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 <ribalda@xxxxxxxxxxxx>
> Reviewed-by: Tomasz Figa <tfiga@xxxxxxxxxxxx>
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> ---
>
> 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.pinchart@xxxxxxxxxxxxxxxx)
> */
>
> +#include <linux/dma-mapping.h>
> +#include <linux/highmem.h>
> #include <linux/kernel.h>
> #include <linux/list.h>
> #include <linux/module.h>
> #include <linux/slab.h>
> #include <linux/usb.h>
> +#include <linux/usb/hcd.h>
> #include <linux/videodev2.h>
> #include <linux/vmalloc.h>
> #include <linux/wait.h>
> @@ -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 <laurent.pinchart@xxxxxxxxxxxxxxxx>
> +{
> + 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(&uvc_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(&stream->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
> - usb_free_coherent(stream->dev->udev, stream->urb_size,
> - uvc_urb->buffer, uvc_urb->dma);
> -#else
> - kfree(uvc_urb->buffer);
> -#endif
> + dma_vunmap_noncontiguous(dma_dev, uvc_urb->buffer);
> + dma_free_noncontiguous(dma_dev, stream->urb_size, uvc_urb->sgt,
> + stream_dir(stream));
> +
> uvc_urb->buffer = NULL;
> + uvc_urb->sgt = NULL;
> }
>
> stream->urb_size = 0;
> }
>
> +static bool uvc_alloc_urb_buffer(struct uvc_streaming *stream,
> + struct uvc_urb *uvc_urb, gfp_t gfp_flags)
> +{
> + struct device *dma_dev = stream_to_dmadev(stream);
> +
> + uvc_urb->sgt = dma_alloc_noncontiguous(dma_dev, stream->urb_size,
> + stream_dir(stream),
> + gfp_flags, 0);
> + if (!uvc_urb->sgt)
> + return false;
> + uvc_urb->dma = uvc_urb->sgt->sgl->dma_address;
> +
> + uvc_urb->buffer = dma_vmap_noncontiguous(dma_dev, stream->urb_size,
> + uvc_urb->sgt);
> + if (!uvc_urb->buffer) {
> + dma_free_noncontiguous(dma_dev, stream->urb_size,
> + uvc_urb->sgt,
> + stream_dir(stream));
> + uvc_urb->sgt = NULL;
> + return false;
> + }
> +
> + return true;
> +}
> +
> /*
> * Allocate transfer buffers. This function can be called with buffers
> * already allocated when resuming from suspend, in which case it will
> @@ -1608,19 +1668,12 @@ static int uvc_alloc_urb_buffers(struct uvc_streaming *stream,
>
> /* Retry allocations until one succeed. */
> for (; npackets > 1; npackets /= 2) {
> + stream->urb_size = psize * npackets;
> +
> for (i = 0; i < UVC_URBS; ++i) {
> struct uvc_urb *uvc_urb = &stream->uvc_urb[i];
>
> - stream->urb_size = psize * npackets;
> -#ifndef CONFIG_DMA_NONCOHERENT
> - uvc_urb->buffer = usb_alloc_coherent(
> - stream->dev->udev, stream->urb_size,
> - gfp_flags | __GFP_NOWARN, &uvc_urb->dma);
> -#else
> - uvc_urb->buffer =
> - kmalloc(stream->urb_size, gfp_flags | __GFP_NOWARN);
> -#endif
> - if (!uvc_urb->buffer) {
> + if (!uvc_alloc_urb_buffer(stream, uvc_urb, gfp_flags)) {
> uvc_free_urb_buffers(stream);
> break;
> }
> @@ -1730,12 +1783,8 @@ static int uvc_init_video_isoc(struct uvc_streaming *stream,
> urb->context = uvc_urb;
> urb->pipe = usb_rcvisocpipe(stream->dev->udev,
> ep->desc.bEndpointAddress);
> -#ifndef CONFIG_DMA_NONCOHERENT
> urb->transfer_flags = URB_ISO_ASAP | URB_NO_TRANSFER_DMA_MAP;
> urb->transfer_dma = uvc_urb->dma;
> -#else
> - urb->transfer_flags = URB_ISO_ASAP;
> -#endif
> urb->interval = ep->desc.bInterval;
> urb->transfer_buffer = uvc_urb->buffer;
> urb->complete = uvc_video_complete;
> @@ -1795,10 +1844,8 @@ static int uvc_init_video_bulk(struct uvc_streaming *stream,
>
> usb_fill_bulk_urb(urb, stream->dev->udev, pipe, uvc_urb->buffer,
> size, uvc_video_complete, uvc_urb);
> -#ifndef CONFIG_DMA_NONCOHERENT
> urb->transfer_flags = URB_NO_TRANSFER_DMA_MAP;
> urb->transfer_dma = uvc_urb->dma;
> -#endif
>
> uvc_urb->urb = urb;
> }
> @@ -1895,6 +1942,7 @@ static int uvc_video_start_transfer(struct uvc_streaming *stream,
>
> /* Submit the URBs. */
> for_each_uvc_urb(uvc_urb, stream) {
> + uvc_urb_dma_sync(uvc_urb, true);
> ret = usb_submit_urb(uvc_urb->urb, gfp_flags);
> if (ret < 0) {
> dev_err(&stream->intf->dev,
> diff --git a/drivers/media/usb/uvc/uvcvideo.h b/drivers/media/usb/uvc/uvcvideo.h
> index 97df5ecd66c9..cce5e38133cd 100644
> --- a/drivers/media/usb/uvc/uvcvideo.h
> +++ b/drivers/media/usb/uvc/uvcvideo.h
> @@ -219,6 +219,7 @@
> */
>
> struct gpio_desc;
> +struct sg_table;
> struct uvc_device;
>
> /* TODO: Put the most frequently accessed fields at the beginning of
> @@ -545,7 +546,8 @@ struct uvc_copy_op {
> * @urb: the URB described by this context structure
> * @stream: UVC streaming context
> * @buffer: memory storage for the URB
> - * @dma: DMA coherent addressing for the urb_buffer
> + * @dma: Allocated DMA handle
> + * @sgt: sgt_table with the urb locations in memory
> * @async_operations: counter to indicate the number of copy operations
> * @copy_operations: work descriptors for asynchronous copy operations
> * @work: work queue entry for asynchronous decode
> @@ -556,6 +558,7 @@ struct uvc_urb {
>
> char *buffer;
> dma_addr_t dma;
> + struct sg_table *sgt;
>
> unsigned int async_operations;
> struct uvc_copy_op copy_operations[UVC_MAX_PACKETS];
--
Regards,
Laurent Pinchart
