cdns3/cdns3-gadget.c
index 5f51215..b4955ce 100644
--- a/drivers/usb/cdns3/cdns3-gadget.c
+++ b/drivers/usb/cdns3/cdns3-gadget.c
@@ -818,10 +818,26 @@ void cdns3_gadget_giveback(struct cdns3_endpoint *priv_ep,
 	usb_gadget_unmap_request_by_dev(priv_dev->sysdev, request,
 					priv_ep->dir);
 
-	if ((priv_req->flags & REQUEST_UNALIGNED) &&
-	    priv_ep->dir == USB_DIR_OUT && !request->status)
-		memcpy(request->buf, priv_req->aligned_buf->buf,
-		       request->length);
+	if ((priv_req->flags & REQUEST_UNALIGNED) && priv_req->aligned_buf) {
+		struct cdns3_aligned_buf *buf;
+
+		buf = priv_req->aligned_buf;
+		dma_unmap_single(priv_dev->sysdev, buf->dma, buf->size,
+			buf->dir);
+		priv_req->flags &= ~REQUEST_UNALIGNED;
+
+		if (priv_ep->dir == USB_DIR_OUT && !request->status) {
+			memcpy(request->buf, priv_req->aligned_buf->buf,
+			       request->length);
+		}
+
+		trace_cdns3_free_aligned_request(priv_req);
+		priv_req->aligned_buf->in_use = 0;
+		queue_work(system_freezable_wq,
+			   &priv_dev->aligned_buf_wq);
+		priv_req->aligned_buf = NULL;
+
+	}
 
 	priv_req->flags &= ~(REQUEST_PENDING | REQUEST_UNALIGNED);
 	/* All TRBs have finished, clear the counter */
@@ -883,8 +899,7 @@ static void cdns3_free_aligned_request_buf(struct work_struct *work)
 			 * interrupts.
 			 */
 			spin_unlock_irqrestore(&priv_dev->lock, flags);
-			dma_free_coherent(priv_dev->sysdev, buf->size,
-					  buf->buf, buf->dma);
+			kfree(buf->buf);
 			kfree(buf);
 			spin_lock_irqsave(&priv_dev->lock, flags);
 		}
@@ -910,27 +925,16 @@ static int cdns3_prepare_aligned_request_buf(struct cdns3_request *priv_req)
 		if (!buf)
 			return -ENOMEM;
 
-		buf->size = priv_req->request.length;
+		buf->size = usb_endpoint_dir_out(priv_ep->endpoint.desc) ?
+				usb_ep_align(&(priv_ep->endpoint), priv_req->request.length)
+				: priv_req->request.length;
 
-		buf->buf = dma_alloc_coherent(priv_dev->sysdev,
-					      buf->size,
-					      &buf->dma,
-					      GFP_ATOMIC);
+		buf->buf = kmalloc(buf->size, GFP_ATOMIC);
 		if (!buf->buf) {
 			kfree(buf);
 			return -ENOMEM;
 		}
 
-		if (priv_req->aligned_buf) {
-			trace_cdns3_free_aligned_request(priv_req);
-			priv_req->aligned_buf->in_use = 0;
-			queue_work(system_freezable_wq,
-				   &priv_dev->aligned_buf_wq);
-		}
-
-		buf->in_use = 1;
-		priv_req->aligned_buf = buf;
-
 		list_add_tail(&buf->list,
 			      &priv_dev->aligned_buf_list);
 	}
@@ -940,6 +944,27 @@ static int cdns3_prepare_aligned_request_buf(struct cdns3_request *priv_req)
 		       priv_req->request.length);
 	}
 
+	if (priv_req->aligned_buf) {
+		trace_cdns3_free_aligned_request(priv_req);
+		priv_req->aligned_buf->in_use = 0;
+		queue_work(system_freezable_wq,
+			   &priv_dev->aligned_buf_wq);
+	}
+
+	buf->dir =  priv_ep->dir ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
+	buf->in_use = 1;
+	priv_req->aligned_buf = buf;
+
+	buf->dma = dma_map_single(priv_dev->sysdev, buf->buf, buf->size,
+				buf->dir);
+
+	if (dma_mapping_error(priv_dev->sysdev, buf->dma)) {
+		dev_err(priv_dev->dev, "Failed to map buffer\n");
+		kfree(buf->buf);
+		kfree(buf);
+		return -EFAULT;
+	}
+
 	priv_req->flags |= REQUEST_UNALIGNED;
 	trace_cdns3_prepare_aligned_request(priv_req);
 
@@ -3088,11 +3113,11 @@ static void cdns3_gadget_exit(struct cdns *cdns)
 		struct cdns3_aligned_buf *buf;
 
 		buf = cdns3_next_align_buf(&priv_dev->aligned_buf_list);
-		dma_free_coherent(priv_dev->sysdev, buf->size,
-				  buf->buf,
-				  buf->dma);
+		dma_unmap_single(priv_dev->sysdev, buf->dma, buf->size,
+			buf->dir);
 
 		list_del(&buf->list);
+		kfree(buf->buf);
 		kfree(buf);
 	}
 
diff --git a/drivers/usb/cdns3/cdns3-gadget.h b/drivers/usb/cdns3/cdns3-gadget.h
index ecf9b91..c5660f2 100644
--- a/drivers/usb/cdns3/cdns3-gadget.h
+++ b/drivers/usb/cdns3/cdns3-gadget.h
@@ -12,6 +12,7 @@
 #ifndef __LINUX_CDNS3_GADGET
 #define __LINUX_CDNS3_GADGET
 #include <linux/usb/gadget.h>
+#include <linux/dma-direction.h>
 
 /*
  * USBSS-DEV register interface.
@@ -1205,6 +1206,7 @@ struct cdns3_aligned_buf {
 	void			*buf;
 	dma_addr_t		dma;
 	u32			size;
+	enum dma_data_direction dir;
 	unsigned		in_use:1;
 	struct list_head	list;
 };
-- 
1.7.1

Re: [PATCH 2/2] usb: cdns3: Optimize DMA request buffer allocation

[Christoph Hellwig]

On Tue, Mar 09, 2021 at 06:19:40AM +0100, Sanket Parmar wrote:
> dma_alloc_coherent() might fail on the platform with a small DMA region.
> 
> To avoid such failure in cdns3_prepare_aligned_request_buf(),
> dma_alloc_coherent() is replaced with kmalloc and dma_map API to
> allocate aligned request buffer of dynamic length.

dma_alloc_noncoherent is the proper API instead of using kmalloc, which
can lead to unaddressable memory that might require bounce buffering.

RE: [PATCH 2/2] usb: cdns3: Optimize DMA request buffer allocation

[Sanket Parmar]

> On Tue, Mar 09, 2021 at 06:19:40AM +0100, Sanket Parmar wrote:
> > dma_alloc_coherent() might fail on the platform with a small DMA region.
> >
> > To avoid such failure in cdns3_prepare_aligned_request_buf(),
> > dma_alloc_coherent() is replaced with kmalloc and dma_map API to
> > allocate aligned request buffer of dynamic length.
> 
> dma_alloc_noncoherent is the proper API instead of using kmalloc, which
> can lead to unaddressable memory that might require bounce buffering.

cdns3 device required DMA coherent buffer to perform operations. So 
dma_alloc_noncoherent will not help here.

Also all gadget classes(except g_ether) use kmalloc to allocated request buffer,
and device driver uses usb_gadget_map_request_by_dev to map the request
buffer. Similar approach is used to allocate aligned buffer. 

Thanks,
Sanket

Re: [PATCH 2/2] usb: cdns3: Optimize DMA request buffer allocation

[Christoph Hellwig]

On Tue, Mar 09, 2021 at 10:18:43AM +0000, Sanket Parmar wrote:
> > On Tue, Mar 09, 2021 at 06:19:40AM +0100, Sanket Parmar wrote:
> > > dma_alloc_coherent() might fail on the platform with a small DMA region.
> > >
> > > To avoid such failure in cdns3_prepare_aligned_request_buf(),
> > > dma_alloc_coherent() is replaced with kmalloc and dma_map API to
> > > allocate aligned request buffer of dynamic length.
> > 
> > dma_alloc_noncoherent is the proper API instead of using kmalloc, which
> > can lead to unaddressable memory that might require bounce buffering.
> 
> cdns3 device required DMA coherent buffer to perform operations. So 
> dma_alloc_noncoherent will not help here.
> 
> Also all gadget classes(except g_ether) use kmalloc to allocated request buffer,
> and device driver uses usb_gadget_map_request_by_dev to map the request
> buffer. Similar approach is used to allocate aligned buffer. 

If you can use kmalloc and dma_map_single you can use
dma_alloc_noncoherent per definition.

RE: [PATCH 2/2] usb: cdns3: Optimize DMA request buffer allocation

[Sanket Parmar]

> On Tue, Mar 09, 2021 at 10:18:43AM +0000, Sanket Parmar wrote:
> > > On Tue, Mar 09, 2021 at 06:19:40AM +0100, Sanket Parmar wrote:
> > > > dma_alloc_coherent() might fail on the platform with a small DMA
> region.
> > > >
> > > > To avoid such failure in cdns3_prepare_aligned_request_buf(),
> > > > dma_alloc_coherent() is replaced with kmalloc and dma_map API to
> > > > allocate aligned request buffer of dynamic length.
> > >
> > > dma_alloc_noncoherent is the proper API instead of using kmalloc, which
> > > can lead to unaddressable memory that might require bounce buffering.
> >
> > cdns3 device required DMA coherent buffer to perform operations. So
> > dma_alloc_noncoherent will not help here.
> >
> > Also all gadget classes(except g_ether) use kmalloc to allocated request
> buffer,
> > and device driver uses usb_gadget_map_request_by_dev to map the
> request
> > buffer. Similar approach is used to allocate aligned buffer.
> 
> If you can use kmalloc and dma_map_single you can use
> dma_alloc_noncoherent per definition.

Okay. I was not aware of it. I will test it. 
Thank you for your feedback.

--
Sanket

Re: [PATCH 1/2] usb: cdns3: Use dma_pool_* api to alloc trb pool

[Peter Chen]

On 21-03-09 06:19:39, Sanket Parmar wrote:
> Allocation of DMA coherent memory in atomic context using
> dma_alloc_coherent() might fail on platforms with smaller
> DMA region.
> 
> To fix it, dma_alloc_coherent() is replaced with dma_pool
> API to allocate a smaller chunk of DMA coherent memory for
> TRB rings.
> 
> Fixes: 7733f6c32e36 ("usb: cdns3: Add Cadence USB3 DRD Driver")

The patch is ok, but I don't think it is a bug-fix, it is an
improvement for smaller DMA region use case.

I will apply it with deletion of above Fixes tag if you have no
more opinion.

Peter
> Reported-by: Aswath Govindraju <a-govindraju@ti.com>
> Signed-off-by: Sanket Parmar <sparmar@cadence.com>
> ---
>  drivers/usb/cdns3/cdns3-gadget.c |   42 +++++++++++++++++--------------------
>  drivers/usb/cdns3/cdns3-gadget.h |    1 +
>  2 files changed, 20 insertions(+), 23 deletions(-)
> 
> diff --git a/drivers/usb/cdns3/cdns3-gadget.c b/drivers/usb/cdns3/cdns3-gadget.c
> index 582bfec..5f51215 100644
> --- a/drivers/usb/cdns3/cdns3-gadget.c
> +++ b/drivers/usb/cdns3/cdns3-gadget.c
> @@ -59,6 +59,7 @@
>  #include <linux/dma-mapping.h>
>  #include <linux/usb/gadget.h>
>  #include <linux/module.h>
> +#include <linux/dmapool.h>
>  #include <linux/iopoll.h>
>  
>  #include "core.h"
> @@ -190,29 +191,13 @@ dma_addr_t cdns3_trb_virt_to_dma(struct cdns3_endpoint *priv_ep,
>  	return priv_ep->trb_pool_dma + offset;
>  }
>  
> -static int cdns3_ring_size(struct cdns3_endpoint *priv_ep)
> -{
> -	switch (priv_ep->type) {
> -	case USB_ENDPOINT_XFER_ISOC:
> -		return TRB_ISO_RING_SIZE;
> -	case USB_ENDPOINT_XFER_CONTROL:
> -		return TRB_CTRL_RING_SIZE;
> -	default:
> -		if (priv_ep->use_streams)
> -			return TRB_STREAM_RING_SIZE;
> -		else
> -			return TRB_RING_SIZE;
> -	}
> -}
> -
>  static void cdns3_free_trb_pool(struct cdns3_endpoint *priv_ep)
>  {
>  	struct cdns3_device *priv_dev = priv_ep->cdns3_dev;
>  
>  	if (priv_ep->trb_pool) {
> -		dma_free_coherent(priv_dev->sysdev,
> -				  cdns3_ring_size(priv_ep),
> -				  priv_ep->trb_pool, priv_ep->trb_pool_dma);
> +		dma_pool_free(priv_dev->eps_dma_pool,
> +			      priv_ep->trb_pool, priv_ep->trb_pool_dma);
>  		priv_ep->trb_pool = NULL;
>  	}
>  }
> @@ -226,7 +211,7 @@ static void cdns3_free_trb_pool(struct cdns3_endpoint *priv_ep)
>  int cdns3_allocate_trb_pool(struct cdns3_endpoint *priv_ep)
>  {
>  	struct cdns3_device *priv_dev = priv_ep->cdns3_dev;
> -	int ring_size = cdns3_ring_size(priv_ep);
> +	int ring_size = TRB_RING_SIZE;
>  	int num_trbs = ring_size / TRB_SIZE;
>  	struct cdns3_trb *link_trb;
>  
> @@ -234,10 +219,10 @@ int cdns3_allocate_trb_pool(struct cdns3_endpoint *priv_ep)
>  		cdns3_free_trb_pool(priv_ep);
>  
>  	if (!priv_ep->trb_pool) {
> -		priv_ep->trb_pool = dma_alloc_coherent(priv_dev->sysdev,
> -						       ring_size,
> -						       &priv_ep->trb_pool_dma,
> -						       GFP_DMA32 | GFP_ATOMIC);
> +		priv_ep->trb_pool = dma_pool_alloc(priv_dev->eps_dma_pool,
> +						   GFP_DMA32 | GFP_ATOMIC,
> +						   &priv_ep->trb_pool_dma);
> +
>  		if (!priv_ep->trb_pool)
>  			return -ENOMEM;
>  
> @@ -3113,6 +3098,7 @@ static void cdns3_gadget_exit(struct cdns *cdns)
>  
>  	dma_free_coherent(priv_dev->sysdev, 8, priv_dev->setup_buf,
>  			  priv_dev->setup_dma);
> +	dma_pool_destroy(priv_dev->eps_dma_pool);
>  
>  	kfree(priv_dev->zlp_buf);
>  	usb_put_gadget(&priv_dev->gadget);
> @@ -3185,6 +3171,14 @@ static int cdns3_gadget_start(struct cdns *cdns)
>  	/* initialize endpoint container */
>  	INIT_LIST_HEAD(&priv_dev->gadget.ep_list);
>  	INIT_LIST_HEAD(&priv_dev->aligned_buf_list);
> +	priv_dev->eps_dma_pool = dma_pool_create("cdns3_eps_dma_pool",
> +						 priv_dev->sysdev,
> +						 TRB_RING_SIZE, 8, 0);
> +	if (!priv_dev->eps_dma_pool) {
> +		dev_err(priv_dev->dev, "Failed to create TRB dma pool\n");
> +		ret = -ENOMEM;
> +		goto err1;
> +	}
>  
>  	ret = cdns3_init_eps(priv_dev);
>  	if (ret) {
> @@ -3235,6 +3229,8 @@ static int cdns3_gadget_start(struct cdns *cdns)
>  err2:
>  	cdns3_free_all_eps(priv_dev);
>  err1:
> +	dma_pool_destroy(priv_dev->eps_dma_pool);
> +
>  	usb_put_gadget(&priv_dev->gadget);
>  	cdns->gadget_dev = NULL;
>  	return ret;
> diff --git a/drivers/usb/cdns3/cdns3-gadget.h b/drivers/usb/cdns3/cdns3-gadget.h
> index 21fa461..ecf9b91 100644
> --- a/drivers/usb/cdns3/cdns3-gadget.h
> +++ b/drivers/usb/cdns3/cdns3-gadget.h
> @@ -1298,6 +1298,7 @@ struct cdns3_device {
>  
>  	struct cdns3_usb_regs		__iomem *regs;
>  
> +	struct dma_pool			*eps_dma_pool;
>  	struct usb_ctrlrequest		*setup_buf;
>  	dma_addr_t			setup_dma;
>  	void				*zlp_buf;
> -- 
> 1.7.1
> 

-- 

Thanks,
Peter Chen

Re: [PATCH 2/2] usb: cdns3: Optimize DMA request buffer allocation

[Peter Chen]

On 21-03-09 06:19:40, Sanket Parmar wrote:
> dma_alloc_coherent() might fail on the platform with a small DMA region.
> 
> To avoid such failure in cdns3_prepare_aligned_request_buf(),
> dma_alloc_coherent() is replaced with kmalloc and dma_map API to
> allocate aligned request buffer of dynamic length.
> 
> Fixes: 7733f6c32e36 ("usb: cdns3: Add Cadence USB3 DRD Driver")

The comment with the 1st patch, it is not a bug-fix.

> Reported-by: Aswath Govindraju <a-govindraju@ti.com>
> Signed-off-by: Sanket Parmar <sparmar@cadence.com>
> ---
>  drivers/usb/cdns3/cdns3-gadget.c |   73 +++++++++++++++++++++++++------------
>  drivers/usb/cdns3/cdns3-gadget.h |    2 +
>  2 files changed, 51 insertions(+), 24 deletions(-)
> 
> diff --git a/drivers/usb/cdns3/cdns3-gadget.c b/drivers/usb/cdns3/cdns3-gadget.c
> index 5f51215..b4955ce 100644
> --- a/drivers/usb/cdns3/cdns3-gadget.c
> +++ b/drivers/usb/cdns3/cdns3-gadget.c
> @@ -818,10 +818,26 @@ void cdns3_gadget_giveback(struct cdns3_endpoint *priv_ep,
>  	usb_gadget_unmap_request_by_dev(priv_dev->sysdev, request,
>  					priv_ep->dir);
>  
> -	if ((priv_req->flags & REQUEST_UNALIGNED) &&
> -	    priv_ep->dir == USB_DIR_OUT && !request->status)
> -		memcpy(request->buf, priv_req->aligned_buf->buf,
> -		       request->length);
> +	if ((priv_req->flags & REQUEST_UNALIGNED) && priv_req->aligned_buf) {
> +		struct cdns3_aligned_buf *buf;
> +
> +		buf = priv_req->aligned_buf;
> +		dma_unmap_single(priv_dev->sysdev, buf->dma, buf->size,
> +			buf->dir);
> +		priv_req->flags &= ~REQUEST_UNALIGNED;
> +
> +		if (priv_ep->dir == USB_DIR_OUT && !request->status) {
> +			memcpy(request->buf, priv_req->aligned_buf->buf,
> +			       request->length);
> +		}
> +
> +		trace_cdns3_free_aligned_request(priv_req);
> +		priv_req->aligned_buf->in_use = 0;
> +		queue_work(system_freezable_wq,
> +			   &priv_dev->aligned_buf_wq);
> +		priv_req->aligned_buf = NULL;
> +
> +	}
>  
>  	priv_req->flags &= ~(REQUEST_PENDING | REQUEST_UNALIGNED);
>  	/* All TRBs have finished, clear the counter */
> @@ -883,8 +899,7 @@ static void cdns3_free_aligned_request_buf(struct work_struct *work)
>  			 * interrupts.
>  			 */
>  			spin_unlock_irqrestore(&priv_dev->lock, flags);
> -			dma_free_coherent(priv_dev->sysdev, buf->size,
> -					  buf->buf, buf->dma);
> +			kfree(buf->buf);
>  			kfree(buf);
>  			spin_lock_irqsave(&priv_dev->lock, flags);
>  		}
> @@ -910,27 +925,16 @@ static int cdns3_prepare_aligned_request_buf(struct cdns3_request *priv_req)
>  		if (!buf)
>  			return -ENOMEM;
>  
> -		buf->size = priv_req->request.length;
> +		buf->size = usb_endpoint_dir_out(priv_ep->endpoint.desc) ?
> +				usb_ep_align(&(priv_ep->endpoint), priv_req->request.length)
> +				: priv_req->request.length;
>  
> -		buf->buf = dma_alloc_coherent(priv_dev->sysdev,
> -					      buf->size,
> -					      &buf->dma,
> -					      GFP_ATOMIC);
> +		buf->buf = kmalloc(buf->size, GFP_ATOMIC);
>  		if (!buf->buf) {
>  			kfree(buf);
>  			return -ENOMEM;
>  		}
>  
> -		if (priv_req->aligned_buf) {
> -			trace_cdns3_free_aligned_request(priv_req);
> -			priv_req->aligned_buf->in_use = 0;
> -			queue_work(system_freezable_wq,
> -				   &priv_dev->aligned_buf_wq);
> -		}
> -
> -		buf->in_use = 1;
> -		priv_req->aligned_buf = buf;
> -
>  		list_add_tail(&buf->list,
>  			      &priv_dev->aligned_buf_list);
>  	}
> @@ -940,6 +944,27 @@ static int cdns3_prepare_aligned_request_buf(struct cdns3_request *priv_req)
>  		       priv_req->request.length);
>  	}
>  
> +	if (priv_req->aligned_buf) {
> +		trace_cdns3_free_aligned_request(priv_req);
> +		priv_req->aligned_buf->in_use = 0;
> +		queue_work(system_freezable_wq,
> +			   &priv_dev->aligned_buf_wq);

@Pawel, do you remember when this condition is met?

> +	}
> +
> +	buf->dir =  priv_ep->dir ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
> +	buf->in_use = 1;
> +	priv_req->aligned_buf = buf;
> +
> +	buf->dma = dma_map_single(priv_dev->sysdev, buf->buf, buf->size,
> +				buf->dir);
> +
> +	if (dma_mapping_error(priv_dev->sysdev, buf->dma)) {
> +		dev_err(priv_dev->dev, "Failed to map buffer\n");
> +		kfree(buf->buf);
> +		kfree(buf);
> +		return -EFAULT;
> +	}
> +
>  	priv_req->flags |= REQUEST_UNALIGNED;
>  	trace_cdns3_prepare_aligned_request(priv_req);
>  
> @@ -3088,11 +3113,11 @@ static void cdns3_gadget_exit(struct cdns *cdns)
>  		struct cdns3_aligned_buf *buf;
>  
>  		buf = cdns3_next_align_buf(&priv_dev->aligned_buf_list);
> -		dma_free_coherent(priv_dev->sysdev, buf->size,
> -				  buf->buf,
> -				  buf->dma);
> +		dma_unmap_single(priv_dev->sysdev, buf->dma, buf->size,
> +			buf->dir);

It only needs to DMA unmap after DMA has completed, this buf will not be
used, otherwise, the kfree below will cause issue.

>  
>  		list_del(&buf->list);
> +		kfree(buf->buf);
>  		kfree(buf);
>  	}
>  
> diff --git a/drivers/usb/cdns3/cdns3-gadget.h b/drivers/usb/cdns3/cdns3-gadget.h
> index ecf9b91..c5660f2 100644
> --- a/drivers/usb/cdns3/cdns3-gadget.h
> +++ b/drivers/usb/cdns3/cdns3-gadget.h
> @@ -12,6 +12,7 @@
>  #ifndef __LINUX_CDNS3_GADGET
>  #define __LINUX_CDNS3_GADGET
>  #include <linux/usb/gadget.h>
> +#include <linux/dma-direction.h>
>  
>  /*
>   * USBSS-DEV register interface.
> @@ -1205,6 +1206,7 @@ struct cdns3_aligned_buf {
>  	void			*buf;
>  	dma_addr_t		dma;
>  	u32			size;
> +	enum dma_data_direction dir;
>  	unsigned		in_use:1;
>  	struct list_head	list;
>  };
> -- 
> 1.7.1
> 

-- 

Thanks,
Peter Chen

RE: [PATCH 1/2] usb: cdns3: Use dma_pool_* api to alloc trb pool

[Sanket Parmar]

Hi Peter,
> 
> On 21-03-09 06:19:39, Sanket Parmar wrote:
> > Allocation of DMA coherent memory in atomic context using
> > dma_alloc_coherent() might fail on platforms with smaller
> > DMA region.
> >
> > To fix it, dma_alloc_coherent() is replaced with dma_pool
> > API to allocate a smaller chunk of DMA coherent memory for
> > TRB rings.
> >
> > Fixes: 7733f6c32e36 ("usb: cdns3: Add Cadence USB3 DRD Driver")
> 
> The patch is ok, but I don't think it is a bug-fix, it is an
> improvement for smaller DMA region use case.
> 
> I will apply it with deletion of above Fixes tag if you have no
> more opinion.
> 

Please go ahead.

Thanks,
Sanket

> Peter
> > Reported-by: Aswath Govindraju <a-govindraju@ti.com>
> > Signed-off-by: Sanket Parmar <sparmar@cadence.com>
> > ---
> >  drivers/usb/cdns3/cdns3-gadget.c |   42 +++++++++++++++++---------------
> -----
> >  drivers/usb/cdns3/cdns3-gadget.h |    1 +
> >  2 files changed, 20 insertions(+), 23 deletions(-)
> >
> > diff --git a/drivers/usb/cdns3/cdns3-gadget.c b/drivers/usb/cdns3/cdns3-
> gadget.c
> > index 582bfec..5f51215 100644
> > --- a/drivers/usb/cdns3/cdns3-gadget.c
> > +++ b/drivers/usb/cdns3/cdns3-gadget.c
> > @@ -59,6 +59,7 @@
> >  #include <linux/dma-mapping.h>
> >  #include <linux/usb/gadget.h>
> >  #include <linux/module.h>
> > +#include <linux/dmapool.h>
> >  #include <linux/iopoll.h>
> >
> >  #include "core.h"
> > @@ -190,29 +191,13 @@ dma_addr_t cdns3_trb_virt_to_dma(struct
> cdns3_endpoint *priv_ep,
> >  	return priv_ep->trb_pool_dma + offset;
> >  }
> >
> > -static int cdns3_ring_size(struct cdns3_endpoint *priv_ep)
> > -{
> > -	switch (priv_ep->type) {
> > -	case USB_ENDPOINT_XFER_ISOC:
> > -		return TRB_ISO_RING_SIZE;
> > -	case USB_ENDPOINT_XFER_CONTROL:
> > -		return TRB_CTRL_RING_SIZE;
> > -	default:
> > -		if (priv_ep->use_streams)
> > -			return TRB_STREAM_RING_SIZE;
> > -		else
> > -			return TRB_RING_SIZE;
> > -	}
> > -}
> > -
> >  static void cdns3_free_trb_pool(struct cdns3_endpoint *priv_ep)
> >  {
> >  	struct cdns3_device *priv_dev = priv_ep->cdns3_dev;
> >
> >  	if (priv_ep->trb_pool) {
> > -		dma_free_coherent(priv_dev->sysdev,
> > -				  cdns3_ring_size(priv_ep),
> > -				  priv_ep->trb_pool, priv_ep-
> >trb_pool_dma);
> > +		dma_pool_free(priv_dev->eps_dma_pool,
> > +			      priv_ep->trb_pool, priv_ep->trb_pool_dma);
> >  		priv_ep->trb_pool = NULL;
> >  	}
> >  }
> > @@ -226,7 +211,7 @@ static void cdns3_free_trb_pool(struct
> cdns3_endpoint *priv_ep)
> >  int cdns3_allocate_trb_pool(struct cdns3_endpoint *priv_ep)
> >  {
> >  	struct cdns3_device *priv_dev = priv_ep->cdns3_dev;
> > -	int ring_size = cdns3_ring_size(priv_ep);
> > +	int ring_size = TRB_RING_SIZE;
> >  	int num_trbs = ring_size / TRB_SIZE;
> >  	struct cdns3_trb *link_trb;
> >
> > @@ -234,10 +219,10 @@ int cdns3_allocate_trb_pool(struct
> cdns3_endpoint *priv_ep)
> >  		cdns3_free_trb_pool(priv_ep);
> >
> >  	if (!priv_ep->trb_pool) {
> > -		priv_ep->trb_pool = dma_alloc_coherent(priv_dev->sysdev,
> > -						       ring_size,
> > -						       &priv_ep-
> >trb_pool_dma,
> > -						       GFP_DMA32 |
> GFP_ATOMIC);
> > +		priv_ep->trb_pool = dma_pool_alloc(priv_dev-
> >eps_dma_pool,
> > +						   GFP_DMA32 |
> GFP_ATOMIC,
> > +						   &priv_ep->trb_pool_dma);
> > +
> >  		if (!priv_ep->trb_pool)
> >  			return -ENOMEM;
> >
> > @@ -3113,6 +3098,7 @@ static void cdns3_gadget_exit(struct cdns *cdns)
> >
> >  	dma_free_coherent(priv_dev->sysdev, 8, priv_dev->setup_buf,
> >  			  priv_dev->setup_dma);
> > +	dma_pool_destroy(priv_dev->eps_dma_pool);
> >
> >  	kfree(priv_dev->zlp_buf);
> >  	usb_put_gadget(&priv_dev->gadget);
> > @@ -3185,6 +3171,14 @@ static int cdns3_gadget_start(struct cdns *cdns)
> >  	/* initialize endpoint container */
> >  	INIT_LIST_HEAD(&priv_dev->gadget.ep_list);
> >  	INIT_LIST_HEAD(&priv_dev->aligned_buf_list);
> > +	priv_dev->eps_dma_pool =
> dma_pool_create("cdns3_eps_dma_pool",
> > +						 priv_dev->sysdev,
> > +						 TRB_RING_SIZE, 8, 0);
> > +	if (!priv_dev->eps_dma_pool) {
> > +		dev_err(priv_dev->dev, "Failed to create TRB dma pool\n");
> > +		ret = -ENOMEM;
> > +		goto err1;
> > +	}
> >
> >  	ret = cdns3_init_eps(priv_dev);
> >  	if (ret) {
> > @@ -3235,6 +3229,8 @@ static int cdns3_gadget_start(struct cdns *cdns)
> >  err2:
> >  	cdns3_free_all_eps(priv_dev);
> >  err1:
> > +	dma_pool_destroy(priv_dev->eps_dma_pool);
> > +
> >  	usb_put_gadget(&priv_dev->gadget);
> >  	cdns->gadget_dev = NULL;
> >  	return ret;
> > diff --git a/drivers/usb/cdns3/cdns3-gadget.h b/drivers/usb/cdns3/cdns3-
> gadget.h
> > index 21fa461..ecf9b91 100644
> > --- a/drivers/usb/cdns3/cdns3-gadget.h
> > +++ b/drivers/usb/cdns3/cdns3-gadget.h
> > @@ -1298,6 +1298,7 @@ struct cdns3_device {
> >
> >  	struct cdns3_usb_regs		__iomem *regs;
> >
> > +	struct dma_pool			*eps_dma_pool;
> >  	struct usb_ctrlrequest		*setup_buf;
> >  	dma_addr_t			setup_dma;
> >  	void				*zlp_buf;
> > --
> > 1.7.1
> >
> 
> --
> 
> Thanks,
> Peter Chen

RE: [PATCH 2/2] usb: cdns3: Optimize DMA request buffer allocation

[Sanket Parmar]

> 
> On 21-03-09 06:19:40, Sanket Parmar wrote:
> > dma_alloc_coherent() might fail on the platform with a small DMA region.
> >
> > To avoid such failure in cdns3_prepare_aligned_request_buf(),
> > dma_alloc_coherent() is replaced with kmalloc and dma_map API to
> > allocate aligned request buffer of dynamic length.
> >
> > Fixes: 7733f6c32e36 ("usb: cdns3: Add Cadence USB3 DRD Driver")
> 
> The comment with the 1st patch, it is not a bug-fix.

I will remove this. 

> 
> > Reported-by: Aswath Govindraju <a-govindraju@ti.com>
> > Signed-off-by: Sanket Parmar <sparmar@cadence.com>
> > ---
> >  drivers/usb/cdns3/cdns3-gadget.c |   73 +++++++++++++++++++++++++--
> ----------
> >  drivers/usb/cdns3/cdns3-gadget.h |    2 +
> >  2 files changed, 51 insertions(+), 24 deletions(-)
> >
> > diff --git a/drivers/usb/cdns3/cdns3-gadget.c b/drivers/usb/cdns3/cdns3-
> gadget.c
> > index 5f51215..b4955ce 100644
> > --- a/drivers/usb/cdns3/cdns3-gadget.c
> > +++ b/drivers/usb/cdns3/cdns3-gadget.c
> > @@ -818,10 +818,26 @@ void cdns3_gadget_giveback(struct
> cdns3_endpoint *priv_ep,
> >  	usb_gadget_unmap_request_by_dev(priv_dev->sysdev, request,
> >  					priv_ep->dir);
> >
> > -	if ((priv_req->flags & REQUEST_UNALIGNED) &&
> > -	    priv_ep->dir == USB_DIR_OUT && !request->status)
> > -		memcpy(request->buf, priv_req->aligned_buf->buf,
> > -		       request->length);
> > +	if ((priv_req->flags & REQUEST_UNALIGNED) && priv_req-
> >aligned_buf) {
> > +		struct cdns3_aligned_buf *buf;
> > +
> > +		buf = priv_req->aligned_buf;
> > +		dma_unmap_single(priv_dev->sysdev, buf->dma, buf->size,
> > +			buf->dir);
> > +		priv_req->flags &= ~REQUEST_UNALIGNED;
> > +
> > +		if (priv_ep->dir == USB_DIR_OUT && !request->status) {
> > +			memcpy(request->buf, priv_req->aligned_buf->buf,
> > +			       request->length);
> > +		}
> > +
> > +		trace_cdns3_free_aligned_request(priv_req);
> > +		priv_req->aligned_buf->in_use = 0;
> > +		queue_work(system_freezable_wq,
> > +			   &priv_dev->aligned_buf_wq);
> > +		priv_req->aligned_buf = NULL;
> > +
> > +	}
> >
> >  	priv_req->flags &= ~(REQUEST_PENDING | REQUEST_UNALIGNED);
> >  	/* All TRBs have finished, clear the counter */
> > @@ -883,8 +899,7 @@ static void cdns3_free_aligned_request_buf(struct
> work_struct *work)
> >  			 * interrupts.
> >  			 */
> >  			spin_unlock_irqrestore(&priv_dev->lock, flags);
> > -			dma_free_coherent(priv_dev->sysdev, buf->size,
> > -					  buf->buf, buf->dma);
> > +			kfree(buf->buf);
> >  			kfree(buf);
> >  			spin_lock_irqsave(&priv_dev->lock, flags);
> >  		}
> > @@ -910,27 +925,16 @@ static int
> cdns3_prepare_aligned_request_buf(struct cdns3_request *priv_req)
> >  		if (!buf)
> >  			return -ENOMEM;
> >
> > -		buf->size = priv_req->request.length;
> > +		buf->size = usb_endpoint_dir_out(priv_ep->endpoint.desc)
> ?
> > +				usb_ep_align(&(priv_ep->endpoint),
> priv_req->request.length)
> > +				: priv_req->request.length;
> >
> > -		buf->buf = dma_alloc_coherent(priv_dev->sysdev,
> > -					      buf->size,
> > -					      &buf->dma,
> > -					      GFP_ATOMIC);
> > +		buf->buf = kmalloc(buf->size, GFP_ATOMIC);
> >  		if (!buf->buf) {
> >  			kfree(buf);
> >  			return -ENOMEM;
> >  		}
> >
> > -		if (priv_req->aligned_buf) {
> > -			trace_cdns3_free_aligned_request(priv_req);
> > -			priv_req->aligned_buf->in_use = 0;
> > -			queue_work(system_freezable_wq,
> > -				   &priv_dev->aligned_buf_wq);
> > -		}
> > -
> > -		buf->in_use = 1;
> > -		priv_req->aligned_buf = buf;
> > -
> >  		list_add_tail(&buf->list,
> >  			      &priv_dev->aligned_buf_list);
> >  	}
> > @@ -940,6 +944,27 @@ static int
> cdns3_prepare_aligned_request_buf(struct cdns3_request *priv_req)
> >  		       priv_req->request.length);
> >  	}
> >
> > +	if (priv_req->aligned_buf) {
> > +		trace_cdns3_free_aligned_request(priv_req);
> > +		priv_req->aligned_buf->in_use = 0;
> > +		queue_work(system_freezable_wq,
> > +			   &priv_dev->aligned_buf_wq);
> 
> @Pawel, do you remember when this condition is met?
> 
> > +	}
> > +
> > +	buf->dir =  priv_ep->dir ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
> > +	buf->in_use = 1;
> > +	priv_req->aligned_buf = buf;
> > +
> > +	buf->dma = dma_map_single(priv_dev->sysdev, buf->buf, buf-
> >size,
> > +				buf->dir);
> > +
> > +	if (dma_mapping_error(priv_dev->sysdev, buf->dma)) {
> > +		dev_err(priv_dev->dev, "Failed to map buffer\n");
> > +		kfree(buf->buf);
> > +		kfree(buf);
> > +		return -EFAULT;
> > +	}
> > +
> >  	priv_req->flags |= REQUEST_UNALIGNED;
> >  	trace_cdns3_prepare_aligned_request(priv_req);
> >
> > @@ -3088,11 +3113,11 @@ static void cdns3_gadget_exit(struct cdns
> *cdns)
> >  		struct cdns3_aligned_buf *buf;
> >
> >  		buf = cdns3_next_align_buf(&priv_dev->aligned_buf_list);
> > -		dma_free_coherent(priv_dev->sysdev, buf->size,
> > -				  buf->buf,
> > -				  buf->dma);
> > +		dma_unmap_single(priv_dev->sysdev, buf->dma, buf->size,
> > +			buf->dir);
> 
> It only needs to DMA unmap after DMA has completed, this buf will not be
> used, otherwise, the kfree below will cause issue.

This part is not clear.  Aligned DMA buffer is allocated and mapped in cdns3_prepare_aligned_request_buf()
and put into aligned_buf_list. While unloading the gadget, We need to undo the same if aligned_buf_list is not
empty.  Am I missing something here? 

Also, I will post v2 of this patch which uses dma_*_noncoherent APIs suggested by Christoph Hellwig.
 
> 
> >
> >  		list_del(&buf->list);
> > +		kfree(buf->buf);
> >  		kfree(buf);
> >  	}
> >
> > diff --git a/drivers/usb/cdns3/cdns3-gadget.h b/drivers/usb/cdns3/cdns3-
> gadget.h
> > index ecf9b91..c5660f2 100644
> > --- a/drivers/usb/cdns3/cdns3-gadget.h
> > +++ b/drivers/usb/cdns3/cdns3-gadget.h
> > @@ -12,6 +12,7 @@
> >  #ifndef __LINUX_CDNS3_GADGET
> >  #define __LINUX_CDNS3_GADGET
> >  #include <linux/usb/gadget.h>
> > +#include <linux/dma-direction.h>
> >
> >  /*
> >   * USBSS-DEV register interface.
> > @@ -1205,6 +1206,7 @@ struct cdns3_aligned_buf {
> >  	void			*buf;
> >  	dma_addr_t		dma;
> >  	u32			size;
> > +	enum dma_data_direction dir;
> >  	unsigned		in_use:1;
> >  	struct list_head	list;
> >  };
> > --
> > 1.7.1
> >
> 
> --
> 
> Thanks,
> Peter Chen


Thanks,
Sanket

Re: [PATCH 2/2] usb: cdns3: Optimize DMA request buffer allocation

[Peter Chen]

On 21-03-15 15:51:04, Sanket Parmar wrote:
> > > +
> > >  	priv_req->flags |= REQUEST_UNALIGNED;
> > >  	trace_cdns3_prepare_aligned_request(priv_req);
> > >
> > > @@ -3088,11 +3113,11 @@ static void cdns3_gadget_exit(struct cdns
> > *cdns)
> > >  		struct cdns3_aligned_buf *buf;
> > >
> > >  		buf = cdns3_next_align_buf(&priv_dev->aligned_buf_list);
> > > -		dma_free_coherent(priv_dev->sysdev, buf->size,
> > > -				  buf->buf,
> > > -				  buf->dma);
> > > +		dma_unmap_single(priv_dev->sysdev, buf->dma, buf->size,
> > > +			buf->dir);
> > 
> > It only needs to DMA unmap after DMA has completed, this buf will not be
> > used, otherwise, the kfree below will cause issue.
> 
> This part is not clear.  Aligned DMA buffer is allocated and mapped in cdns3_prepare_aligned_request_buf()
> and put into aligned_buf_list. While unloading the gadget, We need to undo the same if aligned_buf_list is not
> empty.  Am I missing something here? 

My point is this unmap operation is useless since there is no user for
aligned buf, and it calls kfree afterwards. You could also keep it as it has
no harm.

> 
> Also, I will post v2 of this patch which uses dma_*_noncoherent APIs suggested by Christoph Hellwig.

-- 

Thanks,
Peter Chen