[Gta04-owner] Kernel update for gta04

[NeilBrown]

On Sun, 08 Jul 2012 13:30:00 +0200 Neil Jerram <neil at ossau.homelinux.net>
wrote:

> NeilBrown <neilb at suse.de> writes:
> 
> > Neither.
> > To get race-free suspend you need support in the kernel and in user-space.
> > The kernel side requires that wakeup-events identify themselves.  Last I
> > checked, *nothing* in the kernel identifies as a "wakeup source".  My kernel
> > changes told the power-button, aux-button, RTC alarm and 3G interrupt to
> > identify themselves as wakeup sources.
> >
> > The user-space side is what I was trying to outline here.
> 
> Ah, many thanks for explaining that, and to Radek for the LWN link.
> 
> >> >  The simplest way to avoid races with suspend is:
> >> >  -  read /sys/power/wakeup_count and record the number
> >> >  -  final check that no buttons are pressed etc
> >> >  -  write the recorded number to /sys/power/wakeup_count
> >> >  -  If that succeeds, write "mem" to /sys/power/state
> 
> But isn't it obvious that (1) there's still a window here, between the
> last two lines, and (2) this could and should all be handled within the
> kernel?

1/ no.  When you write a number successfully to wakeup_count it changes the
  behaviour of writing 'mem' to 'state' slightly.  It means that any wakeup
  event should abort that suspend.  If you don't write to wakeup_count, then
  wakeup events won't abort the suspend.
  i.e. the semantics are defined to close the race.

2/ again, no.  The 'final check' must be in use-space.  The kernel doesn't
  keep track of whether wakeup events have been consumed by user-space.
  Doing that was considered  but it introduces too many more problems.

>  
> > As of 3.5 there is code in the kernel which does the core part of this.
> > Whether that is easier or harder to use remains to be seen.  It was added
> > because that is more like how Android works.
> 
> So in 3.5 the race is correctly handled within the kernel, and the user
> space complexity disappears again?

There is still complexity in user-space, it is just different.
To work with in-kernel autosuspend you need to use 'epoll' to listen for
wakeup events, and use a new magic flag that instigates the required
handshake with the kernel.

Achieving race-free suspend requires a certain amount of complexity.  You can
move it around a bit, but you cannot get rid of it completely.

NeilBrown
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 828 bytes
Desc: not available
URL: <http://lists.goldelico.com/pipermail/gta04-owner/attachments/20120708/154e0dda/attachment.bin>