Re: [PATCH 1/11] Add generic helpers for arch IPI function calls

[Mark Lord <lkml@rtr.ca>]

Jens Axboe wrote:
> On Wed, Apr 23 2008, Mark Lord wrote:
>> Jens Axboe wrote:
>>> On Wed, Apr 23 2008, Mark Lord wrote:
>>> ..
>>>> The second bug, is that for the halt case at least,
>>>> nobody waits for the other CPU to actually halt
>>>> before continuing.. so we sometimes enter the shutdown
>>>> code while other CPUs are still active.
>>>>
>>>> This causes some machines to hang at shutdown,
>>>> unless CPU_HOTPLUG is configured and takes them offline
>>>> before we get here.
>>> I'm guessing there's a reason it doesn't pass '1' as the last argument,
>>> because that would fix that issue?
>> ..
>>
>> Undoubtedly -- perhaps the called CPU halts, and therefore cannot reply.  :)
> 
> Uhm yes, I guess stop_this_cpu() does exactly what the name implies :-)
> 
>> But some kind of pre-halt ack, perhaps plus a short delay by the caller
>> after receipt of the ack, would probably suffice to kill that bug.
>>
>> But I really haven't studied this code enough to know,
>> other than that it historically has been a sticky area
>> to poke around in.
> 
> Something like this will close the window to right up until the point
> where the other CPUs have 'almost' called halt().
> 
> diff --git a/arch/x86/kernel/smp.c b/arch/x86/kernel/smp.c
> index 5398385..94ec9bf 100644
> --- a/arch/x86/kernel/smp.c
> +++ b/arch/x86/kernel/smp.c
> @@ -155,8 +155,9 @@ static void stop_this_cpu(void *dummy)
>  	/*
>  	 * Remove this CPU:
>  	 */
> -	cpu_clear(smp_processor_id(), cpu_online_map);
>  	disable_local_APIC();
> +	cpu_clear(smp_processor_id(), cpu_online_map);
> +	smp_wmb();
>  	if (hlt_works(smp_processor_id()))
>  		for (;;) halt();
>  	for (;;);
> @@ -175,6 +176,12 @@ static void native_smp_send_stop(void)
>  
>  	local_irq_save(flags);
>  	smp_call_function(stop_this_cpu, NULL, 0, 0);
> +
> +	while (cpus_weight(cpu_online_map) > 1) {
> +		cpu_relax();
> +		smp_rmb();
> +	}
> +
>  	disable_local_APIC();
>  	local_irq_restore(flags);
>  }
..

Yup, that looks like it oughta work consistently.
Now we just need to hear from some of the folks who
have danced around this code in the past.

(added Pavel & Rafael to Cc:).