Ben-san,

quoted

It should have been set in setup_64.c, in setup_paca() (which is

called

quoted

twice) in that statement:

	local_paca = &paca[cpu];

As local_paca is defined as being a variable held in register r13.

Maybe

quoted

something bad's happening with the compiler ?

Can you check early_setup disassembly, see if r13 is assigned (it

should

be in two spots) and if it's clobbered by the compiler somewhere ?

Also, you may want to try adding --ffixed-r13 to the CFLAGS in the
makefile and let us know if it makes a difference... r13 is marked
reserved by the ABI but segher seems to imply that gcc may decide to

ues

it anyway (ouch !)

I reviewed early_setup() disassembly sets r13 correctly. Now I found
the kernel works well without our hack, so I think I can decline the
patch [1/7].

Commit edd0622bd2e8f755c960827e15aa6908c3c5aa94 seems to break the
kernel and commit 175587cca7447daf5a13e4a53d32360ed8cba804 backs out
the symptom.  Calling slb_flush_and_rebolt() at slb_initialize()
function breaks the SLB table since it is called before setting up
PACA.

Thank you for your hint.

Best regards,
Kou Ishizaki