On 8/23/2022 4:14 PM, David Hildenbrand wrote:
On 23.08.22 09:50, Baolin Wang wrote:Considering the pmd entries of a CONT-PMD hugetlb can not span on multiple PMDs, we can change to use the PMD page lock, which can be much finer grain that lock in the mm. Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> --- include/linux/hugetlb.h | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h index 3a96f67..d4803a89 100644 --- a/include/linux/hugetlb.h +++ b/include/linux/hugetlb.h @@ -892,9 +892,17 @@ static inline gfp_t htlb_modify_alloc_mask(struct hstate *h, gfp_t gfp_mask) static inline spinlock_t *huge_pte_lockptr(struct hstate *h, struct mm_struct *mm, pte_t *pte) { - VM_BUG_ON(huge_page_size(h) == PAGE_SIZE); + unsigned long hp_size = huge_page_size(h);- if (huge_page_size(h) == PMD_SIZE) {+ VM_BUG_ON(hp_size == PAGE_SIZE); + + /* + * Considering CONT-PMD size hugetlb, since the CONT-PMD entry + * can not span multiple PMDs, then we can use the fine grained + * PMD page lock. + */ + if (hp_size == PMD_SIZE || + (hp_size > PMD_SIZE && hp_size < PUD_SIZE)) { return pmd_lockptr(mm, (pmd_t *) pte); } else if (huge_page_size(h) < PMD_SIZE) { unsigned long mask = ~(PTRS_PER_PTE * sizeof(pte_t) - 1);Is there a measurable performance gain? IOW, do we really care?
IMO, It's just a theoretical analysis now:) Let me think about how to measure the performance gain.
