]> git.ipfire.org Git - thirdparty/linux.git/commitdiff
PM: sleep: Use complete() in device_pm_sleep_init()
authorJiakai Xu <xujiakai24@mails.ucas.ac.cn>
Sat, 23 May 2026 02:23:14 +0000 (02:23 +0000)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Tue, 26 May 2026 11:02:34 +0000 (13:02 +0200)
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@mails.ucas.ac.cn>
[ rjw: Subject adjustment ]
Link: https://patch.msgid.link/20260523022314.2657232-1-xujiakai24@mails.ucas.ac.cn
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/base/power/main.c

index e1b550664babbf245c7280385e513012bd99f6cd..ed48c292f57576c2ab8637c56039de89ae67504b 100644 (file)
@@ -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);
 }