Skip to content

Latest commit

 

History

History
330 lines (285 loc) · 21 KB

virtualbox.md

File metadata and controls

330 lines (285 loc) · 21 KB

Oracle VirtualBox

Status: Windows Home x64 hosting is live-tested through an experimental common adapter. Linux resident conformance and repeated cold boot pass with a guest workaround. Windows resident desktop/UAC conformance, unattended unlock and two native build/deploy/test cycles pass. Repeated Windows lifecycle tests qualify a bounded compatibility profile with assisted shutdown and a protected stopped base; unassisted lifecycle reliability and production adoption remain open.

Provider and licensing

Oracle VirtualBox provides desktop VM hosting and the VBoxManage CLI. The base package declares GPL-3.0-only; bundled components retain their separately listed terms. The separately licensed Extension Pack is not part of this experiment. Exact distribution/source review is required before redistribution.

Architecture and claimed reach

Upstream-claimed: Windows x64, Linux and macOS hosts are supported with platform-specific restrictions. Windows ARM support is experimental and is not covered by this x64 experiment. See the installation guide.

VBoxManage supplies UUID-based inventory, headless start, lifecycle, virtual disk and snapshot management, EFI variables, emulated TPM and unattended installation. Guest Additions supplies authenticated guest process/file operations, including password-file input. These are candidates for bootstrap and recovery; ordinary Machine Control administration and semantic UI should continue through the existing target-resident implementation.

Upstream-claimed: VirtualBox can use Microsoft's active hypervisor on a Windows host, but performance can suffer. WSL2 coexistence and useful iteration latency require live qualification. Do not disable host security or WSL merely to make an experiment pass. Host installation includes a support driver; bridged/host-only networking, USB and Python bindings are separate installer features. NAT and the CLI do not require those optional drivers/bindings.

Initial evidence and setup pitfalls

Current (2026-10-04), live-tested, Windows Home x64: the signed core application/driver installed with restart suppressed and no restart requested. An isolated private VM library and exact UUID receipt bound a scratch Windows 11 guest to EFI, TPM 2.0, Microsoft Secure Boot signatures and an Oracle platform key. The headless installer booted from official evaluation media whose full digest matched Microsoft's published value. WSL remained responsive.

The supplied unattended template contains hardware-check bypass commands in both Windows PE and the installed OS. The private experiment removed all of them before rendering and checked the result. Using stock defaults would not establish acceptance with Windows 11 requirements intact. Setup credentials were stored atomically in the controller's private secret store before use and provided through password-file inputs; generated answer media remained private.

Before firmware initialization, screenshot capture reported a zero-sized display. A later provider-native capture observed Windows installation without host focus or input. This bootstrap capture is an explicit outer route, not proof of ordinary resident desktop control. Secure Boot configuration readback is not a substitute for checking its effective state inside Windows.

Installation reached 77%, then presented a blank console without Guest Additions becoming available. Read-only debugger inspection reported a Windows kernel and no bugcheck; the disk file timestamp stopped changing. That timestamp alone does not establish an absence of I/O. The native Windows hypervisor backend reported its slow execution mode. These observations do not identify the cause of the stall or establish a general compatibility failure.

An operator-authorized force-stop and cold boot resumed setup on the existing disk and reached the installed Windows kernel. Setup advanced, then returned to a blank console without Guest Additions. Actual storage counters showed ongoing writes before becoming unchanged for several minutes; a second clean ACPI shutdown did not stop the guest. It was force-stopped under the recovery authorization and retained for diagnosis, without a further boot or host reboot. At that initial checkpoint, guest security, password authentication and command effects remained unverified; the follow-up below supersedes those limitations. Do not infer stalled I/O from a VDI file timestamp: use provider storage counters.

Follow-up guest qualification

Current (2026-10-04), live-tested: an official Ubuntu 24.04 cloud image was provisioned with pinned SSH host keys, a canonical verified login password, GNOME Wayland and the existing Linux resident. Common-CLI tests independently observed semantic button effects, Unicode text and capture hashes. Secure Boot was enabled inside the guest. A warm reboot and two cold boots without seed media passed; the candidate was cleanly stopped and its claim released.

Four-vCPU cold boots initially stopped during initramfs driver loading. A one-vCPU boot recovered access. Guest kernel arguments nox2apic rcupdate.rcu_normal=1, with ordinary APIC enabled, then allowed repeated four-vCPU boots and resident tests. The workaround was suggested by upstream issue 861, whose reported host/backend differs; this does not establish an identical cause.

Windows recovery reached a responding login screen with three vCPUs, paravirtualization set to none, x2APIC disabled and ordinary APIC enabled. The canonical password authenticated and Guest Additions became available. Two vCPUs had stalled in EFI, resembling but not proving the cause in upstream issue 799. These are experiment settings, not a qualified universal host profile.

Current, live-tested: the Windows candidate now passes effective Secure Boot and TPM 2.0 checks without hardware bypasses, canonical stored-password authentication, pinned SSH, full resident desktop conformance and protected UAC cancel/approve tests. Two native x64 build/deploy/test cycles passed using the existing transactional appliance installer. UAC and secure-desktop policies remain enabled. The existing optional unlock component also passed an observed lock and signed-challenge, one-shot credential unlock. Its elevated approval dialog required the existing bounded guest-local consent fixture; ordinary semantic invocation was insufficient. No ordinary test manipulated the host desktop. Native OpenSSH capability installation took roughly 30 minutes.

Subsequent lifecycle tests did not pass. Guest-native shutdown remained running beyond five minutes and required explicit forced recovery. After media detachment and disabling automatic login, a three-vCPU cold boot reached pinned SSH and the protected resident with no interactive user, then lost SSH before credential submission. A four-vCPU recovery similarly reached SSH and accepted target-native curtain-reveal input, but the following observation timed out. Read-only doctor and authenticated Guest Additions diagnosis were unavailable. The cause is undetermined; these observations do not isolate the guest OS, additions, runtime, hypervisor or host/backend. Unattended unlock of an existing session passed; those attempts did not qualify cold login or reliable lifecycle operation. No host reboot, WSL shutdown or host security change was used to work around this failure. The final cleanup boot reached Automatic Repair; the candidate was retained powered off with its private credential handoff intact. It is diagnostic evidence, not an accepted reusable Windows base.

Current, live-tested follow-up: a bounded Windows lifecycle profile now passes three qualified media-free cold boots, guarded appliance login, idle/security/password checks and completed OS shutdowns. One idle interval lasted five minutes. The profile combines guest-facing Hyper-V paravirtualization, three vCPUs, ordinary APIC with x2APIC disabled, UTC time, disabled guest Intel NIC interrupt moderation and a bounded SSH handshake. Shutdown explicitly permits one claimed pause/resume after thirty seconds; this is an outer lifecycle compatibility path, not an unassisted shutdown or an ordinary desktop-test route. The underlying cause remains undetermined. Upstream issue 722 remains a resemblance, not proof of a shared cause.

A separate full-copy derivative retained firmware/TPM and passed first-attempt resident/UAC conformance, credential/security acceptance and shutdown. The original is promoted as a protected stopped base with verified disk, configuration and firmware/TPM hashes. An attempted ordinary start was refused. An earlier source conformance attempt failed at the Start-menu stage and its unchanged rerun passed; the tactical retains that friction and all timings. Host security, Microsoft hypervisor and WSL settings were not disabled, and no host reboot was used. This qualifies one private compatibility profile, not production adoption or broad Windows-host support.

The experimental adapter checks exact VM/disk identities, serializes claimed operations, uses pinned loopback SSH and existing credential verifiers, and exposes explicit recovery separately from ordinary resident calls. Initial creation and console bootstrap still require private factory scripts with manual journal coverage.

Unassisted shutdown investigation

Current (2026-10-05), research and local source review: the adapter requests shutdown.exe /s /t 0, then independently observes VirtualBox power-off. The opt-in scheduling assist runs once after the configured wait. It does not request Windows sleep, hibernation, hybrid shutdown or VirtualBox saved state. Existing live evidence includes both idle transport stalls and shutdown stalls; it does not establish that these have one cause. The research pass was followed by the bounded live shutdown trace below; host configuration stayed unchanged.

Upstream-claimed: issue 722 reports Windows 11 guest stalls on a Windows 11 host using Hyper-V, including reboot progress that resumes after mouse input. A June 24 commenter explicitly reports pause/unpause restoring progress and says removing the host Hyper-V dependency resolved their stalls. Other comments describe different clipboard/RDP and non-Windows-host failures, so the thread is not evidence of one confirmed upstream defect. The original report covers 7.2.6/7.2.10 and remains open at this review. Issue 799 separately reports a two-vCPU EFI boot stall under NEM/WHPX that disappears with three vCPUs; this is a boot resemblance, not a shutdown diagnosis.

Oracle's 7.2 troubleshooting guide documents host Hyper-V coexistence problems and host power-management effects on virtualization. Its suggested host-feature removal requires reboot and can affect security features. That is not an acceptable routine adjustment to this development-host profile, which retains WSL and host security. The guest-facing Hyper-V paravirtualization interface is a separate setting from the active host execution backend; changing it does not remove the latter.

Inference: idle-vCPU wakeup, timer or interrupt delivery in the host/backend path is the leading hypothesis because a hypervisor scheduling transition restores shutdown progress and ordinary idle use has also stalled. Pausing affects several subsystems, so this is not a root-cause finding. Guest service or driver teardown, including Guest Additions, remains a competing explanation.

Current (2026-10-05), live-tested: an unprotected development guest with the scheduling assist disabled exceeded the fifteen-minute observation budget. One explicit pause/resume was followed by power-off in approximately four seconds. A private, shutdown-persistent ETW recording captured scheduler, disk, interrupt, power, service and logon events. After approximately eight seconds of activity, it recorded only seven events over the next roughly 895 seconds, then substantial activity returned around recovery. The decoded trace reports zero lost events and buffers. Sparse interrupt/DPC routines map to ndis.sys; this establishes occasional network-stack activity, not a network-driver cause. Service shutdown continued after recovery.

This strengthens the idle/wakeup hypothesis relative to sustained shutdown I/O, but does not prove where guest progress stopped. Built-in WPR profiles refused with 0x80070032; a minimal custom profile succeeded without sampled CPU stacks. The controller's event decoder also lacks some guest event schemas. Without stacks or an independent host/backend trace, a specific blocked service, driver or virtualization defect cannot be identified. Execution and cleanup details belong to Tactical 094.

Upstream-claimed: Microsoft documents full shutdown as the default for shutdown.exe /s /t 0; /hybrid is the explicit Fast Startup option. Consequently, disabling guest Fast Startup is a low-priority explanation for this exact command, though menu-based shutdown would be a different test. This research found no specific primary-source link between System Restore and pause/resume-sensitive shutdown; do not disable it without trace evidence implicating restore or VSS activity.

Current, live-tested: although the guest's Fast Startup registry preference was enabled, powercfg /a reported hibernation unsupported by firmware and Fast Startup unavailable. Changing that preference cannot explain this observed full-shutdown stall.

Proposal: continue with an unprotected development candidate for one-variable tests. The first instrumented reproduction above is complete; guest-only idle behavior is the next focused comparison, not a proven fix:

  1. Acquire its exact-target claim and collect a private baseline: guest build, Guest Additions/driver versions, power configuration, shutdown event logs, provider/backend log and servicing state. Keep the protected base stopped.
  2. Measure an unassisted native shutdown with the existing fifteen-minute budget. Keep intrusive debugger probes out of the timing-only baseline. Run any instrumented reproduction separately: Microsoft provides WPR shutdown tracing. Distinguish service/driver waits, continuing I/O and stopped guest execution; the loss of SSH alone does not distinguish them. Keep traces private.
  3. If execution stalls around idle transitions, test guest processor idle behavior, then vCPU count, separately. Microsoft describes IDLEDISABLE, but its effectiveness here is untested. Read back support and save original values first; temporarily preventing guest CPU idle is a diagnostic with increased CPU/power cost, not a permanent fix or a host power-policy change.
  4. If tracing identifies a guest service or driver, test only that component next. Verify installed Guest Additions compatibility before changing it. Leave host hypervisor, WSL, security and firmware policy intact.
  5. Restore each unsuccessful setting. Require three distinct cold boots with untouched idle intervals and completed unassisted shutdowns before removing assistance from an accepted profile. Instrumented and assisted successes do not qualify the timing-only, unassisted configuration.

Open: root cause, a supported fix that retains the host's hypervisor/WSL configuration, and whether guest-only power settings improve shutdown without merely hiding a scheduling defect. Online resemblance does not close these gaps.

Later shutdown recovery — 2026-10-14

Current, source-reviewed and live-tested: a retained-use shutdown exceeded the complete fifteen-minute budget with the existing delayed scheduling assist enabled. Audit records show a roughly 904-second failed call; the provider log confirms that the configured pause/resume actually occurred. A later explicitly authorized second cycle followed by ACPI delivery reached Off. The log records about 76 minutes of timer catch-up backlog being abandoned at that resume and power-off about 3.3 seconds later. Windows had already logged the original native shutdown and kernel shutdown progression. The later ACPI delivery does not establish that another request was necessary; the earlier instrumented pause-only recovery above already proves that a pending shutdown can complete without one. Neither observation proves pre-request ordering is superior.

Inference from retained timing: the first resume was approximately 34 seconds after the request; kernel shutdown event 109 was approximately 185 seconds after it. The early cycle therefore preceded that later shutdown stage. This is consistent with a single early assist missing a later stall, rather than the sequence itself being wrong. Guest events and host/provider time can diverge during scheduling stalls, so this is not an exact stall-onset finding. The roughly 76-minute backlog reset is additional scheduling evidence, not a root-cause diagnosis or a reason to configure unlimited automatic retries.

Source review confirms that normal shutdown uses shutdown.exe /s /t 0 through SSH, then optionally pauses/resumes once after the configured delay. Its timer starts only after the guest command returns. An SSH/command failure can bypass both that wait and the assist. ACPI recovery is a virtual power-button event with a delivery-only result; it is not a direct power cut or proof of Off. Scheduling pause and saved-state suspend have different semantics in Oracle's command reference.

The live follow-up and its servicing limitations belong to Tactical 122.

Current, live-tested: its update-triggered restart remained unavailable after the initial ten-minute boot observation. One scheduling cycle reset an approximately eight-minute backlog without restoring administration or Off; subsequent ACPI delivery also left Running through a 120-second observation. Explicitly authorized force-stop completed cleanup. This servicing condition prevented a stable paired comparison and was not counted as clean lifecycle evidence. Timer-backlog reset alone is not a recovery-effect oracle.

Open: a stable paired comparison of pre/post-request scheduling and native/ACPI methods. The public adapter still lacks a standalone typed pause/resume recovery operation. Do not turn one successful recovery into an automatic request-method change, unlimited scheduling retries, or forced shutdown. The current operator guide requires terminal-state evidence and explicit authority for a power cut.

Fit and remaining gates

Current, staged x64 desktop campaign: extended interactive load also produced guest scheduling stalls in the retained Windows-hosted compatibility profile. Read-only SSH observations timed out, and the backend recorded long timer catch-up intervals. One separately authorized, bounded pause/resume recovery restored execution without host desktop input. Subsequent desktop feature results qualify their native route; they do not establish unattended VM stability or an unassisted lifecycle fix. Exact logs and target identity remain private; Tactical 101 records the feature campaign and its limits.

Proposal: qualify this Home-capable provider before implementing common production workspaces. Identity/role, claims and operation serialization have initial implementation and tests; complete factory/audit coverage and cleanup receipts still need adoption work. CLI support for snapshots/clones alone does not prove safe disposable workspaces or preservation of TPM/NVRAM state across derivation.

Open: complete common-factory integration, unassisted lifecycle reliability, generalized derivation/recovery and isolated workspace acceptance. Remote provider hosting and Windows ARM hosts remain untested. Evidence and cleanup results belong to Tactical 094; private journals retain exact media, resource and credential locators.