Re: [markgross@thengar.org: [RFC] wake up notifications and suspendblocking (aka more wakelock stuff)]

[Alan Stern]

On Sat, 8 Oct 2011, Rafael J. Wysocki wrote:

> On Saturday, October 08, 2011, mark gross wrote:
> > On Sat, Oct 08, 2011 at 10:14:39PM +1100, NeilBrown wrote:
> > > On Sun, 2 Oct 2011 09:44:56 -0700 mark gross <markgross@xxxxxxxxxxx> wrote:
> > > 
> > > > resending to wider list for discussion
> > > > ----- Forwarded message from mark gross <markgross@xxxxxxxxxxx> -----
> > > > 
> > > > Subject: [RFC] wake up notifications and suspend blocking (aka more wakelock stuff)
> > > > Date: Tue, 20 Sep 2011 13:33:05 -0700
> > > > From: mark gross <markgross@xxxxxxxxxxx>
> > > > To: linux-pm@xxxxxxxxxxxxxxxxxxxxxxxxxx
> > > > Reply-To: markgross@xxxxxxxxxxx
> > > > Cc: arve@xxxxxxxxxxx, markgross@xxxxxxxxxxx, Alan Stern <stern@xxxxxxxxxxxxxxxxxxx>, amit.kucheria@xxxxxxxxxx, farrowg@xxxxxxxxxx, "Rafael J. Wysocki" <rjw@xxxxxxx>
> > > > 
> > > > The following patch set implement an (untested) solution to the
> > > > following problems.
> > > > 
> > > > 1) a method for making a system unable to suspend for critical sections
> > > > of time.
> > > 
> > > We already have this.  A properly requested suspend (following wakeup_count
> > > protocol) is unable to complete between wakeup_source_activate() and
> > > wake_source_deactivate() - these delimit the critical sections.
> > > 
> > > What more than this do you need?
> > 
> > sometimes devices that are not wake up sources need critical sections
> > where suspend is a problem.

Mark, can you give some examples?  This isn't the same as the firmware 
update thing, is it?  That deserves to be handled by a completely 
separate mechanism.

> > > If user-space wants to prevent suspend, it just needs some sort of protocol
> > > for talking to the user-space process which follows the correct protocol to
> > > initiate suspend.  That isn't a kernel problem.
> > 
> > The devices that I've seen that need to block suspend don't have a
> > communication interface to user mode. 
> > 
> > But, you are right the devices that need this sort of thing could
> > register as wakeup sources and block suspend as well.
> > 
> > FWIW This part of the patch set was to wrap up a proposal I got last
> > year from some folks to try to provide somewhat compatible semantics to
> > wakelock's for the android and linux kernel community.
> > 
> > I include it for completeness. 

Although many of these problems could be solved by adopting a suitable
protocol in userspace, currently there is no such protocol.  Even if
one did exist, the process of getting all the relevant programs to
adopt it would take quite a while.  This is a case where a problem can
be solved either in the kernel or in userspace, and the in-kernel
solution may be simpler.

> > > > 2) providing a race free method for the acknowledgment of wake event
> > > > processing before re-entry into suspend can happen.
> > > 
> > > Again, this is a user-space problem.  It is user-space which requests
> > > suspend.  It shouldn't request it until it has checked that there are no wake
> > > events that need processing - and should use the wakeup_count protocol to
> > > avoid races with wakeup events happening after it has checked.
> > 
> > Here you are wrong, or missing the point.  The kernel needs to be
> > notified from user mode that an update event has been consumed by
> > whoever cares about it before the next suspend can happen.
> 
> This, in fact, isn't correct.  I have tried to turn your (and John's)
> attention to this for quite a few times already.
> 
> The point is that the entity about to trigger suspend (that need not be the
> kernel!) has to communicate with the processes that consume wakeup events
> beforehand.  In theory this communication can happen entirely in user
> space, but that would involve quite complicated interactions between
> processes, so nobody does that in practice.
> 
> The only "problem" that can't be solved entirely in user space, which is
> what John turned my attention to during the LPC, is that it may be
> possible to suspend when processes that should be asked about whether or
> not to suspend are sleeping and that may be done _without_ actually asking
> those processes for permission.

Essentially, there has to be a way for these processes to say "It's 
okay to suspend until I tell you otherwise."  That could also be part 
of the userspace protocol (if it existed).

>  The difficult part is, if we suspend in
> such a situation, we need to wait until all of those processes have a chance
> to run before attempting to suspend again.

Of course, a power-management daemon could handle that part easily.  
Doing it in the kernel is more difficult.

> > The fact that there are time outs in the existing wake event code points to
> > this shortcoming in the current implementation.
> 
> Actually, the timeouts serve a different purpose.  Namely, there are wakeup
> events that aren't actually consumed by anyone above the layer signaling the
> event (think about Wake-on-LAN via a magic packet) and if such an event
> happens, we can't suspend at once, because we need to assume that it happened
> for a reason, so whoever triggered the event has to be given a chance to do
> whatever he needed to wake up the system for.  This cannot be achieved without
> timeouts.

Also there are events that may or may not get consumed, but it is
impossible to tell at the time whether or not a consumer exists.  If
one does exist then good, the system can suspend again when the
consumer is finished.  But if not, the system can only wait until a
timeout.

> > I suppose one could rig up the user mode suspend daemon with
> > notification callbacks between event consumers across the user mode
> > stack but its really complex to get it right and forces a solution to a
> > problem better solved in kernel mode be done with hacky user mode
> > gyrations that may ripple wildly across user mode.
> 
> Agreed.
> 
> > Also it is the kernel that is currently deciding when to unblock the
> > suspend daemon for the next suspend attempt.  Why not build on that and
> > make is so we don't need the time outs?
> > 
> > > i.e. there is no kernel-space problem to solve here (except for possible
> > > bugs).
> > 
> > Just a race between the kernel allowing a suspend and the user mode code
> > having time to consume the last wake event.
> 
> That's correct.

Alan Stern

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