odbc_api/

cursor.rs

mod block_cursor;
mod concurrent_block_cursor;
mod polling_cursor;

use log::warn;
use odbc_sys::HStmt;

use crate::{
    Error, ResultSetMetadata,
    buffers::Indicator,
    error::ExtendResult,
    handles::{
        AsStatementRef, CDataMut, DiagnosticStream, SqlResult, State, Statement, StatementRef,
    },
    parameter::{Binary, CElement, Text, VarCell, VarKind, WideText},
};

use std::{
    mem::{MaybeUninit, size_of},
    ptr,
    thread::panicking,
};

pub use self::{
    block_cursor::{BlockCursor, BlockCursorIterator},
    concurrent_block_cursor::ConcurrentBlockCursor,
    polling_cursor::{BlockCursorPolling, CursorPolling},
};

/// Cursors are used to process and iterate the result sets returned by executing queries.
///
/// # Example: Fetching result in batches
///
/// ```rust
/// use odbc_api::{Cursor, buffers::{BufferDesc, ColumnarAnyBuffer}, Error};
///
/// /// Fetches all values from the first column of the cursor as i32 in batches of 100 and stores
/// /// them in a vector.
/// fn fetch_all_ints(cursor: impl Cursor) -> Result<Vec<i32>, Error> {
///     let mut all_ints = Vec::new();
///     // Batch size determines how many values we fetch at once.
///     let batch_size = 100;
///     // We expect the first column to hold INTEGERs (or a type convertible to INTEGER). Use
///     // the metadata on the result set, if you want to investige the types of the columns at
///     // runtime.
///     let description = BufferDesc::I32 { nullable: false };
///     // This is the buffer we bind to the driver, and repeatedly use to fetch each batch
///     let buffer = ColumnarAnyBuffer::from_descs(batch_size, [description]);
///     // Bind buffer to cursor
///     let mut row_set_buffer = cursor.bind_buffer(buffer)?;
///     // Fetch data batch by batch
///     while let Some(batch) = row_set_buffer.fetch()? {
///         all_ints.extend_from_slice(batch.column(0).as_slice().unwrap())
///     }
///     Ok(all_ints)
/// }
/// ```
pub trait Cursor: ResultSetMetadata {
    /// Advances the cursor to the next row in the result set. This is **Slow**. Bind
    /// [`crate::buffers`] instead, for good performance.
    ///
    /// ⚠ While this method is very convenient due to the fact that the application does not have
    /// to declare and bind specific buffers, it is also in many situations extremely slow. Concrete
    /// performance depends on the ODBC driver in question, but it is likely it performs a roundtrip
    /// to the datasource for each individual row. It is also likely an extra conversion is
    /// performed then requesting individual fields, since the C buffer type is not known to the
    /// driver in advance. Consider binding a buffer to the cursor first using
    /// [`Self::bind_buffer`].
    ///
    /// That being said, it is a convenient programming model, as the developer does not need to
    /// prepare and allocate the buffers beforehand. It is also a good way to retrieve really large
    /// single values out of a data source (like one large text file). See [`CursorRow::get_text`].
    fn next_row(&mut self) -> Result<Option<CursorRow<'_>>, Error> {
        let row_available = unsafe {
            self.as_stmt_ref()
                .fetch()
                .into_result_bool(&self.as_stmt_ref())?
        };
        let ret = if row_available {
            Some(unsafe { CursorRow::new(self.as_stmt_ref()) })
        } else {
            None
        };
        Ok(ret)
    }

    /// Binds this cursor to a buffer holding a row set.
    fn bind_buffer<B>(self, row_set_buffer: B) -> Result<BlockCursor<Self, B>, Error>
    where
        Self: Sized,
        B: RowSetBuffer;

    /// For some datasources it is possible to create more than one result set at once via a call to
    /// execute. E.g. by calling a stored procedure or executing multiple SQL statements at once.
    /// This method consumes the current cursor and creates a new one representing the next result
    /// set should it exist.
    fn more_results(self) -> Result<Option<Self>, Error>
    where
        Self: Sized;
}

/// An individual row of an result set. See [`crate::Cursor::next_row`].
pub struct CursorRow<'s> {
    statement: StatementRef<'s>,
}

impl<'s> CursorRow<'s> {
    /// # Safety
    ///
    /// `statement` must be in a cursor state.
    unsafe fn new(statement: StatementRef<'s>) -> Self {
        CursorRow { statement }
    }
}

impl CursorRow<'_> {
    /// Fills a suitable target buffer with a field from the current row of the result set. This
    /// method drains the data from the field. It can be called repeatedly to if not all the data
    /// fit in the output buffer at once. It should not called repeatedly to fetch the same value
    /// twice. Column index starts at `1`.
    ///
    /// You can use [`crate::Nullable`] to fetch nullable values.
    ///
    /// # Example
    ///
    /// ```
    /// # use odbc_api::{Cursor, Error, Nullable};
    /// # fn fetch_values_example(cursor: &mut impl Cursor) -> Result<(), Error> {
    /// // Declare nullable value to fetch value into. ODBC values layout is different from Rusts
    /// // option. We can not use `Option<i32>` directly.
    /// let mut field = Nullable::<i32>::null();
    /// // Move cursor to next row
    /// let mut row = cursor.next_row()?.unwrap();
    /// // Fetch first column into field
    /// row.get_data(1, &mut field)?;
    /// // Convert nullable value to Option for convinience
    /// let field = field.into_opt();
    /// if let Some(value) = field {
    ///     println!("Value: {}", value);
    /// } else {
    ///     println!("Value is NULL");
    /// }
    /// # Ok(())
    /// # }
    /// ```
    pub fn get_data(
        &mut self,
        col_or_param_num: u16,
        target: &mut (impl CElement + CDataMut),
    ) -> Result<(), Error> {
        self.statement
            .get_data(col_or_param_num, target)
            .into_result(&self.statement)
            .provide_context_for_diagnostic(|record, function| {
                if record.state == State::INDICATOR_VARIABLE_REQUIRED_BUT_NOT_SUPPLIED {
                    Error::UnableToRepresentNull(record)
                } else {
                    Error::Diagnostics { record, function }
                }
            })
    }

    /// Retrieves arbitrary large character data from the row and stores it in the buffer. Column
    /// index starts at `1`. The used encoding is accordig to the ODBC standard determined by your
    /// system local. Ultimatly the choice is up to the implementation of your ODBC driver, which
    /// often defaults to always UTF-8.
    ///
    /// # Example
    ///
    /// Retrieve an arbitrary large text file from a database field.
    ///
    /// ```
    /// use odbc_api::{Connection, Error, IntoParameter, Cursor};
    ///
    /// fn get_large_text(name: &str, conn: &mut Connection<'_>) -> Result<Option<String>, Error> {
    ///     let query = "SELECT content FROM LargeFiles WHERE name=?";
    ///     let parameters = &name.into_parameter();
    ///     let timeout_sec = None;
    ///     let mut cursor = conn
    ///         .execute(query, parameters, timeout_sec)?
    ///         .expect("Assume select statement creates cursor");
    ///     if let Some(mut row) = cursor.next_row()? {
    ///         let mut buf = Vec::new();
    ///         row.get_text(1, &mut buf)?;
    ///         let ret = String::from_utf8(buf).unwrap();
    ///         Ok(Some(ret))
    ///     } else {
    ///         Ok(None)
    ///     }
    /// }
    /// ```
    ///
    /// # Return
    ///
    /// `true` indicates that the value has not been `NULL` and the value has been placed in `buf`.
    /// `false` indicates that the value is `NULL`. The buffer is cleared in that case.
    pub fn get_text(&mut self, col_or_param_num: u16, buf: &mut Vec<u8>) -> Result<bool, Error> {
        self.get_variadic::<Text>(col_or_param_num, buf)
    }

    /// Retrieves arbitrary large character data from the row and stores it in the buffer. Column
    /// index starts at `1`. The used encoding is UTF-16.
    ///
    /// # Return
    ///
    /// `true` indicates that the value has not been `NULL` and the value has been placed in `buf`.
    /// `false` indicates that the value is `NULL`. The buffer is cleared in that case.
    pub fn get_wide_text(
        &mut self,
        col_or_param_num: u16,
        buf: &mut Vec<u16>,
    ) -> Result<bool, Error> {
        self.get_variadic::<WideText>(col_or_param_num, buf)
    }

    /// Retrieves arbitrary large binary data from the row and stores it in the buffer. Column index
    /// starts at `1`.
    ///
    /// # Return
    ///
    /// `true` indicates that the value has not been `NULL` and the value has been placed in `buf`.
    /// `false` indicates that the value is `NULL`. The buffer is cleared in that case.
    pub fn get_binary(&mut self, col_or_param_num: u16, buf: &mut Vec<u8>) -> Result<bool, Error> {
        self.get_variadic::<Binary>(col_or_param_num, buf)
    }

    fn get_variadic<K: VarKind>(
        &mut self,
        col_or_param_num: u16,
        buf: &mut Vec<K::Element>,
    ) -> Result<bool, Error> {
        if buf.capacity() == 0 {
            // User did just provide an empty buffer. So it is fair to assume not much domain
            // knowledge has been used to decide its size. We just default to 256 to increase the
            // chance that we get it done with one alloctaion. The buffer size being 0 we need at
            // least 1 anyway. If the capacity is not `0` we'll leave the buffer size untouched as
            // we do not want to prevent users from providing better guessen based on domain
            // knowledge.
            // This also implicitly makes sure that we can at least hold one terminating zero.
            buf.reserve(256);
        }
        // Utilize all of the allocated buffer.
        buf.resize(buf.capacity(), K::ZERO);

        // Did we learn how much capacity we need in the last iteration? We use this only to panic
        // on erroneous implementations of get_data and avoid endless looping until we run out of
        // memory.
        let mut remaining_length_known = false;
        // We repeatedly fetch data and add it to the buffer. The buffer length is therefore the
        // accumulated value size. The target always points to the last window in buf which is going
        // to contain the **next** part of the data, whereas buf contains the entire accumulated
        // value so far.
        let mut target =
            VarCell::<&mut [K::Element], K>::from_buffer(buf.as_mut_slice(), Indicator::NoTotal);
        self.get_data(col_or_param_num, &mut target)?;
        while !target.is_complete() {
            // Amount of payload bytes (excluding terminating zeros) fetched with the last call to
            // get_data.
            let fetched = target
                .len_in_bytes()
                .expect("ODBC driver must always report how many bytes were fetched.");
            match target.indicator() {
                // If Null the value would be complete
                Indicator::Null => unreachable!(),
                // We do not know how large the value is. Let's fetch the data with repeated calls
                // to get_data.
                Indicator::NoTotal => {
                    let old_len = buf.len();
                    // Use an exponential strategy for increasing buffer size.
                    buf.resize(old_len * 2, K::ZERO);
                    let buf_extend = &mut buf[(old_len - K::TERMINATING_ZEROES)..];
                    target = VarCell::<&mut [K::Element], K>::from_buffer(
                        buf_extend,
                        Indicator::NoTotal,
                    );
                }
                // We did not get all of the value in one go, but the data source has been friendly
                // enough to tell us how much is missing.
                Indicator::Length(len) => {
                    if remaining_length_known {
                        panic!(
                            "SQLGetData has been unable to fetch all data, even though the \
                            capacity of the target buffer has been adapted to hold the entire \
                            payload based on the indicator of the last part. You may consider \
                            filing a bug with the ODBC driver you are using."
                        )
                    }
                    remaining_length_known = true;
                    // Amount of bytes missing from the value using get_data, excluding terminating
                    // zero.
                    let still_missing_in_bytes = len - fetched;
                    let still_missing = still_missing_in_bytes / size_of::<K::Element>();
                    let old_len = buf.len();
                    buf.resize(old_len + still_missing, K::ZERO);
                    let buf_extend = &mut buf[(old_len - K::TERMINATING_ZEROES)..];
                    target = VarCell::<&mut [K::Element], K>::from_buffer(
                        buf_extend,
                        Indicator::NoTotal,
                    );
                }
            }
            // Fetch binary data into buffer.
            self.get_data(col_or_param_num, &mut target)?;
        }
        // We did get the complete value, including the terminating zero. Let's resize the buffer to
        // match the retrieved value exactly (excluding terminating zero).
        if let Some(len_in_bytes) = target.indicator().length() {
            // Since the indicator refers to value length without terminating zero, and capacity is
            // including the terminating zero this also implicitly drops the terminating zero at the
            // end of the buffer.
            let shrink_by_bytes = target.capacity_in_bytes() - len_in_bytes;
            let shrink_by_chars = shrink_by_bytes / size_of::<K::Element>();
            buf.resize(buf.len() - shrink_by_chars, K::ZERO);
            Ok(true)
        } else {
            // value is NULL
            buf.clear();
            Ok(false)
        }
    }
}

/// Cursors are used to process and iterate the result sets returned by executing queries. Created
/// by either a prepared query or direct execution. Usually utilized through the [`crate::Cursor`]
/// trait.
#[derive(Debug)]
pub struct CursorImpl<Stmt: AsStatementRef> {
    /// A statement handle in cursor mode.
    statement: Stmt,
}

impl<S> Drop for CursorImpl<S>
where
    S: AsStatementRef,
{
    fn drop(&mut self) {
        let mut stmt = self.statement.as_stmt_ref();
        if let Err(e) = stmt.close_cursor().into_result(&stmt) {
            // Avoid panicking, if we already have a panic. We don't want to mask the original
            // error.
            if !panicking() {
                panic!("Unexpected error closing cursor: {e:?}")
            }
        }
    }
}

impl<S> AsStatementRef for CursorImpl<S>
where
    S: AsStatementRef,
{
    fn as_stmt_ref(&mut self) -> StatementRef<'_> {
        self.statement.as_stmt_ref()
    }
}

impl<S> ResultSetMetadata for CursorImpl<S> where S: AsStatementRef {}

impl<S> Cursor for CursorImpl<S>
where
    S: AsStatementRef,
{
    fn bind_buffer<B>(mut self, mut row_set_buffer: B) -> Result<BlockCursor<Self, B>, Error>
    where
        B: RowSetBuffer,
    {
        let stmt = self.statement.as_stmt_ref();
        unsafe {
            bind_row_set_buffer_to_statement(stmt, &mut row_set_buffer)?;
        }
        Ok(BlockCursor::new(row_set_buffer, self))
    }

    fn more_results(self) -> Result<Option<Self>, Error>
    where
        Self: Sized,
    {
        // Consume self without calling drop to avoid calling close_cursor.
        let mut statement = self.into_stmt();
        let mut stmt = statement.as_stmt_ref();

        let has_another_result = unsafe { stmt.more_results() }.into_result_bool(&stmt)?;
        let next = if has_another_result {
            Some(CursorImpl { statement })
        } else {
            None
        };
        Ok(next)
    }
}

impl<S> CursorImpl<S>
where
    S: AsStatementRef,
{
    /// Users of this library are encouraged not to call this constructor directly but rather invoke
    /// [`crate::Connection::execute`] or [`crate::Prepared::execute`] to get a cursor and utilize
    /// it using the [`crate::Cursor`] trait. This method is public so users with an understanding
    /// of the raw ODBC C-API have a way to create a cursor, after they left the safety rails of the
    /// Rust type System, in order to implement a use case not covered yet, by the safe abstractions
    /// within this crate.
    ///
    /// # Safety
    ///
    /// `statement` must be in Cursor state, for the invariants of this type to hold.
    pub unsafe fn new(statement: S) -> Self {
        Self { statement }
    }

    /// Deconstructs the `CursorImpl` without calling drop. This is a way to get to the underlying
    /// statement, while preventing a call to close cursor.
    pub fn into_stmt(self) -> S {
        // We want to move `statement` out of self, which would make self partially uninitialized.
        let dont_drop_me = MaybeUninit::new(self);
        let self_ptr = dont_drop_me.as_ptr();

        // Safety: We know `dont_drop_me` is valid at this point so reading the ptr is okay
        unsafe { ptr::read(&(*self_ptr).statement) }
    }

    pub(crate) fn as_sys(&mut self) -> HStmt {
        self.as_stmt_ref().as_sys()
    }
}

/// A Row set buffer binds row, or column wise buffers to a cursor in order to fill them with row
/// sets with each call to fetch.
///
/// # Safety
///
/// Implementers of this trait must ensure that every pointer bound in `bind_to_cursor` stays valid
/// even if an instance is moved in memory. Bound members should therefore be likely references
/// themselves. To bind stack allocated buffers it is recommended to implement this trait on the
/// reference type instead.
pub unsafe trait RowSetBuffer {
    /// Declares the bind type of the Row set buffer. `0` Means a columnar binding is used. Any non
    /// zero number is interpreted as the size of a single row in a row wise binding style.
    fn bind_type(&self) -> usize;

    /// The batch size for bulk cursors, if retrieving many rows at once.
    fn row_array_size(&self) -> usize;

    /// Mutable reference to the number of fetched rows.
    ///
    /// # Safety
    ///
    /// Implementations of this method must take care that the returned referenced stays valid, even
    /// if `self` should be moved.
    fn mut_num_fetch_rows(&mut self) -> &mut usize;

    /// Binds the buffer either column or row wise to the cursor.
    ///
    /// # Safety
    ///
    /// It's the implementation's responsibility to ensure that all bound buffers are valid until
    /// unbound or the statement handle is deleted.
    unsafe fn bind_colmuns_to_cursor(&mut self, cursor: StatementRef<'_>) -> Result<(), Error>;

    /// Find an indicator larger than the maximum element size of the buffer.
    fn find_truncation(&self) -> Option<TruncationInfo>;
}

/// Returned by [`RowSetBuffer::find_truncation`]. Contains information about the truncation found.
#[derive(Clone, Copy, PartialEq, Eq, Debug)]
pub struct TruncationInfo {
    /// Length of the untruncated value if known
    pub indicator: Option<usize>,
    /// Zero based buffer index of the column in which the truncation occurred.
    pub buffer_index: usize,
}

unsafe impl<T: RowSetBuffer> RowSetBuffer for &mut T {
    fn bind_type(&self) -> usize {
        (**self).bind_type()
    }

    fn row_array_size(&self) -> usize {
        (**self).row_array_size()
    }

    fn mut_num_fetch_rows(&mut self) -> &mut usize {
        (*self).mut_num_fetch_rows()
    }

    unsafe fn bind_colmuns_to_cursor(&mut self, cursor: StatementRef<'_>) -> Result<(), Error> {
        unsafe { (*self).bind_colmuns_to_cursor(cursor) }
    }

    fn find_truncation(&self) -> Option<TruncationInfo> {
        (**self).find_truncation()
    }
}

/// Binds a row set buffer to a statment. Implementation is shared between synchronous and
/// asynchronous cursors.
unsafe fn bind_row_set_buffer_to_statement(
    mut stmt: StatementRef<'_>,
    row_set_buffer: &mut impl RowSetBuffer,
) -> Result<(), Error> {
    unsafe {
        stmt.set_row_bind_type(row_set_buffer.bind_type())
            .into_result(&stmt)?;
        let size = row_set_buffer.row_array_size();
        let sql_result = stmt.set_row_array_size(size);

        // Search for "option value changed". A QODBC driver reported "option value changed", yet
        // set the value to `723477590136`. We want to panic if something like this happens.
        //
        // See: <https://github.com/pacman82/odbc-api/discussions/742#discussioncomment-13887516>
        let mut diagnostic_stream = DiagnosticStream::new(&stmt);
        // We just rememeber that we have seen "option value changed", before asking for the array
        // size, in order to not mess with other diagnostic records.
        let mut option_value_changed = false;
        while let Some(record) = diagnostic_stream.next() {
            warn!("{record}");
            if record.state == State::OPTION_VALUE_CHANGED {
                option_value_changed = true;
            }
        }
        if option_value_changed {
            // Now rejecting a too large buffer size is save, but not if the value is something
            // even larger after. Zero is also suspicious.
            let actual_size = stmt.row_array_size().into_result(&stmt)?;
            warn!(
                "Row array size set by the driver to: {actual_size}. Desired size had been: {size}"
            );
            if actual_size > size || actual_size == 0 {
                panic!(
                    "Your ODBC buffer changed the array size for bulk fetchin in an unsound way. \
                    To prevent undefined behavior the application must panic. You can try \
                    different batch sizes for bulk fetching, or report a bug with your ODBC driver \
                    provider. This behavior has been observed with QODBC drivers. If you are using \
                    one try fetching row by row rather than the faster bulk fetch."
                )
            }
        }

        sql_result
            // We already logged diagnostic records then we were looking for Option value changed
            .into_result_without_logging(&stmt)
            // SAP anywhere has been seen to return with an "invalid attribute" error instead of
            // a success with "option value changed" info. Let us map invalid attributes during
            // setting row set array size to something more precise.
            .provide_context_for_diagnostic(|record, function| {
                if record.state == State::INVALID_ATTRIBUTE_VALUE {
                    Error::InvalidRowArraySize { record, size }
                } else {
                    Error::Diagnostics { record, function }
                }
            })?;
        stmt.set_num_rows_fetched(row_set_buffer.mut_num_fetch_rows())
            .into_result(&stmt)?;
        row_set_buffer.bind_colmuns_to_cursor(stmt)?;
        Ok(())
    }
}

/// Error handling for bulk fetching is shared between synchronous and asynchronous usecase.
fn error_handling_for_fetch(
    result: SqlResult<()>,
    mut stmt: StatementRef,
    buffer: &impl RowSetBuffer,
    error_for_truncation: bool,
) -> Result<bool, Error> {
    // Only check for truncation if a) the user indicated that he wants to error instead of just
    // ignoring it and if there is at least one diagnostic record. ODBC standard requires a
    // diagnostic record to be there in case of truncation. Sadly we can not rely on this particular
    // record to be there, as the driver could generate a large amount of diagnostic records,
    // while we are limited in the amount we can check. The second check serves as an optimization
    // for the happy path.
    if error_for_truncation && result == SqlResult::SuccessWithInfo(()) {
        if let Some(TruncationInfo {
            indicator,
            buffer_index,
        }) = buffer.find_truncation()
        {
            return Err(Error::TooLargeValueForBuffer {
                indicator,
                buffer_index,
            });
        }
    }

    let has_row = result
        .on_success(|| true)
        .on_no_data(|| false)
        .into_result(&stmt.as_stmt_ref())
        // Oracle's ODBC driver does not support 64Bit integers. Furthermore, it does not
        // tell it to the user when binding parameters, but rather now then we fetch
        // results. The error code returned is `HY004` rather than `HY003` which should
        // be used to indicate invalid buffer types.
        .provide_context_for_diagnostic(|record, function| {
            if record.state == State::INVALID_SQL_DATA_TYPE {
                Error::OracleOdbcDriverDoesNotSupport64Bit(record)
            } else {
                Error::Diagnostics { record, function }
            }
        })?;
    Ok(has_row)
}

/// Unbinds buffer and num_rows_fetched from the cursor. This implementation is shared between
/// unbind and the drop handler, and the synchronous and asynchronous variant.
fn unbind_buffer_from_cursor(cursor: &mut impl AsStatementRef) -> Result<(), Error> {
    // Now that we have cursor out of block cursor, we need to unbind the buffer.
    let mut stmt = cursor.as_stmt_ref();
    stmt.unbind_cols().into_result(&stmt)?;
    stmt.unset_num_rows_fetched().into_result(&stmt)?;
    Ok(())
}