[PATCH] Accurate task and cpu time accounting

[Benjamin Herrenschmidt]

> CPU_FTR_CI_LARGE_PAGE seems to be a weird case, it's checked only in
> one location (hash code that enables 64K pages) but never actually set
> anywhere in current sources. It never seems to have been.

Yup, because it's not yet clear when to set it... There is a new
device-tree property being architected but I didn't yet have a chance to
test a machine with a firmware that provides it. Setting it based on the
PVR would cause problems in the case of machines mixing multiple CPU
revisions.

In general, we have a problem with our cputable model vs. IBM cpu
feature model. The PAPR architecture considers that any feature not
explicitely exposed in the device-tree should not be used (like altivec
for example). We currently use the PVR for almost everything
however.... 

We need to be able to identify properly a PAPR machine early and clear
out a load of feature bits from what was provided by the table, and then
only set back in the bits that are advertised by the various device-tree
properties defined by IBM. In addition, we need to make sure we don't
break bare-metal in the process. To make things difficult, identifying a
PAPR machine is a bit dodgy since they are still specified as simply
having "chrp" in / device_type ...

Ben.