Re: [PATCH v2] i2c: xiic: restore runtime PM teardown in remove to fix clk WARN flood

Hi Abdurraham,

> On Sun, Aug 16, 2026 at 06:14:08PM +0200, Andi Shyti wrote:
> > isn't devm_pm_runtime_set_active_enabled() calling
> > pm_runtime_disable() + pm_runtime_set_suspended() +
> > pm_runtime_dont_use_autosuspend() at teardown?
> 
> It does call all three, but the ordering defeats it, in two places:
> 
> 1. Within the helper's own release, pm_runtime_disable_action() calls
>    pm_runtime_dont_use_autosuspend() *before* pm_runtime_disable() --
>    i.e. while runtime PM is still enabled.
> 
> 2. devres is LIFO, and devm_clk_get_enabled() is registered before
>    devm_pm_runtime_set_active_enabled() in probe(), so the clock's
>    clk_disable_unprepare() release runs *after* the runtime-PM releases.
> 
> pm_runtime_put_sync() in remove() leaves the device RPM_ACTIVE with the
> autosuspend timer armed: rpm_idle() falls through to
> rpm_suspend(RPM_AUTO), the delay hasn't expired, so it just re-arms the
> timer -- runtime_status stays ACTIVE and the clock stays enabled.
> 
> Then the devres teardown runs. pm_runtime_dont_use_autosuspend() clears
> use_autosuspend; update_autosuspend() takes the else branch and calls
> rpm_idle(RPM_AUTO), but now pm_runtime_autosuspend_expiration() returns
> 0, so rpm_suspend() no longer defers and suspends immediately --
> xiic_i2c_runtime_suspend() clk_disable()s the clock, all while runtime
> PM is still enabled. A moment later, further down the LIFO chain,
> devm_clk_get_enabled()'s release does clk_disable_unprepare() on the
> already-disabled clock and clk_core_disable() WARNs, once per device.
> 
> Calling pm_runtime_disable() in remove(), before any devres release
> runs, closes that window: with runtime PM disabled the helper's
> dont_use_autosuspend() can no longer trigger a suspend, so the clock
> stays enabled and the clk release balances it cleanly.
> 
> You're right that set_suspended() and dont_use_autosuspend() are then
> redundant with the helper -- pm_runtime_disable() is the only call that
> has to happen early. So for v3 I'll drop those two and keep just
> pm_runtime_disable() (with a short comment, and the full reasoning in
> the commit log). Does that work for you?

I understand the logic, but this still leaves runtime PM
unbalanced. The explicit pm_runtime_disable() is followed by the
managed cleanup calling pm_runtime_disable() again.

A solution would be to drop the managed runtime-PM helper and
restore the non managed runtime PM setup and teardown. Otherwise,
this ordering issue needs to be addressed in the runtime-PM/clock
framework.

Does that work?

Andi