Re: [PATCH v8 4/6] rust: rbtree: add mutable iterator

[Benno Lossin <benno.lossin@proton.me>]

On 06.08.24 10:30, Alice Ryhl wrote:
> On Mon, Aug 5, 2024 at 9:22 PM Benno Lossin <benno.lossin@proton.me> wrote:
>>
>> On 27.07.24 22:30, Matt Gilbride wrote:
>>> From: Wedson Almeida Filho <wedsonaf@gmail.com>
>>>
>>> Add mutable Iterator implementation for `RBTree`,
>>> allowing iteration over (key, value) pairs in key order. Only values are
>>> mutable, as mutating keys implies modifying a node's position in the tree.
>>>
>>> Mutable iteration is used by the binder driver during shutdown to
>>> clean up the tree maintained by the "range allocator" [1].
>>>
>>> Link: https://lore.kernel.org/rust-for-linux/20231101-rust-binder-v1-6-08ba9197f637@google.com/ [1]
>>> Signed-off-by: Wedson Almeida Filho <wedsonaf@gmail.com>
>>> Signed-off-by: Matt Gilbride <mattgilbride@google.com>
>>> Reviewed-by: Alice Ryhl <aliceryhl@google.com>
>>> Tested-by: Alice Ryhl <aliceryhl@google.com>
>>> ---
>>>  rust/kernel/rbtree.rs | 98 ++++++++++++++++++++++++++++++++++++++++++++-------
>>>  1 file changed, 86 insertions(+), 12 deletions(-)
>>>
>>> diff --git a/rust/kernel/rbtree.rs b/rust/kernel/rbtree.rs
>>> index d10074e4ac58..d7514ebadfa8 100644
>>> --- a/rust/kernel/rbtree.rs
>>> +++ b/rust/kernel/rbtree.rs
>>> @@ -197,8 +197,26 @@ pub fn iter(&self) -> Iter<'_, K, V> {
>>>          // INVARIANT: `bindings::rb_first` returns a valid pointer to a tree node given a valid pointer to a tree root.
>>
>> This INVARIANT is out of place, `Iter` doesn't have any INVARIANT any
>> more.
> 
> We can delete it.
> 
>>>          Iter {
>>>              _tree: PhantomData,
>>> -            // SAFETY: `self.root` is a valid pointer to the tree root.
>>> -            next: unsafe { bindings::rb_first(&self.root) },
>>> +            iter_raw: IterRaw {
>>
>> This `IterRaw` construction is missing an INVARIANT comment. I think you
>> can copy paste from below.
> 
> We can copy from below.
> 
>>> +                // SAFETY: by the invariants, all pointers are valid.
>>> +                next: unsafe { bindings::rb_first(&self.root) },
>>> +                _phantom: PhantomData,
>>> +            },
>>> +        }
>>> +    }
>>> +
>>> +    /// Returns a mutable iterator over the tree nodes, sorted by key.
>>> +    pub fn iter_mut(&mut self) -> IterMut<'_, K, V> {
>>> +        IterMut {
>>> +            _tree: PhantomData,
>>> +            // INVARIANT:
>>> +            //   - `self.root` is a valid pointer to a tree root.
>>> +            //   - `bindings::rb_first` produces a valid pointer to a node given `root` is valid.
>>> +            iter_raw: IterRaw {
>>> +                // SAFETY: by the invariants, all pointers are valid.
>>> +                next: unsafe { bindings::rb_first(&self.root) },
>>
>> Does this really derive a mutable reference? Ie shouldn't this be:?
>>
>>     next: unsafe { bindings::rb_first(&mut self.root) },
> 
> Let's change this to:
> 
> next: unsafe { bindings::rb_first(ptr::from_mut(&mut self.root)) }
> 
> This way, the pointer will be derived from a mutable reference even if
> it becomes a `*const` through intermediate operations.

SGTM

> 
> 
>>> +                _phantom: PhantomData,
>>> +            },
>>>          }
>>>      }
>>>
>>> @@ -211,6 +229,11 @@ pub fn keys(&self) -> impl Iterator<Item = &'_ K> {
>>>      pub fn values(&self) -> impl Iterator<Item = &'_ V> {
>>>          self.iter().map(|(_, v)| v)
>>>      }
>>> +
>>> +    /// Returns a mutable iterator over the values of the nodes in the tree, sorted by key.
>>> +    pub fn values_mut(&mut self) -> impl Iterator<Item = &'_ mut V> {
>>> +        self.iter_mut().map(|(_, v)| v)
>>> +    }
>>>  }
>>>
>>>  impl<K, V> RBTree<K, V>
>>> @@ -414,13 +437,9 @@ fn into_iter(self) -> Self::IntoIter {
>>>  /// An iterator over the nodes of a [`RBTree`].
>>>  ///
>>>  /// Instances are created by calling [`RBTree::iter`].
>>> -///
>>> -/// # Invariants
>>> -/// - `self.next` is a valid pointer.
>>> -/// - `self.next` points to a node stored inside of a valid `RBTree`.
>>>  pub struct Iter<'a, K, V> {
>>>      _tree: PhantomData<&'a RBTree<K, V>>,
>>> -    next: *mut bindings::rb_node,
>>> +    iter_raw: IterRaw<K, V>,
>>>  }
>>>
>>>  // SAFETY: The [`Iter`] gives out immutable references to K and V, so it has the same
>>> @@ -434,21 +453,76 @@ unsafe impl<'a, K: Sync, V: Sync> Sync for Iter<'a, K, V> {}
>>>  impl<'a, K, V> Iterator for Iter<'a, K, V> {
>>>      type Item = (&'a K, &'a V);
>>>
>>> +    fn next(&mut self) -> Option<Self::Item> {
>>> +        self.iter_raw.next().map(|(k, v)|
>>> +            // SAFETY: Due to `self._tree`, `k` and `v` are valid for the lifetime of `'a`.
>>> +            unsafe { (&*k, &*v) })
>>
>> I don't really like the formatting here, can you move the SAFETY one
>> line upwards? It should format nicely then.
> 
> You suggested exactly the reverse formatting change on RBTreeCursor?

Do you mean on this version or in a previous one? If you mean in this
one, then I would argue that they are not "reverses" of each other. For
this instance I would prefer

        // SAFETY: Due to `self._tree`, `k` and `v` are valid for the lifetime of `'a`.
        self.iter_raw.next().map(|(k, v)| unsafe { (&*k, &*v) })

or

        self.iter_raw.next().map(|(k, v)| {
            // SAFETY: Due to `self._tree`, `k` and `v` are valid for the lifetime of `'a`.
            unsafe { (&*k, &*v) }
        })

I hope that this seems consistent, my motivation behind the suggestions 
are that I don't like the comment splitting the single line.

---
Cheers,
Benno