Hi,
This is the seventh version of the MFC 5.1 driver. It is still a work in progress.
This version currently supports decoding only, but encoding is heavily developed.
For nearly two months new features were added and fixes and suggestions were
implemented.
The main purpose for posting this patch is to show an example driver that
uses the new IOMMU vb2 allocator. In addition this is the first try to
create a V4L2 codec driver that I know of. I hope to discuss issues related
to this type of hardware on the upcoming brainstorming session in Warsaw.
Major new features include:
- proper handling for stream seeking (stream off + stream on and resume from
another point of the stream).
- support for resolution change in the stream
- power domain handling
- clock gating
- support for IOMMU allocator
- setting buffer flags according to decoded frame type:
V4L2_BUF_FLAG_(KEY|P|B)FRAME
- proper display delay handling for H264
Things that still need to be done:
- move to the control framework
- rebase onto Exynos4 platform name change
* Stream seeking scenario is as follows:
1. Stream off on OUTPUT
2. Stream off on CAPTURE
3. Queue at least one new src buffer on OUTPUT (after stream off they all have
been dequeued and at least one buffer has to be queued to successfully do a
stream on)
4. Queue dst buffers on CAPTURE (after stream off they all have been dequeued)
5. Stream on on CAPTURE queue
6. Stream on on OUTPUT queue
* Resolution change scenario:
1. Stream off on CAPTURE
2. Unmap the destination buffers
Note: To successfully unmap the buffer they cannot be used
by any other hardware, such as FIMC. Thus it necessary to
shutdown FIMC - either close, or streamoff and reqbufs(0).
3. Call Reqbufs (0) on CAPTURE
4. Same as after parsing header
a) Run G_FMT on CAPTURE
b) Run G_CROP on CAPTURE
c) Get the minimum number of buffers (by reading the
V4L2_CID_CODEC_REQ_NUM_BUFS control)
5. Init CAPTURE stream
a) Run Reqbufs with appropriate number of buffers (minimum + extra)
b) Run Querybuf and Mmap on each buffer
c) Run Q_BUF on the new buffers
6. Stream on on CAPTURE
7. Initialize FIMC
Note: application is notified about resolution change in the same way as the
notification that all decoded buffers have been read. The value of bytesused
in the capture buffer is set to 0. If the application has sent last encoded
stream frame then it is waiting for last decoded frame. If the application
receives a frame with bytesused set to 0 with no error set in the middle of a
stream then it should treat is as the notification of resolution change.
Best regards,
Kamil
* Changelog:
==================
Changes since v6
==================
1) Stream seeking handling
- done by running stream off and then stream on from another point of the
stream
2) Support for streams during which the resolution is changed
- done by calling stream off, reallocating the buffers and stream on again to
resume
3) Power domain handling
4) Clock gating hw related operations
- This has introduced a large reduction in power use
5) Support for IOMMU allocator
- Using IOMMU as the memory allocator removes the cache problem
and the need for reserving continuous memory at system boot
6) Flags of v4l2_buffer are set accrodingly to the returned buffer frame type
V4L2_BUF_FLAG_(KEY|P|B)FRAME
7) Fixed display delay handling of H264. Now dealy of 0 frames is possible,
although it may cause that the frames are returned out of order.
8) Minor changes
- global s5p_mfc_dev variable has been removed
- improved Packed PB handling
- fixed error handling - separate for decoding and display frames
- some cosmetic changes to simplify the code and make it more readable
==================
Changes since v5
==================
1) Changes suggested by Hans Verkuil:
- small change in videodev2.h - corrected control offsets
- made the code more readable by simplifying if statements and using temporary
pointers
- mfc_mutex is now included in s5p_mfc_dev structure
- after discussion with Peter Oh modification of fourcc defintions
(replaced DX52 and DX53 with DX50)
2) Changes suggested by JongHun Han:
- comsmetic changed of defines in regs-mfc5.h
- in buffers that have no width adn height, such as the buffer for compressed
stream, those values are set to 0 instead of 1
- remove redundant pointer to MFC registers
- change name of the union in s5p_mfc_buf from paddr to cookie
- removed global variable (struct s5p_mfc_dev *dev) and moved to use video_drvdata
3) Other changes:
- added check for values returned after parsing header - in rare circumstances MFC
hw returned 0x0 as image size and this could cause problems
==================
Changes since v4
==================
1) Changes suggested by Kukjin Kim from:
- removed comment arch/arm/mach-s5pv210/include/mach/map.h
- changed device name to s5p-mfc (removed "5", MFC version number)
also removed the version number from the name of MFC device file
- added GPL license to arch/arm/plat-s5p/dev-mfc.c
- removed unused include file from dev-mfc.c and unnecessary comments
2) Cache handling improvement:
- changed cache handling to use dma_map_single and dma_unmap_single
==================
Changes since v3
==================
1) Update to the v6 videobuf2 API (here thanks go to Marek Szyprowski)
- s5p_mfc_buf_negotiate and s5p_mfc_buf_setup_plane functions
have been merged
- queue initialization has been adapted to the new API
- use of the allocator memops has been changed, now there are single
memops for all the allocator contexts
2) Split of the s5p_mfc_try_run and s5p_mfc_handle_frame_int functions
- parts of the s5p_mfc_try_run function have been moved to separate
functions (s5p_mfc_get_new_ctx, s5p_mfc_run_dec_last_frames,
s5p_mfc_run_dec_frame, s5p_mfc_run_get_inst_no, s5p_mfc_run_return_inst
s5p_mfc_run_init_dec,s5p_mfc_run_init_dec_buffers)
- s5p_mfc_handle_frame_int has been split to the following functions:
s5p_mfc_handle_frame_all_extracted, s5p_mfc_handle_frame_new
and s5p_mfc_handle_frame to handle different cases
3) Remove remaining magic numbers and tidy up comments
==================
Changes since v2
==================
1) Update to newest videobuf2 API
This is the major change from v2. The videobuf2 API will hopefully have no more
major API changes. Buffer initialization has been moved from buf_prepare
callback to buf_init to simplify the process. Locking mechanism has been
modified to the requirements of new videobuf2 version.
2) Code cleanup
Removed more magic contants and replaced them with appropriate defines. Changed
code to use unlocked_ioctl instead of ioctl in v4l2 file ops.
3) Allocators
All internal buffer allocations are done using the selected vb2 allocator,
instead of using CMA functions directly.
==================
Changes since v1
==================
1) Cleanup accoridng to Peter Oh suggestions on the mailing list (Thanks).
* Original cover letter:
==============
Introduction
==============
The purpose of this RFC is to discuss the driver for a hw video codec
embedded in the new Samusng's SoCs. Multi Format Codec 5.0 is able to
handle video decoding of in a range of formats.
So far no hardware codec was supported in V4L2 and this would be the
first one. I guess there are more similar device that would benefit from
a V4L2 unified interface. I suggest a separate control class for codec
devices - V4L2_CTRL_CLASS_CODEC.
Internally the driver uses videobuf2 framework and CMA memory allocator.
I am aware that those have not yet been merged, but I wanted to start
discussion about the driver earlier so it could be merged sooner. The
driver posted here is the initial version, so I suppose it will require
more work.
==================
Device interface
==================
The driver principle is based on the idea of memory-to-memory devices:
it provides a single video node and each opened file handle gets its own
private context with separate buffer queues. Each context consist of 2
buffer queues: OUTPUT (for source buffers, i.e. encoded video frames)
and CAPTURE (for destination buffers, i.e. decoded raw video frames).
The process of decoding video data from stream is a bit more complicated
than typical memory-to-memory processing, that's why the m2m framework
is not directly used (it is too limited for this case). The main reason
for this is the fact that the CAPTURE buffers can be dequeued in a
different order than they queued. The hw block decides which buffer has
been completely processed. This is due to the structure of most
compressed video streams - use of B frames causes that decoding and
display order may be different.
==============================
Decoding initialization path
==============================
First the OUTPUT queue is initialized. With S_FMT the application
chooses which video format to decode and what size should be the input
buffer. Fourcc values have been defined for different codecs e.g.
V4L2_PIX_FMT_H264 for h264. Then the OUTPUT buffers are requested and
mmaped. The stream header frame is loaded into the first buffer, queued
and streaming is enabled. At this point the hardware is able to start
processing the stream header and afterwards it will have information
about the video dimensions and the size of the buffers with raw video
data.
The next step is setting up the CAPTURE queue and buffers. The width,
height, buffer size and minimum number of buffers can be read with G_FMT
call. The application can request more output buffer if necessary. After
requesting and mmaping buffers the device is ready to decode video
stream.
The stream frames (ES frames) are written to the OUTPUT buffers, and
decoded video frames can be read from the CAPTURE buffers. When no more
source frames are present a single buffer with bytesused set to 0 should
be queued. This will inform the driver that processing should be
finished and it can dequeue all video frames that are still left. The
number of such frames is dependent on the stream and its internal
structure (how many frames had to be kept as reference frames for
decoding, etc).
===============
Usage summary
===============
This is a step by step summary of the video decoding (from user
application point of view, with 2 treads and blocking api):
01. S_FMT(OUTPUT, V4L2_PIX_FMT_H264, ...)
02. REQ_BUFS(OUTPUT, n)
03. for i=1..n MMAP(OUTPUT, i)
04. put stream header to buffer #1
05. QBUF(OUTPUT, #1)
06. STREAM_ON(OUTPUT)
07. G_FMT(CAPTURE)
08. REQ_BUFS(CAPTURE, m)
09. for j=1..m MMAP(CAPTURE, j)
10. for j=1..m QBUF(CAPTURE, #j)
11. STREAM_ON(CAPTURE)
display thread:
12. DQBUF(CAPTURE) -> got decoded video data in buffer #j
13. display buffer #j
14. QBUF(CAPTURE, #j)
15. goto 12
parser thread:
16. put next ES frame to buffer #i
17. QBUF(OUTPUT, #i)
18. DQBUF(OUTPUT) -> get next empty buffer #i 19. goto 16
...
Similar usage sequence can be achieved with single threaded application
and non-blocking api with poll() call.
Branch with MFC, CMA and videobuf2 will be soon available at
http://git.infradead.org/users/kmpark/linux-2.6-samsung/shortlog/refs/heads/vb2-mfc-fimc
This tree is based on 2.6.37 rc4.
Please have a look at the code and the idea of how to introduce codec
devices to V4L2. Comments will be very much appreciated.
Patch summary:
Kamil Debski (5):
Changes in include/linux/videodev2.h for MFC
MFC: Add MFC 5.1 driver to plat-s5p, mach-s5pv210 and mach-s5pv210
MFC: Added MFC 5.1 V4L2 driver
s5pv310: Enable MFC on C210_universal
v4l: Documentation for the codec interface
Documentation/DocBook/v4l/dev-codec.xml | 169 ++-
arch/arm/mach-s5pv310/Kconfig | 1 +
arch/arm/mach-s5pv310/clock.c | 28 +-
arch/arm/mach-s5pv310/include/mach/map.h | 2 +
arch/arm/mach-s5pv310/include/mach/regs-clock.h | 3 +
arch/arm/mach-s5pv310/mach-universal_c210.c | 8 +
arch/arm/plat-s5p/Kconfig | 5 +
arch/arm/plat-s5p/Makefile | 2 +-
arch/arm/plat-s5p/dev-mfc.c | 49 +
arch/arm/plat-samsung/include/plat/devs.h | 1 +
drivers/media/video/Kconfig | 8 +
drivers/media/video/Makefile | 1 +
drivers/media/video/s5p-mfc/Makefile | 3 +
drivers/media/video/s5p-mfc/regs-mfc5.h | 346 ++++
drivers/media/video/s5p-mfc/s5p_mfc.c | 2253 +++++++++++++++++++++++
drivers/media/video/s5p-mfc/s5p_mfc_common.h | 240 +++
drivers/media/video/s5p-mfc/s5p_mfc_ctrls.h | 182 ++
drivers/media/video/s5p-mfc/s5p_mfc_debug.h | 47 +
drivers/media/video/s5p-mfc/s5p_mfc_intr.c | 92 +
drivers/media/video/s5p-mfc/s5p_mfc_intr.h | 26 +
drivers/media/video/s5p-mfc/s5p_mfc_memory.h | 43 +
drivers/media/video/s5p-mfc/s5p_mfc_opr.c | 913 +++++++++
drivers/media/video/s5p-mfc/s5p_mfc_opr.h | 142 ++
include/linux/videodev2.h | 39 +
24 files changed, 4587 insertions(+), 16 deletions(-)
create mode 100644 arch/arm/plat-s5p/dev-mfc.c
create mode 100644 drivers/media/video/s5p-mfc/Makefile
create mode 100644 drivers/media/video/s5p-mfc/regs-mfc5.h
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc.c
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc_common.h
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc_ctrls.h
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc_debug.h
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc_intr.c
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc_intr.h
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc_memory.h
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc_opr.c
create mode 100644 drivers/media/video/s5p-mfc/s5p_mfc_opr.h
--
To unsubscribe from this list: send the line "unsubscribe linux-samsung-soc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
