On 9/6/23 19:37, Charles Keepax wrote: > On Tue, Sep 05, 2023 at 10:15:46PM +0300, Cristian Ciocaltea wrote: >> On 9/5/23 12:45, Charles Keepax wrote: >>> On Sun, Sep 03, 2023 at 12:06:21AM +0300, Cristian Ciocaltea wrote: >>>> Simplify runtime PM during probe by converting pm_runtime_enable() to >>>> the managed version. >>>> >>>> Signed-off-by: Cristian Ciocaltea <cristian.ciocaltea@xxxxxxxxxxxxx> >>>> --- >>>> @@ -1376,7 +1379,6 @@ void cs35l41_remove(struct cs35l41_private *cs35l41) >>>> cancel_work_sync(&cs35l41->mdsync_up_work); >>>> >>>> pm_runtime_get_sync(cs35l41->dev); >>>> - pm_runtime_disable(cs35l41->dev); >>>> >>>> regmap_write(cs35l41->regmap, CS35L41_IRQ1_MASK1, 0xFFFFFFFF); >>>> if (cs35l41->hw_cfg.bst_type == CS35L41_SHD_BOOST_PASS || >>> >>> Are we sure this is safe? The remove handler appears to be >>> written to disable pm_runtime at the start presumably to stop the >>> resume/suspend handler running during the remove callback. >>> Whereas after this change the pm_runtime isn't disabled until >>> after the remove callback has run. Does this open a window were >>> we could get an erroneous pm_runtime suspend after the >>> pm_runtime_put_noidle? >> >> I've just made a test adding a 6s sleep before returning from the remove >> handler: >> >> [14444.894316] cs35l41 spi-VLV1776:00: Runtime resume >> [14444.894469] cs35l41 spi-VLV1776:00: sleep 6s before return of cs35l41_remove() >> [14448.338994] cs35l41 spi-VLV1776:00: Runtime suspend >> [14451.079649] cs35l41 spi-VLV1776:00: return from cs35l41_remove() >> [14451.080129] cs35l41 spi-VLV1776:00: Runtime resume >> [14451.080165] cs35l41 spi-VLV1776:00: ASoC: Unregistered DAI 'cs35l41-pcm' >> [14451.080181] cs35l41 spi-VLV1776:00: Runtime suspend >> [14451.813639] acp5x_i2s_playcap acp5x_i2s_playcap.0: ASoC: Unregistered DAI 'acp5x_i2s_playcap.0' >> >> As expected, suspend triggered, but a resume was issued later, before DAI >> got unregistered. >> >> I didn't notice any issues while repeating the test several times, hence >> I wonder what would be the reason to prevent getting suspend/resume events >> at this point? > > The enter/exit hibernate code might run, which at the very > least might result in a bunch of unexpected and failing bus > traffic. Having a bit of a poke through the code, I guess the > most dangerous thing would if you actually got as far as an > extra runtime resume. This might cause cs35l41_init_boost > to run which would undo the work done by the call to > cs35l41_safe_reset in remove, which could leave the boost in a > dangerous state when we enable reset/power down the supplies, > which I think was not considered good. But its just likely > simpler/cleaner if we don't have to think about all the > possible implications of such things by just not allowing > it to happen. Agree, let's keep it simple. I will revert the change and instead ensure a proper cleanup of pm_runtime_use_autosuspend(), according to the documentation: "It's important to undo this with pm_runtime_dont_use_autosuspend() at driver exit time unless your driver initially enabled pm_runtime with devm_pm_runtime_enable() (which handles it for you)." Thanks for the clarifications, Cristian
