On Fri, Apr 28, 2023 at 07:02:22PM +0200, David Hildenbrand wrote:
> On 28.04.23 18:56, Kirill A . Shutemov wrote:
> > On Fri, Apr 28, 2023 at 06:51:46PM +0200, David Hildenbrand wrote:
> > > On 28.04.23 18:39, Peter Xu wrote:
> > > > On Fri, Apr 28, 2023 at 07:22:07PM +0300, Kirill A . Shutemov wrote:
> > > > > On Fri, Apr 28, 2023 at 06:13:03PM +0200, David Hildenbrand wrote:
> > > > > > On 28.04.23 18:09, Kirill A . Shutemov wrote:
> > > > > > > On Fri, Apr 28, 2023 at 05:43:52PM +0200, David Hildenbrand wrote:
> > > > > > > > On 28.04.23 17:34, David Hildenbrand wrote:
> > > > > > > > > On 28.04.23 17:33, Lorenzo Stoakes wrote:
> > > > > > > > > > On Fri, Apr 28, 2023 at 05:23:29PM +0200, David Hildenbrand wrote:
> > > > > > > > > > > > >
> > > > > > > > > > > > > Security is the primary case where we have historically closed uAPI
> > > > > > > > > > > > > items.
> > > > > > > > > > > >
> > > > > > > > > > > > As this patch
> > > > > > > > > > > >
> > > > > > > > > > > > 1) Does not tackle GUP-fast
> > > > > > > > > > > > 2) Does not take care of !FOLL_LONGTERM
> > > > > > > > > > > >
> > > > > > > > > > > > I am not convinced by the security argument in regard to this patch.
> > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > > If we want to sells this as a security thing, we have to block it
> > > > > > > > > > > > *completely* and then CC stable.
> > > > > > > > > > >
> > > > > > > > > > > Regarding GUP-fast, to fix the issue there as well, I guess we could do
> > > > > > > > > > > something similar as I did in gup_must_unshare():
> > > > > > > > > > >
> > > > > > > > > > > If we're in GUP-fast (no VMA), and want to pin a !anon page writable,
> > > > > > > > > > > fallback to ordinary GUP. IOW, if we don't know, better be safe.
> > > > > > > > > >
> > > > > > > > > > How do we determine it's non-anon in the first place? The check is on the
> > > > > > > > > > VMA. We could do it by following page tables down to folio and checking
> > > > > > > > > > folio->mapping for PAGE_MAPPING_ANON I suppose?
> > > > > > > > >
> > > > > > > > > PageAnon(page) can be called from GUP-fast after grabbing a reference.
> > > > > > > > > See gup_must_unshare().
> > > > > > > >
> > > > > > > > IIRC, PageHuge() can also be called from GUP-fast and could special-case
> > > > > > > > hugetlb eventually, as it's table while we hold a (temporary) reference.
> > > > > > > > Shmem might be not so easy ...
> > > > > > >
> > > > > > > page->mapping->a_ops should be enough to whitelist whatever fs you want.
> > > > > > >
> > > > > >
> > > > > > The issue is how to stabilize that from GUP-fast, such that we can safely
> > > > > > dereference the mapping. Any idea?
> > > > > >
> > > > > > At least for anon page I know that page->mapping only gets cleared when
> > > > > > freeing the page, and we don't dereference the mapping but only check a
> > > > > > single flag stored alongside the mapping. Therefore, PageAnon() is fine in
> > > > > > GUP-fast context.
> > > > >
> > > > > What codepath you are worry about that clears ->mapping on pages with
> > > > > non-zero refcount?
> > > > >
> > > > > I can only think of truncate (and punch hole). READ_ONCE(page->mapping)
> > > > > and fail GUP_fast if it is NULL should be fine, no?
> > > > >
> > > > > I guess we should consider if the inode can be freed from under us and the
> > > > > mapping pointer becomes dangling. But I think we should be fine here too:
> > > > > VMA pins inode and VMA cannot go away from under GUP.
> > > >
> > > > Can vma still go away if during a fast-gup?
> > > >
> > >
> > > So, after we grabbed the page and made sure the the PTE didn't change (IOW,
> > > the PTE was stable while we processed it), the page can get unmapped (but
> > > not freed, because we hold a reference) and the VMA can theoretically go
> > > away (and as far as I understand, nothing stops the file from getting
> > > deleted, truncated etc).
> > >
> > > So we might be looking at folio->mapping and the VMA is no longer there.
> > > Maybe even the file is no longer there.
> >
> > No. VMA cannot get away before PTEs are unmapped and TLB is flushed. And
> > TLB flushing is serialized against GUP_fast().
>
>
> The whole CONFIG_MMU_GATHER_RCU_TABLE_FREE handling makes the situation more
> complicated.
I think I found relevant snippet of code that solves similar issue.
get_futex_key() uses RCU to stabilize page->mapping after GUP_fast:
/*
* The associated futex object in this case is the inode and
* the page->mapping must be traversed. Ordinarily this should
* be stabilised under page lock but it's not strictly
* necessary in this case as we just want to pin the inode, not
* update the radix tree or anything like that.
*
* The RCU read lock is taken as the inode is finally freed
* under RCU. If the mapping still matches expectations then the
* mapping->host can be safely accessed as being a valid inode.
*/
rcu_read_lock();
if (READ_ONCE(page->mapping) != mapping) {
rcu_read_unlock();
put_page(page);
goto again;
}
inode = READ_ONCE(mapping->host);
if (!inode) {
rcu_read_unlock();
put_page(page);
goto again;
}
I think something similar can be used inside GUP_fast too.
--
Kiryl Shutsemau / Kirill A. Shutemov
