Re: How to handle PTE tables with non contiguous entries ?

From: Christophe LEROY
Date: Tue Sep 18 2018 - 07:53:39 EST

Le 18/09/2018 Ã 13:47, Aneesh Kumar K.V a ÃcritÂ:
Christophe LEROY <christophe.leroy@xxxxxx> writes:

Le 17/09/2018 Ã 11:03, Aneesh Kumar K.V a ÃcritÂ:
Christophe Leroy <christophe.leroy@xxxxxx> writes:


I'm having a hard time figuring out the best way to handle the following

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

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;
typedef struct { pte_basic_t pte; } pte_t;

@@ -181,7 +192,13 @@ static inline unsigned long pte_update(pte_t *p,
: "cc" );
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;
+ *p = __pte(new);
#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
* 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);
*ptep = pte;

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 ?


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.