[PATCH v4 1/5] mm: add mkwrite param to vm_insert_mixed()

[ross.zwisler@linux.intel.com (Ross Zwisler)]

On Mon, Jul 24, 2017 at 01:25:30PM +0200, Jan Kara wrote:
> > @@ -1658,14 +1658,28 @@ static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
> >  	if (!pte)
> >  		goto out;
> >  	retval = -EBUSY;
> > -	if (!pte_none(*pte))
> > -		goto out_unlock;
> > +	if (!pte_none(*pte)) {
> > +		if (mkwrite) {
> > +			if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
> 
> Is the WARN_ON_ONCE() really appropriate here? Your testcase with private
> mappings has triggered this situation if I'm right...

Yep, I think this WARN_ON_ONCE() is correct.  The test with private mappings
had collisions between read-only DAX mappings which were being faulted in via
insert_pfn(), and read/write COW page cache mappings which were being faulted
in by wp_page_copy().

I was hitting a false-positive warning when I had the WARN_ON_ONCE() in
insert_pfn() outside of the mkwrite case, i.e.:

	if (!pte_none(*pte)) {
		if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
			goto out_unlock;
		if (mkwrite) {
			entry = *pte;
			goto out_mkwrite;
		} else
			goto out_unlock;
	}

This was triggering when one thread was faulting in a read-only DAX mapping
when another thread had already faulted in a read-write COW page cache page.

The patches I sent out have the warning in the mkwrite case, which would mean
that we were getting a fault for a read/write PTE in insert_pfn() and the PFN
didn't match what was already in the PTE.

This can't ever happen in the private mapping case because we will never
install a read/write PTE for normal storage, only for COW page cache pages.
Essentially I don't think we should ever be able to hit this warning, and if
we do I'd like to get the bug report so that I can track down how it was
happening and make sure that it's safe.  It is in the mkwrite path of
insert_pfn() which is currently only used by the DAX code.

Does that make sense to you, or would you recommend leaving it out?  (If so,
why?)