virtio-gpu device hardcodes the number of capsets instead of querying Rutabaga

[valpackett]

(From #560 (review))

As we're going for the ability to eventually configure which capsets Rutabaga would be configured with, I was shocked to discover that…

libkrun/src/devices/src/virtio/gpu/device.rs

Lines 162 to 166 in 757b080

	let config = virtio_gpu_config {
	events_read: 0,
	events_clear: 0,
	num_scanouts: self.displays.len() as u32,
	num_capsets: 5,

num_capsets in the MMIO-readable config is hardcoded to 5 and has been ever since rutabaga was imported into libkrun!

And this is what Linux uses to determine how many capsets to query. If we have fewer capsets, it all blows up.

[    0.119895] [drm] cap set 0: id 6, max-version 0, max-size 592
[    0.119946] [drm] cap set 1: id 5, max-version 0, max-size 16
[    0.119992] [drm:virtio_gpu_dequeue_ctrl_func] *ERROR* response 0x1200 (command 0x108)
[    5.129781] [drm:virtio_gpu_get_capsets] *ERROR* timed out waiting for cap set 2
[    5.130347] [drm] Initialized virtio_gpu 0.1.0 for a005000.virtio_mmio on minor 0

[2026-02-27T03:41:29Z WARN  rutabaga_gfx::rutabaga_core] CapSets [RutabagaCapsetInfo { capset_id: 6, component: VirglRenderer, name: "drm" }, RutabagaCapsetInfo { capset_id: 5, component: CrossDomain, name: "cross-domain" }]
[2026-02-27T03:41:29Z WARN  rutabaga_gfx::rutabaga_core] get_capset_info 0
[2026-02-27T03:41:29Z WARN  rutabaga_gfx::virgl_renderer] virgl_renderer_get_cap_set..
[2026-02-27T03:41:29Z INFO  rutabaga_gfx::virgl_renderer] drm: drm_renderer_capset:135: c=(nil)
[2026-02-27T03:41:29Z WARN  rutabaga_gfx::virgl_renderer] virgl_renderer_get_cap_set: 0 592
[2026-02-27T03:41:29Z WARN  rutabaga_gfx::rutabaga_core] get_capset_info 0: 0 592
[2026-02-27T03:41:29Z WARN  rutabaga_gfx::rutabaga_core] get_capset_info 1
[2026-02-27T03:41:29Z WARN  rutabaga_gfx::cross_domain] CroDo get_capset_info
[2026-02-27T03:41:29Z WARN  rutabaga_gfx::rutabaga_core] get_capset_info 1: 0 16
[2026-02-27T03:41:29Z WARN  rutabaga_gfx::rutabaga_core] get_capset_info 2

…with all the control via flags i.e. muvm disabling legacy virgl etc. we never had fewer than 5 virtio-gpu caps?!?! 😮

This seems somewhat tricky to fix since Rutabaga is located in the worker thread, and this config area is served via the synchronous read_config in the.. main thread? It shooould be fine to use a sync oneshot channel right?

[mtjhrc]

Yeah, this is a known issue but I forgot about it, I never got around to fixing this here.

I also ran into this problem when working on rust-vmm/vhost-user-gpu. There we just manually count the number of capset based whatever we pass into Rutabaga:

https://github.com/rust-vmm/vhost-device/blob/93cdc9458d3e196443da262e844f8ec7e2da8c44/vhost-device-gpu/src/device.rs#L173

https://github.com/rust-vmm/vhost-device/blob/93cdc9458d3e196443da262e844f8ec7e2da8c44/vhost-device-gpu/src/lib.rs#L123-L128

So does crosvm:
https://github.com/google/crosvm/blob/2cd1f5096cf8450410a98b9711e65af8a75d559e/devices/src/virtio/gpu/mod.rs#L1745-L1753

I wonder if we could somehow improve this in Rutabaga?

It shooould be fine to use a sync oneshot channel right

Should be fine, but maybe put a timeout on the recv so we can't hang?