Re: [PATCH v3 04/10] rust: list: add struct with prev/next pointers

From: Benno Lossin
Date: Thu Aug 01 2024 - 06:47:13 EST


On 01.08.24 11:42, Alice Ryhl wrote:
> On Wed, Jul 31, 2024 at 8:41 PM Benno Lossin <benno.lossin@xxxxxxxxx> wrote:
>>
>> On 23.07.24 10:22, Alice Ryhl wrote:
>>> +/// The prev/next pointers for an item in a linked list.
>>> +///
>>> +/// # Invariants
>>> +///
>>> +/// The fields are null if and only if this item is not in a list.
>>> +#[repr(transparent)]
>>> +pub struct ListLinks<const ID: u64 = 0> {
>>> + #[allow(dead_code)]
>>> + inner: Opaque<ListLinksFields>,
>>
>> Do you really need `Opaque`? Or would `UnsafeCell` be enough? (If it is
>> enough and you change this, be aware that `Opaque` is `!Unpin`, so if
>> you intend for `ListLinks` to also be `!Unpin`, then you need a
>> `PhantomPinned`)
>
> I need the `!Unpin` part for aliasing.

Oh good point, do you mind adding a comment for that?

>>> +}
>>> +
>>> +// SAFETY: The next/prev fields of a ListLinks can be moved across thread boundaries.
>>
>> Why? This is not a justification.
>
> What would you say?

While trying to come up with a safety comment I thought about the
following: this impl does not depend on the type that is behind the
pointer (ie the type containing the `ListLinks`). Thus this `ListLinks`
will always implement `Send` even if the pointed-to value does not.
What we could do (and what definitely would be correct) is this:
`List` can only be used with `Send` types, then we could implement
`Send` for `ListLinks`. But I haven't actually come up with a problem,
so there might a more permissive solution.
Do you have a use-case where you need `!Send` types in a list?

Here is a part of my reasoning: If the pointed-to value is `!Send`, then
the `List` item type must also be `!Send`. Thus all list operations take
place on the same thread (since the `List` will be `!Send`). Therefore
nobody can access the `prev`/`next` pointers from another thread.

But this does not justify that `ListLinks` can be made `Send`. (although
there isn't actually a problem)

>>> +unsafe impl<const ID: u64> Send for ListLinks<ID> {}
>>> +// SAFETY: The type is opaque so immutable references to a ListLinks are useless. Therefore, it's
>>> +// okay to have immutable access to a ListLinks from several threads at once.
>>
>> You don't need to argue via `Opaque`, the type doesn't expose any
>> `&self` functions, so there are no functions to consider.
>
> I'm not arguing via the `Opaque` type. I'm just using "opaque" as a
> normal english word with its normal meaning.

Oh I see, then it's fine.

---
Cheers,
Benno