Re: Totally broken PCI PM calls

[Dmitry Torokhov <dtor_core@ameritech.net>]

On Monday 11 October 2004 05:28 pm, Benjamin Herrenschmidt wrote:
> > A "hang" sounds like the pmcore bug I reported about a year ago...
> > 
> > It's rather foolish of the PM core to use the same semaphore to
> > protect system-wide suspend/resume operations that it uses to
> > for mutual exclusion on the device add/remove (which suspend
> > and resume callbacks did happily in 2.4) ... since it's routine to
> > unplug peripherals on suspended systems!
> 
> Definitely. One thing is: how to do it instead ? I've been thinking
> about it for a while and am still wondering... do we want a list
> mecanism with add/remove notifiers so the PM walk can keep in sync
> with devices added/removed ? or should addition/removal be simply
> postponed until the end of the sleep/wakeup process (I tend to vote
> for that).
> 

Yes, I think that devices that failed to resume (and all their children)
have to be moved by the core resume function into a separate list and
then destroyed (again by the driver core). For that we might need to add
bus_type->remove_device() handler as it seems that all buses do alot
of work outside of driver->remove handlers. The remove_device should
accept additional argument - something like dead_device that would
suggest that driver should not be alarmed by any errors during unbind/
removal process as the device (or rather usually its parent) is simply
not there anynore.

Just my $.02

-- 
Dmitry