On Wed, Aug 07, 2024 at 11:59:47PM +0200, Alice Ryhl wrote:
> On 8/7/24 4:46 PM, Boqun Feng wrote:
> > On Wed, Aug 07, 2024 at 10:50:32AM +0200, Alice Ryhl wrote:
> > > On Tue, Aug 6, 2024 at 9:30 PM Boqun Feng <boqun.feng@xxxxxxxxx> wrote:
> > > >
> > > > On Tue, Aug 06, 2024 at 10:48:11AM +0200, Alice Ryhl wrote:
> > > > [...]
> > > > > > > + /// Returns the flags associated with the file.
> > > > > > > + ///
> > > > > > > + /// The flags are a combination of the constants in [`flags`].
> > > > > > > + #[inline]
> > > > > > > + pub fn flags(&self) -> u32 {
> > > > > > > + // This `read_volatile` is intended to correspond to a READ_ONCE call.
> > > > > > > + //
> > > > > > > + // SAFETY: The file is valid because the shared reference guarantees a nonzero refcount.
> > > > > > > + //
> > > > > > > + // FIXME(read_once): Replace with `read_once` when available on the Rust side.
> > > > > >
> > > > > > Do you know the status of this?
> > > > >
> > > > > It's still unavailable.
> > > > >
> > > >
> > > > I think with our own Atomic API, we can just use atomic_read() here:
> > > > yes, I know that to make this is not a UB, we need the C side to also do
> > > > atomic write on this `f_flags`, however, my reading of C code seems to
> > > > suggest that FS relies on writes to this field is atomic, therefore
> > > > unless someone is willing to convert all writes to `f_flags` in C into
> > > > a WRITE_ONCE(), nothing more we can do on Rust side. So using
> > > > atomic_read() is the correct thing to begin with.
> > >
> > > Huh? The C side uses atomic reads for this?
> > >
> >
> > Well, READ_ONCE(->f_flags) is atomic, so I thought you want to use
> > atomic here. However, after a quick look of `->f_flags` accesses, I find
> > out they should be protected by `->f_lock` (a few cases rely on
> > data race accesses, see p4_fd_open()), so I think what you should really
> > do here is the similar: make sure Rust code only accesses `->f_flags`
> > if `->f_lock` is held. Unless that's not the case for binder?
>
>
> Binder just has an `if (filp->f_flags & O_NONBLOCK)` block somewhere in the
> ioctl, where filp is the `struct file *` passed to the ioctl. Binder doesn't
> take the lock.
>
Yep, that's my point, I think binder C driver relies on the behaviors of
data race today (or probably all `->f_flags`s accessed by binder don't
have any concurrent write to them). Either way, what you do here is
better than C code if there was a data race. I was simply suggesting
instead of `read_once`, we could just do a `atomic_read` on `->f_flags`
once we support *unsafe` usage of doing atomic accesses on normal data
fields (of course, such a usage will be limited).
In other words, nothing needs to be changed here right now.
Regards,
Boqun
> Alice
