Genlocked alsa loopback problem with Wolfram wm8731 on 6.6.40

[rafael2k]

Describe the bug

After upgrading to kernel 6.6.40 from 6.6.31 (same problem with 6.6.42), I can not use the alsa loopback devices genlocked to the Wolfram 8731 sound device.

Steps to reproduce the behaviour

Just update kernel and run an application which opens the alsa loopback devices.

Device (s)

Raspberry Pi 5, Raspberry Pi 4

System

cat /etc/rpi-issue

Generated using pi-gen, https://github.com/RPi-Distro/pi-gen, 2acf7afcba7d11500313a7b93bb55a2aae20b2d6, stage2

vcgencmd version

2024/07/30 15:25:46
Copyright (c) 2012 Broadcom
version 790da7ef (release) (embedded)

uname -a

Linux gateway.hermes.radio 6.6.42-v8+ #1787 SMP PREEMPT Thu Jul 25 17:44:41 BST 2024 aarch64 GNU/Linux

Logs

Initial dmesh output. Nothing suspicious to me.

(...)
[ 4.826923] wm8731 1-001a: Assuming static MCLK
[ 4.826953] wm8731 1-001a: supply AVDD not found, using dummy regulator
[ 4.827050] wm8731 1-001a: supply HPVDD not found, using dummy regulator
[ 4.827074] wm8731 1-001a: supply DCVDD not found, using dummy regulator
[ 4.827101] wm8731 1-001a: supply DBVDD not found, using dummy regulator
[ 4.831445] audioinjector-stereo soc:sound: ASoC: driver name too long 'audioinjector-pi-soundcard' -> 'audioinjector-p'
[ 4.868650] audioinjector-stereo soc:sound: successfully loaded
(...)

After opening the alsa loopback, thousands per second of:
[ 242.922102] snd_aloop snd_aloop.1: snd_timer_start(0,0,0) failed with -22
[ 242.922117] snd_aloop snd_aloop.1: snd_timer_stop(0,0,0) failed with -16
[ 242.922152] snd_aloop snd_aloop.1: snd_timer_start(0,0,0) failed with -22
[ 242.922157] snd_aloop snd_aloop.1: snd_timer_stop(0,0,0) failed with -16
[ 242.922178] snd_aloop snd_aloop.1: snd_timer_start(0,0,0) failed with -22
[ 242.922182] snd_aloop snd_aloop.1: snd_timer_stop(0,0,0) failed with -16
[ 242.922203] snd_aloop snd_aloop.1: snd_timer_start(0,0,0) failed with -22
[ 242.922207] snd_aloop snd_aloop.1: snd_timer_stop(0,0,0) failed with -16

Then the soundcard itself starts to misbehave (crazy sounds heard):
[ 614.169627] dma dma2chan0: BUG: dma2chan0, IRQ with no descriptors
[ 629.590104] dma dma2chan0: BUG: dma2chan0, IRQ with no descriptors
[ 662.121953] dma dma2chan0: BUG: dma2chan0, IRQ with no descriptors
[ 675.205707] dma dma2chan0: BUG: dma2chan0, IRQ with no descriptors

snd-aloop module parameters:
options snd-aloop enable=1,1,1 index=1,2,3 timer_source=hw:0,0

Additional context

No response

[pelwell]

Could you try some of the intervening releases using the hashes on this page?:
https://github.com/raspberrypi/rpi-firmware/commits/master/
(sudo rpi-update <hash>)

[rafael2k]

Yes. Will do it.

[rafael2k]

Kernel 6.6.32 works. Kernel 6.6.33 is broken with the described issue.

[pelwell]

I've looked through the 760 commits in that range and, although I have some suspicions, there are no obvious culprits. It would help me if you could unpack the phrase "alsa loopback devices genlocked to the Wolfram 8731 sound device":

1. How is this different from normal play/record activity?
2. Does normal play/record activity work?
3. Is this configuration the simplest that demonstrates the problem?

Without having one of these soundcards, I think we'll end up with a pull-request with sets of commits reverted until we can pinpoint the offending one.

[rafael2k]

Right. I'll do more tests, but answering the questions:

1. The only difference is that the alsa loopback is being driven by the wm8731 (timer_source=hw:0,0 to the alsa loopback module, where hw:0,0 is the wm8731)
2. It works ok until I start to see "dma dma2chan0: BUG: dma2chan0, IRQ with no descriptors" in kernel log
3. I'll do more tests and write a simple C code to demonstrate (the production software is not so small, available here: https://github.com/Rhizomatica/hermes-net/tree/main/trx_v2-userland )

I'll also do more tests in the Pi 4 to confirm it is Pi5-specific, in case it helps.

[pelwell]

1. So by "genlocked alsa loopback" you mean the system is capturing samples from the soundcard input and replaying them from the output, using the same clock for both so that there should be a one-to-one correspondence?
2. Understood.
3. That would be helpful, especially if the problem is shared with other soundcards.

[rafael2k]

1. So "genlocked alsa loopback" I just mean that I'm using alsa loopback "genlock" feature (use another sound card as clock source) by adding ¨timer_source=hw:0,0". The DSP pipeline is a bit more complicated than just a replay of the samples. You are right about the "one-to-one" correspondence.
3. Right. I'll work on it.

[rafael2k]

Btw, I think the Pi5 has enough juice to compile kernel. If I want to bisect the issue, which is the first commit after 6.6.32 you would guess to start with?

[pelwell]

I'll prepare a branch with all the relevant commits as cherry-picks, then you'll be able to use git bisect.

[rafael2k]

thanks!

[pelwell]

If you git fetch now you should find you have origin/wm8731test.

$ git checkout origin/wm8731test
$ git bisect start
$ git bisect good 5882bc5db17dd 
$ git bisect bad HEAD
Bisecting: 376 revisions left to test after this (roughly 9 steps)
[bc095ce83c5d4d44392c5f497f73fdfab2d2053d] RDMA/bnxt_re: Adds MSN table capability for Gen P7 adapters

Build and test each kernel, and after each test run either git bisect good or git bisect bad, and with a bit of luck (for no broken builds) it will converge on the culprit.

[rafael2k]

Thanks @pelwell!!

[rafael2k]

I'm still working on it. Please hang on.
: )

[pelwell]

I'm not going anywhere (this week)

[rafael2k]

I already found a commit that the software works, but I still get...

[ 2773.338718] dma dma2chan0: BUG: dma2chan0, IRQ with no descriptors

now at [f00c159] nvme: cancel pending I/O if nvme controller is in terminal state