Re: [PATCH v11 2/3] rust: add dma coherent allocator abstraction.

[Alice Ryhl <aliceryhl@google.com>]

On Mon, Jan 27, 2025 at 5:59 PM Boqun Feng <boqun.feng@gmail.com> wrote:
>
> On Mon, Jan 27, 2025 at 11:43:39AM +0100, Alice Ryhl wrote:
> > On Mon, Jan 27, 2025 at 11:37 AM Danilo Krummrich <dakr@kernel.org> wrote:
> > >
> > > On Fri, Jan 24, 2025 at 08:27:36AM +0100, Alice Ryhl wrote:
> > > > On Thu, Jan 23, 2025 at 11:43 AM Abdiel Janulgue
> > > > <abdiel.janulgue@gmail.com> wrote:
> > > > > +    /// Reads data from the region starting from `offset` as a slice.
> > > > > +    /// `offset` and `count` are in units of `T`, not the number of bytes.
> > > > > +    ///
> > > > > +    /// Due to the safety requirements of slice, the data returned should be regarded by the
> > > > > +    /// caller as a snapshot of the region when this function is called, as the region could
> > > > > +    /// be modified by the device at anytime. For ringbuffer type of r/w access or use-cases
> > > > > +    /// where the pointer to the live data is needed, `start_ptr()` or `start_ptr_mut()`
> > > > > +    /// could be used instead.
> > > > > +    ///
> > > > > +    /// # Safety
> > > > > +    ///
> > > > > +    /// Callers must ensure that no hardware operations that involve the buffer are currently
> > > > > +    /// taking place while the returned slice is live.
> > > > > +    pub unsafe fn as_slice(&self, offset: usize, count: usize) -> Result<&[T]> {
> > > >
> > > > You were asked to rename this function because it returns a slice, but
> > > > I wonder if it's better to take an `&mut [T]` argument and to have
> > > > this function copy data into that argument. That way, we could make
> > > > the function itself safe. Perhaps the actual copy needs to be
> > > > volatile?
> > >
> > > Why do we consider the existing one unsafe?
> > >
> > > Surely, it's not desirable that the contents of the buffer are modified by the
> > > HW unexpectedly, but is this a concern in terms of Rust safety requirements?
> > >
> > > And if so, how does this go away with the proposed approach?
> >
> > In Rust, it is undefined behavior if the value behind an immutable
> > reference changes (unless the type uses UnsafeCell / Opaque or
> > similar). That is, any two consecutive reads of the same immutable
> > reference must return the same byte value no matter what happened in
> > between those reads.
> >
> > If we manually perform the read as a volatile read, then it is
> > arguably allowed for the value to be modified by the hardware while we
> > read the value.
> >
>
> Do you also assume that volatile read/write provide some sort of
> atomicity? Because otherwise even though the read/write may not be
> considered as UB, then results can be load/store teared.
>
> I asked because I think in case that we need atomicity, we should just
> use atomic APIs.

No, I'm not assuming that. I think it's like uaccess. Under normal
cases, it's not going to be concurrently modified, but it shouldn't
trigger UB if it is.

Alice