Re: [RFC 00/15] x86_64: Optimize percpu accesses

[Jeremy Fitzhardinge <jeremy@goop.org>]

Christoph Lameter wrote:
> What is the point of the pda_X interface? It does not exist on 32 bit.
> The pda wastes the GS segment register on a small memory area. This patchset
> makes the GS segment usable to reach all of the per cpu area by placing
> the pda into the per cpu area. Thus the pda_X interface becomes obsolete
> and the 32 bit per cpu stuff becomes usable under 64 bit unifying both
> architectures.
>   

I think we agree on the desired outcome.  I just disagree with the path 
to getting there.

>> I think we should start devolving things out of the pda in the other
>> direction: make a series where each patch takes a member of struct
>> x8664_pda, converts it to a per-cpu variable (where possible, the same
>> one that 32-bit uses), and updates all the references accordingly.  When
>> the pda is as empty as it can be, we can look at removing the
>> pda-specific interfaces.
>>     
>
> This patchset places the whole x8664_pda structure into the per cpu area and makes the pda macros operate on the x8664_pda structure in the per cpu area. Not sure why you want to go through the churn of doing it for each object separately.
>   

No, it's not churn doing it object at a time.  If you convert 
pda.pcurrent into a percpu current_task variable, then at one stroke 
you've 1) shrunk the pda, 2) unified with i386.  If you go through the 
process of converting all the read_pda(pcurrent) references into 
x86_read_percpu(pda.pcurrent) then that's a pure churn patch.  It 
doesn't get rid of the pda variable, it doesn't unify with i386.  All it 
does is remove a reference to a macro which was fairly inoffensive in 
the first place.

Once the pda has shrunk as much as it can (which remove everything 
except stack_canary, I think), then remove all the X_pda macros, since 
there won't be any users anyway.

    J