[RFC PATCH v2 01/12] Add sys_hotplug.h for system device hotplug framework

[Rafael J. Wysocki]

On Friday, February 01, 2013 08:23:12 AM Greg KH wrote:
> On Thu, Jan 31, 2013 at 09:54:51PM +0100, Rafael J. Wysocki wrote:
> > > > But, again, I'm going to ask why you aren't using the existing cpu /
> > > > memory / bridge / node devices that we have in the kernel.  Please use
> > > > them, or give me a _really_ good reason why they will not work.
> > > 
> > > We cannot use the existing system devices or ACPI devices here.  During
> > > hot-plug, ACPI handler sets this shp_device info, so that cpu and memory
> > > handlers (drivers/cpu.c and mm/memory_hotplug.c) can obtain their target
> > > device information in a platform-neutral way.  During hot-add, we first
> > > creates an ACPI device node (i.e. device under /sys/bus/acpi/devices),
> > > but platform-neutral modules cannot use them as they are ACPI-specific.
> > 
> > But suppose we're smart and have ACPI scan handlers that will create
> > "physical" device nodes for those devices during the ACPI namespace scan.
> > Then, the platform-neutral nodes will be able to bind to those "physical"
> > nodes.  Moreover, it should be possible to get a hierarchy of device objects
> > this way that will reflect all of the dependencies we need to take into
> > account during hot-add and hot-remove operations.  That may not be what we
> > have today, but I don't see any *fundamental* obstacles preventing us from
> > using this approach.
> 
> I would _much_ rather see that be the solution here as I think it is the
> proper one.
> 
> > This is already done for PCI host bridges and platform devices and I don't
> > see why we can't do that for the other types of devices too.
> 
> I agree.
> 
> > The only missing piece I see is a way to handle the "eject" problem, i.e.
> > when we try do eject a device at the top of a subtree and need to tear down
> > the entire subtree below it, but if that's going to lead to a system crash,
> > for example, we want to cancel the eject.  It seems to me that we'll need some
> > help from the driver core here.
> 
> I say do what we always have done here, if the user asked us to tear
> something down, let it happen as they are the ones that know best :)
> 
> Seriously, I guess this gets back to the "fail disconnect" idea that the
> ACPI developers keep harping on.  I thought we already resolved this
> properly by having them implement it in their bus code, no reason the
> same thing couldn't happen here, right?

Not really. :-)  We haven't ever resolved that particular issue I'm afraid.

> I don't think the core needs to do anything special, but if so, I'll be glad
> to review it.

OK, so this is the use case.  We have "eject" defined for something like
a container with a number of CPU cores, PCI host bridge, and a memory
controller under it.  And a few pretty much arbitrary I/O devices as a bonus.

Now, there's a button on the system case labeled as "Eject" and if that button
is pressed, we're supposed to _try_ to eject all of those things at once.  We
are allowed to fail that request, though, if that's problematic for some
reason, but we're supposed to let the BIOS know about that.

Do you seriously think that if that button is pressed, we should just proceed
with removing all that stuff no matter what?  That'd be kind of like Russian
roulette for whoever pressed that button, because s/he could only press it and
wait for the system to either crash or not.  Or maybe to crash a bit later
because of some delayed stuff that would hit one of those devices that had just
gone.  Surely not a situation any admin of a high-availability system would
like to be in. :-)

Quite frankly, I have no idea how that can be addressed in a single bus type,
let alone ACPI (which is not even a proper bus type, just something pretending
to be one).

Thanks,
Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.