What is the lowest level open/public API through which Linux graphics system talks to the GPU?

Question

I've got the idea that in old days, XFree86 used /dev/fb* framebuffer devices. But now looks like the GUI system uses OpenGL, which is an open standard. So is OpenGL the lowest level open API? By 'open', I mean no obscure private ioctl or closed-source stuff.

Update:

I've got what I need after discussions with @datenwolf, who also provides a great answers. In case anyone wants more coding detail: The X server driver is the lowest common level for the X system on top of any GPUs. To know what kind of interface an X server driver must implement: refer to DDX Design, it is a detailed doc for the latest X.org that tells you how to write an X driver.

Answer 1

In Linux there is no "lowest level API" exposed by the kernel that does graphics operations. All graphics drivers are actually implemented in the user space by so called "state trackers", which use special kernel functions to talk directly with the graphics hardware¹. The open source drivers implemented as part of the Mesa project use the Linux specific DRM API to talk to the hardware. The proprietary drivers from NVidia and AMD each use their very specific kernel module instead.

Now when it comes to end program usability Mesa and the proprietary drivers differ a bit:

NVidia's and AMD/ATI's proprietary drivers offer no API at all to end user programs to use for. Instead they're implemented as modules to load by the X server; the X server expects the driver modules to follow a specific scheme, which usually changes with major changes in the X server, so each X server major version bump usually required to update the driver modules as well.

The X server in turn provides a well known command stream based graphics API. Graphics commands sent over this API are scheduled by the X server and dispatched into called to the right functions of the driver module. The driver module in turn contains the whole intelligence for talking with the GPU and turns the commands coming from the X server into commands toward the GPU.

So to speak the X server is the lowest level universal graphics API currently available to Linux programs (except for programs that would go the length to implement everything to use the X driver modules directly).

When it comes to X11, the drivers that are part of the Mesa project are no different than the proprietary drivers. However because Mesa is open source its developers began implementing ways that programs could use Mesa and its drivers directly without having talk to it through the X server. For that they choose to expose an API conformant with the EGL specification. Unfortunately EGL is rather useless on its own because it requires an external display system to be present (or you can do only off-screen rendering).

This is where Wayland enters the picture. Wayland is not a display system. It is however a protocol that allows the building blocks of a display system to talk with each other. One central component of a Wayland based display system is the compositor which actually takes hold of the so called "seat" (= display device and associated input devices). Programs that want to display something on the screen use Wayland to open a connection with the compositor, which gives them a display to use EGL on. That they then can use to create actual drawing surfaces to do their graphics on.

---

[1]: Actually for some drivers (of legacy hardware) the X server must be started with root privileges, so that it can use the special function ioperm(…) to gain direct access to the hardware using the out*(…) and in*(…) functions and open /dev/mem for memory access. In this case there is zero support by the kernel regarding the communication with the graphics hardware. But for security and performance reasons nobody does (program) that anymore.