Re: [PATCH] rust_binder: disallow binder fd transfers
From: Carlos Llamas
Date: Sat Nov 21 2026 - 13:22:00 EST
On Wed, Nov 18, 2026 at 11:29:32AM +0000, Alice Ryhl wrote:
> On Tue, Nov 17, 2026 at 11:19:26PM +0000, Carlos Llamas wrote:
> > @@ -182,6 +182,13 @@ pub(crate) fn info_add_fd(
> > buffer_offset: usize,
> > close_on_free: bool,
> > ) -> Result {
> > + // SAFETY: `file` is valid; read `f_op` with `READ_ONCE` semantics.
> > + let f_op = unsafe { (&raw const (*file.as_ptr()).f_op).read_volatile() };
> > + if core::ptr::eq(f_op, &crate::rust_binder_fops.0) {
> > + // Binder FD transfers are not allowed.
> > + return Err(EINVAL);
> > + }
>
> We have moved to using atomic_load instead of a volatile read when the C side
> uses READ_ONCE. For example:
>
> // SAFETY: `file` is valid, so `f_op` is valid and aligned, and any
> // concurrent writes are atomic under LKMM.
> let f_op = unsafe { atomic_load(&raw mut (*file.as_ptr()).f_op, Relaxed) };
I see, ok I'll switch to this for v2.
>
> however the f_op field isn't read/modified with READ/WRITE_ONCE on the C
> side, so I think we can use a bare access.
>
> let f_op = unsafe { (*file.as_ptr()).f_op };
We don't but there are other fd's in other drivers that do. This was
flagged by my local LLM review. It was in some tty driver and somewhere
else that override the f_ops. It seems pretty far-fetched that a false
rust_binder_fops would trigger from this, but at least it would protect
against KCSAN reports and similar. So probably best to keep it.
--
Carlos Llamas