Re: Totally broken PCI PM calls

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

> > In the first case, it makes even sense to keep the driver operating
> > while the device is D1, the driver would then just wake the device on an
> > incoming request (provided this is allowed by the policy). In the later
> > case, the driver state is the only thing that matters.
> 
> Not quite - you have to be able to get the device into a matching state
> at resume time. Probably not a problem in most cases, I realise, but
> thought it was worth a mention.

Wait, I'm not talking about swsusp here, that's exactly my distinction
between driver state and device state :) Only system suspend like swsusp
and suspend-to-ram require a completely coherent and frozen memory
"image".

All sorts of dynamic PM, including system-wide "idle" can deal with
scenarios where the driver can stay functional and decide by itself to
wake up the device upon incoming requests.

> [...]
> 
> > Yup, with the exception that it becomes hell when those devices are
> > anywhere on the VM path... which makes userspace policy unuseable for
> > system suspend.
> 
> A device on the VM path? I don't follow here. Can I have a hand please?

Any block device basically, I don't even want to think about the issues
between NFS & suspend :)

That is anything that userspace potentially requires to operate (device on
the path to an mmap'd file, swap, whatever)...

Ben.