Message-ID: <CAJZ5v0g3qOvESqvqiCnwVz2BYGHzrG8=nRQ8j36Qd_LC0io_Tw@mail.gmail.com>
Date: Thu, 27 Feb 2025 17:23:41 +0100
From: "Rafael J. Wysocki" <rafael@...nel.org>
To: Saravana Kannan <saravanak@...gle.com>
Cc: "Rafael J. Wysocki" <rjw@...ysocki.net>, Linux PM <linux-pm@...r.kernel.org>, 
	LKML <linux-kernel@...r.kernel.org>, Alan Stern <stern@...land.harvard.edu>, 
	Ulf Hansson <ulf.hansson@...aro.org>, Johan Hovold <johan@...nel.org>, 
	Manivannan Sadhasivam <manivannan.sadhasivam@...aro.org>
Subject: Re: [PATCH v1 0/5] PM: sleep: Improvements of async suspend and
 resume of devices

On Thu, Feb 27, 2025 at 4:45 PM Saravana Kannan <saravanak@...gle.com> wrote:
>
> On Tue, Feb 25, 2025 at 8:46 AM Rafael J. Wysocki <rjw@...ysocki.net> wrote:
> >
> > Hi Everyone,
> >
> > Initially, this was an attempt to address the problems described by
> > Saravana related to spawning async work for any async device upfront
> > in the resume path:
> >
> > https://lore.kernel.org/linux-pm/20241114220921.2529905-1-saravanak@google.com/
> >
> > but then I realized that it could be extended to the suspend path and
> > used for speeding it up, which it really does.
>
> Btw, maybe I didn't  word it correctly, but my patch series was meant
> to speed up the non-async case too.

If "the non-async case" means the case with "async" suspend/resume
disabled entirely, I don't think that the ordering in which devices
are processed can be changed just because there are no known
dependencies.

> I was going to get around sending a v2 of my series, but was caught up
> with some other work. But I'm okay if you want to finish up my effort
> -- less work for me and I can focus on the other aspects of suspend :)
>
> Maybe add a Suggested-by: to the patches?

Yeah, I can do that.

> I definitely want to review the series, but very busy this week with
> some other work. I'll get to this next week for sure.

That should be fine.

> > Overall, the idea is that instead of starting an async work item for every
> > async device upfront, which is not very efficient because the majority of
> > those devices will not be able to make progress due to dependencies anyway,
> > the async handling is only started upfront for the devices that are likely
> > to be able to make progress.  That is, devices without parents in the resume
> > path and leaf devices (ie. devices without children or consumers) in the
> > suspend path (the underlying observation here is that devices without parents
> > are likely to have no suppliers too whereas devices without children that
> > have consumers are not unheard of).  This allows to reduce the amount of
> > processing that needs to be done to start with.
> >
> > Then, after processing every device ("async" or "sync"), "async" processing
> > is started for some devices that have been "unblocked" by it, which are its
> > children in the resume path or its parent and its suppliers in the suspend
> > path.  This allows asynchronous handling to start as soon as it makes sense
> > without delaying the "async" devices unnecessarily.
> >
> > Fortunately, the additional plumbing needed to implement this is not
> > particularly complicated.
> >
> > The first two patches in the series are preparatory.
> >
> > Patch [3/5] deals with the resume path for all device resume phases.
> >
> > Patch [4/5] optimizes the "suspend" phase which has the most visible effect (on
> > the systems in my office the speedup is in the 100 ms range which is around 20%
> > of the total device resume time).
> >
> > Patch [5/5] extend this to the "suspend late" and "suspend noirq" phases.
> >
> > Thanks!

Powered by blists - more mailing lists