I'm building a compositor now, and I've been thinking about multi-seat from the beginning, but I know there are lots of places that assume a single seat, and so this sort of thing won't work for me. It's actually a lot more work than you'd think to support this...
I keep thinking of trying to see how far I get re-doing Niri in River.
The one other constraint that I'd really like to see relaxed is about displays, about being able to merge and subdivide screens as I please. Having a huge ultra-wide display really would be so much better if I could divide it up somehow. This is kind of the headline feature of stilch, which is a pretty neat compositor,
https://github.com/wegel/stilch .
There's been all kinds of rough incredibly hard things going on with Niri just trying to figure out how to handle some of the more interesting copy-paste scenarios, and with really weird sequencing issues across basically ephemeral inputs, that have really plagued the project and made life incredibly bad for downstream apps. This is just my opinion, man, but I think the lack of ambition about these ridiculous over the top features eventually starts hampering what turns out to be relatively prosaic asks. Better models for multi-seat, multi-cursor help scenarios like remote-desktop, and video-chat remote control, that at first don't seem needed, but are fundamentals that pay dividends if you shoot for them.
To add to the long list of caveats: This also breaks assumptions GUI toolkit and applications make about only one of their windows being focused at a time: `gtk_application_get_active_window`, `QApplication::activeWindow`, etc.
It's not as weird as you might think: laptops may have touchscreens (which can be multi-touch) on top of the "normal" cursor. IIRC, X.org can't really handle such cases cleanly.
Sometimes that's true, but sometimes (imo the superior approach) the touchscreen is independent from the cursor. I think touchscreen Chromebooks are like that, for example.
I've always thought this was a cool idea. It's obviously been on the minds of Wayland protocol developers since the beginning, otherwise it wouldn't be possible!
The original title, "State of multi-player Wayland" is a much better summary of the article. This ain't just about the pointers! In fact, the author is very explicit about that:
> I think there could be a term that’s describing this concept even better – in the title of this post, I picked “multi-player”, to make clear that it’s mostly about multiple people who are collaborating. I’ve also seen “multi-cursor” or “multi-pointer” (which are missing the keyboard component).
On the note of the two different kinds of seats, a tangent: better multi physical seat support is something I am passionate about, specifically, I want Windows-style "Fast User Switching" on Linux with RDP. Today's Linux RDP support usually acts like screen mirroring; you get a mirror of a physical seat. With GNOME I believe there is also now support to do remote login: you can log in remotely, and even disconnect and reconnect to a remote login session. But what you can't do is continue a physical session remotely or a remote session physically, and the reason why is sorta to do with seats. Sort of.
See, when you start a graphical session on modern systemd-based Linux, a new logind session is created, and it is immutably tied to a seat. So you can't just remove the seat from a running logind session and then make it remote, or add the seat to a running logind session and make it physical. I actually tried patching systemd-logind to do this, but when I proposed it upstream, Lennart Poettering himself actually took the time to explain that this wasn't the intended way to do it: instead, you would actually just create a new logind session and drop the old one entirely, or literally whatever you want.
Which to me sounded perposterous, but then I realized that it really wasn't that ridiculous. As it was, when you boot a desktop using systemd-logind, the desktop itself doesn't typically run much of anything directly in the session context anyway - the compositor and almost all services actually run as user services, as in systemd-land having more than one active desktop session per user isn't really supported. Okay so we can do this!
And so I set out to prototype this with KDE, but there are still a few issues...
- First and foremost, $XDG_SESSION_ID. Believe it gets set by the systemd PAM module somehow? I don't fully recall. Either way, it gets used by various programs to determine the current logind session, but if our graphical session can "roam" between them we need a mechanism that can handle changing over time. I propose we stuff the current logind session ID into a file in $XDG_RUNTIME_DIR instead.
- Secondly, compositors need to be redesigned to handle this. Compositors today can hotplug displays and input devices - that's a good start. Unfortunately, at least kwin can't switch backends at runtime. You can really go one of two ways, and I tried a bit of both and either can work but both are a bit ugly: one is to make a single backend that can handle either case and the other is to allow switching the backend at runtime. Either way, when running remotely the card* DRM nodes are unavailable so you need to switch to renderD* nodes instead. You can also use both, and only use card* for scanout when available, but it seems a bit trickier.
- Finally, I genuinely have no idea how this could or should be handled in non-systemd environments. I have admittedly forgotten how that all works. It would be rather lame to simply require logind for this to work, but sadly it is the path of least resistance at least at first.
And maybe some more, but otherwise the concept does indeed broadly work. I mainly have only tested it in virtual machines with DRM native context, but I can see everything working switching in either direction at least once before things start to get screwed up. Even OpenGL clients and audio playback can move seamlessly! It's just like what Windows accomplished over 20 years ago!
Aside from convenience and being cool, it would also solve a couple problems. No need to worry about the limit of only one session per user: just move your session with you. No need to worry about your physical session being unlocked when remotely connected: when you log in this way your physical seat gets disconnected from the session entirely and the display manager takes it back over.
I actually would really like to at least make some progress on ground work that would be needed to make this happen, although I'm not really that sure that anybody cares. But either way this post's brief discussion about physical vs logical seats activated a few neurons so I couldn't help but dump this out.
I'm building a compositor now, and I've been thinking about multi-seat from the beginning, but I know there are lots of places that assume a single seat, and so this sort of thing won't work for me. It's actually a lot more work than you'd think to support this...
I keep thinking of trying to see how far I get re-doing Niri in River.
The one other constraint that I'd really like to see relaxed is about displays, about being able to merge and subdivide screens as I please. Having a huge ultra-wide display really would be so much better if I could divide it up somehow. This is kind of the headline feature of stilch, which is a pretty neat compositor, https://github.com/wegel/stilch .
Niri has some tickets on it too. Merging monitors into a big display, and splitting a display (moments before the multi-pointer ticket was filed in Niri), https://github.com/niri-wm/niri/discussions/1285 https://github.com/niri-wm/niri/discussions/3160
There's been all kinds of rough incredibly hard things going on with Niri just trying to figure out how to handle some of the more interesting copy-paste scenarios, and with really weird sequencing issues across basically ephemeral inputs, that have really plagued the project and made life incredibly bad for downstream apps. This is just my opinion, man, but I think the lack of ambition about these ridiculous over the top features eventually starts hampering what turns out to be relatively prosaic asks. Better models for multi-seat, multi-cursor help scenarios like remote-desktop, and video-chat remote control, that at first don't seem needed, but are fundamentals that pay dividends if you shoot for them.
To add to the long list of caveats: This also breaks assumptions GUI toolkit and applications make about only one of their windows being focused at a time: `gtk_application_get_active_window`, `QApplication::activeWindow`, etc.
I guess they could add a parameter cursorIndex defaulting to 0!
It's not as weird as you might think: laptops may have touchscreens (which can be multi-touch) on top of the "normal" cursor. IIRC, X.org can't really handle such cases cleanly.
That's not what this article is about. You still only have one mouse cursor in that case, but two things (the touchscreen and touchpad) drive it.
This is about actually having multiple, independent mouse cursors on the screen at once, with different input devices controlling different cursors.
Sometimes that's true, but sometimes (imo the superior approach) the touchscreen is independent from the cursor. I think touchscreen Chromebooks are like that, for example.
I've always thought this was a cool idea. It's obviously been on the minds of Wayland protocol developers since the beginning, otherwise it wouldn't be possible!
On the web, PointerEvents and touch events both support multiple different cursors. https://developer.mozilla.org/en-US/docs/Web/API/Touch_event... https://developer.mozilla.org/en-US/docs/Web/API/PointerEven...
The original title, "State of multi-player Wayland" is a much better summary of the article. This ain't just about the pointers! In fact, the author is very explicit about that:
> I think there could be a term that’s describing this concept even better – in the title of this post, I picked “multi-player”, to make clear that it’s mostly about multiple people who are collaborating. I’ve also seen “multi-cursor” or “multi-pointer” (which are missing the keyboard component).
Another possible use case might be mixed reality/VR displays, for simultaneous mouse cursor + eye tracking/control.
On the note of the two different kinds of seats, a tangent: better multi physical seat support is something I am passionate about, specifically, I want Windows-style "Fast User Switching" on Linux with RDP. Today's Linux RDP support usually acts like screen mirroring; you get a mirror of a physical seat. With GNOME I believe there is also now support to do remote login: you can log in remotely, and even disconnect and reconnect to a remote login session. But what you can't do is continue a physical session remotely or a remote session physically, and the reason why is sorta to do with seats. Sort of.
See, when you start a graphical session on modern systemd-based Linux, a new logind session is created, and it is immutably tied to a seat. So you can't just remove the seat from a running logind session and then make it remote, or add the seat to a running logind session and make it physical. I actually tried patching systemd-logind to do this, but when I proposed it upstream, Lennart Poettering himself actually took the time to explain that this wasn't the intended way to do it: instead, you would actually just create a new logind session and drop the old one entirely, or literally whatever you want.
Which to me sounded perposterous, but then I realized that it really wasn't that ridiculous. As it was, when you boot a desktop using systemd-logind, the desktop itself doesn't typically run much of anything directly in the session context anyway - the compositor and almost all services actually run as user services, as in systemd-land having more than one active desktop session per user isn't really supported. Okay so we can do this!
And so I set out to prototype this with KDE, but there are still a few issues...
- First and foremost, $XDG_SESSION_ID. Believe it gets set by the systemd PAM module somehow? I don't fully recall. Either way, it gets used by various programs to determine the current logind session, but if our graphical session can "roam" between them we need a mechanism that can handle changing over time. I propose we stuff the current logind session ID into a file in $XDG_RUNTIME_DIR instead.
- Secondly, compositors need to be redesigned to handle this. Compositors today can hotplug displays and input devices - that's a good start. Unfortunately, at least kwin can't switch backends at runtime. You can really go one of two ways, and I tried a bit of both and either can work but both are a bit ugly: one is to make a single backend that can handle either case and the other is to allow switching the backend at runtime. Either way, when running remotely the card* DRM nodes are unavailable so you need to switch to renderD* nodes instead. You can also use both, and only use card* for scanout when available, but it seems a bit trickier.
- Finally, I genuinely have no idea how this could or should be handled in non-systemd environments. I have admittedly forgotten how that all works. It would be rather lame to simply require logind for this to work, but sadly it is the path of least resistance at least at first.
And maybe some more, but otherwise the concept does indeed broadly work. I mainly have only tested it in virtual machines with DRM native context, but I can see everything working switching in either direction at least once before things start to get screwed up. Even OpenGL clients and audio playback can move seamlessly! It's just like what Windows accomplished over 20 years ago!
Aside from convenience and being cool, it would also solve a couple problems. No need to worry about the limit of only one session per user: just move your session with you. No need to worry about your physical session being unlocked when remotely connected: when you log in this way your physical seat gets disconnected from the session entirely and the display manager takes it back over.
I actually would really like to at least make some progress on ground work that would be needed to make this happen, although I'm not really that sure that anybody cares. But either way this post's brief discussion about physical vs logical seats activated a few neurons so I couldn't help but dump this out.