[dnm] ASoC: SOF: pcm/pm/Intel: Fix VoW during system suspend

[ujfalusi]

The Wake on Voice flow was broken (supported via IPC3 only atm) because on suspend we attempted to tear down the pipelines
and RESUME was not supported by the PCM:
We had errors on suspend due to failing to free widgets and the user space restarted to capture during resume.

To fix this:
If the target suspend level is SOF_DSP_PM_D0 then we must not tear down the pipelines
we need to set the SNDRV_PCM_INFO_RESUME for the VoW capture PCM, so applications can do the 'resume'
and on RESUME trigger we do nothing as the DSP was left on and everything has been left running as they were before.

Tested on sof-adl-max98357a-rt5682.tplg with:

arecord -Dhw:0,100 -M -N -c 2 -f S16_LE -r 16000 --buffer-size=96000 tmp.wav -d 10 -vvv

and

echo mem > /sys/power/state

then clapping to wake the device up: no errors observed anymore.

[Copilot]

Copilot review overview

🟡 Changes recommended

Resolve the HDA RESUME handling and restrict or implement resume support for non-retained suspend paths.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1

Open (1)

• Advertised RESUME fails after S3/D3 suspend · New

What changed in this PR

Fixes SOF Wake on Voice suspend/resume by retaining pipelines during D0 suspend and enabling ALSA resume support.

Changes:

• Avoids pipeline teardown when suspending to D0.
• Handles RESUME for retained streams.
• Advertises RESUME for compatible HDA capture streams.

File	Summary	Review status
sound/​soc/​sof/​pm.c	Preserves pipelines during D0 suspend.	No issue noted.
sound/​soc/​sof/​pcm.c	Handles RESUME for suspend-ignored streams.	Moderate issue: the HDA DAI path can still return -EINVAL for RESUME.
sound/​soc/​sof/​intel/​hda-pcm.c	Advertises resume capability for VoW capture streams.	Moderate issue: capability is advertised for paths that do not reliably support RESUME.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[ujfalusi]

Changes since v1:

• incorporate ASoC: SOF: Bring WoV support to ipc4 #5878
• update the definition of WoV stream: capture+CAPTURE_COMPATIBLE_D0I3+not deep buffer

[Copilot]

Copilot review overview

🟡 Changes recommended

HDA DAI resume handling is missing and can return -EINVAL on VoW resume.

Review effort: Lite
Findings: 2

Open (2)

• Comment misspells identified · New
• Comment misspells identified · New

Resolved since last review (1)

• Advertised RESUME fails after S3/D3 suspend

[lgirdwood]

Testing/updates/comments from Gemini. @ujfalusi some may be needed, tested both with ALSA and tinyalsa, depend on how you plan to upstream.

Tested PR #5951 on Intel Panther Lake (PTL / ACE 3.0 DSP) using our Wake-on-Voice test suites (10-run S0 sequence, 10-run S2idle sleep/wake sequence, and audio capture validation suite).

The general direction and cleanup of WoV stream definition in PR #5951 is great, but when tested on IPC4 hardware, several pieces are missing to make IPC4 WoV function across system suspend and wake.

Issues Identified on IPC4 with PR #5951 Standalone:

1. Missing IPC4 D0i3 Topology Token Parsing (ipc4-topology.c):
   Commit 2 (036dbd7ff63e) defines a WoV stream as:

   substream->stream == SNDRV_PCM_STREAM_CAPTURE &&
   spcm->stream[substream->stream].d0i3_compatible &&
   spcm->stream[substream->stream].dsp_max_burst_size_in_ms <= 1

   However, on upstream topic/sof-dev, ipc4-topology.c never sets d0i3_compatible (only IPC3 topology.c parses SOF_TKN_STREAM_CAPTURE_COMPATIBLE_D0I3). The topology parsing from PR ASoC: SOF: Bring WoV support to ipc4 #5878 was omitted in PR [dnm] ASoC: SOF: pcm/pm/Intel: Fix VoW during system suspend #5951. Because d0i3_compatible is always false on IPC4:

   • snd_sof_is_vow_stream() always returns false.
   • runtime->hw.info |= SNDRV_PCM_INFO_RESUME is never set on IPC4.
   • spcm->stream[direction].suspend_ignored is never set, so ALSA core suspends the WoV stream into SNDRV_PCM_STATE_SUSPENDED, returning -EBADFD (-86) to userspace on resume.

2. DAPM DAI Link Suspend Exemption (ignore_suspend):
   While suspend_ignored prevents sof_pcm_trigger() from stopping the stream, generic ASoC DAPM suspend power-down sequencing still attempts to tear down the DAI link widgets while the DSP pipeline remains alive. Setting rtd->dai_link->ignore_suspend = 1 in hda_dsp_pcm_open() (and clearing on close) prevents this host/DSP widget desync and eliminates IPC timeouts on resume.

3. DSP Power Target Retention in S0ix (pm.c):
   snd_sof_dsp_power_target() in pm.c only retains the DSP in D0 if snd_sof_stream_suspend_ignored(sdev) is true. If userspace has the WoV PCM open but it has not taken a suspend trigger, target_state evaluates to SOF_DSP_PM_D3, causing ctx_save IPC to be sent to a running DSP which returns -EBUSY (-16). Adding snd_sof_dsp_only_d0i3_compatible_stream_active(sdev) ensures the DSP stays in D0 whenever an open WoV stream is active.

4. Unhandled Phrase Detected Notification (SOF_IPC4_NOTIFY_PHRASE_DETECTED):
   When the DSP detects a wake phrase (or test trigger), firmware emits SOF_IPC4_NOTIFY_PHRASE_DETECTED (0x1b040000). In ipc4.c, this must be handled to call snd_sof_pcm_period_elapsed(), otherwise blocking reads running with NO_PERIOD_WAKEUP (such as tinyalsa and alsa-lib set_period_wakeup(0)) never receive avail wakeups.

5. Pipeline State Reset on STOP/SUSPEND (hda-dai-ops.c):
   Resetting pipeline->state = SOF_IPC4_PIPE_RESET on STOP/SUSPEND trigger ensures subsequent START cycles cleanly transition through PAUSED to RUNNING.

6. Multi-Pin Process Module Format Fallback (ipc4-topology.c):
   In multi-pin topologies (e.g. ECNS modules with 1ch out on pin 0 / pin 1), sof_ipc4_init_output_audio_fmt() needs a fallback to Pin 0 format when no reference format matches channel counts, otherwise widget preparation fails with -22 (-EINVAL).

---

Working Reference Branch:

I have pushed a working branch rebased directly on topic/sof-dev on top of PR #5951 with these additions:
👉 https://github.com/lgirdwood/linux/tree/wov-pr5951-ipc4-reference

With this branch on Panther Lake (PTL):

• 10 / 10 WoV S0 runs passed (~5.65s elapsed per run)
• 10 / 10 WoV S2idle sleep/wake cycles passed (all 10 woke the host via DSP IRQ 193 and read 4000 frames)
• 12 / 12 audio capture validation tests passed (raw DMIC, ECNS, and WoV slots 0/1/2 across arecord, tinycap, and MMAP/NOIRQ)
• tinyalsa blocking read (wov_blocking_read_tinyalsa) passed

[lgirdwood]

For additional context, here is the empirical pass vs fail breakdown from testing PR #5951 standalone on Panther Lake (PTL / ACE 3.0 DSP) before applying the IPC4 enablement delta:

What Passed with PR #5951 Standalone:

• Raw DMIC Capture (PCM 10):
  • arecord, tinycap, and tinycap -M (MMAP) all passed 100% (4ch 16kHz audio captured cleanly without any xruns or errors).
• Initial S0 WoV Capture (Runs 1–5):
  • While the host remained in S0, early runs of wov_blocking_read and wov_blocking_read_tinyalsa (using PCM_MMAP | PCM_NOIRQ) passed, capturing 4,000 frames after ~5.65s and 5.45s upon synthetic wake trigger.
  • NO_PERIOD_WAKEUP was successfully negotiated because sof_hda_common_ops already sets SNDRV_PCM_INFO_NO_PERIOD_WAKEUP in the base kernel.

---

What Failed with PR #5951 Standalone:

• WoV S2idle Suspend & Wake (run_10_wov_sleep_wake.sh) — 0 / 10 Passed (100% Failure):
  • On every suspend attempt, wov_blocking_read immediately failed on resume with wait error or timeout: -86 (-EBADFD / stream suspended by the ALSA core) because without d0i3_compatible parsed from IPC4 topology, snd_sof_is_vow_stream() returned false and the stream was not exempted from suspend.
  • The DSP crashed during suspend/resume with a register and stack dump (sof-audio-pci-intel-ptl 0000:00:1f.3: stack dump from 0x00000000).
  • On resumption, DAPM attempted to power down widgets that were supposed to remain active, causing module unbind IPC timeouts (-19 / -ENODEV).
• S0 WoV Longevity (Runs 6–10):
  • Later S0 runs began timing out after 15s (wait error or timeout: 0).
  • Because SOF_IPC4_NOTIFY_PHRASE_DETECTED (0x1b040000) is ignored by PR [dnm] ASoC: SOF: pcm/pm/Intel: Fix VoW during system suspend #5951, the kernel never calls snd_sof_pcm_period_elapsed(). Without periodic IRQs, userspace snd_pcm_wait() never woke up once the DMA buffer pacing desynced.
• Topologies with Multi-Pin Process Modules:
  • Topologies containing modules that change channel counts between pins (e.g. ECNS with 1ch out on pin 0 / pin 1 in sof-ptl-dmic-wov-multi-4ch.tplg) failed hw_params with -22 (-EINVAL) in sof_ipc4_init_output_audio_fmt() without the Pin 0 format fallback.

---

Summary:

In S0 (awake), basic capture and initial WoV reads function. However, actual Wake-on-Voice from system sleep (s2idle) fails 100% on IPC4 under PR #5951 standalone because d0i3_compatible is never set, so the suspend bypass and DAPM retention never take effect.