generic EGL/EGLstreams based video driver (to replace Nvidia proprietary xorg driver)

[metux]

Describe the idea

• similar to generic modesetting, but entirely via EGL (+EGLstreams)
• driving the display quite the same way like wayland compositors doing it
• should work since driver version 390

It should be done because

1. proprietary drivers are just a huge mess, causing a lot of trouble and hindering refactoring
2. future driver version might stop supporting X

What are the alternatives?

full reverse engineering (finish up nouveau for recent gpu models)

Additional context

No response

Extra fields

• I have checked the existing New Ideas, RFCs Of The Core Team, Ideas Soon To Be Addressed and Someday Ideas
• I have read the Contributing Guidelines
• I'd like to work on this idea

[metux]

@callmetango @cepelinas9000 @mikedld @josephcrowell @jonhermansen @probonopd @artist4xlibre @algrid @b-aaz
what are your thoughts ?

[metux]

@felipe19930

[callmetango]

The idea sounds reasonable to me. I just can't comment on the possible technical pitfalls.

[metux]

Had some brainstorming w/ chatgpt on how we can do this. Several interesting insights:

1. a simple (framebuffer-only) driver based on Nvidia's libGL.so isn't complicated
2. extending it to use glamor for 2d rendering shouldn't be difficult
3. nvidia uses it's own NV-GLX extension instead of GLX (context creation, buffer transfers, sync, ...) --> we would need to RE and implement it on our own
4. they're using EGLstreams for buffer fd sharing (not DRM fd's, their own ones)
5. new drivers also can use dmabuf and DRI3 for Xserver communication

Possible project plan:

1. write a simple modesetting-like driver that's operating on proprietary libGL.so (no DRI / NV-GL yet), using glamor for 2d rendering
2. pick a specific driver version (eg. 470)
3. reverse engineer the NV-GLX protocol and implement it in the Xserver
4. wire up the driver with our own NV-GLX extension
5. once it's working on the chosen driver version, proceed others (maybe opcodes changed ?)

[metux]

chatgpt created some testing / dumping tool for the NV-GLX handshake between client side vs Xserver side driver parts:
https://chatgpt.com/share/68a481e3-b1fc-8005-ae0a-50051a002f9d

[SuperDuperDeou]

I'd say that Modesetting works for Maxwell (580 driver), there are just a few bugs to iron out (namely the cursor bug (X11Libre/xserver#734, the texture_from_pixmap (X11Libre/xserver#373 (comment)), and I'm pretty sure fullscreen windows overscan, but I have to double check), but a driver that specifically targets older GPUs would be neat.

I'm not sure exactly how relevant point 2 is, because 470 and older aren't going to get any updates, 580 (which is only getting bug fixes) and later work with modesetting, so, as soon as the bugs are ironed out, we don't care if Nvidia stops shipping an X11 driver altogether, and looking at how Nvidia's feature deprecation works, that should happen when an architecture goes end of live, Maxwell, Pascal, and Volta just got their end of support in 580 and 585 will still support X11, just not ancient versions of XOrg, the next architecture on the chopping block is Turing, and that's going to be years down the line. In other words, if Nvidia drops X11, it's gonna be years from now, and on stuff that already works with modesetting.

[metux]

I'd say that Modesetting works for Maxwell (580 driver), there are just a few bugs to iron out (namely the cursor bug (X11Libre/xserver#734, the texture_from_pixmap (X11Libre/xserver#373 (comment)), and I'm pretty sure fullscreen windows overscan, but I have to double check), but a driver that specifically targets older GPUs would be neat.

Great to hear. Guess that's w/o acceleration. Or can it already use NVidia's libGL ?

I'm not sure exactly how relevant point 2 is, because 470 and older aren't going to get any updates, 580 (which is only getting bug fixes) and later work with modesetting, so, as soon as the bugs are ironed out, we don't care if Nvidia stops shipping an X11 driver altogether,

It's not just about getting older drivers running, but also keep it running, and still being able to refactor the ABI.
The best scenario would be if wouldn't need them at all (inside the Xserver). That's why I'm looking for a way to write our own free xf86 drivers, that's using their proprietary libGL (until hopefully mesa/nouveau might keep up some day and make even this obsolete).

[SuperDuperDeou]

I'd say that Modesetting works for Maxwell (580 driver), there are just a few bugs to iron out (namely the cursor bug (X11Libre/xserver#734, the texture_from_pixmap (X11Libre/xserver#373 (comment)), and I'm pretty sure fullscreen windows overscan, but I have to double check), but a driver that specifically targets older GPUs would be neat.

Great to hear. Guess that's w/o acceleration. Or can it already use NVidia's libGL ?

It works with GPU acceleration since X11Libre/xserver#373 , I get roughly the same fps from glxgears between Nvidia proprietary and modesetting, if it wasn't for the texture_from_pixmap bug I'd probably be daily driving it. The catch is that it doesn't work on the 470 branch and older.

I'm not sure exactly how relevant point 2 is, because 470 and older aren't going to get any updates, 580 (which is only getting bug fixes) and later work with modesetting, so, as soon as the bugs are ironed out, we don't care if Nvidia stops shipping an X11 driver altogether,

It's not just about getting older drivers running, but also keep it running, and still being able to refactor the ABI. The best scenario would be if wouldn't need them at all (inside the Xserver). That's why I'm looking for a way to write our own free xf86 drivers, that's using their proprietary libGL (until hopefully mesa/nouveau might keep up some day and make even this obsolete).

And I'm not against the idea exclusively because of legacy hardware, since newer hardware works with modesetting.

[callmetango]

It works with GPU acceleration since X11Libre/xserver#373 , I get roughly the same fps from glxgears between Nvidia proprietary and modesetting,

Wow, this is really good news! Note to self: We need to document this and tell the world about it. ;-)

[metux]

It works with GPU acceleration since X11Libre/xserver#373 , I get roughly the same fps from glxgears between Nvidia proprietary and modesetting,

Which libGL ? Proprietary or mesa ?
Is the support matrix in the wiki already up-to-date ?

if it wasn't for the texture_from_pixmap bug I'd probably be daily driving it.

Which bug ?

The catch is that it doesn't work on the 470 branch and older.

hmm, so we still need extra care for the older ones.

It's not just about getting older drivers running, but also keep it running, and still being able to refactor the ABI. The best scenario would be if wouldn't need them at all (inside the Xserver). That's why I'm looking for a way to write our own free xf86 drivers, that's using their proprietary libGL (until hopefully mesa/nouveau might keep up some day and make even this obsolete).

And I'm not against the idea exclusively because of legacy hardware, since newer hardware works with modesetting.

But newer HW still requires proprietary libGL ?

[SuperDuperDeou]

It works with GPU acceleration since X11Libre/xserver#373 , I get roughly the same fps from glxgears between Nvidia proprietary and modesetting,

Which libGL ? Proprietary or mesa ? Is the support matrix in the wiki already up-to-date ?

I don't know, but I assume it's the proprietary libGL.
The support matrix is up to date.

if it wasn't for the texture_from_pixmap bug I'd probably be daily driving it.

Which bug ?

X11Libre/xserver#373 (comment)

[pugly-tech]

It sounds like Generic Buffer Management is a bit more popular than EGLstreams with Gnome supporting it by default and Kwin outright dropping support for EGLstreams. How difficult would it be to make a X11 driver that supports video out using GBM? Is there a standard Wayland driver protocol where if we implement that we can directly run any driver built solely for Wayland?

[metux]

This is specifically about replacing the proprietary Xserver driver, just using Nvidia's libGL.
Newer models and driver versions already working via KMS, so generic modesetting would be an option (perhaps w/ glamor via proprietary libGL). But older ones (needed to replace older drivers, to support older cards) only offering EGLstreams.

If we had our own free drivers for these cards (even if it still needs proprietary libGL), this would be a big step forward - we could get rid of all the ABI issues at once.

[darkhog]

Don't we have Nouveau already?

[metux]

NVidia never actually cared much about Unix/Xorg drivers. The one (and only ?) responsible developer takes a whole year to recognize something's changed - that already had been the case on Xorg. He never replies to mails.
And instead of fixing bugs in their driver, they're instead sending PRs with workarounds in the Xserver.
They don't even have a CI running against a not-so-ancient Xserver.

Feel free trying to motivate them to support us. I don't wanna discourage you.

[darkhog]

NVidia never actually cared much about Unix/Xorg drivers.

Which means they care even less for Wayland ones. Also, it's important to note that a single person can't convince them. The entire community has to make enough noise for them to notice that a) xlibre exist and has future b) is better than xorg c) supporting xlibre is barely different from supporting xorg so they don't have to rewrite the entire thing from scratch like with Wayland.

Finally, I don't think private e-mails are the way to go. A public stink needs to be raised about this.

[metux]

A lot noise in their public user forum might perhaps help.

[smj-cc]

Well, considering the fact that in wayland it DOESN'T work, maybe we should try to get someone at nvidia first to officially support xlibre.

If you're going there, why not just get them to open source their drivers? ;-)

[darkhog]

It would be nice in the long run, however let's pick our battles for now and get them to officially support xlibre.

[YusufKhan-gamedev]

Why not just rewrite the proprietary X driver then? If you do it this way you wont have GPU initialization that the blob driver needs(iirc the device node stuff needs to be done sperately). I think I started on this a bit but I haven't worked on it much.

I could maybe extend xf86-video-nv to include this and then push it on the freedesktop gitlab...

[metux]

Why not just rewrite the proprietary X driver then?

Because we have no idea how exactly to drive these cards. That's why using the proprietary libGL.

If you do it this way you wont have GPU initialization that the blob driver needs(iirc the device node stuff needs to be done sperately). I think I started on this a bit but I haven't worked on it much.

To me it looks like all the magic must be in libGL. Otherwise Wayland couldn't work here.

All we need to do is:
a) properly initialize the GL context (quite like Wayland does it)
b) wire up glamor (for 2d render)
c) find out how buffer passing between proprietary libGL and proprietary Xorg driver works and reimplement that

I could maybe extend xf86-video-nv to include this and then push it on the freedesktop gitlab...

No, it will be a very different driver.

And no, I won't even try to f.d.o - I even can't - because Redhat/IBM banned me.
(btw, the very same Karol Herbst who recently celebrated political murder and called for even more)

[YusufKhan-gamedev]

Because we have no idea how exactly to drive these cards. That's why using the proprietary libGL.

I've read the code others have used to drive the card, and I can drive the card. I think the only read issue will be RANDR and since the CTRL bits are unstable that might be a minor challenge(with version stability), NVRM can probably be targeted for the v5xx+ easily though.

No, it will be a very different driver.

Sure, but there are already 3 very different drivers there(g80, riva, nv), and the common code does checks which are still applicable today. why would we copy/paste or rewrite the common code to be more modern when we can just use it.