This small series adds support for non-coherent video capture buffers on Rockchip ISP V1. Patch 1 fixes cache management for dmabuf's allocated by dma-contig allocator. Patch 2 allows non-coherent allocations on the rkisp1 capture queue. Some timing measurements are provided in the commit message of patch 2.
Signed-off-by: Mikhail Rudenko mike.rudenko@gmail.com --- Changes in v4: - rebase to media/next - use `direction` instead of `buf->dma_dir` in dma_sync_sgtable_* - Link to v3: https://lore.kernel.org/r/20250128-b4-rkisp-noncoherent-v3-0-baf39c997d2a@gm...
Changes in v3: - ignore skip_cache_sync_* flags in vb2_dc_dmabuf_ops_{begin,end}_cpu_access - invalidate/flush kernel mappings as appropriate if they exist - use dma_sync_sgtable_* instead of dma_sync_sg_* - Link to v2: https://lore.kernel.org/r/20250115-b4-rkisp-noncoherent-v2-0-0853e1a24012@gm...
Changes in v2: - Fix vb2_dc_dmabuf_ops_{begin,end}_cpu_access() for non-coherent buffers. - Add cache management timing information to patch 2 commit message. - Link to v1: https://lore.kernel.org/r/20250102-b4-rkisp-noncoherent-v1-1-bba164f7132c@gm...
--- Mikhail Rudenko (2): media: videobuf2: Fix dmabuf cache sync/flush in dma-contig media: rkisp1: Allow non-coherent video capture buffers
.../media/common/videobuf2/videobuf2-dma-contig.c | 22 ++++++++++++++++++++++ .../platform/rockchip/rkisp1/rkisp1-capture.c | 1 + 2 files changed, 23 insertions(+) --- base-commit: b2c4bf0c102084e77ed1b12090d77a76469a6814 change-id: 20241231-b4-rkisp-noncoherent-ad6e7c7a68ba
Best regards,
When support for V4L2_FLAG_MEMORY_NON_CONSISTENT was removed in commit 129134e5415d ("media: media/v4l2: remove V4L2_FLAG_MEMORY_NON_CONSISTENT flag"), vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions were made no-ops. Later, when support for V4L2_MEMORY_FLAG_NON_COHERENT was introduced in commit c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag"), the above functions remained no-ops, making cache maintenance for non-coherent dmabufs allocated by dma-contig impossible.
Fix this by reintroducing dma_sync_sgtable_for_{cpu,device} and {flush,invalidate}_kernel_vmap_range calls to vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions for non-coherent buffers.
Fixes: c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag") Cc: stable@vger.kernel.org Signed-off-by: Mikhail Rudenko mike.rudenko@gmail.com --- .../media/common/videobuf2/videobuf2-dma-contig.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
diff --git a/drivers/media/common/videobuf2/videobuf2-dma-contig.c b/drivers/media/common/videobuf2/videobuf2-dma-contig.c index a13ec569c82f6da2d977222b94af32e74c6c6c82..d41095fe5bd21faf815d6b035d7bc888a84a95d5 100644 --- a/drivers/media/common/videobuf2/videobuf2-dma-contig.c +++ b/drivers/media/common/videobuf2/videobuf2-dma-contig.c @@ -427,6 +427,17 @@ static int vb2_dc_dmabuf_ops_begin_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) { + struct vb2_dc_buf *buf = dbuf->priv; + struct sg_table *sgt = buf->dma_sgt; + + if (!buf->non_coherent_mem) + return 0; + + if (buf->vaddr) + invalidate_kernel_vmap_range(buf->vaddr, buf->size); + + dma_sync_sgtable_for_cpu(buf->dev, sgt, direction); + return 0; }
@@ -434,6 +445,17 @@ static int vb2_dc_dmabuf_ops_end_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) { + struct vb2_dc_buf *buf = dbuf->priv; + struct sg_table *sgt = buf->dma_sgt; + + if (!buf->non_coherent_mem) + return 0; + + if (buf->vaddr) + flush_kernel_vmap_range(buf->vaddr, buf->size); + + dma_sync_sgtable_for_device(buf->dev, sgt, direction); + return 0; }
Hi Mikhail,
Le lundi 03 mars 2025 à 14:40 +0300, Mikhail Rudenko a écrit :
When support for V4L2_FLAG_MEMORY_NON_CONSISTENT was removed in commit 129134e5415d ("media: media/v4l2: remove V4L2_FLAG_MEMORY_NON_CONSISTENT flag"), vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions were made no-ops. Later, when support for V4L2_MEMORY_FLAG_NON_COHERENT was introduced in commit c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag"), the above functions remained no-ops, making cache maintenance for non-coherent dmabufs allocated by dma-contig impossible.
Fix this by reintroducing dma_sync_sgtable_for_{cpu,device} and {flush,invalidate}_kernel_vmap_range calls to vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions for non-coherent buffers.
Fixes: c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag") Cc: stable@vger.kernel.org Signed-off-by: Mikhail Rudenko mike.rudenko@gmail.com
.../media/common/videobuf2/videobuf2-dma-contig.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
diff --git a/drivers/media/common/videobuf2/videobuf2-dma-contig.c b/drivers/media/common/videobuf2/videobuf2-dma-contig.c index a13ec569c82f6da2d977222b94af32e74c6c6c82..d41095fe5bd21faf815d6b035d7 bc888a84a95d5 100644 --- a/drivers/media/common/videobuf2/videobuf2-dma-contig.c +++ b/drivers/media/common/videobuf2/videobuf2-dma-contig.c @@ -427,6 +427,17 @@ static int vb2_dc_dmabuf_ops_begin_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) {
- struct vb2_dc_buf *buf = dbuf->priv;
- struct sg_table *sgt = buf->dma_sgt;
- if (!buf->non_coherent_mem)
return 0;- if (buf->vaddr)
invalidate_kernel_vmap_range(buf->vaddr, buf->size);
What would make me a lot more confortable with this change is if you enable kernel mappings for one test. This will ensure you cover the call to "invalidate" in your testing. I'd like to know about the performance impact. With this implementation it should be identical to the VB2 one.
What I was trying to say in previous comments, is that my impression is that we can skip this for CPU read access, since we don't guaranty concurrent access anyway. Both address space can keep their cache in that case. Though, I see RKISP does not use kernel mapping plus I'm not reporting a bug, but checking if we should leave a comment for possible users of kernel mapping in the future ?
- dma_sync_sgtable_for_cpu(buf->dev, sgt, direction);
return 0; } @@ -434,6 +445,17 @@ static int vb2_dc_dmabuf_ops_end_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) {
- struct vb2_dc_buf *buf = dbuf->priv;
- struct sg_table *sgt = buf->dma_sgt;
- if (!buf->non_coherent_mem)
return 0;- if (buf->vaddr)
flush_kernel_vmap_range(buf->vaddr, buf->size);- dma_sync_sgtable_for_device(buf->dev, sgt, direction);
return 0; }
Hi Nicolas,
On 2025-03-03 at 10:24 -05, Nicolas Dufresne nicolas@ndufresne.ca wrote:
Hi Mikhail,
Le lundi 03 mars 2025 à 14:40 +0300, Mikhail Rudenko a écrit :
When support for V4L2_FLAG_MEMORY_NON_CONSISTENT was removed in commit 129134e5415d ("media: media/v4l2: remove V4L2_FLAG_MEMORY_NON_CONSISTENT flag"), vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions were made no-ops. Later, when support for V4L2_MEMORY_FLAG_NON_COHERENT was introduced in commit c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag"), the above functions remained no-ops, making cache maintenance for non-coherent dmabufs allocated by dma-contig impossible.
Fix this by reintroducing dma_sync_sgtable_for_{cpu,device} and {flush,invalidate}_kernel_vmap_range calls to vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions for non-coherent buffers.
Fixes: c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag") Cc: stable@vger.kernel.org Signed-off-by: Mikhail Rudenko mike.rudenko@gmail.com
.../media/common/videobuf2/videobuf2-dma-contig.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
diff --git a/drivers/media/common/videobuf2/videobuf2-dma-contig.c b/drivers/media/common/videobuf2/videobuf2-dma-contig.c index a13ec569c82f6da2d977222b94af32e74c6c6c82..d41095fe5bd21faf815d6b035d7 bc888a84a95d5 100644 --- a/drivers/media/common/videobuf2/videobuf2-dma-contig.c +++ b/drivers/media/common/videobuf2/videobuf2-dma-contig.c @@ -427,6 +427,17 @@ static int vb2_dc_dmabuf_ops_begin_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) {
- struct vb2_dc_buf *buf = dbuf->priv;
- struct sg_table *sgt = buf->dma_sgt;
- if (!buf->non_coherent_mem)
return 0;- if (buf->vaddr)
invalidate_kernel_vmap_range(buf->vaddr, buf->size);What would make me a lot more confortable with this change is if you enable kernel mappings for one test. This will ensure you cover the call to "invalidate" in your testing. I'd like to know about the performance impact. With this implementation it should be identical to the VB2 one.
I'll enable kernel mappings and rerun my tests later this week.
What I was trying to say in previous comments, is that my impression is that we can skip this for CPU read access, since we don't guaranty concurrent access anyway. Both address space can keep their cache in that case. Though, I see RKISP does not use kernel mapping plus I'm not reporting a bug, but checking if we should leave a comment for possible users of kernel mapping in the future ?
I trust Tomasz here, I'd wait for his comment on v4.
- dma_sync_sgtable_for_cpu(buf->dev, sgt, direction);
return 0; }
@@ -434,6 +445,17 @@ static int vb2_dc_dmabuf_ops_end_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) {
- struct vb2_dc_buf *buf = dbuf->priv;
- struct sg_table *sgt = buf->dma_sgt;
- if (!buf->non_coherent_mem)
return 0;- if (buf->vaddr)
flush_kernel_vmap_range(buf->vaddr, buf->size);- dma_sync_sgtable_for_device(buf->dev, sgt, direction);
return 0; }
-- Best regards, Mikhail Rudenko
On Tue, Mar 4, 2025 at 12:24 AM Nicolas Dufresne nicolas@ndufresne.ca wrote:
Hi Mikhail,
Le lundi 03 mars 2025 à 14:40 +0300, Mikhail Rudenko a écrit :
When support for V4L2_FLAG_MEMORY_NON_CONSISTENT was removed in commit 129134e5415d ("media: media/v4l2: remove V4L2_FLAG_MEMORY_NON_CONSISTENT flag"), vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions were made no-ops. Later, when support for V4L2_MEMORY_FLAG_NON_COHERENT was introduced in commit c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag"), the above functions remained no-ops, making cache maintenance for non-coherent dmabufs allocated by dma-contig impossible.
Fix this by reintroducing dma_sync_sgtable_for_{cpu,device} and {flush,invalidate}_kernel_vmap_range calls to vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions for non-coherent buffers.
Fixes: c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag") Cc: stable@vger.kernel.org Signed-off-by: Mikhail Rudenko mike.rudenko@gmail.com
.../media/common/videobuf2/videobuf2-dma-contig.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
diff --git a/drivers/media/common/videobuf2/videobuf2-dma-contig.c b/drivers/media/common/videobuf2/videobuf2-dma-contig.c index a13ec569c82f6da2d977222b94af32e74c6c6c82..d41095fe5bd21faf815d6b035d7 bc888a84a95d5 100644 --- a/drivers/media/common/videobuf2/videobuf2-dma-contig.c +++ b/drivers/media/common/videobuf2/videobuf2-dma-contig.c @@ -427,6 +427,17 @@ static int vb2_dc_dmabuf_ops_begin_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) {
struct vb2_dc_buf *buf = dbuf->priv;struct sg_table *sgt = buf->dma_sgt;if (!buf->non_coherent_mem)return 0;if (buf->vaddr)invalidate_kernel_vmap_range(buf->vaddr, buf->size);What would make me a lot more confortable with this change is if you enable kernel mappings for one test. This will ensure you cover the call to "invalidate" in your testing. I'd like to know about the performance impact. With this implementation it should be identical to the VB2 one.
I agree that it would be good to test that path as well. I wonder if we could somehow do it with one of the vi* drivers...
What I was trying to say in previous comments, is that my impression is that we can skip this for CPU read access, since we don't guaranty concurrent access anyway. Both address space can keep their cache in that case. Though, I see RKISP does not use kernel mapping plus I'm not reporting a bug, but checking if we should leave a comment for possible users of kernel mapping in the future ?
We can't skip it for CPU read access, because it may be the first read after the DMA writing to the buffer, so we need to invalidate the caches.
That said, on majority of systems this will be a no-op, because it only applies to VIVT and VIPT aliasing caches + only when the kernel mapping is actually used (the buf->vaddr mapping is created on demand).
dma_sync_sgtable_for_cpu(buf->dev, sgt, direction);return 0;}
@@ -434,6 +445,17 @@ static int vb2_dc_dmabuf_ops_end_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) {
struct vb2_dc_buf *buf = dbuf->priv;struct sg_table *sgt = buf->dma_sgt;if (!buf->non_coherent_mem)return 0;if (buf->vaddr)flush_kernel_vmap_range(buf->vaddr, buf->size);dma_sync_sgtable_for_device(buf->dev, sgt, direction);return 0;}
Hi Nicolas, Tomasz,
On 2025-03-03 at 10:24 -05, Nicolas Dufresne nicolas@ndufresne.ca wrote:
Hi Mikhail,
Le lundi 03 mars 2025 à 14:40 +0300, Mikhail Rudenko a écrit :
When support for V4L2_FLAG_MEMORY_NON_CONSISTENT was removed in commit 129134e5415d ("media: media/v4l2: remove V4L2_FLAG_MEMORY_NON_CONSISTENT flag"), vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions were made no-ops. Later, when support for V4L2_MEMORY_FLAG_NON_COHERENT was introduced in commit c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag"), the above functions remained no-ops, making cache maintenance for non-coherent dmabufs allocated by dma-contig impossible.
Fix this by reintroducing dma_sync_sgtable_for_{cpu,device} and {flush,invalidate}_kernel_vmap_range calls to vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions for non-coherent buffers.
Fixes: c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag") Cc: stable@vger.kernel.org Signed-off-by: Mikhail Rudenko mike.rudenko@gmail.com
.../media/common/videobuf2/videobuf2-dma-contig.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
diff --git a/drivers/media/common/videobuf2/videobuf2-dma-contig.c b/drivers/media/common/videobuf2/videobuf2-dma-contig.c index a13ec569c82f6da2d977222b94af32e74c6c6c82..d41095fe5bd21faf815d6b035d7 bc888a84a95d5 100644 --- a/drivers/media/common/videobuf2/videobuf2-dma-contig.c +++ b/drivers/media/common/videobuf2/videobuf2-dma-contig.c @@ -427,6 +427,17 @@ static int vb2_dc_dmabuf_ops_begin_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) {
- struct vb2_dc_buf *buf = dbuf->priv;
- struct sg_table *sgt = buf->dma_sgt;
- if (!buf->non_coherent_mem)
return 0;- if (buf->vaddr)
invalidate_kernel_vmap_range(buf->vaddr, buf->size);What would make me a lot more confortable with this change is if you enable kernel mappings for one test. This will ensure you cover the call to "invalidate" in your testing. I'd like to know about the performance impact. With this implementation it should be identical to the VB2 one.
I have re-run my tests on RK3399, with 1280x720 XRGB capture buffers (1 plane, 3686400 bytes). Capture process was pinned to "big" cores running at fixed frequency of 1.8 GHz. Libcamera was modified to request buffers with V4L2_MEMORY_FLAG_NON_COHERENT flag. DMA_BUF_IOCTL_SYNC ioctls were used as appropriate. For kernel mapping effect test, vb2_plane_vaddr call was inserted into rkisp1_vb2_buf_init.
The timings are as following:
- memcpy coherent buffer: 7570 +/- 63 us - memcpy non-coherent buffer: 1120 +/- 34 us
without kernel mapping:
- ioctl(fd, DMA_BUF_IOCTL_SYNC, {DMA_BUF_SYNC_START|DMA_BUF_SYNC_READ}): 428 +/- 15 us - ioctl(fd, DMA_BUF_IOCTL_SYNC, {DMA_BUF_SYNC_END|DMA_BUF_SYNC_READ}): 422 +/- 28 us
with kernel mapping:
- ioctl(fd, DMA_BUF_IOCTL_SYNC, {DMA_BUF_SYNC_START|DMA_BUF_SYNC_READ}): 526 +/- 13 us - ioctl(fd, DMA_BUF_IOCTL_SYNC, {DMA_BUF_SYNC_END|DMA_BUF_SYNC_READ}): 519 +/- 38 us
So, even with kernel mapping enabled, speedup is 7570 / (1120 + 526 + 519) = 3.5, and the use of noncoherent buffers is justified -- at least on this platform.
-- Best regards, Mikhail Rudenko
On Mon, Mar 10, 2025 at 5:52 PM Mikhail Rudenko mike.rudenko@gmail.com wrote:
Hi Nicolas, Tomasz,
On 2025-03-03 at 10:24 -05, Nicolas Dufresne nicolas@ndufresne.ca wrote:
Hi Mikhail,
Le lundi 03 mars 2025 à 14:40 +0300, Mikhail Rudenko a écrit :
When support for V4L2_FLAG_MEMORY_NON_CONSISTENT was removed in commit 129134e5415d ("media: media/v4l2: remove V4L2_FLAG_MEMORY_NON_CONSISTENT flag"), vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions were made no-ops. Later, when support for V4L2_MEMORY_FLAG_NON_COHERENT was introduced in commit c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag"), the above functions remained no-ops, making cache maintenance for non-coherent dmabufs allocated by dma-contig impossible.
Fix this by reintroducing dma_sync_sgtable_for_{cpu,device} and {flush,invalidate}_kernel_vmap_range calls to vb2_dc_dmabuf_ops_{begin,end}_cpu_access() functions for non-coherent buffers.
Fixes: c0acf9cfeee0 ("media: videobuf2: handle V4L2_MEMORY_FLAG_NON_COHERENT flag") Cc: stable@vger.kernel.org Signed-off-by: Mikhail Rudenko mike.rudenko@gmail.com
.../media/common/videobuf2/videobuf2-dma-contig.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
diff --git a/drivers/media/common/videobuf2/videobuf2-dma-contig.c b/drivers/media/common/videobuf2/videobuf2-dma-contig.c index a13ec569c82f6da2d977222b94af32e74c6c6c82..d41095fe5bd21faf815d6b035d7 bc888a84a95d5 100644 --- a/drivers/media/common/videobuf2/videobuf2-dma-contig.c +++ b/drivers/media/common/videobuf2/videobuf2-dma-contig.c @@ -427,6 +427,17 @@ static int vb2_dc_dmabuf_ops_begin_cpu_access(struct dma_buf *dbuf, enum dma_data_direction direction) {
- struct vb2_dc_buf *buf = dbuf->priv;
- struct sg_table *sgt = buf->dma_sgt;
- if (!buf->non_coherent_mem)
return 0;- if (buf->vaddr)
invalidate_kernel_vmap_range(buf->vaddr, buf->size);What would make me a lot more confortable with this change is if you enable kernel mappings for one test. This will ensure you cover the call to "invalidate" in your testing. I'd like to know about the performance impact. With this implementation it should be identical to the VB2 one.
I have re-run my tests on RK3399, with 1280x720 XRGB capture buffers (1 plane, 3686400 bytes). Capture process was pinned to "big" cores running at fixed frequency of 1.8 GHz. Libcamera was modified to request buffers with V4L2_MEMORY_FLAG_NON_COHERENT flag. DMA_BUF_IOCTL_SYNC ioctls were used as appropriate. For kernel mapping effect test, vb2_plane_vaddr call was inserted into rkisp1_vb2_buf_init.
The timings are as following:
- memcpy coherent buffer: 7570 +/- 63 us
- memcpy non-coherent buffer: 1120 +/- 34 us
without kernel mapping:
- ioctl(fd, DMA_BUF_IOCTL_SYNC, {DMA_BUF_SYNC_START|DMA_BUF_SYNC_READ}): 428 +/- 15 us
- ioctl(fd, DMA_BUF_IOCTL_SYNC, {DMA_BUF_SYNC_END|DMA_BUF_SYNC_READ}): 422 +/- 28 us
with kernel mapping:
- ioctl(fd, DMA_BUF_IOCTL_SYNC, {DMA_BUF_SYNC_START|DMA_BUF_SYNC_READ}): 526 +/- 13 us
- ioctl(fd, DMA_BUF_IOCTL_SYNC, {DMA_BUF_SYNC_END|DMA_BUF_SYNC_READ}): 519 +/- 38 us
So, even with kernel mapping enabled, speedup is 7570 / (1120 + 526 + 519) = 3.5, and the use of noncoherent buffers is justified -- at least on this platform.
Thanks a lot for the additional testing.
Acked-by: Tomasz Figa tfiga@chromium.org
Best regards, Tomasz
linux-stable-mirror@lists.linaro.org
-
Mikhail Rudenko -
Nicolas Dufresne -
Tomasz Figa