Re: [PATCH v3 1/2] mm: migration: fix migration of huge PMD shared pages

[Michal Hocko <mhocko@kernel.org>]

On Wed 22-08-18 09:48:16, Mike Kravetz wrote:
> On 08/22/2018 05:28 AM, Michal Hocko wrote:
> > On Tue 21-08-18 18:10:42, Mike Kravetz wrote:
> > [...]
> >> diff --git a/mm/rmap.c b/mm/rmap.c
> >> index eb477809a5c0..8cf853a4b093 100644
> >> --- a/mm/rmap.c
> >> +++ b/mm/rmap.c
> >> @@ -1362,11 +1362,21 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
> >>  	}
> >>  
> >>  	/*
> >> -	 * We have to assume the worse case ie pmd for invalidation. Note that
> >> -	 * the page can not be free in this function as call of try_to_unmap()
> >> -	 * must hold a reference on the page.
> >> +	 * For THP, we have to assume the worse case ie pmd for invalidation.
> >> +	 * For hugetlb, it could be much worse if we need to do pud
> >> +	 * invalidation in the case of pmd sharing.
> >> +	 *
> >> +	 * Note that the page can not be free in this function as call of
> >> +	 * try_to_unmap() must hold a reference on the page.
> >>  	 */
> >>  	end = min(vma->vm_end, start + (PAGE_SIZE << compound_order(page)));
> >> +	if (PageHuge(page)) {
> >> +		/*
> >> +		 * If sharing is possible, start and end will be adjusted
> >> +		 * accordingly.
> >> +		 */
> >> +		(void)huge_pmd_sharing_possible(vma, &start, &end);
> >> +	}
> >>  	mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
> > 
> > I do not get this part. Why don't we simply unconditionally invalidate
> > the whole huge page range?
> 
> In this routine, we are only unmapping a single page.  The existing code
> is limiting the invalidate range to that page size: 4K or 2M.  With shared
> PMDs, we have the possibility of unmapping a PUD_SIZE area: 1G.  I don't
> think we want to unconditionally invalidate 1G.  Is that what you are asking?

But we know that huge_pmd_unshare unmapped a shared pte so we know when
to flush 2MB or 1GB. I really do not like how huge_pmd_sharing_possible
a) duplicates some checks and b) it updates start/stop out of line.

> I do not know how often PMD sharing is exercised.  It certainly is used by
> DBs for large shared areas.  I suspect it is less frequent than hugtlb pages
> in general, and certainly less frequent than THP or base pages.
> 
> >>  
> >>  	while (page_vma_mapped_walk(&pvmw)) {
> >> @@ -1409,6 +1419,32 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
> >>  		subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
> >>  		address = pvmw.address;
> >>  
> >> +		if (PageHuge(page)) {
> >> +			if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
> > 
> > huge_pmd_unshare is documented to require a pte lock. Where do we take
> > it?
> 
> It is somewhat hidden, but we are in the loop:
> 
> 	while (page_vma_mapped_walk(&pvmw)) {
> 
> The routine page_vma_mapped_walk will acquire the lock, and it correctly
> checks for huge pages and uses huge_pte_lockptr().
> 
> page_vma_mapped_walk_done() will release the lock.

OK, I can see it now. Thanks for the clarification. page_vma_mapped_walk
is quite hard to follow.
-- 
Michal Hocko
SUSE Labs