Re: [PATCH v16 2/5] fs/proc/task_mmu: Implement IOCTL to get and optionally clear info about PTEs

[Peter Xu <peterx@redhat.com>]

Muhammad,

Sorry, I probably can only review the non-interface part, and leave the
interface/buffer handling, etc. review for others and real potential users
of it..

On Thu, May 25, 2023 at 01:55:14PM +0500, Muhammad Usama Anjum wrote:
> +static inline void make_uffd_wp_huge_pte(struct vm_area_struct *vma,
> +					 unsigned long addr, pte_t *ptep,
> +					 pte_t ptent)
> +{
> +	pte_t old_pte;
> +
> +	if (!huge_pte_none(ptent)) {
> +		old_pte = huge_ptep_modify_prot_start(vma, addr, ptep);
> +		ptent = huge_pte_mkuffd_wp(old_pte);
> +		ptep_modify_prot_commit(vma, addr, ptep, old_pte, ptent);

huge_ptep_modify_prot_start()?

The other thing is what if it's a pte marker already?  What if a hugetlb
migration entry?  Please check hugetlb_change_protection().

> +	} else {
> +		set_huge_pte_at(vma->vm_mm, addr, ptep,
> +				make_pte_marker(PTE_MARKER_UFFD_WP));
> +	}
> +}
> +#endif

[...]

> +static int pagemap_scan_pmd_entry(pmd_t *pmd, unsigned long start,
> +				  unsigned long end, struct mm_walk *walk)
> +{
> +	struct pagemap_scan_private *p = walk->private;
> +	struct vm_area_struct *vma = walk->vma;
> +	unsigned long addr = end;
> +	pte_t *pte, *orig_pte;
> +	spinlock_t *ptl;
> +	bool is_written;
> +	int ret = 0;
> +
> +	arch_enter_lazy_mmu_mode();
> +
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	ptl = pmd_trans_huge_lock(pmd, vma);
> +	if (ptl) {
> +		unsigned long n_pages = (end - start)/PAGE_SIZE;
> +
> +		if (p->max_pages && n_pages > p->max_pages - p->found_pages)
> +			n_pages = p->max_pages - p->found_pages;
> +
> +		is_written = !is_pmd_uffd_wp(*pmd);
> +
> +		/*
> +		 * Break huge page into small pages if the WP operation need to
> +		 * be performed is on a portion of the huge page.
> +		 */
> +		if (is_written && IS_PM_SCAN_WP(p->flags) &&
> +		    n_pages < HPAGE_SIZE/PAGE_SIZE) {
> +			spin_unlock(ptl);
> +
> +			split_huge_pmd(vma, pmd, start);
> +			goto process_smaller_pages;
> +		}
> +
> +		if (IS_PM_SCAN_GET(p->flags))
> +			ret = pagemap_scan_output(is_written, vma->vm_file,
> +						  pmd_present(*pmd),
> +						  is_swap_pmd(*pmd),
> +						  p, start, n_pages);
> +
> +		if (ret >= 0 && is_written && IS_PM_SCAN_WP(p->flags))
> +			make_uffd_wp_pmd(vma, addr, pmd);
> +
> +		if (IS_PM_SCAN_WP(p->flags))
> +			flush_tlb_range(vma, start, end);
> +
> +		spin_unlock(ptl);
> +
> +		arch_leave_lazy_mmu_mode();
> +		return ret;
> +	}
> +
> +process_smaller_pages:
> +	if (pmd_trans_unstable(pmd)) {
> +		arch_leave_lazy_mmu_mode();
> +		return 0;

I'm not sure whether this is right..  Shouldn't you return with -EAGAIN and
let the user retry?  Returning 0 means you'll move on with the next pmd
afaict and ignoring this one.

> +	}
> +#endif
> +
> +	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);

Just a heads-up that this may start to fail at some point if Hugh's work
will land earlier:

https://lore.kernel.org/linux-mm/68a97fbe-5c1e-7ac6-72c-7b9c6290b370@google.com/

> +	for (addr = start; addr < end && !ret; pte++, addr += PAGE_SIZE) {
> +		is_written = !is_pte_uffd_wp(*pte);
> +
> +		if (IS_PM_SCAN_GET(p->flags))
> +			ret = pagemap_scan_output(is_written, vma->vm_file,
> +						  pte_present(*pte),
> +						  is_swap_pte(*pte),
> +						  p, addr, 1);
> +
> +		if (ret >= 0 && is_written && IS_PM_SCAN_WP(p->flags))
> +			make_uffd_wp_pte(vma, addr, pte);
> +	}
> +
> +	if (IS_PM_SCAN_WP(p->flags))
> +		flush_tlb_range(vma, start, addr);
> +
> +	pte_unmap_unlock(orig_pte, ptl);
> +	arch_leave_lazy_mmu_mode();
> +
> +	cond_resched();
> +	return ret;
> +}
> +
> +#ifdef CONFIG_HUGETLB_PAGE
> +static int pagemap_scan_hugetlb_entry(pte_t *ptep, unsigned long hmask,
> +				      unsigned long start, unsigned long end,
> +				      struct mm_walk *walk)
> +{
> +	unsigned long n_pages = (end - start)/PAGE_SIZE;
> +	struct pagemap_scan_private *p = walk->private;
> +	struct vm_area_struct *vma = walk->vma;
> +	struct hstate *h = hstate_vma(vma);
> +	spinlock_t *ptl;
> +	bool is_written;
> +	int ret = 0;
> +	pte_t pte;
> +
> +	arch_enter_lazy_mmu_mode();

This _seems_ to be not needed for hugetlb entries.

> +
> +	if (p->max_pages && n_pages > p->max_pages - p->found_pages)
> +		n_pages = p->max_pages - p->found_pages;
> +
> +	if (IS_PM_SCAN_WP(p->flags)) {
> +		i_mmap_lock_write(vma->vm_file->f_mapping);
> +		ptl = huge_pte_lock(h, vma->vm_mm, ptep);
> +	}
> +
> +	pte = huge_ptep_get(ptep);
> +	is_written = !is_huge_pte_uffd_wp(pte);
> +
> +	/*
> +	 * Partial hugetlb page clear isn't supported
> +	 */
> +	if (is_written && IS_PM_SCAN_WP(p->flags) &&
> +	    n_pages < HPAGE_SIZE/PAGE_SIZE) {
> +		ret = -EPERM;
> +		goto unlock_and_return;
> +	}
> +
> +	if (IS_PM_SCAN_GET(p->flags)) {
> +		ret = pagemap_scan_output(is_written, vma->vm_file,
> +					  pte_present(pte), is_swap_pte(pte),
> +					  p, start, n_pages);
> +		if (ret < 0)
> +			goto unlock_and_return;
> +	}
> +
> +	if (is_written && IS_PM_SCAN_WP(p->flags)) {
> +		make_uffd_wp_huge_pte(vma, start, ptep, pte);
> +		flush_hugetlb_tlb_range(vma, start, end);
> +	}
> +
> +unlock_and_return:
> +	if (IS_PM_SCAN_WP(p->flags)) {
> +		spin_unlock(ptl);
> +		i_mmap_unlock_write(vma->vm_file->f_mapping);
> +	}
> +
> +	arch_leave_lazy_mmu_mode();

Same here.

> +
> +	return ret;
> +}

[...]

-- 
Peter Xu