Re: [PATCH V6 01/16] perf/core: Add PERF_SAMPLE_DATA_PAGE_SIZE

[Dave Hansen]

On 8/10/20 2:24 PM, Kan Liang wrote:
> +static u64 __perf_get_page_size(struct mm_struct *mm, unsigned long addr)
> +{
> +	struct page *page;
> +	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 (defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE))
> +	if (p4d_leaf(*p4d)) {
> +		page = p4d_page(*p4d);
> +
> +		if (PageCompound(page))
> +			return page_size(compound_head(page));
> +
> +		return P4D_SIZE;
> +	}
> +#endif
> +
> +	pud = pud_offset(p4d, addr);
> +	if (!pud_present(*pud))
> +		return 0;
> +
> +#if (defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE))
> +	if (pud_leaf(*pud)) {
> +		page = pud_page(*pud);
> +
> +		if (PageCompound(page))
> +			return page_size(compound_head(page));
> +
> +		return PUD_SIZE;
> +	}
> +#endif
> +
> +	pmd = pmd_offset(pud, addr);
> +	if (!pmd_present(*pmd))
> +		return 0;
> +
> +#if (defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE))
> +	if (pmd_leaf(*pmd)) {
> +		page = pmd_page(*pmd);
> +
> +		if (PageCompound(page))
> +			return page_size(compound_head(page));
> +
> +		return PMD_SIZE;
> +	}
> +#endif
> +
> +	pte = pte_offset_map(pmd, addr);
> +	if (!pte_present(*pte)) {
> +		pte_unmap(pte);
> +		return 0;
> +	}
> +
> +	pte_unmap(pte);
> +	return PAGE_SIZE;
> +}

It's probably best if we very carefully define up front what is getting
reported here.  For instance, I believe we already have some fun cases
with huge tmpfs where a compound page is mapped with 4k PTEs.  Kirill
also found a few drivers doing this as well.  I think there were also
some weird cases for ARM hugetlbfs where there were multiple hardware
page table entries mapping a single hugetlbfs page.  These would be
cases where compound_head() size would be greater than the size of the
leaf paging structure entry.

This is also why we have KerelPageSize and MMUPageSize in /proc/$pid/smaps.

So, is this returning the kernel software page size or the MMU size?