The installers write two allow-lists, /etc/sudoers.d/ledmatrix_web and
ledmatrix_wifi. Anything the code runs under sudo that is not in one of them
needs a password, which a service cannot supply, so the call fails.
Five commands were being run and none of them granted:
sysctl -w net.ipv4.ip_forward=0|1 wifi_manager.py:788, 883
nft add|delete table ip ledmatrix wifi_manager.py:835, 895
rfkill unblock wifi wifi_manager.py:1811
iptables ... wifi_manager.py:796, 813, 818, 871
mkdir -p .../dnsmasq-shared.d wifi_manager.py:922
Together these are the captive portal: unblock the radio, bring up the AP,
add the redirect, turn on forwarding, and undo all of it afterwards. Without
the grants a hardened install would associate clients to the access point and
then fail to route them.
Why it has gone unnoticed: a stock Raspberry Pi image ships
/etc/sudoers.d/010_pi-nopasswd granting the default user
<user> ALL=(ALL) NOPASSWD: ALL
which satisfies every one of these regardless of what the allow-lists say.
Confirmed on a live rig -- `sudo -n -l` permits sysctl there, and the blanket
rule is why. The allow-lists are effectively decorative on a default image and
only start mattering once that rule is removed or the service runs as another
user.
test_sudo_allowlist_covers_calls.py extracts every argv-style sudo call in
src/ and web_interface/ and asserts an installer grants it, so the next command
added without a rule fails here rather than on someone's hardened box.
Getting that test honest took three passes, each worth recording:
- Matching the literal "systemctl" against rules written as
`$SYSTEMCTL_PATH enable ...` reported six gaps that did not exist. Binary
path variables are now normalised before comparing.
- Scanning the whole installer let `NFT_PATH=$(command -v nft)` -- a variable
definition, not a grant -- satisfy the check on its own, so deleting the
actual nft rules still passed. Only NOPASSWD lines are considered now.
- `sudo -n <tool>` reported "-n" as the binary. sudo's own flags are skipped.
Each of the five grants is individually mutation-checked: removing any one
fails the suite.
(cherry picked from commit a372b43cd1)
Nothing re-applies systemd units after the first install. `git pull` -- which
is what the web UI's update button runs -- brings a new template into the
checkout, but no code in web_interface/ or src/ copies it to
/etc/systemd/system, and nothing anywhere runs `systemctl daemon-reload`. The
unit that actually runs is whatever first_time_install.sh wrote on day one.
So every hardening added to a unit is inert on existing installs, silently.
Measured on a live rig:
installed /etc/systemd/system/ledmatrix.service 2026-08-06
template systemd/ledmatrix.service 2026-08-19
contents differ
with the practical result that the MemoryMax=85% the repo's template specifies
was not being enforced at all -- `systemctl show` reported
MemoryMax=infinity. Anyone reading the template would reasonably believe the
service was capped.
Startup now compares each installed unit against its substituted template and
warns when they differ, naming install_service.sh as the remedy.
A warning, not an error, and deliberately not a silent rewrite: editing files
under /etc and restarting services is the installer's job, not something a
display process should do to a machine while it is booting. Making it fatal
would also brick every development checkout whose unit is legitimately absent
or hand-edited.
Comparison ignores comments, blank lines and ordering. The template carries
explanatory comments the installed copy will not have, and systemd does not
care about order within a section, so a literal comparison would warn on every
boot and be ignored within a week.
Mutation-checked three ways: never reporting drift fails, making it fatal
fails, and -- after the first attempt missed it -- comparing raw text now fails
too. That last gap is worth noting: the comment-insensitivity tests originally
exercised the helper directly, so a comparison that stopped calling the helper
passed them all. The test that catches it goes through _validate_systemd_units.
29 startup-validator tests pass.
(cherry picked from commit cf521bdfd8)
Review follow-up. The range check accepted 1, 3 and 4, so a change to 4 --
which hands most of the resident saving back -- passed a test whose whole
purpose is to notice that.
Pinned to the value the unit ships, in one named constant. Raising it is still
a legitimate response to a frame-time regression, but it should be a visible
edit here rather than silent drift, and the failure message says so.
Mutation-checked: changing the unit to 4 now fails.
(cherry picked from commit 73fff8d2d5)
Measured on a live rig 2.5 hours after start:
RSS 1030 MB
Private_Dirty 988 MB
anonymous mappings > 10 MB 23
largest 104, 79, 66, 63, 63 MB, on 64 MB-aligned addresses
threads 9
cores 3 -> glibc ceiling = 8 x 3 = 24 arenas
23 against a ceiling of 24, all 64 MB-aligned: these are glibc's per-thread
malloc arenas, not live objects. The data the process was actually holding
accounts for perhaps 15 MB -- the widest scroll strip observed was 35,746 x 64,
about 7 MB as RGB and the same again for its numpy mirror.
It is bloat rather than a leak: sampled four times over 135 seconds, RSS sat
between 990 and 1030 MB rather than climbing. glibc gives each allocating
thread its own arena, grows them to hold peak demand, and never gives them
back. A process that builds and drops large images across several threads is
exactly the shape that produces this.
The device had 59 MB free at the time, on 1845 MB total.
MALLOC_ARENA_MAX=2 trades a little allocator concurrency for that resident
memory. It is a tuning knob rather than a fix for a defect, so the rationale
and the measurements sit next to it in the unit file, and a test asserts they
stay there -- a bare environment variable invites removal by whoever meets it
next.
Two things this is NOT, both checked rather than assumed:
- Not an OOM problem today. A grep for "oom" in the service journal returned
24 matches, all of which were the radar logging zoom=9 and zoom=7. The kernel
OOM killer has not fired: dmesg has zero matches.
- Not currently capped by the unit's MemoryMax=85% either. That directive is in
this file but absent from the unit actually installed on the rig, which
reports MemoryMax=infinity, so nothing is enforcing a ceiling there.
The saving is unmeasured on hardware: applying it needs a service restart,
which blanks the panel, so that is the user's call rather than something to do
mid-audit. If p99 frame time regresses -- it sits at 18.4 ms against a 16.7 ms
budget for 60 FPS, so there is not much headroom -- raise the value rather than
remove it.
(cherry picked from commit 446207ffbc)
The 3.2.0 section described the unified sports library but none of the
install-path work that landed in #428 and #431 -- which matters more than a
normal changelog omission, because the sunset rule keys on this section to
tell plugin authors what a given floor buys them.
The headline addition: 3.2.0 is the first release that *enforces*
ledmatrix_min_version. Before it the floor was advisory, so a plugin could
declare one and still be delivered to a core that could not run it. That is
the property B6 waits on, and it is now stated where a plugin author will
look for it -- along with the caveat that a core reporting below 2.0.0 is
treated as unknown rather than old and is never blocked.
Also records compatibility.py (and that it does not yet read
compatible_versions), check_release_version.py and its workflow, the
install-preservation fix, the reentrant-lock deadlock fix, and the
web_interface version re-export.
No version bump: 3.2.0 is unreleased, so this describes the release being
cut rather than a new one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Udr6MfaFLUPhX5Fgo67Jf5
ChuckBuilds