Re: interaction of MADV_PAGEOUT with CoW anonymous mappings?

[Minchan Kim <minchan@kernel.org>]

On Fri, Mar 13, 2020 at 09:59:50AM -0700, Dave Hansen wrote:
> On 3/12/20 7:00 PM, Minchan Kim wrote:
> > On Thu, Mar 12, 2020 at 02:41:07PM -0700, Dave Hansen wrote:
> >> One other fun thing.  I have a "victim" thread sitting in a loop doing:
> >>
> >> 	sleep(1)
> >> 	memcpy(&garbage, buffer, sz);
> >>
> >> The "attacker" is doing
> >>
> >> 	madvise(buffer, sz, MADV_PAGEOUT);
> >>
> >> in a loop.  That, oddly enough doesn't cause the victim to page fault.
> >> But, if I do:
> >>
> >> 	memcpy(&garbage, buffer, sz);
> >> 	madvise(buffer, sz, MADV_PAGEOUT);
> >>
> >> It *does* cause the memory to get paged out.  The MADV_PAGEOUT code
> >> actually has a !pte_present() check.  It will punt on a PTE if it sees
> >> it.  In other words, if a page is in the swap cache but not mapped by a
> >> pte_present() PTE, MADV_PAGEOUT won't touch it.
> >>
> >> Shouldn't MADV_PAGEOUT be able to find and reclaim those pages?  Patch
> >> attached.
> > 
> >>
> >>
> >> ---
> >>
> >>  b/mm/madvise.c |   38 +++++++++++++++++++++++++++++++-------
> >>  1 file changed, 31 insertions(+), 7 deletions(-)
> >>
> >> diff -puN mm/madvise.c~madv-pageout-find-swap-cache mm/madvise.c
> >> --- a/mm/madvise.c~madv-pageout-find-swap-cache	2020-03-12 14:24:45.178775035 -0700
> >> +++ b/mm/madvise.c	2020-03-12 14:35:49.706773378 -0700
> >> @@ -248,6 +248,36 @@ static void force_shm_swapin_readahead(s
> >>  #endif		/* CONFIG_SWAP */
> >>  
> >>  /*
> >> + * Given a PTE, find the corresponding 'struct page'.  Also handles
> >> + * non-present swap PTEs.
> >> + */
> >> +struct page *pte_to_reclaim_page(struct vm_area_struct *vma,
> >> +				 unsigned long addr, pte_t ptent)
> >> +{
> >> +	swp_entry_t entry;
> >> +
> >> +	/* Totally empty PTE: */
> >> +	if (pte_none(ptent))
> >> +		return NULL;
> >> +
> >> +	/* A normal, present page is mapped: */
> >> +	if (pte_present(ptent))
> >> +		return vm_normal_page(vma, addr, ptent);
> >> +
> > 
> > Please check is_swap_pte first.
> 
> Why?
> 
> is_swap_pte() duplicates the first two checks.  But, I need an explicit
> pte_present() check somewhere because I need to call vm_normal_page()
> only on present PTEs.
> 
> I guess the pte_present() check could be:
> 
> 	if (!is_swap_pte(ptent))
> 		return vm_normal_page(...);
> 
> *after* the pte_none() check.

Yub, I thought is_swap_pte looks more readable and maintainable for
the change of pte encoding in future. Anyway, I am not insisting.

> 
> >> +	entry = pte_to_swp_entry(vmf->orig_pte);
> >> +	/* Is it one of the "swap PTEs" that's not really swap? */
> >> +	if (non_swap_entry(entry))
> >> +		return false;
> >> +
> >> +	/*
> >> +	 * The PTE was a true swap entry.  The page may be in the
> >> +	 * swap cache.  If so, find it and return it so it may be
> >> +	 * reclaimed.
> >> +	 */
> >> +	return lookup_swap_cache(entry, vma, addr);
> > 
> > If we go with handling only exclusived owned page for anon,
> > I think we should apply the rule to swap cache, too.
> 
> I'm going back and forth on it.  If we're just trying to avoid causing
> faults in other processes, we could add a mapcount>0 check here in
> addition to the mapcount>1 checks that were added in the other patch.
> 
> But, if we want a check for true exclusivity: no other swap entries or
> mappings, we need to check swap_count() too.  It's getting quite a bit
> uglier as I add that it, but I guess we'll see how it looks in the end.

If we go to the map_count > 1 check here and follows the Daniel's suggestion
of MADV_PAGEOUT_ALL to make shared page paging out, that means it clearly
makes semantic change for MADV_PAGEOUT: "paging out only exclusive owned page"
so it would be rather weired if we reclaim swap_count() > 1 of swap cache.
> 
> > Do you mind posting it as formal patch?
> 
> Yeah, I'll send something out.

Thanks for bring up the issue, Dave!