Re: [PATCH v3 14/28] userfaultfd: wp: handle COW properly for uffd-wp

[Jerome Glisse <jglisse@redhat.com>]

On Tue, Apr 23, 2019 at 11:00:30AM +0800, Peter Xu wrote:
> On Mon, Apr 22, 2019 at 10:54:02AM -0400, Jerome Glisse wrote:
> > On Mon, Apr 22, 2019 at 08:20:10PM +0800, Peter Xu wrote:
> > > On Fri, Apr 19, 2019 at 11:02:53AM -0400, Jerome Glisse wrote:
> > > 
> > > [...]
> > > 
> > > > > > > +			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) {
> > > > > > > +					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;
> > > > > > 
> > > > > > This is wrong if ret == 0 you still need to remap the pte before
> > > > > > continuing as otherwise you will go to next pte without the page
> > > > > > table lock for the directory. So 0 case must be handled after
> > > > > > arch_enter_lazy_mmu_mode() below.
> > > > > > 
> > > > > > Sorry i should have catch that in previous review.
> > > > > 
> > > > > My fault to not have noticed it since the very beginning... thanks for
> > > > > spotting that.
> > > > > 
> > > > > I'm squashing below changes into the patch:
> > > > 
> > > > 
> > > > Well thinking of this some more i think you should use do_wp_page() and
> > > > not wp_page_copy() it would avoid bunch of code above and also you are
> > > > not properly handling KSM page or page in the swap cache. Instead of
> > > > duplicating same code that is in do_wp_page() it would be better to call
> > > > it here.
> > > 
> > > Yeah it makes sense to me.  Then here's my plan:
> > > 
> > > - I'll need to drop previous patch "export wp_page_copy" since then
> > >   it'll be not needed
> > > 
> > > - I'll introduce another patch to split current do_wp_page() and
> > >   introduce function "wp_page_copy_cont" (better suggestion on the
> > >   naming would be welcomed) which contains most of the wp handling
> > >   that'll be needed for change_pte_range() in this patch and isolate
> > >   the uffd handling:
> > > 
> > > static vm_fault_t do_wp_page(struct vm_fault *vmf)
> > > 	__releases(vmf->ptl)
> > > {
> > > 	struct vm_area_struct *vma = vmf->vma;
> > > 
> > > 	if (userfaultfd_pte_wp(vma, *vmf->pte)) {
> > > 		pte_unmap_unlock(vmf->pte, vmf->ptl);
> > > 		return handle_userfault(vmf, VM_UFFD_WP);
> > > 	}
> > > 
> > > 	return do_wp_page_cont(vmf);
> > > }
> > > 
> > > Then I can probably use do_wp_page_cont() in this patch.
> > 
> > Instead i would keep the do_wp_page name and do:
> >     static vm_fault_t do_userfaultfd_wp_page(struct vm_fault *vmf) {
> >         ... // what you have above
> >         return do_wp_page(vmf);
> >     }
> > 
> > Naming wise i think it would be better to keep do_wp_page() as
> > is.
> 
> In case I misunderstood... what I've proposed will be simply:
> 
> diff --git a/mm/memory.c b/mm/memory.c
> index 64bd8075f054..ab98a1eb4702 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -2497,6 +2497,14 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
>                 return handle_userfault(vmf, VM_UFFD_WP);
>         }
> 
> +       return do_wp_page_cont(vmf);
> +}
> +
> +vm_fault_t do_wp_page_cont(struct vm_fault *vmf)
> +       __releases(vmf->ptl)
> +{
> +       struct vm_area_struct *vma = vmf->vma;
> +
>         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
>         if (!vmf->page) {
>                 /*
> 
> And the other proposal is:
> 
> diff --git a/mm/memory.c b/mm/memory.c
> index 64bd8075f054..a73792127553 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -2469,6 +2469,8 @@ static vm_fault_t wp_page_shared(struct vm_fault *vmf)
>         return VM_FAULT_WRITE;
>  }
> 
> +static vm_fault_t do_wp_page(struct vm_fault *vmf);
> +
>  /*
>   * This routine handles present pages, when users try to write
>   * to a shared page. It is done by copying the page to a new address
> @@ -2487,7 +2489,7 @@ static vm_fault_t wp_page_shared(struct vm_fault *vmf)
>   * but allow concurrent faults), with pte both mapped and locked.
>   * We return with mmap_sem still held, but pte unmapped and unlocked.
>   */
> -static vm_fault_t do_wp_page(struct vm_fault *vmf)
> +static vm_fault_t do_userfaultfd_wp_page(struct vm_fault *vmf)
>         __releases(vmf->ptl)
>  {
>         struct vm_area_struct *vma = vmf->vma;
> @@ -2497,6 +2499,14 @@ static vm_fault_t do_wp_page(struct vm_fault *vmf)
>                 return handle_userfault(vmf, VM_UFFD_WP);
>         }
> 
> +       return do_wp_page(vmf);
> +}
> +
> +static vm_fault_t do_wp_page(struct vm_fault *vmf)
> +       __releases(vmf->ptl)
> +{
> +       struct vm_area_struct *vma = vmf->vma;
> +
>         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
>         if (!vmf->page) {
>                 /*
> @@ -2869,7 +2879,7 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
>         }
> 
>         if (vmf->flags & FAULT_FLAG_WRITE) {
> -               ret |= do_wp_page(vmf);
> +               ret |= do_userfaultfd_wp_page(vmf);
>                 if (ret & VM_FAULT_ERROR)
>                         ret &= VM_FAULT_ERROR;
>                 goto out;
> @@ -3831,7 +3841,7 @@ static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
>                 goto unlock;
>         if (vmf->flags & FAULT_FLAG_WRITE) {
>                 if (!pte_write(entry))
> -                       return do_wp_page(vmf);
> +                       return do_userfaultfd_wp_page(vmf);
>                 entry = pte_mkdirty(entry);
>         }
>         entry = pte_mkyoung(entry);
> 
> I would prefer the 1st approach since it not only contains fewer lines
> of changes because it does not touch callers, and also the naming in
> the 2nd approach can be a bit confusing (calling
> do_userfaultfd_wp_page in handle_pte_fault may let people think of an
> userfault-only path but actually it covers the general path).  But if
> you really like the 2nd one I can use that too.

Maybe move the userfaultfd code to a small helper, call it first in
call site of do_wp_page() and do_wp_page() if it does not fire ie:

bool do_userfaultfd_wp(struct vm_fault *vmf, int ret)
{
    if (handleuserfault) return true;
    return false;
}

then
     if (vmf->flags & FAULT_FLAG_WRITE) {
            if (do_userfaultfd_wp(vmf, tmp)) {
                ret |= tmp;
            } else
                ret |= do_wp_page(vmf);
            if (ret & VM_FAULT_ERROR)
                ret &= VM_FAULT_ERROR;
            goto out;

and:
    if (vmf->flags & FAULT_FLAG_WRITE) {
        if (!pte_write(entry)) {
            if (do_userfaultfd_wp(vmf, ret))
                return ret;
            else
                return do_wp_page(vmf);
        }

Cheers,
Jérôme