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

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Thursday 03 June 2010, Neil Brown wrote:
> On Wed, 2 Jun 2010 22:41:14 +0200
> "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> 
> > On Wednesday 02 June 2010, Neil Brown wrote:
> > > - Would this fix the "bug"??
> > > - and address the issues that suspend-blockers was created to address?
> > > - or are the requirements on user-space too onerous?
> > 
> > In theory wakeup events can also happen after  wait_for_blockers() has returned
> > 0 and I guess we should rollback the suspend in such cases.
> > 
> 
> I naively assumed this was already the case, but on a slightly closer look at
> the code it seems not.
> 
> Presumably there is some point deep in the suspend code, probably after the
> call to sysdev_suspend, where interrupts are disabled and we are about to
> actually suspend.  At that point a simple "is a roll-back required" test
> could abort the suspend.

Yes.

> Then any driver that handles wake-up events, if it gets and event that (would
> normally cause a wakeup) PM_SUSPEND_PREPARE and PM_POST_SUSPEND, could set
> the "roll-back is required" flag.

That's my idea, but instead of setting a flag, I'd use a counter increased
every time there is a wakeup event.  Then, the core suspend core code
would store a pre-suspend value of the counter and compare it with
the current value after all wakeup event sources had been set.  If they
were different, the suspend would be aborted.

Rafael