How to handle PTE tables with non contiguous entries ?

[Christophe LEROY]

Le 18/09/2018 à 13:47, Aneesh Kumar K.V a écrit :
> Christophe LEROY <christophe.leroy at c-s.fr> writes:
> 
>> Le 17/09/2018 à 11:03, Aneesh Kumar K.V a écrit :
>>> Christophe Leroy <christophe.leroy at c-s.fr> writes:
>>>
>>>> Hi,
>>>>
>>>> I'm having a hard time figuring out the best way to handle the following
>>>> situation:
>>>>
>>>> On the powerpc8xx, handling 16k size pages requires to have page tables
>>>> with 4 identical entries.
>>>
>>> I assume that hugetlb page size? If so isn't that similar to FSL hugetlb
>>> page table layout?
>>
>> No, it is not for 16k hugepage size with a standard page size of 4k.
>>
>> Here I'm trying to handle the case of CONFIG_PPC_16K_PAGES.
>> As of today, it is implemented by using the standard Linux page layout,
>> ie one PTE entry for each 16k page. This forbids the use the 8xx HW
>> assistance.
>>
>>>
>>>>
>>>> Initially I was thinking about handling this by simply modifying
>>>> pte_index() which changing pte_t type in order to have one entry every
>>>> 16 bytes, then replicate the PTE value at *ptep, *ptep+1,*ptep+2 and
>>>> *ptep+3 both in set_pte_at() and pte_update().
>>>>
>>>> However, this doesn't work because many many places in the mm core part
>>>> of the kernel use loops on ptep with single ptep++ increment.
>>>>
>>>> Therefore did it with the following hack:
>>>>
>>>>     /* PTE level */
>>>> +#if defined(CONFIG_PPC_8xx) && defined(CONFIG_PPC_16K_PAGES)
>>>> +typedef struct { pte_basic_t pte, pte1, pte2, pte3; } pte_t;
>>>> +#else
>>>>     typedef struct { pte_basic_t pte; } pte_t;
>>>> +#endif
>>>>
>>>> @@ -181,7 +192,13 @@ static inline unsigned long pte_update(pte_t *p,
>>>>            : "cc" );
>>>>     #else /* PTE_ATOMIC_UPDATES */
>>>>            unsigned long old = pte_val(*p);
>>>> -       *p = __pte((old & ~clr) | set);
>>>> +       unsigned long new = (old & ~clr) | set;
>>>> +
>>>> +#if defined(CONFIG_PPC_8xx) && defined(CONFIG_PPC_16K_PAGES)
>>>> +       p->pte = p->pte1 = p->pte2 = p->pte3 = new;
>>>> +#else
>>>> +       *p = __pte(new);
>>>> +#endif
>>>>     #endif /* !PTE_ATOMIC_UPDATES */
>>>>
>>>>     #ifdef CONFIG_44x
>>>>
>>>>
>>>> @@ -161,7 +161,11 @@ static inline void __set_pte_at(struct mm_struct
>>>> *mm, unsigned long addr,
>>>>            /* Anything else just stores the PTE normally. That covers all
>>>> 64-bit
>>>>             * cases, and 32-bit non-hash with 32-bit PTEs.
>>>>             */
>>>> +#if defined(CONFIG_PPC_8xx) && defined(CONFIG_PPC_16K_PAGES)
>>>> +       ptep->pte = ptep->pte1 = ptep->pte2 = ptep->pte3 = pte_val(pte);
>>>> +#else
>>>>            *ptep = pte;
>>>> +#endif
>>>>
>>>>
>>>>
>>>> But I'm not too happy with it as it means pte_t is not a single type
>>>> anymore so passing it from one function to the other is quite heavy.
>>>>
>>>>
>>>> Would someone have an idea of an elegent way to handle that ?
>>>>
>>>> Thanks
>>>> Christophe
>>>
>>> Why would pte_update bother about updating all the 4 entries?. Can you
>>> help me understand the issue?
>>
>> Because the 8xx HW assistance expects 4 identical entries for each 16k
>> page, so everytime a PTE is updated the 4 entries have to be updated.
>>
> 
> What you suggested in the original mail is what matches that best isn't it?
> That is a linux pte update involves updating 4 slot. Hence a linux pte
> consist of 4 unsigned long?
> 

Yes indeed.
It seems hopeless to avoid carrying the 4 longs from one function to the 
other allthough that's four times the same thing.

Christophe