[BUG] SOF freezes when running SRC pipeline conversion 44.1khz -> 48khz

[dbaluta]

Describe the bug

SOF freezes when playing file with following parameters: 44.1Khz, 32bits, 2 channels. Only FW compiled with Xtensa toolchain see this issues. GCC builds are ok.

To Reproduce

Use sof-imx8mp-src-wm8960.m4 and do:

aplay -Dhw:0,0 audio41k2c32b.wav.

aplay freezes and no sound can be heard.

Reproduction Rate

always

Expected behavior

aplay should play the wav file with accurate sound.

Environment

1. Branch name and commit hash of the 2 repositories: sof (firmware/topology) and linux (kernel driver).
   • Kernel: latest
   • SOF: stable-v2.2 (anything newer reproduces the issue)
2. Name of the topology file
   • Topology: sof-imx8mp-src-wm8960.m4
3. Name of the platform(s) on which the bug is observed.
   • Platform: imx8, imx8x, imx8m

Root cause for this bug is:

commit 3aff4598f7ebd79d8cfb0398d85a85bbf4ba37e1
Author: Seppo Ingalsuo <[email protected]>
Date:   Fri Apr 1 21:22:03 2022 +0300

    Audio: SRC: Increase length of internal buffer
    
    The too small size of this buffer caused a new 11025 to 8000 Hz
    conversion pipeline to freeze due to impossible condition in
    source/internal/sink buffers.
    
    Signed-off-by: Seppo Ingalsuo <[email protected]>

[singalsu]

Thanks for report @dbaluta ! There seems to be another freeze possibility that that the larger internal buffer triggered.

Have you tried if 11.025 to 8 kHz freezes now? Though I fully agree 44.1 kHz freeze is more critical, so good that the commit revert worked around it.

[singalsu]

"Only FW compiled with Xtensa toolchain see this issues. GCC builds are ok." --> this is a very good pointer for starting debugging, it sounds like a bug in xcc version

[singalsu]

In my test with UP2 with modified sof-apl-pcm512x.tplg to add support for s32_le (#6109) I don't get freeze with HiFi3 xtos xt-xcc build with 8881a406af67536b9ac6bfbb53d1c712aacdae81 that is just before @dbaluta 's revert.

I think I need to test with Zephyr since I assume imx8 used it, apply to my test topology similar pipeline buffers parameters as on imx8. Also imx8 seems to be HiFi4, and there's a brand new src_hifi4.c also to inspect.

@RanderWang Did you observe the bug with a HiFi4 platform?

[singalsu]

Zephyr build is not triggering this issue for me. Has anyone replicated this issue with HiFi3?

[dbaluta]

In my test with UP2 with modified sof-apl-pcm512x.tplg to add support for s32_le (#6109) I don't get freeze with HiFi3 xtos xt-xcc build with 8881a40 that is just before @dbaluta 's revert.

I think I need to test with Zephyr since I assume imx8 used it, apply to my test topology similar pipeline buffers parameters as on imx8. Also imx8 seems to be HiFi4, and there's a brand new src_hifi4.c also to inspect.

I have this replicated with XTOS not Zephyr. Also, have not tried 11.025 to 8 kHz as our base is 48Khz and we convert from X to 48Khz always.

I might find some time at the end of the week to test this scenario.

Also, as you noticed i.MX used HIFI4.

[RanderWang]

I can reproduce it with Linux kernel + cSOF + cavs-src-gain-mixin.tplg

[RanderWang]

two issues: (1) need to revert

    Revert "Audio: SRC: Fix incorrect and insufficient buffer size check"
    
    This reverts commit 40b6c80f18b4e4f173a0b6ebc2376d4b9bfc2e7a.

a error happen src/audio/src/src.c:709 INFO src_params(), source_rate = 44100, sink_rate = 48000 src/audio/src/src.c:711 INFO src_params(), sourceb->channels = 2, sinkb->channels = 2, dev->frames = 48 src/audio/src/src.c:720 INFO src_params(), blk_in = 0, blk_out = 0 ....../pipeline-params.c:340 INFO pipe prepare src/audio/src/src.c:967 INFO src_prepare() src/audio/src/src.c:944 ERROR Source size 720 is less than required 744 src/audio/src/src.c:990 ERROR src_prepare(): source or sink buffer size is insufficient ....../pipeline-params.c:347 ERROR pipeline_prepare(): ret = -105, dev->comp.id = 262144 src/ipc/ipc4/handler.c:142 ERROR ipc: pipe 1 comp 0 prepare failed -105

(2) with step (1) we will find fw stuck

[singalsu]

I can reproduce it with Linux kernel + cSOF + cavs-src-gain-mixin.tplg

This was with xt-xcc? Can you try if gcc built firmware has this issue?

[singalsu]

@dbaluta Can this bug be closed now after your recent fix?

[dbaluta]

@singalsu Let me test this today and get back to you.

[dbaluta]

i should have close this much earlier, it works not because of my recent fix but because we have reverted

    Revert "Audio: SRC: Fix incorrect and insufficient buffer size check"
    
    This reverts commit 40b6c80f18b4e4f173a0b6ebc2376d4b9bfc2e7a.