Re: [PATCH v4 3/3] mm: fix double page fault on arm64 if PTE_AF is cleared

[Kirill A. Shutemov]

On Thu, Sep 19, 2019 at 10:16:34AM +0800, Jia He wrote:
> Hi Kirill
> 
> [On behalf of justin.he@xxxxxxx because some mails are filted...]
> 
> On 2019/9/18 22:00, Kirill A. Shutemov wrote:
> > On Wed, Sep 18, 2019 at 09:19:14PM +0800, Jia He wrote:
> > > When we tested pmdk unit test [1] vmmalloc_fork TEST1 in arm64 guest, there
> > > will be a double page fault in __copy_from_user_inatomic of cow_user_page.
> > > 
> > > Below call trace is from arm64 do_page_fault for debugging purpose
> > > [  110.016195] Call trace:
> > > [  110.016826]  do_page_fault+0x5a4/0x690
> > > [  110.017812]  do_mem_abort+0x50/0xb0
> > > [  110.018726]  el1_da+0x20/0xc4
> > > [  110.019492]  __arch_copy_from_user+0x180/0x280
> > > [  110.020646]  do_wp_page+0xb0/0x860
> > > [  110.021517]  __handle_mm_fault+0x994/0x1338
> > > [  110.022606]  handle_mm_fault+0xe8/0x180
> > > [  110.023584]  do_page_fault+0x240/0x690
> > > [  110.024535]  do_mem_abort+0x50/0xb0
> > > [  110.025423]  el0_da+0x20/0x24
> > > 
> > > The pte info before __copy_from_user_inatomic is (PTE_AF is cleared):
> > > [ffff9b007000] pgd=000000023d4f8003, pud=000000023da9b003, pmd=000000023d4b3003, pte=360000298607bd3
> > > 
> > > As told by Catalin: "On arm64 without hardware Access Flag, copying from
> > > user will fail because the pte is old and cannot be marked young. So we
> > > always end up with zeroed page after fork() + CoW for pfn mappings. we
> > > don't always have a hardware-managed access flag on arm64."
> > > 
> > > This patch fix it by calling pte_mkyoung. Also, the parameter is
> > > changed because vmf should be passed to cow_user_page()
> > > 
> > > [1] https://github.com/pmem/pmdk/tree/master/src/test/vmmalloc_fork
> > > 
> > > Reported-by: Yibo Cai <Yibo.Cai@xxxxxxx>
> > > Signed-off-by: Jia He <justin.he@xxxxxxx>
> > > ---
> > >   mm/memory.c | 35 ++++++++++++++++++++++++++++++-----
> > >   1 file changed, 30 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/mm/memory.c b/mm/memory.c
> > > index e2bb51b6242e..d2c130a5883b 100644
> > > --- a/mm/memory.c
> > > +++ b/mm/memory.c
> > > @@ -118,6 +118,13 @@ int randomize_va_space __read_mostly =
> > >   					2;
> > >   #endif
> > > +#ifndef arch_faults_on_old_pte
> > > +static inline bool arch_faults_on_old_pte(void)
> > > +{
> > > +	return false;
> > > +}
> > > +#endif
> > > +
> > >   static int __init disable_randmaps(char *s)
> > >   {
> > >   	randomize_va_space = 0;
> > > @@ -2140,8 +2147,12 @@ static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
> > >   	return same;
> > >   }
> > > -static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
> > > +static inline void cow_user_page(struct page *dst, struct page *src,
> > > +				 struct vm_fault *vmf)
> > >   {
> > > +	struct vm_area_struct *vma = vmf->vma;
> > > +	unsigned long addr = vmf->address;
> > > +
> > >   	debug_dma_assert_idle(src);
> > >   	/*
> > > @@ -2152,20 +2163,34 @@ static inline void cow_user_page(struct page *dst, struct page *src, unsigned lo
> > >   	 */
> > >   	if (unlikely(!src)) {
> > >   		void *kaddr = kmap_atomic(dst);
> > > -		void __user *uaddr = (void __user *)(va & PAGE_MASK);
> > > +		void __user *uaddr = (void __user *)(addr & PAGE_MASK);
> > > +		pte_t entry;
> > >   		/*
> > >   		 * 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.
> > > +		 * zeroes. On architectures with software "accessed" bits,
> > > +		 * we would take a double page fault here, so mark it
> > > +		 * accessed here.
> > >   		 */
> > > +		if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
> > > +			spin_lock(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);
> > > +			}
> > I don't follow.
> > 
> > So if pte has changed under you, you don't set the accessed bit, but never
> > the less copy from the user.
> > 
> > What makes you think it will not trigger the same problem?
> > 
> > I think we need to make cow_user_page() fail in this case and caller --
> > wp_page_copy() -- return zero. If the fault was solved by other thread, we
> > are fine. If not userspace would re-fault on the same address and we will
> > handle the fault from the second attempt.
> 
> Thanks for the pointing. How about make cow_user_page() be returned
> 
> VM_FAULT_RETRY? Then in do_page_fault(), it can retry the page fault?

No. VM_FAULT_RETRY has different semantics: we have to drop mmap_sem(), so
let's try to take it again and handle the fault. In this case the more
likely scenario is that 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.

-- 
 Kirill A. Shutemov