Linux Power Management development
 help / color / mirror / Atom feed
From: Brian Norris <briannorris@chromium.org>
To: "Rafael J. Wysocki" <rafael@kernel.org>
Cc: Linux PM <linux-pm@vger.kernel.org>,
	Alan Stern <stern@rowland.harvard.edu>,
	Sakari Ailus <sakari.ailus@linux.intel.com>,
	Ulf Hansson <ulf.hansson@oss.qualcomm.com>,
	Linux Driver Core Development <driver-core@lists.linux.dev>
Subject: Re: [PATCH v1 1/4] PM: runtime: Update last_busy in all rpm_idle() paths
Date: Fri, 20 Nov 2026 12:59:33 -0800	[thread overview]
Message-ID: <awC0tcqgq2KvCyxO@google.com> (raw)
In-Reply-To: <3076901.e9J7NaK4W3@rafael.j.wysocki>

On Thu, Nov 19, 2026 at 09:16:30PM +0100, Rafael J. Wysocki wrote:
> From: "Rafael J. Wysocki" <rafael.j.wysocki@intel.com>
> 
> The changelog of commit 26236a3e8803 ("PM: runtime: Update last_busy
> before calling rpm_suspend() in rpm_idle()") stated that rpm_idle() only
> needed to update last_busy when it was about to call rpm_suspend(), but
> that is incorrect.  There are still cases in which a device may be
> autosuspended earlier than expected.
> 
> The first one is when pm_runtime_put() is called while a leftover
> autosuspend timer is running.  Then, dropping the device's runtime PM
> usage counter to 0 allows rpm_suspend() triggered by the timer to
> suspend the device immediately before __pm_runtime_idle() is able to
> grab the spinlock.  Prevent it from occurring by adding a last_busy
> update before the rpm_drop_usage_count() call in __pm_runtime_idle().

Do consumer->supplier and child->parent refcount drops have the same
problem? e.g., in pm_runtime_release_supplier() and rpm_suspend() /
parent->power.child_count decrement?

> Another one is when rpm_idle() returns without calling rpm_suspend(),
> for example because the callback has returned a transient error, while
> an autosuspend timer is running.  The timer may then suspend the device
> without respecting the autosuspend delay.  Moreover, the timer may fire
> while the callback is running and it will not see the new last_busy
> value in that case.  To address these cases, move the existing
> pm_runtime_mark_last_busy() call in rpm_idle() before the first return
> value check.

I suppose we don't care about: "what if .runtime_idle() is slow?" --
that makes a small delta in the expiration time.

> In both places mentioned above add comments explaining the rationale for
> updating last_busy.
> 
> Fixes: 26236a3e8803 ("PM: runtime: Update last_busy before calling rpm_suspend() in rpm_idle()")
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> ---
> 
> Commit 26236a3e8803 is present in linux-next.

Semi-unrelated cleanup: pm_runtime_put_sync() some stale kerneldoc now,
saying "this does not update the last access time".

> ---
>  drivers/base/power/runtime.c | 16 +++++++++++++++-
>  1 file changed, 15 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
> index 0b15c718c1bb..02e7b4953c38 100644
> --- a/drivers/base/power/runtime.c
> +++ b/drivers/base/power/runtime.c
> @@ -518,6 +518,14 @@ static int rpm_idle(struct device *dev, int rpmflags)
>  	else if (dev->power.idle_notification)
>  		retval = -EINPROGRESS;
>  
> +	/*
> +	 * Update last_busy to defer autosuspend by the full autosuspend delay.
> +	 * Since an autosuspend timer may be running and the new last_busy value
> +	 * needs to be seen by it, update last_busy before returning an error or
> +	 * dropping the spinlock and allowing the timer function to grab it.
> +	 */
> +	pm_runtime_mark_last_busy(dev);
> +
>  	if (retval)
>  		goto out;
>  
> @@ -563,7 +571,6 @@ static int rpm_idle(struct device *dev, int rpmflags)
>  	if (retval)
>  		return retval;
>  
> -	pm_runtime_mark_last_busy(dev);
>  	return rpm_suspend(dev, rpmflags | RPM_AUTO);
>  }
>  
> @@ -1132,6 +1139,13 @@ int __pm_runtime_idle(struct device *dev, int rpmflags)
>  	int retval;
>  
>  	if (rpmflags & RPM_GET_PUT) {
> +		/*
> +		 * Once the usage counter drops to 0, a leftover autosuspend
> +		 * timer may fire before the spinlock below is acquired.  Update
> +		 * last_busy first so it does not find the expiration time in
> +		 * the past and suspend the device immediately.
> +		 */
> +		pm_runtime_mark_last_busy(dev);
>  		retval = rpm_drop_usage_count(dev);

This is somewhat of a pre-existing issue (e.g., in
pm_runtime_put_autosuspend()), but weakly ordered architectures don't
really enforce the dependencies you want.

rpm_drop_usage_count() uses an atomic decrement, which has a full
barrier. But this is supposed to influence the rpm_suspend() behavior of
rpm_check_suspend_allowed() (atomic_read) + READ_ONCE(last_busy), which
includes only a conditional dependency and no barrier.

https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0124r6.html

  "Control dependencies can order prior loads against later dependent
  stores, however, they do not order prior loads against later dependent
  loads. (Use memory_order_consume or memory_order_acquire if you
  require this behavior.)"

So, rpm_check_suspend_allowed() should probably use
atomic_read_acquire(&dev->power.usage_count).

(This is probably an extremely narrow, practically impossible issue.)

Brian

>  		if (retval < 0) {
>  			return retval;
> -- 
> 2.51.0
> 
> 
> 
> 

  reply	other threads:[~2026-11-20 20:59 UTC|newest]

Thread overview: 11+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-19 20:13 [PATCH v1 0/4] PM: runtime: Close some gaps related to autosuspend Rafael J. Wysocki
2026-11-19 20:16 ` [PATCH v1 1/4] PM: runtime: Update last_busy in all rpm_idle() paths Rafael J. Wysocki
2026-11-20 20:59   ` Brian Norris [this message]
2026-11-19 20:17 ` [PATCH v1 2/4] PM: runtime: Drop space characters preceding labels Rafael J. Wysocki
2026-11-20  2:47   ` Alan Stern
2026-11-20 12:11     ` Rafael J. Wysocki (Intel)
2026-11-20 12:30       ` Rafael J. Wysocki (Intel)
2026-11-20 16:11         ` Alan Stern
2026-11-19 20:17 ` [PATCH v1 3/4] PM: runtime: Let devices suspend after an attempted resume Rafael J. Wysocki
2026-11-20 21:04   ` Brian Norris
2026-11-19 20:18 ` [PATCH v1 4/4] PM: runtime: Drop direct last_busy updates from rpm_resume() Rafael J. Wysocki

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=awC0tcqgq2KvCyxO@google.com \
    --to=briannorris@chromium.org \
    --cc=driver-core@lists.linux.dev \
    --cc=linux-pm@vger.kernel.org \
    --cc=rafael@kernel.org \
    --cc=sakari.ailus@linux.intel.com \
    --cc=stern@rowland.harvard.edu \
    --cc=ulf.hansson@oss.qualcomm.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox