Re: [PATCH v2 6/6] media: uvcvideo: Use dma_alloc_noncontiguous API
Hi Laurent Thanks a lot for the review On Fri, Mar 12, 2021 at 10:19 PM Laurent Pinchart wrote: > > Hi Ricardo, > > Thank you for the patch. > > On Fri, Mar 12, 2021 at 01:57:09PM +0100, Ricardo Ribalda wrote: > > On architectures where there is no coherent caching such as ARM use the > > dma_alloc_noncontiguous API and handle manually the cache flushing using > > dma_sync_sgtable(). > > You're actually switching to dma_alloc_noncontiguous() unconditionally, > not only on those architectures :-) Do I assume correctly that > dma_alloc_noncontiguous() will do the right thing on x86 too ? Yes dma_alloc_noncontiguous does all the magic. It falls back to alloc_dma_pages. Christoph has done a great job. I tried it in my notebook and although the images are nothing but pretty it is not the APIs fault. It is because the barbers are closed due to the pandemic ;). > > > With this patch on the affected architectures we can measure up to 20x > > performance improvement in uvc_video_copy_data_work(). > > Have you measured performances on x86 to ensure there's no regression ? Yes in the early stages. I am adding the latest x86 measurements in the commit message in v3 to make it more clear. > > > Eg: aarch64 with an external usb camera > > > > NON_CONTIGUOUS > > frames: 999 > > packets: 999 > > empty: 0 (0 %) > > errors: 0 > > invalid: 0 > > pts: 0 early, 0 initial, 999 ok > > scr: 0 count ok, 0 diff ok > > sof: 2048 <= sof <= 0, freq 0.000 kHz > > bytes 67034480 : duration 33303 > > FPS: 29.99 > > URB: 523446/4993 uS/qty: 104.836 avg 132.532 std 13.230 min 831.094 max (uS) > > header: 76564/4993 uS/qty: 15.334 avg 15.229 std 3.438 min 186.875 max (uS) > > latency: 468945/4992 uS/qty: 93.939 avg 132.577 std 9.531 min 824.010 max > > (uS) > > decode: 54161/4993 uS/qty: 10.847 avg 6.313 std 1.614 min 111.458 max (uS) > > raw decode speed: 9.931 Gbits/s > > raw URB handling speed: 1.025 Gbits/s > > throughput: 16.102 Mbits/s > > URB decode CPU usage 0.162600 % > > > > COHERENT > > frames: 999 > > packets: 999 > > empty: 0 (0 %) > > errors: 0 > > invalid: 0 > > pts: 0 early, 0 initial, 999 ok > > scr: 0 count ok, 0 diff ok > > sof: 2048 <= sof <= 0, freq 0.000 kHz > > bytes 54683536 : duration 33302 > > FPS: 29.99 > > URB: 1478135/4000 uS/qty: 369.533 avg 390.357 std 22.968 min 3337.865 max > > (uS) > > header: 79761/4000 uS/qty: 19.940 avg 18.495 std 1.875 min 336.719 max (uS) > > latency: 281077/4000 uS/qty: 70.269 avg 83.102 std 5.104 min 735.000 max > > (uS) > > decode: 1197057/4000 uS/qty: 299.264 avg 318.080 std 1.615 min 2806.667 max > > (uS) > > raw decode speed: 365.470 Mbits/s > > raw URB handling speed: 295.986 Mbits/s > > throughput: 13.136 Mbits/s > > URB decode CPU usage 3.594500 % > > > > Signed-off-by: Ricardo Ribalda > > Reviewed-by: Tomasz Figa > > Signed-off-by: Christoph Hellwig > > --- > > > > Changelog from v2: (Thanks Laurent) > > > > - Fix typos > > - Use the right dma direction if not capturing > > - Clear sgt during free > > > > drivers/media/usb/uvc/uvc_video.c | 92 +++ > > drivers/media/usb/uvc/uvcvideo.h | 5 +- > > 2 files changed, 74 insertions(+), 23 deletions(-) > > > > diff --git a/drivers/media/usb/uvc/uvc_video.c > > b/drivers/media/usb/uvc/uvc_video.c > > index f2f565281e63..8e60f81e2257 100644 > > --- a/drivers/media/usb/uvc/uvc_video.c > > +++ b/drivers/media/usb/uvc/uvc_video.c > > @@ -6,11 +6,14 @@ > > * Laurent Pinchart (laurent.pinch...@ideasonboard.com) > > */ > > > > +#include > > +#include > > #include > > #include > > #include > > #include > > #include > > +#include > > #include > > #include > > #include > > @@ -1096,6 +1099,34 @@ static int uvc_video_decode_start(struct > > uvc_streaming *stream, > > return data[0]; > > } > > > > +static inline enum dma_data_direction stream_dir(struct uvc_streaming > > *stream) > > +{ > > + if (stream->type == V4L2_BUF_TYPE_VIDEO_CAPTURE) > > + return DMA_FROM_DEVICE; > > + else > > + return DMA_TO_DEVICE; > > +} > > + > > +static inline struct device *stream_to_dmadev(struct uvc_streaming *stream) > > +{ > > + return bus_to_hcd(stream->dev->udev->bus)->self.sysdev; > > +} > > + > > +static void uvc_urb_dma_sync(struct uvc_urb *uvc_urb, bool for_device) > > Maybe nitpicking a little bit, but wouldn't the code be clearer if you > created uvc_urb_dma_sync_for_cpu() and uvc_urb_dma_sync_for_device() ? > When reading > > uvc_urb_dma_sync(uvc_urb, true); > > I have to constantly look up the definition of the function to figure > out what boolean value corresponds to what direction. > > Given that uvc_urb_dma_sync(..., true) is always called right before > submitting the URB, we could even create a uvc_submit_urb() function > that groups the dma_sync_sgtable_for_device() and usb_submit_urb() > calls, and do without uvc_urb_dma_sync_for_device(). Up to you on this > one. I Like it! thanks for the idea > >
Re: [PATCH v2 6/6] media: uvcvideo: Use dma_alloc_noncontiguous API
Hi Ricardo, On Fri, Mar 12, 2021 at 11:15:46PM +0100, Ricardo Ribalda Delgado wrote: > On Fri, Mar 12, 2021 at 10:19 PM Laurent Pinchart wrote: > > On Fri, Mar 12, 2021 at 01:57:09PM +0100, Ricardo Ribalda wrote: > > > On architectures where there is no coherent caching such as ARM use the > > > dma_alloc_noncontiguous API and handle manually the cache flushing using > > > dma_sync_sgtable(). > > > > You're actually switching to dma_alloc_noncontiguous() unconditionally, > > not only on those architectures :-) Do I assume correctly that > > dma_alloc_noncontiguous() will do the right thing on x86 too ? > > Yes dma_alloc_noncontiguous does all the magic. It falls back to > alloc_dma_pages. Christoph has done a great job. It's a really nice work yes. Maybe the commit message could be reworded accordingly then, to explain that we switch to dma_alloc_noncontiguous() unconditionally as it handles the differences behind the scenes ? > I tried it in my notebook and although the images are nothing but > pretty it is not the APIs fault. It is because the barbers are closed > due to the pandemic ;). :-) > > > With this patch on the affected architectures we can measure up to 20x > > > performance improvement in uvc_video_copy_data_work(). > > > > Have you measured performances on x86 to ensure there's no regression ? > > Yes in the early stages. I am adding the latest x86 measurements in > the commit message in v3 to make it more clear. > > > > Eg: aarch64 with an external usb camera > > > > > > NON_CONTIGUOUS > > > frames: 999 > > > packets: 999 > > > empty: 0 (0 %) > > > errors: 0 > > > invalid: 0 > > > pts: 0 early, 0 initial, 999 ok > > > scr: 0 count ok, 0 diff ok > > > sof: 2048 <= sof <= 0, freq 0.000 kHz > > > bytes 67034480 : duration 33303 > > > FPS: 29.99 > > > URB: 523446/4993 uS/qty: 104.836 avg 132.532 std 13.230 min 831.094 max > > > (uS) > > > header: 76564/4993 uS/qty: 15.334 avg 15.229 std 3.438 min 186.875 max > > > (uS) > > > latency: 468945/4992 uS/qty: 93.939 avg 132.577 std 9.531 min 824.010 max > > > (uS) > > > decode: 54161/4993 uS/qty: 10.847 avg 6.313 std 1.614 min 111.458 max (uS) > > > raw decode speed: 9.931 Gbits/s > > > raw URB handling speed: 1.025 Gbits/s > > > throughput: 16.102 Mbits/s > > > URB decode CPU usage 0.162600 % > > > > > > COHERENT > > > frames: 999 > > > packets: 999 > > > empty: 0 (0 %) > > > errors: 0 > > > invalid: 0 > > > pts: 0 early, 0 initial, 999 ok > > > scr: 0 count ok, 0 diff ok > > > sof: 2048 <= sof <= 0, freq 0.000 kHz > > > bytes 54683536 : duration 33302 > > > FPS: 29.99 > > > URB: 1478135/4000 uS/qty: 369.533 avg 390.357 std 22.968 min 3337.865 max > > > (uS) > > > header: 79761/4000 uS/qty: 19.940 avg 18.495 std 1.875 min 336.719 max > > > (uS) > > > latency: 281077/4000 uS/qty: 70.269 avg 83.102 std 5.104 min 735.000 max > > > (uS) > > > decode: 1197057/4000 uS/qty: 299.264 avg 318.080 std 1.615 min 2806.667 > > > max (uS) > > > raw decode speed: 365.470 Mbits/s > > > raw URB handling speed: 295.986 Mbits/s > > > throughput: 13.136 Mbits/s > > > URB decode CPU usage 3.594500 % > > > > > > Signed-off-by: Ricardo Ribalda > > > Reviewed-by: Tomasz Figa > > > Signed-off-by: Christoph Hellwig > > > --- > > > > > > Changelog from v2: (Thanks Laurent) > > > > > > - Fix typos > > > - Use the right dma direction if not capturing > > > - Clear sgt during free > > > > > > drivers/media/usb/uvc/uvc_video.c | 92 +++ > > > drivers/media/usb/uvc/uvcvideo.h | 5 +- > > > 2 files changed, 74 insertions(+), 23 deletions(-) > > > > > > diff --git a/drivers/media/usb/uvc/uvc_video.c > > > b/drivers/media/usb/uvc/uvc_video.c > > > index f2f565281e63..8e60f81e2257 100644 > > > --- a/drivers/media/usb/uvc/uvc_video.c > > > +++ b/drivers/media/usb/uvc/uvc_video.c > > > @@ -6,11 +6,14 @@ > > > * Laurent Pinchart (laurent.pinch...@ideasonboard.com) > > > */ > > > > > > +#include > > > +#include > > > #include > > > #include > > > #include > > > #include > > > #include > > > +#include > > > #include > > > #include > > > #include > > > @@ -1096,6 +1099,34 @@ static int uvc_video_decode_start(struct > > > uvc_streaming *stream, > > > return data[0]; > > > } > > > > > > +static inline enum dma_data_direction stream_dir(struct uvc_streaming > > > *stream) > > > +{ > > > + if (stream->type == V4L2_BUF_TYPE_VIDEO_CAPTURE) > > > + return DMA_FROM_DEVICE; > > > + else > > > + return DMA_TO_DEVICE; > > > +} > > > + > > > +static inline struct device *stream_to_dmadev(struct uvc_streaming > > > *stream) > > > +{ > > > + return bus_to_hcd(stream->dev->udev->bus)->self.sysdev; > > > +} > > > + > > > +static void uvc_urb_dma_sync(struct uvc_urb *uvc_urb, bool for_device) > > > > Maybe nitpicking a little bit, but wouldn't the code be clearer if you > > created uvc_urb_dma_sync_for_cpu() and uvc_urb_dma_sync_for_device() ? > > When reading > > > >
Re: [PATCH v2 6/6] media: uvcvideo: Use dma_alloc_noncontiguous API
Hi Ricardo, Thank you for the patch. On Fri, Mar 12, 2021 at 01:57:09PM +0100, Ricardo Ribalda wrote: > On architectures where there is no coherent caching such as ARM use the > dma_alloc_noncontiguous API and handle manually the cache flushing using > dma_sync_sgtable(). You're actually switching to dma_alloc_noncontiguous() unconditionally, not only on those architectures :-) Do I assume correctly that dma_alloc_noncontiguous() will do the right thing on x86 too ? > With this patch on the affected architectures we can measure up to 20x > performance improvement in uvc_video_copy_data_work(). Have you measured performances on x86 to ensure there's no regression ? > Eg: aarch64 with an external usb camera > > NON_CONTIGUOUS > frames: 999 > packets: 999 > empty: 0 (0 %) > errors: 0 > invalid: 0 > pts: 0 early, 0 initial, 999 ok > scr: 0 count ok, 0 diff ok > sof: 2048 <= sof <= 0, freq 0.000 kHz > bytes 67034480 : duration 33303 > FPS: 29.99 > URB: 523446/4993 uS/qty: 104.836 avg 132.532 std 13.230 min 831.094 max (uS) > header: 76564/4993 uS/qty: 15.334 avg 15.229 std 3.438 min 186.875 max (uS) > latency: 468945/4992 uS/qty: 93.939 avg 132.577 std 9.531 min 824.010 max (uS) > decode: 54161/4993 uS/qty: 10.847 avg 6.313 std 1.614 min 111.458 max (uS) > raw decode speed: 9.931 Gbits/s > raw URB handling speed: 1.025 Gbits/s > throughput: 16.102 Mbits/s > URB decode CPU usage 0.162600 % > > COHERENT > frames: 999 > packets: 999 > empty: 0 (0 %) > errors: 0 > invalid: 0 > pts: 0 early, 0 initial, 999 ok > scr: 0 count ok, 0 diff ok > sof: 2048 <= sof <= 0, freq 0.000 kHz > bytes 54683536 : duration 33302 > FPS: 29.99 > URB: 1478135/4000 uS/qty: 369.533 avg 390.357 std 22.968 min 3337.865 max (uS) > header: 79761/4000 uS/qty: 19.940 avg 18.495 std 1.875 min 336.719 max (uS) > latency: 281077/4000 uS/qty: 70.269 avg 83.102 std 5.104 min 735.000 max (uS) > decode: 1197057/4000 uS/qty: 299.264 avg 318.080 std 1.615 min 2806.667 max > (uS) > raw decode speed: 365.470 Mbits/s > raw URB handling speed: 295.986 Mbits/s > throughput: 13.136 Mbits/s > URB decode CPU usage 3.594500 % > > Signed-off-by: Ricardo Ribalda > Reviewed-by: Tomasz Figa > Signed-off-by: Christoph Hellwig > --- > > Changelog from v2: (Thanks Laurent) > > - Fix typos > - Use the right dma direction if not capturing > - Clear sgt during free > > drivers/media/usb/uvc/uvc_video.c | 92 +++ > drivers/media/usb/uvc/uvcvideo.h | 5 +- > 2 files changed, 74 insertions(+), 23 deletions(-) > > diff --git a/drivers/media/usb/uvc/uvc_video.c > b/drivers/media/usb/uvc/uvc_video.c > index f2f565281e63..8e60f81e2257 100644 > --- a/drivers/media/usb/uvc/uvc_video.c > +++ b/drivers/media/usb/uvc/uvc_video.c > @@ -6,11 +6,14 @@ > * Laurent Pinchart (laurent.pinch...@ideasonboard.com) > */ > > +#include > +#include > #include > #include > #include > #include > #include > +#include > #include > #include > #include > @@ -1096,6 +1099,34 @@ static int uvc_video_decode_start(struct uvc_streaming > *stream, > return data[0]; > } > > +static inline enum dma_data_direction stream_dir(struct uvc_streaming > *stream) > +{ > + if (stream->type == V4L2_BUF_TYPE_VIDEO_CAPTURE) > + return DMA_FROM_DEVICE; > + else > + return DMA_TO_DEVICE; > +} > + > +static inline struct device *stream_to_dmadev(struct uvc_streaming *stream) > +{ > + return bus_to_hcd(stream->dev->udev->bus)->self.sysdev; > +} > + > +static void uvc_urb_dma_sync(struct uvc_urb *uvc_urb, bool for_device) Maybe nitpicking a little bit, but wouldn't the code be clearer if you created uvc_urb_dma_sync_for_cpu() and uvc_urb_dma_sync_for_device() ? When reading uvc_urb_dma_sync(uvc_urb, true); I have to constantly look up the definition of the function to figure out what boolean value corresponds to what direction. Given that uvc_urb_dma_sync(..., true) is always called right before submitting the URB, we could even create a uvc_submit_urb() function that groups the dma_sync_sgtable_for_device() and usb_submit_urb() calls, and do without uvc_urb_dma_sync_for_device(). Up to you on this one. With those small changes, and assuming there's no performance regression on x86, Reviewed-by: Laurent Pinchart > +{ > + struct device *dma_dev = stream_to_dmadev(uvc_urb->stream); > + > + if (for_device) { > + dma_sync_sgtable_for_device(dma_dev, uvc_urb->sgt, > + stream_dir(uvc_urb->stream)); > + } else { > + dma_sync_sgtable_for_cpu(dma_dev, uvc_urb->sgt, > + stream_dir(uvc_urb->stream)); > + invalidate_kernel_vmap_range(uvc_urb->buffer, > + uvc_urb->stream->urb_size); > + } > +} > + > /* > * uvc_video_decode_data_work: Asynchronous memcpy processing > * > @@ -1117,6 +1148,8 @@ static void
[PATCH v2 6/6] media: uvcvideo: Use dma_alloc_noncontiguous API
On architectures where there is no coherent caching such as ARM use the dma_alloc_noncontiguous API and handle manually the cache flushing using dma_sync_sgtable(). With this patch on the affected architectures we can measure up to 20x performance improvement in uvc_video_copy_data_work(). Eg: aarch64 with an external usb camera NON_CONTIGUOUS frames: 999 packets: 999 empty: 0 (0 %) errors: 0 invalid: 0 pts: 0 early, 0 initial, 999 ok scr: 0 count ok, 0 diff ok sof: 2048 <= sof <= 0, freq 0.000 kHz bytes 67034480 : duration 33303 FPS: 29.99 URB: 523446/4993 uS/qty: 104.836 avg 132.532 std 13.230 min 831.094 max (uS) header: 76564/4993 uS/qty: 15.334 avg 15.229 std 3.438 min 186.875 max (uS) latency: 468945/4992 uS/qty: 93.939 avg 132.577 std 9.531 min 824.010 max (uS) decode: 54161/4993 uS/qty: 10.847 avg 6.313 std 1.614 min 111.458 max (uS) raw decode speed: 9.931 Gbits/s raw URB handling speed: 1.025 Gbits/s throughput: 16.102 Mbits/s URB decode CPU usage 0.162600 % COHERENT frames: 999 packets: 999 empty: 0 (0 %) errors: 0 invalid: 0 pts: 0 early, 0 initial, 999 ok scr: 0 count ok, 0 diff ok sof: 2048 <= sof <= 0, freq 0.000 kHz bytes 54683536 : duration 33302 FPS: 29.99 URB: 1478135/4000 uS/qty: 369.533 avg 390.357 std 22.968 min 3337.865 max (uS) header: 79761/4000 uS/qty: 19.940 avg 18.495 std 1.875 min 336.719 max (uS) latency: 281077/4000 uS/qty: 70.269 avg 83.102 std 5.104 min 735.000 max (uS) decode: 1197057/4000 uS/qty: 299.264 avg 318.080 std 1.615 min 2806.667 max (uS) raw decode speed: 365.470 Mbits/s raw URB handling speed: 295.986 Mbits/s throughput: 13.136 Mbits/s URB decode CPU usage 3.594500 % Signed-off-by: Ricardo Ribalda Reviewed-by: Tomasz Figa Signed-off-by: Christoph Hellwig --- Changelog from v2: (Thanks Laurent) - Fix typos - Use the right dma direction if not capturing - Clear sgt during free drivers/media/usb/uvc/uvc_video.c | 92 +++ drivers/media/usb/uvc/uvcvideo.h | 5 +- 2 files changed, 74 insertions(+), 23 deletions(-) diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c index f2f565281e63..8e60f81e2257 100644 --- a/drivers/media/usb/uvc/uvc_video.c +++ b/drivers/media/usb/uvc/uvc_video.c @@ -6,11 +6,14 @@ * Laurent Pinchart (laurent.pinch...@ideasonboard.com) */ +#include +#include #include #include #include #include #include +#include #include #include #include @@ -1096,6 +1099,34 @@ static int uvc_video_decode_start(struct uvc_streaming *stream, return data[0]; } +static inline enum dma_data_direction stream_dir(struct uvc_streaming *stream) +{ + if (stream->type == V4L2_BUF_TYPE_VIDEO_CAPTURE) + return DMA_FROM_DEVICE; + else + return DMA_TO_DEVICE; +} + +static inline struct device *stream_to_dmadev(struct uvc_streaming *stream) +{ + return bus_to_hcd(stream->dev->udev->bus)->self.sysdev; +} + +static void uvc_urb_dma_sync(struct uvc_urb *uvc_urb, bool for_device) +{ + struct device *dma_dev = stream_to_dmadev(uvc_urb->stream); + + if (for_device) { + dma_sync_sgtable_for_device(dma_dev, uvc_urb->sgt, + stream_dir(uvc_urb->stream)); + } else { + dma_sync_sgtable_for_cpu(dma_dev, uvc_urb->sgt, +stream_dir(uvc_urb->stream)); + invalidate_kernel_vmap_range(uvc_urb->buffer, +uvc_urb->stream->urb_size); + } +} + /* * uvc_video_decode_data_work: Asynchronous memcpy processing * @@ -1117,6 +1148,8 @@ static void uvc_video_copy_data_work(struct work_struct *work) uvc_queue_buffer_release(op->buf); } + uvc_urb_dma_sync(uvc_urb, true); + ret = usb_submit_urb(uvc_urb->urb, GFP_KERNEL); if (ret < 0) dev_err(_urb->stream->intf->dev, @@ -1541,10 +1574,12 @@ static void uvc_video_complete(struct urb *urb) * Process the URB headers, and optionally queue expensive memcpy tasks * to be deferred to a work queue. */ + uvc_urb_dma_sync(uvc_urb, false); stream->decode(uvc_urb, buf, buf_meta); /* If no async work is needed, resubmit the URB immediately. */ if (!uvc_urb->async_operations) { + uvc_urb_dma_sync(uvc_urb, true); ret = usb_submit_urb(uvc_urb->urb, GFP_ATOMIC); if (ret < 0) dev_err(>intf->dev, @@ -1560,24 +1595,49 @@ static void uvc_video_complete(struct urb *urb) */ static void uvc_free_urb_buffers(struct uvc_streaming *stream) { + struct device *dma_dev = stream_to_dmadev(stream); struct uvc_urb *uvc_urb; for_each_uvc_urb(uvc_urb, stream) { if (!uvc_urb->buffer) continue; -#ifndef CONFIG_DMA_NONCOHERENT -