RE: [PATCH v2] usb: cdns3: Optimize DMA request buffer allocation

[Sanket Parmar]

Hi Peter,

> On 21-03-17 20:13:59, 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 dma_alloc_noncoherent()
> > to allocate aligned request buffer of dynamic length.
> >
> > Reported-by: Aswath Govindraju 
> > Signed-off-by: Sanket Parmar 
> > ---
> >
> > Changelog:
> > v2:
> > - used dma_*_noncoherent() APIs
> > - changed the commit log
> >
> >  drivers/usb/cdns3/cdns3-gadget.c | 30 -
> -
> >  drivers/usb/cdns3/cdns3-gadget.h |  2 ++
> >  2 files changed, 26 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/usb/cdns3/cdns3-gadget.c b/drivers/usb/cdns3/cdns3-
> gadget.c
> > index 0b892a2..126087b 100644
> > --- a/drivers/usb/cdns3/cdns3-gadget.c
> > +++ b/drivers/usb/cdns3/cdns3-gadget.c
> > @@ -819,9 +819,15 @@ void cdns3_gadget_giveback(struct
> cdns3_endpoint *priv_ep,
> > priv_ep->dir);
> >
> > if ((priv_req->flags & REQUEST_UNALIGNED) &&
> > -   priv_ep->dir == USB_DIR_OUT && !request->status)
> > +   priv_ep->dir == USB_DIR_OUT && !request->status) {
> > +   /* Make DMA buffer CPU accessible */
> > +   dma_sync_single_for_cpu(priv_dev->sysdev,
> > +   priv_req->aligned_buf->dma,
> > +   priv_req->aligned_buf->size,
> > +   priv_req->aligned_buf->dir);
> > memcpy(request->buf, priv_req->aligned_buf->buf,
> >request->length);
> > +   }
> >
> > priv_req->flags &= ~(REQUEST_PENDING | REQUEST_UNALIGNED);
> > /* All TRBs have finished, clear the counter */
> > @@ -883,8 +889,8 @@ 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);
> > +   dma_free_noncoherent(priv_dev->sysdev, buf-
> >size,
> > + buf->buf, buf->dma, buf->dir);
> > kfree(buf);
> > spin_lock_irqsave(&priv_dev->lock, flags);
> > }
> > @@ -911,10 +917,13 @@ static int
> cdns3_prepare_aligned_request_buf(struct cdns3_request *priv_req)
> > return -ENOMEM;
> >
> > buf->size = priv_req->request.length;
> > +   buf->dir = usb_endpoint_dir_in(priv_ep->endpoint.desc) ?
> > +   DMA_TO_DEVICE : DMA_FROM_DEVICE;
> >
> > -   buf->buf = dma_alloc_coherent(priv_dev->sysdev,
> > +   buf->buf = dma_alloc_noncoherent(priv_dev->sysdev,
> >   buf->size,
> >   &buf->dma,
> > + buf->dir,
> >   GFP_ATOMIC);
> > if (!buf->buf) {
> > kfree(buf);
> > @@ -936,10 +945,18 @@ static int
> cdns3_prepare_aligned_request_buf(struct cdns3_request *priv_req)
> > }
> >
> > if (priv_ep->dir == USB_DIR_IN) {
> > +   /* Make DMA buffer CPU accessible */
> > +   dma_sync_single_for_cpu(priv_dev->sysdev,
> > +   buf->dma, buf->size, buf->dir);
> > memcpy(buf->buf, priv_req->request.buf,
> >priv_req->request.length);
> > }
> >
> > +   /* Transfer DMA buffer ownership back to device */
> > +   dma_sync_single_for_device(priv_dev->sysdev,
> > +   buf->dma, buf->size, buf->dir);
> > +
> > +
> 
> One more blank line.
> 
> Otherwise, it seems OK for me.

I have remove this blank line. New patch has been posted already.

> 
> > priv_req->flags |= REQUEST_UNALIGNED;
> > trace_cdns3_prepare_aligned_request(priv_req);
> >
> > @@ -3088,9 +3105,10 @@ 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,
> > +   dma_free_noncoherent(priv_dev->sysdev, buf->size,
> >   buf->buf,
> > - buf->dma);
> > + buf->dma,
> > + buf->dir);
> >
> > list_del(&buf->list);
> > 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 
> > +#include 
> >
> >  /*
> >   * USBSS-DEV register interface.
> > @@ -1205,6 +1206,7 @@ struct cdns3_a

Re: [PATCH v2] usb: cdns3: Optimize DMA request buffer allocation

[Peter Chen]

On 21-03-17 20:13:59, 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 dma_alloc_noncoherent()
> to allocate aligned request buffer of dynamic length.
> 
> Reported-by: Aswath Govindraju 
> Signed-off-by: Sanket Parmar 
> ---
> 
> Changelog:
> v2:
> - used dma_*_noncoherent() APIs
> - changed the commit log
> 
>  drivers/usb/cdns3/cdns3-gadget.c | 30 --
>  drivers/usb/cdns3/cdns3-gadget.h |  2 ++
>  2 files changed, 26 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/usb/cdns3/cdns3-gadget.c 
> b/drivers/usb/cdns3/cdns3-gadget.c
> index 0b892a2..126087b 100644
> --- a/drivers/usb/cdns3/cdns3-gadget.c
> +++ b/drivers/usb/cdns3/cdns3-gadget.c
> @@ -819,9 +819,15 @@ void cdns3_gadget_giveback(struct cdns3_endpoint 
> *priv_ep,
>   priv_ep->dir);
>  
>   if ((priv_req->flags & REQUEST_UNALIGNED) &&
> - priv_ep->dir == USB_DIR_OUT && !request->status)
> + priv_ep->dir == USB_DIR_OUT && !request->status) {
> + /* Make DMA buffer CPU accessible */
> + dma_sync_single_for_cpu(priv_dev->sysdev,
> + priv_req->aligned_buf->dma,
> + priv_req->aligned_buf->size,
> + priv_req->aligned_buf->dir);
>   memcpy(request->buf, priv_req->aligned_buf->buf,
>  request->length);
> + }
>  
>   priv_req->flags &= ~(REQUEST_PENDING | REQUEST_UNALIGNED);
>   /* All TRBs have finished, clear the counter */
> @@ -883,8 +889,8 @@ 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);
> + dma_free_noncoherent(priv_dev->sysdev, buf->size,
> +   buf->buf, buf->dma, buf->dir);
>   kfree(buf);
>   spin_lock_irqsave(&priv_dev->lock, flags);
>   }
> @@ -911,10 +917,13 @@ static int cdns3_prepare_aligned_request_buf(struct 
> cdns3_request *priv_req)
>   return -ENOMEM;
>  
>   buf->size = priv_req->request.length;
> + buf->dir = usb_endpoint_dir_in(priv_ep->endpoint.desc) ?
> + DMA_TO_DEVICE : DMA_FROM_DEVICE;
>  
> - buf->buf = dma_alloc_coherent(priv_dev->sysdev,
> + buf->buf = dma_alloc_noncoherent(priv_dev->sysdev,
> buf->size,
> &buf->dma,
> +   buf->dir,
> GFP_ATOMIC);
>   if (!buf->buf) {
>   kfree(buf);
> @@ -936,10 +945,18 @@ static int cdns3_prepare_aligned_request_buf(struct 
> cdns3_request *priv_req)
>   }
>  
>   if (priv_ep->dir == USB_DIR_IN) {
> + /* Make DMA buffer CPU accessible */
> + dma_sync_single_for_cpu(priv_dev->sysdev,
> + buf->dma, buf->size, buf->dir);
>   memcpy(buf->buf, priv_req->request.buf,
>  priv_req->request.length);
>   }
>  
> + /* Transfer DMA buffer ownership back to device */
> + dma_sync_single_for_device(priv_dev->sysdev,
> + buf->dma, buf->size, buf->dir);
> +
> +

One more blank line.

Otherwise, it seems OK for me.

>   priv_req->flags |= REQUEST_UNALIGNED;
>   trace_cdns3_prepare_aligned_request(priv_req);
>  
> @@ -3088,9 +3105,10 @@ 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,
> + dma_free_noncoherent(priv_dev->sysdev, buf->size,
> buf->buf,
> -   buf->dma);
> +   buf->dma,
> +   buf->dir);
>  
>   list_del(&buf->list);
>   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 
> +#include 
>  
>  /*
>   * 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;
>   uns

Re: [PATCH v2] usb: cdns3: Optimize DMA request buffer allocation

[Peter Chen]

On 21-03-18 07:32:45, Christoph Hellwig wrote:
> On Wed, Mar 17, 2021 at 08:13:59PM +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 dma_alloc_noncoherent()
> > to allocate aligned request buffer of dynamic length.
> > 
> > Reported-by: Aswath Govindraju 
> > Signed-off-by: Sanket Parmar 
> 
> Looks good to me:
> 
> Reviewed-by: Christoph Hellwig 

Hi Christoph,

I would like to confirm the dma_alloc_noncoherent allocates the memory
less than PAGE_SIZE if buffer size it would like to allocate is small
(eg, 64 bytes)? 

-- 

Thanks,
Peter Chen

Re: [PATCH v2] usb: cdns3: Optimize DMA request buffer allocation

[Christoph Hellwig]

On Wed, Mar 17, 2021 at 08:13:59PM +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 dma_alloc_noncoherent()
> to allocate aligned request buffer of dynamic length.
> 
> Reported-by: Aswath Govindraju 
> Signed-off-by: Sanket Parmar 

Looks good to me:

Reviewed-by: Christoph Hellwig