Fixes an issue where a dual GPU system would keep allocating dGPU PBOs in
`cpu_copy()` every frame, but `cleanup()` was not being called, since the
surface thread for the builtin output was rendered on the
primary/integrated GPU, targeting the same GPU.
Even if a different monitor was compositing on the dGPU, That wouldn't
help since that thread has it's own `GpuManager` with it's own renderer
and cache.
Running cleanup at the end (or start) of each frame seems like a good
idea. Not sure if it would be best to avoid additional calls, or if
that's desirable/fine.
Instead of choosing a primary GPU the first time `device_added` is
called (and then not updating it on the next call, even if that should
be the primary GPU), set the primary GPU only after all devices have
been initially added, and on future changes.
Alternately, the `was_empty` test can just be removed, but it's probably
best not to select the primary GPU multiple times each time the
compositor starts on a multi-GPU system.
Fixes https://github.com/pop-os/cosmic-comp/issues/1437.
Add more sophisticated code to handle the primary node disappearing.
Also overhaul the selection logic to respect our allow/deny-list and
prefer devices with built-in connectors before using the boot gpu.
This will also allow triggering a primary node switch at runtime
for debugging purposes in the future.
Previously, `Frame` was stored in KMS frame udata, but in some cases the
udata was dropped without a capture happening, and `Frame` did not
implement `Drop`, so `fail` was never sent.
Instead, rename `DropableFrame` to `Frame` and `Frame` to `FrameRef`, so
we can have a single instance of `Frame`, that will send `fail` on drop.
This guarantees either `.success` or `.fail` are send, as long as its
not leaked.
This seems to fix https://github.com/pop-os/cosmic-comp/issues/1305.
xdg-desktop-portal-cosmic prints an error, buy retries (as it should for
an `Unknown` error; though maybe there should be a retry limit) and the
session continues working.
(Not sure if it should be sending `failed`, or queing it with the next
frame so it can send `success` to the client, but this works and is
desirable as a failsafe anyway.)
`Session` and `CursorSession` are similiarly updated.
`.fail()`, `.success()`, and `.stop()` now consume
`Frame`/`Session`/`CursorSession`. So to stop a session, it is now
necessary to call `.remove_session()`, but then simply dropping with
send `.stop()`.
Factoring out some `Request::Capture` handling into a `capture_frame`
function seems to clean up error handling and such a bit.
If `redraw()` returned early, before updating `self.state`, but after
calling `queue_frame`, `on_vblank` would later be called, and reach
`unreachable` since state isn't set to `WaitinfForVBlank`.
In particular, this was happening when the dmabuf from the image copy
frame failed to `bind`.
To avoid this, make sure to update `self.state` immediately after
calling `queue_frame`, before any early return from an error.
The new type bounds seem to be addressable using `AsGlowRenderer`.
Though things like this will need a different solution when we want to
support Pixman rendering.
This was mapping `UnderlyingStorage::Memory` to `None`. I don't see any
reason for this. Though this also shouldn't change behavior since
`TextureRenderElement` doesn't provide `underlying_storage` currently.
Make numlock state on boot configurable
This will expose 3 settings for numlock behavior:
1. Numlock is off on boot (this is the current default behavior)
2. Numlock is on on boot
3. Numlock will restore the state from the last boot
Fixes#369
* Address comments:
Get keyboard after create_seat called rather than returning from
create_seat.
Use constants rather than magic numbers for keypress.
Only save updated modifier state after keypresses are handled/skipped.
* Remove unused import, fold other into existing use block.
To determine if `mirroring_state` is needed, we want to compare the
untransformed dimensions of the source output to the target outputs
mode. The fact the mode comparision previously compared refresh rate
(since it compared the whole `Mode`) seems unintended.
We also re-create the `MirroringState` when the source output
dimensions changes.
Apply inverse of output transform to mode to get render size, and apply
no transform during rendering. The transform of the output being
mirrored from shouldn't affect the final render.
Fixes issues when source output for mirroring has a transform, and also
fixes issues in https://github.com/pop-os/cosmic-comp/pull/1058
when this code is used for postprocessing, where this resulted in the
same transform being applied twice.
