HV: verify VirtioFS DAX end-to-end (mmap → page fault → hv_vm_map) #253

Problem

docs/virtio-improvements-plan.md Phase 4 explicitly flags: "End-to-end DAX test... unverified at the integration level. No test covers the full path: guest mmap() → FUSE_SETUPMAPPING → hv_vm_map → guest page fault → host file read."

DAX is the primary performance differentiator for VirtioFS on HV. Without end-to-end verification, container image layer reads and build workloads may silently fall back to the copy-through path, erasing the win over VZ.

Related landings (why we think it should work now)

• 106a6ca — fixed per-share DAX mapper base IPA collision. All VirtioFS shares previously mapped into share 0's window; now each share gets its own HvDaxMapper with correct base_ipa.
• ABX-342 (Done) — DAX window munmap'd on shutdown / error.
• ABX-328 (Done) — DAX mapper bounds / overlap check on window_offset.

Remaining unverified:

• ABX-329 (Backlog) — DAX remove_mapping ignores length parameter.
• End-to-end path has never been exercised in a test.

Scope

1. Write a guest-side program that mmap()s a file from a VirtioFS share, reads a byte, checks it matches the host-side content.
2. Confirm the host observes a FUSE_SETUPMAPPING request for the read region.
3. Confirm hv_vm_map() is called with the right GPA → host VA → length.
4. Confirm the guest page fault is serviced without a FUSE_READ round trip (observe via FUSE request count).
5. Test remove_mapping on unmap; confirm subsequent access re-establishes mapping.

Acceptance

• Test runs in CI with the HV E2E infrastructure (sibling sub-issue).
• Measured round-trip reduction: DAX read path is ≥10x faster than the FUSE_READ path for the same content.
• ABX-329 fixed or explicitly deferred with a test demonstrating the failure mode.