Re: [PATCH V4 01/13] perf/core, x86: Add PERF_SAMPLE_DATA_PAGE_SIZE

[Liang, Kan]

On 2/1/2019 7:43 AM, Peter Zijlstra wrote:

On Fri, Feb 01, 2019 at 01:36:00PM +0300, Kirill A. Shutemov wrote:

On Fri, Feb 01, 2019 at 11:03:58AM +0100, Peter Zijlstra wrote:

Will just mentioned a lovely feature where some archs have multi entry
large pages.

Possible something like:

        if (pud_large(*pud)) {
                struct page *page = pud_page(*pud);
                int order = PUD_SHIFT;

                if (PageHuge(page)) {
                        page = compound_head(page);
                        order += compound_order(page);
                }

                return 1ULL << order;
        }

works correctly.

For more fun: some compound pages can be mapped withe page table entries
not matching it's compound size, i.e. 2M pages mapped with PTE.

Surely not for PageHuge() ?! I thought the point of hugetlbfs was to
guarantee page granularity.

How is the below?

static u64 __perf_get_page_size(struct mm_struct *mm, unsigned long addr)
{
        pgd_t *pgd;
        p4d_t *p4d;
        pud_t *pud;
        pmd_t *pmd;
        pte_t *pte;

        pgd = pgd_offset(mm, addr);
        if (pgd_none(*pgd))
                return 0;

        p4d = p4d_offset(pgd, addr);
        if (!p4d_present(*p4d))
                return 0;

        if (p4d_large(*p4d)) {

This one looks like x86 specific?

Kan

                struct page *page = p4d_page(*p4d);
                int shift = P4D_SHIFT;

                if (PageHuge(page)) {
                        page = compound_head(page);
                        shift = PAGE_SHIFT + compound_order(page);
                }

                return 1ULL << shift;
        }

        if (!p4d_present(*p4d))
                return 0;

        pud = pud_offset(p4d, addr);
        if (!pud_present(*pud))
                return 0;

        if (pud_large(*pud)) {
                struct page *page = pud_page(*pud);
                int shift = P4D_SHIFT;

                if (PageHuge(page)) {
                        page = compound_head(page);
                        shift = PAGE_SHIFT + compound_order(page);
                }

                return 1ULL << shift;
        }

        pmd = pmd_offset(pud, addr);
        if (!pmd_present(*pmd))
                return 0;

        if (pmd_large(*pmd)) {
                struct page *page = pud_page(*pud);
                int shift = P4D_SHIFT;

                if (PageHuge(page)) {
                        page = compound_head(page);
                        shift = PAGE_SHIFT + compound_order(page);
                }

                return 1ULL << shift;
        }

        pte = pte_offset_map(pmd, addr);
        if (!pte_present(*pte)) {
                pte_unmap(pte);
                return 0;
        }

        pte_unmap(pte);
        return PAGE_SIZE;
}