Contiguous memory allocations

Does anyone know if there's a common infrastructure for allocation of DMA'able memory by drivers and applications above the straight kernel API (dma_alloc_coherent)?

I'm working with Freescale i.MX51 drivers to do 720P video input and output and the embedded calls to dma_alloc_coherent fail except when used right after boot because of fragmentation.

I'm fighting the urge to write yet another special-purpose allocator for video buffers thinking this must be a common problem with a solution already, but I can't seem to locate one.

The closest thing I've found is the bigphysarea patch, which doesn't appear to be supported or headed toward main-line.

Thanks in advance,

Quoted 15 lines · Eric Nelson

Does anyone know if there's a common infrastructure for allocation of DMA'able memory by drivers and applications above the straight kernel API (dma_alloc_coherent)?

I'm working with Freescale i.MX51 drivers to do 720P video input and output and the embedded calls to dma_alloc_coherent fail except when used right after boot because of fragmentation.

I'm fighting the urge to write yet another special-purpose allocator for video buffers thinking this must be a common problem with a solution already, but I can't seem to locate one.

The closest thing I've found is the bigphysarea patch, which doesn't appear to be supported or headed toward main-line.

Quoted 1 line

Thanks in advance,

dma_alloc_coherent is pretty much just a wrapper round get_free_pages, which is the lowest level allocator in the kernel. So, no there is no other option (but see below). The simplest thing is to make sure your driver is loaded at boot time and to grab all the memory you need then and never let it go. That's what I do.

If you are desperate, you can use the bigphysarea patch - it's quite common on streaming video devices - but you will have to port it to your kernel. Or, you can restrict the memory the kernel uses with something like "mem=128M" on the command line and take that above 128M for yourself. You will have to map it in with ioremap(_nocache).

Bye for now,
Chris.

Quoted 16 lines · Chris Simmonds

Quoted 15 lines · Eric Nelson

Does anyone know if there's a common infrastructure for allocation of DMA'able memory by drivers and applications above the straight kernel API (dma_alloc_coherent)?

I'm working with Freescale i.MX51 drivers to do 720P video input and output and the embedded calls to dma_alloc_coherent fail except when used right after boot because of fragmentation.

I'm fighting the urge to write yet another special-purpose allocator for video buffers thinking this must be a common problem with a solution already, but I can't seem to locate one.

The closest thing I've found is the bigphysarea patch, which doesn't appear to be supported or headed toward main-line.

Quoted 8 lines

Quoted 1 line

Thanks in advance,

dma_alloc_coherent is pretty much just a wrapper round get_free_pages, which is the lowest level allocator in the kernel. So, no there is no other option (but see below). The simplest thing is to make sure your driver is loaded at boot time and to grab all the memory you need then and never let it go. That's what I do.

Thanks Chris.

The trouble is always "how much"? If we don't know at startup what kind of video's needed or what size(s) of camera input may be needed, it's impossible
to tune. In the current Freescale kernels, there are at least 4 separate drivers that allocate RAM, sometimes for internal use, but mostly in response
to userspace calls (ioctl).

- frame-buffer driver
- Video Processing Unit (VPU) - video encode/decode
- V4L2 output device - allows access to YUV output layer, color blending
- Image Processing Unit (IPU) - allows userspace bitblts through DMA

With this number of calls, tuning with separate kernel command-line args seems
unworkable.

Quoted 38 lines · Eric Nelson

Quoted 16 lines · Chris Simmonds

Quoted 15 lines · Eric Nelson

Does anyone know if there's a common infrastructure for allocation of DMA'able memory by drivers and applications above the straight kernel API (dma_alloc_coherent)?

I'm working with Freescale i.MX51 drivers to do 720P video input and output and the embedded calls to dma_alloc_coherent fail except when used right after boot because of fragmentation.

I'm fighting the urge to write yet another special-purpose allocator for video buffers thinking this must be a common problem with a solution already, but I can't seem to locate one.

The closest thing I've found is the bigphysarea patch, which doesn't appear to be supported or headed toward main-line.

Quoted 8 lines

Quoted 1 line

Thanks in advance,

dma_alloc_coherent is pretty much just a wrapper round get_free_pages, which is the lowest level allocator in the kernel. So, no there is no other option (but see below). The simplest thing is to make sure your driver is loaded at boot time and to grab all the memory you need then and never let it go. That's what I do.

Thanks Chris.

The trouble is always "how much"? If we don't know at startup what kind of video's needed or what size(s) of camera input may be needed, it's impossible to tune. In the current Freescale kernels, there are at least 4 separate drivers that allocate RAM, sometimes for internal use, but mostly in response to userspace calls (ioctl).

- frame-buffer driver
- Video Processing Unit (VPU) - video encode/decode
- V4L2 output device - allows access to YUV output layer, color blending
- Image Processing Unit (IPU) - allows userspace bitblts through DMA

Quoted 2 lines

With this number of calls, tuning with separate kernel command-line args seems unworkable.

I think the kernel developers don't like this kind of on-the-side allocator because they tend to be dedicated to solving one kind of problem.

Here are a few thoughts about the imx51 specifically, based on my experience. First, the size of the memory pool used for dma_alloc_coherent is set in plat-mxc/include/mach/memory.h where it is hard coded to 64 MiB. You could try bumping that up a bit.

Second, you could re-do the buffer allocation and replace dma_alloc_coherent with kmalloc and then use dma_map_single to lock it down while dma is taking place. This way you avoid the 64M dma pool limit and you speed up buffer access via mmap because the memory is cached. In my case I got a two fold speed improvement reading frames into application memory. I have to admit that my case was a bit specialised though and it may not be worth the effort for you.

Bye for now,
Chris.