powerpc: make 64K huge pages more reliable

This patch adds reliability to the 64K huge page option by making use of the PMD for 64K huge pages when base pages are 4k. So instead of a 12 bit pte it would be 7 bit pmd and a 5 bit pte. The pgd and pud offsets would continue as 9 bits and 7 bits respectively. This will allow the pgtable to fit in one base page. This patch would have to be applied after part 1.

---

hugetlbpage.c | 53 ++++++++++++++++++++++++++++++++++++++---------------

1 files changed, 38 insertions(+), 15 deletions(-)

index f4632ad..c6df45b 100644
@@ -30,15 +30,11 @@

#define NUM_LOW_AREAS (0x100000000UL >> SID_SHIFT)
#define NUM_HIGH_AREAS (PGTABLE_RANGE >> HTLB_AREA_SHIFT)

-#ifdef CONFIG_PPC_64K_PAGES
-#define HUGEPTE_INDEX_SIZE (PMD_SHIFT-HPAGE_SHIFT)
-#else
-#define HUGEPTE_INDEX_SIZE (PUD_SHIFT-HPAGE_SHIFT)
-#endif
-#define PTRS_PER_HUGEPTE (1 << HUGEPTE_INDEX_SIZE)
-#define HUGEPTE_TABLE_SIZE (sizeof(pte_t) << HUGEPTE_INDEX_SIZE)
+unsigned int hugepte_shift;
+#define PTRS_PER_HUGEPTE (1 << hugepte_shift)
+#define HUGEPTE_TABLE_SIZE (sizeof(pte_t) << hugepte_shift)

-#define HUGEPD_SHIFT (HPAGE_SHIFT + HUGEPTE_INDEX_SIZE)
+#define HUGEPD_SHIFT (HPAGE_SHIFT + hugepte_shift)
#define HUGEPD_SIZE (1UL << HUGEPD_SHIFT)
#define HUGEPD_MASK (~(HUGEPD_SIZE-1))

@@ -105,7 +101,14 @@ pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)

			if (!pmd_none(*pm))
				return hugepte_offset((hugepd_t *)pm, addr);

#else
- return hugepte_offset((hugepd_t *)pu, addr);
+ if (HPAGE_SHIFT == HPAGE_SHIFT_64K) {
+ pmd_t *pm;
+ pm = pmd_offset(pu, addr);
+ if (!pmd_none(*pm))
+ return hugepte_offset((hugepd_t *)pm, addr);
+ } else {
+ return hugepte_offset((hugepd_t *)pu, addr);
+ }

#endif
		}
	}
@@ -133,7 +136,14 @@ pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
		if (pm)
			hpdp = (hugepd_t *)pm;

#else
- hpdp = (hugepd_t *)pu;
+ if (HPAGE_SHIFT == HPAGE_SHIFT_64K) {
+ pmd_t *pm;
+ pm = pmd_alloc(mm, pu, addr);
+ if (pm)
+ hpdp = (hugepd_t *)pm;
+ } else {
+ hpdp = (hugepd_t *)pu;
+ }

#endif
	}

@@ -161,7 +171,6 @@ static void free_hugepte_range(struct mmu_gather *tlb, hugepd_t *hpdp)

						 PGF_CACHENUM_MASK));
}

-#ifdef CONFIG_PPC_64K_PAGES
static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
unsigned long addr, unsigned long end, unsigned long floor, unsigned long ceiling)
@@ -194,7 +203,6 @@ static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,

	pud_clear(pud);
	pmd_free_tlb(tlb, pmd);
}

-#endif

static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
unsigned long addr, unsigned long end,
@@ -213,9 +221,15 @@ static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,

			continue;
		hugetlb_free_pmd_range(tlb, pud, addr, next, floor, ceiling);

#else
- if (pud_none(*pud))
- continue;
- free_hugepte_range(tlb, (hugepd_t *)pud);
+ if (HPAGE_SHIFT == HPAGE_SHIFT_64K) {
+ if (pud_none_or_clear_bad(pud))
+ continue;
+ hugetlb_free_pmd_range(tlb, pud, addr, next, floor, ceiling);
+ } else {
+ if (pud_none(*pud))
+ continue;
+ free_hugepte_range(tlb, (hugepd_t *)pud);
+ }

#endif
	} while (pud++, addr = next, addr != end);

@@ -538,6 +552,15 @@ void set_huge_psize(int psize)

			mmu_psize_defs[psize].shift == HPAGE_SHIFT_64K)) {
		HPAGE_SHIFT = mmu_psize_defs[psize].shift;
		mmu_huge_psize = psize;

+#ifdef CONFIG_PPC_64K_PAGES
+ hugepte_shift = (PMD_SHIFT-HPAGE_SHIFT);
+#else
+ if (HPAGE_SHIFT == HPAGE_SHIFT_64K)
+ hugepte_shift = (PMD_SHIFT-HPAGE_SHIFT);
+ else
+ hugepte_shift = (PUD_SHIFT-HPAGE_SHIFT);
+#endif
+

	} else
		HPAGE_SHIFT = 0;
}

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Hrm.. shouldn't we just ban 64K hugepages on a 64K base page size setup? There's not a whole lot of point to it, after all...

David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Banning the base and huge page size from being the same size feels like an artificial barrier. It is probably not the most massively useful combination, but it shouldn't hurt performance.

Jon

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Actually, it sounds to me like an ideal way to benchmark the efficiency of the hugepage implementation and VM effects, without the TLB performance obscuring the results.

I agree that it's not something people will want to do very often, but the same can be said about quite a lot of strange things that we allow just because there's no fundamental reason why they cannot be.

-- Chris

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Well.. maybe. The pagetable format, and therefore the hugepage size, will affect the size of that overhead so even with this its not clear how meaningful test results would be.

Hrm, yeah I guess, since this was a fairly simple fix. I still don't think it's worth jumping through too many hoops to make this possible, though.

David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>