Re: [PATCH v8 3/3] mm: fix double page fault on arm64 if PTE_AF is cleared

[Catalin Marinas]

On Tue, Sep 24, 2019 at 06:43:06AM +0000, Justin He (Arm Technology China) wrote:
> Catalin Marinas wrote:
> > On Sat, Sep 21, 2019 at 09:50:54PM +0800, Jia He wrote:
> > > @@ -2151,21 +2163,53 @@ static inline void cow_user_page(struct page *dst, struct page *src, unsigned lo
> > >  	 * fails, we just zero-fill it. Live with it.
> > >  	 */
> > >  	if (unlikely(!src)) {
> > > -		void *kaddr = kmap_atomic(dst);
> > > -		void __user *uaddr = (void __user *)(va & PAGE_MASK);
> > > +		void *kaddr;
> > > +		pte_t entry;
> > > +		void __user *uaddr = (void __user *)(addr & PAGE_MASK);
> > >
> > > +		/* On architectures with software "accessed" bits, we would
> > > +		 * take a double page fault, so mark it accessed here.
> > > +		 */
[...]
> > > +		if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
> > > +			vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr,
> > > +						       &vmf->ptl);
> > > +			if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
> > > +				entry = pte_mkyoung(vmf->orig_pte);
> > > +				if (ptep_set_access_flags(vma, addr,
> > > +							  vmf->pte, entry, 0))
> > > +					update_mmu_cache(vma, addr, vmf->pte);
> > > +			} else {
> > > +				/* Other thread has already handled the fault
> > > +				 * and we don't need to do anything. If it's
> > > +				 * not the case, the fault will be triggered
> > > +				 * again on the same address.
> > > +				 */
> > > +				pte_unmap_unlock(vmf->pte, vmf->ptl);
> > > +				return false;
> > > +			}
> > > +			pte_unmap_unlock(vmf->pte, vmf->ptl);
> > > +		}
[...]
> > > +
> > > +		kaddr = kmap_atomic(dst);
> > 
> > Since you moved the kmap_atomic() here, could the above
> > arch_faults_on_old_pte() run in a preemptible context? I suggested to
> > add a WARN_ON in patch 2 to be sure.
> 
> Should I move kmap_atomic back to the original line? Thus, we can make sure
> that arch_faults_on_old_pte() is in the context of preempt_disabled?
> Otherwise, arch_faults_on_old_pte() may cause plenty of warning if I add
> a WARN_ON in arch_faults_on_old_pte.  I tested it when I enable the PREEMPT=y
> on a ThunderX2 qemu guest.

So we have two options here:

1. Change arch_faults_on_old_pte() scope to the whole system rather than
   just the current CPU. You'd have to wire up a new arm64 capability
   for the access flag but this way we don't care whether it's
   preemptible or not.

2. Keep the arch_faults_on_old_pte() per-CPU but make sure we are not
   preempted here. The kmap_atomic() move would do but you'd have to
   kunmap_atomic() before the return.

I think the answer to my question below also has some implication on
which option to pick:

> > >  		/*
> > >  		 * This really shouldn't fail, because the page is there
> > >  		 * in the page tables. But it might just be unreadable,
> > >  		 * in which case we just give up and fill the result with
> > >  		 * zeroes.
> > >  		 */
> > > -		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
> > > +		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
> > > +			/* Give a warn in case there can be some obscure
> > > +			 * use-case
> > > +			 */
> > > +			WARN_ON_ONCE(1);
> > 
> > That's more of a question for the mm guys: at this point we do the
> > copying with the ptl released; is there anything else that could have
> > made the pte old in the meantime? I think unuse_pte() is only called on
> > anonymous vmas, so it shouldn't be the case here.

If we need to hold the ptl here, you could as well have an enclosing
kmap/kunmap_atomic (option 2) with some goto instead of "return false".

-- 
Catalin