Fix five real bugs found by actually installing and running the plugin

First live end-to-end install on real Unraid hardware (all 8 built
packages installed via upgradepkg, rc.podman started, containers
pulled/run/networked/port-mapped) — surfaced five genuine bugs no
amount of container-based CI testing could have caught, since none of
them exist inside the vbatts/slackware:15.0 build container:

1. rc.podman never created $PODMAN_LOG_DIR before redirecting the
   podman system service's output into it, so the service failed to
   even start ("No such file or directory"). Added it alongside the
   existing PODMAN_RUN_DIR mkdir.

2. config/storage.conf hardcoded a [storage] table, and
   podman-config.sh's `sync` step appended a second one at boot with
   the real graphroot/runroot — TOML forbids defining the same table
   twice. Removed the template's [storage] entirely; sync already
   generates the whole thing.

3. config/policy.json had a "_comment" pseudo-field for
   documentation, but containers/image's policy parser rejects any
   unknown top-level key outright. JSON has no comment syntax; moved
   the rationale into docs/ARCHITECTURE.md instead.

4. netavark >= 2.0 dropped its iptables firewall driver entirely
   (verified: passing "iptables" is flatly rejected) — nftables or
   firewalld are the only remaining options, and firewalld needs
   systemd/dbus, which Unraid has neither of. Set firewall_driver =
   "nftables" explicitly and documented that Unraid OS doesn't ship
   the `nft` binary this needs (a slackware64 nftables package works;
   not yet wired into the build/install pipeline — see follow-up).

5. Every container failed with "crun: pivot_root: Invalid argument".
   Root cause: Unraid's / is permanently the kernel's initial "rootfs"
   pseudo-filesystem (Unraid never pivots to a real one at boot — the
   whole OS runs from RAM), and pivot_root(2) unconditionally rejects
   that as the old root. This is not new: Docker/runc hits the exact
   same kernel restriction on this exact host and silently falls back
   to an MS_MOVE-based chroot; crun has no such fallback, only a
   --no-pivot flag with no config-file equivalent. Added
   plugin/sbin/crun-no-pivot.sh, a thin wrapper that scans crun's full
   argument list (podman puts global flags before the subcommand, so
   the subcommand isn't reliably $1) and injects --no-pivot right
   after create/run, and pointed containers.conf's crun runtime at it.
   Also fixed the podman.plg postinstall's chmod glob
   (`podman-*.sh` -> `*.sh`), which would have skipped this new
   non-podman-prefixed sbin script.

Verified end-to-end on the real host: pull, run, real network
connectivity (wget through the container's bridge), and a published
port actually serving HTTP (curl through -p 8099:80 to nginx) all
work. --no-pivot's security tradeoff (disabling one particular
container-escape mitigation) was explicitly discussed with and
approved by the user before committing, given it must be the default
for any container to start at all on this platform.

Follow-up not yet done: nftables (needed for #4) is not yet a
packages/ component in the reproducible build pipeline — it was only
installed manually on the test host for this verification run.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

config/containers.conf

@@ -18,6 +18,30 @@
 # TODO: static_dir / volume_path / runroot overrides pointing at the cache-pool
 # backed storage location instead of RAM-root defaults.
 
+[engine.runtimes]
+# Unraid's / is the kernel's initial 'rootfs' pseudo-filesystem — Unraid
+# never pivots to a real one during boot, the whole OS runs from RAM — and
+# pivot_root(2) unconditionally rejects that as the "old root" (EINVAL).
+# runc (what Docker uses) silently falls back to an MS_MOVE-based chroot
+# in that situation; crun has no such fallback, only a --no-pivot flag on
+# `create`/`run` with no config-file equivalent — so podman is pointed at
+# a thin wrapper (installed by the unraid-podman package, see
+# plugin/sbin/crun-no-pivot.sh) that injects it, instead of crun directly.
+# Verified live: without this, every container fails with
+# "crun: pivot_root: Invalid argument: OCI runtime error".
+crun = ["/usr/local/sbin/crun-no-pivot.sh"]
+
 [network]
-# TODO: default network backend (netavark), default subnet range distinct from
-# Docker's docker0 range — see docs/ARCHITECTURE.md section 8.
+# TODO: default subnet range distinct from Docker's docker0 range — see
+# docs/ARCHITECTURE.md section 8.
+#
+# netavark >= 2.0 dropped its iptables firewall driver entirely — only
+# nftables and firewalld remain (verified against the actual netavark
+# binary; "iptables" is rejected with "Must provide a valid firewall
+# backend"). firewalld needs systemd/dbus, which Unraid has neither of, so
+# nftables (netavark's own default — explicit here so that stays true even
+# if netavark's default ever changes) is the only viable driver. Unraid OS
+# does not ship the `nft` binary this needs — see docs/ARCHITECTURE.md
+# section 8 for how it's provisioned.
+network_backend = "netavark"
+firewall_driver = "nftables"

config/policy.json

@@ -1,5 +1,4 @@
 {
-  "_comment": "Default container image signature verification policy. Placeholder: accepts all images without signature verification, matching Docker's default trust model on Unraid today. See docs/ARCHITECTURE.md section 10 (Images). Revisit before a 1.0 release if signed-image verification becomes a goal.",
   "default": [
     { "type": "insecureAcceptAnything" }
   ],

config/storage.conf

@@ -7,14 +7,13 @@
 #
 # Full reference: https://github.com/containers/storage/blob/main/docs/containers-storage.conf.5.md
 
-[storage]
-driver = "overlay"
-# runroot and graphroot are set at runtime by rc.podman based on
-# /boot/config/plugins/podman/podman.cfg (configurable storage location),
-# not hardcoded here. TODO: document the exact substitution mechanism once
-# rc.podman is implemented.
-# graphroot = "/var/lib/containers/storage"
-# runroot = "/var/run/containers/storage"
+# The [storage] table itself (driver, graphroot, runroot) is intentionally
+# NOT defined here — podman-config.sh's `sync` command appends it in full
+# at sync time, since graphroot depends on STORAGE_PATH from
+# /boot/config/plugins/podman/podman.cfg (configurable storage location) and
+# can't be known statically. A second [storage] table here would be a TOML
+# duplicate-key error once sync appends its own — see podman-config.sh's
+# cmd_sync for the generated content.
 
 [storage.options]
 # TODO: overlay-specific mount options once the loopback filesystem

docs/ARCHITECTURE.md

@@ -339,6 +339,12 @@ Unraid-Plugins.
 ## 8. Netzwerke
 
 - **Backend**: `netavark` + `aardvark-dns` (Podman-Default seit 4.x), kein CNI in Phase 1.
+- **Firewall-Treiber**: `netavark` >= 2.0 hat seinen `iptables`-Treiber ersatzlos entfernt —
+  nur noch `nftables` oder `firewalld` (Letzteres braucht systemd/dbus, hat Unraid nicht).
+  `nftables` ist also zwingend, aber Unraid OS bringt das `nft`-Binary selbst nicht mit
+  (nur das ältere `iptables`/`iptables-nft`) — muss vom Plugin bereitgestellt werden
+  (verifiziert per Live-Test: `podman run` mit Port-Publishing scheitert ohne `nft`
+  mit „Must provide a valid firewall backend“).
 - **Default-Bridge**: eigene Bridge `podman0` (nicht `docker0`), eigener privater
   Adressraum (konfigurierbar, Default-Vorschlag außerhalb von Dockers Default-Range,
   um Kollisionen bei Parallelbetrieb zu vermeiden).
@@ -397,6 +403,12 @@ Unraid-Plugins.
 - **Community-Applications-Kompatibilität**: kein automatischer Import von
   Docker-Templates in Phase 1 (eigenes, separates Vorhaben); Podman-Images werden
   zunächst über CLI/eigene, minimale Template-Definition verwaltet.
+- **Signatur-Policy** (`config/policy.json`): `insecureAcceptAnything` als Default —
+  akzeptiert Images ohne Signaturprüfung, entspricht Dockers heutigem
+  Standard-Vertrauensmodell auf Unraid. Vor einem 1.0-Release erneut bewerten, falls
+  signierte Images ein Ziel werden. Die Datei selbst darf keine Kommentarfelder
+  enthalten (`containers/image`s Policy-Parser lehnt unbekannte Top-Level-Keys wie
+  `_comment` strikt ab) — Begründung lebt deshalb hier, nicht in der Datei.
 
 ---
 

plugin/podman.plg

@@ -274,7 +274,7 @@ set -u
 
 echo "Setting permissions..."
 chmod 0755 /etc/rc.d/rc.podman
-chmod 0755 /usr/local/sbin/podman-*.sh
+chmod 0755 /usr/local/sbin/*.sh
 chmod 0755 /usr/local/emhttp/plugins/podman/event/disks_mounted
 chmod 0755 /usr/local/emhttp/plugins/podman/event/stopping
 

plugin/rc.d/rc.podman

@@ -98,7 +98,7 @@ podman_start() {
   #    a shared process gives consistent state and event streaming.
   #    --time=0 disables the idle-shutdown timeout (this is a long-running
   #    daemon under our process management, not an on-demand activation).
-  mkdir -p "$PODMAN_RUN_DIR"
+  mkdir -p "$PODMAN_RUN_DIR" "$PODMAN_LOG_DIR"
   podman_log "start: starting podman system service on unix://$PODMAN_SOCKET"
   nohup podman system service --time=0 "unix://$PODMAN_SOCKET" \
     > "$PODMAN_LOG_DIR/podman-service.log" 2>&1 &

plugin/sbin/crun-no-pivot.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+# =============================================================================
+# plugin/sbin/crun-no-pivot.sh
+#
+# Thin OCI-runtime wrapper around crun, injecting --no-pivot on `create`/
+# `run`. Unraid's / is the kernel's initial "rootfs" pseudo-filesystem —
+# Unraid never pivots to a real one during boot, the whole OS runs from
+# RAM — and pivot_root(2) unconditionally rejects that as the "old root"
+# (EINVAL). runc (what Docker uses) silently falls back to an MS_MOVE-based
+# chroot in that situation; crun has no such fallback and no config-file
+# equivalent, only this per-invocation flag — so config/containers.conf
+# points podman's crun runtime at this wrapper instead of crun directly.
+# See docs/ARCHITECTURE.md section 8.
+#
+# Verified live: without this, every container fails with
+# "crun: pivot_root: Invalid argument: OCI runtime error".
+#
+# --no-pivot is only a valid flag on crun's create/run subcommands (see
+# `crun run --help`) — every other subcommand (delete, exec, kill, list,
+# ...) is passed straight through unmodified. podman invokes crun with its
+# own global flags BEFORE the subcommand (e.g.
+# `crun --log-format=json --log <path> create --bundle <path> ...`), so
+# the subcommand is not reliably $1 — scan every argument instead of only
+# checking the first one, and insert --no-pivot immediately after
+# create/run wherever it actually appears.
+# =============================================================================
+
+args=()
+found=0
+for arg in "$@"; do
+  args+=("$arg")
+  if [ "$found" -eq 0 ] && { [ "$arg" = create ] || [ "$arg" = run ]; }; then
+    args+=("--no-pivot")
+    found=1
+  fi
+done
+exec /usr/bin/crun "${args[@]}"