Re: [PATCH] - race-free suspend. Was: Re: [linux-pm] [PATCH 0/8]Suspend block api (version 8)

[Thomas Gleixner]

On Wed, 2 Jun 2010, Arve Hjønnevåg wrote:

> 2010/6/2 Thomas Gleixner <tglx@xxxxxxxxxxxxx>:
> > On Wed, 2 Jun 2010, Arve Hjønnevåg wrote:
> >> 2010/6/2 Neil Brown <neilb@xxxxxxx>:
> >> > There would still need to be some sort of communication between the the
> >> > suspend daemon on any event daemon to ensure that the events had been
> >> > processed.  This could be very light weight interaction.  The point though is
> >> > that with this patch it becomes possible to avoid races.  Possible is better
> >> > than impossible.
> >> >
> >>
> >> We already have a solution. I don't think rejecting our solution but
> >> merging a worse solution should be the goal.
> >
> > That's not the goal at all. We want a solution which is acceptable for
> > android and OTOH does not get into the way of other approaches.
> >
> 
> I don't actually think the suspend blocker patchset get in the way of
> anything else.
> 
> > The main problem I have is that suspend blockers are only addressing
> > one particular problem space of power management.
> >
> > We have more requirements than that, e.g. an active device transfer
> > requires to prevent the idle code to select a deep power state due to
> > latency requirements.
> >
> > So we then have to implement two mechanisms in the relevant drivers:
> >
> >   1) telling the idle code to limit latency
> >   2) telling the suspend code not to suspend
> 
> And 3) telling the idle code to not enter low power modes that disrupt
> active interrupts or clocks.
>
> Our wakelock code handles 2 and 3, but I removed support for 3 on
> request since you can hack it by specifying a latency value that you
> know the low power mode cannot support.

You are mixing concepts. 

clock domains and power domains are a separate issue which are already
handled by the run time power management code and the clock framework.

The interrupt latency is a QoS requirement and has nothing to do with
power domains and clock domains simply because I can go deeper w/o
violating the clock and power domain constraints when the latency
allows it.

> > My main interest is to limit it to one mechanism, which is QoS based
> > and let idle and suspend make the appropriate decisions based on that
> > information.
> >
> 
> We can use one mechanism for this, but we still have to specify both.
> To me this is just another naming argument and not a good reason to
> not merge the suspend blocker code. You have to modify the same

The main objection against suspend blocker is the user space interface
/ ABI issue, not the driver code which we can fix in no time. But we
cannot fix it once it is glued into a user space interface.

I don't care about adding two empty static inlines into a header file,
which allows to merge the android drivers, but I care much about
giving a guaranteed behaviour to user space.

Thanks,

	tglx