Re: [PATCH v5 0/2] rust: revocable: documentation and refactorings

["Benno Lossin" <lossin@kernel.org>]

On Sat Jul 5, 2025 at 7:09 AM CEST, Marcelo Moreira wrote:
> Em qui., 3 de jul. de 2025 às 05:24, Benno Lossin <lossin@kernel.org> escreveu:
>>
>> On Thu Jun 26, 2025 at 6:59 PM CEST, Marcelo Moreira wrote:
>> > This patch series brings documentation and refactorings to the `Revocable` type.
>> >
>> > Changes include:
>> > - Clarifying the write invariant and updating associated safety comments for `Revocable<T>`.
>> > - Splitting the internal `revoke_internal` function into two distinct, explicit functions: `revoke()` (safe, synchronizing with RCU) and `revoke_nosync()` (unsafe, without RCU synchronization), now returning `bool` to indicate revocation status.
>>
>> Could you wrap your text to a more readable column? Thanks!
>
> Sure! Thanks!
>
>>
>> >
>> > Marcelo Moreira (2):
>> >   rust: revocable: Refactor revocation mechanism to remove generic
>> >     revoke_internal
>> >   rust: revocable: Clarify write invariant and update safety comments
>> >
>> > Changelog
>> > ---------
>> >
>> > Changes since v4:
>> > - Rebased the series onto the latest `rfl/rust-next` to integrate recent changes, specifically the `bool` return for `revoke()` and `revoke_nosync()`.
>> > - Dropped the "rust: revocable: simplify RevocableGuard for internal safety" patch, as the approach of using a direct reference (`&'a T`) for `RevocableGuard` was found to be unsound due to Rust's aliasing rules and LLVM's `dereferencable` attribute guarantees, which require references to remain valid for the entire function call duration, even if the internal RCU guard is dropped earlier.
>> > - Refined the `PinnedDrop::drop` `SAFETY` comment based on Benno Lossin's and Miguel Ojeda's feedback, adopting a more concise and standard Kernel-style bullet point format.
>> > - Corrected a duplicated line in the commit message of the second patch.
>>
>> Now since we had to drop the `RevocableGuard` change, its safety
>> invariant & comment in `deref` is insufficient. It shouldn't have the
>> invariant that the rcu lock is held (since it owns an `rcu::Guard`, that
>> already is guaranteed), but instead it should require that the
>> `data_ref` pointer is valid. That invariant is then used by the safety
>> comment in `deref` to justify dereferencing the pointer.
>>
>> Also, I think it's better to reorder the patches again (since the
>> current first one relies on changes from the second one), the first one
>> should be the change to the invariants section of `Revocable` (so
>> currently the second patch). Then the second and third patches can be
>> the removal of `revoke_internal` and the `RevocableGuard` safety
>> documentation fix.
>
> All right Benno, I'll prepare the comment for `RevocableGuard` and send v6.
>
> The order now is:
> 1- Documentation for invariant and updates associated `SAFETY` comments
> 2- Remove `revoke_internal` (Refactoring)
> 3- `RevocableGuard` safety documentation fix.

Sounds good!

---
Cheers,
Benno