Re: [XenPPC] [PATCH/RFC] Handshake with secondary processors

[Jimi Xenidis <jimix@xxxxxxxxxxxxxx>]

Good, lets take it a little further.
I know you are are building on my original "good enuff loop" but lets  
go all the way here.
On Aug 9, 2006, at 11:12 PM, Amos Waterland wrote:

> diff -r 058f2e27476d xen/arch/powerpc/boot_of.c
> --- a/xen/arch/powerpc/boot_of.c        Mon Aug 07 17:49:16 2006 -0500
> +++ b/xen/arch/powerpc/boot_of.c        Wed Aug 09 23:08:03 2006 -0400
> @@ -42,6 +42,7 @@ static char builtin_cmdline[COMMAND_LINE
>      __attribute__((section("__builtin_cmdline"))) = CMDLINE;
>
>  extern struct ns16550_defaults ns16550;
> +extern unsigned long __spin_ack;
There really is no need for this to be a long, the ID is an int so we  
could make it an int.
This would also make your assembler 32/64 neutral, tho I'm not too  
worried about that, and the barrier (too come) may not be neutral-able.
>  #undef OF_DEBUG
>
>
> @@ -956,7 +957,37 @@ static int __init boot_of_cpus(void)
>      /* FIXME: should not depend on the boot CPU bring the first  
> child */
That is correct.
so you need something like the following (psuedo code, bad style):

       of_getprop(bof_chosen, "cpu", &boot_cpu, sizeof (boot_cpu));
>      cpu = of_getpeer(cpu);
>      while (cpu > 0) {
           if (cpu == boot_cpu)
               continue;

> -        of_start_cpu(cpu, (ulong)spin_start, 0);
> +        unsigned int cpuid;
> +        unsigned long ping, pong, now, then, timeout;
> +        unsigned long *ack = (unsigned long *)&__spin_ack;
> +
> +        result = of_getprop(cpu, "reg", &cpuid, sizeof(cpuid));
> +        if (result == OF_FAILURE) of_panic("cpuid lookup failed\n");
> +
> +        of_printf("spinning up secondary processor #%d: ", cpuid);
> +
> +        *ack = 0x0;
Set this to ~0, cpuid 0 will cause you problems later.

> +        ping = *ack;
> +        of_printf("ping = %lx: ", ping);
> +
> +        mb();
> +        of_start_cpu(cpu, (ulong)spin_start, cpuid);
> +
> +        then = mftb();
> +        timeout = then + 1024;
> +        do {
> +            now = mftb();
> +            if (now >= timeout) {
> +                of_printf("SKIPPING: ");
> +                pong = ~0x0;
> +                break;
> +            }
> +
> +            mb();
> +            pong = *ack;
> +        } while (pong == ping);
> +        of_printf("pong = %lx\n", pong);
> +
>          cpu = of_getpeer(cpu);
>      }
>      return 1;
> diff -r 058f2e27476d xen/arch/powerpc/powerpc64/exceptions.S
> --- a/xen/arch/powerpc/powerpc64/exceptions.S	Mon Aug 07 17:49:16  
> 2006 -0500
> +++ b/xen/arch/powerpc/powerpc64/exceptions.S	Wed Aug 09 21:41:16  
> 2006 -0400
> @@ -178,6 +178,25 @@ zero:
>      li r0, 0x0 /* exception vector for GDB stub */
>      bctr
The space at the beginning of text is precious, please but this back  
to the end of this file.
Extra credit: As you add more code here for the barrier I would not  
discourage a unique file here that is not 64bit specific and may end  
up in the __init section so it can be released.
> +/* Begin secondary processor spin and ack logic.  */
> +    .globl __spin_ack
> +__spin_ack:
> +    .llong 0x0
'.long' here is plenty.
BTW: I can think of no reason this cannot be allocated in "C" rather  
than assembler, it would be easier to find with cscope. :)
> +       
> +    .globl spin_start
> +spin_start:
I almost tricked myself here, so a comment about how we do _not_ use  
_GLOBAL() because we desire a function entry point rather than a  
"function descriptor", would be nice.
> +    /* Our physical cpu number is passed in r3.  */
> +    mr r24, r3
> +    lis r25, __spin_ack@highest
> +    ori r25, r25, __spin_ack@higher
> +    rldicr r25, r25, 32, 31
> +    oris r25, r25, __spin_ack@h
> +    ori r25, r25, __spin_ack@l
> +    stdu r24, 0(r25)
The above assembler can be replaced by:
       LOADADDR(r4, __spin_ack)
       stw r3, 0(r4)

I use r4 because it really doesn't matter what register you use, so  
you have 31 to chose from :)
                
> +    sync
> +    b .
> +
> +/* Begin exception handlers.  */       
>      . = 0x100 # System Reset
>  ex_reset:
>      /* XXX thread initialization */
> @@ -513,7 +532,3 @@ _GLOBAL(sleep)
>      isync
>      mtmsrd r3
>      blr
> -
> -    .globl spin_start
> -spin_start:
> -    b .
>
> _______________________________________________
> Xen-ppc-devel mailing list
> Xen-ppc-devel@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-ppc-devel

_______________________________________________
Xen-ppc-devel mailing list
Xen-ppc-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ppc-devel