For about a year and a half, the Qualcomm Linux team has been working to support the OpenGL ES 3D core in the Snapdragon processor. The hardware made its debut in the Nexus One and has subsequently been used in a few other commercial products since then. To support the 3D GPU we wrote a kernel based driver to manage all the usual graphics concerns - interrupts, command streams, context switching, etc, etc. You can see the latest and greatest incarnation of our driver here:
https://www.codeaurora.org/gitweb/quic/la/?p=kernel/msm.git;a=tree;f=drivers...
I'm writing this email because we think it is high time that we get off the bench, into the game and push support for the Qualcomm graphics cores to the mainline kernel. We are looking for advice and comment from the community on the approach we have taken and what steps we might need to take, if any, to modify the driver so it can be accepted. I'm going to offer up a quick description of the hardware and describe our current approach as well as our current development plan for the summer. Our goal is to start pushing this code upstream as quickly as possible, so comments and flames would be greatly appreciated.
=====
The hardware layout is reasonably straight forward. The GPU is, as expected, a platform device located on the same die as the processor. The registers are mapped into the physical address space of the processor. There device also shares memory with the main processor; there is no dedicated memory on board for the GPU. The GPU has a built in MMU for mapping paged memory.
Some processor variants also have a separate 2D core attached. The 2D core is also a platform device with shared memory and a MM. While the general interface is similar to the 3D core, the 2D GPU has its own separate pipeline and interrupt.
The core of the driver is a home-grown ioctl() API through a character device we call /dev/kgsl. Like other GPU drivers, all the heavy lifting is done by the userspace drivers so the core set of ioctls are mainly used to access the hardware or to manage contexts:
IOCTL_KGSL_DEVICE_GETPROPERTY IOCTL_KGSL_DEVICE_REGREAD IOCTL_KGSL_DEVICE_REGWRITE IOCTL_KGSL_RINGBUFFER_ISSUEIBCMDS IOCTL_KGSL_DEVICE_WAITTIMESTAMP IOCTL_KGSL_CMDSTREAM_READTIMESTAMP IOCTL_KGSL_DRAWCTXT_CREATE IOCTL_KGSL_DRAWCTXT_DESTROY
In the early days of the driver, any memory used for the GPU (command buffers, color buffers, etc) were allocated by the user space driver through PMEM (a simple contiguous memory allocator written for Android; see drivers/misc/pmem.c). PMEM is not ideal because the contiguous memory it uses needs to be carved out of bootmem at init time and is lost to the general system pool, so the driver was switched to use paged memory allocated via vmalloc() and mapped into the GPU MMU. As a result, a handful of additional ioctls were put into the driver to support allocating and managing the memory.
When we started our X11 effort last year, we needed a DRM driver to support DRI2. We added a DRM skeleton to the existing GPU driver (this was the driving force behind the platform DRM patches I've sent out periodically). The DRM driver allocates its own memory for buffers to support GEM, and the buffers are mapped into the GPU MMU prior to rendering. It is important to note that the DRM driver only provides GEM and basic DRM services - the userspace graphics libraries still run rendering through the /dev/kgsl interface.
Then, when support came along for the 2D core, it turned out that most of the support code was identical to that for the 3D GPU, with only a few differences in how the command streams and interrupts were processed. The 2D and 3D code were merged together to form the driver that I linked above. The ioctl calls remained the same, and a "device" member was added to each structure to determine which core the call was destined for. Each device has its own MMU, but each MMU shares the same pagetables.
It has been argued that having the 2D and the 3D together is silly since they are separate platform devices and they should be treated as such - the proposal is to have multiple device nodes, one for each device. Each device will have its iomapped registers, MMU pagetables, etc; while sharing generic support code in the driver:
/dev/kgsl/0 /dev/kgsl/1 etc..
I think that if we did this design we would also need to have an additional device to allocate memory buffers which will make it easier for us to share memory between cores (X11 in particular does a lot of simultaneous 2D and 3D). I also think that the memory allocator should be transitioned to a standard design (probably TTM). Of course for X11 the DRM/GEM driver would still be used with GEM turning into a wrapper for the shared memory allocator.
Thanks for reading, Jordan
PS: About the name: our userspace drivers uses a HAL called GSL (for graphics support layer). Elements of that HAL is what you see today in the kernel driver, so we called it KGSL (Kernel GSL). We used to have the driver in drivers/char/msm_kgsl/ but we were convinced to move it to drivers/gpu/msm, which is already a great improvement in naming and in location. I presume that one of the conditions of upstreaming would be to rename everything to something a little bit more descriptive and a little bit less cryptic.