On Apr 30, 2008, at 1:32 AM, Benjamin Herrenschmidt wrote:

On Tue, 2008-04-29 at 18:52 -0500, Kumar Gala wrote:

quoted

On Apr 29, 2008, at 4:28 PM, Paul Mackerras wrote:

quoted

Kumar Gala writes:

quoted

We need to have unique transfer_to_handler paths for each exception
level
that is supported.  We need to use the proper xSRR0/1 depending on
which
exception level the interrupt was from.  The macro conversion  
lets up
templatize this code path.

It seems to me that this implies you are assuming that you will  
never
ever get a synchronous normal interrupt such as a TLB miss while you
are in a critical or machine check handler.

Grr.. one more thing to fix :)

quoted

Wouldn't it be better and safer to have the exception prolog for
critical interrupts save SRR0/1 in the stack frame, and have the
prolog for machine checks save SRR0/1 and CSRR0/1 likewise?

If we do this I guess we can use SRR0/1 regardless of which level we
came from.

Also consider saving/restoring MAS

got that.  Does 40x/44x have anything similar we need to save/restore?