Message-ID: <20100531090329.786fce79@notabene.brown>
Date:	Mon, 31 May 2010 09:03:29 +1000
From:	Neil Brown <neilb@...e.de>
To:	"Rafael J. Wysocki" <rjw@...k.pl>
Cc:	Arve Hjønnevåg <arve@...roid.com>,
	markgross@...gnar.org, Matthew Garrett <mjg59@...f.ucam.org>,
	Greg KH <gregkh@...e.de>, linux-doc@...r.kernel.org,
	Peter Zijlstra <peterz@...radead.org>,
	Jesse Barnes <jbarnes@...tuousgeek.org>,
	Andi Kleen <ak@...ux.intel.com>,
	"Linux-pm mailing list" <linux-pm@...ts.linux-foundation.org>,
	Len Brown <len.brown@...el.com>,
	James Bottomley <James.Bottomley@...e.de>, tytso@....edu,
	Dmitry Torokhov <dmitry.torokhov@...il.com>,
	Kernel development list <linux-kernel@...r.kernel.org>,
	Tejun Heo <tj@...nel.org>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Wu Fengguang <fengguang.wu@...el.com>
Subject: Re: [linux-pm] [PATCH 1/8] PM: Opportunistic suspend support.

On Sun, 30 May 2010 21:52:14 +0200
"Rafael J. Wysocki" <rjw@...k.pl> wrote:

> On Sunday 30 May 2010, Neil Brown wrote:
> > On Fri, 28 May 2010 21:04:53 -0700
> > Arve Hjønnevåg <arve@...roid.com> wrote:
> > 
> > > On Fri, May 28, 2010 at 7:52 PM, mark gross <640e9920@...il.com> wrote:
> > > > On Thu, May 27, 2010 at 05:23:54PM +1000, Neil Brown wrote:
> > > >> On Wed, 26 May 2010 14:20:51 +0100
> > > >> Matthew Garrett <mjg59@...f.ucam.org> wrote:
> > > >>
> > > >> > On Wed, May 26, 2010 at 02:57:45PM +0200, Peter Zijlstra wrote:
> > > >> >
> > > >> > > I fail to see why. In both cases the woken userspace will contact a
> > > >> > > central governing task, either the kernel or the userspace suspend
> > > >> > > manager, and inform it there is work to be done, and please don't
> > > >> > > suspend now.
> > > >> >
> > > >> > Thinking about this, you're right - we don't have to wait, but that does
> > > >> > result in another problem. Imagine we get two wakeup events
> > > >> > approximately simultaneously. In the kernel-level universe the kernel
> > > >> > knows when both have been handled. In the user-level universe, we may
> > > >> > have one task schedule, bump the count, handle the event, drop the count
> > > >> > and then we attempt a suspend again because the second event handler
> > > >> > hasn't had an opportunity to run yet. We'll then attempt a suspend and
> > > >> > immediately bounce back up. That's kind of wasteful, although it'd be
> > > >> > somewhat mitigated by checking that right at the top of suspend entry
> > > >> > and returning -EAGAIN or similar.
> > > >> >
> > > >>
> > > >> (I'm coming a little late to this party, so excuse me if I say something that
> > > >> has already been covered however...)
> > > >>
> > > >> The above triggers a sequence of thoughts which (When they settled down) look
> > > >> a bit like this.
> > > >>
> > > >> At the hardware level, there is a thing that we could call a "suspend
> > > >> blocker".  It is an interrupt (presumably level-triggered) that causes the
> > > >> processor to come out of suspend, or not to go into it.
> > > >>
> > > >> Maybe it makes sense to export a similar thing from the kernel to user-space.
> > > >> When any event happens that would wake the device (and drivers need to know
> > > >> about these already), it would present something to user-space to say that
> > > >> the event happened.
> > > >>
> > > >> When user-space processes the event, it clears the event indicator.
> > > >
> > > > we did I proposed making the suspend enabling a oneshot type of thing
> > > > and all sorts of weak arguments came spewing forth.  I honestly couldn't
> > > > tell if I was reading valid input or fanboy BS.
> > > >
> > > 
> > > Can you be more specific? If you are talking about only letting
> > > drivers abort suspend, not block it, then the main argument against
> > > that is that you are forcing user-space to poll until the driver stops
> > > aborting suspend (which according to people arguing against us using
> > > suspend would make the power-manager a "bad" process). Or are you
> > > talking about blocking the request from user-space until all other
> > > suspend-blockers have been released and then doing a single suspend
> > > cycle before returning. This would not be as bad, but it would force
> > > the user-space power manager to be multi-threaded since it now would
> > > have way to cancel the request. Either way, what problem are you
> > > trying to solve by making it a one-shot request?
> > > 
> > 
> > I don't know exactly what Mark has in mind, but I would advocate 1-shot
> > simply because what we currently have (echo mem > /sys/power/state) is
> > 1-shot and I don't believe you need to do more than fix the bugs in that.
> > 
> > Your question of whether to abort or block suspend in central I think - the
> > answer to that question will make or break a possible solution. 
> > 
> > Simply aborting the suspend cannot work as you rightly say - the suspend
> > daemon would then spin until other user-space processes get into action.
> > Simply blocking while there are any unhandled 'wakeup events' - then aborting
> > if there were any - is how I think it should work.  However as it
> > doesn't work that way now I don't think it is safe to make it work that way
> > unconditionally.  If we did we could find that existing configurations always
> > block suspend indefinitely with would clearly be a regression.
> > 
> > I think we still need some sort of "suspend_prepare".  This would have two
> > particular effects.
> > 1/ it sets the start time for interpreting the word "were" above.  i.e. the
> >   suspend would abort of there were any unhandled wakeup events since the
> >   "suspend_prepare" was issued.
> > 2/ It would allow unhandled wakeup events to abort the suspend.  If no
> >   suspend_prepare had been issued, then only "new" wakeup events would
> >   be allowed to abort the suspend (i.e. the old racy version of suspend).
> > 
> > So the suspend daemon does:
> > 
> >    wait for there to be no user-space suspend blocks
> >    issue suspend_prepare
> >    check there are still no suspend blocks
> >    if there are, loop (possibly issue suspend_abort if needed)
> >    issue suspend request
> >    loop
> > 
> > processes that handle wakeup events would
> > 
> >    poll for event to be available
> >    request suspend-block
> >    consume event
> >    release suspend-block
> >    loop
> > 
> > (where consuming the event would quite possibly cause some other
> > suspend-block to become active - e.g. it might request that the display
> > be unlocked which would block suspends for a time).
> 
> Well, please have a look at the Alan Stern's proposal here:
> http://lkml.org/lkml/2010/5/29/77
>

Thanks for the reference.
Some of the details are different, but the idea seems almost exactly the same
as mine.
The apparent dependence on signals makes me feel a little uncomfortable
(interfaces that depend on using signals seem to be easy to use wrongly), but
I don't think that is a serious flaw.

Maybe the biggest difference is philosophical.  Alan's proposal appears to be
adding a new feature.  Mine presents as trying to fix an existing feature so
that it can be used reliably.
It is easier to argue against a feature than against a bug-fix (??).

Thanks,
NeilBrown
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists