Implement virtio-input for keyboard and mouse

[Mes0903]

This pull request is derived from #34 and #121, implements VirtIO-Input keyboard and mouse devices for semu, enabling host keyboard and mouse input to reach the guest through an SDL-backed window path, in conformance with the VirtIO specification v1.3.

VirtIO-Input

• Implements separate VirtIO-Input keyboard and mouse devices
• Uses the VirtIO MMIO transport layer
• Supports:
  • keyboard key press and release events
  • mouse button events
  • absolute pointer motion
  • scroll wheel events
• Implements the VirtIO-Input configuration selectors required by the Linux driver:
  • VIRTIO_INPUT_CFG_ID_NAME
  • VIRTIO_INPUT_CFG_ID_SERIAL
  • VIRTIO_INPUT_CFG_ID_DEVIDS
  • VIRTIO_INPUT_CFG_PROP_BITS
  • VIRTIO_INPUT_CFG_EV_BITS
  • VIRTIO_INPUT_CFG_ABS_INFO
• Handles both EVENTQ and STATUSQ
  • EVENTQ is used for keyboard and mouse event delivery
  • STATUSQ is drained and acknowledged so guest LED updates can make forward progress without stalling the queue

Prerequisites

Linux

sudo apt install libsdl2-dev

macOS

brew install sdl2

Running semu

make check

If SDL is available, semu should open a window for input testing.

Basic guest verification

After the guest has booted, verify that the VirtIO-Input devices are present:

ls /dev/input/event*
cat /proc/bus/input/devices

The guest should expose /dev/input/event* nodes, and /proc/bus/input/devices should report VirtIO input devices.

Manual end-to-end verification

1. Launch semu:

   make ENABLE_INPUT_DEBUG=1 check

   If SDL is available, semu should open a window for input testing. With ENABLE_INPUT_DEBUG=1, semu also prints the host input events translated by the SDL backend, including:

   • keyboard press/release events
   • mouse button press/release events
   • mouse motion coordinates
   • wheel deltas
   • queue-full drop messages

2. In the guest, verify that the VirtIO-Input devices are present:

   ls /dev/input/event*
   cat /proc/bus/input/devices

3. Optionally inspect the guest event stream directly:

   hexdump /dev/input/event1

4. Focus the SDL window and:

   • type keys
   • click mouse buttons
   • move the mouse
   • scroll the wheel

You should see the translated host-side events in semu's debug output, and new records should appear in the guest input event stream.

Automated testing

• Adds .ci/test-vinput.sh as a smoke test
• Verifies:
  • /dev/input/event* exists
  • /proc/bus/input/devices reports VirtIO input devices
• Runs in GitHub Actions on both Linux and macOS
• Uses SDL_VIDEODRIVER=offscreen in CI so the guest-visible device
  enumeration path can be validated without requiring a visible desktop
  window

The current automated test covers device enumeration in the guest. It does not yet validate the actual host->guest input injection path, which is still verified manually.

---

Summary by cubic

Implements virtio-input keyboard and mouse over VirtIO‑MMIO with an SDL‑backed window so host input reaches guest evdev. The window runs on the main thread; the emulator runs in a worker and uses a coalesced wake pipe plus bounded per‑device SPSC queues for fast IRQs and clean shutdown (including gdbstub and window close).

• New Features

  • Separate keyboard and mouse; key press/release (EV_REP advertised, host autorepeat dropped), mouse buttons, relative motion (REL_X/REL_Y), and wheel (REL_WHEEL/REL_HWHEEL). Click window to grab; Ctrl+Alt+G releases.
  • Linux-required config selectors (ID_NAME, ID_SERIAL, ID_DEVIDS, PROP_BITS, EV_BITS, ABS_INFO); unknown selectors return size=0. Drains STATUSQ for LED updates (no host LED control).
  • Capabilities match the backend: keyboard EV_KEY bitmap from the SDL key map, INPUT_PROP_POINTER, and relative axes only.
  • SDL/main thread translates host input into bounded per-device SPSC queues; the emulator thread drains events, validates virtqueue bounds/descriptors, coalesces wakeups via a self‑pipe, and clears per‑device queues on reset. Window‑close requests stop the emulator cleanly.
  • Optional input debug logging via ENABLE_INPUT_DEBUG=1 prints injected events and queue‑drop messages.

• Dependencies

  • Build gated on SDL2; if sdl2-config is missing, virtio-input is disabled. CI installs SDL2, exports SDL_VIDEODRIVER=offscreen, adds .ci/test-vinput.sh on Linux/macOS, and extends macOS timeouts.
  • Guest Linux enables CONFIG_INPUT_EVDEV and CONFIG_VIRTIO_INPUT; the device tree adds two VirtIO‑MMIO input nodes with separate IRQs. CI forces fresh Image/rootfs.cpio downloads to avoid stale artifacts.

^{Written for commit 5c24de0. Summary will update on new commits.}

[Mes0903]

The virtio-input feature added in this branch slightly increases boot time: SDL initialization is attempted on every run, and two additional virtio device objects are initialized even in the headless environment used by the netdev test. As a result, the netdev CI test now times out.

The netdev test script runs two network-device tests back to back: on macOS, it exercises both user-mode networking and vmnet, each with a 900-second per-statement expect timeout. However, based on observation, this test sometimes finishes in about 10 minutes, while at other times it exceeds the previous 20-minute limit. In other words, the current setting is still insufficient in slower cases to cover both tests when the guest is slow to configure the network interface or complete the ping exchange.

Although I previously made optimizations for the headless-mode case, they still do not seem to be sufficient. Therefore, for now, I have added a temporary commit to extend the timeout to 30 minutes, so that we can stay focused on the vinput implementation first.

[Mes0903]

The macOS netdev CI test timed out again. From the log, it appears to be stuck at Starting klogd: OK. This makes me suspect that there may be a condition in the macOS environment that can cause the emulator to hang. For now, I have extended the timeout to two hours so I can determine whether this is really just a timeout issue, or whether there is an actual bug that only occurs on macOS.

[Mes0903]

While testing #123 on my own remote, I found that the sound test also experienced timeouts. At the same time, this PR has also shown timeout behavior. These timeouts have occurred in both the netdev and sound tests, but they appear to be random and are difficult to reproduce.

After extending the timeout to two hours, I found that the jobs were not consistently hanging, but rather sometimes just running much more slowly. The difference usually starts to appear only after the init process has been launched, and the gap between the fast and slow cases can be quite large.

For example, the virtio-input test just took 31 minutes, whereas it normally only takes about 5 minutes, such as in this run. The codebase did not change at all between these two runs. Taking the EXT4-fs (vda): mounted filesystem message as an observation point, the former recorded a timestamp of [ 108.175012], while the latter recorded [ 22.946067], which is far smaller.

Based on this, I believe the problem is most likely with the GitHub runner. The fact that #123 also experienced timeouts is a key reason for this conclusion, because #123 does not change the overall architecture at all, it only fixes a value inside virtio-fs, which should not have any impact on execution time.

[jserv]

Drop "DO NOT MERGE" commits. Then, I will merge this PR.

[Mes0903]

The current CI failure is expected because the Image on the main blob branch is still the old one, and CI does not rebuild the kernel Image.

I've manually rebuilt the kernel locally with the current config and tested it by hand. The virtio-input functionality works correctly with the rebuilt Image.

[jserv]

Thank @Mes0903 for contributing!