ffi.rs

// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

//! Contains declarations to bind to the [C Data Interface](https://arrow.apache.org/docs/format/CDataInterface.html).
//!
//! Generally, this module is divided in two main interfaces:
//! One interface maps C ABI to native Rust types, i.e. convert c-pointers, c_char, to native rust.
//! This is handled by [FFI_ArrowSchema] and [FFI_ArrowArray].
//!
//! The second interface maps native Rust types to the Rust-specific implementation of Arrow such as `format` to `Datatype`,
//! `Buffer`, etc. This is handled by `ArrowArray`.
//!
//! ```rust
//! # use std::sync::Arc;
//! # use arrow::array::{Int32Array, Array, ArrayData, export_array_into_raw, make_array, make_array_from_raw};
//! # use arrow::error::{Result, ArrowError};
//! # use arrow::compute::kernels::arithmetic;
//! # use arrow::ffi::{FFI_ArrowArray, FFI_ArrowSchema};
//! # use std::convert::TryFrom;
//! # fn main() -> Result<()> {
//! // create an array natively
//! let array = Int32Array::from(vec![Some(1), None, Some(3)]);
//!
//! // export it
//! let (array_ptr, schema_ptr) = array.to_raw()?;
//!
//! // consumed and used by something else...
//!
//! // import it
//! let array = unsafe { make_array_from_raw(array_ptr, schema_ptr)? };
//!
//! // perform some operation
//! let array = array.as_any().downcast_ref::<Int32Array>().ok_or(
//!     ArrowError::ParseError("Expects an int32".to_string()),
//! )?;
//! let array = arithmetic::add(&array, &array)?;
//!
//! // verify
//! assert_eq!(array, Int32Array::from(vec![Some(2), None, Some(6)]));
//!
//! // Simulate if raw pointers are provided by consumer
//! let array = make_array(Int32Array::from(vec![Some(1), None, Some(3)]).data().clone());
//!
//! let out_array = Box::new(FFI_ArrowArray::empty());
//! let out_schema = Box::new(FFI_ArrowSchema::empty());
//! let out_array_ptr = Box::into_raw(out_array);
//! let out_schema_ptr = Box::into_raw(out_schema);
//!
//! // export array into raw pointers from consumer
//! unsafe { export_array_into_raw(array, out_array_ptr, out_schema_ptr)?; };
//!
//! // import it
//! let array = unsafe { make_array_from_raw(out_array_ptr, out_schema_ptr)? };
//!
//! // perform some operation
//! let array = array.as_any().downcast_ref::<Int32Array>().ok_or(
//!     ArrowError::ParseError("Expects an int32".to_string()),
//! )?;
//! let array = arithmetic::add(&array, &array)?;
//!
//! // verify
//! assert_eq!(array, Int32Array::from(vec![Some(2), None, Some(6)]));
//!
//! // (drop/release)
//! unsafe {
//!     Box::from_raw(out_array_ptr);
//!     Box::from_raw(out_schema_ptr);
//!     Arc::from_raw(array_ptr);
//!     Arc::from_raw(schema_ptr);
//! }
//!
//! Ok(())
//! }
//! ```

/*
# Design:

Main assumptions:
* A memory region is deallocated according it its own release mechanism.
* Rust shares memory regions between arrays.
* A memory region should be deallocated when no-one is using it.

The design of this module is as follows:

`ArrowArray` contains two `Arc`s, one per ABI-compatible `struct`, each containing data
according to the C Data Interface. These Arcs are used for ref counting of the structs
within Rust and lifetime management.

Each ABI-compatible `struct` knowns how to `drop` itself, calling `release`.

To import an array, unsafely create an `ArrowArray` from two pointers using [ArrowArray::try_from_raw].
To export an array, create an `ArrowArray` using [ArrowArray::try_new].
*/

use std::{
    convert::TryFrom,
    ffi::CStr,
    ffi::CString,
    iter,
    mem::size_of,
    os::raw::{c_char, c_void},
    ptr::{self, NonNull},
    sync::Arc,
};

use bitflags::bitflags;

use crate::array::{layout, ArrayData};
use crate::buffer::Buffer;
use crate::datatypes::DataType;
use crate::error::{ArrowError, Result};
use crate::util::bit_util;

bitflags! {
    pub struct Flags: i64 {
        const DICTIONARY_ORDERED = 0b00000001;
        const NULLABLE = 0b00000010;
        const MAP_KEYS_SORTED = 0b00000100;
    }
}

/// ABI-compatible struct for `ArrowSchema` from C Data Interface
/// See <https://arrow.apache.org/docs/format/CDataInterface.html#structure-definitions>
/// This was created by bindgen
#[repr(C)]
#[derive(Debug)]
pub struct FFI_ArrowSchema {
    pub(crate) format: *const c_char,
    pub(crate) name: *const c_char,
    pub(crate) metadata: *const c_char,
    pub(crate) flags: i64,
    pub(crate) n_children: i64,
    pub(crate) children: *mut *mut FFI_ArrowSchema,
    pub(crate) dictionary: *mut FFI_ArrowSchema,
    pub(crate) release: Option<unsafe extern "C" fn(arg1: *mut FFI_ArrowSchema)>,
    pub(crate) private_data: *mut c_void,
}

struct SchemaPrivateData {
    children: Box<[*mut FFI_ArrowSchema]>,
    dictionary: *mut FFI_ArrowSchema,
}

// callback used to drop [FFI_ArrowSchema] when it is exported.
unsafe extern "C" fn release_schema(schema: *mut FFI_ArrowSchema) {
    if schema.is_null() {
        return;
    }
    let schema = &mut *schema;

    // take ownership back to release it.
    drop(CString::from_raw(schema.format as *mut c_char));
    if !schema.name.is_null() {
        drop(CString::from_raw(schema.name as *mut c_char));
    }
    if !schema.private_data.is_null() {
        let private_data = Box::from_raw(schema.private_data as *mut SchemaPrivateData);
        for child in private_data.children.iter() {
            drop(Box::from_raw(*child))
        }
        if !private_data.dictionary.is_null() {
            drop(Box::from_raw(private_data.dictionary));
        }

        drop(private_data);
    }

    schema.release = None;
}

impl FFI_ArrowSchema {
    /// create a new [`FFI_ArrowSchema`]. This fails if the fields'
    /// [`DataType`] is not supported.
    pub fn try_new(
        format: &str,
        children: Vec<FFI_ArrowSchema>,
        dictionary: Option<FFI_ArrowSchema>,
    ) -> Result<Self> {
        let mut this = Self::empty();

        let children_ptr = children
            .into_iter()
            .map(Box::new)
            .map(Box::into_raw)
            .collect::<Box<_>>();

        this.format = CString::new(format).unwrap().into_raw();
        this.release = Some(release_schema);
        this.n_children = children_ptr.len() as i64;

        let dictionary_ptr = dictionary
            .map(|d| Box::into_raw(Box::new(d)))
            .unwrap_or(std::ptr::null_mut());

        let mut private_data = Box::new(SchemaPrivateData {
            children: children_ptr,
            dictionary: dictionary_ptr,
        });

        // intentionally set from private_data (see https://github.com/apache/arrow-rs/issues/580)
        this.children = private_data.children.as_mut_ptr();

        this.dictionary = dictionary_ptr;

        this.private_data = Box::into_raw(private_data) as *mut c_void;

        Ok(this)
    }

    pub fn with_name(mut self, name: &str) -> Result<Self> {
        self.name = CString::new(name).unwrap().into_raw();
        Ok(self)
    }

    pub fn with_flags(mut self, flags: Flags) -> Result<Self> {
        self.flags = flags.bits();
        Ok(self)
    }

    pub fn empty() -> Self {
        Self {
            format: std::ptr::null_mut(),
            name: std::ptr::null_mut(),
            metadata: std::ptr::null_mut(),
            flags: 0,
            n_children: 0,
            children: ptr::null_mut(),
            dictionary: std::ptr::null_mut(),
            release: None,
            private_data: std::ptr::null_mut(),
        }
    }

    /// returns the format of this schema.
    pub fn format(&self) -> &str {
        assert!(!self.format.is_null());
        // safe because the lifetime of `self.format` equals `self`
        unsafe { CStr::from_ptr(self.format) }
            .to_str()
            .expect("The external API has a non-utf8 as format")
    }

    /// returns the name of this schema.
    pub fn name(&self) -> &str {
        assert!(!self.name.is_null());
        // safe because the lifetime of `self.name` equals `self`
        unsafe { CStr::from_ptr(self.name) }
            .to_str()
            .expect("The external API has a non-utf8 as name")
    }

    pub fn flags(&self) -> Option<Flags> {
        Flags::from_bits(self.flags)
    }

    pub fn child(&self, index: usize) -> &Self {
        assert!(index < self.n_children as usize);
        unsafe { self.children.add(index).as_ref().unwrap().as_ref().unwrap() }
    }

    pub fn children(&self) -> impl Iterator<Item = &Self> {
        (0..self.n_children as usize).map(move |i| self.child(i))
    }

    pub fn nullable(&self) -> bool {
        (self.flags / 2) & 1 == 1
    }

    pub fn dictionary(&self) -> Option<&Self> {
        unsafe { self.dictionary.as_ref() }
    }
}

impl Drop for FFI_ArrowSchema {
    fn drop(&mut self) {
        match self.release {
            None => (),
            Some(release) => unsafe { release(self) },
        };
    }
}

// returns the number of bits that buffer `i` (in the C data interface) is expected to have.
// This is set by the Arrow specification
#[allow(clippy::manual_bits)]
fn bit_width(data_type: &DataType, i: usize) -> Result<usize> {
    Ok(match (data_type, i) {
        // the null buffer is bit sized
        (_, 0) => 1,
        // primitive types first buffer's size is given by the native types
        (DataType::Boolean, 1) => 1,
        (DataType::UInt8, 1) => size_of::<u8>() * 8,
        (DataType::UInt16, 1) => size_of::<u16>() * 8,
        (DataType::UInt32, 1) => size_of::<u32>() * 8,
        (DataType::UInt64, 1) => size_of::<u64>() * 8,
        (DataType::Int8, 1) => size_of::<i8>() * 8,
        (DataType::Int16, 1) => size_of::<i16>() * 8,
        (DataType::Int32, 1) | (DataType::Date32, 1) | (DataType::Time32(_), 1) => size_of::<i32>() * 8,
        (DataType::Int64, 1) | (DataType::Date64, 1) | (DataType::Time64(_), 1) => size_of::<i64>() * 8,
        (DataType::Float32, 1) => size_of::<f32>() * 8,
        (DataType::Float64, 1) => size_of::<f64>() * 8,
        (DataType::Decimal(..), 1) => size_of::<i128>() * 8,
        (DataType::Timestamp(..), 1) => size_of::<i64>() * 8,
        (DataType::Duration(..), 1) => size_of::<i64>() * 8,
        // primitive types have a single buffer
        (DataType::Boolean, _) |
        (DataType::UInt8, _) |
        (DataType::UInt16, _) |
        (DataType::UInt32, _) |
        (DataType::UInt64, _) |
        (DataType::Int8, _) |
        (DataType::Int16, _) |
        (DataType::Int32, _) | (DataType::Date32, _) | (DataType::Time32(_), _) |
        (DataType::Int64, _) | (DataType::Date64, _) | (DataType::Time64(_), _) |
        (DataType::Float32, _) |
        (DataType::Float64, _) |
        (DataType::Decimal(..), _) |
        (DataType::Timestamp(..), _) |
        (DataType::Duration(..), _) => {
            return Err(ArrowError::CDataInterface(format!(
                "The datatype \"{:?}\" expects 2 buffers, but requested {}. Please verify that the C data interface is correctly implemented.",
                data_type, i
            )))
        }
        (DataType::FixedSizeBinary(num_bytes), 1) => size_of::<u8>() * (*num_bytes as usize) * 8,
        (DataType::FixedSizeList(f, num_elems), 1) => {
            let child_bit_width = bit_width(f.data_type(), 1)?;
            child_bit_width * (*num_elems as usize)
        },
        (DataType::FixedSizeBinary(_), _) | (DataType::FixedSizeList(_, _), _) => {
            return Err(ArrowError::CDataInterface(format!(
                "The datatype \"{:?}\" expects 2 buffers, but requested {}. Please verify that the C data interface is correctly implemented.",
                data_type, i
            )))
        },
        // Variable-sized binaries: have two buffers.
        // "small": first buffer is i32, second is in bytes
        (DataType::Utf8, 1) | (DataType::Binary, 1) | (DataType::List(_), 1) => size_of::<i32>() * 8,
        (DataType::Utf8, 2) | (DataType::Binary, 2) | (DataType::List(_), 2) => size_of::<u8>() * 8,
        (DataType::Utf8, _) | (DataType::Binary, _) | (DataType::List(_), _)=> {
            return Err(ArrowError::CDataInterface(format!(
                "The datatype \"{:?}\" expects 3 buffers, but requested {}. Please verify that the C data interface is correctly implemented.",
                data_type, i
            )))
        }
        // Variable-sized binaries: have two buffers.
        // LargeUtf8: first buffer is i64, second is in bytes
        (DataType::LargeUtf8, 1) | (DataType::LargeBinary, 1) | (DataType::LargeList(_), 1) => size_of::<i64>() * 8,
        (DataType::LargeUtf8, 2) | (DataType::LargeBinary, 2) | (DataType::LargeList(_), 2)=> size_of::<u8>() * 8,
        (DataType::LargeUtf8, _) | (DataType::LargeBinary, _) | (DataType::LargeList(_), _)=> {
            return Err(ArrowError::CDataInterface(format!(
                "The datatype \"{:?}\" expects 3 buffers, but requested {}. Please verify that the C data interface is correctly implemented.",
                data_type, i
            )))
        }
        _ => {
            return Err(ArrowError::CDataInterface(format!(
                "The datatype \"{:?}\" is still not supported in Rust implementation",
                data_type
            )))
        }
    })
}

/// ABI-compatible struct for ArrowArray from C Data Interface
/// See <https://arrow.apache.org/docs/format/CDataInterface.html#structure-definitions>
/// This was created by bindgen
#[repr(C)]
#[derive(Debug)]
pub struct FFI_ArrowArray {
    pub(crate) length: i64,
    pub(crate) null_count: i64,
    pub(crate) offset: i64,
    pub(crate) n_buffers: i64,
    pub(crate) n_children: i64,
    pub(crate) buffers: *mut *const c_void,
    pub(crate) children: *mut *mut FFI_ArrowArray,
    pub(crate) dictionary: *mut FFI_ArrowArray,
    pub(crate) release: Option<unsafe extern "C" fn(arg1: *mut FFI_ArrowArray)>,
    // When exported, this MUST contain everything that is owned by this array.
    // for example, any buffer pointed to in `buffers` must be here, as well
    // as the `buffers` pointer itself.
    // In other words, everything in [FFI_ArrowArray] must be owned by
    // `private_data` and can assume that they do not outlive `private_data`.
    pub(crate) private_data: *mut c_void,
}

impl Drop for FFI_ArrowArray {
    fn drop(&mut self) {
        match self.release {
            None => (),
            Some(release) => unsafe { release(self) },
        };
    }
}

unsafe impl Send for FFI_ArrowArray {}
unsafe impl Sync for FFI_ArrowArray {}

// callback used to drop [FFI_ArrowArray] when it is exported
unsafe extern "C" fn release_array(array: *mut FFI_ArrowArray) {
    if array.is_null() {
        return;
    }
    let array = &mut *array;

    // take ownership of `private_data`, therefore dropping it`
    let private = Box::from_raw(array.private_data as *mut ArrayPrivateData);
    for child in private.children.iter() {
        let _ = Box::from_raw(*child);
    }
    if !private.dictionary.is_null() {
        let _ = Box::from_raw(private.dictionary);
    }

    array.release = None;
}

struct ArrayPrivateData {
    #[allow(dead_code)]
    buffers: Vec<Option<Buffer>>,
    buffers_ptr: Box<[*const c_void]>,
    children: Box<[*mut FFI_ArrowArray]>,
    dictionary: *mut FFI_ArrowArray,
}

impl FFI_ArrowArray {
    /// creates a new `FFI_ArrowArray` from existing data.
    /// # Safety
    /// This method releases `buffers`. Consumers of this struct *must* call `release` before
    /// releasing this struct, or contents in `buffers` leak.
    pub fn new(data: &ArrayData) -> Self {
        // * insert the null buffer at the start
        // * make all others `Option<Buffer>`.
        let buffers = iter::once(data.null_buffer().cloned())
            .chain(data.buffers().iter().map(|b| Some(b.clone())))
            .collect::<Vec<_>>();
        let data_layout = layout(data.data_type());
        // `n_buffers` is the number of buffers by the spec.
        let n_buffers = {
            data_layout.buffers.len() + {
                // If the layout has a null buffer by Arrow spec.
                // Note that even the array doesn't have a null buffer because it has
                // no null value, we still need to count 1 here to follow the spec.
                if data_layout.can_contain_null_mask {
                    1
                } else {
                    0
                }
            }
        } as i64;

        let buffers_ptr = buffers
            .iter()
            .flat_map(|maybe_buffer| match maybe_buffer {
                // note that `raw_data` takes into account the buffer's offset
                Some(b) => Some(b.as_ptr() as *const c_void),
                // This is for null buffer. We only put a null pointer for
                // null buffer if by spec it can contain null mask.
                None if data_layout.can_contain_null_mask => Some(std::ptr::null()),
                None => None,
            })
            .collect::<Box<[_]>>();

        let empty = vec![];
        let (child_data, dictionary) = match data.data_type() {
            DataType::Dictionary(_, _) => (
                empty.as_slice(),
                Box::into_raw(Box::new(FFI_ArrowArray::new(&data.child_data()[0]))),
            ),
            _ => (data.child_data(), std::ptr::null_mut()),
        };

        let children = child_data
            .iter()
            .map(|child| Box::into_raw(Box::new(FFI_ArrowArray::new(child))))
            .collect::<Box<_>>();
        let n_children = children.len() as i64;

        // create the private data owning everything.
        // any other data must be added here, e.g. via a struct, to track lifetime.
        let mut private_data = Box::new(ArrayPrivateData {
            buffers,
            buffers_ptr,
            children,
            dictionary,
        });

        Self {
            length: data.len() as i64,
            null_count: data.null_count() as i64,
            offset: data.offset() as i64,
            n_buffers,
            n_children,
            buffers: private_data.buffers_ptr.as_mut_ptr(),
            children: private_data.children.as_mut_ptr(),
            dictionary,
            release: Some(release_array),
            private_data: Box::into_raw(private_data) as *mut c_void,
        }
    }

    /// create an empty `FFI_ArrowArray`, which can be used to import data into
    pub fn empty() -> Self {
        Self {
            length: 0,
            null_count: 0,
            offset: 0,
            n_buffers: 0,
            n_children: 0,
            buffers: std::ptr::null_mut(),
            children: std::ptr::null_mut(),
            dictionary: std::ptr::null_mut(),
            release: None,
            private_data: std::ptr::null_mut(),
        }
    }

    /// the length of the array
    pub fn len(&self) -> usize {
        self.length as usize
    }

    /// whether the array is empty
    pub fn is_empty(&self) -> bool {
        self.length == 0
    }

    /// the offset of the array
    pub fn offset(&self) -> usize {
        self.offset as usize
    }

    /// the null count of the array
    pub fn null_count(&self) -> usize {
        self.null_count as usize
    }
}

/// returns a new buffer corresponding to the index `i` of the FFI array. It may not exist (null pointer).
/// `bits` is the number of bits that the native type of this buffer has.
/// The size of the buffer will be `ceil(self.length * bits, 8)`.
/// # Panic
/// This function panics if `i` is larger or equal to `n_buffers`.
/// # Safety
/// This function assumes that `ceil(self.length * bits, 8)` is the size of the buffer
unsafe fn create_buffer(
    owner: Arc<FFI_ArrowArray>,
    array: &FFI_ArrowArray,
    index: usize,
    len: usize,
) -> Option<Buffer> {
    if array.buffers.is_null() {
        return None;
    }
    let buffers = array.buffers as *mut *const u8;

    assert!(index < array.n_buffers as usize);
    let ptr = *buffers.add(index);

    NonNull::new(ptr as *mut u8)
        .map(|ptr| Buffer::from_custom_allocation(ptr, len, owner))
}

fn create_child(
    owner: Arc<FFI_ArrowArray>,
    array: &FFI_ArrowArray,
    schema: &FFI_ArrowSchema,
    index: usize,
) -> ArrowArrayChild<'static> {
    assert!(index < array.n_children as usize);
    assert!(!array.children.is_null());
    assert!(!array.children.is_null());
    unsafe {
        let arr_ptr = *array.children.add(index);
        let schema_ptr = *schema.children.add(index);
        assert!(!arr_ptr.is_null());
        assert!(!schema_ptr.is_null());
        let arr_ptr = &*arr_ptr;
        let schema_ptr = &*schema_ptr;
        ArrowArrayChild::from_raw(arr_ptr, schema_ptr, owner)
    }
}

pub trait ArrowArrayRef {
    fn to_data(&self) -> Result<ArrayData> {
        let data_type = self.data_type()?;
        let len = self.array().len();
        let offset = self.array().offset();
        let null_count = self.array().null_count();
        let buffers = self.buffers()?;
        let null_bit_buffer = self.null_bit_buffer();

        let mut child_data: Vec<ArrayData> = (0..self.array().n_children as usize)
            .map(|i| {
                let child = self.child(i);
                child.to_data()
            })
            .map(|d| d.unwrap())
            .collect();

        if let Some(d) = self.dictionary() {
            // For dictionary type there should only be a single child, so we don't need to worry if
            // there are other children added above.
            assert!(child_data.is_empty());
            child_data.push(d.to_data()?);
        }

        // Should FFI be checking validity?
        Ok(unsafe {
            ArrayData::new_unchecked(
                data_type,
                len,
                Some(null_count),
                null_bit_buffer,
                offset,
                buffers,
                child_data,
            )
        })
    }

    /// returns all buffers, as organized by Rust (i.e. null buffer is skipped)
    fn buffers(&self) -> Result<Vec<Buffer>> {
        (0..self.array().n_buffers - 1)
            .map(|index| {
                // + 1: skip null buffer
                let index = (index + 1) as usize;

                let len = self.buffer_len(index)?;

                unsafe { create_buffer(self.owner().clone(), self.array(), index, len) }
                    .ok_or_else(|| {
                        ArrowError::CDataInterface(format!(
                            "The external buffer at position {} is null.",
                            index - 1
                        ))
                    })
            })
            .collect()
    }

    /// Returns the length, in bytes, of the buffer `i` (indexed according to the C data interface)
    // Rust implementation uses fixed-sized buffers, which require knowledge of their `len`.
    // for variable-sized buffers, such as the second buffer of a stringArray, we need
    // to fetch offset buffer's len to build the second buffer.
    fn buffer_len(&self, i: usize) -> Result<usize> {
        // Special handling for dictionary type as we only care about the key type in the case.
        let t = self.data_type()?;
        let data_type = match &t {
            DataType::Dictionary(key_data_type, _) => key_data_type.as_ref(),
            dt => dt,
        };

        // Inner type is not important for buffer length.
        Ok(match (&data_type, i) {
            (DataType::Utf8, 1)
            | (DataType::LargeUtf8, 1)
            | (DataType::Binary, 1)
            | (DataType::LargeBinary, 1)
            | (DataType::List(_), 1)
            | (DataType::LargeList(_), 1) => {
                // the len of the offset buffer (buffer 1) equals length + 1
                let bits = bit_width(data_type, i)?;
                debug_assert_eq!(bits % 8, 0);
                (self.array().length as usize + 1) * (bits / 8)
            }
            (DataType::Utf8, 2) | (DataType::Binary, 2) | (DataType::List(_), 2) => {
                // the len of the data buffer (buffer 2) equals the last value of the offset buffer (buffer 1)
                let len = self.buffer_len(1)?;
                // first buffer is the null buffer => add(1)
                // we assume that pointer is aligned for `i32`, as Utf8 uses `i32` offsets.
                #[allow(clippy::cast_ptr_alignment)]
                let offset_buffer = unsafe {
                    *(self.array().buffers as *mut *const u8).add(1) as *const i32
                };
                // get last offset
                (unsafe { *offset_buffer.add(len / size_of::<i32>() - 1) }) as usize
            }
            (DataType::LargeUtf8, 2)
            | (DataType::LargeBinary, 2)
            | (DataType::LargeList(_), 2) => {
                // the len of the data buffer (buffer 2) equals the last value of the offset buffer (buffer 1)
                let len = self.buffer_len(1)?;
                // first buffer is the null buffer => add(1)
                // we assume that pointer is aligned for `i64`, as Large uses `i64` offsets.
                #[allow(clippy::cast_ptr_alignment)]
                let offset_buffer = unsafe {
                    *(self.array().buffers as *mut *const u8).add(1) as *const i64
                };
                // get last offset
                (unsafe { *offset_buffer.add(len / size_of::<i64>() - 1) }) as usize
            }
            // buffer len of primitive types
            _ => {
                let bits = bit_width(data_type, i)?;
                bit_util::ceil(self.array().length as usize * bits, 8)
            }
        })
    }

    /// returns the null bit buffer.
    /// Rust implementation uses a buffer that is not part of the array of buffers.
    /// The C Data interface's null buffer is part of the array of buffers.
    fn null_bit_buffer(&self) -> Option<Buffer> {
        // similar to `self.buffer_len(0)`, but without `Result`.
        let buffer_len = bit_util::ceil(self.array().length as usize, 8);

        unsafe { create_buffer(self.owner().clone(), self.array(), 0, buffer_len) }
    }

    fn child(&self, index: usize) -> ArrowArrayChild {
        create_child(self.owner().clone(), self.array(), self.schema(), index)
    }

    fn owner(&self) -> &Arc<FFI_ArrowArray>;
    fn array(&self) -> &FFI_ArrowArray;
    fn schema(&self) -> &FFI_ArrowSchema;
    fn data_type(&self) -> Result<DataType>;
    fn dictionary(&self) -> Option<ArrowArrayChild> {
        unsafe {
            assert!(!(self.array().dictionary.is_null() ^ self.schema().dictionary.is_null()),
                    "Dictionary should both be set or not set in FFI_ArrowArray and FFI_ArrowSchema");
            if !self.array().dictionary.is_null() {
                Some(ArrowArrayChild::from_raw(
                    &*self.array().dictionary,
                    &*self.schema().dictionary,
                    self.owner().clone(),
                ))
            } else {
                None
            }
        }
    }
}

#[allow(rustdoc::private_intra_doc_links)]
/// Struct used to move an Array from and to the C Data Interface.
/// Its main responsibility is to expose functionality that requires
/// both [FFI_ArrowArray] and [FFI_ArrowSchema].
///
/// This struct has two main paths:
///
/// ## Import from the C Data Interface
/// * [ArrowArray::empty] to allocate memory to be filled by an external call
/// * [ArrowArray::try_from_raw] to consume two non-null allocated pointers
/// ## Export to the C Data Interface
/// * [ArrowArray::try_new] to create a new [ArrowArray] from Rust-specific information
/// * [ArrowArray::into_raw] to expose two pointers for [FFI_ArrowArray] and [FFI_ArrowSchema].
///
/// # Safety
/// Whoever creates this struct is responsible for releasing their resources. Specifically,
/// consumers *must* call [ArrowArray::into_raw] and take ownership of the individual pointers,
/// calling [FFI_ArrowArray::release] and [FFI_ArrowSchema::release] accordingly.
///
/// Furthermore, this struct assumes that the incoming data agrees with the C data interface.
#[derive(Debug)]
pub struct ArrowArray {
    pub(crate) array: Arc<FFI_ArrowArray>,
    pub(crate) schema: Arc<FFI_ArrowSchema>,
}

#[derive(Debug)]
pub struct ArrowArrayChild<'a> {
    array: &'a FFI_ArrowArray,
    schema: &'a FFI_ArrowSchema,
    owner: Arc<FFI_ArrowArray>,
}

impl ArrowArrayRef for ArrowArray {
    /// the data_type as declared in the schema
    fn data_type(&self) -> Result<DataType> {
        DataType::try_from(self.schema.as_ref())
    }

    fn array(&self) -> &FFI_ArrowArray {
        self.array.as_ref()
    }

    fn schema(&self) -> &FFI_ArrowSchema {
        self.schema.as_ref()
    }

    fn owner(&self) -> &Arc<FFI_ArrowArray> {
        &self.array
    }
}

impl<'a> ArrowArrayRef for ArrowArrayChild<'a> {
    /// the data_type as declared in the schema
    fn data_type(&self) -> Result<DataType> {
        DataType::try_from(self.schema)
    }

    fn array(&self) -> &FFI_ArrowArray {
        self.array
    }

    fn schema(&self) -> &FFI_ArrowSchema {
        self.schema
    }

    fn owner(&self) -> &Arc<FFI_ArrowArray> {
        &self.owner
    }
}

impl ArrowArray {
    /// creates a new `ArrowArray`. This is used to export to the C Data Interface.
    /// # Safety
    /// See safety of [ArrowArray]
    #[allow(clippy::too_many_arguments)]
    pub unsafe fn try_new(data: ArrayData) -> Result<Self> {
        let array = Arc::new(FFI_ArrowArray::new(&data));
        let schema = Arc::new(FFI_ArrowSchema::try_from(data.data_type())?);
        Ok(ArrowArray { array, schema })
    }

    /// creates a new [ArrowArray] from two pointers. Used to import from the C Data Interface.
    /// # Safety
    /// See safety of [ArrowArray]
    /// Note that this function will copy the content pointed by the raw pointers. Considering
    /// the raw pointers can be from `Arc::into_raw` or other raw pointers, users must be responsible
    /// on managing the allocation of the structs by themselves.
    /// # Error
    /// Errors if any of the pointers is null
    pub unsafe fn try_from_raw(
        array: *const FFI_ArrowArray,
        schema: *const FFI_ArrowSchema,
    ) -> Result<Self> {
        if array.is_null() || schema.is_null() {
            return Err(ArrowError::MemoryError(
                "At least one of the pointers passed to `try_from_raw` is null"
                    .to_string(),
            ));
        };

        let array_mut = array as *mut FFI_ArrowArray;
        let schema_mut = schema as *mut FFI_ArrowSchema;

        let array_data = std::ptr::replace(array_mut, FFI_ArrowArray::empty());
        let schema_data = std::ptr::replace(schema_mut, FFI_ArrowSchema::empty());

        Ok(Self {
            array: Arc::new(array_data),
            schema: Arc::new(schema_data),
        })
    }

    /// creates a new empty [ArrowArray]. Used to import from the C Data Interface.
    /// # Safety
    /// See safety of [ArrowArray]
    pub unsafe fn empty() -> Self {
        let schema = Arc::new(FFI_ArrowSchema::empty());
        let array = Arc::new(FFI_ArrowArray::empty());
        ArrowArray { array, schema }
    }

    /// exports [ArrowArray] to the C Data Interface
    pub fn into_raw(this: ArrowArray) -> (*const FFI_ArrowArray, *const FFI_ArrowSchema) {
        (Arc::into_raw(this.array), Arc::into_raw(this.schema))
    }
}

impl<'a> ArrowArrayChild<'a> {
    fn from_raw(
        array: &'a FFI_ArrowArray,
        schema: &'a FFI_ArrowSchema,
        owner: Arc<FFI_ArrowArray>,
    ) -> Self {
        Self {
            array,
            schema,
            owner,
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::array::{
        export_array_into_raw, make_array, Array, ArrayData, BooleanArray, DecimalArray,
        DictionaryArray, DurationSecondArray, FixedSizeBinaryArray, FixedSizeListArray,
        GenericBinaryArray, GenericListArray, GenericStringArray, Int32Array, NullArray,
        OffsetSizeTrait, Time32MillisecondArray, TimestampMillisecondArray,
    };
    use crate::compute::kernels;
    use crate::datatypes::{Field, Int8Type};
    use std::convert::TryFrom;

    #[test]
    fn test_round_trip() -> Result<()> {
        // create an array natively
        let array = Int32Array::from(vec![1, 2, 3]);

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = array.as_any().downcast_ref::<Int32Array>().unwrap();
        let array = kernels::arithmetic::add(array, array).unwrap();

        // verify
        assert_eq!(array, Int32Array::from(vec![2, 4, 6]));

        // (drop/release)
        Ok(())
    }

    #[test]
    #[cfg(not(feature = "force_validate"))]
    fn test_decimal_round_trip() -> Result<()> {
        // create an array natively
        let original_array = [Some(12345_i128), Some(-12345_i128), None]
            .into_iter()
            .collect::<DecimalArray>()
            .with_precision_and_scale(6, 2)
            .unwrap();

        // export it
        let array = ArrowArray::try_from(original_array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = array.as_any().downcast_ref::<DecimalArray>().unwrap();

        // verify
        assert_eq!(array, &original_array);

        // (drop/release)
        Ok(())
    }
    // case with nulls is tested in the docs, through the example on this module.

    fn test_generic_string<Offset: OffsetSizeTrait>() -> Result<()> {
        // create an array natively
        let array =
            GenericStringArray::<Offset>::from(vec![Some("a"), None, Some("aaa")]);

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = kernels::concat::concat(&[array.as_ref(), array.as_ref()]).unwrap();
        let array = array
            .as_any()
            .downcast_ref::<GenericStringArray<Offset>>()
            .unwrap();

        // verify
        let expected = GenericStringArray::<Offset>::from(vec![
            Some("a"),
            None,
            Some("aaa"),
            Some("a"),
            None,
            Some("aaa"),
        ]);
        assert_eq!(array, &expected);

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_string() -> Result<()> {
        test_generic_string::<i32>()
    }

    #[test]
    fn test_large_string() -> Result<()> {
        test_generic_string::<i64>()
    }

    fn test_generic_list<Offset: OffsetSizeTrait>() -> Result<()> {
        // Construct a value array
        let value_data = ArrayData::builder(DataType::Int32)
            .len(8)
            .add_buffer(Buffer::from_slice_ref(&[0, 1, 2, 3, 4, 5, 6, 7]))
            .build()
            .unwrap();

        // Construct a buffer for value offsets, for the nested array:
        //  [[0, 1, 2], [3, 4, 5], [6, 7]]
        let value_offsets = [0_usize, 3, 6, 8]
            .iter()
            .map(|i| Offset::from_usize(*i).unwrap())
            .collect::<Buffer>();

        // Construct a list array from the above two
        let list_data_type = match std::mem::size_of::<Offset>() {
            4 => DataType::List(Box::new(Field::new("item", DataType::Int32, false))),
            _ => {
                DataType::LargeList(Box::new(Field::new("item", DataType::Int32, false)))
            }
        };

        let list_data = ArrayData::builder(list_data_type)
            .len(3)
            .add_buffer(value_offsets)
            .add_child_data(value_data)
            .build()
            .unwrap();

        // create an array natively
        let array = GenericListArray::<Offset>::from(list_data.clone());

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // downcast
        let array = array
            .as_any()
            .downcast_ref::<GenericListArray<Offset>>()
            .unwrap();

        dbg!(&array);

        // verify
        let expected = GenericListArray::<Offset>::from(list_data);
        assert_eq!(&array.value(0), &expected.value(0));
        assert_eq!(&array.value(1), &expected.value(1));
        assert_eq!(&array.value(2), &expected.value(2));

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_list() -> Result<()> {
        test_generic_list::<i32>()
    }

    #[test]
    fn test_large_list() -> Result<()> {
        test_generic_list::<i64>()
    }

    fn test_generic_binary<Offset: OffsetSizeTrait>() -> Result<()> {
        // create an array natively
        let array: Vec<Option<&[u8]>> = vec![Some(b"a"), None, Some(b"aaa")];
        let array = GenericBinaryArray::<Offset>::from(array);

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = kernels::concat::concat(&[array.as_ref(), array.as_ref()]).unwrap();
        let array = array
            .as_any()
            .downcast_ref::<GenericBinaryArray<Offset>>()
            .unwrap();

        // verify
        let expected: Vec<Option<&[u8]>> = vec![
            Some(b"a"),
            None,
            Some(b"aaa"),
            Some(b"a"),
            None,
            Some(b"aaa"),
        ];
        let expected = GenericBinaryArray::<Offset>::from(expected);
        assert_eq!(array, &expected);

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_binary() -> Result<()> {
        test_generic_binary::<i32>()
    }

    #[test]
    fn test_large_binary() -> Result<()> {
        test_generic_binary::<i64>()
    }

    #[test]
    fn test_bool() -> Result<()> {
        // create an array natively
        let array = BooleanArray::from(vec![None, Some(true), Some(false)]);

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = array.as_any().downcast_ref::<BooleanArray>().unwrap();
        let array = kernels::boolean::not(array)?;

        // verify
        assert_eq!(
            array,
            BooleanArray::from(vec![None, Some(false), Some(true)])
        );

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_time32() -> Result<()> {
        // create an array natively
        let array = Time32MillisecondArray::from(vec![None, Some(1), Some(2)]);

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = kernels::concat::concat(&[array.as_ref(), array.as_ref()]).unwrap();
        let array = array
            .as_any()
            .downcast_ref::<Time32MillisecondArray>()
            .unwrap();

        // verify
        assert_eq!(
            array,
            &Time32MillisecondArray::from(vec![
                None,
                Some(1),
                Some(2),
                None,
                Some(1),
                Some(2)
            ])
        );

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_timestamp() -> Result<()> {
        // create an array natively
        let array = TimestampMillisecondArray::from(vec![None, Some(1), Some(2)]);

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = kernels::concat::concat(&[array.as_ref(), array.as_ref()]).unwrap();
        let array = array
            .as_any()
            .downcast_ref::<TimestampMillisecondArray>()
            .unwrap();

        // verify
        assert_eq!(
            array,
            &TimestampMillisecondArray::from(vec![
                None,
                Some(1),
                Some(2),
                None,
                Some(1),
                Some(2)
            ])
        );

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_fixed_size_binary_array() -> Result<()> {
        let values = vec![
            None,
            Some(vec![10, 10, 10]),
            None,
            Some(vec![20, 20, 20]),
            Some(vec![30, 30, 30]),
            None,
        ];
        let array = FixedSizeBinaryArray::try_from_sparse_iter(values.into_iter())?;

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = kernels::concat::concat(&[array.as_ref(), array.as_ref()]).unwrap();
        let array = array
            .as_any()
            .downcast_ref::<FixedSizeBinaryArray>()
            .unwrap();

        // verify
        assert_eq!(
            array,
            &FixedSizeBinaryArray::try_from_sparse_iter(
                vec![
                    None,
                    Some(vec![10, 10, 10]),
                    None,
                    Some(vec![20, 20, 20]),
                    Some(vec![30, 30, 30]),
                    None,
                    None,
                    Some(vec![10, 10, 10]),
                    None,
                    Some(vec![20, 20, 20]),
                    Some(vec![30, 30, 30]),
                    None,
                ]
                .into_iter()
            )?
        );

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_fixed_size_list_array() -> Result<()> {
        // 0000 0100
        let mut validity_bits: [u8; 1] = [0; 1];
        bit_util::set_bit(&mut validity_bits, 2);

        let v: Vec<i32> = (0..9).into_iter().collect();
        let value_data = ArrayData::builder(DataType::Int32)
            .len(9)
            .add_buffer(Buffer::from_slice_ref(&v))
            .build()?;

        let list_data_type =
            DataType::FixedSizeList(Box::new(Field::new("f", DataType::Int32, false)), 3);
        let list_data = ArrayData::builder(list_data_type.clone())
            .len(3)
            .null_bit_buffer(Some(Buffer::from(validity_bits)))
            .add_child_data(value_data)
            .build()?;

        // export it
        let array = ArrowArray::try_from(list_data)?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = kernels::concat::concat(&[array.as_ref(), array.as_ref()]).unwrap();
        let array = array.as_any().downcast_ref::<FixedSizeListArray>().unwrap();

        // 0010 0100
        let mut expected_validity_bits: [u8; 1] = [0; 1];
        bit_util::set_bit(&mut expected_validity_bits, 2);
        bit_util::set_bit(&mut expected_validity_bits, 5);

        let mut w = vec![];
        w.extend_from_slice(&v);
        w.extend_from_slice(&v);

        let expected_value_data = ArrayData::builder(DataType::Int32)
            .len(18)
            .add_buffer(Buffer::from_slice_ref(&w))
            .build()?;

        let expected_list_data = ArrayData::builder(list_data_type)
            .len(6)
            .null_bit_buffer(Some(Buffer::from(expected_validity_bits)))
            .add_child_data(expected_value_data)
            .build()?;
        let expected_array = FixedSizeListArray::from(expected_list_data);

        // verify
        assert_eq!(array, &expected_array);

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_dictionary() -> Result<()> {
        // create an array natively
        let values = vec!["a", "aaa", "aaa"];
        let dict_array: DictionaryArray<Int8Type> = values.into_iter().collect();

        // export it
        let array = ArrowArray::try_from(dict_array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = kernels::concat::concat(&[array.as_ref(), array.as_ref()]).unwrap();
        let actual = array
            .as_any()
            .downcast_ref::<DictionaryArray<Int8Type>>()
            .unwrap();

        // verify
        let new_values = vec!["a", "aaa", "aaa", "a", "aaa", "aaa"];
        let expected: DictionaryArray<Int8Type> = new_values.into_iter().collect();
        assert_eq!(actual, &expected);

        // (drop/release)
        Ok(())
    }

    #[test]
    fn test_export_array_into_raw() -> Result<()> {
        let array = make_array(Int32Array::from(vec![1, 2, 3]).data().clone());

        // Assume two raw pointers provided by the consumer
        let out_array = Box::new(FFI_ArrowArray::empty());
        let out_schema = Box::new(FFI_ArrowSchema::empty());
        let out_array_ptr = Box::into_raw(out_array);
        let out_schema_ptr = Box::into_raw(out_schema);

        unsafe {
            export_array_into_raw(array, out_array_ptr, out_schema_ptr)?;
        }

        // (simulate consumer) import it
        unsafe {
            let array = ArrowArray::try_from_raw(out_array_ptr, out_schema_ptr).unwrap();
            let data = ArrayData::try_from(array)?;
            let array = make_array(data);

            // perform some operation
            let array = array.as_any().downcast_ref::<Int32Array>().unwrap();
            let array = kernels::arithmetic::add(array, array).unwrap();

            // verify
            assert_eq!(array, Int32Array::from(vec![2, 4, 6]));

            Box::from_raw(out_array_ptr);
            Box::from_raw(out_schema_ptr);
        }
        Ok(())
    }

    #[test]
    fn test_duration() -> Result<()> {
        // create an array natively
        let array = DurationSecondArray::from(vec![None, Some(1), Some(2)]);

        // export it
        let array = ArrowArray::try_from(array.data().clone())?;

        // (simulate consumer) import it
        let data = ArrayData::try_from(array)?;
        let array = make_array(data);

        // perform some operation
        let array = kernels::concat::concat(&[array.as_ref(), array.as_ref()]).unwrap();
        let array = array
            .as_any()
            .downcast_ref::<DurationSecondArray>()
            .unwrap();

        // verify
        assert_eq!(
            array,
            &DurationSecondArray::from(vec![
                None,
                Some(1),
                Some(2),
                None,
                Some(1),
                Some(2)
            ])
        );

        // (drop/release)
        Ok(())
    }

    #[test]
    fn null_array_n_buffers() -> Result<()> {
        let array = NullArray::new(10);
        let data = array.data();

        let ffi_array = FFI_ArrowArray::new(data);
        assert_eq!(0, ffi_array.n_buffers);

        let private_data =
            unsafe { Box::from_raw(ffi_array.private_data as *mut ArrayPrivateData) };

        assert_eq!(0, private_data.buffers_ptr.len());

        Box::into_raw(private_data);

        Ok(())
    }
}