[PATCH v2 0/2] iio: magnetometer: fix the register restore on resume in mmc35240 and mmc5633

[PATCH v2 0/2] iio: magnetometer: fix the register restore on resume in mmc35240 and mmc5633

[Shashank Mohan Jain]

The resume callbacks of mmc35240 and mmc5633 call
regcache_sync_region() before they take the regmap out of cache-only
mode. Since v7.3-rc1 regmap refuses a sync in cache-only mode with a
WARN and -EINVAL (86fc3483a4d6 and afc8e3ee8307), so both drivers now
warn on every resume and no longer restore their control registers.

The kernel-doc added by 86fc3483a4d6 says that the caller has to
disable cache-only mode before the sync, so the two patches change the
order in the drivers.

The Fixes tags name the commit that added each resume callback
(553a776b7908 for mmc35240, 6e5f6bf2e3f0 for mmc5633), as suggested by
Bui Duc Phuc in review of v1: the wrong order predates the regcache
change (86fc3483a4d6 / afc8e3ee8307), which only turned it into a WARN
and a failed restore.

I have no hardware for either chip. The patches were tested in qemu
with the drivers bound to an i2c-stub device and a suspend/resume
cycle through /sys/power/pm_test; the details are in the changelogs.
A scan of drivers/ and sound/ for other resume paths which sync before
leaving cache-only mode in the same function found only these two.

Based on mainline (a5ebb76233b7, v7.3-rc6+); the patches also apply
to next-20261009.

Changes since v1:
 - repoint the Fixes tags to the commits that added the resume
   callbacks (553a776b7908, 6e5f6bf2e3f0), per Bui Duc Phuc's review;
   the code change is unchanged.
 - note in each changelog that the ordering predates the regcache
   change that exposed it.

Shashank Mohan Jain (2):
  iio: magnetometer: mmc35240: leave cache-only mode before the cache
    sync
  iio: magnetometer: mmc5633: leave cache-only mode before the cache
    sync

 drivers/iio/magnetometer/mmc35240.c | 3 +--
 drivers/iio/magnetometer/mmc5633.c  | 3 +--
 2 files changed, 2 insertions(+), 4 deletions(-)


base-commit: a5ebb76233b79db01e82e061173cb73c6d2b5c6b
-- 
2.43.0

[PATCH v2 1/2] iio: magnetometer: mmc35240: leave cache-only mode before the cache sync

[Shashank Mohan Jain]

mmc35240_resume() marks the register cache dirty and calls
regcache_sync_region() while the regmap is still in cache-only mode.
It calls regcache_cache_only(false) only after the sync. This
ordering has been wrong since the resume support was added; it happened
to work until regmap started rejecting a sync in cache-only mode.

Since commit afc8e3ee8307 ("regcache: Validate cache_only state in
regcache_sync_region()") a sync in cache-only mode is refused with a
WARN and -EINVAL. Every resume now gives

  WARNING: drivers/base/regmap/regcache.c:536 at regcache_sync_region+0x1a4/0x1c0
  ...
   mmc35240_resume+0x30/0x70
   dpm_run_callback+0x7d/0x1c0
   device_resume+0x175/0x270
  ...
  mmc35240 1-0030: Failed to restore control registers

and the control registers are not written. If the chip lost its state
during suspend, it runs with the reset value of CTRL1 while the cache,
and with it the sampling frequency attribute, still has the configured
value. Writing the same frequency again does not repair this, because
regmap_update_bits() finds no difference to the cache.

Leave cache-only mode first, then mark the cache dirty and sync it.

Reproduced in qemu with the driver bound to an i2c-stub device: set
in_magn_sampling_frequency to 25 (CTRL1 = 0x02), clear CTRL0 and CTRL1
of the stub behind the driver's back to emulate the power loss, and run
a suspend/resume cycle with "devices" in /sys/power/pm_test. Without
this change CTRL1 is still 0x00 after resume and the warning is
printed. With it, and also with the regcache commit reverted, CTRL1 is
0x02 again and there is no warning. Not tested on real hardware.

Fixes: 553a776b7908 ("iio: magnetometer: mmc35240: Add PM sleep support")
Assisted-by: LLM
Signed-off-by: Shashank Mohan Jain <jain.sm@gmail.com>
---
Prepared with Claude Code (Anthropic), model Claude Opus 5.5
(claude-opus-5-5).

 drivers/iio/magnetometer/mmc35240.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/drivers/iio/magnetometer/mmc35240.c b/drivers/iio/magnetometer/mmc35240.c
index 6f40cba4b1..7f028d7f1f 100644
--- a/drivers/iio/magnetometer/mmc35240.c
+++ b/drivers/iio/magnetometer/mmc35240.c
@@ -532,14 +532,13 @@ static int mmc35240_resume(struct device *dev)
 	struct mmc35240_data *data = iio_priv(indio_dev);
 	int ret;
 
+	regcache_cache_only(data->regmap, false);
 	regcache_mark_dirty(data->regmap);
 	ret = regcache_sync_region(data->regmap, MMC35240_REG_CTRL0,
 				   MMC35240_REG_CTRL1);
 	if (ret < 0)
 		dev_err(dev, "Failed to restore control registers\n");
 
-	regcache_cache_only(data->regmap, false);
-
 	return 0;
 }
 
-- 
2.43.0

[PATCH v2 2/2] iio: magnetometer: mmc5633: leave cache-only mode before the cache sync

[Shashank Mohan Jain]

mmc5633_resume() marks the register cache dirty and calls
regcache_sync_region() while the regmap is still in cache-only mode.
It calls regcache_cache_only(false) only after the sync. This
ordering has been wrong since the driver was added; it happened to work
until regmap started rejecting a sync in cache-only mode.

Since commit afc8e3ee8307 ("regcache: Validate cache_only state in
regcache_sync_region()") a sync in cache-only mode is refused with a
WARN and -EINVAL. Every resume now gives

  WARNING: drivers/base/regmap/regcache.c:536 at regcache_sync_region+0x1a4/0x1c0
  ...
   mmc5633_resume+0x2d/0x70
   dpm_run_callback+0x7d/0x1c0
   device_resume+0x175/0x270
  ...
  mmc5633_i2c 1-0030: Failed to restore control registers

and the control registers are not written. If the chip lost its state
during suspend, it runs with the reset value of CTRL1 while the cache,
and with it the sampling frequency attribute, still has the configured
value. Writing the same frequency again does not repair this, because
regmap_update_bits() finds no difference to the cache.

Leave cache-only mode first, then mark the cache dirty and sync it.

Reproduced in qemu with the I2C driver bound to an i2c-stub device: set
in_magn_sampling_frequency to 3.5 (CTRL1 = 0x02), clear CTRL0 and CTRL1
of the stub behind the driver's back to emulate the power loss, and run
a suspend/resume cycle with "devices" in /sys/power/pm_test. Without
this change CTRL1 is still 0x00 after resume and the warning is
printed. With it, and also with the regcache commit reverted, CTRL1 is
0x02 again and there is no warning. Not tested on real hardware.

Fixes: 6e5f6bf2e3f0 ("iio: magnetometer: Add mmc5633 sensor")
Assisted-by: LLM
Signed-off-by: Shashank Mohan Jain <jain.sm@gmail.com>
---
Prepared with Claude Code (Anthropic), model Claude Opus 5.5
(claude-opus-5-5).

 drivers/iio/magnetometer/mmc5633.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/drivers/iio/magnetometer/mmc5633.c b/drivers/iio/magnetometer/mmc5633.c
index f4dae7ba13..b651f01e1b 100644
--- a/drivers/iio/magnetometer/mmc5633.c
+++ b/drivers/iio/magnetometer/mmc5633.c
@@ -499,13 +499,12 @@ static int mmc5633_resume(struct device *dev)
 	struct regmap *regmap = dev_get_regmap(dev, NULL);
 	int ret;
 
+	regcache_cache_only(regmap, false);
 	regcache_mark_dirty(regmap);
 	ret = regcache_sync_region(regmap, MMC5633_REG_CTRL0, MMC5633_REG_CTRL1);
 	if (ret)
 		dev_err(dev, "Failed to restore control registers\n");
 
-	regcache_cache_only(regmap, false);
-
 	return 0;
 }
 
-- 
2.43.0