devices/tsc/

calibrate.rs

// Copyright 2022 The ChromiumOS Authors
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

use std::collections::HashSet;
use std::iter::FromIterator;
use std::time::Duration;
use std::time::Instant;

use anyhow::anyhow;
use anyhow::Context;
use anyhow::Result;
use base::set_cpu_affinity;
use base::warn;
use remain::sorted;
use thiserror::Error;

use super::grouping::*;
use super::rdtsc_safe;

const TSC_CALIBRATION_SAMPLES: usize = 10;
const TSC_CALIBRATION_DURATION: Duration = Duration::from_millis(100);
// remove data that is outside 3 standard deviations off the median
const TSC_CALIBRATION_STANDARD_DEVIATION_LIMIT: f64 = 3.0;
// We consider two TSC cores to be in sync if they are within 2 microseconds of each other.
// An optimal context switch takes about 1-3 microseconds.
const TSC_OFFSET_GROUPING_THRESHOLD: Duration = Duration::from_micros(2);

#[sorted]
#[derive(Error, Debug)]
pub enum TscCalibrationError {
    /// Received `err` when setting the cpu affinity to `core`
    #[error("failed to set thread cpu affinity to core {core}: {err}")]
    SetCpuAffinityError { core: usize, err: base::Error },
}

/// Get the standard deviation of a `Vec<T>`.
pub fn standard_deviation<T: num_traits::ToPrimitive + num_traits::Num + Copy>(items: &[T]) -> f64 {
    let sum: T = items.iter().fold(T::zero(), |acc: T, elem| acc + *elem);
    let count = items.len();

    let mean: f64 = sum.to_f64().unwrap_or(0.0) / count as f64;

    let variance = items
        .iter()
        .map(|x| {
            let diff = mean - (x.to_f64().unwrap_or(0.0));
            diff * diff
        })
        .sum::<f64>();
    (variance / count as f64).sqrt()
}

fn sort_and_get_bounds(items: &mut [i128], stdev_limit: f64) -> (f64, f64) {
    items.sort_unstable();
    let median = items[items.len() / 2];

    let standard_deviation = standard_deviation(items);
    let lower_bound = median as f64 - stdev_limit * standard_deviation;
    let upper_bound = median as f64 + stdev_limit * standard_deviation;
    (lower_bound, upper_bound)
}

/// Represents the host monotonic time and the TSC value at a single moment in time.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
struct TscMoment {
    time: Instant,
    tsc: u64,
}

impl TscMoment {
    fn now(rdtsc: fn() -> u64) -> Self {
        TscMoment {
            time: Instant::now(),
            tsc: rdtsc(),
        }
    }

    /// Measure the tsc frequency using two `TscMoment`s.
    fn measure_tsc_frequency(first: &TscMoment, second: &TscMoment) -> i128 {
        // handle case where first is actually second in time
        let (first, second) = if first.time > second.time {
            (second, first)
        } else {
            (first, second)
        };

        let time_delta = second.time - first.time;
        let tsc_delta = second.tsc as i128 - first.tsc as i128;

        tsc_delta * 1_000_000_000i128 / time_delta.as_nanos() as i128
    }

    /// Measure the tsc offset using two `TscMoment`s and the TSC frequency.
    fn measure_tsc_offset(first: &TscMoment, second: &TscMoment, tsc_frequency: u64) -> i128 {
        // handle case where first is actually second in time
        let (first, second) = if first.time > second.time {
            (second, first)
        } else {
            (first, second)
        };

        let tsc_delta = second.tsc as i128 - first.tsc as i128;
        let time_delta_as_tsc_ticks =
            (second.time - first.time).as_nanos() * tsc_frequency as u128 / 1_000_000_000u128;
        tsc_delta - time_delta_as_tsc_ticks as i128
    }
}

#[derive(Default, Debug, Clone)]
pub struct TscState {
    pub frequency: u64,
    pub offsets: Vec<(usize, i128)>,
    pub core_grouping: CoreGrouping,
}

impl TscState {
    pub(crate) fn new(
        tsc_frequency: u64,
        offsets: Vec<(usize, i128)>,
        in_sync_threshold: Duration,
    ) -> Result<Self> {
        let core_grouping = group_core_offsets(&offsets, in_sync_threshold, tsc_frequency)
            .context("Failed to group cores by their TSC offsets")?;
        Ok(TscState {
            frequency: tsc_frequency,
            offsets,
            core_grouping,
        })
    }
}

/// Calibrate the TSC frequency of `core`.
///
/// This function first pins itself to `core`, generates `num_samples` start `TscMoment`s, sleeps
/// for `calibration_duration`, and then generates `num_samples` end `TscMoment`s. For each pair
/// of start and end moments, a TSC frequency value is calculated. Any frequencies that are
/// outside of `stddev_limit` standard deviations from the median offset are discarded, because
/// they may represent an interrupt that occurred while a TscMoment was generated. The remaining
/// non-discarded frequencies are then averaged. The function returns the TSC frequency average, as
/// well as a Vec of `TscMoment`s, which are all of the end moments that were associated with at
/// least one non-discarded frequency.
///
/// # Arguments
/// * `core` - Core that this function should run on.
/// * `rdtsc` - Function for reading the TSC value, usually just runs RDTSC instruction.
/// * `num_samples` - Number of start and end `TscMoment`s to generate.
/// * `calibration_duration` - How long to sleep in between gathering start and end moments.
/// * `stdev_limit` - Number of standard deviations outside of which frequencies are discarded.
fn calibrate_tsc_frequency(
    rdtsc: fn() -> u64,
    core: usize,
    num_samples: usize,
    calibration_duration: Duration,
    stdev_limit: f64,
) -> std::result::Result<(i128, Vec<TscMoment>), TscCalibrationError> {
    set_cpu_affinity(vec![core])
        .map_err(|e| TscCalibrationError::SetCpuAffinityError { core, err: e })?;

    let starts: Vec<TscMoment> = (0..num_samples).map(|_| TscMoment::now(rdtsc)).collect();

    std::thread::sleep(calibration_duration);

    let ends: Vec<TscMoment> = (0..num_samples).map(|_| TscMoment::now(rdtsc)).collect();

    let mut freqs = Vec::with_capacity(num_samples * num_samples);
    for start in &starts {
        for end in &ends {
            freqs.push(TscMoment::measure_tsc_frequency(start, end))
        }
    }

    let (lower_bound, upper_bound) = sort_and_get_bounds(&mut freqs, stdev_limit);

    let mut good_samples: Vec<i128> = Vec::with_capacity(num_samples * num_samples);
    let mut good_end_moments: HashSet<TscMoment> = HashSet::new();
    for i in 0..num_samples {
        for j in 0..num_samples {
            let freq = freqs[i * num_samples + j];

            if lower_bound < (freq as f64) && (freq as f64) < upper_bound {
                good_end_moments.insert(ends[j]);
                good_samples.push(freq);
            }
        }
    }

    Ok((
        good_samples.iter().sum::<i128>() / good_samples.len() as i128,
        Vec::from_iter(good_end_moments),
    ))
}

/// Measure the TSC offset for `core` from core 0 where `reference_moments` were gathered.
///
/// This function first pins itself to `core`, then generates `num_samples` `TscMoment`s for this
/// core, and then measures the TSC offset between those moments and all `reference_moments`. Any
/// moments that are outside of `stddev_limit` standard deviations from the median offset are
/// discarded, because they may represent an interrupt that occurred while a TscMoment was
/// generated. The remaining offsets are averaged and returned as nanoseconds.
///
/// # Arguments
/// * `core` - Core that this function should run on.
/// * `rdtsc` - Function for reading the TSC value, usually just runs RDTSC instruction.
/// * `tsc_frequency` - TSC frequency measured from core 0.
/// * `reference_moments` - `TscMoment`s gathered from core 0.
/// * `num_samples` - Number of `TscMoment`s to generate on this thread for measuring the offset.
/// * `stdev_limit` - Number of standard deviations outside of which offsets are discarded.
fn measure_tsc_offset(
    core: usize,
    rdtsc: fn() -> u64,
    tsc_frequency: u64,
    reference_moments: Vec<TscMoment>,
    num_samples: usize,
    stdev_limit: f64,
) -> std::result::Result<i128, TscCalibrationError> {
    set_cpu_affinity(vec![core])
        .map_err(|e| TscCalibrationError::SetCpuAffinityError { core, err: e })?;

    let mut diffs: Vec<i128> = Vec::with_capacity(num_samples);

    for _ in 0..num_samples {
        let now = TscMoment::now(rdtsc);
        for reference_moment in &reference_moments {
            diffs.push(TscMoment::measure_tsc_offset(
                reference_moment,
                &now,
                tsc_frequency,
            ));
        }
    }

    let (lower_bound, upper_bound) = sort_and_get_bounds(&mut diffs, stdev_limit);

    let mut good_samples: Vec<i128> = Vec::with_capacity(num_samples);
    for diff in &diffs {
        if lower_bound < (*diff as f64) && (*diff as f64) < upper_bound {
            good_samples.push(*diff);
        }
    }

    let average_diff = good_samples.iter().sum::<i128>() / good_samples.len() as i128;

    // Convert the diff to nanoseconds using the tsc_frequency
    Ok(average_diff * 1_000_000_000 / tsc_frequency as i128)
}

/// Calibrate the TSC state.
///
/// This function first runs a TSC frequency calibration thread for 100ms, which is pinned to
/// core0. The TSC calibration thread returns both the calibrated frequency, as well as a Vec of
/// TscMoment objects which were validated to be accurate (meaning it's unlikely an interrupt
/// occurred between moment's `time` and `tsc` values). This function then runs a tsc offset
/// measurement thread for each core, which takes the TSC frequency and the Vec of TscMoments and
/// measures whether or not the TSC values for that core are offset from core 0, and by how much.
/// The frequency and the per-core offsets are returned as a TscState.
pub fn calibrate_tsc_state() -> Result<TscState> {
    calibrate_tsc_state_inner(
        rdtsc_safe,
        (0..base::number_of_logical_cores().context("Failed to get number of logical cores")?)
            .collect(),
    )
}

/// Actually calibrate the TSC state.
///
/// This function takes a customizable version of rdtsc and a specific set of cores to calibrate,
/// which is helpful for testing calibration logic and error handling.
///
/// # Arguments
///
/// * `rdtsc` - Function for reading the TSC value, usually just runs RDTSC instruction.
/// * `cores` - Cores to measure the TSC offset of.
fn calibrate_tsc_state_inner(rdtsc: fn() -> u64, cores: Vec<usize>) -> Result<TscState> {
    // For loops can't return values unfortunately
    let mut calibration_contents: Option<(u64, Vec<TscMoment>)> = None;
    for core in &cores {
        // Copy the value of core to a moveable variable now.
        let moved_core = *core;
        let handle = std::thread::Builder::new()
            .name(format!("tsc_calibration_core_{core}").to_string())
            .spawn(move || {
                calibrate_tsc_frequency(
                    rdtsc,
                    moved_core,
                    TSC_CALIBRATION_SAMPLES,
                    TSC_CALIBRATION_DURATION,
                    TSC_CALIBRATION_STANDARD_DEVIATION_LIMIT,
                )
            })
            .map_err(|e| {
                anyhow!(
                    "TSC frequency calibration thread for core {} failed: {:?}",
                    core,
                    e
                )
            })?;

        match handle.join() {
            Ok(calibrate_result) => match calibrate_result {
                Ok((freq, reference_moments)) => {
                    if freq <= 0 {
                        warn!(
                            "TSC calibration on core {} resulted in TSC frequency of {} Hz, \
                    trying on another core.",
                            core, freq
                        );
                        continue;
                    };
                    calibration_contents = Some((freq as u64, reference_moments));
                    break;
                }

                Err(TscCalibrationError::SetCpuAffinityError { core, err }) => {
                    // There are several legitimate reasons why it might not be possible for crosvm
                    // to run on some cores:
                    //  1. Some cores may be offline.
                    //  2. On Windows, the process affinity mask may not contain all cores.
                    //
                    // We thus just warn in this situation.
                    warn!(
                        "Failed to set thread affinity to {} during tsc frequency calibration due \
                            to {}. This core is probably offline.",
                        core, err
                    );
                }
            },
            // thread failed
            Err(e) => {
                return Err(anyhow!(
                    "TSC frequency calibration thread for core {} failed: {:?}",
                    core,
                    e
                ));
            }
        };
    }

    let (freq, reference_moments) =
        calibration_contents.ok_or(anyhow!("Failed to calibrate TSC frequency on all cores"))?;

    let mut offsets: Vec<(usize, i128)> = Vec::with_capacity(cores.len());
    for core in cores {
        let thread_reference_moments = reference_moments.clone();
        let handle = std::thread::Builder::new()
            .name(format!("measure_tsc_offset_core_{core}").to_string())
            .spawn(move || {
                measure_tsc_offset(
                    core,
                    rdtsc,
                    freq,
                    thread_reference_moments,
                    TSC_CALIBRATION_SAMPLES,
                    TSC_CALIBRATION_STANDARD_DEVIATION_LIMIT,
                )
            })
            .map_err(|e| {
                anyhow!(
                    "TSC offset measurement thread for core {} failed: {:?}",
                    core,
                    e
                )
            })?;
        let offset = match handle.join() {
            // thread succeeded
            Ok(measurement_result) => match measurement_result {
                Ok(offset) => Some(offset),
                Err(TscCalibrationError::SetCpuAffinityError { core, err }) => {
                    // There are several legitimate reasons why it might not be possible for crosvm
                    // to run on some cores:
                    //  1. Some cores may be offline.
                    //  2. On Windows, the process affinity mask may not contain all cores.
                    //
                    // We thus just warn in this situation.
                    warn!(
                        "Failed to set thread affinity to {} during tsc offset measurement due \
                        to {}. This core is probably offline.",
                        core, err
                    );
                    None
                }
            },
            // thread failed
            Err(e) => {
                return Err(anyhow!(
                    "TSC offset measurement thread for core {} failed: {:?}",
                    core,
                    e
                ));
            }
        };

        if let Some(offset) = offset {
            offsets.push((core, offset));
        }
    }

    TscState::new(freq, offsets, TSC_OFFSET_GROUPING_THRESHOLD)
}

#[cfg(test)]
mod tests {
    use std::arch::x86_64::__rdtscp;
    use std::arch::x86_64::_rdtsc;

    use super::*;

    const ACCEPTABLE_OFFSET_MEASUREMENT_ERROR: i128 = 2_000i128;

    #[test]
    fn test_handle_offline_core() {
        // This test imitates what would happen if a core is offline, and set_cpu_affinity fails.
        // The calibration should not fail, and the extra core should not appear in the list of
        // offsets.

        let num_cores =
            base::number_of_logical_cores().expect("number of logical cores should not fail");

        let too_may_cores = num_cores + 2;
        let host_state = calibrate_tsc_state_inner(rdtsc_safe, (0..too_may_cores).collect())
            .expect("calibrate tsc state should not fail");

        // First assert that the number of offsets measured is at most num_cores (it might be
        // less if the current host has some offline cores).
        assert!(host_state.offsets.len() <= num_cores);

        for (core, _) in host_state.offsets {
            // Assert that all offsets that we have are for cores 0..num_cores.
            assert!(core < num_cores);
        }
    }

    #[test]
    fn test_frequency_higher_than_u32() {
        // This test is making sure that we're not truncating our TSC frequencies in the case that
        // they are greater than u32::MAX.

        let host_state = calibrate_tsc_state_inner(
            rdtsc_safe,
            (0..base::number_of_logical_cores().expect("number of logical cores should not fail"))
                .collect(),
        )
        .expect("failed to calibrate host freq");

        // We use a static multiplier of 1000 here because the function has to be static (fn).
        // 1000 should work for tsc frequency > 4.2MHz, which should apply to basically any
        // processor. This if statement checks and bails early if that's not the case.
        if host_state.frequency * 1000 < (u32::MAX as u64) {
            return;
        }

        fn rdtsc_frequency_higher_than_u32() -> u64 {
            // SAFETY: trivially safe
            unsafe { _rdtsc() }.wrapping_mul(1000)
        }

        let state = calibrate_tsc_state_inner(
            rdtsc_frequency_higher_than_u32,
            (0..base::number_of_logical_cores().expect("number of logical cores should not fail"))
                .collect(),
        )
        .unwrap();

        let expected_freq = host_state.frequency * 1000;
        let margin_of_error = expected_freq / 100;
        assert!(state.frequency < expected_freq + margin_of_error);
        assert!(state.frequency > expected_freq - margin_of_error);
    }

    #[test]
    #[ignore]
    fn test_offset_identification_core_0() {
        fn rdtsc_with_core_0_offset_by_100_000() -> u64 {
            let mut id = 0u32;
            // SAFETY: trivially safe
            let mut value = unsafe { __rdtscp(&mut id as *mut u32) };
            if id == 0 {
                value += 100_000;
            }

            value
        }

        // This test only works if the host has >=2 logical cores.
        let num_cores =
            base::number_of_logical_cores().expect("Failed to get number of logical cores");
        if num_cores < 2 {
            return;
        }

        let state = calibrate_tsc_state_inner(
            rdtsc_with_core_0_offset_by_100_000,
            (0..base::number_of_logical_cores().expect("number of logical cores should not fail"))
                .collect(),
        )
        .unwrap();

        for core in 0..num_cores {
            let expected_offset_ns = if core > 0 {
                -100_000i128 * 1_000_000_000i128 / state.frequency as i128
            } else {
                0i128
            };
            assert!(
                state.offsets[core].1 < expected_offset_ns + ACCEPTABLE_OFFSET_MEASUREMENT_ERROR
            );
            assert!(
                state.offsets[core].1 > expected_offset_ns - ACCEPTABLE_OFFSET_MEASUREMENT_ERROR
            );
        }
    }

    #[test]
    #[ignore]
    fn test_offset_identification_core_1() {
        fn rdtsc_with_core_1_offset_by_100_000() -> u64 {
            let mut id = 0u32;
            // SAFETY: trivially safe
            let mut value = unsafe { __rdtscp(&mut id as *mut u32) };
            if id == 1 {
                value += 100_000;
            }

            value
        }

        // This test only works if the host has >=2 logical cores.
        let num_cores =
            base::number_of_logical_cores().expect("Failed to get number of logical cores");
        if num_cores < 2 {
            return;
        }

        let state = calibrate_tsc_state_inner(
            rdtsc_with_core_1_offset_by_100_000,
            (0..base::number_of_logical_cores().expect("number of logical cores should not fail"))
                .collect(),
        )
        .unwrap();

        for core in 0..num_cores {
            let expected_offset_ns = if core == 1 {
                100_000i128 * 1_000_000_000i128 / state.frequency as i128
            } else {
                0i128
            };
            assert!(
                state.offsets[core].1 < expected_offset_ns + ACCEPTABLE_OFFSET_MEASUREMENT_ERROR
            );
            assert!(
                state.offsets[core].1 > expected_offset_ns - ACCEPTABLE_OFFSET_MEASUREMENT_ERROR
            );
        }
    }
}