Message-ID: <Z5fc0EFCUdXlkpWB@tardis.local>
Date: Mon, 27 Jan 2025 11:21:52 -0800
From: Boqun Feng <boqun.feng@...il.com>
To: Alice Ryhl <aliceryhl@...gle.com>
Cc: Danilo Krummrich <dakr@...nel.org>,
	Abdiel Janulgue <abdiel.janulgue@...il.com>,
	rust-for-linux@...r.kernel.org, daniel.almeida@...labora.com,
	robin.murphy@....com, Miguel Ojeda <ojeda@...nel.org>,
	Alex Gaynor <alex.gaynor@...il.com>, Gary Guo <gary@...yguo.net>,
	Björn Roy Baron <bjorn3_gh@...tonmail.com>,
	Benno Lossin <benno.lossin@...ton.me>,
	Andreas Hindborg <a.hindborg@...nel.org>,
	Trevor Gross <tmgross@...ch.edu>,
	Valentin Obst <kernel@...entinobst.de>,
	open list <linux-kernel@...r.kernel.org>,
	Christoph Hellwig <hch@....de>,
	Marek Szyprowski <m.szyprowski@...sung.com>, airlied@...hat.com,
	"open list:DMA MAPPING HELPERS" <iommu@...ts.linux.dev>
Subject: Re: [PATCH v11 2/3] rust: add dma coherent allocator abstraction.

On Mon, Jan 27, 2025 at 08:05:46PM +0100, Alice Ryhl wrote:
> On Mon, Jan 27, 2025 at 8:01 PM Boqun Feng <boqun.feng@...il.com> wrote:
> >
> > On Mon, Jan 27, 2025 at 07:46:07PM +0100, Alice Ryhl wrote:
> > > On Mon, Jan 27, 2025 at 7:38 PM Boqun Feng <boqun.feng@...il.com> wrote:
> > > >
> > > > On Mon, Jan 27, 2025 at 07:32:29PM +0100, Alice Ryhl wrote:
> > > > > On Mon, Jan 27, 2025 at 5:59 PM Boqun Feng <boqun.feng@...il.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@...nel.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@...il.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.
> > > > >
> > > >
> > > > Let's say my_alloc[7].foo is a (u64, u64), would
> > > >
> > > >         dma_read!(my_alloc[7].foo)
> > > >
> > > > and
> > > >
> > > >         dma_write!(my_alloc[7].foo, (1u64, 2u64))
> > > >
> > > > trigger any UB when they are concurrent? (Of course, the example here is
> > > > a bit inpropriate because it's DMA buff, but still the question is more
> > > > on whatever atomic expectation we want from read_volatile() and
> > > > write_volatile()).
> > >
> > > I imagine that it would be most convenient for it to not be UB, but I
> > > also don't think people would have an expectation for that to not
> > > involve tearing.
> > >
> >
> > Depending on the granularity that tearing can happen, if .foo is an enum
> > (or any other type that not all bit combinations are valid) and tearing
> > can happen at byte levels, then a racing dma_read() may read invalid
> > data.
> 
> T: FromBytes + ToBytes is already required for these types. You can't
> use these operations with such an enum.
> 

I was talking about a wider problem, but fine ;-) So the assumption is
the read_volatile() or copy_nonoverlapping() provide byte-level
atomicity? Although unlikely, but if tearing happens at sub-byte level,
then even if `T: FromBytes + ToBytes` you can still get invalid data.

> > I think it's fine to expect read_volatile() and write_volatile()
> > themselves don't trigger UB, but we will need to be careful about the
> > atomic granularity that we can expect on them. It would be more clear if
> > we use the atomic API here (and implementation can be read_volatile()
> > and write_volatile()), and it can avoid coding based on tribal knowledge
> > such as "in kernel, read_volatile() and write_volatile() imply atomic".
> 
> Why should we use atomics for operations that don't need to be atomic?
> Most of the time, dma memory is not *actually* changed while you read
> it.
> 

Because the requirement here actually needs atomic at byte level, it's
similar to:

	https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p1478r8.html

also, notice that for byte level atomic, it's actually free on most of
the architectures (i.e. no extra cost). Again, it's more of a "how do we
express our assumption" question. If indeed we expect byte-level
atomicity, then I see no harm to use atomic API here.

Regards,
Boqun

> Alice

Powered by blists - more mailing lists