sync: new internal semaphore based on intrusive lists (#2325)

## Motivation

Many of Tokio's synchronization primitives (`RwLock`, `Mutex`,
`Semaphore`, and the bounded MPSC channel) are based on the internal
semaphore implementation, called `semaphore_ll`. This semaphore type
provides a lower-level internal API for the semaphore implementation
than the public `Semaphore` type, and supports "batch" operations, where
waiters may acquire more than one permit at a time, and batches of
permits may be released back to the semaphore.

Currently, `semaphore_ll` uses an atomic singly-linked list for the
waiter queue. The linked list implementation is specific to the
semaphore. This implementation therefore requires a heap allocation for
every waiter in the queue. These allocations are owned by the semaphore,
rather than by the task awaiting permits from the semaphore. Critically,
they are only _deallocated_ when permits are released back to the
semaphore, at which point it dequeues as many waiters from the front of
the queue as can be satisfied with the released permits. If a task
attempts to acquire permits from the semaphore and is cancelled (such as
by timing out), their waiter nodes remain in the list until they are
dequeued while releasing permits. In cases where large numbers of tasks
are cancelled while waiting for permits, this results in extremely high
memory use for the semaphore (see #2237).

## Solution

@Matthias247 has proposed that Tokio adopt the approach used in his
`futures-intrusive` crate: using an _intrusive_ linked list to store the
wakers of tasks waiting on a synchronization primitive. In an intrusive
list, each list node is stored as part of the entry that node
represents, rather than in a heap allocation that owns the entry.
Because futures must be pinned in order to be polled, the necessary
invariant of such a list --- that entries may not move while in the list
--- may be upheld by making the waiter node `!Unpin`. In this approach,
the waiter node can be stored inline in the future, rather than
requiring  separate heap allocation, and cancelled futures may remove
their nodes from the list.

This branch adds a new semaphore implementation that uses the intrusive
list added to Tokio in #2210. The implementation is essentially a hybrid
of the old `semaphore_ll` and the semaphore used in `futures-intrusive`:
while a `Mutex` around the wait list is necessary, since the intrusive
list is not thread-safe, the permit state is stored outside of the mutex
and updated atomically. 

The mutex is acquired only when accessing the wait list — if a task 
can acquire sufficient permits without waiting, it does not need to
acquire the lock. When releasing permits, we iterate over the wait
list from the end of the queue until we run out of permits to release,
and split off all the nodes that received enough permits to wake up
into a separate list. Then, we can drain the new list and notify those
wakers *after* releasing the lock. Because the split operation only
modifies the pointers on the head node of the split-off list and the
new tail node of the old list, it is O(1) and does not require an
allocation to return a variable length number of waiters to notify.


Because of the intrusive list invariants, the API provided by the new
`batch_semaphore` is somewhat different than that of `semaphore_ll`. In
particular, the `Permit` type has been removed. This type was primarily
intended allow the reuse of a wait list node allocated on the heap.
Since the intrusive list means we can avoid heap-allocating waiters,
this is no longer necessary. Instead, acquiring permits is done by
polling an `Acquire` future returned by the `Semaphore` type. The use of
a future here ensures that the waiter node is always pinned while
waiting to acquire permits, and that a reference to the semaphore is
available to remove the waiter if the future is cancelled.
Unfortunately, the current implementation of the bounded MPSC requires a
`poll_acquire` operation, and has methods that call it while outside of
a pinned context. Therefore, I've left the old `semaphore_ll`
implementation in place to be used by the bounded MPSC, and updated the
`Mutex`, `RwLock`, and `Semaphore` APIs to use the new implementation.
Hopefully, a subsequent change can update the bounded MPSC to use the
new semaphore as well.

Fixes #2237

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

1 files changed, 553 insertions, 0 deletions

diff --git a/tokio/src/sync/batch_semaphore.rs b/tokio/src/sync/batch_semaphore.rs
new file mode 100644
index 00000000..d89ac6ab
--- /dev/null
+++ b/tokio/src/sync/batch_semaphore.rs
@@ -0,0 +1,553 @@
+//! # Implementation Details
+//!
+//! The semaphore is implemented using an intrusive linked list of waiters. An
+//! atomic counter tracks the number of available permits. If the semaphore does
+//! not contain the required number of permits, the task attempting to acquire
+//! permits places its waker at the end of a queue. When new permits are made
+//! available (such as by releasing an initial acquisition), they are assigned
+//! to the task at the front of the queue, waking that task if its requested
+//! number of permits is met.
+//!
+//! Because waiters are enqueued at the back of the linked list and dequeued
+//! from the front, the semaphore is fair. Tasks trying to acquire large numbers
+//! of permits at a time will always be woken eventually, even if many other
+//! tasks are acquiring smaller numbers of permits. This means that in a
+//! use-case like tokio's read-write lock, writers will not be starved by
+//! readers.
+use crate::loom::cell::CausalCell;
+use crate::loom::sync::{atomic::AtomicUsize, Mutex, MutexGuard};
+use crate::util::linked_list::{self, LinkedList};
+
+use std::future::Future;
+use std::marker::PhantomPinned;
+use std::pin::Pin;
+use std::ptr::NonNull;
+use std::sync::atomic::Ordering::*;
+use std::task::Poll::*;
+use std::task::{Context, Poll, Waker};
+use std::{cmp, fmt};
+
+/// An asynchronous counting semaphore which permits waiting on multiple permits at once.
+pub(crate) struct Semaphore {
+    waiters: Mutex<Waitlist>,
+    /// The current number of available permits in the semaphore.
+    permits: AtomicUsize,
+}
+
+struct Waitlist {
+    queue: LinkedList<Waiter>,
+    closed: bool,
+}
+
+/// Error returned by `Semaphore::try_acquire`.
+#[derive(Debug)]
+pub(crate) enum TryAcquireError {
+    Closed,
+    NoPermits,
+}
+/// Error returned by `Semaphore::acquire`.
+#[derive(Debug)]
+pub(crate) struct AcquireError(());
+
+pub(crate) struct Acquire<'a> {
+    node: Waiter,
+    semaphore: &'a Semaphore,
+    num_permits: u16,
+    queued: bool,
+}
+
+/// An entry in the wait queue.
+struct Waiter {
+    /// The current state of the waiter.
+    ///
+    /// This is either the number of remaining permits required by
+    /// the waiter, or a flag indicating that the waiter is not yet queued.
+    state: AtomicUsize,
+
+    /// The waker to notify the task awaiting permits.
+    ///
+    /// # Safety
+    ///
+    /// This may only be accessed while the wait queue is locked.
+    waker: CausalCell<Option<Waker>>,
+
+    /// Intrusive linked-list pointers.
+    ///
+    /// # Safety
+    ///
+    /// This may only be accessed while the wait queue is locked.
+    ///
+    /// TODO: Ideally, we would be able to use loom to enforce that
+    /// this isn't accessed concurrently. However, it is difficult to
+    /// use a `CausalCell` here, since the `Link` trait requires _returning_
+    /// references to `Pointers`, and `CausalCell` requires that checked access
+    /// take place inside a closure. We should consider changing `Pointers` to
+    /// use `CausalCell` internally.
+    pointers: linked_list::Pointers<Waiter>,
+
+    /// Should not be `Unpin`.
+    _p: PhantomPinned,
+}
+
+impl Semaphore {
+    /// The maximum number of permits which a semaphore can hold.
+    ///
+    /// Note that this reserves three bits of flags in the permit counter, but
+    /// we only actually use one of them. However, the previous semaphore
+    /// implementation used three bits, so we will continue to reserve them to
+    /// avoid a breaking change if additional flags need to be aadded in the
+    /// future.
+    pub(crate) const MAX_PERMITS: usize = std::usize::MAX >> 3;
+    const CLOSED: usize = 1;
+    const PERMIT_SHIFT: usize = 1;
+
+    /// Creates a new semaphore with the initial number of permits
+    pub(crate) fn new(permits: usize) -> Self {
+        assert!(
+            permits <= Self::MAX_PERMITS,
+            "a semaphore may not have more than MAX_PERMITS permits ({})",
+            Self::MAX_PERMITS
+        );
+        Self {
+            permits: AtomicUsize::new(permits << Self::PERMIT_SHIFT),
+            waiters: Mutex::new(Waitlist {
+                queue: LinkedList::new(),
+                closed: false,
+            }),
+        }
+    }
+
+    /// Returns the current number of available permits
+    pub(crate) fn available_permits(&self) -> usize {
+        self.permits.load(Acquire) >> Self::PERMIT_SHIFT
+    }
+
+    /// Adds `n` new permits to the semaphore.
+    pub(crate) fn release(&self, added: usize) {
+        if added == 0 {
+            return;
+        }
+
+        // Assign permits to the wait queue, returning a list containing all the
+        // waiters at the back of the queue that received enough permits to wake
+        // up.
+        let notified = self.add_permits_locked(added, self.waiters.lock().unwrap());
+
+        // Once we release the lock, notify all woken waiters.
+        notify_all(notified);
+    }
+
+    /// Closes the semaphore. This prevents the semaphore from issuing new
+    /// permits and notifies all pending waiters.
+    // This will be used once the bounded MPSC is updated to use the new
+    // semaphore implementation.
+    #[allow(dead_code)]
+    pub(crate) fn close(&self) {
+        let notified = {
+            let mut waiters = self.waiters.lock().unwrap();
+            // If the semaphore's permits counter has enough permits for an
+            // unqueued waiter to acquire all the permits it needs immediately,
+            // it won't touch the wait list. Therefore, we have to set a bit on
+            // the permit counter as well. However, we must do this while
+            // holding the lock --- otherwise, if we set the bit and then wait
+            // to acquire the lock we'll enter an inconsistent state where the
+            // permit counter is closed, but the wait list is not.
+            self.permits.fetch_or(Self::CLOSED, Release);
+            waiters.closed = true;
+            waiters.queue.take_all()
+        };
+        notify_all(notified)
+    }
+
+    pub(crate) fn try_acquire(&self, num_permits: u16) -> Result<(), TryAcquireError> {
+        let mut curr = self.permits.load(Acquire);
+        let num_permits = (num_permits as usize) << Self::PERMIT_SHIFT;
+        loop {
+            // Has the semaphore closed?git
+            if curr & Self::CLOSED > 0 {
+                return Err(TryAcquireError::Closed);
+            }
+
+            // Are there enough permits remaining?
+            if curr < num_permits {
+                return Err(TryAcquireError::NoPermits);
+            }
+
+            let next = curr - num_permits;
+
+            match self.permits.compare_exchange(curr, next, AcqRel, Acquire) {
+                Ok(_) => return Ok(()),
+                Err(actual) => curr = actual,
+            }
+        }
+    }
+
+    pub(crate) fn acquire(&self, num_permits: u16) -> Acquire<'_> {
+        Acquire::new(self, num_permits)
+    }
+
+    /// Release `rem` permits to the semaphore's wait list, starting from the
+    /// end of the queue.
+    ///
+    /// This returns a new `LinkedList` containing all the waiters that received
+    /// enough permits to be notified. Once the lock on the wait list is
+    /// released, this list should be drained and the waiters in it notified.
+    ///  
+    /// If `rem` exceeds the number of permits needed by the wait list, the
+    /// remainder are assigned back to the semaphore.
+    fn add_permits_locked(
+        &self,
+        mut rem: usize,
+        mut waiters: MutexGuard<'_, Waitlist>,
+    ) -> LinkedList<Waiter> {
+        // Starting from the back of the wait queue, assign each waiter as many
+        // permits as it needs until we run out of permits to assign.
+        let mut last = None;
+        for waiter in waiters.queue.iter().rev() {
+            // Was the waiter assigned enough permits to wake it?
+            if !waiter.assign_permits(&mut rem) {
+                break;
+            }
+            last = Some(NonNull::from(waiter));
+        }
+
+        // If we assigned permits to all the waiters in the queue, and there are
+        // still permits left over, assign them back to the semaphore.
+        if rem > 0 {
+            let permits = rem << Self::PERMIT_SHIFT;
+            assert!(
+                permits < Self::MAX_PERMITS,
+                "cannot add more than MAX_PERMITS permits ({})",
+                Self::MAX_PERMITS
+            );
+            let prev = self.permits.fetch_add(rem << Self::PERMIT_SHIFT, Release);
+            assert!(
+                prev + permits <= Self::MAX_PERMITS,
+                "number of added permits ({}) would overflow MAX_PERMITS ({})",
+                rem,
+                Self::MAX_PERMITS
+            );
+        }
+
+        // Split off the queue at the last waiter that was satisfied, creating a
+        // new list. Once we release the lock, we'll drain this list and notify
+        // the waiters in it.
+        if let Some(waiter) = last {
+            // Safety: it's only safe to call `split_back` with a pointer to a
+            // node in the same list as the one we call `split_back` on. Since
+            // we got the waiter pointer from the list's iterator, this is fine.
+            unsafe { waiters.queue.split_back(waiter) }
+        } else {
+            LinkedList::new()
+        }
+    }
+
+    fn poll_acquire(
+        &self,
+        cx: &mut Context<'_>,
+        num_permits: u16,
+        node: Pin<&mut Waiter>,
+        queued: bool,
+    ) -> Poll<Result<(), AcquireError>> {
+        let mut acquired = 0;
+
+        let needed = if queued {
+            node.state.load(Acquire) << Self::PERMIT_SHIFT
+        } else {
+            (num_permits as usize) << Self::PERMIT_SHIFT
+        };
+
+        let mut lock = None;
+        // First, try to take the requested number of permits from the
+        // semaphore.
+        let mut curr = self.permits.load(Acquire);
+        let mut waiters = loop {
+            // Has the semaphore closed?
+            if curr & Self::CLOSED > 0 {
+                return Ready(Err(AcquireError::closed()));
+            }
+
+            let mut remaining = 0;
+            let total = curr
+                .checked_add(acquired)
+                .expect("number of permits must not overflow");
+            let (next, acq) = if total >= needed {
+                let next = curr - (needed - acquired);
+                (next, needed >> Self::PERMIT_SHIFT)
+            } else {
+                remaining = (needed - acquired) - curr;
+                (0, curr >> Self::PERMIT_SHIFT)
+            };
+
+            if remaining > 0 && lock.is_none() {
+                // No permits were immediately available, so this permit will
+                // (probably) need to wait. We'll need to acquire a lock on the
+                // wait queue before continuing. We need to do this _before_ the
+                // CAS that sets the new value of the semaphore's `permits`
+                // counter. Otherwise, if we subtract the permits and then
+                // acquire the lock, we might miss additional permits being
+                // added while waiting for the lock.
+                lock = Some(self.waiters.lock().unwrap());
+            }
+
+            match self.permits.compare_exchange(curr, next, AcqRel, Acquire) {
+                Ok(_) => {
+                    acquired += acq;
+                    if remaining == 0 {
+                        if !queued {
+                            return Ready(Ok(()));
+                        } else if lock.is_none() {
+                            break self.waiters.lock().unwrap();
+                        }
+                    }
+                    break lock.expect("lock must be acquired before waiting");
+                }
+                Err(actual) => curr = actual,
+            }
+        };
+
+        if waiters.closed {
+            return Ready(Err(AcquireError::closed()));
+        }
+
+        if node.assign_permits(&mut acquired) {
+            self.add_permits_locked(acquired, waiters);
+            return Ready(Ok(()));
+        }
+
+        assert_eq!(acquired, 0);
+
+        // Otherwise, register the waker & enqueue the node.
+        node.waker.with_mut(|waker| {
+            // Safety: the wait list is locked, so we may modify the waker.
+            let waker = unsafe { &mut *waker };
+            // Do we need to register the new waker?
+            if waker
+                .as_ref()
+                .map(|waker| !waker.will_wake(cx.waker()))
+                .unwrap_or(true)
+            {
+                *waker = Some(cx.waker().clone());
+            }
+        });
+
+        // If the waiter is not already in the wait queue, enqueue it.
+        if !queued {
+            let node = unsafe {
+                let node = Pin::into_inner_unchecked(node) as *mut _;
+                NonNull::new_unchecked(node)
+            };
+
+            waiters.queue.push_front(node);
+        }
+
+        Pending
+    }
+}
+
+impl fmt::Debug for Semaphore {
+    fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
+        fmt.debug_struct("Semaphore")
+            .field("permits", &self.permits.load(Relaxed))
+            .finish()
+    }
+}
+
+/// Pop all waiters from `list`, starting at the end of the queue, and notify
+/// them.
+fn notify_all(mut list: LinkedList<Waiter>) {
+    while let Some(waiter) = list.pop_back() {
+        let waker = unsafe { waiter.as_ref().waker.with_mut(|waker| (*waker).take()) };
+
+        waker
+            .expect("if a node is in the wait list, it must have a waker")
+            .wake();
+    }
+}
+
+impl Waiter {
+    fn new(num_permits: u16) -> Self {
+        Waiter {
+            waker: CausalCell::new(None),
+            state: AtomicUsize::new(num_permits as usize),
+            pointers: linked_list::Pointers::new(),
+            _p: PhantomPinned,
+        }
+    }
+
+    /// Assign permits to the waiter.
+    ///
+    /// Returns `true` if the waiter should be removed from the queue
+    fn assign_permits(&self, n: &mut usize) -> bool {
+        let mut curr = self.state.load(Acquire);
+        loop {
+            let assign = cmp::min(curr, *n);
+            let next = curr - assign;
+            match self.state.compare_exchange(curr, next, AcqRel, Acquire) {
+                Ok(_) => {
+                    *n -= assign;
+                    return next == 0;
+                }
+                Err(actual) => curr = actual,
+            }
+        }
+    }
+}
+
+impl Future for Acquire<'_> {
+    type Output = Result<(), AcquireError>;
+
+    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        let (node, semaphore, needed, queued) = self.project();
+        match semaphore.poll_acquire(cx, needed, node, *queued) {
+            Pending => {
+                *queued = true;
+                Pending
+            }
+            Ready(r) => {
+                r?;
+                *queued = false;
+                Ready(Ok(()))
+            }
+        }
+    }
+}
+
+impl<'a> Acquire<'a> {
+    fn new(semaphore: &'a Semaphore, num_permits: u16) -> Self {
+        Self {
+            node: Waiter::new(num_permits),
+            semaphore,
+            num_permits,
+            queued: false,
+        }
+    }
+
+    fn project(self: Pin<&mut Self>) -> (Pin<&mut Waiter>, &Semaphore, u16, &mut bool) {
+        fn is_unpin<T: Unpin>() {}
+        unsafe {
+            // Safety: all fields other than `node` are `Unpin`
+
+            is_unpin::<&Semaphore>();
+            is_unpin::<&mut bool>();
+            is_unpin::<u16>();
+
+            let this = self.get_unchecked_mut();
+            (
+                Pin::new_unchecked(&mut this.node),
+                &this.semaphore,
+                this.num_permits,
+                &mut this.queued,
+            )
+        }
+    }
+}
+
+impl Drop for Acquire<'_> {
+    fn drop(&mut self) {
+        // If the future is completed, there is no node in the wait list, so we
+        // can skip acquiring the lock.
+        if !self.queued {
+            return;
+        }
+
+        // This is where we ensure safety. The future is being dropped,
+        // which means we must ensure that the waiter entry is no longer stored
+        // in the linked list.
+        let mut waiters = match self.semaphore.waiters.lock() {
+            Ok(lock) => lock,
+            // Removing the node from the linked list is necessary to ensure
+            // safety. Even if the lock was poisoned, we need to make sure it is
+            // removed from the linked list before dropping it --- otherwise,
+            // the list will contain a dangling pointer to this node.
+            Err(e) => e.into_inner(),
+        };
+
+        // remove the entry from the list
+        let node = NonNull::from(&mut self.node);
+        // Safety: we have locked the wait list.
+        unsafe { waiters.queue.remove(node) };
+
+        let acquired_permits = self.num_permits as usize - self.node.state.load(Acquire);
+        if acquired_permits > 0 {
+            let notified = self.semaphore.add_permits_locked(acquired_permits, waiters);
+            notify_all(notified);
+        }
+    }
+}
+
+// ===== impl AcquireError ====
+
+impl AcquireError {
+    fn closed() -> AcquireError {
+        AcquireError(())
+    }
+}
+
+impl fmt::Display for AcquireError {
+    fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(fmt, "semaphore closed")
+    }
+}
+
+impl std::error::Error for AcquireError {}
+
+// ===== impl TryAcquireError =====
+
+impl TryAcquireError {
+    /// Returns `true` if the error was caused by a closed semaphore.
+    #[allow(dead_code)] // may be used later!
+    pub(crate) fn is_closed(&self) -> bool {
+        match self {
+            TryAcquireError::Closed => true,
+            _ => false,
+        }
+    }
+
+    /// Returns `true` if the error was caused by calling `try_acquire` on a
+    /// semaphore with no available permits.
+    #[allow(dead_code)] // may be used later!
+    pub(crate) fn is_no_permits(&self) -> bool {
+        match self {
+            TryAcquireError::NoPermits => true,
+            _ => false,
+        }
+    }
+}
+
+impl fmt::Display for TryAcquireError {
+    fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            TryAcquireError::Closed => write!(fmt, "{}", "semaphore closed"),
+            TryAcquireError::NoPermits => write!(fmt, "{}", "no permits available"),
+        }
+    }
+}
+
+impl std::error::Error for TryAcquireError {}
+
+/// # Safety
+///
+/// `Waiter` is forced to be !Unpin.
+unsafe impl linked_list::Link for Waiter {
+    // XXX: ideally, we would be able to use `Pin` here, to enforce the
+    // invariant that list entries may not move while in the list. However, we
+    // can't do this currently, as using `Pin<&'a mut Waiter>` as the `Handle`
+    // type would require `Semaphore` to be generic over a lifetime. We can't
+    // use `Pin<*mut Waiter>`, as raw pointers are `Unpin` regardless of whether
+    // or not they dereference to an `!Unpin` target.
+    type Handle = NonNull<Waiter>;
+    type Target = Waiter;
+
+    fn as_raw(handle: &Self::Handle) -> NonNull<Waiter> {
+        *handle
+    }
+
+    unsafe fn from_raw(ptr: NonNull<Waiter>) -> NonNull<Waiter> {
+        ptr
+    }
+
+    unsafe fn pointers(mut target: NonNull<Waiter>) -> NonNull<linked_list::Pointers<Waiter>> {
+        NonNull::from(&mut target.as_mut().pointers)
+    }
+}