Re: [PATCH v6 4/4] rust: add abstraction for `struct page`

[Boqun Feng]

On Fri, Apr 19, 2024 at 07:24:31PM +0000, Benno Lossin wrote:
> On 19.04.24 19:23, Boqun Feng wrote:
> > On Fri, Apr 19, 2024 at 08:36:11AM +0000, Benno Lossin wrote:
> >> On 19.04.24 01:04, Boqun Feng wrote:
> >>> On Thu, Apr 18, 2024 at 03:56:11PM -0700, Boqun Feng wrote:
> >>>> On Thu, Apr 18, 2024 at 10:08:40PM +0000, Benno Lossin wrote:
> >>>>> On 18.04.24 20:52, Boqun Feng wrote:
> >>>>>> On Thu, Apr 18, 2024 at 08:59:20AM +0000, Alice Ryhl wrote:
> >>>>>>> +    /// Runs a piece of code with a raw pointer to a slice of this page, with bounds checking.
> >>>>>>> +    ///
> >>>>>>> +    /// If `f` is called, then it will be called with a pointer that points at `off` bytes into the
> >>>>>>> +    /// page, and the pointer will be valid for at least `len` bytes. The pointer is only valid on
> >>>>>>> +    /// this task, as this method uses a local mapping.
> >>>>>>> +    ///
> >>>>>>> +    /// If `off` and `len` refers to a region outside of this page, then this method returns
> >>>>>>> +    /// `EINVAL` and does not call `f`.
> >>>>>>> +    ///
> >>>>>>> +    /// # Using the raw pointer
> >>>>>>> +    ///
> >>>>>>> +    /// It is up to the caller to use the provided raw pointer correctly. The pointer is valid for
> >>>>>>> +    /// `len` bytes and for the duration in which the closure is called. The pointer might only be
> >>>>>>> +    /// mapped on the current thread, and when that is the case, dereferencing it on other threads
> >>>>>>> +    /// is UB. Other than that, the usual rules for dereferencing a raw pointer apply: don't cause
> >>>>>>> +    /// data races, the memory may be uninitialized, and so on.
> >>>>>>> +    ///
> >>>>>>> +    /// If multiple threads map the same page at the same time, then they may reference with
> >>>>>>> +    /// different addresses. However, even if the addresses are different, the underlying memory is
> >>>>>>> +    /// still the same for these purposes (e.g., it's still a data race if they both write to the
> >>>>>>> +    /// same underlying byte at the same time).
> >>>>>>> +    fn with_pointer_into_page<T>(
> >>>>>>> +        &self,
> >>>>>>> +        off: usize,
> >>>>>>> +        len: usize,
> >>>>>>> +        f: impl FnOnce(*mut u8) -> Result<T>,
> >>>>>>
> >>>>>> I wonder whether the way to go here is making this function signature:
> >>>>>>
> >>>>>>        fn with_slice_in_page<T> (
> >>>>>>            &self,
> >>>>>> 	       off: usize,
> >>>>>> 	       len: usize,
> >>>>>> 	       f: iml FnOnce(&UnsafeCell<[u8]>) -> Result<T>
> >>>>>>        ) -> Result<T>
> >>>>>>
> >>>>>> , because in this way, it makes a bit more clear that what memory that
> >>>>>> `f` can access, in other words, the users are less likely to use the
> >>>>>> pointer in a wrong way.
> >>>>>>
> >>>>>> But that depends on whether `&UnsafeCell<[u8]>` is the correct
> >>>>>> abstraction and the ecosystem around it: for example, I feel like these
> >>>>>> two functions:
> >>>>>>
> >>>>>> 	    fn len(slice: &UnsafeCell<[u8]>) -> usize
> >>>>>> 	    fn as_ptr(slice: &UnsafeCell<[u8]>) -> *mut u8
> >>>>>>
> >>>>>> should be trivially safe, but I might be wrong. Again this is just for
> >>>>>> future discussion.
> >>>>>
> >>>>> I think the "better" type would be `&[UnsafeCell<u8>]`. Since there you
> >>>>> can always access the length.
> >>>>>
> >>>>
> >>>> Hmm.. here is the thing, having `&UnsafeCell<[u8]>` means having a `*mut
> >>>> [u8]>`, and it should always be safe to get a "length" of `*mut [u8]`,
> >>>> right? I haven't found any method doing that, but the length should be
> >>>> just a part of fat pointer, so I think getting that is a defined
> >>>> behavior. But maybe I'm missing something.
> >>
> >> There is `to_raw_parts` [1], but that is unstable. (Note that
> >> `<[T] as Pointee>::Metadata = usize`, see [2])
> >>
> > 
> > Oh, that's good to know, thank you! ;-)
> >
> >>> Hmm... but I guess one of the problems of this approach, is how to
> >>> construct a `&UnsafeCell<[u8]>` from a pointer and length...
> >>
> >> We could use `from_raw_parts` [3]. But when making the slice the outer
> >> type, we can use a stable function to convert a pointer and a length to
> >> a slice [4].
> >>
> > 
> > Yes, but there appears no way to get a pointer with larger provenance
> > from a `&[UnsafeCell<u8>]`, right?
> 
> What do you mean by "larger provenance"?
> 

Say you have a `&[UnsafeCell<u8>]` whose length is 64, what's the proper
way to get a `*mut u8` (or any other pointer) has the provenance for the
whole 64 bytes, so that you can pass it to a memcpy like function?
"larger" means the size of the provenance is larger than u8.

> >>>>> Another question would be if page allows for uninitialized bits, in that
> >>>>> case, we would need `&[Opaque<u8>]`.
> >>>>>
> >>>>
> >>>> Yes, or `&Opaque<[u8>]`.
> >>
> >> I don't think that putting the slice on the inside is what we want. Also
> > 
> > Hmm.. why? So in `&UnsafeCell<[u8]>` vs `&[UnsafeCell<u8>]` case, I
> > think the former represent "a slice of u8 that can be modified in the
> > same time" very well, and this is what a pointer-and-length pair usually
> > represents in kernel, I think. But yes, the latter is OK to me as well,
> > just hard to play the provenance game I guess?
> 
> Ultimately it again comes down to missing field projections :)
> 
> The type `&UnsafeCell<[u8]>` is less *useful*, since you cannot even get
> the length of the slice. Also indexing into this type is not easily
> possible. This is because the only way to get/change the inner value of
> an `UnsafeCell` is via `get`.
> 
> Compare this with the slice type. It allows getting the length, indexing
> into it (ie a form of field projections, if we consider slices as having
> a variable amount of fields).
> 
> All those issues would be solved by (good) field projections.
> 
> 
> Field projections also give a reason for why using `&[UnsafeCell<u8>]`
> is not really different from `&UnsafeCell<[u8]>`: At any point in time
> we ought to be able to project `&UnsafeCell<[u8]> -> &[UnsafeCell<u8>]`.
> 

Right, to me there is no significant difference between these two. Maybe
because I'm full field projected minded ;-)

> So it's fine to just use that from the get-go.
> 
> >> note that `Opaque<T>` requires that `T: Sized` and that is not the case
> >> for `[u8]`.
> > 
> > Oh, you're right. In case of MaybeUninit, it requires `T: Sized`, so
> > `Opaque<[u8]>` doesn't quite work.
> > 
> > Moving forward, maybe the first step is to see whether `&[Opaque<u8>]`
> > and `&[UnsafeCell<u8>]` can have a good way to generate a pointer with
> > proper provenance? Time to ping t-opsem maybe?
> 
> Good idea, do you want to do that, or should I do it?
> 

A way to get a larger provenance (explained above) is currently my only
question, so if you think that's something reasonable to ask, i.e.
nothing you know of can help. I will post a message there.

Regards,
Boqun

> -- 
> Cheers,
> Benno
>