rt_history/

lib.rs

//! An RT-safe history log with error checking.
//!
//! This is a bounded wait-free thread synchronization primitive which allows
//! you to record the time evolution of some data on one thread and be able to
//! query the last N data points on other threads.
//!
//! By definition of wait-free synchronization, the producer and consumer
//! threads cannot wait for each other, so in bad conditions (too small a buffer
//! size, too low a thread's priority, too loaded a system...), two race
//! conditions can occur:
//!
//! - The producer can be too fast with respect to consumers, and overwrite
//!   historical data which a consumer was still in the process of reading.
//!   This buffer overrun scenario is reported, along with the degree of overrun
//!   that occurred, which can be used to guide system parameter adjustments so
//!   that the error stops occurring.
//! - The producer can be too slow with respect to consumers, and fail to write
//!   a sufficient amount of new data inbetween two consumer readouts. The
//!   precise definition of this buffer underrun error is workload-dependent, so
//!   we provide the right tool to detect it in the form of a latest data point
//!   timestamp, but we do not handle it ourselves.
//!
//! ```
//! # use rt_history::RTHistory;
//! let (mut input, output) = RTHistory::<u8>::new(8).split();
//!
//! let in_buf = [1, 2, 3];
//! input.write(&in_buf[..]);
//!
//! let mut out_buf = [0; 3];
//! assert_eq!(output.read(&mut out_buf[..]), Ok(3));
//! assert_eq!(in_buf, out_buf);
//! ```

#![deny(missing_docs)]

use atomic::{Atomic, Ordering};
use bytemuck::NoUninit;
use std::sync::Arc;
use thiserror::Error;

/// Re-export of the bytemuck version used by this crate
pub use bytemuck;

/// Data that is shared between the producer and the consumer
struct SharedState<T: NoUninit + Sync> {
    /// Circular buffer
    data: Box<[Atomic<T>]>,

    /// data.len() will always be a power of 2, with this exponent, and the
    /// compiler can use this knowledge to compute modulos much faster
    data_len_pow2: u32,

    /// Timestamp of the last entry that has been published
    readable: Atomic<usize>,

    /// Timestamp of the last entry that has been or is being overwritten
    writing: Atomic<usize>,
}
//
impl<T: NoUninit + Sync> SharedState<T> {
    /// Length of the inner circular buffer
    #[inline]
    fn data_len(&self) -> usize {
        let data_len = 1 << self.data_len_pow2;
        debug_assert_eq!(self.data.len(), data_len);
        data_len
    }
}

/// Producer interface to the history log
pub struct Input<T: NoUninit + Sync>(Arc<SharedState<T>>);
//
impl<T: NoUninit + Sync> Input<T> {
    /// Query the history capacity
    ///
    /// Once this number of entries has been recorded, subsequent writes will
    /// overwrite old entries in a FIFO manner.
    ///
    pub fn capacity(&self) -> usize {
        self.0.data_len()
    }

    /// Record new entries into the history log
    ///
    /// `input` must be shorter than the buffer length.
    ///
    pub fn write(&mut self, input: &[T]) {
        // Check that the request makes sense
        let data_len = self.0.data_len();
        assert!(
            input.len() <= data_len,
            "History is shorter than provided input"
        );

        // Notify the consumer that we are writing new data
        let old_writing = self.0.writing.load(Ordering::Relaxed);
        let new_writing = old_writing.wrapping_add(input.len());
        self.0.writing.store(new_writing, Ordering::Relaxed);

        // Make sure that this notification is not reordered after data writes
        atomic::fence(Ordering::Release);

        // Wrap old and new writing timestamps into circular buffer range
        let old_writing_idx = old_writing % data_len;
        let new_writing_idx = new_writing % data_len;

        // Perform the data writes
        let first_output_len = input.len().min(data_len - old_writing_idx);
        let first_output = &self.0.data[old_writing_idx..old_writing_idx + first_output_len];
        let (first_input, second_input) = input.split_at(first_output_len);
        for (src, dst) in first_input.iter().zip(first_output.iter()) {
            // NOTE: This compiles to a memcpy with N-bit granularity. More
            //       performance can be obtained by using a regular memcpy,
            //       which will fully leverage wide hardware instructions like
            //       REP MOVS and SIMD load/stores. But unfortunately, racing
            //       memcpys are also UB in the Rust memory model. So we'd need
            //       to defer to the hardware memory model for this, which means
            //       volatile memcpy, which is nightly-only.
            dst.store(*src, Ordering::Relaxed);
        }
        if first_output_len < input.len() {
            debug_assert!(old_writing_idx >= new_writing_idx);
            debug_assert_eq!(new_writing_idx, second_input.len());
            let second_output = &self.0.data[..new_writing_idx];
            for (src, dst) in second_input.iter().zip(second_output.iter()) {
                // NOTE: See above
                dst.store(*src, Ordering::Relaxed);
            }
        }

        // Notify the consumer that new data has been published, make sure that
        // this write is ordered after previous data writes
        if cfg!(debug_assertions) {
            assert_eq!(self.0.readable.load(Ordering::Relaxed), old_writing);
        }
        self.0.readable.store(new_writing, Ordering::Release);
    }
}

/// Number of entries that the producer wrote so far (with wraparound)
///
/// Differences between two consecutive readouts of this counter can be used to
/// tell how many new entries arrived between two readouts, and detect buffer
/// underruns where the producer wrote too few data. How few is too few is
/// workload-dependent, so we do not implement this logic ourselves.
///
/// This counter may wrap around from time to time, every couple of hours for a
/// mono audio stream on a 32-bit CPU. Use wrapping_sub() for deltas.
///
pub type Clock = usize;

/// Indication that a buffer overrun occured
///
/// This means that the producer wrote over the data that the consumer was in
/// the process of reading. Possible fixes include making the buffer larger,
/// speeding up the consumer, and slowing down the producer, in order of
/// decreasing preference.
///
#[derive(Clone, Copy, Debug, Error, PartialEq, Eq)]
#[error("A buffer overrun occured while reading history of entry {clock}, writer overwrote {excess_entries} entries")]
pub struct Overrun {
    /// Number of entries that were fully written by the producer at the time
    /// where the consumer started reading out data
    pub clock: Clock,

    /// Number of early history entries that were overwritten by the producer
    /// while the consumer was in the process of reading out data.
    ///
    /// This can be higher than the number of entries which the consumer
    /// requested. The intent is to provide a quantitative indication of how
    /// late the consumer is with respect to the producer, so that the buffer
    /// size can be increased by a matching amount if need be.
    ///
    pub excess_entries: Clock,
}

/// Consumer interface to the history log
#[derive(Clone)]
pub struct Output<T: NoUninit + Sync>(Arc<SharedState<T>>);
//
impl<T: NoUninit + Sync> Output<T> {
    /// Query the history capacity
    ///
    /// Once this number of entries has been recorded, subsequent writes will
    /// overwrite old entries in a FIFO manner.
    ///
    pub fn capacity(&self) -> usize {
        self.0.data_len()
    }

    /// Read the last N entries, check timestamp and errors
    ///
    /// `output` must be shorter than the buffer length.
    ///
    /// On successful readout, this method returns the timestamp of the latest
    /// entry, which can be used to check how many samples were written by the
    /// producer since the previous readout.
    ///
    /// If the producer overwrote some of the entries that were read (a scenario
    /// known as a buffer overrun, this error is reported, along with an
    /// indication of how late the consumer is with respect to the producer).
    ///
    pub fn read(&self, output: &mut [T]) -> Result<Clock, Overrun> {
        // Check that the request makes sense
        let data_len = self.0.data_len();
        assert!(
            output.len() <= data_len,
            "History is shorter than requested output"
        );

        // Check the timestamp of the last published data point, and make sure
        // this read is ordered before subsequent data reads
        let last_readable = self.0.readable.load(Ordering::Acquire);

        // Deduce the timestamp of the first data point that we're interested in
        let first_readable = last_readable.wrapping_sub(output.len());

        // Wrap old and new timestamps into circular buffer range
        let last_readable_idx = last_readable % data_len;
        let first_readable_idx = first_readable % data_len;

        // Perform the data reads
        let output_len = output.len();
        let first_input_len = output_len.min(data_len - first_readable_idx);
        let first_input = &self.0.data[first_readable_idx..first_readable_idx + first_input_len];
        let (first_output, second_output) = output.split_at_mut(first_input_len);
        for (src, dst) in first_input.iter().zip(first_output.iter_mut()) {
            // NOTE: See write()
            *dst = src.load(Ordering::Relaxed);
        }
        if first_input_len < output_len {
            debug_assert!(first_readable_idx >= last_readable_idx);
            debug_assert_eq!(last_readable_idx, second_output.len());
            let second_input = &self.0.data[..last_readable_idx];
            for (src, dst) in second_input.iter().zip(second_output.iter_mut()) {
                // NOTE: See write()
                *dst = src.load(Ordering::Relaxed);
            }
        }

        // Make sure the producer did not concurrently overwrite our data.
        // This overrun check must be ordered after the previous data reads.
        atomic::fence(Ordering::Acquire);
        let last_writing = self.0.writing.load(Ordering::Relaxed);
        let excess_entries = last_writing
            .wrapping_sub(first_readable)
            .saturating_sub(data_len);

        // Produce final result
        if excess_entries > 0 {
            Err(Overrun {
                clock: last_readable,
                excess_entries,
            })
        } else {
            Ok(last_readable)
        }
    }
}

/// A realtime-safe (bounded wait-free) history log
pub struct RTHistory<T: NoUninit + Sync>(Arc<SharedState<T>>);
//
impl<T: NoUninit + Default + Sync> RTHistory<T> {
    /// Build a history log that can hold a certain number of entries
    ///
    /// To avoid data corruption (buffer overruns), the log must be
    /// significantly larger than the size of the largest read to be performed
    /// by the consumer. If you are not tight on RAM, a safe starting point is
    /// to make it three times as large.
    ///
    /// For efficiency reason, the actual history buffer size will be rounded to
    /// the next power of two.
    ///
    pub fn new(min_entries: usize) -> Self {
        assert!(
            Atomic::<T>::is_lock_free(),
            "Cannot build a lock-free history log if type T does not have lock-free atomics"
        );
        let data_len = min_entries.next_power_of_two();
        let data_len_pow2 = data_len.trailing_zeros();
        Self(Arc::new(SharedState {
            data: std::iter::repeat_with(Atomic::<T>::default)
                .take(data_len)
                .collect(),
            data_len_pow2,
            readable: Atomic::<usize>::new(0),
            writing: Atomic::<usize>::new(0),
        }))
    }
}
//
impl<T: NoUninit + Sync> RTHistory<T> {
    /// Query the history capacity
    ///
    /// Once this number of entries has been recorded, subsequent writes will
    /// overwrite old entries in a FIFO manner.
    ///
    pub fn capacity(&self) -> usize {
        self.0.data_len()
    }

    /// Split the history log into a producer and consumer interface
    pub fn split(self) -> (Input<T>, Output<T>) {
        (Input(self.0.clone()), Output(self.0))
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use proptest::prelude::*;
    use std::{
        panic::{self, AssertUnwindSafe},
        sync::atomic::AtomicUsize,
    };

    // Number of entries in an history buffer/query
    //
    // Can't go arbitrarily high, otherwise the test will take forever to run or
    // exhaust all available RAM in trying to do so ...
    fn num_entries() -> impl Strategy<Value = usize> {
        0usize..(1024 * 1024)
    }

    proptest! {
        #[test]
        fn new_split_clock(min_entries in num_entries()) {
            // Check initial state
            let history = RTHistory::<f32>::new(min_entries);
            prop_assert!(history
                .0
                .data
                .iter()
                .map(|x| x.load(Ordering::Relaxed))
                .all(|f| f == 0.0));
            prop_assert_eq!(history.0.data_len(), min_entries.next_power_of_two());
            prop_assert_eq!(history.0.data_len(), history.0.data.len());
            prop_assert_eq!(history.0.readable.load(Ordering::Relaxed), 0);
            prop_assert_eq!(history.0.writing.load(Ordering::Relaxed), 0);

            // Split producer and consumer interface
            let (input, output) = history.split();

            // Check that they point to the same thing
            prop_assert_eq!(&*input.0 as *const _, &*output.0 as *const _);
        }

        #[test]
        fn writes_and_read(
            min_entries in num_entries(),
            write1: Vec<u32>,
            write2: Vec<u32>,
            read_size in num_entries(),
        ) {
            // Set up a history log
            let (mut input, output) = RTHistory::<u32>::new(min_entries).split();

            // Attempt a write, which will fail if and only if the input data is
            // larger than the history log's inner buffer.
            macro_rules! checked_write {
                ($write:expr, $input:expr) => {
                    if let Err(_) = panic::catch_unwind(AssertUnwindSafe(|| $input.write(&$write[..])))
                    {
                        prop_assert!($write.len() > $input.0.data_len());
                        return Ok(());
                    }
                    prop_assert!($write.len() <= $input.0.data_len());
                };
            }
            checked_write!(write1, input);

            // Check the final buffer state
            prop_assert!(input
                .0
                .data
                .iter()
                .take(write1.len())
                .zip(&write1)
                .all(|(a, &x)| a.load(Ordering::Relaxed) == x));
            prop_assert!(input
                .0
                .data
                .iter()
                .skip(write1.len())
                .all(|a| a.load(Ordering::Relaxed) == 0));
            prop_assert_eq!(input.0.readable.load(Ordering::Relaxed), write1.len());
            prop_assert_eq!(input.0.writing.load(Ordering::Relaxed), write1.len());

            // Perform another write, check state again
            checked_write!(write2, input);
            let new_readable = write1.len() + write2.len();
            let data_len = input.0.data_len();
            let overwritten = new_readable.saturating_sub(data_len);
            prop_assert!(input
                .0
                .data
                .iter()
                .take(overwritten)
                .zip(write2[write2.len() - overwritten..].iter())
                .all(|(a, &x2)| a.load(Ordering::Relaxed) == x2));
            prop_assert!(input
                .0
                .data
                .iter()
                .skip(overwritten)
                .take(write1.len().saturating_sub(overwritten))
                .zip(write1.iter().skip(overwritten))
                .all(|(a, &x1)| a.load(Ordering::Relaxed) == x1));
            prop_assert!(input
                .0
                .data
                .iter()
                .skip(write1.len())
                .take(write2.len())
                .zip(&write2)
                .all(|(a, &x2)| a.load(Ordering::Relaxed) == x2));
            prop_assert!(input
                .0
                .data
                .iter()
                .skip(new_readable)
                .all(|a| a.load(Ordering::Relaxed) == 0));
            prop_assert_eq!(input.0.readable.load(Ordering::Relaxed), new_readable);
            prop_assert_eq!(input.0.writing.load(Ordering::Relaxed), new_readable);

            // Back up current ring buffer state
            let data_backup = input
                .0
                .data
                .iter()
                .map(|a| a.load(Ordering::Relaxed))
                .collect::<Box<[_]>>();

            // Read some data back and make sure no overrun happened (it is
            // impossible in single-threaded code)
            let mut read = vec![0; read_size];
            match panic::catch_unwind(AssertUnwindSafe(|| output.read(&mut read[..]))) {
                Err(_) => {
                    prop_assert!(read.len() > output.0.data_len());
                    return Ok(());
                }
                Ok(result) => {
                    prop_assert!(read.len() <= output.0.data_len());
                    prop_assert_eq!(result, Ok(new_readable));
                }
            }

            // Make sure that the "read" did not alter the history data
            prop_assert!(input
                .0
                .data
                .iter()
                .zip(data_backup.iter().copied())
                .all(|(a, x)| a.load(Ordering::Relaxed) == x));
            prop_assert_eq!(input.0.readable.load(Ordering::Relaxed), new_readable);
            prop_assert_eq!(input.0.writing.load(Ordering::Relaxed), new_readable);

            // Check that the collected output is correct
            prop_assert!(read
                .iter()
                .rev()
                .zip(
                    write2
                        .iter()
                        .rev()
                        .chain(write1.iter().rev())
                        .chain(std::iter::repeat(&0))
                )
                .all(|(&x, &y)| x == y));
        }
    }

    // This test is ignored because it needs a special configuration:
    // - Be built in release mode
    // - Be run on a machine with low background load
    // - 2+ physical cores are needed to validate basic 1 producer / 1 consumer
    //   operation, and 3+ physical cores to validate multi-consumer operation.
    #[test]
    #[ignore]
    fn concurrent_test() {
        const PRODUCER_SIZE: usize = 1 << 3; // One 64B cache line
        const CONSUMER_SIZE: usize = PRODUCER_SIZE << 1; // To observe partial overrun
        const BUFFER_SIZE: usize = CONSUMER_SIZE << 2; // 2x = too many overruns

        const NUM_ELEMS: usize = 1 << 31;

        let (mut input, output) = RTHistory::<u64>::new(BUFFER_SIZE).split();

        // The producer simply emits regularly increasing counter values.
        // This makes consumer validation as it means value == associated clock.
        // It also regularly sleeps to trigger some buffer underruns.
        let mut last_emitted = 0;
        let producer = move || {
            while last_emitted < NUM_ELEMS as u64 {
                let mut buf = [0; PRODUCER_SIZE];
                for dst in &mut buf {
                    last_emitted += 1;
                    *dst = last_emitted;
                }
                input.write(&buf[..]);
            }
        };

        // Consumers detect underruns and overruns, and asserts that for data
        // which has not been corrupted by an overrun, value == clock holds.
        #[allow(clippy::declare_interior_mutable_const)]
        const XRUN_CTR_INIT: AtomicUsize = AtomicUsize::new(0);
        let underrun_ctrs = [XRUN_CTR_INIT; 2];
        let overrun_ctrs = [XRUN_CTR_INIT; 2];
        let gen_consumer = |idx| {
            let num_underruns: &AtomicUsize = &underrun_ctrs[idx];
            let num_overruns: &AtomicUsize = &overrun_ctrs[idx];
            let output = output.clone();
            move || {
                let mut last_clock = 0;
                let mut last_underrun = usize::MAX;
                let mut buf = [0; CONSUMER_SIZE];
                while last_clock < NUM_ELEMS {
                    // Fetch latest batch from producer, return clock and the
                    // valid subset of data that hasn't been corrupted by a
                    // buffer overrun.
                    let (clock, valid) = match output.read(&mut buf[..]) {
                        // In absence of overrun, all data is valid
                        Ok(clock) => (clock, &buf[..]),

                        // When an overrun occurs...
                        Err(Overrun {
                            clock,
                            excess_entries,
                        }) => {
                            // Check the number of excess entries makes sense
                            assert!(excess_entries > 0);
                            assert_eq!(excess_entries % PRODUCER_SIZE, 0);

                            // Keep track of number of overruns (increment does
                            // not need to be atomic because the value will only
                            // be read out in the end).
                            num_overruns
                                .store(num_overruns.load(Ordering::Relaxed) + 1, Ordering::Relaxed);

                            // Extract valid subset of data
                            if excess_entries < buf.len() {
                                (clock, &buf[excess_entries + 1..])
                            } else {
                                (clock, &[][..])
                            }
                        }
                    };

                    // Monitor clock progression
                    if clock == last_clock {
                        // An unchanging clock means a buffer underrun occured,
                        // make sure we only record each of them once.
                        if clock != last_underrun {
                            num_underruns.store(
                                num_underruns.load(Ordering::Relaxed) + 1,
                                Ordering::Relaxed,
                            );
                            last_underrun = clock;
                        }
                    } else {
                        // The clock can only move forward
                        assert!(clock > last_clock);
                        last_clock = clock;
                    }

                    // Check that value == clock property is honored
                    for (expected, &actual) in (1..=clock).rev().zip(valid.iter().rev()) {
                        assert_eq!(expected as u64, actual);
                    }
                }
            }
        };

        // Run the test
        testbench::concurrent_test_3(producer, gen_consumer(0), gen_consumer(1));

        // Check that a significant number of xruns were detected, but that
        // there was a significant number of "normal" runs too.
        const NUM_READOUTS: usize = NUM_ELEMS / CONSUMER_SIZE;
        underrun_ctrs
            .into_iter()
            .map(|a| a.load(Ordering::Relaxed))
            .enumerate()
            .for_each(|(idx, underrun_ctr)| {
                println!(
                    "consumer {}: {}/{} underruns",
                    idx, underrun_ctr, NUM_READOUTS
                );
                assert!(underrun_ctr > NUM_READOUTS / 10000);
                assert!(underrun_ctr < NUM_READOUTS / 5);
            });
        overrun_ctrs
            .into_iter()
            .map(|a| a.load(Ordering::Relaxed))
            .enumerate()
            .for_each(|(idx, overrun_ctr)| {
                println!(
                    "consumer {}: {}/{} overruns",
                    idx, overrun_ctr, NUM_READOUTS
                );
                assert!(overrun_ctr > NUM_READOUTS / 10000);
                assert!(overrun_ctr < NUM_READOUTS / 5);
            });
    }
}