Similar to the commit 1161db6776bd ("media: usb: pwc: Don't use coherent
DMA buffers for ISO transfer") [1] for the pwc driver. Use streaming DMA
APIs to transfer buffers and sync them explicitly, because accessing
buffers allocated by usb_alloc_coherent() is slow on platforms without
hardware coherent DMA.
Tested on x86-64 (Intel Celeron 4305U) and armv7l (Rockchip RK3288) with
Logitech Brio 4K camera at 1920x1080 using the WebRTC sample site [3].
| | URB (us) | Decode (Gbps) | CPU (%) |
|------------------|------------|---------------|---------|
| x86-64 coherent | 53 +- 20 | 50.6 | 0.24 |
| x86-64 streaming | 55 +- 19 | 50.1 | 0.25 |
| armv7l coherent | 342 +- 379 | 1.8 | 2.16 |
| armv7l streaming | 99 +- 98 | 11.0 | 0.36 |
The performance characteristics are improved a lot on armv7l, and
remained (almost) unchanged on x86-64. The code used for measurement can
be found at [2].
[1] https://git.kernel.org/torvalds/c/1161db6776bd
[2] https://crrev.com/c/1729133
[3] https://webrtc.github.io/samples/src/content/getusermedia/resolution/
Signed-off-by: Shik Chen <shik@xxxxxxxxxxxx>
---
The allocated buffer could be as large as 768K when streaming 4K video.
Ideally we should use some generic helper to allocate non-coherent
memory in a more efficient way, such as https://lwn.net/Articles/774429/
("make the non-consistent DMA allocator more userful").
But I cannot find any existing helper so a simple kzalloc() is used in
this patch. The logic to figure out the DMA addressable GFP flags is
similar to __dma_direct_alloc_pages() without the optimistic retries:
https://elixir.bootlin.com/linux/v5.2.5/source/kernel/dma/direct.c#L96
drivers/media/usb/uvc/uvc_video.c | 65 +++++++++++++++++++++----------
1 file changed, 45 insertions(+), 20 deletions(-)
diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c
index 8fa77a81dd7f2c..962c35478896c4 100644
--- a/drivers/media/usb/uvc/uvc_video.c
+++ b/drivers/media/usb/uvc/uvc_video.c
@@ -1539,6 +1539,8 @@ 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.
*/
+ dma_sync_single_for_cpu(&urb->dev->dev, urb->transfer_dma,
+ urb->transfer_buffer_length, DMA_FROM_DEVICE);
stream->decode(uvc_urb, buf, buf_meta);
/* If no async work is needed, resubmit the URB immediately. */
@@ -1565,18 +1567,51 @@ static void uvc_free_urb_buffers(struct uvc_streaming *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
+ dma_unmap_single(&stream->dev->udev->dev, uvc_urb->dma,
+ stream->urb_size, DMA_FROM_DEVICE);
kfree(uvc_urb->buffer);
-#endif
- uvc_urb->buffer = NULL;
}
stream->urb_size = 0;
}
+static gfp_t uvc_alloc_gfp_flags(struct device *dev)
+{
+ u64 mask = dma_get_mask(dev);
+
+ if (dev->bus_dma_mask)
+ mask &= dev->bus_dma_mask;
+
+ if (mask < DMA_BIT_MASK(32) && IS_ENABLED(CONFIG_ZONE_DMA))
+ return GFP_DMA;
+
+ if (mask < DMA_BIT_MASK(64)) {
+ if (IS_ENABLED(CONFIG_ZONE_DMA32))
+ return GFP_DMA32;
+ if (IS_ENABLED(CONFIG_ZONE_DMA))
+ return GFP_DMA;
+ }
+
+ return 0;
+}
+
+static char *uvc_alloc_urb_buffer(struct device *dev, size_t size,
+ gfp_t gfp_flags, dma_addr_t *dma_handle)
+{
+ void *buffer = kzalloc(size, gfp_flags | uvc_alloc_gfp_flags(dev));
+
+ if (!buffer)
+ return NULL;
+
+ *dma_handle = dma_map_single(dev, buffer, size, DMA_FROM_DEVICE);
+ if (dma_mapping_error(dev, *dma_handle)) {
+ kfree(buffer);
+ return NULL;
+ }
+
+ return buffer;
+}
+
/*
* Allocate transfer buffers. This function can be called with buffers
* already allocated when resuming from suspend, in which case it will
@@ -1607,18 +1642,14 @@ 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,
+ uvc_urb->buffer = uvc_alloc_urb_buffer(
+ &stream->dev->udev->dev, 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) {
uvc_free_urb_buffers(stream);
break;
@@ -1728,12 +1759,8 @@ 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;
@@ -1793,10 +1820,8 @@ 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;
}
--
2.22.0.770.g0f2c4a37fd-goog
