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. It's a first shot at real problems encountered in real situations with real hardware (namely high resolution still image capture on OMAP3). It's far from perfect, and I'm open to completely different solutions if someone thinks of one. Introduction ============ The V4L2 video buffers handling API makes use of a queue of video buffers to 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 and large video resolutions. This RFC will list the current limitations of the API 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. Limitations =========== Large buffers allocation ------------------------ Many video devices still require physically contiguous memory. The introduction of IOMMUs on high-end systems will probably make that a distant 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 memory). Allocating large amounts of physically contiguous memory needs to be done as 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 few drivers allocate a large chunk of memory when they are loaded and then use it when a userspace application requests video buffers to be allocated. However, that method requires guessing how much memory will be needed, and can lead to waste of system memory (if the guess was too large) or allocation failures (if the guess was too low). Buffer queuing latency ----------------------- VIDIOC_QBUF is becoming a performance bottleneck when capturing large images 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 all 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 invalidate (for video capture) or flush (for video output) the cache (either a range, or the whole cache) every time a buffer is queued. This ensures that stale data in the cache will not be written back to memory during or after DMA and that 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 cache. When video buffers are passed between two devices (for instance when passing 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 be skipped. - 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 when allocated (VIDIOC_REQBUFS). However, when using the USERPTR streaming method, the buffers can only be locked the first time they are queued, adding to the VIDIOC_QBUF latency. A similar issue arises when using IOMMUs. The IOMMU needs to be programmed to translate physically scattered pages into a contiguous memory range on the bus. This operation is done the first time buffers are queued for USERPTR buffers. Sharing buffers between devices ------------------------------- Video buffers memory can be shared between several devices when at most one of 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 case). However, the use of USERPTR buffers comes with restrictions compared to MMAP. Most architectures don't offer any API to DMA data to/from userspace buffers. 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 kernel-allocated 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 a 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 contiguous) - 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 might want to restrict the amount of RAM used by the Linux kernel to reserve memory 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 is not required, as buffers in that range will be allocated by applications that 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 the requested number of buffers (v4l2_requestbuffers::count) but does not allocate any buffer. Later, the userspace application calls VIDIOC_QBUF to queue buffers from the pool to the video node queue. It sets v4l2_buffer::memory to V4L2_MEMORY_POOL 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 is not limited to, verifying the buffer size. Some devices might require contiguous memory, in which case the driver must check if the buffer is contiguous. 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 with MMAP or USERPTR buffers. The ability to perform those operations when allocating the group speeds up the VIDIOC_QBUF operation, decreasing the still picture shot latency. Once a buffer from a group is queued, the group is market as active and can't 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 driver 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 touch 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 driver to skip cache invalidation and cleaning. Another option would be to base the decision whether to invalidate/flush the 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 thus not required. However, this wouldn't work for USERPTR-like buffer groups, but 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 includes - 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 both 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 majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html