Re: [Kgdb-bugreport] Re: problem with kgdb_roundup_cpus() and ppc64

[George A. <ge...@mv...>]

Tom Rini wrote:
> On Mon, Aug 29, 2005 at 04:39:23PM -0700, Tom Rini wrote:
> 
>>On Mon, Aug 29, 2005 at 04:29:48PM -0700, Geoff Levand wrote:
>>
>>
>>>I have a problem around kgdb_roundup_cpus() on ppc64 SMP.  This 
>>>is the current implementation:
>>
>>[snip]
>>
>>>I did the following, and it seems to work OK, but I don't know enough of 
>>>the details to know the proper fix.  Help would be appreciated.
>>>
>>>void kgdb_roundup_cpus(unsigned long flags)
>>>{
>>>	const unsigned long saved_flags = flags;
>>>	local_irq_restore((flags | MSR_EE));
>>>	smp_call_function(kgdb_call_nmi_hook, 0, 0, 0);
>>>	local_irq_restore(flags);
>>>}

I clearly remember writting, yesterday, that smp_call_function() is a 
very poor choice here.  This is what the orgional x86 code did and which 
I changed to send_nmi_all_but_self() because the smp_call_function() may 
not return.  This is especially true if one of the other cpus is in a 
loop with ints off, which it may well be if it is trying to get a 
spin_lockirq which the current cpu holds.  We ARE kgdb! We want results NOW.

George
-- 

>>
>>This sounds like the same issue MIPS SMP has, and they fix it similarly.
>>Ddo we need to or in MSR_EE 'tho?  If not, I think the more correct fix
>>is to provide a default, weak, kgdb_roundup_cpus that does this, take it
>>out of MIPS as well, and be happy.  I bet ppc32 would also be able to
>>use this code, if KGDB&&SMP was tested & supported there.
> 
> 
> Der, clearly, what I said doesn't make sense.
> 
> But, why do we save flags to saved_flags ?  Is there macro expansion I'm
> missing?
> 

-- 
George Anzinger   ge...@mv...
HRT (High-res-timers):  http://sourceforge.net/projects/high-res-timers/