word_lock.rs

// Copyright 2016 Amanieu d'Antras
//
// Licensed under the Apache License, Version 2.0, <LICENSE-APACHE or
// http://apache.org/licenses/LICENSE-2.0> or the MIT license <LICENSE-MIT or
// http://opensource.org/licenses/MIT>, at your option. This file may not be
// copied, modified, or distributed except according to those terms.

use crate::spinwait::SpinWait;
use crate::thread_parker::{ThreadParker, ThreadParkerT, UnparkHandleT};
use core::{
    cell::Cell,
    mem, ptr,
    sync::atomic::{fence, AtomicUsize, Ordering},
};

struct ThreadData {
    parker: ThreadParker,

    // Linked list of threads in the queue. The queue is split into two parts:
    // the processed part and the unprocessed part. When new nodes are added to
    // the list, they only have the next pointer set, and queue_tail is null.
    //
    // Nodes are processed with the queue lock held, which consists of setting
    // the prev pointer for each node and setting the queue_tail pointer on the
    // first processed node of the list.
    //
    // This setup allows nodes to be added to the queue without a lock, while
    // still allowing O(1) removal of nodes from the processed part of the list.
    // The only cost is the O(n) processing, but this only needs to be done
    // once for each node, and therefore isn't too expensive.
    queue_tail: Cell<*const ThreadData>,
    prev: Cell<*const ThreadData>,
    next: Cell<*const ThreadData>,
}

impl ThreadData {
    #[inline]
    fn new() -> ThreadData {
        assert!(mem::align_of::<ThreadData>() > !QUEUE_MASK);
        ThreadData {
            parker: ThreadParker::new(),
            queue_tail: Cell::new(ptr::null()),
            prev: Cell::new(ptr::null()),
            next: Cell::new(ptr::null()),
        }
    }
}

// Invokes the given closure with a reference to the current thread `ThreadData`.
#[inline]
fn with_thread_data<T>(f: impl FnOnce(&ThreadData) -> T) -> T {
    let mut thread_data_ptr = ptr::null();
    // If ThreadData is expensive to construct, then we want to use a cached
    // version in thread-local storage if possible.
    if !ThreadParker::IS_CHEAP_TO_CONSTRUCT {
        thread_local!(static THREAD_DATA: ThreadData = ThreadData::new());
        if let Ok(tls_thread_data) = THREAD_DATA.try_with(|x| x as *const ThreadData) {
            thread_data_ptr = tls_thread_data;
        }
    }
    // Otherwise just create a ThreadData on the stack
    let mut thread_data_storage = None;
    if thread_data_ptr.is_null() {
        thread_data_ptr = thread_data_storage.get_or_insert_with(ThreadData::new);
    }

    f(unsafe { &*thread_data_ptr })
}

const LOCKED_BIT: usize = 1;
const QUEUE_LOCKED_BIT: usize = 2;
const QUEUE_MASK: usize = !3;

// Word-sized lock that is used to implement the parking_lot API. Since this
// can't use parking_lot, it instead manages its own queue of waiting threads.
pub struct WordLock {
    state: AtomicUsize,
}

impl WordLock {
    /// Returns a new, unlocked, `WordLock`.
    pub const fn new() -> Self {
        WordLock {
            state: AtomicUsize::new(0),
        }
    }

    #[inline]
    pub fn lock(&self) {
        if self
            .state
            .compare_exchange_weak(0, LOCKED_BIT, Ordering::Acquire, Ordering::Relaxed)
            .is_ok()
        {
            return;
        }
        self.lock_slow();
    }

    /// Must not be called on an already unlocked `WordLock`!
    #[inline]
    pub unsafe fn unlock(&self) {
        let state = self.state.fetch_sub(LOCKED_BIT, Ordering::Release);
        if state.is_queue_locked() || state.queue_head().is_null() {
            return;
        }
        self.unlock_slow();
    }

    #[cold]
    fn lock_slow(&self) {
        let mut spinwait = SpinWait::new();
        let mut state = self.state.load(Ordering::Relaxed);
        loop {
            // Grab the lock if it isn't locked, even if there is a queue on it
            if !state.is_locked() {
                match self.state.compare_exchange_weak(
                    state,
                    state | LOCKED_BIT,
                    Ordering::Acquire,
                    Ordering::Relaxed,
                ) {
                    Ok(_) => return,
                    Err(x) => state = x,
                }
                continue;
            }

            // If there is no queue, try spinning a few times
            if state.queue_head().is_null() && spinwait.spin() {
                state = self.state.load(Ordering::Relaxed);
                continue;
            }

            // Get our thread data and prepare it for parking
            state = with_thread_data(|thread_data| {
                // The pthread implementation is still unsafe, so we need to surround `prepare_park`
                // with `unsafe {}`.
                #[allow(unused_unsafe)]
                unsafe {
                    thread_data.parker.prepare_park();
                }

                // Add our thread to the front of the queue
                let queue_head = state.queue_head();
                if queue_head.is_null() {
                    thread_data.queue_tail.set(thread_data);
                    thread_data.prev.set(ptr::null());
                } else {
                    thread_data.queue_tail.set(ptr::null());
                    thread_data.prev.set(ptr::null());
                    thread_data.next.set(queue_head);
                }
                if let Err(x) = self.state.compare_exchange_weak(
                    state,
                    state.with_queue_head(thread_data),
                    Ordering::AcqRel,
                    Ordering::Relaxed,
                ) {
                    return x;
                }

                // Sleep until we are woken up by an unlock
                // Ignoring unused unsafe, since it's only a few platforms where this is unsafe.
                #[allow(unused_unsafe)]
                unsafe {
                    thread_data.parker.park();
                }

                // Loop back and try locking again
                spinwait.reset();
                self.state.load(Ordering::Relaxed)
            });
        }
    }

    #[cold]
    fn unlock_slow(&self) {
        let mut state = self.state.load(Ordering::Relaxed);
        loop {
            // We just unlocked the WordLock. Just check if there is a thread
            // to wake up. If the queue is locked then another thread is already
            // taking care of waking up a thread.
            if state.is_queue_locked() || state.queue_head().is_null() {
                return;
            }

            // Try to grab the queue lock
            match self.state.compare_exchange_weak(
                state,
                state | QUEUE_LOCKED_BIT,
                Ordering::Acquire,
                Ordering::Relaxed,
            ) {
                Ok(_) => break,
                Err(x) => state = x,
            }
        }

        // Now we have the queue lock and the queue is non-empty
        'outer: loop {
            // First, we need to fill in the prev pointers for any newly added
            // threads. We do this until we reach a node that we previously
            // processed, which has a non-null queue_tail pointer.
            let queue_head = state.queue_head();
            let mut queue_tail;
            let mut current = queue_head;
            loop {
                queue_tail = unsafe { (*current).queue_tail.get() };
                if !queue_tail.is_null() {
                    break;
                }
                unsafe {
                    let next = (*current).next.get();
                    (*next).prev.set(current);
                    current = next;
                }
            }

            // Set queue_tail on the queue head to indicate that the whole list
            // has prev pointers set correctly.
            unsafe {
                (*queue_head).queue_tail.set(queue_tail);
            }

            // If the WordLock is locked, then there is no point waking up a
            // thread now. Instead we let the next unlocker take care of waking
            // up a thread.
            if state.is_locked() {
                match self.state.compare_exchange_weak(
                    state,
                    state & !QUEUE_LOCKED_BIT,
                    Ordering::Release,
                    Ordering::Relaxed,
                ) {
                    Ok(_) => return,
                    Err(x) => state = x,
                }

                // Need an acquire fence before reading the new queue
                fence_acquire(&self.state);
                continue;
            }

            // Remove the last thread from the queue and unlock the queue
            let new_tail = unsafe { (*queue_tail).prev.get() };
            if new_tail.is_null() {
                loop {
                    match self.state.compare_exchange_weak(
                        state,
                        state & LOCKED_BIT,
                        Ordering::Release,
                        Ordering::Relaxed,
                    ) {
                        Ok(_) => break,
                        Err(x) => state = x,
                    }

                    // If the compare_exchange failed because a new thread was
                    // added to the queue then we need to re-scan the queue to
                    // find the previous element.
                    if state.queue_head().is_null() {
                        continue;
                    } else {
                        // Need an acquire fence before reading the new queue
                        fence_acquire(&self.state);
                        continue 'outer;
                    }
                }
            } else {
                unsafe {
                    (*queue_head).queue_tail.set(new_tail);
                }
                self.state.fetch_and(!QUEUE_LOCKED_BIT, Ordering::Release);
            }

            // Finally, wake up the thread we removed from the queue. Note that
            // we don't need to worry about any races here since the thread is
            // guaranteed to be sleeping right now and we are the only one who
            // can wake it up.
            unsafe {
                (*queue_tail).parker.unpark_lock().unpark();
            }
            break;
        }
    }
}

// Thread-Sanitizer only has partial fence support, so when running under it, we
// try and avoid false positives by using a discarded acquire load instead.
#[inline]
fn fence_acquire(a: &AtomicUsize) {
    if cfg!(tsan_enabled) {
        let _ = a.load(Ordering::Acquire);
    } else {
        fence(Ordering::Acquire);
    }
}

trait LockState {
    fn is_locked(self) -> bool;
    fn is_queue_locked(self) -> bool;
    fn queue_head(self) -> *const ThreadData;
    fn with_queue_head(self, thread_data: *const ThreadData) -> Self;
}

impl LockState for usize {
    #[inline]
    fn is_locked(self) -> bool {
        self & LOCKED_BIT != 0
    }

    #[inline]
    fn is_queue_locked(self) -> bool {
        self & QUEUE_LOCKED_BIT != 0
    }

    #[inline]
    fn queue_head(self) -> *const ThreadData {
        (self & QUEUE_MASK) as *const ThreadData
    }

    #[inline]
    fn with_queue_head(self, thread_data: *const ThreadData) -> Self {
        (self & !QUEUE_MASK) | thread_data as *const _ as usize
    }
}