Re: [PATCH v3 09/10] rust: list: support heterogeneous lists

[Benno Lossin]

On 01.08.24 11:38, Alice Ryhl wrote:
> On Thu, Aug 1, 2024 at 11:24 AM Benno Lossin <benno.lossin@xxxxxxxxx> wrote:
>>
>> On 23.07.24 10:22, Alice Ryhl wrote:
>>> @@ -181,6 +185,47 @@ unsafe fn from_fields(me: *mut ListLinksFields) -> *mut Self {
>>>      }
>>>  }
>>>
>>> +/// Similar to [`ListLinks`], but also contains a pointer to the full value.
>>> +///
>>> +/// This type can be used instead of [`ListLinks`] to support lists with trait objects.
>>> +#[repr(C)]
>>> +pub struct ListLinksSelfPtr<T: ?Sized, const ID: u64 = 0> {
>>> +    /// The `ListLinks` field inside this value.
>>> +    ///
>>> +    /// This is public so that it can be used with `impl_has_list_links!`.
>>> +    pub inner: ListLinks<ID>,
>>> +    self_ptr: UnsafeCell<MaybeUninit<*const T>>,
>>
>> Why do you need `MaybeUninit`?
> 
> Right now the constructor initializes it to MaybeUninit::zeroed().
> What would you initialize it to without MaybeUninit? Remember that the
> vtable pointer in a fat pointer has strict validity requirements.

Oh... I forgot about that, can you add a comment about that? Also why
not use `Opaque` in that case then?

>>> +}
>>> +
>>> +// SAFETY: The fields of a ListLinksSelfPtr can be moved across thread boundaries.
>>> +unsafe impl<T: ?Sized + Send, const ID: u64> Send for ListLinksSelfPtr<T, ID> {}
>>> +// SAFETY: The type is opaque so immutable references to a ListLinksSelfPtr are useless. Therefore,
>>> +// it's okay to have immutable access to a ListLinks from several threads at once.
>>> +//
>>> +// Note that `inner` being a public field does not prevent this type from being opaque, since
>>> +// `inner` is a opaque type.
>>> +unsafe impl<T: ?Sized + Sync, const ID: u64> Sync for ListLinksSelfPtr<T, ID> {}
>>
>> [...]
>>
>>> @@ -135,5 +178,91 @@ unsafe fn post_remove(me: *mut $crate::list::ListLinks<$num>) -> *const Self {
>>>              }
>>>          }
>>>      };
>>> +
>>> +    (
>>> +        impl$({$($generics:tt)*})? ListItem<$num:tt> for $t:ty {
>>> +            using ListLinksSelfPtr;
>>> +        } $($rest:tt)*
>>> +    ) => {
>>> +        // SAFETY: See GUARANTEES comment on each method.
>>> +        unsafe impl$(<$($generics)*>)? $crate::list::ListItem<$num> for $t {
>>> +            // GUARANTEES:
>>> +            // This implementation of `ListItem` will not give out exclusive access to the same
>>> +            // `ListLinks` several times because calls to `prepare_to_insert` and `post_remove`
>>> +            // must alternate and exclusive access is given up when `post_remove` is called.
>>> +            //
>>> +            // Other invocations of `impl_list_item!` also cannot give out exclusive access to the
>>> +            // same `ListLinks` because you can only implement `ListItem` once for each value of
>>> +            // `ID`, and the `ListLinks` fields only work with the specified `ID`.
>>> +            unsafe fn prepare_to_insert(me: *const Self) -> *mut $crate::list::ListLinks<$num> {
>>> +                // SAFETY: The caller promises that `me` points at a valid value of type `Self`.
>>> +                let links_field = unsafe { <Self as $crate::list::ListItem<$num>>::view_links(me) };
>>> +
>>> +                let spoff = $crate::list::ListLinksSelfPtr::<Self, $num>::LIST_LINKS_SELF_PTR_OFFSET;
>>> +                // SAFETY: The constant is equal to `offset_of!(ListLinksSelfPtr, self_ptr)`, so
>>> +                // the pointer stays in bounds of the allocation.
>>> +                let self_ptr = unsafe { (links_field as *const u8).add(spoff) }
>>> +                    as *const ::core::cell::UnsafeCell<*const Self>;
>>
>> A bit confused why you need to do it this way, can't you just do this?:
>>
>>     let links_self_field = links_field.cast::<$crate::list::ListLinksSelfPtr>();
>>     // SAFETY: ...
>>     let self_ptr = unsafe { ::core::ptr::addr_of_mut!((*links_self_field).self_ptr) };
> 
> If nothing else, the field is not public. I can't remember if there
> was another reason.

Oh yeah that's true, then you have to go via the offset.

---
Cheers,
Benno