Re: [PATCH v5 05/26] mm/swap: Introduce the idea of special swap ptes

[Peter Xu <peterx@redhat.com>]

On Thu, Jul 22, 2021 at 11:08:53AM +1000, Alistair Popple wrote:
> On Thursday, 22 July 2021 7:35:32 AM AEST Peter Xu wrote:
> > On Wed, Jul 21, 2021 at 09:28:49PM +1000, Alistair Popple wrote:
> > > On Saturday, 17 July 2021 5:11:33 AM AEST Peter Xu wrote:
> > > > On Fri, Jul 16, 2021 at 03:50:52PM +1000, Alistair Popple wrote:
> > > > > Hi Peter,
> > > > > 
> > > > > [...]
> > > > > 
> > > > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > > > > index ae1f5d0cb581..4b46c099ad94 100644
> > > > > > --- a/mm/memcontrol.c
> > > > > > +++ b/mm/memcontrol.c
> > > > > > @@ -5738,7 +5738,7 @@ static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
> > > > > >  
> > > > > >  	if (pte_present(ptent))
> > > > > >  		page = mc_handle_present_pte(vma, addr, ptent);
> > > > > > -	else if (is_swap_pte(ptent))
> > > > > > +	else if (pte_has_swap_entry(ptent))
> > > > > >  		page = mc_handle_swap_pte(vma, ptent, &ent);
> > > > > >  	else if (pte_none(ptent))
> > > > > >  		page = mc_handle_file_pte(vma, addr, ptent, &ent);
> > > > > 
> > > > > As I understand things pte_none() == False for a special swap pte, but
> > > > > shouldn't this be treated as pte_none() here? Ie. does this need to be
> > > > > pte_none(ptent) || is_swap_special_pte() here?
> > > > 
> > > > Looks correct; here the page/swap cache could hide behind the special pte just
> > > > like a none pte.  Will fix it.  Thanks!
> > > > 
> > > > > 
> > > > > > diff --git a/mm/memory.c b/mm/memory.c
> > > > > > index 0e0de08a2cd5..998a4f9a3744 100644
> > > > > > --- a/mm/memory.c
> > > > > > +++ b/mm/memory.c
> > > > > > @@ -3491,6 +3491,13 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
> > > > > >  	if (!pte_unmap_same(vmf))
> > > > > >  		goto out;
> > > > > >  
> > > > > > +	/*
> > > > > > +	 * We should never call do_swap_page upon a swap special pte; just be
> > > > > > +	 * safe to bail out if it happens.
> > > > > > +	 */
> > > > > > +	if (WARN_ON_ONCE(is_swap_special_pte(vmf->orig_pte)))
> > > > > > +		goto out;
> > > > > > +
> > > > > >  	entry = pte_to_swp_entry(vmf->orig_pte);
> > > > > >  	if (unlikely(non_swap_entry(entry))) {
> > > > > >  		if (is_migration_entry(entry)) {
> > > > > 
> > > > > Are there other changes required here? Because we can end up with stale special
> > > > > pte's and a special pte is !pte_none don't we need to fix some of the !pte_none
> > > > > checks in these functions:
> > > > > 
> > > > > insert_pfn() -> checks for !pte_none
> > > > > remap_pte_range() -> BUG_ON(!pte_none)
> > > > > apply_to_pte_range() -> didn't check further but it tests for !pte_none
> > > > > 
> > > > > In general it feels like I might be missing something here though. There are
> > > > > plenty of checks in the kernel for pte_none() which haven't been updated. Is
> > > > > there some rule that says none of those paths can see a special pte?
> > > > 
> > > > My rule on doing this was to only care about vma that can be backed by RAM,
> > > > majorly shmem/hugetlb, so the special pte can only exist there within those
> > > > vmas.  I believe in most pte_none() users this special pte won't exist.
> > > > 
> > > > So if it's not related to RAM backed memory at all, maybe it's fine to keep the
> > > > pte_none() usage like before.
> > > > 
> > > > Take the example of insert_pfn() referenced first - I think it can be used to
> > > > map some MMIO regions, but I don't think we'll call that upon a RAM region
> > > > (either shmem or hugetlb), nor can it be uffd wr-protected.  So I'm not sure
> > > > adding special pte check there would be helpful.
> > > > 
> > > > apply_to_pte_range() seems to be a bit special - I think the pte_fn_t matters
> > > > more on whether the special pte will matter.  I had a quick look, it seems
> > > > still be used mostly by all kinds of driver code not mm core.  It's used in two
> > > > forms:
> > > > 
> > > >         apply_to_page_range
> > > >         apply_to_existing_page_range
> > > > 
> > > > The first one creates ptes only, so it ignores the pte_none() check so I skipped.
> > > > 
> > > > The second one has two call sites:
> > > > 
> > > > *** arch/powerpc/mm/pageattr.c:
> > > > change_memory_attr[99]         return apply_to_existing_page_range(&init_mm, start, size,
> > > > set_memory_attr[132]           return apply_to_existing_page_range(&init_mm, start, sz, set_page_attr,
> > > > 
> > > > *** mm/kasan/shadow.c:
> > > > kasan_release_vmalloc[485]     apply_to_existing_page_range(&init_mm,
> > > > 
> > > > I'll leave the ppc callers for now as uffd-wp is not even supported there.  The
> > > > kasan_release_vmalloc() should be for kernel allocated memories only, so should
> > > > not be a target for special pte either.
> > > > 
> > > > So indeed it's hard to 100% cover all pte_none() users to make sure things are
> > > > used right.  As stated above I still believe most callers don't need that, but
> > > > the worst case is if someone triggered uffd-wp issues with a specific feature,
> > > > we can look into it.  I am not sure whether it's good we add this for all the
> > > > pte_none() users, because mostly they'll be useless checks, imho.
> > > 
> > > I wonder then - should we make pte_none() return true for these special pte's
> > > as well? It seems if we do miss any callers it could result in some fairly hard
> > > to find bugs if the code follows a different path due to the presence of an
> > > unexpected special pte changing the result of pte_none().
> > 
> > I thought about something similar before, but I didn't dare to change
> > pte_none() as it's been there for ages and I'm afraid people will get confused
> > when it's meaning changed.  So even if we want to have some helper identifying
> > "either none pte or the swap special pte" it should use a different name.
> > 
> > Modifying the meaning of pte_none() could also have other risks that when we
> > really want an empty pte to be doing something else now.  It turns out there's
> > no easy way to not identify the case one by one, at least to me.  I'm always
> > open to good suggestions.
> 
> I'm not convinced it's changing the behaviour of pte_none() though and my
> concern is that introducing special swap ptes does change it. Prior to this
> clearing a pte would result in pte_none()==True. After this series clearing a
> pte can some sometimes result in pte_none()==False because it doesn't really
> get cleared.

The thing is the uffd special pte is not "none" literally; there's something
inside.  That's what makes it feel not right to me.  I'm not against trapping
all of pte_none(), but as I mentioned I think at least it needs to be renamed
to something else (maybe pte_none_mostly(), but I don't know..).

> 
> Now as you say it's hard to cover 100% of pte_none() uses, so it's possible we
> have missed cases that may now encounter a special pte and take a different
> path (get_mctgt_type() is one example, I stopped looking for other possible
> ones after mm/memory.c).
> 
> So perhaps if we want to keep pte_none() to check for really clear pte's then
> what is required is converting all callers to a new helper
> (pte_none_not_special()?) that treats special swap ptes as pte_none() and warns
> if a special pte is encountered?

By double check all core memory calls to pte_none()?

The special swap pte shouldn't exist for most cases but only for shmem and
hugetlbfs so far.  So we can sensibly drop a lot of pte_none() users IMHO
depending on the type of memory.

> 
> > Btw, as you mentioned before, we can use a new number out of MAX_SWAPFILES,
> > that'll make all these easier a bit here, then we don't need to worry on
> > pte_none() issues too.  Two days ago Hugh has raised some similar concern on
> > whether it's good to implement this uffd-wp special pte like this.  I think we
> > can discuss this separately.
> 
> Yes, I saw that and personally I still prefer that approach.

Yes I see your preference.  Let's hold off a bit on the pte_none() discussions;
I'll re-raise this in the cover letter soon.  If everyone is okay that we use
yet another MAX_SWAPFILES and that's preferred, then I can switch the design.
Then I think I can also avoid touching the pte_none() bits at all, which seems
to be controversial here.

But still, I am also not convinced that we can blindly replace pte_none() into
"either none pte or some special pte", either in this series or (if this series
will switch to swp_entry) in the future when we want to use !pte_present and
!swp_entry ptes.  If we want to replace that, we may still want to check over
all the users of pte_none then it's the same as what we should do now, and do a
proper rename of it.

Thanks,

-- 
Peter Xu