ASoC: tas2783-sdw: port prepare never completes after S0i3, no audio and no error (AMD ACP7.0, ASUS ProArt PX13)
Hi,
On an ASUS ProArt PX13 HN7306EAC (Ryzen AI MAX+ 395, AMD ACP rev 0x70, two TAS2783 at unique_id 0x8/0xB plus RT721 on SoundWire link 1), audio is silent after resume from s2idle whenever the platform actually reaches S0i3. Playback runs with no error whatsoever - no XRUN at first, no prepare failure, no bus error - and the speakers stay quiet until the PCM is fully torn down and recreated.
This is with v7.2-rc4 plus 5893013efabb ("ASoC: amd: ps: disable MSI on resume in ACP PCI driver"), which is what makes the peripherals re-attach at all on this board [1], and with a local fix that drops the stale tas2783 regcache on re-attach [2]. Both are necessary here; neither is sufficient.
I measured where the audio stops, and the answer is not where I expected.
1. The ACP side is healthy
--------------------------
I dumped the ACP registers that the driver programs (ring buffer address and size, FIFO address and size, DMA size, watermark, stream enable, plus the SoundWire manager block) during actual playback, once in the silent state after resume and once after the workaround restored audio.
72 of 78 registers are bit-identical. The only differences are ACP_EXTERNAL_INTR_CNTL (one extra bit, 0x10000 = PDM_DMA_INTR_MASK, i.e. the digital mics, unrelated to this stream) and the last immediate command/response pair on the bus.
That is expected: acp63_sdw_pcm_resume() reprograms the PTE, the ring buffer, the watermark and the DMA interrupt masks for every live substream on system resume. The ACP DMA configuration is restored correctly.
2. The DMA is running at the correct rate while silent
------------------------------------------------------
Sampling ACP_P1_AUDIO1_TX_LINEARPOSITIONCNTR every 500 ms during silent playback:
0x000002FBF880 +96064 0x000002FD6F80 +96000 0x000002FEE6C0 +96064 0x000003005E00 +96064 0x00000301D500 +96000
96000 bytes per 500 ms = 192000 B/s = 48000 Hz x 2 ch x 2 bytes, exactly nominal. The ACP is fetching the buffer and feeding the SoundWire FIFO the whole time the speakers are silent.
3. The peripheral port never finishes preparing
-----------------------------------------------
Dumping /sys/kernel/debug/soundwire/master-0-1/sdw:*/registers during playback, silent state versus working state, the entire difference across all three peripherals is two registers - and it is the same register on both amplifiers:
DP1 0x104 (DPn_PrepareStatus): silent = 0x3 working = 0x0
with DPn_PrepareCtrl (0x105) = 0x3 and DPn_ChannelEn = 0x3 in both cases. (The third difference is SCP_Int1 on one amp, i.e. transient interrupt status.)
Per the core's own reading of that register - "Poll for NOT_PREPARED==0" in sdw_prep_deprep_slave_ports() - a set bit means the channel is *not* prepared. So after the power gate both amps sit with both channels unprepared, indefinitely, while the manager streams data at them.
4. Why nothing complains
------------------------
tas2783-sdw declares simple_ch_prep_sm, so the core skips its entire prepare block: it neither writes DPn_PrepareCtrl nor polls DPn_PrepareStatus. The driver writes DPn_PrepareCtrl itself from tas_port_prep() - as the comment there says, "TAS2783 requires explicit port prepare during playback stream setup even when simple_ch_prep_sm is enabled. Without this, the port fails to enter the prepared state resulting in no audio output" - but nobody ever checks whether the port actually reached the prepared state.
The result is a completely silent failure: the stream is enabled, the DMA runs, the log is clean, and there is no sound.
What restores it is a full teardown: hw_free (which de-prepares the port, writing DPn_PrepareCtrl = 0) followed by hw_params and a fresh prepare. Re-preparing without the de-prepare - which is what a plain userspace resume does - leaves DPn_PrepareStatus stuck. On this board I work around it by cycling the card profile after every resume, which forces that teardown.
Questions
---------
- Should TAS2783 declare simple_ch_prep_sm at all? The port evidently does not prepare instantaneously, which is the premise of that property; letting the core do its normal write-and-poll would at least turn this into a visible error instead of silence.
- If the property is correct, should tas_port_prep() poll DPn_PrepareStatus after writing DPn_PrepareCtrl, mirroring what the core does?
- Either way: why does a fresh prepare not complete after the device has lost power, when a de-prepare followed by a prepare does? If the amp needs the port explicitly de-prepared before it will accept a new prepare after a power cycle, that sequencing has to happen somewhere on resume.
Happy to test patches, dump additional registers, or run traces on this hardware.
[1] https://lore.kernel.org/all/[email protected]/
[2] https://lore.kernel.org/all/[email protected]/
Thanks,
Andrey
Quoted 81 lines · Andrey Golovko
Hi,
On an ASUS ProArt PX13 HN7306EAC (Ryzen AI MAX+ 395, AMD ACP rev
0x70, two TAS2783 atunique_id0x8/0xB plus RT721 on SoundWire link 1), audio is silent after resume from s2idle whenever the platform actually reaches S0i3. Playback runs with no error whatsoever - no XRUN at first, no prepare failure, no bus error - and the speakers stay quiet until the PCM is fully torn down and recreated.This is with v7.2-rc4 plus
5893013efabb("ASoC: amd: ps: disable MSI on resume in ACP PCI driver"), which is what makes the peripherals re-attach at all on this board [1], and with a local fix that drops the stale tas2783 regcache on re-attach [2]. Both are necessary here; neither is sufficient.I measured where the audio stops, and the answer is not where I expected.
1. The ACP side is healthy
--------------------------
I dumped the ACP registers that the driver programs (ring buffer address and size, FIFO address and size, DMA size, watermark, stream enable, plus the SoundWire manager block) during actual playback, once in the silent state after resume and once after the workaround restored audio.72 of 78 registers are bit-identical. The only differences are
ACP_EXTERNAL_INTR_CNTL(one extra bit,0x10000=PDM_DMA_INTR_MASK, i.e. the digital mics, unrelated to this stream) and the last immediate command/response pair on the bus.That is expected:
acp63_sdw_pcm_resume()reprograms the PTE, the ring buffer, the watermark and the DMA interrupt masks for every live substream on system resume. The ACP DMA configuration is restored correctly.2. The DMA is running at the correct rate while silent
------------------------------------------------------
SamplingACP_P1_AUDIO1_TX_LINEARPOSITIONCNTRevery 500 ms during silent playback:0x000002FBF880 +96064 0x000002FD6F80 +96000 0x000002FEE6C0 +96064 0x000003005E00 +96064 0x00000301D500 +9600096000 bytes per 500 ms = 192000 B/s = 48000 Hz x 2 ch x 2 bytes, exactly nominal. The ACP is fetching the buffer and feeding the SoundWire FIFO the whole time the speakers are silent.
3. The peripheral port never finishes preparing
-----------------------------------------------
Dumping /sys/kernel/debug/soundwire/master-0-1/sdw:*/registers during playback, silent state versus working state, the entire difference across all three peripherals is two registers - and it is the same register on both amplifiers:DP1
0x104(DPn_PrepareStatus): silent =0x3working =0x0with
DPn_PrepareCtrl(0x105) =0x3andDPn_ChannelEn=0x3in both cases. (The third difference isSCP_Int1on one amp, i.e. transient interrupt status.)Per the core's own reading of that register - "Poll for NOT_PREPARED==0" in
sdw_prep_deprep_slave_ports()- a set bit means the channel is *not* prepared. So after the power gate both amps sit with both channels unprepared, indefinitely, while the manager streams data at them.4. Why nothing complains
------------------------
tas2783-sdw declaressimple_ch_prep_sm, so the core skips its entire prepare block: it neither writesDPn_PrepareCtrlnor pollsDPn_PrepareStatus. The driver writesDPn_PrepareCtrlitself fromtas_port_prep()- as the comment there says, "TAS2783 requires explicit port prepare during playback stream setup even whensimple_ch_prep_smis enabled. Without this, the port fails to enter the prepared state resulting in no audio output" - but nobody ever checks whether the port actually reached the prepared state.The result is a completely silent failure: the stream is enabled, the DMA runs, the log is clean, and there is no sound.
Quoted 6 lines
What restores it is a full teardown:
hw_free(which de-prepares the port, writingDPn_PrepareCtrl= 0) followed byhw_paramsand a fresh prepare. Re-preparing without the de-prepare - which is what a plain userspace resume does - leavesDPn_PrepareStatusstuck. On this board I work around it by cycling the card profile after every resume, which forces that teardown.
Just to be clear: are you referring to a scenario where audio playback is on-going, the device does a system suspend during the audio playback due to a user request, and on resume audio is supposed to restart playing?
That's not an easy capability to test/automate. When I worked with Intel, we had scripts to force a suspend-resume with aplay running in the background. They worked most of the time except when they didn't...
Quoted 9 lines
Questions
---------
- Should TAS2783 declaresimple_ch_prep_smat all? The port evidently does not prepare instantaneously, which is the premise of that property; letting the core do its normal write-and-poll would at least turn this into a visible error instead of silence.
- If the property is correct, shouldtas_port_prep()pollDPn_PrepareStatusafter writingDPn_PrepareCtrl, mirroring what the core does?
It'd be surprising, these registers are typically updated quickly and if the sequence works for one playback it should work timing wise after resume.
Anyways, adding Niranjan H Y who contributed the patch "ASoC: tas2783A:
add explicit port prepare handling". The topic of this prepare bit was discussed in the thread "SoundWire: Allow Prepare command for Simplified_CP_SM" and we ended-up with the use of this port_prep callback. IIRC the use of the simple_ch_prep_sm was due to Windows/BIOS precedent.
Note that there was a change in the full port-prepare programming sequence, we've since stopped using this interrupt due to other deadlock problems (see "soundwire: stream: Poll for DP prepare to avoid interrupt deadlock"). It could very well be that this driver can now use the full state machine, who knows.
Quoted 5 lines
- Either way: why does a fresh prepare not complete after the device has lost power, when a de-prepare followed by a prepare does? If the amp needs the port explicitly de-prepared before it will accept a new prepare after a power cycle, that sequencing has to happen somewhere on resume.
It could be that the use of the port_prep callback is restricted to the initial stream setup. I am not sure if that stream callback is invoked a second time after a suspend-resume cycle, this driver is the only one who makes use of that callback...
Alternatively it could be that something else remains stuck and prevents the port prepare sequence from completing.
Quoted 14 lines · Pierre-Louis Bossart
Subject: Re: ASoC: tas2783-sdw: port prepare never completes after S0i3, no audio and no error (AMD ACP7.0, ASUS ProArt PX13)
Quoted 11 lines · Andrey Golovko
Questions
---------
- Should TAS2783 declaresimple_ch_prep_smat all? The port evidently does not prepare instantaneously, which is the premise of that property; letting the core do its normal write-and-poll would at least turn this into a visible error instead of silence.
- If the property is correct, shouldtas_port_prep()pollDPn_PrepareStatusafter writingDPn_PrepareCtrl, mirroring what the core does?
We need to use the same BIOS between Windows and Linux, the BIOS sets the device as Simplified_CP_SM.
The device expects the DPn_PrepareCtrl bits to be set to be functional - but doesn't update DPn_PrepareStatus.
If the device is configured as generic CP_SM, the DPn_PrepareStatus bits will not be updated. Hence we end up with port prepare fail errors even for normal scenario where it could produce audio. So, the changes were added in the codec driver to explicitly write DPn_PrepareCtrl even in simplified CP_SM and ignore the DPn_PrepareStatus.
Regards
Niranjan
Hi Niranjan, Pierre-Louis,
thanks to both - answers and new data inline.
Quoted 4 lines · Pierre-Louis
Just to be clear: are you referring to a scenario where audio playback is on-going, the device does a system suspend during the audio playback due to a user request, and on resume audio is supposed to restart playing?
Not necessarily on-going playback - that's the point that makes this hit ordinary users. PipeWire keeps the PCM substream open the whole session, so the substream survives the suspend regardless of whether anything was audibly playing. After resume, even a stream started *fresh* by an application plays into that surviving substream and is silent. The only way to get sound back is to force the substream through hw_free + hw_params (I cycle the card profile, which PipeWire translates into exactly that).
So the reproduction is deterministic and needs no aplay-in-background
race: boot, let PipeWire open the device, systemctl suspend for
Quoted 1 line
= 1 min (enough for
amd_pmcto report real S0i3 residency; without
residency the bug does not reproduce), resume, start any playback. Silent every time here.
Quoted 3 lines
It could be that the use of the
port_prepcallback is restricted to the initial stream setup. I am not sure if that stream callback is invoked a second time after a suspend-resume cycle
I can answer that with a trace rather than a guess, with one caveat:
stock ps-sdw-dma.c advertises SNDRV_PCM_INFO_RESUME, so a plain resume takes TRIGGER_RESUME and skips the prepare path entirely - no stream callbacks at all. With a local patch dropping INFO_RESUME (so userspace does a full snd_pcm_prepare recovery), the resume-spanning ftrace shows the complete sequence running after resume: sdw_prepare_stream -> sdw_prep_deprep_slave_ports -> tas_port_prep() invoked again, 4 calls (2 amps x 2 callbacks), DPn_PrepareCtrl written with the right masks, then enable. So the callback *is* invoked on the recovery path - and the amps are still silent.
The one thing that separates every silent case from every working case in my traces is the de-prepare: the working path (hw_free) writes DPn_PrepareCtrl = 0 first, then a fresh prepare writes 3. The silent path re-writes 3 over 3. After the device has lost power and been re-initialized (SW_RESET + firmware re-download on re-attach), a rewrite of the same value evidently does not re-arm anything; the 0 -> 3 edge does.
Quoted 2 lines · Niranjan
The device expects the
DPn_PrepareCtrlbits to be set to be functional - but doesn't updateDPn_PrepareStatus.
Thanks, that explains the design - I'll withdraw the "port prepare never completes" framing then, since it leaned on reading DPn_PrepareStatus, and per your description those bits are not to be interpreted in Simplified_CP_SM.
One empirical note, for what it's worth: on this hardware 0x104 is not static. Across all my dumps, on both amplifiers, it reads 0x3 in every silent state and 0x0 in every working state, tracking the audible state exactly. I'm not suggesting polling it - just that on this part the register does reflect something about the port state.
Which leaves the practical question for TI:
After the amplifier loses power while the manager is power-gated (attach as UNINITIALIZED, SW_RESET, firmware re-download), what does the device need before a port prepare takes effect again? Empirically writing DPn_PrepareCtrl = 0 and then 3 works, while writing 3 alone does not. If the 0 -> 3 sequence is genuinely required, where should it live - tas_port_prep() unconditionally de-preparing before preparing, or the codec driver de-preparing its ports as part of the re-attach re-init?
If TI can confirm the required sequence I'm happy to write and test the patch on this hardware - I already have the register dumps and traces automated.
Thanks,
Andrey
I have the root cause, and it is neither the host DMA nor the data port
itself: after the peripheral loses power in S0i3 its SDCA Power Domain Entity PDE23 comes back at PS3, and a Data Port cannot complete channel preparation while the Function is powered down. The only place the driver powers PDE23 on is tas_sdw_hw_params(), which the resume path never calls.
Below is the evidence, and at the end a question about where the fix belongs, which I do not think I should answer alone.
Test setup: ASUS ProArt PX13 HN7306EAC, AMD ACP7.0, two TAS2783 at unique 0x8/0xB plus an rt721-sdca on link 1. Kernel: broonie/sound for-next (7.2.0-rc6 base) with b627da430357, Peter Ujfalusi's two tas2783 reg_defaults fixes, the regcache sort fixes, Antoine's stereo patch and three unrelated display fixes. One s2idle cycle with 99.98 s of S0i3 residency out of 105 s of sleep, i.e. the ACP island really was power-gated.
After resume, with no workaround service running: all three peripherals Attached, both amps re-download firmware, MSI is re-disabled by 5893013efabb, suspend_stats clean, zero errors of any kind in the log -- and no audio, until the PCM is fully torn down and re-created.
The host side is not at fault, and I was wrong in July
=====================================================
I dumped the ACP and SoundWire manager registers in the silent state and again after a card-profile cycle had restored audio (78 registers, mmap of the PCI BAR). The playback stream configuration is identical in both:
ACP_P1_AUDIO1_TX_RINGBUFADDR 0x048E0000 both ACP_P1_AUDIO1_TX_RINGBUFSIZE 0x00008000 both ACP_P1_AUDIO1_TX_FIFOADDR/SIZE 0x700/0x100 both ACP_P1_AUDIO1_TX_DMA_SIZE 0x00000040 both ACP_P1_AUDIO1_TX_INTR_WATERMARK 0x00001000 both ACP70_SW1_AUDIO1_TX_EN 0x00000001 both
The only meaningful difference is the linear position counter, and it shows the DMA is running in the silent state too: 0x017F9440 = 25.1 MB, which at 48 kHz/2ch/S16 is ~131 s, matching the time since resume.
This retracts what I suggested in July, that ACP loses its SoundWire DMA ring-buffer configuration across S0i3 and would need it reprogrammed in .prepare. That hypothesis is wrong: the registers are already correct while there is no sound.
The peripheral reports the failure correctly
============================================
Slave-side registers, silent versus working, taken during playback:
DP1 PortCtrl (0x102) 0x20 both DP1 BlockCtrl1 (0x103) 0x0f both
bank SampleCtrl/Offset/HCtrl (0x122..0x126) identical
DPn_PrepareStatus (0x104) amp 0x8: 0x1 silent, 0x0 working
amp 0xB: 0x2 silent, 0x0 working
Niranjan, this is the point you doubted on 28 July: the device does update DPn_PrepareStatus, and it does so per channel -- each amp reports exactly the channel it owns as not prepared, 0x1 for the one on channel 0 and 0x2 for the one on channel 1 (single-channel masks, per Antoine's patch). The transport parameters are intact. Only the prepare state is lost.
Minimal reproduction, no ALSA involved
======================================
With a debug module that talks to the peripherals through sdw_write_no_pm()/sdw_read_no_pm() only, in the silent state after resume:
1. PrepareCtrl (0x105) already holds the channel mask, so writing the
same value again changes nothing: PrepareStatus stays 0x1/0x2.
2. PrepareCtrl <= 0 clears PrepareStatus to 0x0 immediately. De-prepare works, and the device is answering us.
3. PrepareCtrl <= mask sets PrepareStatus to the mask again, and it never clears. I polled for 50 ms, the same way sdw_prep_deprep_slave_port() does.
4. PDE23 Requested and Actual Power State both read 0x3, i.e. PS3.
5. Write PDE23 Requested = PS0. Actual reads 0x0 immediately. Repeat step 3 and PrepareStatus clears within 1 ms. Audio is back, without touching the PCM, without hw_params, without re-enumeration.
So the port is not broken and the state machine is not stuck: it is waiting for power that nobody restores.
Why nothing is logged
=====================
tas_sdw_hw_params() writes PDE23 Requested = PS0, with a retry loop whose comment already states the dependency:
/*
* Sometimes, there is error returned during power on.
* So added retry logic to ensure power on so that
* port prepare succeeds
*/
and tas_sdw_pcm_hw_free() writes PS3 on the way out. After S0i3 the peripheral is back at its register defaults, where PDE23 is PS3. Userspace resumes the surviving PCM with TRIGGER_RESUME rather than tearing it down, so hw_params never runs, so PDE23 is never powered up again.
The silence is total because tas2783 sets simple_ch_prep_sm, and sdw_prep_deprep_slave_port() then skips both the PrepareCtrl write and the NOT_PREPARED poll. The driver compensates with tas_port_prep(), whose comment says the same thing -- "the port fails to enter the prepared state resulting in no audio output" -- but that callback only runs during a real prepare. Nobody ever asks the peripheral whether it is ready, so a port that never prepares looks exactly like a healthy one.
Pierre-Louis, this also answers your question from 27 July about whether playback is ongoing during suspend: it does not matter. What matters is that a PCM stays open across the cycle, which PipeWire does by default on an idle sink.
Where should this be fixed?
===========================
Three candidates, and I would rather hear TI and the maintainers than pick one myself:
(a) ACP stops advertising SNDRV_PCM_INFO_RESUME on the SoundWire DMA PCMs. That flag promises a resume with no stream re-setup, which SoundWire cannot honour when the peripheral loses power; Intel's SoundWire DMA does not set it. Userspace would then have to do a full hw_free/hw_params recovery, which restores PDE23 as a side effect. I have this patch and will report whether it is sufficient on its own on a tree that carries everything above.
(b) tas2783 restores the SDCA power state itself, in tas2783_sdca_dev_resume() or on the uninitialised re-attach path, when a stream is active. This makes the codec track stream state that ASoC already owns, which I do not much like.
(c) The SoundWire core stops letting a stream continue silently across a peripheral that went UNATTACHED and came back: mark the runtime as needing re-prepare, and fail loudly if it is not.
My own preference is (a), because the promise in the PCM flag is the thing that is actually untrue. But (a) fixes it by making userspace do the right thing, which is not the same as the kernel keeping its own state consistent, so I may well be missing the intended design.
One more question for TI while we are here: given that tas2783 needs an explicit PrepareCtrl write anyway, and has tas_port_prep() for exactly that reason, is simple_ch_prep_sm the right property to declare? Dropping it would put the core back in charge of the write and, more importantly, of the NOT_PREPARED poll -- which would have turned this silent failure into a "Chn prep failed for port 1" error line months ago.
Happy to test patches, and to run the register-level probe again on any variant that is useful.
Thanks,
Andrey
Answering my own question from a few hours ago, since I said I would report back on option (a) and the answer turned out to be more definite than I expected: (a) cannot work, and (b) is enough on its own.
A patch is posted here:
https://lore.kernel.org/all/[email protected]/
(a) does not work, and not by a small margin
============================================
I built a kernel with SNDRV_PCM_INFO_RESUME dropped from the SoundWire DMA PCMs and nothing else changed. Still silent after resume. The register state says why:
DPn_PrepareStatus 0x1 / 0x2 still not prepared ChannelEn now in bank 1 so the stream really was re-prepared PDE23 req=0x3 act=0x3
So dropping the flag did make userspace stop trying to resume the stream in place: the port was prepared afresh, the bank flipped, PrepareCtrl was written. It just did not help, because after resume userspace calls snd_pcm_prepare(), which reaches .prepare and sdw_prepare_stream() and never hw_params() again -- ALSA only requires hw_params() after hw_free(). The Function power-up lives in tas_sdw_hw_params(), so it is skipped on every recovery path, flag or no flag.
That is worth stating plainly because it kills a whole family of fixes:
no PCM capability flag and no change of userspace behaviour can put the power-up back, as long as it sits on a callback that recovery does not re-run.
(b) is enough
=============
The patch powers PDE23 up in tas_port_prep(), in the PRE_PREP case, immediately before the PrepareCtrl write that function already does. The dependency is thereby satisfied on every path that prepares a port, which is what the hardware actually requires.
No ACP change is involved: the tested kernel does not carry the INFO_RESUME patch. Same machine, s2idle for 8 min 51 s with 526 s of S0i3 residency, px13-audio-restore workaround service disabled:
PDE23 req=0x0 act=0x0 DPn_PrepareStatus 0x0 / 0x0 audio after
resume, with no user intervention at all
Two remarks
===========
Niranjan, Shenghao -- the hw_free() path still writes PS3, and with this patch the power-up now happens in port prepare while the power-down stays in hw_free(). That asymmetry works, but if you would rather have the power-down move to the POST_DEPREP case for symmetry, say so and I will respin. I left it alone because hw_free() is also the point where the stream is removed from the slave, and I did not want to change the power-down timing while fixing the power-up.
Pierre-Louis, my question about simple_ch_prep_sm from the previous mail stands, and this result sharpens it: the failure was silent for months only because the core skips the NOT_PREPARED poll for this codec. Had it polled, sdw_prep_deprep_slave_port() would have printed "Chn prep failed for port 1" on the first resume. Whatever happens to this patch, TI may want to reconsider that property.
Thanks,
Andrey