DMA-BUF Is Ready To Push Forward Linux Drivers
Besides the Intel chatter, Nouveau banging on OpenCL, Coreboot's less than interesting laptop, one of the other exciting meetings this past weekend in Brussels at FOSDEM was about DMA-BUF. DMA-BUF will help a range of Linux hardware drivers from embedded devices to multi-GPU desktops.
DMA-BUF is the buffer sharing mechanism being introduced in the Linux 3.3 kernel with the original design coming out of Linaro. DMA-BUF is meant to offer zero-copy buffer sharing between multiple Linux kernel device drivers, with the original developers seeking this support being many ARM SoC vendors. However, there's now a whole range of potential DMA-BUF users.
Besides for using DMA-BUF in the embedded space for easily sharing a buffer between multiple drivers all touching the same SoC, DMA-BUF can be used on the Linux desktop for sharing buffers between multiple graphics drivers in the case of multi-GPU technologies like AMD CrossFire and NVIDIA SLI. DMA-BUF could also be implemented for use in cases like NVIDIA Optimus with sharing a buffer between a driver for an integrated GPU and discrete graphics processor. With graphics drivers it can also come in to play for GPGPU/OpenCL buffer sharing. There's also other interesting niche cases to use DMA-BUF for like wireless displays.
Daniel Vetter, the Intel employee since last year, was one of the DRM developers involved with DMA-BUF and talked about it at FOSDEM 2012 this past weekend.
Besides going over DMA-BUF and what it can do, he did talk about some of the current missing features like CPU access from kernel-space, user-space access (i.e. mmap support), eviction support, etc. There's also some user-space features he's wanting to add like a GBM/EGL extension interfacing with DMA-BUF, format/layout aribtrage, and some hooks for Wayland.
The initial DMA-BUF implementation landed in the Linux 3.3 kernel, but none of the kernel drivers in 3.3 are currently relying upon this infrastructure for the upstream release. Hopefully with the Linux 3.4 kernel is where we'll see some drivers using DMA-BUF for buffer sharing. There's also a problem though in that the binary Linux drivers can't use DMA-BUF, so right now it's out of the question for hooking in the AMD Catalyst or NVIDIA graphics drivers from using this infrastructure to play better with the open-source drivers, since all of the DMA-BUF exported symbols are marked for GPL-only use. There's been no signalling by the developers they are willing to make this change.
Below is my video recording of Daniel's DMA-BUF presentation.
DMA-BUF is the buffer sharing mechanism being introduced in the Linux 3.3 kernel with the original design coming out of Linaro. DMA-BUF is meant to offer zero-copy buffer sharing between multiple Linux kernel device drivers, with the original developers seeking this support being many ARM SoC vendors. However, there's now a whole range of potential DMA-BUF users.
Besides for using DMA-BUF in the embedded space for easily sharing a buffer between multiple drivers all touching the same SoC, DMA-BUF can be used on the Linux desktop for sharing buffers between multiple graphics drivers in the case of multi-GPU technologies like AMD CrossFire and NVIDIA SLI. DMA-BUF could also be implemented for use in cases like NVIDIA Optimus with sharing a buffer between a driver for an integrated GPU and discrete graphics processor. With graphics drivers it can also come in to play for GPGPU/OpenCL buffer sharing. There's also other interesting niche cases to use DMA-BUF for like wireless displays.
Daniel Vetter, the Intel employee since last year, was one of the DRM developers involved with DMA-BUF and talked about it at FOSDEM 2012 this past weekend.
Besides going over DMA-BUF and what it can do, he did talk about some of the current missing features like CPU access from kernel-space, user-space access (i.e. mmap support), eviction support, etc. There's also some user-space features he's wanting to add like a GBM/EGL extension interfacing with DMA-BUF, format/layout aribtrage, and some hooks for Wayland.
The initial DMA-BUF implementation landed in the Linux 3.3 kernel, but none of the kernel drivers in 3.3 are currently relying upon this infrastructure for the upstream release. Hopefully with the Linux 3.4 kernel is where we'll see some drivers using DMA-BUF for buffer sharing. There's also a problem though in that the binary Linux drivers can't use DMA-BUF, so right now it's out of the question for hooking in the AMD Catalyst or NVIDIA graphics drivers from using this infrastructure to play better with the open-source drivers, since all of the DMA-BUF exported symbols are marked for GPL-only use. There's been no signalling by the developers they are willing to make this change.
Below is my video recording of Daniel's DMA-BUF presentation.
1 Comment