Re: [PATCH] Add mmap_assert_locked() annotations to find_vma*()

[Jann Horn <jannh@google.com>]

On Wed, Aug 4, 2021 at 1:07 AM Liam Howlett <liam.howlett@oracle.com> wrote:
> * Luigi Rizzo <lrizzo@google.com> [210803 17:49]:
> > On Tue, Aug 3, 2021 at 6:08 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
> > >
> > > On Sat, Jul 31, 2021 at 10:53:41AM -0700, Luigi Rizzo wrote:
> > > > find_vma() and variants need protection when used.
> > > > This patch adds mmap_assert_lock() calls in the functions.
> > > >
> > > > To make sure the invariant is satisfied, we also need to add a
> > > > mmap_read_loc() around the get_user_pages_remote() call in
> > > > get_arg_page(). The lock is not strictly necessary because the mm
> > > > has been newly created, but the extra cost is limited because
> > > > the same mutex was also acquired shortly before in __bprm_mm_init(),
> > > > so it is hot and uncontended.
> > > >
> > > > Signed-off-by: Luigi Rizzo <lrizzo@google.com>
> > > >  fs/exec.c | 2 ++
> > > >  mm/mmap.c | 2 ++
> > > >  2 files changed, 4 insertions(+)
> > > >
> > > > diff --git a/fs/exec.c b/fs/exec.c
> > > > index 38f63451b928..ac7603e985b4 100644
> > > > +++ b/fs/exec.c
> > > > @@ -217,8 +217,10 @@ static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
> > > >        * We are doing an exec().  'current' is the process
> > > >        * doing the exec and bprm->mm is the new process's mm.
> > > >        */
> > > > +     mmap_read_lock(bprm->mm);
> > > >       ret = get_user_pages_remote(bprm->mm, pos, 1, gup_flags,
> > > >                       &page, NULL, NULL);
> > > > +     mmap_read_unlock(bprm->mm);
> > > >       if (ret <= 0)
> > > >               return NULL;
> > >
> > > Wasn't Jann Horn working on something like this too?
> > >
> > > https://lore.kernel.org/linux-mm/20201016225713.1971256-1-jannh@google.com/
> > >
> > > IIRC it was very tricky here, are you sure it is OK to obtain this lock
> > > here?
> >
> > I cannot comment on Jann's patch series but no other thread knows
> > about this mm at this point in the code so the lock is definitely
> > safe to acquire (shortly before there was also a write lock acquired
> > on the same mm, in the same conditions).
>
> If there is no other code that knows about this mm, then does one need
> the lock at all?  Is this just to satisfy the new check you added?
>
> If you want to make this change, I would suggest writing it in a way to
> ensure the call to expand_downwards() in the same function also holds
> the lock.  I believe this is technically required as well?  What do you
> think?

The call to expand_downwards() takes a VMA pointer as argument, and
the mmap lock is the only thing that normally prevents concurrent
freeing of VMA structs. Taking a lock there would be of limited utility - either
the lock is not necessary because nobody else can access the MM, or
the lock is insufficient because someone could have freed the VMA
pointer before the lock was taken. So I think that taking a lock
around the expand_downwards() call would just be obfuscating things,
unless you specifically want to prevent concurrent *reads* while
concurrent *writes* are impossible.

Since I haven't sent a new version of my old series for almost a year,
I think it'd be fine to take Luigi's patch for now, and undo it at a
later point when/if we want to actually use proper locking here
because we're worried about concurrent access to the MM.