[PATCH v6 5/9] arm64: hugetlb: Handle swap entries in huge_pte_offset() for contiguous hugepages

Catalin Marinas [off-list ref] writes:

On Thu, Aug 10, 2017 at 06:09:02PM +0100, Punit Agrawal wrote:

quoted

diff --git a/arch/arm64/mm/hugetlbpage.c b/arch/arm64/mm/hugetlbpage.c
index d3a6713048a2..09e79785c019 100644
--- a/arch/arm64/mm/hugetlbpage.c
+++ b/arch/arm64/mm/hugetlbpage.c

@@ -210,6 +210,7 @@ pte_t *huge_pte_offset(struct mm_struct *mm,
 	pgd_t *pgd;
 	pud_t *pud;
 	pmd_t *pmd;
+	pte_t *pte;
 
 	pgd = pgd_offset(mm, addr);
 	pr_debug("%s: addr:0x%lx pgd:%p\n", __func__, addr, pgd);

@@ -217,19 +218,29 @@ pte_t *huge_pte_offset(struct mm_struct *mm,
 		return NULL;
 
 	pud = pud_offset(pgd, addr);
-	if (pud_none(*pud))
+	if (pud_none(*pud) && sz != PUD_SIZE)
 		return NULL;
 	/* swap or huge page */
 	if (!pud_present(*pud) || pud_huge(*pud))
 		return (pte_t *)pud;
 	/* table; check the next level */

So if sz == PUD_SIZE and we have pud_none(*pud) == true, it returns the
pud. Isn't this different from what you proposed for the generic
huge_pte_offset()? [1]

I think I missed this case in the generic version.

As hugetlb_fault() deals with p*d_none() entries by calling
hugetlb_no_page(), the thinking was that returning the p*d saves us an
extra round trip by avoiding the call to huge_pte_alloc().

quoted

 
+	if (sz == CONT_PMD_SIZE)
+		addr &= CONT_PMD_MASK;
+
 	pmd = pmd_offset(pud, addr);
-	if (pmd_none(*pmd))
+	if (pmd_none(*pmd) &&
+	    !(sz == PMD_SIZE || sz == CONT_PMD_SIZE))
 		return NULL;

Again, if sz == PMD_SIZE, you no longer return NULL. The generic
proposal in [1] looks like:

	if (pmd_none(*pmd))
		return NULL;

and that's even when sz == PMD_SIZE.

Anyway, I think we need to push for [1] again to be accepted before we
go ahead with these changes.

[1] is already queued in Andrew's tree. I'll send an update - hopefully
it can be picked up for the next merge.

[1] http://lkml.kernel.org/r/20170725154114.24131-2-punit.agrawal at arm.com