Re: [RFC -tip] x86,io-apic: Do not map IO-APIC direct registers twice

[Yinghai Lu <yinghai@kernel.org>]

Cyrill Gorcunov wrote:
> On Thu, Nov 12, 2009 at 03:22:55PM -0800, Yinghai Lu wrote:
>> Cyrill Gorcunov wrote:
>>> Please review, I didn't manage to test (emulate actually since I don't
>>>
> ...
> 
> First of all -- thanks a lot for review Yinghai!
> 
>>> Also insert_resourse will not fail anymore on 1K aligned io-apics.
>> looks that we don't need that ...
>> not io_apic_base already have that + &. left problems are
>> 1. display.
>> 2. insert resource problem.
>>
>> YH
>>
>> diff --git a/arch/x86/kernel/apic/io_apic.c b/arch/x86/kernel/apic/io_apic.c
>> index 90e8bc5..6a9379b 100644
>> --- a/arch/x86/kernel/apic/io_apic.c
>> +++ b/arch/x86/kernel/apic/io_apic.c
>> @@ -4140,11 +4140,13 @@ fake_ioapic_page:
>>  		set_fixmap_nocache(idx, ioapic_phys);
>>  		apic_printk(APIC_VERBOSE,
>>  			    "mapped IOAPIC to %08lx (%08lx)\n",
>> -			    __fix_to_virt(idx), ioapic_phys);
>> +			    __fix_to_virt(idx) + (ioapic_phys & ~PAGE_MASK),
> 
> Don't understand, why? What is wrong with physical address,
> could you elaborate please?

ioapic_phys could be 1k aligned, but __fix_to_virt(idx) will always return 4k aligned.

> 
>> +			    ioapic_phys);
>>  		idx++;
>>  
>> +		/* spec says size is 1024 */
> 
> Hmm, MP says nothing about size of IO-APIC direct registers
> cound and as a result -- the size of MMIO. It will (and
> is) differ between IO-APIC versions. An example -- IO-APIC EOI register
> which 82489DX just dont have at all. At moment (ICH-10) the lenght is 68
> bytes so you may note the comment in the former patch that we "hope" such
> a size will be enough for quite a long time to cover all direct register
> space an IO-APIC provides (though to be precise from this 68 bytes only
> index,data,eoi registers specified).
> 
>>  		ioapic_res->start = ioapic_phys;
>> -		ioapic_res->end = ioapic_phys + PAGE_SIZE-1;
>> +		ioapic_res->end = ioapic_phys + (1<<10) - 1;
>>  		ioapic_res++;
>>  	}
>>  }
>>
> 
> I think I've compicated the patch/idea too much indeed :)
> Since we have fixmap for all io_apics build time reserved
> even if some io-apic is 1K aligned we still may use new
> fixmap index. So only issue remains -- resource allocation.
> Here is an updated patch. Please review.

2. print out ...?

YH