Re: [RFC PATCH v2 06/20] powerpc/8xx: Fix size given to set_huge_pte_at()

[Christophe Leroy <christophe.leroy@csgroup.eu>]

+Peter Z. who added that commit.

Le 22/05/2024 à 10:32, Christophe Leroy a écrit :
> 
> 
> Le 21/05/2024 à 11:26, Oscar Salvador a écrit :
>> On Tue, May 21, 2024 at 10:48:21AM +1000, Michael Ellerman wrote:
>>> Yeah I can. Does it actually cause a bug at runtime (I assume so)?
>>
>> No, currently set_huge_pte_at() from 8xx ignores the 'sz' parameter.
>> But it will be used after this series.
>>
> 
> Ah yes, I mixed things up with something else in my mind.
> 
> So this patch doesn't qualify as a fix and doesn't need to be handled 
> separately from the series and doesn't really need to go on top of the 
> series either, I think it is better to keep it grouped with other 8xx 
> changes.
> 

I remember now, what I had in mind was commit c5eecbb58f65 
("powerpc/8xx: Implement pXX_leaf_size() support")

That commit is buggy, because pgd_leaf() will always return false on 
8xx. First of all pgd_leaf() could only return true on a target with 
P4Ds. Without P4Ds it should just return 0 like pgd_none(), pgd_bad(), 
... as defined in include/asm-generic/pgtable-nop4d.h

So it is pmd_leaf_size() that could eventually return something for 8xx.
But as 8xx is using hugepd, at the best case it will return crap, worst 
case the read will go in the weed.

To be correct we should had support of hugepd in perf_get_pgtable_size() 
but that's not trivial and this series is aiming at removing hugepd 
completely so there is no point in fixing stuff here, except maybe for 
stable ?