virtio_p9: modernize to VirtioDevice trait

[jstarks]

Replace LegacyVirtioDevice impl with direct VirtioDevice impl:

• Add driver, worker, exit_event fields for direct queue management
• Implement enable()/disable() following virtiofs pattern
• Constructor now takes VmTaskDriverSource parameter
• Remove LegacyWrapper from resolver, construct device directly
• Use Option instead of Vec since there is only one queue
• Add deps: vmcore, task_control, event-listener, pal_async

The VirtioPlan9Worker and process_work() logic is unchanged. This is one step toward removing LegacyVirtioDevice entirely.

[Copilot]

Pull request overview

Modernizes the virtio_p9 (virtio-9p) device to implement the VirtioDevice trait directly (matching the newer direct-queue pattern used by devices like virtiofs), removing the legacy wrapper usage in the resolver.

Changes:

• Replace LegacyVirtioDevice + LegacyWrapper usage with a direct VirtioDevice implementation in virtio_p9.
• Add direct queue worker lifecycle management (enable()/disable()) using TaskControl, VmTaskDriver, and event-listener.
• Update the resolver and crate dependencies to construct the device directly and support the new worker model.

Reviewed changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated 1 comment.

File	Description
vm/devices/virtio/virtio_p9/src/resolver.rs	Stop wrapping with LegacyWrapper; construct VirtioPlan9Device directly.
vm/devices/virtio/virtio_p9/src/lib.rs	Implement VirtioDevice with explicit queue worker management and updated constructor signature.
vm/devices/virtio/virtio_p9/Cargo.toml	Add new dependencies required for direct queue/task management.
Cargo.lock	Lockfile updates reflecting new dependencies for virtio_p9.

Comments suppressed due to low confidence (1)

vm/devices/virtio/virtio_p9/src/lib.rs:52

• size_of::<u16>() is used in the constructor but size_of isn’t imported or qualified, which will fail to compile. Import std::mem::size_of (or use std::mem::size_of::<u16>()).

        // The tag uses the same format as 9p protocol strings (2 byte length followed by string).
        let length = tag.len() + size_of::<u16>();

        // Round the length up to a multiple of 4 to make the read function simpler.
        let length = (length + 3) & !3;
        let mut tag_buffer = vec![0u8; length];

[Brian-Perkins]

may want to also implement drop

impl Drop for VirtioPlan9Device {
    fn drop(&mut self) {
        self.disable();
    }
}

[jstarks]

Ah, the drop thing isn't really correct, because disable really needs to be async to work for many devices (such as virtio-blk)--we need to wait for the device to finish accessing memory or whatever. I want to add another state transition to devices to disable them before they're torn down... for now I am just eliminating the drops.

[smalis-msft]

Why use an iter when queues is a Vec? Just call .get(0). And also should probably validate that length is exactly 1, what if we get many queues?

[jstarks]

We don't really need to be that defensive--we specify the queue count in traits and can rely on the infra to pass the right number back to us.

.get(0) doesn't work because we need ownership. I guess .remove(0) would work. This seems fine though.

[smalis-msft]

If we know the queue count ahead of time, could it be made a const-sized generic array? Not in this PR ofc, just food for thought.

[jstarks]

For some devices it will be dynamic.