Re: [PATCH v8 5/6] rust: rbtree: add `RBTreeCursor`

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

On 27.07.24 22:30, Matt Gilbride wrote:
> Add a cursor interface to `RBTree`, supporting the following use cases:
> - Inspect the current node pointed to by the cursor, inspect/move to
>   it's neighbors in sort order (bidirectionally).
> - Mutate the tree itself by removing the current node pointed to by the
>   cursor, or one of its neighbors.
> 
> Add functions to obtain a cursor to the tree by key:
> - The node with the smallest key
> - The node with the largest key
> - The node matching the given key, or the one with the next larger key
> 
> The cursor abstraction is needed by the binder driver to efficiently
> search for nodes and (conditionally) modify them, as well as their
> neighbors [1].
> 
> Link: https://lore.kernel.org/rust-for-linux/20231101-rust-binder-v1-6-08ba9197f637@google.com/ [1]
> Co-developed-by: Alice Ryhl <aliceryhl@google.com>
> Signed-off-by: Alice Ryhl <aliceryhl@google.com>
> Reviewed-by: Alice Ryhl <aliceryhl@google.com>
> Tested-by: Alice Ryhl <aliceryhl@google.com>
> Signed-off-by: Matt Gilbride <mattgilbride@google.com>
> ---
>  rust/kernel/rbtree.rs | 536 ++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 536 insertions(+)
> 
> diff --git a/rust/kernel/rbtree.rs b/rust/kernel/rbtree.rs
> index d7514ebadfa8..5d37aa373685 100644
> --- a/rust/kernel/rbtree.rs
> +++ b/rust/kernel/rbtree.rs
> @@ -234,6 +234,40 @@ pub fn values(&self) -> impl Iterator<Item = &'_ V> {
>      pub fn values_mut(&mut self) -> impl Iterator<Item = &'_ mut V> {
>          self.iter_mut().map(|(_, v)| v)
>      }
> +
> +    /// Returns a cursor over the tree nodes, starting with the smallest key.
> +    pub fn cursor_front(&mut self) -> Option<RBTreeCursor<'_, K, V>> {
> +        let root = addr_of_mut!(self.root);
> +        // SAFETY: `self.root` is always a valid root node
> +        let current = unsafe { bindings::rb_first(root) };
> +        NonNull::new(current).map(|current| {
> +            // INVARIANT:
> +            // - `current` is a valid node in the [`RBTree`] pointed to by `self`.
> +            // - Due to the type signature of this function, the returned [`RBTreeCursor`]
> +            //   borrows mutably from `self`.
> +            RBTreeCursor {
> +                current,
> +                tree: self,
> +            }
> +        })
> +    }
> +
> +    /// Returns a cursor over the tree nodes, starting with the largest key.
> +    pub fn cursor_back(&mut self) -> Option<RBTreeCursor<'_, K, V>> {
> +        let root = addr_of_mut!(self.root);
> +        // SAFETY: `self.root` is always a valid root node
> +        let current = unsafe { bindings::rb_last(root) };
> +        NonNull::new(current).map(|current| {
> +            // INVARIANT:
> +            // - `current` is a valid node in the [`RBTree`] pointed to by `self`.
> +            // - Due to the type signature of this function, the returned [`RBTreeCursor`]
> +            //   borrows mutably from `self`.
> +            RBTreeCursor {
> +                current,
> +                tree: self,
> +            }
> +        })
> +    }
>  }
> 
>  impl<K, V> RBTree<K, V>
> @@ -394,6 +428,75 @@ fn remove_node(&mut self, key: &K) -> Option<RBTreeNode<K, V>> {
>      pub fn remove(&mut self, key: &K) -> Option<V> {
>          self.remove_node(key).map(|node| node.node.value)
>      }
> +
> +    /// Returns a cursor over the tree nodes based on the given key.
> +    ///
> +    /// If the given key exists, the cursor starts there.
> +    /// Otherwise it starts with the first larger key in sort order.
> +    /// If there is no larger key, it returns [`None`].
> +    pub fn cursor_lower_bound(&mut self, key: &K) -> Option<RBTreeCursor<'_, K, V>>
> +    where
> +        K: Ord,
> +    {
> +        let mut node = self.root.rb_node;
> +        let mut best_match: Option<NonNull<Node<K, V>>> = None;
> +        while !node.is_null() {
> +            // SAFETY: By the type invariant of `Self`, all non-null `rb_node` pointers stored in `self`
> +            // point to the links field of `Node<K, V>` objects.
> +            let this = unsafe { container_of!(node, Node<K, V>, links) }.cast_mut();
> +            // SAFETY: `this` is a non-null node so it is valid by the type invariants.
> +            let this_key = unsafe { &(*this).key };
> +            // SAFETY: `node` is a non-null node so it is valid by the type invariants.
> +            let left_child = unsafe { (*node).rb_left };
> +            // SAFETY: `node` is a non-null node so it is valid by the type invariants.
> +            let right_child = unsafe { (*node).rb_right };
> +            if key == this_key {
> +                return NonNull::new(node).map(|current| {
> +                    // INVARIANT:
> +                    // - `node` is a valid node in the [`RBTree`] pointed to by `self`.
> +                    // - Due to the type signature of this function, the returned [`RBTreeCursor`]
> +                    //   borrows mutably from `self`.
> +                    RBTreeCursor {
> +                        current,
> +                        tree: self,
> +                    }
> +                });
> +            } else {
> +                node = if key > this_key {
> +                    right_child
> +                } else {
> +                    let is_better_match = match best_match {
> +                        None => true,
> +                        Some(best) => {
> +                            // SAFETY: `best` is a non-null node so it is valid by the type invariants.
> +                            let best_key = unsafe { &(*best.as_ptr()).key };
> +                            best_key > this_key
> +                        }
> +                    };
> +                    if is_better_match {
> +                        best_match = NonNull::new(this);
> +                    }
> +                    left_child
> +                };
> +            }
> +        }
> +
> +        let best = best_match?;
> +
> +        // SAFETY: `best` is a non-null node so it is valid by the type invariants.
> +        let links = unsafe { addr_of_mut!((*best.as_ptr()).links) };
> +
> +        NonNull::new(links).map(|current| {

Why would `links` be a null pointer? AFAIK it just came from `best`
which is non-null. (I don't know if we want to use `new_unchecked`
instead, but wanted to mention it)

> +            // INVARIANT:
> +            // - `current` is a valid node in the [`RBTree`] pointed to by `self`.
> +            // - Due to the type signature of this function, the returned [`RBTreeCursor`]
> +            //   borrows mutably from `self`.
> +            RBTreeCursor {
> +                current,
> +                tree: self,
> +            }
> +        })
> +    }

[...]

> +/// // Calling `remove_next` removes and returns the last element.
> +/// assert_eq!(cursor.remove_next().unwrap().to_key_value(), (30, 300));
> +///
> +/// # Ok::<(), Error>(())
> +/// ```

I would put a newline here.

> +/// # Invariants
> +/// - `current` points to a node that is in the same [`RBTree`] as `tree`.
> +pub struct RBTreeCursor<'a, K, V> {

I think we can name it just `Cursor`, since one can refer to it as
`rbtree::Cursor` and then it also follows the naming scheme for `Iter`
etc.

> +    tree: &'a mut RBTree<K, V>,
> +    current: NonNull<bindings::rb_node>,
> +}
> +
> +// SAFETY: The [`RBTreeCursor`] gives out immutable references to K and mutable references to V,
> +// so it has the same thread safety requirements as mutable references.
> +unsafe impl<'a, K: Send, V: Send> Send for RBTreeCursor<'a, K, V> {}

Again, do we want to use `K: Sync` here instead?

> +
> +// SAFETY: The [`RBTreeCursor`] gives out immutable references to K and mutable references to V,
> +// so it has the same thread safety requirements as mutable references.
> +unsafe impl<'a, K: Sync, V: Sync> Sync for RBTreeCursor<'a, K, V> {}
> +
> +impl<'a, K, V> RBTreeCursor<'a, K, V> {
> +    /// The current node
> +    pub fn current(&self) -> (&K, &V) {
> +        // SAFETY:
> +        // - `self.current` is a valid node by the type invariants.
> +        // - We have an immutable reference by the function signature.
> +        unsafe { Self::to_key_value(self.current) }
> +    }
> +
> +    /// The current node, with a mutable value
> +    pub fn current_mut(&mut self) -> (&K, &mut V) {
> +        // SAFETY:
> +        // - `self.current` is a valid node by the type invariants.
> +        // - We have an mutable reference by the function signature.
> +        unsafe { Self::to_key_value_mut(self.current) }
> +    }
> +
> +    /// Remove the current node from the tree.
> +    ///
> +    /// Returns a tuple where the first element is a cursor to the next node, if it exists,
> +    /// else the previous node, else [`None`] (if the tree becomes empty). The second element
> +    /// is the removed node.
> +    pub fn remove_current(self) -> (Option<Self>, RBTreeNode<K, V>) {
> +        let prev = self.get_neighbor_raw(Direction::Prev);
> +        let next = self.get_neighbor_raw(Direction::Next);
> +        // SAFETY: By the type invariant of `Self`, all non-null `rb_node` pointers stored in `self`
> +        // point to the links field of `Node<K, V>` objects.
> +        let this = unsafe { container_of!(self.current.as_ptr(), Node<K, V>, links) }.cast_mut();
> +        // SAFETY: `this` is valid by the type invariants as described above.
> +        let node = unsafe { Box::from_raw(this) };
> +        let node = RBTreeNode { node };
> +        // SAFETY: The reference to the tree used to create the cursor outlives the cursor, so
> +        // the tree cannot change. By the tree invariant, all nodes are valid.
> +        unsafe { bindings::rb_erase(&mut (*this).links, addr_of_mut!(self.tree.root)) };
> +
> +        let current = match (prev, next) {
> +            (_, Some(next)) => next,
> +            (Some(prev), None) => prev,
> +            (None, None) => {
> +                return (None, node);
> +            }
> +        };
> +
> +        (
> +            // INVARIANT:
> +            // - `current` is a valid node in the [`RBTree`] pointed to by `self.tree`.
> +            // - Due to the function signature, `self` is an owned [`RBTreeCursor`],
> +            //   and [`RBTreeCursor`]s are only created via functions with a mutable reference
> +            //   to an [`RBTree`].
> +            Some(Self {
> +                current,
> +                tree: self.tree,
> +            }),
> +            node,
> +        )
> +    }
> +
> +    /// Remove the previous node, returning it if it exists.
> +    pub fn remove_prev(&mut self) -> Option<RBTreeNode<K, V>> {
> +        self.remove_neighbor(Direction::Prev)
> +    }
> +
> +    /// Remove the next node, returning it if it exists.
> +    pub fn remove_next(&mut self) -> Option<RBTreeNode<K, V>> {
> +        self.remove_neighbor(Direction::Next)
> +    }
> +
> +    fn remove_neighbor(&mut self, direction: Direction) -> Option<RBTreeNode<K, V>> {
> +        if let Some(neighbor) = self.get_neighbor_raw(direction) {
> +            let neighbor = neighbor.as_ptr();
> +            // SAFETY: The reference to the tree used to create the cursor outlives the cursor, so
> +            // the tree cannot change. By the tree invariant, all nodes are valid.
> +            unsafe { bindings::rb_erase(neighbor, addr_of_mut!(self.tree.root)) };
> +            // SAFETY: By the type invariant of `Self`, all non-null `rb_node` pointers stored in `self`
> +            // point to the links field of `Node<K, V>` objects.
> +            let this = unsafe { container_of!(neighbor, Node<K, V>, links) }.cast_mut();
> +            // SAFETY: `this` is valid by the type invariants as described above.
> +            let node = unsafe { Box::from_raw(this) };
> +            return Some(RBTreeNode { node });
> +        }
> +        None
> +    }
> +
> +    /// Move the cursor to the previous node, returning [`None`] if it doesn't exist.
> +    pub fn move_prev(self) -> Option<Self> {
> +        self.mv(Direction::Prev)
> +    }
> +
> +    /// Move the cursor to the next node, returning [`None`] if it doesn't exist.
> +    pub fn move_next(self) -> Option<Self> {
> +        self.mv(Direction::Next)
> +    }
> +
> +    fn mv(self, direction: Direction) -> Option<Self> {
> +        // INVARIANT:
> +        // - `neighbor` is a valid node in the [`RBTree`] pointed to by `self.tree`.
> +        // - Due to the function signature, `self` is an owned [`RBTreeCursor`],
> +        //   and [`RBTreeCursor`]s are only created via functions with a mutable reference
> +        //   to an [`RBTree`].
> +        self.get_neighbor_raw(direction).map(|neighbor| Self {
> +            tree: self.tree,
> +            current: neighbor,
> +        })
> +    }
> +
> +    /// Access the previous node without moving the cursor.
> +    pub fn peek_prev(&self) -> Option<(&K, &V)> {
> +        self.peek(Direction::Prev)
> +    }
> +
> +    /// Access the previous node without moving the cursor.
> +    pub fn peek_next(&self) -> Option<(&K, &V)> {
> +        self.peek(Direction::Next)
> +    }
> +
> +    fn peek(&self, direction: Direction) -> Option<(&K, &V)> {
> +        self.get_neighbor_raw(direction)
> +            // SAFETY:
> +            // - `neighbor` is a valid tree node.
> +            // - By the function signature, we have an immutable reference to `self`.
> +            .map(|neighbor| unsafe { Self::to_key_value(neighbor) })

Alternative way of formatting this:

        self.get_neighbor_raw(direction).map(|neighbor| {
            // SAFETY:
            // - `neighbor` is a valid tree node.
            // - By the function signature, we have an immutable reference to `self`.
            unsafe { Self::to_key_value(neighbor) }
        })

I think it looks nicer, but we should probably have a written
preference.

> +    }
> +
> +    /// Access the previous node mutably without moving the cursor.
> +    pub fn peek_prev_mut(&mut self) -> Option<(&K, &mut V)> {
> +        self.peek_mut(Direction::Prev)
> +    }
> +
> +    /// Access the next node mutably without moving the cursor.
> +    pub fn peek_next_mut(&mut self) -> Option<(&K, &mut V)> {
> +        self.peek_mut(Direction::Next)
> +    }
> +
> +    fn peek_mut(&mut self, direction: Direction) -> Option<(&K, &mut V)> {
> +        self.get_neighbor_raw(direction)
> +            // SAFETY:
> +            // - `neighbor` is a valid tree node.
> +            // - By the function signature, we have a mutable reference to `self`.
> +            .map(|neighbor| unsafe { Self::to_key_value_mut(neighbor) })

Ditto.

---
Cheers,
Benno

> +    }
> +
> +    fn get_neighbor_raw(&self, direction: Direction) -> Option<NonNull<bindings::rb_node>> {
> +        // SAFETY: `self.current` is valid by the type invariants.
> +        let neighbor = unsafe {
> +            match direction {
> +                Direction::Prev => bindings::rb_prev(self.current.as_ptr()),
> +                Direction::Next => bindings::rb_next(self.current.as_ptr()),
> +            }
> +        };
> +
> +        NonNull::new(neighbor)
> +    }
> +
> +    /// SAFETY:
> +    /// - `node` must be a valid pointer to a node in an [`RBTree`].
> +    /// - The caller has immutable access to `node` for the duration of 'b.
> +    unsafe fn to_key_value<'b>(node: NonNull<bindings::rb_node>) -> (&'b K, &'b V) {
> +        // SAFETY: the caller guarantees that `node` is a valid pointer in an `RBTree`.
> +        let (k, v) = unsafe { Self::to_key_value_raw(node) };
> +        // SAFETY: the caller guarantees immutable access to `node`.
> +        (k, unsafe { &*v })
> +    }
> +
> +    /// SAFETY:
> +    /// - `node` must be a valid pointer to a node in an [`RBTree`].
> +    /// - The caller has mutable access to `node` for the duration of 'b.
> +    unsafe fn to_key_value_mut<'b>(node: NonNull<bindings::rb_node>) -> (&'b K, &'b mut V) {
> +        // SAFETY: the caller guarantees that `node` is a valid pointer in an `RBTree`.
> +        let (k, v) = unsafe { Self::to_key_value_raw(node) };
> +        // SAFETY: the caller guarantees mutable access to `node`.
> +        (k, unsafe { &mut *v })
> +    }
> +
> +    /// SAFETY:
> +    /// - `node` must be a valid pointer to a node in an [`RBTree`].
> +    /// - The caller has immutable access to the key for the duration of 'b.
> +    unsafe fn to_key_value_raw<'b>(node: NonNull<bindings::rb_node>) -> (&'b K, *mut V) {
> +        // SAFETY: By the type invariant of `Self`, all non-null `rb_node` pointers stored in `self`
> +        // point to the links field of `Node<K, V>` objects.
> +        let this = unsafe { container_of!(node.as_ptr(), Node<K, V>, links) }.cast_mut();
> +        // SAFETY: The passed `node` is the current node or a non-null neighbor,
> +        // thus `this` is valid by the type invariants.
> +        let k = unsafe { &(*this).key };
> +        // SAFETY: The passed `node` is the current node or a non-null neighbor,
> +        // thus `this` is valid by the type invariants.
> +        let v = unsafe { addr_of_mut!((*this).value) };
> +        (k, v)
> +    }
> +}
> +
> +/// Direction for [`RBTreeCursor`] operations.
> +enum Direction {
> +    /// the node immediately before, in sort order
> +    Prev,
> +    /// the node immediately after, in sort order
> +    Next,
> +}
> +
>  impl<'a, K, V> IntoIterator for &'a RBTree<K, V> {
>      type Item = (&'a K, &'a V);
>      type IntoIter = Iter<'a, K, V>;
> @@ -583,6 +1114,11 @@ impl<K, V> RBTreeNode<K, V> {
>      pub fn new(key: K, value: V, flags: Flags) -> Result<RBTreeNode<K, V>> {
>          Ok(RBTreeNodeReservation::new(flags)?.into_node(key, value))
>      }
> +
> +    /// Get the key and value from inside the node.
> +    pub fn to_key_value(self) -> (K, V) {
> +        (self.node.key, self.node.value)
> +    }
>  }
> 
>  // SAFETY: If K and V can be sent across threads, then it's also okay to send [`RBTreeNode`] across
> 
> --
> 2.46.0.rc1.232.g9752f9e123-goog
>