Re: Reworking suspend-resume sequence (was: Re: PCI PM: Restore standard config registers of all devices early)

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

> (Newer) ACPI says that devices should be put into low power states (presumably
> with the help of appropriate ACPI AML routines) before the _PTS method is
> called.  In turn, we're supposed to disable nonboot CPUs after calling _PTS.
> There is analogous requirement for the _WAK method during resume.
> 
> Currently, the suspend code ordering follows these rules, but if we move
> the putting of devices into low power states into the suspend_late part, they
> will have to be done after _PTS and that is likely to break things (we've
> already had this problem once and I have really bad memories related to it).

Wait wait wait ... the -whole- point of the exercise, wether using
local_irq_save or disable_irq, -is- to put the ACPI bit -after- setting
the device in low power state and before the restore on wakeup...

So basically, that isn't changing.

The -one- thing that indeed conflicts here is that we disable nonboot
CPUs earlier. Right ?

Now, I doubt that would be a big issue, ie, we are supposedly capable of
dynamically disabling/enabling CPUs anyway, but if it is, then indeed I
see how the using of higher level PIC irq disabling would allow to move
the whole suspend_late() over to before disabling non-boot CPUs.

That does introduce a significant change in semantics for drivers in the
sense that now, suspend_late will be called with timers running, things
scheduling, requests coming in, etc... One of the big reasons for doing
suspend_late with IRQs off was precisely that drivers wouldn't have to
synchronize with all these things. They now do.

Thus I tend to think that keeping the disabling of nonboot CPUs earlier
than the suspending of devices is the least of two evils. _BUT_ As I
said, I'm no ACPI expert and not -that- familiar with x86 land, it might
indeed be a can of worms.

Cheers,
Ben.