V4L2_MEMORY_USERPTR example for iMX6

[erezsteinberg]

Hello experts,

On my iMX6DL with Linux 3.14.28 I have a camera connected using MIPI.  The video streams needs to be processed by the CPU, but the buffers are non-cacheable (Using V4L2_MEMORY_MMAP).

Is there a solution for this issue?

Some ideas -

1. Can I replace the dma_alloc_coherent() uses to allocate buffer with another call to make them cacheable?

2. Is there an example of how to get V4L2_MEMORY_USERPTR working properly? (I couldn't...)

   Maybe a unit-test used to verify the Driver implementation?

Any help would be appreciated.

Sincerely,

Erez

[Yuri]

  Please try to fix ENGR00234387 to support V4L2_MEMORY_USERPTR.

https://bitbucket.org/devonit/linux-2.6-imx/branch/imx_3.0.35_1.1.0

https://bitbucket.org/devonit/linux-2.6-imx/commits/4eed6a080a4d4a453d976a4fb2dc6f5d90fbb827?at=mast...

As for an example :

http://linuxtv.org/downloads/v4l-dvb-apis/capture-example.html

Have a great day,
Yuri

-----------------------------------------------------------------------------------------------------------------------
Note: If this post answers your question, please click the Correct Answer button. Thank you!
-----------------------------------------------------------------------------------------------------------------------

[erezsteinberg]

Hi Yuri,

Thanks for the reply.

The link you sent is from 2012, and it looks like this code is already implemented in 3.14.28.

Anyway -- it doesn't seem to work.

When using USERPTR, the user allocates buffers in user-space, and passes a pointer via m.userptr member of v4l2_buffer. However, none of the mxc-capture code in drivers/media/platform/mxc/capture/  accesses this struct member.

Can you please check?

Regards,

Erez

[Yuri]

Can you try the patch

diff --git a/drivers/media/video/mxc/capture/mxc_v4l2_capture.c b/drivers/media/video/mxc/capture/mxc_v4l2_capture.c

index 9130388..dddf670 100644

--- a/drivers/media/video/mxc/capture/mxc_v4l2_capture.c

+++ b/drivers/media/video/mxc/capture/mxc_v4l2_capture.c

@@ -324,8 +324,10 @@ static int mxc_v4l2_prepare_bufs(cam_data *cam, struct v4l2_buffer *buf)

{

  pr_debug("In MVC:mxc_v4l2_prepare_bufs\n");

  if (buf->index < 0 || buf->index >= FRAME_NUM || buf->length <

- PAGE_ALIGN(cam->v2f.fmt.pix.sizeimage)) {

+ cam->v2f.fmt.pix.sizeimage) {

  pr_err("ERROR: v4l2 capture: mxc_v4l2_prepare_bufs buffers "

  "not allocated,index=%d, length=%d\n", buf->index,

  buf->length);

[erezsteinberg]

Hi Yuri,

Removing the 'PAGE_ALIGN' macro?

Is that the whole patch? I don't see how that would make a difference.

With USERPTR the user-space application provide the virtual address of the video buffer in buf.m.userptr.   This member is ignored in the function -- so how can it work?

From what I know, for userptr the v4l2 driver should copy the virtual address, and then verify the physical pages are continuous (If not, the driver should rearrange the pages). I think this is done in videobuf2-memops.c this is done in vb2_get_contig_userptr()

Regards,

Erez

[Yuri]

From app team :

"

It need user space allocate  physical continuous memory , such use  ipu alloc , or other method.

I attached v4l2 unit test code , mxc_v4l2_output.c and mx6s_v4l2_capture.c , one use ipu alloc , the other use pxp alloc .

Search memalloc in these code.

"

Regards,

Yuri.

[erezsteinberg]

Hi Yuri,

Thanks for the examples. This is good information. However, it does not resolve the problem.

I looked at the kernel sources to see how the IPU and PXP drivers implement the memory allocation, they still use dma_alloc_coherent() ... so, the memory buffers are non cachable.

Moreover, the purpose of USERPTR in V4L2 is to allow users to pass buffers allocated in user-space directly by malloc or statically. Using PXP or IPU ot perform the allocation is not really the way it was intended.

Also -- I'm not sure my problem is clear.  I don't have to use USERPTR (MMAP is okay), but using dma_alloc_coherent() give very bad performance.

What I need is a way to allocate the video buffers and to have good performance in user-space

I wrote a DDR benchmark application to show the issue. The code allocates a buffer of 4MB and reads it several times in a loop.

When the buffer is allocated with malloc() it takes 49.8msec to read 10MB.

However, when using IPU_ALLOC, it takes 338.4msec (x6.8 longer!)

root@imx6qsabresd:~# ./ddr_benchmark

Test start

Test complete (dummy 0)

Time taken (nanoseconds): 49827333

root@imx6qsabresd:~# ./ddr_benchmark_ipu

USRP: alloc bufs offset 0x24b00000 size 4149248

Test start

Test complete (dummy 0)

Time taken (nanoseconds): 338425669

Buffer allocated using malloc()

-------------------------------

MMDC new Profiling results:

***********************

Measure time: 1000ms

Total cycles count: 396050646

Busy cycles count: 240362999

Read accesses count: 7001555

Write accesses count: 9174

Read bytes count: 447536316

Write bytes count: 293466

Avg. Read burst size: 63

Avg. Write burst size: 31

Read: 426.80 MB/s /  Write: 0.28 MB/s  Total: 427.08 MB/s

Utilization: 11%

Bus Load: 60%

Bytes Access: 63

Buffer allocated using IPU_ALLOC

---------------------------------

MMDC new Profiling results:

***********************

Measure time: 1001ms

Total cycles count: 396043446

Busy cycles count: 256915721

Read accesses count: 8802699

Write accesses count: 14923

Read bytes count: 78915456

Write bytes count: 252306

Avg. Read burst size: 8

Avg. Write burst size: 16

Read: 75.18 MB/s /  Write: 0.24 MB/s  Total: 75.42 MB/s

Utilization: 1%

Bus Load: 64%

Bytes Access: 8

The benchmark code is attached.

To select IPU_ALLOC, uncomment line 12.

Build command: arm-linux-gnueabihf-gcc -O3 -mcpu=cortex-a9 -mfloat-abi=hard ddr_benchmark.c -o ddr_benchmark

[Yuri]

From app team :

"The current driver used  is coherent mapping .

For customer's use case which need  buffer cache-able for
cpu process  the captured data buffer , they need  implement
this new feature by their own."

Regards,

Yuri.

[erezsteinberg]

Hi Yuri-

I found an interesting discussion about performance in user-space for buffers allocated with kmalloc (kmalloc memory slower than malloc )

That lead me to check the mmap function in mxc_v4l2_capture.c, and there I found a solution -

To get normal performance in user-space I did the following changes -

in mxc_mmap()   -- comment out:   vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);

in mxc_allocate_frame_buf()   --  replace allocation with dma_alloc_coherent() to kmalloc()

in mxc_free_frame_buf() -- replace dma_free_coherent() with kfree().

I still use V4L2_MEMORY_MMAP.

With this change I can see normal memory loads:

MMDC new Profiling results:

***********************

Measure time: 1000ms

Total cycles count: 396071850

Busy cycles count: 155735196

Read accesses count: 2254063

Write accesses count: 4653481

Read bytes count: 138436492

Write bytes count: 211011038

Avg. Read burst size: 61

Avg. Write burst size: 45

Read: 132.02 MB/s /  Write: 201.24 MB/s  Total: 333.26 MB/s

Utilization: 14%

Bus Load: 39%

Bytes Access: 50

Thanks for the support!

Regards,

Erez

[kennywang]

Hi  Erez

Did you solve the issue? I modify  the the functions as you written, but it also slower. Would you share your your code ?

Thanks ,

Kenny

[yeti425]

Hi Erez

I'm having a similar issue with memory access times from the user space. Would it be possible for you to share the source code for the three functions please?

Many thanks

Andy