push_mac_acl wrote macfilter='allow' for "whitelist" and 'deny' for any
other mode but "off", so a mode the driver does not know -- a typo, or a
new mode from the caller -- became a deny list. It now raises ValueError
naming the mode before the first command is sent.
Closes#39
push_mac_acl and push_radio_channel sent each step through the
interactive shell and never looked at the result: a refused uci set,
add_list, commit or wifi reload passed silently, an SSID no wifi-iface
carried did nothing, and wifi was restarted after a refused channel.
Each step now runs on an SSH exec channel of its own, as root, with its
exit status checked (uci_steps.UciStepsMixin). A refused step raises
CommandErrorException naming the command and what the router said; a
refusal before the commit reverts what was staged, so a later commit
does not apply half of it, and wifi is not restarted. push_mac_acl
reads the sections once, raises when none carries the SSID, and deletes
macfilter/maclist only where a section has them instead of hiding each
delete's status behind `|| true`.
Closes#35
DHCP option 42 counts (NetOrk/netork#2380): a router that gets the
site's time giver from its lease is aligned. The first time sync commit
set use_dhcp=0 on every write, which cut a router off from the giver its
DHCP server hands out.
A write now leaves use_dhcp as the router has it, and the "already has
these servers" check no longer rewrites a router only because use_dhcp
is on. get_time_sync() reads what sysntpd adds to ntpd's list while
use_dhcp is on -- each interface's data.ntpserver in ubus call
network.interface dump, which /lib/netifd/dhcp.script fills from
udhcpc's ntpsrv, only the dhcp_interface list's where it names any -- in
the same round trip as uci, into dhcp_ntp_servers: [] while use_dhcp or
the client is off, None where netifd gives no answer.
Needs napalm-device-types 4.36.0 (TimeSyncDict.dhcp_ntp_servers).
netOrk aligns every device's time sources and zone through the driver
(NetOrk/netork#2360); for APs that replaces the UCI commands netOrk
built itself for the AP profile fields (NetOrk/netork#2364).
Both values live in UCI's system config:
- The servers are system.ntp.server. A write replaces the list, enables
the client, sets use_dhcp=0 so the servers of a DHCP lease are not
added, and restarts sysntpd: the router ends up with exactly these.
A disabled client reads as no servers.
- The zone is zonename (the IANA name, in LuCI's spelling with spaces)
and timezone, the POSIX rule from posix_tz that libc reads as TZ. The
rule is what counts (NetOrk/netork#1849), so the name is read as the
zone only where the rule in force is that zone's; a rule alone reads
as the one zone that has it, or None where several share it. A write
sets both, reloads /etc/init.d/system and restarts logd, which reads
the zone at start.
netOrk aligns on every poll, so each step compares first and writes
and restarts nothing when the router already has the values. GMT0 (the
stock rule) counts as UTC's time.
get_ntp_servers now uses the same reading, so it agrees with
get_time_sync; it used to keep only the first server of a list, as uci
prints a list as 'a' 'b' on one line. synchronized is None: BusyBox
ntpd keeps no state to read.
The floor on napalm-device-types is 4.35.0, the first with
TimeSyncMixin.
The router's host status, read by napalm-device-types' command over an
SSH exec channel: whether it needs a reboot, how full its filesystems
are, and, with napalm-device-types 4.32, whether apk or opkg was left
halfway.
Refs NetOrk/netork#851
get_scheduled_jobs reads every crontab in /etc/crontabs in one round trip
as root, with the time zone from UCI and each crontab's package from opkg;
manage_scheduled_job enables, disables or runs one now; put_ and
delete_scheduled_job write and remove an application's own job as marked
lines in the user's crontab. BusyBox crond has no /etc/crontab, cron.d,
run-parts, anacron or systemd, so none is read.
OpenWrt's BusyBox differs from what the POSIX mixin assumes, so a few
steps are the driver's own:
- crontab -l says only "can't open" for a missing crontab, as for an
unreadable one: the file is read directly, a missing one is empty, and
one that cannot be read is never written.
- crontab FILE exits 0 when it could not write: every write is read back.
- cron is enabled and restarted after each write; OpenWrt does not start
crond while /etc/crontabs is empty.
- crond has no special times: aliases are written as five fields and
@reboot is refused.
- crond hands the whole command to sh and takes only SHELL and PATH from
the crontab, so a run now does the same (% is no stdin).
- there is no su and no timeout before 25.x: another user's job runs
through start-stop-daemon -c, and the run is bounded by timeout only
where the router has it.
Every write changes only the job's own line, so the reboot failback line
in /etc/crontabs/root stays as it is.
Requires napalm-device-types 4.26.0.
Closes#32
lldpd tells the chassis MAC, but get_lldp_neighbors reported only name
and port, so netOrk, which matches a neighbour by MAC first, could match
OpenWrt's neighbours by name only. napalm-device-types' lldp_neighbors
projection now carries mac (lower case) when the chassis id is a MAC
address, as napalm-linux reports it.
Closes#31
lldpctl's error text has no "=", so a missing lldpctl or a stopped
lldpd read as a device without neighbours. lldpctl now runs as
napalm-device-types' LLDPCTL_KEYVALUE_COMMAND, which brings its exit
status and message back, and the shared parser raises on a failure.
get_environment swallowed an unreadable /proc/stat or /proc/meminfo
and reported 0 % CPU and 0 bytes of memory, a plausible value instead
of a failure. Both now raise ValueError; 0 % stays for counters that
did not move between the two reads.
Closes#27
The local interface was the second dotted part of the key, so the
neighbours of eth0.1 landed under eth0 with every field missed. The
remote port was read from port.ifname or port.id.value, and lldpd
prints a port given by MAC or local id as port.mac / port.local, never
as id.value, so that port came back empty.
lldpd's output now goes through napalm-device-types'
parse_lldpctl_keyvalue and its two NAPALM projections, which keep the
interface name up to the neighbour's first field and read the ids under
whatever type lldpd prints. The detail gains the system description and
capabilities on the way. The module is under strict mypy now; the floor
on napalm-device-types rises to 4.24.0.
Closes#26
install_package and uninstall_package searched the output for keywords.
opkg's dependency failure ("Collected errors:") contains none of them and
read as success; a successful install naming a "failed" file read as a
failure. Append the exit-status marker and let the status alone decide.
Closes#28
The LLDP getters started lldpd and rewrote its interface list on every
poll; with read-only getters that setup needs an action. The driver
mixes in napalm-device-types' LldpAgentMixin and writes the step the
template leaves to it: lldpd.config.interface in UCI. A VLAN device is
written as the device under it, as the old helper did, since lldpd
sends on the L2 device. Which interfaces is the caller's choice.
Refs #8
get_lldp_neighbors, which a poll calls every time, enabled and started a
stopped lldpd and rewrote lldpd's interface list with uci, restarting
the daemon. A getter must not change the device. Both LLDP getters now
only run lldpctl; installing and starting lldpd are ensure_package and
manage_service.
Closes#8
The keyvalue output was grouped by local interface only, so with several
neighbours on one port their keys overwrote each other and one mixed
entry came back. A neighbour now ends where the next one opens (alias,
then via, as lldpd prints it); a key repeated inside one neighbour, such
as one mgmt-ip per address, does not start a new one.
Closes#9
get_environment promised two samples one second apart but read
/proc/stat once, so the figure was the average since boot: a box at
100 % for the last minute showed its lifetime average. It now reads the
cpu line twice, a second apart, and uses the difference.
Closes#10
apply_updates ran apk/opkg upgrade and returned success whatever
happened, so a failed upgrade was reported as done. The command now
prints its exit status after the marker get_available_updates already
uses; a non-zero or missing status is a failure, named in "error".
Closes#11
manage_service returned success for every action, so a failed start or
a missing init script showed up as done. The command now prints its
exit status after a marker, and only that status decides.
Closes#3
- Long lines are wrapped: the get_environment awk command is split into
adjacent string literals, the rollback message and the system mixin's
docstring are wrapped; the command and the text are unchanged.
- The two except Exception: pass (the snmpd community in get_snmp_config,
the network VLANs in get_ssids) now log at debug level (B110), so a
swallowed failure leaves a trace.
- tests: raise ... from exc, unused loop variable.
Mechanical: ruff format, and ruff check --fix with the rule set every NAPALM
driver uses (E, F, I, UP, B at line length 100, Python 3.10). What ruff
cannot fix comes in the next commit.
The driver accepted any SSH host key. It now verifies through
napalm-device-types' HostKeyCheck (>= 4.19.0): with a known key in
optional_args["ssh_host_keys"] any other key is refused before login, and
every key seen is reported to optional_args["ssh_host_key_observer"].
Without either argument it connects as before (first use).
HostRebootMixin's reboot_host over the command channel: reboot detached
and two seconds late, so the answer comes back before the session goes
down, and a refusal is raised instead of being mistaken for the dropped
session. netOrk sent /sbin/reboot through the private _send_command and
could not tell the two apart (netOrk #1113). A login other than root is
refused before anything runs.
run_command and open_stream (napalm-device-types CommandChannelMixin,
which AccessPointDriver carries) run on an SSH exec channel next to
netmiko's shell: stdin, stderr apart, a real exit code. netOrk sent UCI
scripts and the wifi channel through the private _send_command, typing
a file in base64 chunks or printf lines; with stdin it writes it with
cat, and base64 is needed for nothing (netOrk #1113). OpenWrt has no
sudo, so a privileged command from a login other than root is refused
before it runs.
get_device_warnings goes with it: no_base64 was its last code.
get_device_warnings no longer names lldpd_not_installed,
updates_available or update_notifications_disabled. The packages and
get_available_updates are the facts; netOrk derives the warnings with the
same reach (netOrk #1111). The old update read swallowed failures,
get_available_updates raises them.
no_base64 stays: BusyBox may provide base64 as an applet, so no package
tells the caller whether it is there.
fix_snmp set no community at all and opened the management zone's
udp/161 to every source, with a rule named after an application; a
second fix_snmp in the system mixin was shadowed and never ran (#5).
install_lldpd, install_auc and install_coreutils_base64 were actions of
their own.
The driver implements napalm-device-types 4.13's templates instead:
- configure_snmp_agent(settings): snmpd installed when missing;
community, contact and location in UCI with one com2sec (com2sec6)
per allowed source; the zone-resolving firewall step of the old fix,
now limited to the sources (src_ip) and named Allow-SNMP; snmpd
enabled, restarted and listening;
- get_management_peer() from SSH_CONNECTION;
- ensure_package(name) from the base package, over install_package.
A community, contact or location that would break UCI is refused.
BREAKING CHANGE: run_device_action is removed. Requires
napalm-device-types 4.13.
Closes#5. Part of #13.
Mixes in napalm-device-types' ListeningSocketsMixin (2.5.0): without ss
the shared command reads busybox netstat, and each process's cgroup names
its procd service. The driver only carries the command across.
Over an SSH exec channel of its own, not the interactive shell: busybox
ash cuts a typed line at 512 characters, the command is longer, and the
cut line left the shell waiting for a closing quote. OpenWrt logs in as
root; a login other than root cannot become root, so its reading says it
is not attributed.
Checked against a real OpenWrt 25.12.2 access point. For netOrk#673.
OpenWrt keeps its package index in RAM. After a reboot `apk version` only
warns that it cannot open its cached index and exits 0, and `opkg
list-upgradable` prints nothing: both read as "no updates". netOrk keeps
"pending since" per package (MVP 5), so the reader now raises instead:
- apk: on apk's own warning ("opening from cache ... No such file"), or a
failed exit status.
- opkg: when /var/opkg-lists is empty.
- refresh_available_updates(): apk update / opkg update, success from the
exit status.
get_system_config() already parsed the whole system section but returned
only a slice of it, so netOrk had no IST side for three AP-profile fields
and could not compare them at all — the fields were editable, stored and
silently ineffective.
Adds:
* syslog_remote / syslog_ip / syslog_port / syslog_proto — read from the
system section that was already being parsed. log_remote gates the
others: OpenWrt ships nothing without it, so a leftover log_ip must
not read as an active target.
* luci_enabled — from the uhttpd init script rather than its listener
config, so the answer stays "is the web UI served" and re-enabling
restores whatever was configured before.
* bridge_stp — None when there is no br-ap at all, which is a different
statement from "bridge without STP" and has to stay distinguishable.
netOrk #164
fix_snmp reported success on APs where the rule never reached nftables.
Five defects stacked up:
1. Zone detection required ".src=" and "ssh" in the same `uci show` line.
UCI prints one option per line, so anonymous rules never matched and
every device fell through to the hardcoded "lan" fallback.
2. That fallback was never checked against the zones that actually exist.
On an AP whose zone section has no `option name`, fw4 skips the section,
so `src='lan'` referenced a zone that was not there and the rule was
dropped with it.
3. The "already present" guard was a substring test, so a rule written by
an earlier broken run was skipped forever instead of repaired.
4. Stale-rule deletion never committed — the only `uci commit firewall`
sat in the add branch that the guard had just skipped.
5. `fw4 reload` errors were swallowed by `|| true`, and with no local
snmpget the action hardcoded success = True.
Now: the management address comes from $SSH_CONNECTION and is mapped to
its network section (via ipaddr, or via `ip -o -4 addr` -> device when the
interface is DHCP-addressed) and from there to the owning zone. A zone
section without a name aborts the action with the repair command rather
than writing a dead rule — naming it is left to the operator, since an
inert zone becoming active changes what the AP filters. Rules are written
in full every run, stale ones are deleted highest anonymous index first
(uci renumbers @rule[n] on delete) and committed, reload output is no
longer truncated or ignored, and success is verified on the device via
`ss -lun` and a udp/161 lookup in the live ruleset.
get_vlans() relied on `bridge vlan show` for port membership. On devices
whose BusyBox ships without the bridge/ip-full packages the command does
not exist, _send_command returns the shell error, and the parser silently
finds nothing in it. All that survived was the sub-interface fallback,
which restates the bridge topology (br-ap tagged, br-ap.10 untagged) and
never names the uplink port — the only port whose tagging matters.
Measured on two Sophos AP100 (BusyBox 1.37.0): eth0 was absent from the
output entirely, while UCI held ports='eth0:u*' for the management VLAN
and 'eth0:t' for the rest.
UCI bridge-vlan sections are now parsed as a second source. They are
readable without the bridge binary and describe the configured state.
Section collection goes through the type declaration, so named sections
(network.apbr_vlan10) are recognised alongside anonymous ones — the old
regex matched only @bridge-vlan[N], so a hand-built bridge was invisible
even for name lookup. Runtime data keeps precedence where it exists,
since that is what the kernel actually enforces.
Option values are kept raw through collection; stripping quotes there
would collapse ports='lan1:t' 'lan2:t' into a single mangled item.
Also repairs TestGetVlans, red on master since get_vlans() moved to
separate tagged/untagged lists while the tests still asserted the old
'interfaces' key, and TestGetFacts, which never learned about the
number_of_interfaces key get_facts() sets deliberately. Both had left
the interesting behaviour uncovered.
OpenWrt's wifi-scripts validator rejects any macfilter value other than
"allow"/"deny" outright — confirmed on real hardware, setting
macfilter='disable' puts netifd in a permanent restart crash loop with the
radio stuck down. The only way to disable filtering is to delete the option
entirely. Also guards against pushing an empty whitelist (macfilter='allow'
with zero MACs blocks every client outright) by treating it as equivalent
to "off".
Adds MAC ACL (whitelist/blacklist) read+write support for wireless SSIDs,
mirroring push_radio_channel's UCI write style. Backs the new Global MAC
ACL feature in netOrk.
'wifi reload' only reapplies the running config without physically
changing the channel on many ath9k/ath10k/mt76 hardware+driver
combinations. A full 'wifi' (down + up) cycle is required for channel
changes to take effect.
Also updated the test assertion accordingly.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>