Re: [PATCH v2 14/26] userfaultfd: wp: handle COW properly for uffd-wp

[Jerome Glisse <jglisse@redhat.com>]

On Fri, Feb 22, 2019 at 04:46:03PM +0800, Peter Xu wrote:
> On Thu, Feb 21, 2019 at 01:04:24PM -0500, Jerome Glisse wrote:
> > On Tue, Feb 12, 2019 at 10:56:20AM +0800, Peter Xu wrote:
> > > This allows uffd-wp to support write-protected pages for COW.

[...]

> > > diff --git a/mm/mprotect.c b/mm/mprotect.c
> > > index 9d4433044c21..ae93721f3795 100644
> > > --- a/mm/mprotect.c
> > > +++ b/mm/mprotect.c
> > > @@ -77,14 +77,13 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
> > >  		if (pte_present(oldpte)) {
> > >  			pte_t ptent;
> > >  			bool preserve_write = prot_numa && pte_write(oldpte);
> > > +			struct page *page;
> > >  
> > >  			/*
> > >  			 * Avoid trapping faults against the zero or KSM
> > >  			 * pages. See similar comment in change_huge_pmd.
> > >  			 */
> > >  			if (prot_numa) {
> > > -				struct page *page;
> > > -
> > >  				page = vm_normal_page(vma, addr, oldpte);
> > >  				if (!page || PageKsm(page))
> > >  					continue;
> > > @@ -114,6 +113,46 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
> > >  					continue;
> > >  			}
> > >  
> > > +			/*
> > > +			 * Detect whether we'll need to COW before
> > > +			 * resolving an uffd-wp fault.  Note that this
> > > +			 * includes detection of the zero page (where
> > > +			 * page==NULL)
> > > +			 */
> > > +			if (uffd_wp_resolve) {
> > > +				/* If the fault is resolved already, skip */
> > > +				if (!pte_uffd_wp(*pte))
> > > +					continue;
> > > +				page = vm_normal_page(vma, addr, oldpte);
> > > +				if (!page || page_mapcount(page) > 1) {
> > 
> > This is wrong, if you allow page to be NULL then you gonna segfault
> > in wp_page_copy() down below. Are you sure you want to test for
> > special page ? For anonymous memory this should never happens ie
> > anon page always are regular page. So if you allow userfaulfd to
> > write protect only anonymous vma then there is no point in testing
> > here beside maybe a BUG_ON() just in case ...
> 
> It's majorly for zero pages where page can be NULL.  Would this be
> clearer:
> 
>   if (is_zero_pfn(pte_pfn(old_pte)) || (page && page_mapcount(page)))
> 
> ?
> 
> Now we treat zero pages as normal COW pages so we'll do COW here even
> for zero pages.  I think maybe we can do special handling on all over
> the places for zero pages (e.g., we don't write protect a PTE if we
> detected that this is the zero PFN) but I'm uncertain on whether
> that's what we want, so I chose to start with current solution at
> least to achieve functionality first.

You can keep the vm_normal_page() in that case but split the if
between page == NULL and page != NULL with mapcount > 1. As other-
wise you will segfault below.


> 
> > 
> > > +					struct vm_fault vmf = {
> > > +						.vma = vma,
> > > +						.address = addr & PAGE_MASK,
> > > +						.page = page,
> > > +						.orig_pte = oldpte,
> > > +						.pmd = pmd,
> > > +						/* pte and ptl not needed */
> > > +					};
> > > +					vm_fault_t ret;
> > > +
> > > +					if (page)
> > > +						get_page(page);
> > > +					arch_leave_lazy_mmu_mode();
> > > +					pte_unmap_unlock(pte, ptl);
> > > +					ret = wp_page_copy(&vmf);
> > > +					/* PTE is changed, or OOM */
> > > +					if (ret == 0)
> > > +						/* It's done by others */
> > > +						continue;
> > > +					else if (WARN_ON(ret != VM_FAULT_WRITE))
> > > +						return pages;
> > > +					pte = pte_offset_map_lock(vma->vm_mm,
> > > +								  pmd, addr,
> > > +								  &ptl);
> > 
> > Here you remap the pte locked but you are not checking if the pte is
> > the one you expect ie is it pointing to the copied page and does it
> > have expect uffd_wp flag. Another thread might have raced between the
> > time you called wp_page_copy() and the time you pte_offset_map_lock()
> > I have not check the mmap_sem so maybe you are protected by it as
> > mprotect is taking it in write mode IIRC, if so you should add a
> > comments at very least so people do not see this as a bug.
> 
> Thanks for spotting this.  With nornal uffd-wp page fault handling
> path we're only with read lock held (and I would suspect it's racy
> even with write lock...).  I agree that there can be a race right
> after the COW has done.
> 
> Here IMHO we'll be fine as long as it's still a present PTE, in other
> words, we should be able to tolerate PTE changes as long as it's still
> present otherwise we'll need to retry this single PTE (e.g., the page
> can be quickly marked as migrating swap entry, or even the page could
> be freed beneath us).  Do you think below change look good to you to
> be squashed into this patch?

Ok, but below if must be after arch_enter_lazy_mmu_mode(); not before.

> 
> diff --git a/mm/mprotect.c b/mm/mprotect.c
> index 73a65f07fe41..3423f9692838 100644
> --- a/mm/mprotect.c
> +++ b/mm/mprotect.c
> @@ -73,6 +73,7 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,                                                              
>         flush_tlb_batched_pending(vma->vm_mm);
>         arch_enter_lazy_mmu_mode();
>         do {
> +retry_pte:
>                 oldpte = *pte;
>                 if (pte_present(oldpte)) {
>                         pte_t ptent;
> @@ -149,6 +150,13 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,                                                           
>                                         pte = pte_offset_map_lock(vma->vm_mm,
>                                                                   pmd, addr,
>                                                                   &ptl);
> +                                       if (!pte_present(*pte))
> +                                               /*
> +                                                * This PTE could have
> +                                                * been modified when COW;
> +                                                * retry it
> +                                                */
> +                                               goto retry_pte;
>                                         arch_enter_lazy_mmu_mode();
>                                 }
>                         }