audit(Q1c): NAS automount trigger does NOT survive a guest reboot (empty-dir FAIL) + follow-up
Supervised live proof on the demo (agent 0.82.0 + controller 0.115.0): after pct reboot 9201 the autofs trigger does not re-propagate into the guest; an in-guest access sees an empty dir and does not trigger the host mount. Root cause: shared->slave propagation + no agent network-mount reassert on guest reboot. FAIL protocol honored (nothing repaired live); fix spec'd in backlog. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PSK5g6qYLknKj8u3QAFEr6
This commit is contained in:
@@ -59,6 +59,45 @@ the UI invokes — anti-F9).
|
||||
(192.168.0.180:/mnt/5_hdd/felhom-data, health ok) — untouched by this spike; it is the Part E
|
||||
(Q1c restart) subject.
|
||||
|
||||
## Q1c — NAS automount trigger survival across a guest restart: **FAIL (the dangerous empty-dir case)**
|
||||
|
||||
**Supervised (operator GO 2026-07-11), on the fully-updated stack (agent 0.82.0 + controller
|
||||
0.115.0), using the operator's own `nas-media` share (read-only).** This closes the open Q1c item
|
||||
from SPIKE-nas-verify-2026-07-11 §Q1.
|
||||
|
||||
**Method + result:**
|
||||
1. Pre-state: `nas-media` idle (host autofs trigger active-waiting; **the trigger WAS visible in the
|
||||
guest's /proc/mounts**, having propagated when the automount was enabled while the guest ran);
|
||||
apps healthy.
|
||||
2. `pct reboot 9201` (the sudoers-granted verb).
|
||||
3. Post-boot, **no access yet** → `pct exec 9201 -- grep nas-media /proc/mounts` = **NO_GUEST_TRIGGER**
|
||||
(the host trigger is intact + active, but it did NOT re-propagate into the guest's fresh namespace).
|
||||
4. In-guest `ls /mnt/felhom-drives/nas-media` → **an EMPTY directory** (`nobody:nogroup`, just `./..`)
|
||||
and the access **did NOT trigger the host mount** (no nfs4 host-side afterward). This is exactly
|
||||
the WRONG outcome the June spike flagged: a media app (jellyfin/*arr) restarting after a guest
|
||||
reboot sees an **empty library**, not the NAS content — and an app that "cleans up" a
|
||||
suddenly-empty library could destroy metadata.
|
||||
5. Apps recovered healthy (controller/traefik/cloudflared/filebrowser Up); the controller's NAS list
|
||||
still reports `nas-media` **health: idle** because it reads the HOST automount state — so the UI
|
||||
looks fine while the guest cannot actually reach the share (a silent, misleading gap).
|
||||
|
||||
**Root cause:** `mp8 /mnt/felhom-drives` is `shared` (host) → `shared,slave` (guest). Slave
|
||||
propagation only carries mount events that happen AFTER the slave relationship is (re)established.
|
||||
The host's autofs trigger for `nas-media` was mounted BEFORE the guest's post-reboot bind, so it is
|
||||
not present in the guest's new namespace; and the agent has **no network-mount reassert on guest
|
||||
reboot** (no `netmount`/reassert lines in the boot journal) — unlike the drive path's
|
||||
`ReassertGuestBinds`.
|
||||
|
||||
**FAIL protocol honored — NOTHING repaired live.** State left as found: the `nas-media` host
|
||||
automount stays configured + active; the guest sees the empty bind dir until the share is re-added
|
||||
(remove+add re-runs `EnsureNetworkMount`; the fresh `enable --now` propagates the trigger into the
|
||||
running guest) or an agent-side reassert lands. **Follow-up task** (evidence = this section):
|
||||
felhom-agent should re-propagate/re-establish network automounts into the guest after a guest
|
||||
restart — mirror `ReassertGuestBinds` for the network-storage class (a `ReassertNetworkMounts` on
|
||||
the guest-reboot path). Until then, document that a customer guest reboot requires re-touching NAS
|
||||
shares, and consider having the controller's per-share health cross-check the GUEST-visible mount,
|
||||
not only the host automount, so the UI stops showing idle/ok for a guest-unreachable share.
|
||||
|
||||
## Teardown — verified
|
||||
|
||||
`dsmsmb`/`dsmnfs` removed via the real remove endpoint (list shows only `nas-media`); felhom-pve:
|
||||
|
||||
@@ -0,0 +1,49 @@
|
||||
# FOLLOW-UP — NAS automount trigger does not survive a guest reboot (agent reassert gap)
|
||||
|
||||
**Opened:** 2026-07-11 · **Severity:** HIGH (data-safety-adjacent for media apps) · **Class:**
|
||||
implementation (felhom-agent, likely a small felhom-controller health-surface follow-on).
|
||||
**Evidence:** `documentation/audits/SPIKE-nas-dsm-2026-07-11.md` §Q1c (supervised live proof on the
|
||||
demo, agent 0.82.0 + controller 0.115.0, operator GO).
|
||||
|
||||
## The bug
|
||||
|
||||
After a **customer guest reboot**, a configured NAS share's autofs trigger is NOT present in the
|
||||
guest's fresh mount namespace, and an in-guest access **does not** trigger the host mount — the app
|
||||
sees an **empty directory** at the NAS path instead of the share content. The host-side automount is
|
||||
intact and `active (waiting)`; the controller's per-share health still reads `idle`/`ok` (it inspects
|
||||
the HOST automount, not the guest-visible mount), so the UI looks healthy while the guest cannot
|
||||
reach the share.
|
||||
|
||||
## Why
|
||||
|
||||
`mp8 /mnt/felhom-drives` is `shared` (host) → `shared,slave` (guest). Slave propagation only carries
|
||||
mount events that occur AFTER the slave bind is (re)established. The host autofs trigger was mounted
|
||||
BEFORE the guest's post-reboot bind, so it is absent from the guest's new namespace. The agent has no
|
||||
network-mount reassert on the guest-reboot path (contrast the drive path's `ReassertGuestBinds`).
|
||||
|
||||
## Impact
|
||||
|
||||
A media app (jellyfin/*arr/immich) restarting with the guest sees an empty library; an app that
|
||||
prunes/cleans a suddenly-empty library could destroy metadata. Silent because the operator UI shows
|
||||
the share healthy.
|
||||
|
||||
## Proposed fix
|
||||
|
||||
1. **felhom-agent — `ReassertNetworkMounts` on guest reboot** (mirror `ReassertGuestBinds`): after a
|
||||
guest (re)start, re-establish each configured network automount so a fresh `enable --now` (or an
|
||||
explicit re-propagation) lands the trigger in the running guest's namespace. This is the real fix.
|
||||
2. **felhom-controller — health cross-check (follow-on):** per-share health should also verify the
|
||||
GUEST-visible mount, not only the host automount, so the UI stops reporting idle/ok for a
|
||||
guest-unreachable share.
|
||||
|
||||
## Interim workaround (manual, until #1 ships)
|
||||
|
||||
Re-add the share (remove + add via the UI): `EnsureNetworkMount` re-runs and the fresh automount
|
||||
enable propagates the trigger into the running guest. A guest reboot on a customer box currently
|
||||
requires re-touching NAS shares.
|
||||
|
||||
## Current demo state (left as-found per the FAIL protocol — NOT repaired live)
|
||||
|
||||
The demo's `nas-media` share (192.168.0.180:/mnt/5_hdd/felhom-data) has its host automount active but
|
||||
the guest sees the empty bind dir. No media app is bound to it (experimental). Re-add to restore, or
|
||||
leave for the fix's live validation.
|
||||
Reference in New Issue
Block a user