Re: [PATCH 02/11] mm/memory: Remove page fault assumption of compound page size

[Matthew Wilcox <willy@infradead.org>]

On Wed, Sep 09, 2020 at 05:29:04PM +0300, Kirill A. Shutemov wrote:
> On Tue, Sep 08, 2020 at 08:55:29PM +0100, Matthew Wilcox (Oracle) wrote:
> > A compound page in the page cache will not necessarily be of PMD size,
> > so check explicitly.
> > 
> > Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
> > ---
> >  mm/memory.c | 7 ++++---
> >  1 file changed, 4 insertions(+), 3 deletions(-)
> > 
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 602f4283122f..4b35b4e71e64 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -3562,13 +3562,14 @@ static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
> >  	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
> >  	pmd_t entry;
> >  	int i;
> > -	vm_fault_t ret;
> > +	vm_fault_t ret = VM_FAULT_FALLBACK;
> >  
> >  	if (!transhuge_vma_suitable(vma, haddr))
> > -		return VM_FAULT_FALLBACK;
> > +		return ret;
> >  
> > -	ret = VM_FAULT_FALLBACK;
> >  	page = compound_head(page);
> > +	if (page_order(page) != HPAGE_PMD_ORDER)
> > +		return ret;
> 
> Maybe also VM_BUG_ON_PAGE(page_order(page) > HPAGE_PMD_ORDER, page)?
> Just in case.

In the patch where I actually start creating THPs, I limit the order to
HPAGE_PMD_ORDER, so we're not going to see this today.  At some point
in the future, I can imagine that we allow THPs larger than PMD size,
and what we'd want alloc_set_pte() to look like is:

	if (pud_none(*vmf->pud) && PageTransCompound(page)) {
		ret = do_set_pud(vmf, page);
		if (ret != VM_FAULT_FALLBACK)
			return ret;
	}
	if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
		ret = do_set_pmd(vmf, page);
		if (ret != VM_FAULT_FALLBACK)
			return ret;
	}

Once we're in that situation, in do_set_pmd(), we'd want to figure out
which sub-page of the >PMD-sized page to insert.  But I don't want to
write code for that now.

So, what's the right approach if somebody does call alloc_set_pte()
with a >PMD sized page?  It's not exported, so the only two ways to get
it called with a >PMD sized page is to (1) persuade filemap_map_pages()
to call it, which means putting it in the page cache or (2) return it
from vm_ops->fault.  If someone actually does that (an interesting
device driver, perhaps), I don't think hitting it with a BUG is the
right response.  I think it should actually be to map the right PMD-sized
chunk of the page, but we don't even do that today -- we map the first
PMD-sized chunk of the page.

With this patch, we'll simply map the appropriate PAGE_SIZE chunk at the
requested address.  So this would be a bugfix for such a demented driver.
At some point, it'd be nice to handle this with a PMD, but I don't want
to write that code without a test-case.  We could probably simulate
it with the page cache THP code and be super-aggressive about creating
order-10 pages ... but this is feeling more and more out of scope for
this patch set, which today hit 99 patches.