A New Proposal For Supporting DRM Linux Drivers In User-Space

[starshipeleven]

"Uhm, not the same thing."

how so?

Already said above, this isn't a high-speed interface suitable for good GPUs (just look at the latencies stated), but an auxiliary path for crappy 2D-only or similar (dongles or embedded-grade GPUs) where the lower performance isn't going to be a bottleneck anyway.

On Windows you have all GPU drivers in kinda userspace, which is a different thing.

[starshipeleven]

Products that allow you to connect a full GPU to a laptop will be available for purchase before this is merged into the kernel. I think at least 2 have already been announced at CES this week.

I could buy the same stuff since at least 2 years anyway (direct from china).
If you just need a screen for office-grade stuff you don't need a true GPU to run it, and such usb-to-hdmi adapters cost peanuts and are tiny.

[starshipeleven]

All drivers should be in userspace. Microkernel for the win!

userspace drivers = blobs = Windows = nopenopenopenope.

[Staffan]

userspace drivers = blobs = Windows = nopenopenopenope.

That has nothing to do with it. It's perfectly possible to use blobs in kernel space, as both NVidia and AMD do and I'm sure several others.

[starshipeleven]

That has nothing to do with it. It's perfectly possible to use blobs in kernel space, as both NVidia and AMD do and I'm sure several others.

It's not a matter of possibility (you can blob all you want, you may violate GPL but who cares, embedded device manufacturers do this all the time), but of not becoming a crappier windows.

Allowing stable kernel interfaces for drivers (userspace or not) would mean one of the main benefits of linux stops (i.e. drivers that can be recompiled for all archs, modified or fixed by third parties and so on), and then it becomes a crappier Windows, that makes no sense to use at all.

If this interface is not stable like other kernel-driver interfaces, you are only sacrificing performance (as stated in the article, a latency of 13 becomes a latency of 33 ms or something), for the sake of letting simpler stuff have a easier time working at all with Linux.
Which is NOT what a microkernel does.