Re: [PATCH v2 1/2] x86_32: summit_32, use BAD_APICID

[Jiri Slaby <jirislaby@gmail.com>]

On 25.2.2009 12:10, Ingo Molnar wrote:
> * Jiri Slaby<jirislaby@gmail.com>  wrote:
>> In that case, the callers code is buggy, since it passes
>> online_cpu masks even on machines, where apics are not on the
>> same clusters.
>
> It's most likely confusion in the old code. This used to be
> copy&paste-ed versions of different snapshots of the
> mach-default-code, hacked to make work on weird platforms.
> Mainline fixes/updates werent merged in consistently.
>
> So could you please send a patch that fixes this?

I've sent 4 more patches, but there are still issues:
* es7000 + summit: I haven't solved calling with all bits set (only all 
online is sufficient to trigger this). Some of the processors needn't be 
on the same apic cluster. It will scream now (again -- it did before 
adding the "optimisation"). Actually I don't know how to solve this. How 
the caller would know the correct mask, ANDing with a apic->target_cpus 
retval?

* es7000: target_cpus_cluster returns CPU_MASK_ALL and hence it will 
choke itself, because es7000_cpu_mask_to_apicid doesn't count with that. 
Invoked by setup_timer_IRQ0_pin this way.

* set_desc_affinity doesn't expect apic->cpu_mask_to_apicid_and to 
return BAD_APICID and silently sets desc->affinity. Again, I see no 
straightforward solution (rollback of assign_irq_vector and 
set_extra_move_desc needed).