Repository navigation
Description
Goal
Make MVM's verifiable secure-execution contract explicit across lifecycle transitions: an authenticated launch is not a permanent statement about a workload after replacement, resume, policy changes, or revocation. Keep the current trusted-host boundary (ADR-001); independent hardware-backed remote verification is separate work.
Owner decisions: Policy changes must affect already-running workloads. Choose atomic apply: editing policy files stages a candidate; applying it with MVM withdraws affected authority before the new generation becomes effective. Prefer a prepared rollover: old workload serves while the replacement is built, then one host-owned fenced cutover and fresh admission. Stop or refuse when a safe handoff is unavailable. Local mvmctl automatically manages its coordinator; no separate daemon command, mvmd, or mandatory enrollment for local use.
Delivery sequence: Restricted first slice for local project mvm.toml on Firecracker is acceptable. HVF follows next. All backends are eventual requirements, including libkrun and the dev/portability tiers under their actual capability contracts; an unsupported backend must refuse policy apply until its own live witness, not silently downgrade or count as complete.
Acceptance checklist
- Review proposed ADR-111 defining evidence freshness and the host-owned atomic policy transition, including policy sources, old/new generations, affected-instance binding, in-flight effects, failed handoff, deadlines, cancellation and crash recovery.
- Design the authoring-to-committed-policy boundary and
mvmctl policy apply: direct edits remain drafts, every affected launch/restore consumes the committed generation, and CLI/SDK use one host-interface authority. No silent fallback to mutable source files. - Connect the local coordinator to real launch admission, source/digest/generation binding, and lifecycle ownership; serialize apply with launches, resumes and other applies; refuse when affected running instances cannot be identified or controlled.
- Firecracker first: prepare an isolated candidate while the old workload remains authorized, fence old effects, admit and ready the candidate under the new signed plan, transfer one host-owned owner/route, retire old, and recover an interrupted handoff. Test invalid input, stale policy/epoch, active and in-flight effects, concurrent launch/apply, failed candidate, failed stop, audit and crash recovery at the production boundary; add a live Firecracker witness before claiming that backend.
- HVF next: implement and live-witness the same policy generation, fenced handoff and recovery without reusing Firecracker's backend-specific mechanism as assumed evidence.
- Complete backend matrix: implement and witness libkrun, QEMU dev/test, and explicit Wasm/browser portability targets where they expose workload lifecycle, preserving their tier limits. Refuse until supported; do not promote an all-backend claim from a mock, fixture, or only one platform. Keep unfinished items open.
- Reconcile the lifecycle inventory with actual production paths for local launch, durable agent-session resume, snapshot resume, replacement and revocation; do not duplicate gates already enforced.
- Update applicable user documentation and ADR-001 claims only to the extent observed tests support; preserve the host-trusted posture and mark unimplemented measurement/freshness promises as proposed.
- Record scoped and required validation plus any unrun live-backend/hardware witness in the PR. Leave incomplete items unchecked.
Known starting evidence
- Durable agent-session resume creates and admits a fresh plan (
crates/mvm-hostd/src/session_resume.rs); checkpoint revert also re-admits (crates/mvm-cli/src/commands/vm/checkpoint/revert.rs). - Snapshot resume has distinct replay, reseed and admission checks in
crates/mvm-client/src/local.rs; itsfc.admittedmarker is not a new signed-plan admission. - The per-VM endpoint builds a static
EgressGatefrom admitted policy; destination checks do not observe a later file edit (crates/mvm-hostd/src/supervisor/network_endpoint.rs). - CLI authoring/profile policy and tenant policy bundles are resolved at launch/admission, not under one resident committed generation. Host-interface
policy.jsonis a distinct host-level policy. Nomvmctl policy applyoperation exists today. mvmctl trust attest exportcurrently emits a zero boot-measurement sentinel andverifychecks signature/trust, not a verifier challenge or current workload state (crates/mvm-client/src/attestation.rs). Do not ship a generic re-attestation claim from this report.- Design and implement the optional distributed host interface #5097 owns the local/remote host-interface contract; this work implements its local policy/lifecycle owner, not a second authority. Carry a workload execution epoch in the signed plan, and exercise quarantine against a running VM #5158 separately tracks a real-VM quarantine witness; coordinate rather than duplicate.
Non-goals
No malicious-host confidentiality claim, no universal time-to-live for all evidence, no second local authority, no new artifact wire format without versioned migration, and no wholesale adoption of another project's runtime or codec.
Activity
auser commentedon Nov 12, 2026
Source trace for the proposed ADR (local branch
docs/evidence-freshness-contract,specs/adrs/111-host-trusted-evidence-freshness.md):mvm-client::LocalBackend::resume_machinecallsplain_resume/warm_resume; sealed Firecrackerresume_sealedverifies snapshot replay and waits for post-restore reseed, but does not re-resolve signed-plan/policy authority.fc.admittedrecords reseed completion, not renewed plan admission (crates/mvm-client/src/local.rs:227-350,crates/mvm-runtime/src/vm/admission.rs:1-24).Arc<EgressGate>from its admitted policy at assembly; each OpenTcp checks that gate. Rechecking a destination against an unchanged gate does not detect a changed policy file (crates/mvm-hostd/src/supervisor/network_endpoint.rs:364-388,crates/mvm-vmm/src/vsock_egress_bridge/egress_gate.rs:261-329).crates/mvm-hostd/src/session_resume.rs:319-408).Recommended division: pre-vCPU resume preflight for any authority that must still be valid before guest effects, plus existing per-effect gates for destinations/capabilities/expiry. A running instance's revocation requires a separate interruption/re-admission protocol; a resume-only guard cannot solve it. Do not use the signed plan's short launch validity window as a workload lease. Before implementation, identify the actual pause-persistent endpoint and policy/revocation source whose contract makes a stale state invalid, then choose a narrowly testable production transition. No security claim promoted or code path changed by this investigation.
auser commentedon Nov 12, 2026
Owner clarification supersedes the earlier recommendation in my first issue comment: policy changes must affect already-running workloads, including stopping them if changed permissions cannot be safely enforced in place. The local ADR draft has been updated to say this; no live-policy enforcement is shipped yet.
The current policy surfaces are read at launch:
mvmctl policy resolve/show/validate/diffare read-only (crates/mvm-cli/src/commands/policy.rs), CLI run folds named profiles and project policy into launch flags before the signed plan (crates/mvm-cli/src/commands/vm/run_policy.rs), persistent-machine create resolves policy into a persisted spec (crates/mvm-cli/src/commands/machine/create_policy.rs), and tenant bundle refs resolve at admission (crates/mvm-client/src/admission/policy_resolver.rs). There is no common resident file-to-instance owner or atomic policy-update operation. The per-VM egress projection is assembled once, so a later file edit is not itself enforced by each request's destination check.Design boundary to settle before code: arbitrary editor writes are observed only after the write by a watcher, with a nonzero/missable interval; an atomic no-stale-effect promise needs an owner-mediated update transaction that withholds affected effects before committing a new policy generation. If direct edits are supported, define a bounded, measured containment delay and label it honestly rather than claiming instantaneous enforcement. Keep the issue open; no checklist item is complete yet.
auser commentedon Nov 12, 2026
Owner refinement: prefer a rollover, not an immediate cold stop. Proposed ADR-111 now describes prepare-then-cutover: old workload keeps serving under the old committed policy while a replacement is built and verified; an early candidate guest must already satisfy signed launch preflight/isolation and has no external authority before its applicable admission completes. At cutover, fence old-generation effects, drain/cancel in-flight work, complete the candidate's signed-plan admission and readiness, switch the single host-owned route/mailbox owner, then stop the old instance. A failed candidate before cutover leaves the old instance only while its old authority remains valid; after commit it cannot be resurrected under stale authority. This is proposed design, not a shipped zero-downtime claim. Acceptance must include crash/failed-stop and accepted still-authorized mailbox preservation; if no safe handoff exists, apply cannot report success while old effects remain possible.
auser commentedon Nov 12, 2026
Production-path audit for the first implementation slice: there is no current local single-owner policy rollover seam to attach
policy applyto.mvmctl policyis read-only; the CLI resolves profiles/project policy into launch flags (crates/mvm-cli/src/commands/policy.rs,commands/vm/run_policy.rs), and ADR-014 records that the CLI callsplan_admissionandbackend.startdirectly instead of delegating the whole boot to a resident supervisor.ResolvedPolicy.layersnames contributing files but not a committed generation (crates/mvm-client/src/policy_profiles/merge.rs:45-70). The per-VM endpoint'sEgressGateis static.hot_revision_replaces_firewall_for_same_workloadinmvm-hostd::supervisorresets a mock state machine between launches (crates/mvm-hostd/src/supervisor/aggregate.rs:1563-1593); it is not a production cutover. The HVF resident handoff in the workload runner claims a standby for one new workload, not an old/new policy authority transfer.I expanded proposed ADR-111's dependency boundary: first connect committed source/generation binding to real launch+restore, a host-owned old-effect fence and single-owner route/mailbox switch, and crash-recovery evidence. A new policy-generation struct or CLI verb with no production caller would be a dormant control. No rollout claim is ready for promotion; retain all checklist items open.
auser commentedon Nov 12, 2026
Alignment with #5097: the proposed atomic policy rollover is a local implementation of its transport-independent HostInterface policy/lifecycle responsibilities, not another local admission authority.
mvm-hostd::plan_admission::admit_host_interfacealready calls the local adapter for launch preflight (crates/mvm-hostd/src/plan_admission/host_gates.rs). Keep CLI/pack UX unchanged andmvmdoptional; the managed adapter receives authenticated policy/epoch/lease authority frommvmd, while local mode resolves locally. The host-owned cutover coordinator must consume the same HostInterface contract and avoid parallel authorization gates. Atomic apply is per host; an unreachable node cannot claim immediate fleet-wide adoption and remains bounded by its signed offline lease until expiry/reconciliation. Proposed ADR-111 now cites #5097 and records this boundary; this is compatibility analysis, not a shipped implementation.