mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
From: "Rafael J. Wysocki" <rafael@kernel.org>
To: Saravana Kannan <saravanak@google.com>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>,
	Linux PM <linux-pm@vger.kernel.org>,
	 LKML <linux-kernel@vger.kernel.org>,
	Alan Stern <stern@rowland.harvard.edu>,
	 Ulf Hansson <ulf.hansson@linaro.org>,
	Johan Hovold <johan@kernel.org>,
	 Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>,
	Jon Hunter <jonathanh@nvidia.com>
Subject: Re: [PATCH v2 2/3] PM: sleep: Suspend parents and suppliers after suspending subordinates
Date: Fri, 14 Mar 2025 11:23:26 +0100	[thread overview]
Message-ID: <CAJZ5v0htBwvcTFYcLVYYQ0EfunHc3yebJFw2yCSS7Ch4R5XdjQ@mail.gmail.com> (raw)
In-Reply-To: <CAJZ5v0j0kD_HjbVzHb7=5nEYx_5fzwoBupPBQs2Fpfzho_amfg@mail.gmail.com>

On Thu, Mar 13, 2025 at 11:36 PM Rafael J. Wysocki <rafael@kernel.org> wrote:
>
> On Thu, Mar 13, 2025 at 10:17 PM Saravana Kannan <saravanak@google.com> wrote:
> >
> > On Thu, Mar 13, 2025 at 1:35 PM Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
> > >
> > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > >
> > > In analogy with the previous change affecting the resume path,
> > > make device_suspend() start the async suspend of the device's parent
> > > and suppliers after the device itself has been processed and make
> > > dpm_suspend() start processing "async" leaf devices (that is, devices
> > > without children or consumers) upfront because they don't need to wait
> > > for any other devices.
> > >
> > > On the Dell XPS13 9360 in my office, this change reduces the total
> > > duration of device suspend by approximately 100 ms (over 20%).
> > >
> > > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > > Suggested-by: Saravana Kannan <saravanak@google.com>
> > > ---
> > >
> > > v1 -> v2:
> > >    * Adjust for the changes in patch [1/3].
> > >    * Fix walking suppliers in dpm_async_suspend_superior().
> > >    * Use device links read locking in dpm_async_suspend_superior() (Saravana).
> > >    * Move all devices to the target list even if there are errors in
> > >      dpm_suspend() so they are properly resumed during rollback (Saravana).
> > >
> > > ---
> > >  drivers/base/power/main.c |   78 ++++++++++++++++++++++++++++++++++++++++++----
> > >  1 file changed, 72 insertions(+), 6 deletions(-)
> > >
> > > --- a/drivers/base/power/main.c
> > > +++ b/drivers/base/power/main.c
> > > @@ -1231,6 +1231,50 @@
> > >
> > >  /*------------------------- Suspend routines -------------------------*/
> > >
> > > +static bool dpm_leaf_device(struct device *dev)
> > > +{
> > > +       struct device *child;
> > > +
> > > +       lockdep_assert_held(&dpm_list_mtx);
> > > +
> > > +       child = device_find_any_child(dev);
> > > +       if (child) {
> > > +               put_device(child);
> > > +
> > > +               return false;
> > > +       }
> > > +
> > > +       /*
> > > +        * Since this function is required to run under dpm_list_mtx, the
> > > +        * list_empty() below will only return true if the device's list of
> > > +        * consumers is actually empty before calling it.
> > > +        */
> > > +       return list_empty(&dev->links.consumers);
> > > +}
> >
> > We need the equivalent of this for resume.
>
> Maybe.
>
> > > +
> > > +static void dpm_async_suspend_superior(struct device *dev, async_func_t func)
> > > +{
> > > +       struct device_link *link;
> > > +       int idx;
> > > +
> > > +       mutex_lock(&dpm_list_mtx);
> > > +
> > > +       /* Start processing the device's parent if it is "async". */
> > > +       if (dev->parent)
> > > +               dpm_async_with_cleanup(dev->parent, func);
> > > +
> > > +       mutex_unlock(&dpm_list_mtx);
> > > +
> > > +       idx = device_links_read_lock();
> > > +
> > > +       /* Start processing the device's "async" suppliers. */
> > > +       list_for_each_entry_rcu(link, &dev->links.suppliers, c_node)
> > > +               if (READ_ONCE(link->status) != DL_STATE_DORMANT)
> > > +                       dpm_async_with_cleanup(link->supplier, func);
> >
> > We should check that the rest of the consumers of the supplier are
> > "done" before we queue the supplier. With 386 device links (and the
> > number only increases as we add support for more properties), there's
> > no doubt that we'll hit this often.
>
> And I'm not doing this until I see any data confirming that it makes a
> difference in the order of 10% or more.
>
> > > +
> > > +       device_links_read_unlock(idx);
> >
> > Is passing idx to unlock a new (within the past 6 months) thing? I
> > don't remember having to do this in the past.
>
> It's SRCU and it's been there forever.
>
> > > +}
> > > +
> > >  /**
> > >   * resume_event - Return a "resume" message for given "suspend" sleep state.
> > >   * @sleep_state: PM message representing a sleep state.
> > > @@ -1656,6 +1700,8 @@
> > >         device_links_read_unlock(idx);
> > >  }
> > >
> > > +static void async_suspend(void *data, async_cookie_t cookie);
> > > +
> > >  /**
> > >   * device_suspend - Execute "suspend" callbacks for given device.
> > >   * @dev: Device to handle.
> > > @@ -1785,7 +1831,13 @@
> > >
> > >         complete_all(&dev->power.completion);
> > >         TRACE_SUSPEND(error);
> > > -       return error;
> > > +
> > > +       if (error || async_error)
> > > +               return error;
> > > +
> > > +       dpm_async_suspend_superior(dev, async_suspend);
> > > +
> > > +       return 0;
> > >  }
> > >
> > >  static void async_suspend(void *data, async_cookie_t cookie)
> > > @@ -1803,6 +1855,7 @@
> > >  int dpm_suspend(pm_message_t state)
> > >  {
> > >         ktime_t starttime = ktime_get();
> > > +       struct device *dev;
> > >         int error = 0;
> > >
> > >         trace_suspend_resume(TPS("dpm_suspend"), state.event, true);
> > > @@ -1816,12 +1869,28 @@
> > >
> > >         mutex_lock(&dpm_list_mtx);
> > >
> > > +       /*
> > > +        * Start processing "async" leaf devices upfront because they don't need
> > > +        * to wait.
> > > +        */
> > > +       list_for_each_entry_reverse(dev, &dpm_prepared_list, power.entry) {
> > > +               dpm_clear_async_state(dev);
> > > +               if (dpm_leaf_device(dev))
> > > +                       dpm_async_with_cleanup(dev, async_suspend);
> > > +       }
> > > +
> > >         while (!list_empty(&dpm_prepared_list)) {
> > > -               struct device *dev = to_device(dpm_prepared_list.prev);
> > > +               dev = to_device(dpm_prepared_list.prev);
> > >
> > >                 list_move(&dev->power.entry, &dpm_suspended_list);
> > >
> > > -               dpm_clear_async_state(dev);
> > > +               /*
> > > +                * Move all devices to the target list to resume them properly
> > > +                * on errors.
> > > +                */
> >
> > I did this initially on my end, but we have so many devices that
> > looping through them had a measurable impact.
>
> Which I guess is super-important for error handling.  Come on.
>
> > It's better to just splice the lists on error.
>
> On top of this change, yes.

Actually, adding a list_splice() after the error check would be a
simpler change than moving the check, so I'm doing this in v3.

  reply	other threads:[~2025-03-14 10:23 UTC|newest]

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-03-13 20:17 [PATCH v2 0/3] PM: sleep: Improvements of async suspend and resume of devices Rafael J. Wysocki
2025-03-13 20:26 ` [PATCH v2 1/3] PM: sleep: Resume children after resuming the parent Rafael J. Wysocki
2025-03-13 21:16   ` Saravana Kannan
2025-03-13 22:29     ` Rafael J. Wysocki
2025-03-13 20:34 ` [PATCH v2 2/3] PM: sleep: Suspend parents and suppliers after suspending subordinates Rafael J. Wysocki
2025-03-13 21:16   ` Saravana Kannan
2025-03-13 22:36     ` Rafael J. Wysocki
2025-03-14 10:23       ` Rafael J. Wysocki [this message]
2025-03-13 20:35 ` [PATCH v2 3/3] PM: sleep: Make suspend of devices more asynchronous 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=CAJZ5v0htBwvcTFYcLVYYQ0EfunHc3yebJFw2yCSS7Ch4R5XdjQ@mail.gmail.com \
    --to=rafael@kernel.org \
    --cc=johan@kernel.org \
    --cc=jonathanh@nvidia.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-pm@vger.kernel.org \
    --cc=manivannan.sadhasivam@linaro.org \
    --cc=rjw@rjwysocki.net \
    --cc=saravanak@google.com \
    --cc=stern@rowland.harvard.edu \
    --cc=ulf.hansson@linaro.org \
    /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

all inboxes | Powered by JetHome®