c630/docs/firmware.md

# Firmware

The C630's firmware splits into two groups, and the split is about licensing,
not difficulty.

## What ships in the image

Qualcomm permits redistribution of these, so Fedora packages them and the build
installs them:

| Package | Files | Enables |
|---|---|---|
| `atheros-firmware` | `ath10k/WCN3990/hw1.0/{firmware-5.bin,board-2.bin,wlanmdsp.mbn}` | WiFi and Bluetooth |
| `qcom-firmware` | `qcom/a630_gmu.bin`, `qcom/a630_sqe.fw` | The generic half of the Adreno 630 — necessary but not sufficient |

Networking therefore works on a freshly written image. Graphics does not: see
below.

## What you have to supply

These are signed against per-model keys and Fedora does not ship them, so a
stock install has none of them. Nothing about them is unredistributable in
principle — the equivalent files for the
ThinkPad X13s *are* in linux-firmware — it is that nobody has obtained
permission for this model. See [If you have no
backup](#if-you-have-no-backup).

The list is not a guess: it is every `firmware-name` property in the mainline
device tree, `arch/arm64/boot/dts/qcom/sdm850-lenovo-yoga-c630.dts`:

| File | Device tree node | Lost without it |
|---|---|---|
| `qcdxkmsuc850.mbn` | `gpu_zap_shader` | **Accelerated graphics** |
| `qcadsp850.mbn` | `adsp_pas` | Audio — speakers, microphone |
| `qcslpi850.mbn` | `slpi_pas` | Sensor hub — lid switch, accelerometer, auto-rotate |
| `qccdsp850.mbn` | `cdsp_pas` | Compute DSP offload |
| `qcvss850.mbn` | `venus` | Hardware video decode |
| `qcdsp1v2850.mbn` | `mss_pil` | Modem, stage 1 |
| `qcdsp2850.mbn` | `mss_pil` | Modem, stage 2 |
| `ipa_fws.elf` | `ipa` | IPA datapath, needed by the modem |

They belong under `/lib/firmware/updates/qcom/sdm850/LENOVO/81JL/`.

The GPU one is worth calling out, because it is easy to assume otherwise:
Fedora does ship an `sdm845/a630_zap.mbn`, but that is the generic
Snapdragon 845 zap shader and **not** what this device asks for. The C630's
device tree names `qcom/sdm850/LENOVO/81JL/qcdxkmsuc850.mbn` specifically, so
until you supply it you get an unaccelerated display.

Each file is independent, so a partial recovery is still worth having —
`qcdxkmsuc850.mbn` on its own buys you working graphics.

### Confirmed on the hardware

This is not inferred from the device tree. On a machine with the generic
Adreno firmware present and the model-signed blob absent, the GPU gets two
files in and then stops:

```
[drm:adreno_request_fw]   loaded qcom/a630_sqe.fw from new location
[drm:adreno_request_fw]   loaded qcom/a630_gmu.bin from new location
[drm:zap_shader_load_mdt] *ERROR* Unable to load qcom/sdm850/LENOVO/81JL/qcdxkmsuc850.mbn
[drm:adreno_load_gpu]     *ERROR* gpu hw init failed: -2
```

So the generic pieces are necessary and not sufficient, and the zap shader is a
hard requirement rather than an optimisation. Two details follow from that
trace:

The driver never falls back to `sdm845/a630_zap.mbn`, even though the file is
installed — the device tree's `firmware-name` pins the model-specific path. Nor
would falling back help: the zap shader is signed against the device's own
secure-boot chain, and the Snapdragon 845 development board's copy will not
validate here.

The display keeps working throughout. `msm_dpu` binds the panel and the console
lands on `msmdrmfb` regardless, so a failed GPU costs you acceleration, not a
picture. A desktop still runs on llvmpipe — though `/dev/dri/renderD128` is
present even when `hw init` has failed, so mesa may try freedreno, fail, and
take the display manager down with it. `LIBGL_ALWAYS_SOFTWARE=1` in
`/etc/environment` avoids that.

### These have to be in the initramfs

`msm_dpu` probes about six seconds in, while the initramfs is still the root
filesystem, and does not retry. Firmware that only exists on the real root is
therefore invisible to it:

```
msm_dpu ae01000.display-controller: Direct firmware load for qcom/a630_sqe.fw failed with error -2
```

`overlay/etc/dracut.conf.d/10-c630.conf` puts the Adreno files in the initramfs
for this reason, and `c630-firmware` adds `qcdxkmsuc850.mbn` once it exists.

Only the GPU needs this. The ADSP, CDSP, SLPI, venus and modem all probe around
forty seconds in, well after the real root is mounted, and load from it
perfectly happily. That distinction is worth keeping: `qcdsp2850.mbn` alone is
60 MB, so sweeping the whole firmware directory into the initramfs would inflate
it roughly twentyfold for no benefit.

## Finding them on a backup

If Windows is long gone from the machine but you kept a copy of the files, one
search catches all eight, since seven share a suffix:

```sh
find /mnt/backup -type f \
    \( -iname 'qc*850.mbn' -o -iname 'ipa_fws.elf' -o -iname 'bdwlan*.bin' \) \
    -printf '%10s  %p\n'
```

They originally lived in `C:\Windows\System32\DriverStore\FileRepository\`, in
directories named `qcdx850.inf_arm64_*` and `qcadsp*.inf_arm64_*`. If the loose
files turn up nothing, look for those directories wholesale:

```sh
find /mnt/backup -type d -iname 'qc*.inf_arm64_*'
```

Sizes range from 14 KB (`qcdxkmsuc850.mbn`) to 60 MB (`qcdsp2850.mbn`).
Despite the `.mbn` extension they are ELF images, and `file` should say so:

```
qcdxkmsuc850.mbn: ELF 32-bit LSB executable, QUALCOMM DSP6, version 1 (SYSV)
```

Anything reporting as text or HTML is a download that went wrong.

Lenovo does not appear to publish a driver package for the 81JL containing
these. If the backups turn up nothing, see [If you have no
backup](#if-you-have-no-backup) — the files are not tied to your individual
machine, so yours is not the only possible source.

## If you have no backup

Losing the Windows partition is not the end of it, because **these blobs are
signed per model, not per machine**. A copy from any other Yoga C630 works on
yours.

The evidence is in linux-firmware itself, which already ships the exact
analogous set for the ThinkPad X13s — one copy serving every unit of that
model:

```
qcom/sc8280xp/LENOVO/21BX/qcdxkmsuc8280.mbn     zap shader
qcom/sc8280xp/LENOVO/21BX/qcadsp8280.mbn        audio
qcom/sc8280xp/LENOVO/21BX/qccdsp8280.mbn        compute DSP
qcom/sc8280xp/LENOVO/21BX/qcslpi8280.mbn        sensors
qcom/sc8280xp/LENOVO/21BX/qcvss8280.mbn         video
```

A one-to-one match with the C630's list, differing only in the model. So these
files are redistributable once Lenovo and Qualcomm permit it — it is a
licensing question, not a technical one.

`qcom/sdm850/LENOVO/81JL/` has simply never been contributed. Fedora ships no
sdm850 model firmware at all, so it will not turn up in a `linux-firmware`
update. Four sources are known:

Two projects publish C630 firmware. They overlap but neither is a superset of
the other, and both are worth knowing about.

**WOA-Project** — <https://github.com/WOA-Project/Qualcomm-Reference-Drivers>

Windows-on-ARM driver packages by vendor and model, under `Lenovo/YogaC630/`.
Nine versions are published, `200.0.1.0` through `200.0.19.0`; the cabinets
differ between them. This is the only source here for the GPU, DSP, sensor and
venus blobs:

| Cabinet | Contains |
|---|---|
| `qcdx850.cab` | `qcdxkmsuc850.mbn` (zap shader), `qcvss850.mbn` (venus) |
| `qcipa850.cab` | `ipa_fws.elf` |
| `qcsubsys850.cab` | `qcadsp850.mbn`, `qccdsp850.mbn`, `qcslpi850.mbn`, `qcdsp1v2850.mbn`, `qcdsp2850.mbn`, `WLANMDSP.MBN` |
| `qcwlan850.cab` | `bdwlan.bin`, `bdwlan.b3*` — ath10k board data |
| `mcfg_subsys_ext850.cab` | `MCFG/*` modem configuration |
| `qcsubsys_extension850.cab` | ADSP/CDSP runtime modules, `fastrpc_shell_*` |

`cabextract` or `7z x` opens them. Ignore the `_nm` and `_CLS` variants; the
device tree asks for the plain names. Fetch via the GitHub API's
`download_url` — the `raw.githubusercontent.com` paths return 404.

**aarch64-laptops** — <https://github.com/aarch64-laptops/build>

Files checked directly into `misc/lenovo-yoga-c630/`, no extraction needed.
A narrower set, but with things WOA-Project does not carry:

| Path | Contains |
|---|---|
| `wifi/` | `qcdsp1v2850.mbn`, `qcdsp2850.mbn`, `wlanmdsp.mbn`, `bdwlan.*` |
| `wifi/create-board-2.bin/` | a prebuilt `board-2.bin`, and the script that makes it |
| `gpu/firmware/` | `a630_gmu.bin` |
| `acpi/`, `laptop-lenovo-yoga-c630.dts` | the machine's ACPI tables and a device tree |

It has none of `qcdxkmsuc850.mbn`, `qcadsp850.mbn`, `qccdsp850.mbn`,
`qcslpi850.mbn`, `qcvss850.mbn` or `ipa_fws.elf`. Note `wifi/firmware-5.bin` is
an HTML error page in the checkout, not firmware — Fedora's `atheros-firmware`
ships a real one.

**Which to use.** For the GPU, DSPs, sensors and venus, WOA-Project is the only
option of the two. For the modem and WLAN trio both publish files, and they are
different builds — same sizes for the modem pair, different checksums. On this
machine both produce identical results, so nothing here distinguishes them; see
[What the firmware does not fix](#what-the-firmware-does-not-fix).

**Lenovo's Digital Download Recovery Service.** Lenovo builds a Windows
recovery image against a machine's serial number and lets the owner download
it. That image contains `Windows/System32/DriverStore/FileRepository`, and
therefore the whole set as shipped for your specific machine. Slower than
either repository, and untested here.

**Another C630 owner.** Any 81JL's copy is as good as your own.

Whichever route, drop the result into `firmware/local/` and rebuild — see
[Baking firmware into the image](#baking-firmware-into-the-image).

## What the firmware does not fix

Installing all of it still leaves two things not working out of the box, for
reasons that are not about firmware. Audio turns out to be a missing Fedora
kernel module and is fixable today; onboard WiFi remains genuinely broken.
Recorded here so nobody repeats the search.

### Onboard WiFi: everything is in place except a working modem

The name gives it away: **wlan*mdsp*** is the WLAN *Modem* DSP. On SDM845 the
WiFi firmware runs on the modem, so `ath10k_snoc` binds to `18800000.wifi` and
then sits silent. It logs nothing and fails nothing — it has registered a QMI
client and is waiting for service 69, `wlfw`, which only appears once the modem
is up. `qrtr-lookup` shows what is and is not there.

Everything on the WiFi side checks out:

- `ath10k_snoc` binds successfully and is not in deferred probe
- the device tree carries `qcom,calibration-variant = "Lenovo_C630"`
- Fedora's `board-2.bin` contains a matching `variant=Lenovo_C630` entry
- `modemst1` and `modemst2` hold ~2 MB of real NV data — not wiped

#### The userspace services are required, and mostly missing by default

[aarch64-laptops' WiFi README][aal-wifi] lists four services. On a current
Fedora two of them are now in the kernel, and the other two are packaged but
inactive:

| Service | Status here |
|---|---|
| `pd-mapper` | in-kernel as `qcom_pd_mapper` — nothing to install |
| `qrtr-ns` | in-kernel — the `qrtr` package supplies only `qrtr-lookup`/`qrtr-cfg` |
| `rmtfs` | packaged, **disabled by default** — serves the modem's NV partitions |
| `tqftpserv` | packaged, **not installed** — serves files to the modem over TFTP |

```sh
sudo dnf install rmtfs tqftpserv qrtr
sudo systemctl enable --now rmtfs tqftpserv
```

Enabling these took `qrtr-lookup` from 19 services to 25. A working setup is
said to show around 40.

[aal-wifi]: https://github.com/aarch64-laptops/build/tree/master/misc/lenovo-yoga-c630/wifi

#### The modem fetches wlanmdsp over TFTP

This is the step that is easy to miss, and the reason `tqftpserv` exists.
The modem asks it for the WLAN firmware by path — but where tqftpserv serves
from depends on the version, and the layout below (from aarch64-laptops) does
**not** work with Fedora's current build; see [The tqftpserv paths above are
version-dependent](#the-tqftpserv-paths-above-are-version-dependent). On
current Fedora:

```sh
sudo ln -sf updates/qcom/sdm850/LENOVO/81JL/wlanmdsp.mbn /lib/firmware/wlanmdsp.mbn
```

The modem also wants a writable area, and says so in the `tqftpserv` journal:

```
[TQFTP] RRQ: /readwrite/ota_firewall/ruleset → unable to open (2), reject
[TQFTP] WRQ: /readwrite/server_check.txt
```

On current Fedora that lives in `/tmp/tqftpserv` (recreate after each boot, or
add a tmpfiles.d entry):

```sh
sudo mkdir -p /tmp/tqftpserv/ota_firewall
sudo touch /tmp/tqftpserv/ota_firewall/ruleset
sudo chmod -R 0777 /tmp/tqftpserv
```

Watching `journalctl -u tqftpserv` while starting the modem is the best
available feedback channel — it shows exactly which paths the modem wants.

#### Where it actually stops

The modem boots and then dies at RF front-end initialisation, roughly every 42
seconds until stopped:

```
qcom-q6v5-mss 4080000.remoteproc: MBA booted without debug policy, loading mpss
qcom-q6v5-mss 4080000.remoteproc: fatal error received:
    RFLM@rflm_qlnk.cpp:1087 [3,3] RF stuck in QLINK start state: 0x0 after maximum
remoteproc remoteproc0: crash detected ... recovering
```

It never gets far enough to request `wlanmdsp`, so WLAN never starts.

Three firmware pairings, three distinct failures — worth recording, because the
differences are informative:

| MBA (stage 1) | mpss (stage 2) | Result |
|---|---|---|
| full | full | boots, dies at `RF stuck in QLINK start state` |
| `_nm` | `_nm` | `PBL returned unexpected status` — rejected by the boot ROM |
| full | `_nm` | `MPSS header authentication failed: -3` |

The `_nm` ("no modem") variants are the obvious idea — Fedora ships the
equivalent `modem_nm.mbn` for generic sdm845, and skipping cellular RF is
exactly what is wanted here. They are not signed for this device and cannot be
made to load.

#### Firmware has been eliminated as the variable

The nine driver versions under `Lenovo/YogaC630/` contain five distinct
modem-firmware builds — `200.0.1.0` through `200.0.4.0` share one 2018-vintage
pair (a different mpss size from all later builds, and the closest match to
this machine's 2018-10 manufacture date), `200.0.10.0`, `200.0.11.0`,
`200.0.14.0`/`200.0.16.0`, and `200.0.19.0` are the others, and
aarch64-laptops' copies differ again by checksum. Every one of them was booted
on this machine, with `rmtfs` and `tqftpserv` running. Every one produces the
**identical** QLINK failure, crashing about 0.4 s after mpss comes up.

So neither the firmware nor its vintage is the variable. Nor is the kernel
version, at least not recently: Fedora's 6.12.15 (`kernel-core` from koji,
booted with the same device tree) produces the identical crash. The
aarch64-laptops setup that ticked WiFi ran a 2019 5.x branch, so a very old
regression is not excluded — but their issue #51 records modem crash/recovery
lockups on that setup too, so "WiFi worked" and "the modem crashed" may well
have coexisted, with `wlanmdsp` loading in the window before the first crash.
An earlier note in this project's history has the crash loop at roughly 42 s;
today it is ~3.4 s. If those cadences are both real, what changed — and
whether a long first window is enough for `wlfw` to appear — is an open
question.

#### What QLINK is, and what the modem reads before dying

QLINK is the SerDes link between the SoC and the SDR845 RF transceiver chip;
`rflm_qlnk.cpp` is the modem firmware's manager for it. "Stuck in start state
0x0" means the transceiver never answered — the signature of a chip that is
absent, unpowered, or misconfigured. This unit is an LTE SKU (SIM slot,
factory-label IMEI, modem worked under Windows), so the chip is present and
functional; under Windows, ACPI and the Qualcomm PEP handle subsystem power,
and under device-tree boot whatever they did for the RF front end has no
equivalent. That is a hypothesis, not a finding — the DSDT keeps its power
management in opaque PEP tables, so it could not be confirmed from the tables
alone.

`rmtfs -v` shows the modem's storage dialogue in the half-second before the
crash: it opens `modem_fs1`, `modem_fs2`, `modem_fsg` and `modem_fsc`
successfully, requests two partitions Linux's rmtfs has never heard of —
`modem_fsg_oem_1` and `modem_fsg_oem_2`, names that appear inside the mpss
image itself, with no provider found in any WOA-Project cabinet — allocates
its 2 MB buffer, reads a superblock's worth, and dies. On this machine
`modemst1`/`modemst2` carry valid `IMGEFS1`/`IMGEFS2` filesystems, while
**`fsg` and `fsc` are entirely blank**. Byte copies of all four live in
`firmware/local/nv-backup/` (gitignored), taken before any experimentation.

#### The tqftpserv paths above are version-dependent

The `/lib/firmware/readonly/firmware/image/` and `/lib/firmware/readwrite/`
layout described earlier matches the old aarch64-laptops tqftpserv. Fedora's
current build is different: `/readwrite/*` is served from `/tmp/tqftpserv`
(tmpfs — anything created there is gone on reboot), and
`/readonly/firmware/image/<file>` resolves through the normal firmware search
path, i.e. `/lib/firmware/<file>`. So `wlanmdsp.mbn` needs to be reachable as
`/lib/firmware/wlanmdsp.mbn` (a symlink to the copy under `updates/` works),
and the writable tree belongs under `/tmp/tqftpserv/`.

The full Windows-side TFTP tree can be reconstructed, and was: the driver
INFs are TFTP-mapping manifests, and `mcfg_subsys_ext850.inf` maps 159 modem
configuration files — carrier configs and, notably, `mcfg_hw` hardware RF
configs for the SDX20 — into `readonly/firmware/image/modem_pr/mcfg/...`.
Serving all of it changes nothing: the modem crashes before requesting any of
it.

Things that sound like they should help and do not:

- **Unplugging the USB WiFi dongle.** Different bus, different driver.
  `ath10k_snoc` is waiting on the modem, not competing for a radio.
- **`rmtfs` alone.** Necessary, not sufficient. Its "failed to update start
  state" warnings are its optional remoteproc-control helper and are harmless.
- **The `_nm` variants.** See above.
- **Any published firmware version.** All five distinct builds fail
  identically; see above.
- **The MCFG configuration tree.** Served in full over TFTP; the modem never
  gets far enough to ask for it.
- **`clk_ignore_unused pd_ignore_unused`.** Already on the command line
  (postmarketOS ships the same for this machine).

Untried, in rough order of promise: reporting upstream — the error string
appears nowhere public, and linux-arm-msm (Bjorn Andersson, Dmitry Baryshkov,
Steev Klimaszewski all have C630 history) is where the RF-power hypothesis
could be confirmed or killed; Lenovo's Digital Download Recovery Service image
for this serial, which would show the exact original driver set and whatever
provides `modem_fsg_oem_*` under Windows; building `board-2.bin` from the
C630's own `bdwlan.*` files (only relevant once the modem stays up); and the
5.x aarch64-laptops kernel itself, to test whether its crash cadence leaves a
usable window.

### Audio: Fedora does not build the speaker-amp driver

The ADSP runs, the audio services register, and the codec answers:

```
wcd934x-slim 217:250:1:0: WCD934x chip id major 0x108, minor 0x1
```

So SLIMbus, the ADSP and the hardware are all fine. The trail of suspicion
used to end at this SoundWire message:

```
qcom-soundwire wcd934x-soundwire.12.auto: din-ports (2) mismatch with controller (6)
platform sound: deferred probe pending: msm-snd-sdm845: SLIM Playback: codec dai not found
```

The first line is a red herring. Since kernel 6.18 (commit `9e53a66a2f2f`,
"soundwire: qcom: deprecate qcom,din/out-ports") it is a warning: the
device-tree value overrides the register value afterwards and the probe
continues.

The real blocker is the second line. The device tree's `SLIM Playback` link
names four codec DAIs: the WCD9340, the SoundWire master inside it, and the
two WSA881x speaker amplifiers hanging off that master. Fedora's aarch64
kernel has never set `CONFIG_SND_SOC_WSA881X` — while the newer siblings
WSA883X and WSA884X, which the ThinkPad X13s needs, are both `=m`. With no
amplifier driver the two SoundWire slaves never bind, the machine driver
cannot find its DAIs, and the card sits in deferred probe forever.

Building the one missing module proves it. `sound/soc/codecs/wsa881x.c` is a
single self-contained file; compiled against the running kernel
(`kernel-devel`, `gcc`, `make`, then `make -C /lib/modules/$(uname -r)/build
M=$PWD modules` with `obj-m += snd-soc-wsa881x.o` and `snd-soc-wsa881x-objs :=
wsa881x.o`), installed under `/lib/modules/$(uname -r)/extra/` and picked up
by `depmod -a`, it autoloads on the SoundWire modalias at boot and the
`Lenovo-YOGA-C630-13Q50` card registers, headphone-jack detection included.
The UCM profile has been in alsa-ucm-conf since 2020; `alsaucm` itself is in
`alsa-ucm-utils`, so on the minimal image:

```sh
alsaucm -c Lenovo-YOGA-C630-13Q50 set _verb HiFi set _enadev Speaker
speaker-test -D plughw:0,0 -c2 -t sine -f 440 -l 1
```

The `overflow error on RX port` and register-read `-16` messages logged when a
stream stops are noise, seen on working sdm845 devices too.

**PipeWire needs one more thing.** The ADSP front-end advertises S24_LE
(24-bit samples in 32-bit containers) and then garbles it: playback runs
without any error, consumes data at the right rate, and powers every DAPM
widget — a byte-for-byte identical `/sys/kernel/debug/asoc` widget dump
against a working stream — while producing ~-48 dB of near-silence,
consistent with the samples being bit-shifted inside the container. PipeWire
prefers the widest format a device offers, so every PipeWire client (mpv,
GNOME, `pw-play`) is inaudible while `aplay` and `speaker-test`, which
default to S16_LE, work perfectly. This cost most of a day: the failure is
per-format, invisible in every log, and easy to misattribute to routing,
suspend/resume or per-stream volumes, all of which look guilty and are not.
`overlay/etc/wireplumber/wireplumber.conf.d/51-c630-alsa.conf` pins the
built-in card to S16_LE, which fixes every client. The underlying q6asm
format bug is worth a report to alsa-devel.

Unrelated to the hardware, two pieces of routing politics are worth knowing
on a machine like this: PipeWire's RAOP discovery will happily make a
neighbour's AirPlay speaker the default sink, and a newly connected
Bluetooth device also takes over the default. If sound "does not work",
check `wpctl status` for where the default actually points before blaming
the drivers.

The durable fix is one line of Fedora kernel config — flipping
`CONFIG_SND_SOC_WSA881X` to `=m` for aarch64, an easy case given the in-tree
device tree requires it and both newer WSA drivers are already enabled. Until
that lands, the module has to be built out of tree per kernel.

One thing worth knowing if you go looking: `slim_qcom_ngd_ctrl` can probe before
the ADSP is ready and log `QMI wait timeout`. Reloading the module afterwards
gets the codec detected, which is how the chip ID above was obtained. A series
of NGD probe-ordering fixes (Bjorn Andersson, mid-2026, Cc: stable) targets
exactly this class of race and should reach Fedora with the 7.2 rebase.

## Extracting them

The image ships Fedora's `qcom-firmware-extract`. Run it once:

```sh
sudo c630-firmware
sudo reboot
```

That wraps `qcom-firmware-extract` and regenerates the initramfs afterwards,
which matters — the ADSP firmware has to be available before the root
filesystem is mounted, and `overlay/etc/dracut.conf.d/10-c630.conf` only takes
effect on a rebuild.

If it cannot find Windows, mount it yourself and point the tool at it:

```sh
lsblk -o NAME,SIZE,FSTYPE,LABEL       # find the NTFS partition
sudo mkdir -p /mnt/windows
sudo mount /dev/sda4 /mnt/windows
sudo qcom-firmware-extract --windows-dir /mnt/windows
```

`qcom-firmware-extract` was written for the Snapdragon 8cx and X Elite laptops.
If it does not recognise the SDM850, copy the files by hand — they are in
`Windows/System32/DriverStore/FileRepository/` under a `qcdx*.inf_arm64_*`
directory:

```sh
sudo find /mnt/windows/Windows/System32/DriverStore/FileRepository \
    -iname 'qc*850.mbn' -o -iname 'ipa_fws.elf' -o -iname 'qcdxkmsuc850.mbn'

sudo mkdir -p /lib/firmware/updates/qcom/sdm850/LENOVO/81JL
sudo cp <each file> /lib/firmware/updates/qcom/sdm850/LENOVO/81JL/
sudo dracut --force --regenerate-all
```

Check it took:

```sh
dmesg | grep -iE 'remoteproc|adsp|cdsp'
```

## Baking firmware into the image

If you would rather not repeat the extraction on every reinstall, drop the files
into `firmware/local/` in this repo, mirroring the `/usr/lib/firmware/updates/`
layout:

```
firmware/local/qcom/sdm850/LENOVO/81JL/qcadsp850.mbn
firmware/local/qcom/sdm850/LENOVO/81JL/qccdsp850.mbn
...
```

`build/stage2.sh` picks them up automatically. They are gitignored, and should
stay that way — this repository is public and the blobs are not yours to
publish.