Re: [PATCH] call drv->shutdown at rmmod

[ebiederm@xmission.com (Eric W. Biederman)]

Benjamin Herrenschmidt <benh@kernel.crashing.org> writes:

> > You're assuming that a driver can always bring a device out a shutdown
> > state. That's one of the things we talked about at OLS, and the other half
> > of the justification behind such a feature, not just making sure the
> > device is queisced. Your argument against my suggestion are some of the
> > same arguments for a feature like you're introducing. 
> 
> There is a problem of semantics here. Is shutdown() supposed to shutdown
> the hardware device (ie. low power) or just the driver ? If yes, then
> it's duplicate of the PM callbacks. My understanding of the shutdown()
> callback is that it was more than "stop driver activity, put device into
> idle state" to prepare for a shutdown/reboot (though we do also sleep
> IDE drives in this case, but this is because of that nasty cache flush
> issue).
> 
> The problem with kexec is just that. What it needs isn't low power devices,
> it needs device back in "idle" state, but if possible powered up (or at
> least in whatever state the driver found them on boot). The most important
> thing is to actually stop pending bus mastering activities.
> 
> On PPC, we have a name for that which comes from Open Firmware (since we
> need to ask the firmware to stop bus mastering & idle devices the same way
> when we take over it and before we get control of the system memory) and
> it's called "quiesce".

Even if kexec is not brought into the picture the devices need to be quiesed on
reboot.  On x86 and probably other architectures there are 2 ways a reboot can go.
1) The firmware when it regains control toggles the motherboard reset
   line resetting all of the devices, so nothing we do really makes a difference.
2) The firmware when it regains control tweaks a few things and
   pretends it was never out of control, and restarts the boot
   process.

When the firmware does not toggle the motherboard reset line during a
reboot the firmware case is exactly equivalent to the kexec one.

So shutdown needs to quiese things.

Eric