Re: [PATCH v2 1/2] omap: add bits for future 3430/3630 ES revisions

[Nishanth Menon <nm@ti.com>]

Paul Walmsley had written, on 10/20/2009 06:14 PM, the following:
> Hi Vikram, Nishanth, Richard,
> 
> a few comments on this:
> 
> On Tue, 20 Oct 2009, Vikram Pandita wrote:
> 
>> Add bits for future expansion of omap_chip_id type field.
>> This is needed to accomodate 3630ES1 chip id which is bit10
> 
> ...
> 
>> diff --git a/arch/arm/plat-omap/include/plat/cpu.h b/arch/arm/plat-omap/include/plat/cpu.h
>> index 7cb0556..922bf1c 100644
>> --- a/arch/arm/plat-omap/include/plat/cpu.h
>> +++ b/arch/arm/plat-omap/include/plat/cpu.h
>> @@ -45,7 +45,7 @@ int omap_type(void);
>>  
>>  struct omap_chip_id {
>>  	u8 oc;
>> -	u8 type;
>> +	u32 type;
>>  };
> 
> Just wanted to understand the motivation for using the u32.  
> 
> Earlier in the life of these patches, two comments were mentioned: the 
> desire to 'futureproof' and the desire to reserve space for other 
> 34xx-family parts.
> 
> Regarding 'futureproofing:' that's part of the reason that a separate 
> struct was defined for this: to prevent code that uses it from depending 
> on the size of the type.  (Originally it was a typedef, but Linus hates 
> typedefs...)  So it shouldn't matter how big or small the type is here, as 
> long as it can handle all of the bits allocated for it.
> 
> Also mentioned was the idea of reserving space for other 34xx-family 
> chips.  I'd suggest simply renumbering the bits when and if those versions 
> appear.  Code that uses the omap_chip_id system should always use the 
> macros (e.g. CHIP_IS_OMAP3430) and not encode separate bit shift values, 
> so renumbering should be completely safe and transparent for core code.  
> Module code shouldn't be using the omap_chip code, it's for core usage 
> only.
> 
> So, since only one bit is being added, why not continue the use of the u8?  
> Then when the next bits need to be added, the type can be expanded at that 
> point, and the bits renumbered if necessary.  This should be a completely 
> transparent operation for code that uses it.  Vikram's original patch:
> 
>     http://patchwork.kernel.org/patch/54847/
> 
> should be fine.

Assumptions:
a) omap_chip_id is supposedly constant for all devices within the same 
family. 3630, 3430 rev x will belong to the same family.
b) we prefer bitops to using cpu_is_blahblah() for reasons of 
performance (if it could be argued so)..

Issues with the strategy of restricting to the current 8 bits:
a) Why extrabits now:
we have 8 bits now and we would have used all 8 bits with 3630 with the 
mentioned patch. What do we do with the next revision of 3430? Do we 
want to increase the size once it comes along? OR Do we want to do it 
right now? Why wait till we get additional silicons to go figure how to 
add those bits as Richard pointed out, when there could be one more in 
the pipeline?

b) How much extra bits should we add?
It does not matter when you pull in extra bits - u'd have to go u16 for 
next rev of 3430, it is gonna be a single entry, what if we end up (god 
forbid) further revisions of 3630 or even a new gen of omap3 family - 
call it 3XYZ devices?
Further how expensive is it to add the u16 Vs u32 size from a memory 
usage perspective?

Either way, we could choose:
a) We can choose to stick with u8 and take the next rev of 3430 ES 
revision when the time is right
b) we can choose to stick with u16 for the known future (+allowing for 5 
more revisions of silicons) of OMAP3.
c) We can choose to say - no one really knows how many ES/variants of a 
silicon family could happen for OMAP3/OMAP4/5/6/7 etc.. and choose u32 
giving us a limit beyond which we might choose to rearchitect things.

The choice is upto this community to make.. options are very easy to 
implement ;).

-- 
Regards,
Nishanth Menon