[PATCH mmotm/next] powerpc: fix powernv boot breakage on G5???

[Benjamin Herrenschmidt]

On Sun, 2014-01-12 at 00:46 -0800, Hugh Dickins wrote:
> My PowerMac G5 cannot boot mmotm these days: different symptoms
> (starting /sbin/init failed? or ATA errors and hang?), with unrelated
> bugs adding to the confusion; but a bisection led to b5ff4211a829
> "powerpc/book3s: Queue up and process delayed MCE events".  Since that
> series seems to be mostly about powernv, I tried changing BOOK3S_64
> to POWERNV in entry_64.S, which has got it back to working for me.
> 
> Signed-off-by: Hugh Dickins <hughd at google.com>
> just in case this happens to be right, but it's well beyond me!

Mahesh, I think that code is broken. It calls into C code after R3 is
loaded which will clobber it, thus clobbering the return from syscalls.
Also, are you really adding that overhead to every syscall in the
system ? That doesn't sound great...

Why do you do something in syscall_exit to begin with anyway ? That
doesn't seen like a great way to process delayed mchecks... When do you
want to process them ? Any interruptible code ? Task level ?

The former, you might be better off just shooting something like a local
doorbell and use some IPI message bit or similar to indicate the
presence of a pending mcheck. The local doorbell will trigger whenever
interrupts are enabled.

Ben.

> ---
> 
>  arch/powerpc/kernel/entry_64.S |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> --- mmotm/arch/powerpc/kernel/entry_64.S	2014-01-10 18:24:56.940448828 -0800
> +++ linux/arch/powerpc/kernel/entry_64.S	2014-01-10 18:29:24.276455182 -0800
> @@ -184,7 +184,7 @@ syscall_exit:
>  	bl	.do_show_syscall_exit
>  	ld	r3,RESULT(r1)
>  #endif
> -#ifdef CONFIG_PPC_BOOK3S_64
> +#ifdef CONFIG_PPC_POWERNV
>  BEGIN_FTR_SECTION
>  	bl	.machine_check_process_queued_event
>  END_FTR_SECTION_IFSET(CPU_FTR_HVMODE)