Repository navigation
Conversation
linux-firmware ships brcmfmac43602-pcie.bin but not the board NVRAM, and without it the 43602 firmware's country detection breaks: only 2.4GHz networks appear, signal collapses, and the WPA 4-way handshake times out — which NetworkManager misreports as a wrong password (kernel bug 193121). Vendor the known-good NVRAM (Apple Boot Camp driver lineage, with ccode=0/regrev=1 to bypass the broken detection) and install it for Apple machines carrying a BCM43602 (14e4:43ba/43bb/43bc), at install time and via migration for existing installs. The live interface MAC is substituted in; the line is stripped when no MAC is discoverable. An existing NVRAM is never overwritten, and the migration marks reboot-required instead of reloading brcmfmac out from under the connection carrying the update. Verified on MacBookPro14,3: 5GHz networks appear, stronger signal and handshake completes instantly. Complements the firmware-supplicant quirk from omacom#6652, which was active on the affected machine but not sufficient on its own.
|
Can confirm this fixes it on a MacBookPro14,3 (15" 2017 Touch Bar, BCM43602 14e4:43ba, Omarchy 4.0.1, kernel 7.1.9-arch1-2). Out of the box I had exactly the symptoms described here: no Installed the NVRAM file from this PR into Two small notes:
|
|
Reviewed on a disposable worker VM, with an independent second opinion from Codex at xhigh reasoning. Three things still need a decision from you or from the maintainer, and the confirmation report above — welcome as it is — does not settle any of them. Tests
Still open1. The destination sits inside linux-firmware's own directory. 2. One board's calibration is installed on boards that are not that board. The vendored file declares 3. The MAC baked into the NVRAM may be a temporary one. On the confirmation reportThank you for it — before-and-after on real hardware is the part nobody else can supply, and the Second opinionCodex at xhigh reasoning agreed that all three findings remain open and that a single field report answers none of them; its independence is not currently guaranteed, since it can read this session's own working files, so read that as agreement rather than a separate discovery. It did contribute two things this review had not reasoned out: the arithmetic behind the antenna point — that Where this standsWaiting on both of us. The maintainer has a call no amount of review settles: whether Omarchy vendors a Broadcom-proprietary blob at all. You flagged that yourself in the description and were right to. Waiting on you are findings 1 and 2 — the destination path, and either narrowing the gate to PR #7671 is an open, competing implementation of the same fix, built on a different NVRAM dump. Choosing between the two, or asking for one merged approach, is the maintainer's call; there is nothing here for you to act on. |
|
A/B test on MacBookPro11,5 (BCM43602 14e4:43ba, same chip, different board): baseline has no NVRAM (no clm_blob in journal) yet 5GHz already works here (5180MHz, -30dBm, 400/400Mbit/s, 14x5GHz nets visible), so this tests cross-board safety, not efficacy. Ran the actual leaf: installs a 7131-byte file with macaddr correctly substituted for this NIC's real (non-random) address. Reload recipe that works: nmcli disconnect + ip link down + modprobe -r brcmfmac_wcc, then brcmfmac (wcc must go first; plain modprobe -r brcmfmac fails 'in use'). Note reconnect needs explicit nmcli dev connect -- disconnect blocks autoconnect by design. With NVRAM loaded: identical link (5180, 400/400), scans no worse (20x5GHz, within rescan variance), clm_blob still absent (expected -- NVRAM is not CLM). Reverted afterwards; baseline restored exactly. No degradation on this 3x3 board -- with one caveat: at VHT 400Mbit/s the HT MCS-cap concern from review is not exercised, so this does not settle that half of the cross-board question, only that nothing regresses at these rates. One correction to the thread discussion: /sys/class/net/wlp4s0/perm_address does not exist at all on brcmfmac here (not all-zero, absent), so a perm_address-first fallback would need to handle ENOENT, not just zero content -- and would land back on address anyway, which is what the leaf already reads. |
|
Addendum -- HT/VHT capability check for the cross-board question, same MacBookPro11,5. With the NVRAM loaded and the driver reloaded, iw phy advertises byte-identical capabilities to baseline: HT MCS 0-23 and VHT MCS 0-9 on all 3 streams, link still 400/400. So the file's rxchain/txchain=3 mask does not shrink the advertised stream support on this 3x3 board -- the predicted MCS cap does not materialize at the advertisement level. Caveat: advertisement is not runtime enforcement; forcing HT MCS 16-23 would be the final proof, but that needs a 3SS AP and band manipulation, and with identical advertisements plus identical link rates there is no signal pointing at hidden harm. |
|
Another MacBookPro14,3 data point, and one result that bears on the cross-board question @rand0mdud3 raised. First, confirming @bernardcosta: this PR's file does fix 5 GHz here. MacBookPro14,3 (15-inch 2017, T1), Omarchy 4.0.3, kernel 7.2.3-arch1-3, BCM43602 On this board the file does halve advertised spatial streams. @rand0mdud3's MacBookPro11,5 result was that
Raw
|
|
Measured data from a This NVRAM file is from subsystem Measured in both directions, after a full reboot each time (
The falsifier throughout was the AP-side registration table, not the client's own status.
The file declares Two further measured facts:
I am not proposing a fix and not asserting a cause beyond the table above. Whether correct srom11 data for |
|
@gunars — a third board, and on this one MacBookPro14,3, Apple subsystem
So the three boards in this thread now read: the file originates from On your instrument warning. Agreed that nmcli is not evidence, and I did not rely on it — but my falsifier is weaker than yours and I want that on the record. I do not control this AP, so I have no registration table. What I have is the supplicant's own DFS: I can neither confirm nor refute. There is no AP on a DFS channel here to test against. For what it is worth, on this board those channels are present in the phy list rather than missing:
One methodology note that may save you reboots. You avoided module reloads because The interface returns and reassociates in 7 seconds, with nothing in dmesg beyond the expected |
|
Fourth board for the matrix: MacBookPro11,5, Apple subsystem 106b:0152 (BCM43602 14e4:43ba). Both bands work with and without the file here (associated 5180 MHz, 12+ 5GHz nets visible stock), and with the file loaded the advertised caps stayed byte-identical (HT 0-23, VHT 3x0-9) with unchanged link rates -- so 0152 neither needs the file nor is harmed by it at VHT rates. |
|
@bjarneo @csvenke Following the duplicate closure of #7671, could we agree on the consolidation requirements before treating the two PRs as interchangeable? At current heads
Consolidation is reasonable, but please retain the required-key/unique-MAC behavior, board-scoped calibration selection, existing-file protections and failure-path tests. The missing-MAC strip case is not just duplicate implementation style. Calibration redistribution/provenance remains a separate maintainer decision. Our draft #10141 contains the shared #7671 work plus its 13,3 extension. I have marked that dependency as unresolved after the closure; I am not rebasing it onto this broader installer as-is or changing installed firmware. Can maintainers confirm these differences should be reconciled here, or reconsider keeping #7671 open until that is done? |
|
Hey, sorry about this being part of the automated flow. Let us open the PRs again! Still tweaking the duplication runner so I appreciate the feedback. |
|
Re-reviewed at Recommendation: #7671 is the better of the two to bring to the maintainer, and this one should not also be marked ready. The reasons are the field data posted on this thread since the last review:
Two findings from the last review still stand at this head: the destination Codex, briefed on both pull requests without this review's answer, also chose #7671, for the same reasons: the board-scoped gate, the override directory and the MAC handling. Its independence is not guaranteed, since it can read this session's working files. It also named what remains open on #7671 — a live-address fallback when the wiphy address is unavailable, and an override that a later packaged board file cannot replace without a cleanup migration — which belongs on that pull request's own review. Nothing was pushed to this branch. This waits on the maintainer, who has two decisions no review settles: whether Omarchy ships a Broadcom-proprietary calibration file at all, which you flagged yourself in the description,, and which of #7487 and #7671 to carry. There is nothing for you to act on until they decide. |
|
Thank you for this, @csvenke. #7671 fixes BCM43602 Wi-Fi on these Macs without the board-dependent regressions field reports showed for this approach, so the work continues there. Closed at the maintainer's request. The review above has the details; it was done by Claude Opus 5.5 with Codex Medium as a second opinion, whose agreement is not independent. |
Issue
14e4:43ba/43bb/43bc) ship withbrcmfmac43602-pcie.binbut no board NVRAM (brcmfmac43602-pcie.txt) —linux-firmwarehas never carried it.deauth reason=15).wpa_supplicantmisreports that timeout as "pre-shared key may be incorrect" and NetworkManager as "psk mismatch" — so affected users chase their passphrase instead of the real cause.feature_disable=0x82000) fixes the WPA-offload failure against WPA2/WPA3 transition APs; it was active on the machine below the whole time and was not sufficient. Both defects surface as "wrong password", which is why this one hides behind it.Fix
Vendor the known-good NVRAM (Apple Boot Camp driver lineage, from bug 193121) and install it for Apple machines carrying a BCM43602:
default/firmware/apple/brcmfmac43602-pcie.txt— the NVRAM withccode=0/regrev=1, which bypasses the broken country detection, plus a provenance header.install/hardware/apple/fix-brcmfmac-nvram.sh— install-time leaf, gated on Apple DMI vendor +14e4:(43ba|43bb|43bc)(narrower than the supplicant quirk's list: this NVRAM is 43602-family only, and BCM4360 Macs run the out-of-treewldriver and are excluded for free). Substitutes the live interface MAC into themacaddr=line; strips it when no MAC is discoverable (the OTP fallback is proven). Never overwrites an existing file.migrations/<epoch>.sh— the same repair for existing installs: idempotent, env-overridable paths, never clobbers, setsreboot-requiredonly when it actually installed — it does not reloadbrcmfmac, which would drop the Wi-Fi connection carrying the update.test/shell.d/brcmfmac-nvram-test.sh— 13 assertions covering gates, exclusions, MAC substitution/strip, idempotency, no-clobber, and the escalation contract.Evidence
MacBookPro14,3 (2017, BCM43602), before → after installing the NVRAM and reloading
brcmfmac:reason=15Independent confirmations on 13,2 and 14,3: mbp-2016-linux#172, mbp-2016-linux#208.
Provenance & licensing
The NVRAM is Broadcom-proprietary, extracted from Apple's Windows driver package — which is exactly why
linux-firmwarecan't ship it and bug 193121 has been open for years. Vendoring it here is the same trade-off, made at the distro layer; if you'd rather not carry the file in-repo, the alternatives are fetch-at-install (the canonical bugzilla source is now behind an anti-bot wall, and the gists are a supply-chain risk) or detect-and-instruct (weak). Flagging it explicitly so the call is yours.