Re: [PATCH v2 2/3] binder: do not initialize locals passed to copy_from_user()

[Jann Horn]

On Mon, Mar 2, 2020 at 7:17 PM Alexander Potapenko <glider@xxxxxxxxxx> wrote:
> On Mon, Mar 2, 2020 at 3:00 PM Joe Perches <joe@xxxxxxxxxxx> wrote:
> > On Mon, 2020-03-02 at 14:25 +0100, Alexander Potapenko wrote:
> > > On Mon, Mar 2, 2020 at 2:11 PM Joe Perches <joe@xxxxxxxxxxx> wrote:
> > > > On Mon, 2020-03-02 at 14:04 +0100, glider@xxxxxxxxxx wrote:
> > > > > Certain copy_from_user() invocations in binder.c are known to
> > > > > unconditionally initialize locals before their first use, like e.g. in
> > > > > the following case:
> > > > []
> > > > > diff --git a/drivers/android/binder.c b/drivers/android/binder.c
> > > > []
> > > > > @@ -3788,7 +3788,7 @@ static int binder_thread_write(struct binder_proc *proc,
> > > > >
> > > > >               case BC_TRANSACTION_SG:
> > > > >               case BC_REPLY_SG: {
> > > > > -                     struct binder_transaction_data_sg tr;
> > > > > +                     struct binder_transaction_data_sg tr __no_initialize;
> > > > >
> > > > >                       if (copy_from_user(&tr, ptr, sizeof(tr)))
> > > >
> > > > I fail to see any value in marking tr with __no_initialize
> > > > when it's immediately written to by copy_from_user.
> > >
> > > This is being done exactly because it's immediately written to by copy_to_user()
> > > Clang is currently unable to figure out that copy_to_user() initializes memory.
> > > So building the kernel with CONFIG_INIT_STACK_ALL=y basically leads to
> > > the following code:
> > >
> > >   struct binder_transaction_data_sg tr;
> > >   memset(&tr, 0xAA, sizeof(tr));
> > >   if (copy_from_user(&tr, ptr, sizeof(tr))) {...}
> > >
> > > This unnecessarily slows the code down, so we add __no_initialize to
> > > prevent the compiler from emitting the redundant initialization.
> >
> > So?  CONFIG_INIT_STACK_ALL by design slows down code.
> Correct.
>
> > This marking would likely need to be done for nearly all
> > 3000+ copy_from_user entries.
> Unfortunately, yes. I was just hoping to do so for a handful of hot
> cases that we encounter, but in the long-term a compiler solution must
> supersede them.
>
> > Why not try to get something done on the compiler side
> > to mark the function itself rather than the uses?
> This is being worked on in the meantime as well (see
> http://lists.llvm.org/pipermail/cfe-dev/2020-February/064633.html)
> Do you have any particular requisitions about how this should look on
> the source level?

Just thinking out loud: Should this be a function attribute, or should
it be a builtin - something like __builtin_assume_initialized(ptr,
len)? That would make it also work for macros, and it might simplify
the handling of inlining in the compiler. And you wouldn't need such a
complicated attribute that refers to function arguments by index and
such. The downside would be that it wouldn't work for non-inlined
functions without creating inline wrappers around them...