On Mon, 2011-06-27 at 11:57 +0200, Sebastian Andrzej Siewior wrote:

I see. No, actually I don't. xics is pseries where I don't see the .dts. 
So you are saying that we have one irq_domain but 2+ different
interrupt-parents nodes?

Yes. Basically the various PCI bridges and other interrupt sources like
HEA contain what are called "source controllers" that control the
targetting of interrupts (toward a CPU thread), priority, masking etc...

Those act as device-tree parents as well, there's pretty much one per
PCI host bridge for example since that's where all of the
configuration/masking/etc... happens.

However, they then turn interrupts into a special bus message that
reaches eventually a presentation controller (there's one per HW
thread). Those messages are basically the "HW interrupt numbers" (along
with priority info etc...) and that number encodes a "BUID" which
identifies the source controller that shot the message.

Thus the "interrupt numbers" are unique accross the fabric and live in a
unique number space. It's one domain for all intend and purposes. But
several device-nodes.

Now as to whether it's several irq_chip or not .. well, it depends :-)
On pHyp and old style pseries, it's a single set of FW call, it's
abstracted, so it's also basically one chip. On WSP, the separate source
controllers (ICS) are exposed as individual chips.

How do you distinguish then between two different controllers lets say
xics and a gpio based controller? This implementation calls ->dt_translate
until one controller returns 0 which looks like brute force.

It's a bit brute force but would work if the xics implementation of that
translate call checks that the device-node is indeed a XICS source
controller (which can be identified by its compatible property).

Now I have lost track a bit with what Grant is doing, is this the old DT
stuff I objected to ? I basically asked him to make the remapping
orthogonal from the DT matching.

xics_host_xlate() returns always zero so you would have to go for
the compatible and check it.
Every device has an interrupt-parent node. Shouldn't the code call exact 
this irq controller xlate function instead of trying them all?

Well, my powerpc code iterates the domains with "match" to check which
one claims to own the parent device-node, then calls xlate for that one,
but I see why one could collapse those two action after all.

Cheers,
Ben.

quoted

 
Cheers,
Ben.

Sebastian