Skip to contentSkip to conversation

ASoC: tas2783-sdw: drop stale regcache on uninitialized re-attach

PATCH Superseded Andrey Golovko 3 participants 6 messages Jul 27 – Sep 2 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ASoC: tas2783-sdw: drop stale regcache on uninitialized re-attach

When the peripheral re-attaches after the SoundWire controller was power-gated during system suspend (s2idle reaching S0i3 on AMD ACP), the amplifier has lost all of its register and DSP state. tas_update_status() handles that by re-running tas_io_init(), which soft-resets the device and re-downloads the firmware, but before doing so it syncs back a register cache that still holds the pre-suspend values.

That sync is useless, since the soft reset immediately wipes whatever it wrote, and it leaves the cache claiming that the amplifier is already powered up and unmuted. Subsequent read-modify-write updates - DAPM amplifier power-up, SDCA PDE transitions at stream start - then see "no change" and skip the hardware write. Playback runs without a single error while the speakers stay silent. Unbinding and rebinding the driver restores audio, since probe starts from a fresh cache.

Drop the cache instead of syncing it when an uninitialized device attaches, so that later accesses see the real hardware state. regcache_mark_dirty() + regcache_sync() is not an option here: the cache can also hold registers outside the SDCA MBQ map, written during the init sequence, which the MBQ backend refuses to write back. The sync then fails with -EINVAL and takes initialization down with it.

Cached user settings fall back to hardware defaults across such a power loss, which seems clearly preferable to a silent amplifier - the device is being reset and its firmware reloaded at this point anyway.

Tested on an ASUS ProArt PX13 HN7306EAC (AMD Strix Halo, ACP7.0, two TAS2783 amplifiers plus RT721 on SoundWire link 1): the speakers work after an s2idle resume with ~51 s of S0i3 residency, where previously they stayed silent despite a complete firmware re-download.

Fixes 4cc9bd8d7b32 ("ASoc: tas2783A: Add soundwire based codec driver")

---
Based on broonie/sound for-next (asoc-next), i.e. on top of 0d6b2d6f93a6 ("ASoC: codecs: tas2783-sdw: Propagate regcache_sync() errors"), which touches the same call site.

Tested on 7.2-rc4 plus the ACP MSI-on-resume fix 5893013efabb, which is a prerequisite for the peripherals to re-attach at all on this board:
https://lore.kernel.org/all/[email protected]/

 sound/soc/codecs/tas2783-sdw.c | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)
sound/soc/codecs/tas2783-sdw.c +16 -8
index db58c50e8a83..e62470671951 100644
@@ -1216,7 +1216,6 @@ static s32 tas_update_status(struct sdw_slave *slave,
 {
 	struct tas2783_prv *tas_dev = dev_get_drvdata(&slave->dev);
 	struct device *dev = &slave->dev;
-	int ret;
 
 	dev_dbg(dev, "Peripheral status = %s",
 		status == SDW_SLAVE_UNATTACHED ? "unattached" :
@@ -1232,14 +1231,23 @@ static s32 tas_update_status(struct sdw_slave *slave,
 	if (tas_dev->hw_init || tas_dev->status != SDW_SLAVE_ATTACHED)
 		return 0;
 
-	/* updated the cache data to device */
 	regcache_cache_only(tas_dev->regmap, false);
-	ret = regcache_sync(tas_dev->regmap);
-	if (ret) {
-		regcache_cache_only(tas_dev->regmap, true);
-		regcache_mark_dirty(tas_dev->regmap);
-		return ret;
-	}
+
+	/*
+	 * The device is attaching uninitialized: either this is the first
+	 * attach, or it lost power (and with it all register and DSP state)
+	 * while the controller was power-gated during system suspend. The
+	 * cache still holds the pre-suspend values, and tas_io_init() below
+	 * soft-resets the device anyway, so syncing it back is both useless
+	 * and harmful: later read-modify-write updates would compare against
+	 * stale data and skip the hardware write.
+	 *
+	 * Drop the cache instead, so that subsequent accesses see the real
+	 * hardware state. regcache_mark_dirty() + regcache_sync() cannot be
+	 * used here: the cache may hold registers outside the SDCA MBQ map,
+	 * which the MBQ backend refuses to write back.
+	 */
+	regcache_drop_region(tas_dev->regmap, 0, UINT_MAX);
 
 	/* perform I/O transfers required for Slave initialization */
 	return tas_io_init(&slave->dev, slave);
Signature · 1 line
2.53.0
#1 lore raw
Quoted 61 lines · Andrey Golovko

When the peripheral re-attaches after the SoundWire controller was power-gated during system suspend (s2idle reaching S0i3 on AMD ACP), the amplifier has lost all of its register and DSP state. tas_update_status() handles that by re-running tas_io_init(), which soft-resets the device and re-downloads the firmware, but before doing so it syncs back a register cache that still holds the pre-suspend values.

That sync is useless, since the soft reset immediately wipes whatever it wrote, and it leaves the cache claiming that the amplifier is already powered up and unmuted. Subsequent read-modify-write updates - DAPM amplifier power-up, SDCA PDE transitions at stream start - then see "no change" and skip the hardware write. Playback runs without a single error while the speakers stay silent. Unbinding and rebinding the driver restores audio, since probe starts from a fresh cache.

Drop the cache instead of syncing it when an uninitialized device attaches, so that later accesses see the real hardware state. regcache_mark_dirty() + regcache_sync() is not an option here: the cache can also hold registers outside the SDCA MBQ map, written during the init sequence, which the MBQ backend refuses to write back. The sync then fails with -EINVAL and takes initialization down with it.

Cached user settings fall back to hardware defaults across such a power loss, which seems clearly preferable to a silent amplifier - the device is being reset and its firmware reloaded at this point anyway.

Tested on an ASUS ProArt PX13 HN7306EAC (AMD Strix Halo, ACP7.0, two TAS2783 amplifiers plus RT721 on SoundWire link 1): the speakers work after an s2idle resume with ~51 s of S0i3 residency, where previously they stayed silent despite a complete firmware re-download.

Fixes 4cc9bd8d7b32 ("ASoc: tas2783A: Add soundwire based codec driver")

---
Based on broonie/sound for-next (asoc-next), i.e. on top of 0d6b2d6f93a6 ("ASoC: codecs: tas2783-sdw: Propagate regcache_sync() errors"), which touches the same call site.

Tested on 7.2-rc4 plus the ACP MSI-on-resume fix 5893013efabb, which is a prerequisite for the peripherals to re-attach at all on this board:
https://lore.kernel.org/all/[email protected]/

 sound/soc/codecs/tas2783-sdw.c | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)
sound/soc/codecs/tas2783-sdw.c +0 -1
index db58c50e8a83..e62470671951 100644
@@ -1216,7 +1216,6 @@ static s32 tas_update_status(struct sdw_slave *slave,
 {
 	struct tas2783_prv *tas_dev = dev_get_drvdata(&slave->dev);
 	struct device *dev = &slave->dev;
-	int ret;
 
 	dev_dbg(dev, "Peripheral status = %s",
 		status == SDW_SLAVE_UNATTACHED ? "unattached" :
@@ -1232,14 +1231,23 @@ static s32 tas_update_status(struct sdw_slave *slave,
 	if (tas_dev->hw_init || tas_dev->status != SDW_SLAVE_ATTACHED)

sound/soc/codecs/tas2783-sdw.c static s32 tas_update_status(struct sdw_slave *slave, · line 1242

Quoted 10 lines
sound/soc/codecs/tas2783-sdw.c +0 -7
 		return 0;
 
-	/* updated the cache data to device */
 	regcache_cache_only(tas_dev->regmap, false);
-	ret = regcache_sync(tas_dev->regmap);
-	if (ret) {
-		regcache_cache_only(tas_dev->regmap, true);
-		regcache_mark_dirty(tas_dev->regmap);
-		return ret;
-	}

Agree that this sequence didn't make sense, but you have a set of comments below that could be clearer.

sound/soc/codecs/tas2783-sdw.c static s32 tas_update_status(struct sdw_slave *slave, · line 1241

Quoted 7 lines
sound/soc/codecs/tas2783-sdw.c +7 -0
+
+	/*
+	 * The device is attaching uninitialized: either this is the first
+	 * attach, or it lost power (and with it all register and DSP state)
+	 * while the controller was power-gated during system suspend. The
+	 * cache still holds the pre-suspend values, and tas_io_init() below
+	 * soft-resets the device anyway, so syncing it back is both useless

you may want to clarify what 'soft-reset' means. This isn't a SoundWire term, all forms of reset defined in the standard will require re-enumeration. Some devices from Cirrus Logic perform a 'device reset' and a second enumeration, if that was the case here then you could end-up in a boot loop.

sound/soc/codecs/tas2783-sdw.c static s32 tas_update_status(struct sdw_slave *slave, · line 1248

Quoted 7 lines
sound/soc/codecs/tas2783-sdw.c +7 -0
+	 * and harmful: later read-modify-write updates would compare against
+	 * stale data and skip the hardware write.
+	 *
+	 * Drop the cache instead, so that subsequent accesses see the real
+	 * hardware state. regcache_mark_dirty() + regcache_sync() cannot be
+	 * used here: the cache may hold registers outside the SDCA MBQ map,
+	 * which the MBQ backend refuses to write back.

Not following this comment, there's a single cache with specific registers tagged as requiring the MBQ-specific sequence with multiple ordered read/writes. it doesn't matter whether the registers are in the MBQ area and I don't know what the 'MBQ backend' refers to.

Quoted 5 lines
sound/soc/codecs/tas2783-sdw.c +2 -0
+	 */
+	regcache_drop_region(tas_dev->regmap, 0, UINT_MAX);
 
 	/* perform I/O transfers required for Slave initialization */
 	return tas_io_init(&slave->dev, slave);
#2 lore raw

PATCH v2 1/1 ASoC: tas2783-sdw: drop stale regcache on uninitialized re-attach

In reply to Andrey Golovko

When the peripheral re-attaches after the SoundWire controller was power-gated during system suspend (s2idle reaching S0i3 on AMD ACP), the amplifier has lost all of its register and DSP state. tas_update_status() handles that by re-running tas_io_init(), which writes the device's TAS2783_SW_RESET register - a vendor register write that clears the device's register file and DSP state, not a SoundWire reset, so no re-enumeration is involved - and re-downloads the firmware. Before doing any of that, it syncs back a register cache that still holds the pre-suspend values.

That sync is useless, since the reset immediately wipes whatever it wrote, and it leaves the cache claiming that the amplifier is already powered up and unmuted. Subsequent read-modify-write updates - DAPM amplifier power-up, SDCA PDE transitions at stream start - then see "no change" and skip the hardware write. Playback runs without a single error while the speakers stay silent. Unbinding and rebinding the driver restores audio, since probe starts from a fresh cache.

Drop the cache instead of syncing it when an uninitialized device attaches, so that later accesses see the real hardware state.

Reordering the sync after tas_io_init() and marking the cache dirty is not a workable alternative here: tas_regmap has no .writeable_reg, so the cache accepts every register up to .max_register, including ones for which tas2783_sdca_mbq_size() returns 0. regmap_sdw_mbq_size() rejects those with -EINVAL, so the replay fails on the first such register and takes initialization down with it.

Cached user settings fall back to hardware defaults across such a power loss, which seems clearly preferable to a silent amplifier - the device is being reset and its firmware reloaded at this point anyway.

Tested on an ASUS ProArt PX13 HN7306EAC (AMD Strix Halo, ACP7.0, two TAS2783 amplifiers plus RT721 on SoundWire link 1): the speakers work after an s2idle resume with ~51 s of S0i3 residency, where previously they stayed silent despite a complete firmware re-download.

Fixes 4cc9bd8d7b32 ("ASoc: tas2783A: Add soundwire based codec driver")

---
Changes since v1 (thanks Pierre-Louis for the review):
- describe the reset precisely: tas_io_init() writes the vendor TAS2783_SW_RESET register, which is not a SoundWire reset and does not trigger re-enumeration. 'soft reset' is gone from both the commit message and the code comment.
- drop the vague 'MBQ backend' wording. The reason a sync is not usable is concrete: tas_regmap has no .writeable_reg, so the cache takes registers for which tas2783_sdca_mbq_size() returns 0, and regmap_sdw_mbq_size() then returns -EINVAL for them.
- no functional change; the diff is the same modulo the comment.

Based on broonie/sound for-next, on top of 0d6b2d6f93a6 ("ASoC: codecs:
tas2783-sdw: Propagate regcache_sync() errors").

 sound/soc/codecs/tas2783-sdw.c | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)
sound/soc/codecs/tas2783-sdw.c +16 -8
index db58c50e8a83..f96a53a08175 100644
@@ -1216,7 +1216,6 @@ static s32 tas_update_status(struct sdw_slave *slave,
 {
 	struct tas2783_prv *tas_dev = dev_get_drvdata(&slave->dev);
 	struct device *dev = &slave->dev;
-	int ret;
 
 	dev_dbg(dev, "Peripheral status = %s",
 		status == SDW_SLAVE_UNATTACHED ? "unattached" :
@@ -1232,14 +1231,23 @@ static s32 tas_update_status(struct sdw_slave *slave,
 	if (tas_dev->hw_init || tas_dev->status != SDW_SLAVE_ATTACHED)
 		return 0;
 
-	/* updated the cache data to device */
 	regcache_cache_only(tas_dev->regmap, false);
-	ret = regcache_sync(tas_dev->regmap);
-	if (ret) {
-		regcache_cache_only(tas_dev->regmap, true);
-		regcache_mark_dirty(tas_dev->regmap);
-		return ret;
-	}
+
+	/*
+	 * The device is attaching uninitialized: either this is the first
+	 * attach, or it lost power (and with it all register and DSP state)
+	 * while the controller was power-gated during system suspend. The
+	 * cache still holds the pre-suspend values, and tas_io_init() below
+	 * resets the device via TAS2783_SW_RESET anyway, so syncing it back
+	 * is both useless and harmful: later read-modify-write updates would
+	 * compare against stale data and skip the hardware write.
+	 *
+	 * Drop the cache instead, so that subsequent accesses see the real
+	 * hardware state. Syncing after the reset is not an option either:
+	 * the cache accepts registers for which tas2783_sdca_mbq_size()
+	 * returns 0, and writing those back fails with -EINVAL.
+	 */
+	regcache_drop_region(tas_dev->regmap, 0, UINT_MAX);
 
 	/* perform I/O transfers required for Slave initialization */
 	return tas_io_init(&slave->dev, slave);
Signature · 1 line
2.53.0
maintainer
#3 lore raw
Quoted 3 lines · Andrey Golovko

When the peripheral re-attaches after the SoundWire controller was power-gated during system suspend (s2idle reaching S0i3 on AMD ACP), the amplifier has lost all of its register and DSP state. tas_update_status()

Please don't send new patches in reply to old patches or serieses, this makes it harder for both people and tools to understand what is going on - it can bury things in mailboxes and make it difficult to keep track of what current patches are, both for the new patches and the old ones.

maintainer
#4 lore raw
In reply to Andrey Golovko
Quoted 1 line · Andrey Golovko

ASoC: tas2783-sdw: drop stale regcache on uninitialized re-attach

Applied to

https://git.kernel.org/pub/scm/linux/kernel/git/broonie/sound.git for-7.3

Thanks!

[1/1] ASoC: tas2783-sdw: drop stale regcache on uninitialized re-attach

      https://git.kernel.org/broonie/sound/c/b627da430357

All being well this means that it will be integrated into the linux-next tree (usually sometime in the next 24 hours) and sent to Linus during the next merge window (or sooner if it is a bug fix), however if problems are discovered then the patch may be dropped or reverted.

You may get further e-mails resulting from automated or manual testing and review of the tree, please engage with people reporting problems and send followup patches addressing any issues that are reported if needed.

If any updates are required or you are submitting further changes they should be sent as incremental updates against current git, existing patches will not be replaced.

Please add any relevant lists and maintainers to the CCs when replying to this mail.

Thanks,
Mark

#6 lore raw
In reply to a message not in the archive
Quoted 2 lines · Baojun

Based on my test, this modification is also needed in tas2783_sdca_dev_resume().

(Restoring linux-sound and the rest of the Cc list, since the patch was posted there.)

Thank you for looking at it. I agree the same sync is a problem there, and the ordering makes that unambiguous: sdw_handle_slave_status() calls the driver's update_status() callback - where the cache is now dropped and tas_io_init() re-downloads the firmware - and only afterwards does complete_all(&slave->initialization_complete). That completion is what sdw_slave_wait_for_init() at the top of tas2783_sdca_dev_resume() is waiting for, so by the time regcache_sync() runs the part has already been soft-reset and re-initialized. Syncing there writes the cache back onto a device that has just been brought up from scratch.

What I would not do is drop unconditionally in dev_resume(), because that function is also the RUNTIME_PM_OPS resume callback. On a resume where the peripheral kept its context, or came back through clock stop without losing state, regcache_sync() is the only thing that restores the user's settings, and dropping the cache there would silently reset volume and mute on every runtime resume.

So the shape I have in mind is a flag rather than a second drop: set it in tas_update_status() on the uninitialized-attach path where the cache is dropped today, consume it in dev_resume(), and simply skip the sync when it is set - after tas_io_init() the cache already mirrors the hardware, so there is nothing worth syncing, only registers that can fail. Something like:

	if (test_and_clear_bit(TAS_REINIT, &tas_dev->flags))
		return 0;	/* re-initialized from scratch, cache is fresh */

	regcache_cache_only(tas_dev->regmap, false);
	ret = regcache_sync(tas_dev->regmap);

Would you prefer that, or do you have a different fix in progress on your side? I am happy to write and test it either way - I just do not want us to post two versions of the same thing.

Before I write the changelog, though, I need to describe a failure I can actually point at, and this is where I have to ask what you saw. On the board I have here - ASUS ProArt PX13, AMD ACP7.0, two TAS2783 plus RT721 on link 1 - I cannot reproduce a failure on that path:

- the codec never reaches runtime suspend at all (runtime_status stays active, the usage count never drops to zero), so the runtime resume path is not exercised;

- on the system resume path, across s2idle cycles where both amplifiers genuinely lose power, re-attach and re-download the firmware, the journal shows no resume error at all. My reading is that regcache_sync() returns early: regcache_cache_only(true) does not by itself set cache_dirty, and if nothing writes through the cache while the device is suspended, sync takes the "if (!map->cache_dirty) goto out" exit and never touches the bus. Which would mean the bug is latent here and armed only when something does dirty the cache during suspend.

That it is armed at all is easy to show: when I force the sync on this part (through a small debug module, outside of any suspend), it aborts at 0x40400108, FU23 Mute ch0, with -ENODATA - reg_defaults claims 0x1 while the init sequence writes 0x00, so sync tries to "restore" a value the device refuses. A dev_resume() that reaches the sync on this hardware would therefore fail outright and return -ENODATA to the PM core, not merely leave the amplifier stale. That is one more instance of the reg_defaults question in my other mail of 24 August [1].

So could you tell me a bit more about your test:

1. Which resume path - runtime resume, or system resume from s2idle/S3?
2. Which tree, and does it already contain b627da430357 in update_status()?
3. What did you observe - a sync error code in the log, or silent speakers with no error at all?
4. Does your board's controller power-gate the link across suspend, so the amplifiers re-attach uninitialized, or do they keep context?

With that I can write the patch against a failure that is described rather than assumed, and test it here by forcing the cache dirty across suspend.

[1] https://lore.kernel.org/linux-sound/[email protected]/

Thanks,
Andrey