On architectures where there is no coherent caching such as ARM it's
proved that using dma_alloc_noncontiguous API and handling manually
the cache flushing will significantly improve performance.

Refer to:
commit 20e1dbf2bbe2 ("media: uvcvideo: Use dma_alloc_noncontiguous API")
commit 68d0c3311ec1 ("media: stk1160: use dma_alloc_noncontiguous API")

However, it's obvious that there is significant code duplication between
these two commits. Besides, a potential user USB Monitor may read outdated
data before the driver do DMA sync for CPU which will make the data
unreliable.

To reduce code duplication and avoid USB Monitor result unreliable, this
series will introduce DMA noncoherent API to USB core. And the USB core
layer will manage synchronization itself.

Then the last 2 patches have used the API.

I have tested uvcvideo driver. But I haven't tested stk1160 driver as I
don't have such boards. @Ezequiel Garcia, @Dafna Hirschfeld do you have
time to test it? Your support on this would be greatly appreciated.

Changes in v5:
 - improve if-else logic as suggested by Andy and Alan.
 - add Reviewed-by tag

Changes in v4:
 - https://lore.kernel.org/all/20250703103811.4048542-1-xu.yang_2@nxp.com/
 - improve if-else logic
 - remove uvc_stream_to_dmadev()

Changes in v3:
 - https://lore.kernel.org/all/20250702110222.3926355-1-xu.yang_2@nxp.com/
 - put Return section at the end of description
 - correct some abbreviations
 - remove usb_dma_noncoherent_sync_for_cpu() and
   usb_dma_noncoherent_sync_for_device()
 - do DMA sync in usb_hcd_map_urb_for_dma() and
   usb_hcd_unmap_urb_for_dma()
 - call flush_kernel_vmap_range() for OUT transfers
   and invalidate_kernel_vmap_range() for IN transfers 

Changes in v2:
 - https://lore.kernel.org/all/20250627101939.3649295-1-xu.yang_2@nxp.com/
 - handle it in USB core

v1:
 - https://lore.kernel.org/linux-usb/20250614132446.251218-1-xu.yang_2@nxp.com/

Xu Yang (3):
  usb: core: add dma-noncoherent buffer alloc and free API
  media: uvcvideo: use usb_alloc_noncoherent/usb_free_noncoherent()
  media: stk1160: use usb_alloc_noncoherent/usb_free_noncoherent()

 drivers/media/usb/stk1160/stk1160-v4l.c   |  4 --
 drivers/media/usb/stk1160/stk1160-video.c | 43 ++++--------
 drivers/media/usb/stk1160/stk1160.h       |  7 --
 drivers/media/usb/uvc/uvc_video.c         | 61 ++++-------------
 drivers/usb/core/hcd.c                    | 29 +++++---
 drivers/usb/core/usb.c                    | 80 +++++++++++++++++++++++
 include/linux/usb.h                       | 11 ++++
 7 files changed, 137 insertions(+), 98 deletions(-)

-- 
2.34.1

Hi all,

On 4-Jul-25 11:57, Xu Yang wrote:
> On architectures where there is no coherent caching such as ARM it's
> proved that using dma_alloc_noncontiguous API and handling manually
> the cache flushing will significantly improve performance.
> 
> Refer to:
> commit 20e1dbf2bbe2 ("media: uvcvideo: Use dma_alloc_noncontiguous API")
> commit 68d0c3311ec1 ("media: stk1160: use dma_alloc_noncontiguous API")
> 
> However, it's obvious that there is significant code duplication between
> these two commits. Besides, a potential user USB Monitor may read outdated
> data before the driver do DMA sync for CPU which will make the data
> unreliable.
> 
> To reduce code duplication and avoid USB Monitor result unreliable, this
> series will introduce DMA noncoherent API to USB core. And the USB core
> layer will manage synchronization itself.
> 
> Then the last 2 patches have used the API.
> 
> I have tested uvcvideo driver. But I haven't tested stk1160 driver as I
> don't have such boards. @Ezequiel Garcia, @Dafna Hirschfeld do you have
> time to test it? Your support on this would be greatly appreciated.

It seems that patches 1 + 2 are ready for merging now
(for patch 3 we should probably wait for testing).

I think that it would be best for both patches 1 + 2 to
be merged through the USB tree. The changed code in the UVC
driver is not touched that often so I do not expect any
conflicts.

Regards,

Hans




> Changes in v5:
>  - improve if-else logic as suggested by Andy and Alan.
>  - add Reviewed-by tag
> 
> Changes in v4:
>  - https://lore.kernel.org/all/20250703103811.4048542-1-xu.yang_2@nxp.com/
>  - improve if-else logic
>  - remove uvc_stream_to_dmadev()
> 
> Changes in v3:
>  - https://lore.kernel.org/all/20250702110222.3926355-1-xu.yang_2@nxp.com/
>  - put Return section at the end of description
>  - correct some abbreviations
>  - remove usb_dma_noncoherent_sync_for_cpu() and
>    usb_dma_noncoherent_sync_for_device()
>  - do DMA sync in usb_hcd_map_urb_for_dma() and
>    usb_hcd_unmap_urb_for_dma()
>  - call flush_kernel_vmap_range() for OUT transfers
>    and invalidate_kernel_vmap_range() for IN transfers 
> 
> Changes in v2:
>  - https://lore.kernel.org/all/20250627101939.3649295-1-xu.yang_2@nxp.com/
>  - handle it in USB core
> 
> v1:
>  - https://lore.kernel.org/linux-usb/20250614132446.251218-1-xu.yang_2@nxp.com/
> 
> Xu Yang (3):
>   usb: core: add dma-noncoherent buffer alloc and free API
>   media: uvcvideo: use usb_alloc_noncoherent/usb_free_noncoherent()
>   media: stk1160: use usb_alloc_noncoherent/usb_free_noncoherent()
> 
>  drivers/media/usb/stk1160/stk1160-v4l.c   |  4 --
>  drivers/media/usb/stk1160/stk1160-video.c | 43 ++++--------
>  drivers/media/usb/stk1160/stk1160.h       |  7 --
>  drivers/media/usb/uvc/uvc_video.c         | 61 ++++-------------
>  drivers/usb/core/hcd.c                    | 29 +++++---
>  drivers/usb/core/usb.c                    | 80 +++++++++++++++++++++++
>  include/linux/usb.h                       | 11 ++++
>  7 files changed, 137 insertions(+), 98 deletions(-)
>

On Mon, Jul 07, 2025 at 12:02:41PM +0200, Hans de Goede wrote:
> Hi all,
> 
> On 4-Jul-25 11:57, Xu Yang wrote:
> > On architectures where there is no coherent caching such as ARM it's
> > proved that using dma_alloc_noncontiguous API and handling manually
> > the cache flushing will significantly improve performance.
> > 
> > Refer to:
> > commit 20e1dbf2bbe2 ("media: uvcvideo: Use dma_alloc_noncontiguous API")
> > commit 68d0c3311ec1 ("media: stk1160: use dma_alloc_noncontiguous API")
> > 
> > However, it's obvious that there is significant code duplication between
> > these two commits. Besides, a potential user USB Monitor may read outdated
> > data before the driver do DMA sync for CPU which will make the data
> > unreliable.
> > 
> > To reduce code duplication and avoid USB Monitor result unreliable, this
> > series will introduce DMA noncoherent API to USB core. And the USB core
> > layer will manage synchronization itself.
> > 
> > Then the last 2 patches have used the API.
> > 
> > I have tested uvcvideo driver. But I haven't tested stk1160 driver as I
> > don't have such boards. @Ezequiel Garcia, @Dafna Hirschfeld do you have
> > time to test it? Your support on this would be greatly appreciated.
> 
> It seems that patches 1 + 2 are ready for merging now
> (for patch 3 we should probably wait for testing).
> 
> I think that it would be best for both patches 1 + 2 to
> be merged through the USB tree. The changed code in the UVC
> driver is not touched that often so I do not expect any
> conflicts.

Ok, thanks, I'll take them through the USB tree now.

greg k-h