Re: [PATCH v3] PM: sleep: Use complete() in device_pm_sleep_init() and skip no_pm devices in dpm_wait()

[Rafael J. Wysocki]

On Sat, May 23, 2026 at 4:23 AM Jiakai Xu <xujiakai24@xxxxxxxxxxxxxxxx> wrote:
>
> Replace complete_all() with complete() in device_pm_sleep_init() to allow
> it to be called in atomic contexts without triggering a false-positive
> WARNING from lockdep_assert_RT_in_threaded_ctx() when
> CONFIG_PROVE_RAW_LOCK_NESTING is enabled.
>
> device_pm_sleep_init() may be called during device initialization while
> holding a raw_spinlock (e.g., from within device_initialize()), and
> complete_all() is unsafe in atomic contexts on PREEMPT_RT kernels.
> complete(), which is safe to call from any context, is sufficient here.
>
> complete_all() sets the completion count to UINT_MAX/2 (permanently
> signaled), while complete() increments it by 1. Since no threads can be
> waiting during device initialization, both are functionally equivalent.
> The completion is always reinitialized via reinit_completion() in
> dpm_clear_async_state() before each suspend/resume cycle.
>
> However, changing to complete() introduces a potential deadlock for
> devices with no PM support (dev->power.no_pm = true). Such devices are
> never added to the dpm_list and never go through dpm_clear_async_state(),
> so their completion is never reinitialized. A parent device waiting on a
> no_pm child across multiple suspend phases would consume the single-use
> token in the first phase and block forever in the second.
>
> Fix this by adding an early return in dpm_wait() when dev->power.no_pm is
> set, since no_pm devices do not participate in system suspend/resume.
>
> Fixes: 152e1d592071 ("PM: Prevent waiting forever on asynchronous resume after failing suspend")
> Signed-off-by: Jiakai Xu <xujiakai24@xxxxxxxxxxxxxxxx>
> ---
> V1 -> V2:
> - Updated dpm_wait() to skip devices with no PM support,
>   as suggested by Rafael J. Wysocki.
>
> V2 -> V3:
> - Revised commit message to clarify that the lockdep warning originates
>   from device_pm_sleep_init() called in atomic context, not from
>   device_pm_remove().  No code changes relative to v2.
> ---
>  drivers/base/power/main.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
> index e1b550664bab..ed48c292f575 100644
> --- a/drivers/base/power/main.c
> +++ b/drivers/base/power/main.c
> @@ -115,7 +115,7 @@ void device_pm_sleep_init(struct device *dev)
>         dev->power.is_noirq_suspended = false;
>         dev->power.is_late_suspended = false;
>         init_completion(&dev->power.completion);
> -       complete_all(&dev->power.completion);
> +       complete(&dev->power.completion);
>         dev->power.wakeup = NULL;
>         INIT_LIST_HEAD(&dev->power.entry);
>  }
> @@ -252,6 +252,10 @@ static void dpm_wait(struct device *dev, bool async)
>         if (!dev)
>                 return;
>
> +       /* Devices with no PM support don't use the completion. */
> +       if (dev->power.no_pm)
> +               return;
> +
>         if (async || (pm_async_enabled && dev->power.async_suspend))
>                 wait_for_completion(&dev->power.completion);
>  }
> --

Applied as 7.2 material, thanks!