virtio: reduce guest memory accesses on queue hot path

[jstarks]

The virtio queue implementation was doing expensive guest memory writes (kick arming/suppression) on every call to is_available(), even when the device was just draining a batch and had no intention of sleeping. This added unnecessary overhead on the hot path, particularly for high-throughput devices like virtio-net.

Restructure notification handling so that is_available() is a lightweight read-only check, kick arming is deferred to the actual sleep boundary (poll_kick), and suppression happens lazily when work is found while armed. The split queue caches avail_index to drain batches with a single read, and the packed queue reads only the 2-byte flags field instead of the full 16-byte descriptor for availability checks.

Burette TCP tx throughput improves by 31% on nested KVM, UDP tx by 48%. TCP rx is unchanged.

[Copilot]

Pull request overview

This PR optimizes virtio queue hot-path performance by restructuring notification (kick) arming/suppression so that availability checks avoid unnecessary guest-memory writes, reducing overhead for high-throughput virtio devices.

Changes:

• Refactors split-queue availability checks to cache avail_index and defer kick arming to the sleep boundary.
• Refactors packed-queue availability checks to read only the descriptor flags (instead of the full descriptor) and adds explicit kick arm/suppress APIs.
• Introduces core-level tracking of whether kicks are currently armed, and updates poll_kick() to arm only when actually sleeping.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
vm/devices/virtio/virtio_spec/src/lib.rs	Adds/clarifies documentation for packed-queue event suppression flags.
vm/devices/virtio/virtio/src/queue/split.rs	Caches avail_index and splits “availability” from “arm/suppress kicks”.
vm/devices/virtio/virtio/src/queue/packed.rs	Makes is_available() a lightweight flags-only read; adds arm/suppress kick methods.
vm/devices/virtio/virtio/src/queue.rs	Tracks armed state in QueueCoreGetWork and suppresses kicks lazily when work is found.
vm/devices/virtio/virtio/src/common.rs	Changes poll_kick() to arm for kick only at the sleep boundary.

[chris-oo]

Is there value in returning something other than an option to make this behavior more explicit? This sounds like a bug waiting to happen, IE caller forgot to call arm_kick.

[chris-oo]

Or is the intention that most callers should not call this directly, and use the poll wrappers in common.rs?