Features Download
From: Laurent Pinchart <laurent.pinchart <at> ideasonboard.com>
Subject: [RFC] Global video buffers pool
Newsgroups: gmane.linux.drivers.video-input-infrastructure
Date: Wednesday 16th September 2009 15:46:39 UTC (over 8 years ago)
Hi everybody,

I didn't want to miss this year's pretty flourishing RFC season, so here's 
another one about a global video buffers pool.

All comments are welcome, but please don't trash this proposal too fast.
a first shot at real problems encountered in real situations with real 
hardware (namely high resolution still image capture on OMAP3). It's far
perfect, and I'm open to completely different solutions if someone thinks


The V4L2 video buffers handling API makes use of a queue of video buffers
exchange data between video devices and userspace applications (the read 
method don't expose the buffers objects directly but uses them underneath).

Although quite efficient for simple video capture and output use cases, the

current implementation doesn't scale well when used with complex hardware
large video resolutions. This RFC will list the current limitations of the
and propose a possible solution.

The document is at this stage a work in progress. Its main purpose is to be

used as support material for discussions at the Linux Plumbers Conference.


Large buffers allocation

Many video devices still require physically contiguous memory. The 
introduction of IOMMUs on high-end systems will probably make that a
nightmare in the future, but we have to deal with this situation for the 
moment (I'm not sure if the most recent PCI devices support scatter-gather 
lists, but many embedded systems still require physically contiguous

Allocating large amounts of physically contiguous memory needs to be done
soon as possible after (or even during) system bootup, otherwise memory 
fragmentation will cause the allocation to fail.

As the amount of required video memory depends on the frame size and the 
number of buffers, the driver can't pre-allocate the buffers beforehand. A
drivers allocate a large chunk of memory when they are loaded and then use
when a userspace application requests video buffers to be allocated.
that method requires guessing how much memory will be needed, and can lead
waste of system memory (if the guess was too large) or allocation failures
the guess was too low).

Buffer queuing latency

VIDIOC_QBUF is becoming a performance bottleneck when capturing large
on some systems (especially in the embedded world). When capturing high 
resolution still pictures, the VIDIOC_QBUF delay adds to the shot latency, 
making the camera appear slow to the user.

The delay is caused by several operations required by DMA transfers that
happen when queuing buffers.

- Cache coherency management

When the processor has a non-coherent cache (which is the case with most 
embedded devices, especially ARM-based) the device driver needs to
(for video capture) or flush (for video output) the cache (either a range,
the whole cache) every time a buffer is queued. This ensures that stale
in the cache will not be written back to memory during or after DMA and
all data written by the CPU is visible to the device.

Invalidating the cache for large resolutions take a considerable amount of 
time. Preliminary tests showed that cache invalidation for a 5MP buffer 
requires several hundreds of milliseconds on an OMAP3 platform for range 
invalidation, or several tens of milliseconds when invalidating the whole D


When video buffers are passed between two devices (for instance when
the same USERPTR buffer to a video capture device and a hardware codec) 
without any userspace access to the memory, CPU cache invalidation/flushing

isn't required on either side (video capture and hardware codec) and could

- Memory locking and IOMMU

Drivers need to lock the video buffer pages in memory to make sure that the

physical pages will not be freed while DMA is in progress under low-memory 
conditions. This requires looping over all pages (typically 4kB long) that 
back the video buffer (10MB for a 5MP YUV image) and takes a considerable 
amount of time.

When using the MMAP streaming method, the buffers can be locked in memory
allocated (VIDIOC_REQBUFS). However, when using the USERPTR streaming
the buffers can only be locked the first time they are queued, adding to
VIDIOC_QBUF latency.

A similar issue arises when using IOMMUs. The IOMMU needs to be programmed
translate physically scattered pages into a contiguous memory range on the 
bus. This operation is done the first time buffers are queued for USERPTR 

Sharing buffers between devices

Video buffers memory can be shared between several devices when at most one
them uses the MMAP method, and the others the USERPTR method. This avoids 
memcpy() operations when transferring video data from one device to another

through memory (video acquisition -> hardware codec is the most common use 

However, the use of USERPTR buffers comes with restrictions compared to
Most architectures don't offer any API to DMA data to/from userspace
Beside, kernel-allocated buffers could be fine-tuned by the driver (making 
them non-cacheable when it makes sense for instance), which is not possible

when allocating the buffers in userspace.

For that reason it would be interesting to be able to share
video buffers between devices.

Video buffers pool

Instead of having separate buffer queues at the video node level, this RFC 
proposes the creation of a video buffers pool at the media controller level

that can be used to pre-allocate and pre-queue video buffers shared by all 
video devices created by the media controller.

Depending on the implementation complexity, the pool could even be made 
system-wide and shared by all video nodes.

Allocating buffers

The video buffers pool will handle independent groups of video buffers.

        allocate               request
(NULL)   ----->   (ALLOCATED)   ----->   (ACTIVE)
         <----                  <-----
          free                 release

Video buffers groups allocation is controlled by userspace. When allocating
buffers group, an application will specify

- the number of buffers
- the buffer size (all buffers in a group have the same size)
- what type of physical memory to allocate (virtual or physically
- whether to lock the pages in memory
- whether to invalidate the cache

Once allocated, a group becomes ALLOCATED and is given an ID by the kernel.

When dealing with really large video buffers, embedded system designers
want to restrict the amount of RAM used by the Linux kernel to reserve
for video buffers. This use case should be supported. One possible solution

would be to set the reserved RAM address and size as module parameters, and

let the video buffers pool manage that memory. A full-blown memory manager
not required, as buffers in that range will be allocated by applications
know what they're doing.

Queuing the buffers

Buffers can be used by any video node that belongs to the same media 
controller as the buffer pool.

To use buffers from the video buffers pool, a userspace application calls 
VIDIOC_REQBUFS on the video node and sets the memory field to 
V4L2_MEMORY_POOL. The video node driver creates a video buffers queue with
requested number of buffers (v4l2_requestbuffers::count) but does not
any buffer.

Later, the userspace application calls VIDIOC_QBUF to queue buffers from
pool to the video node queue. It sets v4l2_buffer::memory to
and v4l2_buffer::m to the ID of the buffers group in the pool.

The driver must check if the buffer fulfills its needs. This includes, but
not limited to, verifying the buffer size. Some devices might require 
contiguous memory, in which case the driver must check if the buffer is 

Depending whether the pages have been locked in memory and the cache 
invalidated when allocating the buffers group in the pool, the driver might

need to lock pages and invalidate the cache at this point, is it would do
MMAP or USERPTR buffers. The ability to perform those operations when 
allocating the group speeds up the VIDIOC_QBUF operation, decreasing the
picture shot latency.

Once a buffer from a group is queued, the group is market as active and
be freed until all its buffers are released.

Dequeuing and using the buffers

V4L2_MEMORY_POOL buffers are dequeued similarly to MMAP or USERPTR buffers.

Applications must set v4l2_buffer::memory to V4L2_MEMORY_POOL and the
will set v4l2_buffer::m to the buffers group ID.

The buffer can then be used by the application and queued back to the same 
video node, or queued to another video node. If the application doesn't
the buffer memory (neither reads from nor writes to memory) it can set 
v4l2_buffer::flags to the new V4L2_BUF_FLAG_NO_CACHE value to tell the
to skip cache invalidation and cleaning.

Another option would be to base the decision whether to invalidate/flush
cache on whether to buffer is currently mmap'ed in userspace. A non-mmap'ed

buffer can't be touched by userspace, and cache invalidation/flushing is
not required. However, this wouldn't work for USERPTR-like buffer groups,
those are not supported at the moment.

Freeing the buffers

A buffer group can only be freed if all its buffers are not in use. This 

- all buffers that have been mmap'ed must have been unmap'ed
- no buffer can be queued to a video node

If both conditions are fulfilled, all buffers in the group are unused by
userspace and kernelspace. They can then be freed.

Laurent Pinchart
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
CD: 5ms