Review 4727756972 (A2): the backout no longer trusts gpasswd — it re-reads the live group database after removal; verified-absent gets the safe message, anything else screams ROLLBACK INCOMPLETE, exits nonzero, and names the exact remediation. The concurrent-login window (a session begun between usermod and backout keeps the group) is CLOSED to the extent the database can't reach: the backout detects live processes and names loginctl terminate-user, and the success wording claims only what was verified. Review 4727641752 (A1): a failed grant for a user whose membership predates the run (the hand-added-user scenario) now fails LOUDLY — they retain socket access on part-converged policy, and the message says so with both remediations (box revoke now, or fix and re-run). Their membership is not stripped: breaking a working user over a failed re-grant is its own hazard. The default-profile eth0 removal is deliberately not restored on failure — that mutation only reduces capability, and restoring it would move the failure state AWAY from fail-closed. Injected-failure coverage is criterion (n), both flavors: fresh-user backout (fault at the LAST mutation, so the rollback runs after every earlier one) with the group's absence verified and a converging re-run; blocked narrowing staged for real with an instance-local NIC parked on the private bridge. Review A3, resolution 3 with the measurement demanded: criterion (m) launches exactly 'incus launch --network boxnet' as the restricted user and probes the raw NIC from inside — egress works, RFC1918 dropped (the ACL is the network's), sibling probes dropped BOTH directions (the nft drop is the host's), name enumeration blocked. The scoped guarantee is now stated in box-design.md and measured on every run: box-minted instances carry per-NIC port_isolation; raw attachments keep every network- and host-owned control, losing only that redundant L2 layer. Instrument lesson kept as MU-5: the probe's first cut minted the non-cloud image — no DHCP client, no lease, and a dead NIC passes every negative probe vacuously; it now requires the lease before believing its own answers. Rehearsal: 54/54 (containers). test/cli.sh: 82 checks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
7.9 KiB
Restricted incus tier — design and measured results (#74)
Status: implemented and rehearsed. 54/54 rehearsal criteria green on the
design host (Debian 13 trixie, Incus 6.0.4, nested KVM), in container mode
and VM mode — and green in CI on ubuntu-latest / incus 6.0.0 (whose one
version-drift false FAIL is MU-4 in drill/RUNS.md). This doc records the design, what was measured, and why each
decision fell where it did. It supersedes the vetoed #72 design
(docs/plans/2026-07-17-multiuser-hosts.md on feat/restricted-tier-wip).
What #74 asked
A restricted (incus-group) user can box new/list/shell/snapshot/rm their
own boxes, on a network carrying box's full isolation contract, seeing no
one else's; the convergence is a documented, idempotent path; the rehearsal
passes criteria (a)–(f); the admin tier is unchanged.
The three facts that shaped the design
Task-0 (#72) measured one: incus-user confines users to user-<uid> projects
(sound), but pins them to a private auto-created bridge incusbr-<uid> and
restricted.networks.access: incusbr-<uid> — they cannot even see boxnet.
This round measured two more:
- The private bridge is worse than unhardened.
incusbr-<uid>is a fully functional NAT bridge —ipv4.nat=true,ipv6.nat=true— with no ACL, nodns.mode=none, no resolver pin, no port isolation, and IPv6 egress box's contract explicitly forbids. Any instance placed on it holds a door to the host's LAN. - incus-user projects block snapshots (
Project "user-<uid>" doesn't allow for snapshot creation) — box's entire reuse workflow.
And two open questions from #74, answered from incus-user's own source
(cmd/incus-user/server.go, stable-6.0) and then confirmed live:
- Daemon-level project template? None exists — the project config is
hardcoded in
serverSetupUser(). A per-user admin hook is the only path. - Does widening survive a re-sync? Yes. Setup runs only when the project
does not exist (and early-outs when the user's certificate is already
trusted); incus-user never rewrites an existing project. Confirmed live:
systemctl restart incus-user.socketleaves the convergence intact.
The decision: option 1, tightened
#74 offered (1) converge users onto the shared boxnet or (2) harden each
private bridge. Option 2 multiplies every mechanism per user (ACL, resolver
pin, dnsmasq, nft rules, firewall coexistence) and turns the shipped static
profile into N generated ones. Option 1 keeps one hardened network and one
shipped profile — and the existing box↔box mechanisms already make
cross-user isolation free: the nft bridge-family drop and dns.mode=none
are host/network-owned, so they bind every instance on boxnet no matter
whose project it lives in.
One tightening beyond the issue's sketch: the issue proposed
restricted.networks.access boxnet,incusbr-<uid> ("must list both" — true
as long as the default profile still references the private bridge). Listing
both leaves fact 1's unhardened bridge one --network flag away, forever.
Instead, box grant:
- removes
eth0from the project's default profile (nothing references the private bridge anymore, so the narrowing validates), and - sets
restricted.networks.access boxnet— only.
After which the hardened network is not the user's default placement but the
only placement their certificate can express. Measured: incus launch --network incusbr-<uid> as the user → Network not found; the user cannot
widen their own project (Error: Certificate is restricted); they cannot
touch boxnet's config or the ACL (no permission for project "default").
A restricted user CAN edit the box-net profile copy in their own project
(they own project profiles — features.profiles=true), including stripping
security.port_isolation — and CAN attach boxnet raw with --network boxnet, no profile at all (the network must be in
restricted.networks.access for the profile to work; there is no
allow-via-profile-only lever). That is why the host-owned nft bridge drop is
the second layer: meta ibrname boxnet obrname boxnet drop fires on every
port-to-port frame regardless of per-NIC flags. The documented guarantee is
scoped accordingly (see box-design.md): box-minted instances carry per-NIC
port isolation; raw attachments keep every network- and host-owned control,
losing only that redundant L2 layer. Both shapes are measured from inside
the instances (rehearsal criteria g and m).
Two grant-failure contracts, both injected in the rehearsal (criterion n): a fresh user is backed out of the group with the removal VERIFIED against the live group database (and any session begun mid-grant is named, with the loginctl remedy — the one window the database cannot close); a pre-existing member is never stripped by a failed re-grant, but the failure states out loud that they retain socket access on part-converged policy, with both remediations. The default-profile eth0 removal is deliberately NOT restored on failure: that mutation only reduces capability, and restoring it would move the failure state away from fail-closed. Every step is check-then- converge, which is what makes re-run-to-repair deterministic.
What box grant <user> converges (idempotent, re-run to refresh)
usermod -aG incus(notincus-admin— that is the tier)- first-touch incus-user as the user (
runuser/sudo -u, stdin pinned) — the project is created lazily and cannot be pre-created by an admin - remove the default profile's private-bridge
eth0 restricted.networks.access boxnetrestricted.snapshots allow- install/refresh the shipped
box-netprofile into the project - verify from the user's side of the socket
box revoke <user> is the inverse, two strengths: bare = group removal (the
socket closes; boxes keep running; re-grant restores), --purge = boxes,
images, project, private bridge, trust-store certificate, incus-user state —
then asserts the absence (the wipe.sh discipline).
The tier in the CLI
box_tier() — UID 0 / incus-admin → admin, incus alone → restricted,
neither → none — decided from live process credentials (argless id -nG),
byte-identical in bin/box and host/setup-host.sh (diffed by a test).
Tier-aware surface: new pre-flights the profile and names the right fix per
tier; expose refuses before any daemon call (its plumbing is daemon-global;
without the guard the failure is a lie — a restricted user cannot read
boxnet's redacted config, so box_net_ip would claim their running box has no
address); setup-host exits 0 with the honest note; doctor runs a
restricted check-set (is the tier granted, does their box resolve/route)
instead of judging host state they cannot see.
Found along the way, fixed for every tier: box restore dispatched
incus restore, which does not exist in Incus 6 (incus snapshot restore).
It had never worked.
Rehearsal and CI
drill/multiuser.sh (root, opt-in via BOX_MULTIUSER_REHEARSAL=1) proves
criteria (a)–(f) from #74 plus the measured extensions (g)–(n): the in-box
isolation contract (egress, DNS, box→host, RFC1918, cross-user sibling drop,
name enumeration, IPv6-off), the closed escape hatches, re-sync survival, and
scoped revoke, the raw-attach scoped guarantee (m) and the grant-failure
injections (n). Real users, real grants, real mints, probes from inside;
--container for CI, VM mode on real hardware; cleanup deletes everything it
made.
CI gains a rehearsal job (ubuntu-latest): install incus, stage the checkout
at /opt/box (the #71 layout), setup-host, doctor, then the rehearsal in
container mode. The tier's semantics are proven on every PR against a live
daemon; the VM trust boundary stays a real-hardware ritual, like the drill.
Environment
Debian 13 (trixie), Incus 6.0.4, incus-user.socket shipped in the incus
package (Debian 13 and Ubuntu 24.04 both), /dev/kvm present (VM-mode run),
btrfs pool. Companion work: rig#24 (box role), rig#12/#25 (host-class).