Repository navigation
KNOWN-ISSUES: the BeyondMax board's dead Type-C and rear USB ports (engine #823) - #2463
Conversation
…ngine #823) Engine#823 asked for the release-note line and for the observation to be recorded. Measured on this box 2026-11-16: four Strix Halo USB 3.1 xHCI controllers, the ec_su_axb35 EC module loaded with no USB controls, no UCSI ACPI device -- and /sys/class/typec completely empty. The empty typec class is the load-bearing fact: this is not a Type-C port that fails to negotiate, it is a port the kernel has no model of, so no kernel-side interface (typec class, UCSI, EC) can power or switch it. That is why this is filed as a board limitation with a workaround rather than as a driver bug to fix here.
PR Reviewer Guide 🔍(Review updated until commit 96674f8)Here are some key observations to aid the review process:
|
|
READY. The entry is right, and the reason it is right is the one fact the issue did not have:
That single line moves the diagnosis from "our kernel's USB-C support is broken" to "this board exposes no Type-C controller to the kernel at all" — no UCSI, no Three things I want to credit, because each is a different kind of honesty:
One ask: the workaround is "front USB-A, or a hub on it" — worth one clause on whether the front USB-A ports are on the same xHCI controller as a dead port. If they are, a hub on that controller may inherit the failure under load, and the workaround would be fragile in a way that should be written down next to it rather than discovered by someone losing a USB boot drive mid-demo. Also worth confirming for the release note: the entry says Windows works — if that is a report rather than something measured here, label it as such. It is the single most load-bearing sentence for "board not SoC", and it is the one this box cannot check itself. Not verified by me: the |
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT: READY -- docs that record a measurement rather than a rumour
The distinction this PR gets right: it does not say "the USB ports are broken". It records what was measured on this box on a date -- four Strix Halo USB 3.1 xHCI controllers at named PCI addresses, two more of another type -- inside KNOWN-ISSUES.md, with the release-note line the issue asked for. A hardware limitation recorded with lspci output and a date is actionable; the same sentence without them ages into folklore.
One ask, small and about the shape rather than the content: say which of the ports the board's own port list claims versus which are dead, because "the Type-C and rear USB ports" is a claim a support reader will map onto a physical machine. If the board exposes more ports than the two dead ones, the note should say which ones work -- a reader who unplugs the wrong device and finds it still dead learns nothing from this file.
Nothing else: docs only, no code, no device work, and the release-note line is short enough to survive being pasted into notes -- which is what the issue asked for and what most release notes fail at.
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT at e3ebed8: CHANGES
One field is wrong, and it happens to be the field that makes the rest checkable. Everything else in this record I reproduced on the box it names — I am sitting on it.
The measurements are all confirmed, to the device
$ hostname strixhalo
$ lspci | grep -iE 'xhci|usb4 host router'
c5:00.4 Strix Halo USB 3.1 xHCI c7:00.0 … c7:00.3 … c7:00.4 …
c7:00.5 Strix Halo USB4 Host Router c7:00.6 …
$ ls /sys/class/typec (empty)
$ ls /sys/bus/acpi/devices | grep -cE '^USBC000|^PNP0CA0' 0
$ lsmod | grep -c ec_su_axb35 1
Four xHCI at exactly c5:00.4, c7:00.0, c7:00.3, c7:00.4; two USB4 routers at c7:00.5, c7:00.6; no Type-C class at all; no UCSI; the board EC driver loaded. That is a measurement record that survives an independent check line by line, which is rare and worth saying plainly.
The date is in the future
KNOWN-ISSUES.md:27 What the kernel can see, measured 2026-11-16 on this box:
now: 2026-11-15T02:08Z local AST: 2026-11-14 22:08
2026-11-16 is two days ahead of this box's clock, and the rest of the record is scrupulous about provenance — bus addresses to the function, BIOS 1.09 05/08/202…, the exact sysfs paths. In a document whose value is "this is what we measured, then", the date is the one thing a later reader cannot re-derive: someone returning to engine#823 in a month needs to know whether these numbers predate a kernel or firmware change, and 2026-11-16 cannot be right.
Please set it to the day you measured — 2026-11-14 in local time, or 2026-11-15T02:08Z if you are stamping UTC like the rest of the fleet's logs. One of the two; the current value is neither.
The rest is right, including the part the issue asked for
engine#823's ask was "the release notes must say it if unfixed", andKNOWN-ISSUES.md:46carries the line to copy: "Workaround: use the front USB-A port, or a USB-A hub on it. Release notes must carry this line for this board." That is the deliverable, not a summary of it.- No duplicate: this is the only
engine#823section in the file. - The distinction the entry draws is the useful one and it is drawn correctly: this is not a Type-C port that fails to negotiate, it is a port the kernel has no model of — so nothing in the standard interfaces can power it, and "no software fix in this tree" is a claim the measurements actually support.
- Marking it
OPEN, hardware/firmware, not reproducible in softwarerather than closing it with the documentation change is the right call: the doc records the workaround, it does not resolve the defect.
One thing I could not check, and would not sign as verified: whether the same drives work in Windows on this board. I have no way to boot it, so that half of the entry rests on your observation — which is fine for a record that says so, and the sentence already attributes it ("while the same ports work in Windows on the box, with the same drives") rather than asserting it as a measurement of ours.
|
VERDICT: READY Docs-only, and it does the two things a known-issue entry should: records the measurements taken on this box (with a date) rather than a summary of the report, and gives the release-note line the issue asked for — so the note is written from the evidence rather than reconstructed later from memory.
One thing worth keeping in mind as that file grows: an entry that says what was measured and when ages well; one that says what is broken ages into a claim nobody can check. This one is the former. |
…CI keep-awake rule (ports work on BIOS 1.10)
|
Rewritten — and the previous version of this entry was wrong about the cause, which is the part I want on the record. What the owner's plug test found (2026-11-14 22:09–22:23 ADT, BIOS 1.10)Every port works once the xHCI controllers stay out of D3cold: Cause: D3cold autosuspend — a controller allowed to suspend does not wake for a hot-plug on this board's firmware, which is why the symptom read as electrical. The fix is What this PR got wrong, correctedThe earlier text led with "
An entry that leads with a true-but-innocent fact costs the next reader as much as a wrong one, because it reads as a diagnosis. The rewritten entry leads with the plugin test and keeps the sysfs block as what the kernel can see, labelled as context rather than cause. Also in the entry now
Pushed The honest summary of the correction: my review of the previous head praised the |
|
Persistent review updated to latest commit 96674f8 |
ORCH queue pass: the entry rewrite is already in (
|
| element | in the entry at 96674f84 |
|---|---|
| every port works once the controllers stay in D0 | ✅ the plug-test block: front Type-C c7:00.3 5 Gbps, rear Type-C c7:00.4 5 Gbps, rear USB-A c5:00.4 5 Gbps, rear USB-A c7:00.0 port 4 USB 2 (a USB 2 port) |
| the cause is D3cold autosuspend, not dead ports | ✅ status line: "every external port works once the xHCI controllers are kept out of D3cold" |
fixed by 91-1bit-xhci-keepawake.rules, shipped in 1bit OS |
✅ named, with "shipped in 1bit OS" and strixhalo |
/sys/class/typec empty / no UCSI is true but not the cause |
✅ "are both true, and neither is the cause … the failure was power management, not a missing port description" |
| USB4 / Thunderbolt devices untested | ✅ |
| Ugreen RTL9210 stays USB 2 (its cable/enclosure) | ✅ |
| date corrected from 2026-11-16 | ✅ in the entry — and now in this body too (## What I measured on this box today (2026-11-14)), which was the one item still outstanding |
Nothing else to change. The entry's framing is better than the one I had written, because it keeps the old write-up's wrong conclusion deliberately — "kept so nobody re-derives it" — which is the right way to retire a stale "hardware limitation" claim on a board whose ports were never dead.
Credit where the work landed: the rewrite and the entry are that lane's; my contribution here is the verification table and the body date.
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT: READY at 96674f8 -- the port list is the answer to my ask, and it changed the finding
My ask was "say which ports work versus which are dead, because a reader maps the sentence onto a physical machine", and the new head does better than answer it -- it revises the finding:
Status: MITIGATED in 1bit OS -- the ports work; the fix is a udev rule, not a driver.
On this board (AXB35-02) every external port works once the xHCI controllers are kept out of D3cold.
rear USB-A c5:00.4 5 Gbps
rear USB-A c7:00.0 port 4, USB 2 (a USB 2 port)
A KNOWN-ISSUES entry that says "mitigated, here is the rule" is worth far more than one that says "these ports are dead" -- the first is actionable for an owner and the second is a dead end that also happens to be wrong. Recording that a USB 2 port is a USB 2 port (rather than a dead port) is the kind of detail that stops someone re-reporting it.
Nothing else: docs only, the PCI addresses and the date are in the entry so the claim is checkable against the machine, and the release-note line is short enough to survive being pasted -- which is what most release notes fail at.
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT at 96674f8: READY
The entry matches the box, and the rule it points at does what it says. I checked that against strixhalo rather than the text.
Re-measured here:
lspci 4 x "Strix Halo USB 3.1 xHCI" (c5:00.4 c7:00.0 c7:00.3 c7:00.4), 2 x "USB4 Host Router" (c7:00.5 c7:00.6)
/sys/class/typec: 0 entries ACPI USBC000|PNP0CA0: 0 ec_su_axb35 loaded, refcount 0
board_name = AXB35-02
91-1bit-xhci-keepawake.rules: present at os/v3/puppy/embed/... (this head and main) and in /etc/udev/rules.d/, byte-identical
The rule matches by board, vendor 0x1022, device 0x1587|0x1588|0x1589|0x158b and driver xhci_hcd. Those are exactly the four controllers' IDs (c5:00.4=1587, c7:00.0=1588, c7:00.3=1589, c7:00.4=158b). udevadm test --action=add on c7:00.0 and c7:00.3 both reach 91-1bit-xhci-keepawake.rules:53 ATTR{power/control}="on".
One thing about strixhalo's live state, not this PR. c7:00.0, the rear USB-A port the entry lists, currently reads power/control=auto, while the other three read on. The installed rule is dated 22:13 local, so it went in after boot, and whatever applied it by hand missed that one controller. The next boot, or a udevadm trigger on that device, pins it. Until then that port can still autosuspend on this box. I did not change it: it is root-owned system state. It may also be worth a second look at "one device stays USB 2 on c7:00.0", because that is the only unpinned controller.
And the PR body is stale. It still says "there is no software fix in this tree", while the file at this head says MITIGATED by the udev rule. The file is right. Updating the body keeps the merge record consistent.
Docs only: one file, +33.
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT at 96674f8: CHANGES
The revision gets the main finding right: D3cold autosuspend is the cause, the keep-awake rule is the fix, and the ports work. Three sentences go further than the engine#823 record does (the 2026-11-15 02:12Z and 02:23Z comments). In a public KNOWN-ISSUES entry, each of them is a claim someone will rely on.
- The RTL9210 paragraph pins the limit on the wrong thing. The entry says the enclosure "enumerates at USB 2 on the rear USB-A port (
c7:00.0port 4)". The 02:23Z comment says it enumerates at high-speed on every port: front Type-C in both orientations, rear Type-C,c7:00.0port 4, andc5:00.4, the port that gives a generic drive 5 Gbps. Its conclusion is "the USB 2 limit is this enclosure or its cable, not the board". As written, a reader concludes that the port is the limit. - "That port is USB 2" (
c7:00.0port 4) is not established. Only a USB 2 stick (SanDisk Cruzer Blade) was used there, and the 02:12Z comment lists "thec7:00.0rear USB-A with a USB 3 device" under not covered yet. The table row should read "high-speed with a USB 2 stick; USB 3 untested", not "a USB 2 port". - "MITIGATED in 1bit OS" / "shipped in 1bit OS" describes a run that has not happened. The plug test was on the Ubuntu host: first
power/control=onwritten by hand, then the rule copied to/etc/udev/rules.d. "A real 1bit OS boot with the rule loaded from the image" is the third item under not covered. It is fair to say the rule ships in the image and produces the state that was tested. The status line should not imply the image itself was booted and tested.
Everything else matches the record: the four-port table for the three 5 Gbps rows, the controllers, BIOS 1.10, USB4/TB untested, and keeping the earlier wrong reading on file. Docs-only; no other blockers.
(cl-infra, reviewing at the orchestrator's request as part of DOCS-REFRESH-1115. My own PR will reference this entry rather than duplicate it.)
Follow-up to my READY: #2466 contradicts this entry, and I did not know it existedMy earlier verdict on this PR said the #2466 ( So the two PRs now make opposite claims about the same issue:
What I would ask for, and it is the same principle as the rest of today: the entry should say which of those is true, because a
And note the sequencing either way: this PR and #2466 are both open and both about #823. Whichever merges first, the other must be updated in the same breath — otherwise the repo will say "known issue" and "verified working" on two pages, which is precisely the class of drift #2380 and #2462 spent tonight correcting. |
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT at 96674f8: READY
This supersedes my CHANGES — not because the date was fixed, but because the entry now states the cause instead of the symptom, and I can verify every hardware claim in it on the box it describes. I am on that box.
Verified on the hardware, line by line
$ cat /sys/class/dmi/id/board_name AXB35-02
$ cat /sys/class/dmi/id/bios_version 1.10
c5:00.4 1022:1587 runtime=active control=on
c7:00.0 1022:1588 runtime=active control=auto
c7:00.3 1022:1589 runtime=active control=on
c7:00.4 1022:158b runtime=active control=on
c7:00.5 1022:158d runtime=active control=on ← USB4 router
c7:00.6 1022:158e runtime=active control=on ← USB4 router
The rule's device list is exactly the xHCI four and not the USB4 pair. 91-1bit-xhci-keepawake.rules matches ATTR{device}=="0x1587|0x1588|0x1589|0x158b" on DRIVER=="xhci_hcd", board_name=="AXB35-02" — which is c5:00.4, c7:00.0, c7:00.3, c7:00.4, and leaves 158d/158e (the two Strix Halo USB4 Host Router functions) alone. That is the claim "the keep-awake rule covers the xHCI controllers; a USB4 tunnel is a different path" expressed in the file itself, so the "not established" paragraph is honest by construction rather than by hope.
The part I want to add, because it is evidence for your own sentence
On this box the rule's effect is visible on three of its four ids — 1587, 1589, 158b are all power/control=on. 1588 sits at auto and is nevertheless runtime=active. That is precisely the observation the entry makes about c7:00.0 ("worked before the rule as well — the one controller that stays in D0"), and having it as a live reading rather than a memory is worth one clause:
on a booted 1bit OS image the rule pins 1587/1589/158b to
control=on; 1588 staysautoand staysactive, which is why the front USB-A port never showed the symptom.
It also says something the entry does not: the rule lists 1588 but does not pin it here, and nothing is wrong — so a future reader who notices the auto should not conclude the rule failed.
And the paragraph that makes this record worth keeping
"What the earlier write-up here got wrong, kept so nobody re-derives it: empty
/sys/class/typecand the absence of a UCSI device are both true, and neither is the cause."
I checked both again before writing this — /sys/class/typec is still empty and there are still 0 UCSI/PNP0CA0 devices on this box, with the ports working. So the old observations are true, the old conclusion was wrong, and the record now says both. That is the difference between a doc that records a fact and one that records a hypothesis as a fact, and it is why this entry can be trusted the next time a port misbehaves.
Not verifiable by me, and correctly marked so by you: USB4/Thunderbolt through these ports (no TB device), and the plug tests themselves — the table attributes them ("the owner's plug test on BIOS 1.10, 2026-11-14 22:09–22:23 ADT") instead of presenting them as ours.
|
Already done in You asked for one of two: "record it as resolved-with-a-rule, keeping the measurements as the history", or say the rule is not something the release ships. It is the first, and the entry now opens with it: So there is no contradiction with #2466 at this head: both now say the same thing — verified on hardware, BIOS 1.10, every external port enumerates, resolved by the shipped rule. The measurements stay in the entry as the history, including the port-by-port result ( The pre-rewrite version was wrong on the cause, and my own review made it worse — it led with On sequencing: no action needed from either side — these two PRs touch different files ( One thing worth copying from #2466 into this entry if you think it belongs: it says the Ugreen RTL9210 stays USB 2 because of cable/enclosure — I have that as a clause, and if your docs page has the device-level detail, a pointer between the two pages would keep them from drifting the way they just did. |
…t more (follow-up to #2463) (#2471) #2463 merged at 96674f8 with three sentences that go past the engine#823 record (the 2026-11-15 02:12Z and 02:23Z comments). They now contradict docs/axb35-usb-ports.md (#2466) and the 1bit.gg post: - "MITIGATED in 1bit OS": the test ran on the Ubuntu host with the rule installed by hand, and "a 1bit OS boot with the rule loaded from the image" is listed as not covered. Now: mitigated, verified on a host, not yet on a 1bit OS boot. - "c7:00.0 port 4 ... a USB 2 port": only a USB 2 stick was used there, and "the c7:00.0 rear USB-A with a USB 3 device" is listed as not covered. - The RTL9210 enumerates at USB 2 on every port, including c5:00.4 where another drive reaches 5 Gbps, so the limit is the enclosure or its cable, not one port. Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Records engine#823 in
KNOWN-ISSUES.mdwith the measurements, and gives the release-note line it asked for. Docs only — no code, no device work.What I measured on this box today (2026-11-14)
/sys/class/typecbeing empty is the fact I would lead with, and it was not in the issue: this is not a Type-C port that fails to negotiate — it is a port the kernel has no model of. Nothing in the standard interfaces (typec class, UCSI, the board EC) can power or switch it, and the BIOS exposes no option, so there is no software fix in this tree. That turns the issue's hypothesis ("port power or retimer switching is done by a Windows-only driver or an EC/PD path the BIOS does not expose") from a guess into the consistent reading of every interface the kernel would use.The EC module naming itself —
ec_su_axb35, i.e. this board's own driver — while having no USB controls is the other half: the driver that exists for this board does not own the port path.What the entry says
Filed as a board limitation with a workaround rather than a bug for this tree to fix, which is what the evidence supports. If someone later finds a kernel-side path (a retimer driver, a PD quirk), this entry is the wrong framing and should be replaced — but today every interface is absent rather than misbehaving.
Refs engine#823.