shaper: detect a cake_mq root qdisc instead of assuming cake (#353)

[ooonea]

Closes the gap reported in #353. @lynxthecat asked me in that thread whether I saw a good way to address it; this is the shape I proposed there, now written out.

The failure mode

set_shaper_rate() hardcodes cake in its tc qdisc change root call. If the root qdisc is cake_mq, the kernel rejects the change with Invalid qdisc name — but the call sends stderr to /dev/null and its exit status is never checked, so cake-autorate keeps emitting SHAPER log lines while shaping nothing. It fails silently, which is worse than failing loudly: from the logs everything looks fine.

This is reachable today, not hypothetical: sqm-scripts picks cake_mq whenever USE_MQ=1 and the device has more than one queue (functions.sh:966-968), and OpenWrt carries the kernel side in backport-6.12 (700-02..700-07) plus the iproute2 side in 001-1-tc-cake-add-cake_mq-support. Configuring the cake_mq instance directly and letting the config reach the sub-qdiscs is the designed behaviour — 700-04 is literally "Share config across cake_mq sub-qdiscs".

The change

~15 lines. get_root_qdisc_kind() reads tc qdisc show root dev <if> once per direction at startup and stores the kind in a new qdisc_kind[dl|ul]; set_shaper_rate() uses that instead of the literal. Anything unrecognised — htb, noqueue, empty output, a leaf line — falls back to cake, so a conventional setup is byte-identical to today. No per-iteration cost, no new config knob.

What I tested, and what I could not

Tested:

• 6/6 detection cases against real tc output strings: cake_mq root → cake_mq; cake root, htb root, noqueue, a parent 1:1 leaf line, and empty output → cake
• against a real tc qdisc show root, in a network namespace with a dummy interface: a genuine cake root is detected as cake; fq_codel and htb roots fall back to cake
• bash -n clean, shellcheck -S warning reports nothing on the new lines

Not tested, and this is the part that needs someone with the hardware: I have no cake_mq-capable box, so I have not observed an actual tc qdisc change root dev <if> cake_mq bandwidth <X> taking effect end to end. The detection is verified; that the subsequent rate change propagates to the sub-qdiscs rests on reading 700-04/701-02 rather than on my having seen it work.

Marking this a draft for exactly that reason. @Snuupy, @CHC383 — you raised #353 and suggested the grep -q cake_mq approach; if either of you can run this on a multi-queue setup with USE_MQ=1 and confirm the shaper actually tracks, I will flip it out of draft. If it turns out change on a cake_mq root needs different handling, better to find that now than after merge.

[lynxthecat]

This looks pretty good to me. Pedantic me wondered about qdisc_type rather than qdisc_kind. Any thoughts on which is a better fit?

[ooonea]

I went with kind because it is iproute2's own word for this rather than a preference of mine. tc qdisc help ends its usage with [ [ QDISC_KIND ] [ help | OPTIONS ] ] and spells out QDISC_KIND := { [p|b]fifo | tbf | prio | red | etc. }; the parser function behind it is get_qdisc_kind(), and on the netlink side the attribute carrying the string is TCA_KIND. tc qdisc show then prints it immediately after the word qdisc, which is exactly what the regex here anchors on.

There is a local reason too: type is already spoken for twice in cake-autorate.sh — the message class in log_msg (local type=${1}) and the shell variable type in the config validation (typeof, user_type, valid_type). qdisc_kind keeps the three apart at a glance. That said it is your codebase, so if you prefer qdisc_type say the word and I will rename it.

Still keeping this in draft until someone with a multi-queue box can confirm the tc qdisc change actually lands on a cake_mq root — the detection is covered by fixtures and a real tc qdisc show root in a netns here, but I have no hardware to prove the change propagates.

[lynxthecat]

Perhaps I can test on my RT3200 actually?

[ooonea]

That would be great! The detection half is already covered (fixtures plus a live tc qdisc show parse), but whether the change calls behave on a real cake_mq root is exactly the part I could not verify without multi-queue hardware. If it works on your RT3200, I'll mark this ready for review.