runtime: reduce the lock contention in task spawn (#6001)

Author: wathenjiang

benches/Cargo.toml

@@ -12,6 +12,7 @@ tokio = { version = "1.5.0", path = "../tokio", features = ["full"] }
 criterion = "0.5.1"
 rand = "0.8"
 rand_chacha = "0.3"
+num_cpus = "1.16.0"
 
 [dev-dependencies]
 tokio-util = { version = "0.7.0", path = "../tokio-util", features = ["full"] }

tokio/src/runtime/builder.rs

@@ -1279,7 +1279,6 @@ cfg_rt_multi_thread! {
                 use crate::runtime::scheduler::MultiThreadAlt;
 
                 let core_threads = self.worker_threads.unwrap_or_else(num_cpus);
-
                 let (driver, driver_handle) = driver::Driver::new(self.get_cfg())?;
 
                 // Create the blocking pool

tokio/src/runtime/id.rs

@@ -1,5 +1,5 @@
 use std::fmt;
-use std::num::NonZeroU64;
+use std::num::{NonZeroU32, NonZeroU64};
 
 /// An opaque ID that uniquely identifies a runtime relative to all other currently
 /// running runtimes.
@@ -39,6 +39,12 @@ impl From<NonZeroU64> for Id {
     }
 }
 
+impl From<NonZeroU32> for Id {
+    fn from(value: NonZeroU32) -> Self {
+        Id(value.into())
+    }
+}
+
 impl fmt::Display for Id {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         self.0.fmt(f)

tokio/src/runtime/scheduler/current_thread/mod.rs

@@ -132,7 +132,7 @@ impl CurrentThread {
         let handle = Arc::new(Handle {
             shared: Shared {
                 inject: Inject::new(),
-                owned: OwnedTasks::new(),
+                owned: OwnedTasks::new(1),
                 woken: AtomicBool::new(false),
                 config,
                 scheduler_metrics: SchedulerMetrics::new(),
@@ -248,7 +248,7 @@ fn shutdown2(mut core: Box<Core>, handle: &Handle) -> Box<Core> {
     // Drain the OwnedTasks collection. This call also closes the
     // collection, ensuring that no tasks are ever pushed after this
     // call returns.
-    handle.shared.owned.close_and_shutdown_all();
+    handle.shared.owned.close_and_shutdown_all(0);
 
     // Drain local queue
     // We already shut down every task, so we just need to drop the task.
@@ -614,7 +614,7 @@ impl Schedule for Arc<Handle> {
                             // If `None`, the runtime is shutting down, so there is no need to signal shutdown
                             if let Some(core) = core.as_mut() {
                                 core.unhandled_panic = true;
-                                self.shared.owned.close_and_shutdown_all();
+                                self.shared.owned.close_and_shutdown_all(0);
                             }
                         }
                         _ => unreachable!("runtime core not set in CURRENT thread-local"),

tokio/src/runtime/scheduler/multi_thread/worker.rs

@@ -287,7 +287,7 @@ pub(super) fn create(
             remotes: remotes.into_boxed_slice(),
             inject,
             idle,
-            owned: OwnedTasks::new(),
+            owned: OwnedTasks::new(size),
             synced: Mutex::new(Synced {
                 idle: idle_synced,
                 inject: inject_synced,
@@ -548,7 +548,6 @@ impl Context {
         }
 
         core.pre_shutdown(&self.worker);
-
         // Signal shutdown
         self.worker.handle.shutdown_core(core);
         Err(())
@@ -955,8 +954,16 @@ impl Core {
     /// Signals all tasks to shut down, and waits for them to complete. Must run
     /// before we enter the single-threaded phase of shutdown processing.
     fn pre_shutdown(&mut self, worker: &Worker) {
+        // Start from a random inner list
+        let start = self
+            .rand
+            .fastrand_n(worker.handle.shared.owned.get_shard_size() as u32);
         // Signal to all tasks to shut down.
-        worker.handle.shared.owned.close_and_shutdown_all();
+        worker
+            .handle
+            .shared
+            .owned
+            .close_and_shutdown_all(start as usize);
 
         self.stats
             .submit(&worker.handle.shared.worker_metrics[worker.index]);

tokio/src/runtime/scheduler/multi_thread_alt/worker.rs

@@ -307,7 +307,7 @@ pub(super) fn create(
             remotes: remotes.into_boxed_slice(),
             inject,
             idle,
-            owned: OwnedTasks::new(),
+            owned: OwnedTasks::new(num_cores),
             synced: Mutex::new(Synced {
                 assigned_cores: (0..num_workers).map(|_| None).collect(),
                 shutdown_cores: Vec::with_capacity(num_cores),
@@ -1460,7 +1460,9 @@ impl Shared {
     }
 
     pub(super) fn shutdown_core(&self, handle: &Handle, mut core: Box<Core>) {
-        self.owned.close_and_shutdown_all();
+        // Start from a random inner list
+        let start = core.rand.fastrand_n(self.owned.get_shard_size() as u32);
+        self.owned.close_and_shutdown_all(start as usize);
 
         core.stats.submit(&self.worker_metrics[core.index]);
 

tokio/src/runtime/task/id.rs

@@ -24,7 +24,7 @@ use std::fmt;
 #[cfg_attr(docsrs, doc(cfg(all(feature = "rt", tokio_unstable))))]
 #[cfg_attr(not(tokio_unstable), allow(unreachable_pub))]
 #[derive(Clone, Copy, Debug, Hash, Eq, PartialEq)]
-pub struct Id(u64);
+pub struct Id(pub(crate) u64);
 
 /// Returns the [`Id`] of the currently running task.
 ///
@@ -74,11 +74,22 @@ impl fmt::Display for Id {
 
 impl Id {
     pub(crate) fn next() -> Self {
-        use crate::loom::sync::atomic::{Ordering::Relaxed, StaticAtomicU64};
+        use crate::loom::sync::atomic::Ordering::Relaxed;
+        use crate::loom::sync::atomic::StaticAtomicU64;
 
-        static NEXT_ID: StaticAtomicU64 = StaticAtomicU64::new(1);
+        #[cfg(all(test, loom))]
+        {
+            crate::loom::lazy_static! {
+                static ref NEXT_ID: StaticAtomicU64 = StaticAtomicU64::new(1);
+            }
+            Self(NEXT_ID.fetch_add(1, Relaxed))
+        }
 
-        Self(NEXT_ID.fetch_add(1, Relaxed))
+        #[cfg(not(all(test, loom)))]
+        {
+            static NEXT_ID: StaticAtomicU64 = StaticAtomicU64::new(1);
+            Self(NEXT_ID.fetch_add(1, Relaxed))
+        }
     }
 
     pub(crate) fn as_u64(&self) -> u64 {

tokio/src/runtime/task/list.rs

@@ -8,10 +8,11 @@
 
 use crate::future::Future;
 use crate::loom::cell::UnsafeCell;
-use crate::loom::sync::Mutex;
 use crate::runtime::task::{JoinHandle, LocalNotified, Notified, Schedule, Task};
-use crate::util::linked_list::{CountedLinkedList, Link, LinkedList};
+use crate::util::linked_list::{Link, LinkedList};
+use crate::util::sharded_list;
 
+use crate::loom::sync::atomic::{AtomicBool, Ordering};
 use std::marker::PhantomData;
 use std::num::NonZeroU64;
 
@@ -25,7 +26,7 @@ use std::num::NonZeroU64;
 // mixed up runtimes happen to have the same id.
 
 cfg_has_atomic_u64! {
-    use std::sync::atomic::{AtomicU64, Ordering};
+    use std::sync::atomic::AtomicU64;
 
     static NEXT_OWNED_TASKS_ID: AtomicU64 = AtomicU64::new(1);
 
@@ -40,7 +41,7 @@ cfg_has_atomic_u64! {
 }
 
 cfg_not_has_atomic_u64! {
-    use std::sync::atomic::{AtomicU32, Ordering};
+    use std::sync::atomic::AtomicU32;
 
     static NEXT_OWNED_TASKS_ID: AtomicU32 = AtomicU32::new(1);
 
@@ -55,30 +56,30 @@ cfg_not_has_atomic_u64! {
 }
 
 pub(crate) struct OwnedTasks<S: 'static> {
-    inner: Mutex<CountedOwnedTasksInner<S>>,
+    list: List<S>,
     pub(crate) id: NonZeroU64,
+    closed: AtomicBool,
 }
-struct CountedOwnedTasksInner<S: 'static> {
-    list: CountedLinkedList<Task<S>, <Task<S> as Link>::Target>,
-    closed: bool,
-}
+
+type List<S> = sharded_list::ShardedList<Task<S>, <Task<S> as Link>::Target>;
+
 pub(crate) struct LocalOwnedTasks<S: 'static> {
     inner: UnsafeCell<OwnedTasksInner<S>>,
     pub(crate) id: NonZeroU64,
     _not_send_or_sync: PhantomData<*const ()>,
 }
+
 struct OwnedTasksInner<S: 'static> {
     list: LinkedList<Task<S>, <Task<S> as Link>::Target>,
     closed: bool,
 }
 
 impl<S: 'static> OwnedTasks<S> {
-    pub(crate) fn new() -> Self {
+    pub(crate) fn new(num_cores: usize) -> Self {
+        let shard_size = Self::gen_shared_list_size(num_cores);
         Self {
-            inner: Mutex::new(CountedOwnedTasksInner {
-                list: CountedLinkedList::new(),
-                closed: false,
-            }),
+            list: List::new(shard_size),
+            closed: AtomicBool::new(false),
             id: get_next_id(),
         }
     }
@@ -112,24 +113,23 @@ impl<S: 'static> OwnedTasks<S> {
             task.header().set_owner_id(self.id);
         }
 
-        let mut lock = self.inner.lock();
-        if lock.closed {
-            drop(lock);
-            drop(notified);
+        let shard = self.list.lock_shard(&task);
+        // Check the closed flag in the lock for ensuring all that tasks
+        // will shut down after the OwnedTasks has been closed.
+        if self.closed.load(Ordering::Acquire) {
+            drop(shard);
             task.shutdown();
-            None
-        } else {
-            lock.list.push_front(task);
-            Some(notified)
+            return None;
         }
+        shard.push(task);
+        Some(notified)
     }
 
     /// Asserts that the given task is owned by this OwnedTasks and convert it to
     /// a LocalNotified, giving the thread permission to poll this task.
     #[inline]
     pub(crate) fn assert_owner(&self, task: Notified<S>) -> LocalNotified<S> {
         debug_assert_eq!(task.header().get_owner_id(), Some(self.id));
-
         // safety: All tasks bound to this OwnedTasks are Send, so it is safe
         // to poll it on this thread no matter what thread we are on.
         LocalNotified {
@@ -140,34 +140,34 @@ impl<S: 'static> OwnedTasks<S> {
 
     /// Shuts down all tasks in the collection. This call also closes the
     /// collection, preventing new items from being added.
-    pub(crate) fn close_and_shutdown_all(&self)
+    ///
+    /// The parameter start determines which shard this method will start at.
+    /// Using different values for each worker thread reduces contention.
+    pub(crate) fn close_and_shutdown_all(&self, start: usize)
     where
         S: Schedule,
     {
-        // The first iteration of the loop was unrolled so it can set the
-        // closed bool.
-        let first_task = {
-            let mut lock = self.inner.lock();
-            lock.closed = true;
-            lock.list.pop_back()
-        };
-        match first_task {
-            Some(task) => task.shutdown(),
-            None => return,
+        self.closed.store(true, Ordering::Release);
+        for i in start..self.get_shard_size() + start {
+            loop {
+                let task = self.list.pop_back(i);
+                match task {
+                    Some(task) => {
+                        task.shutdown();
+                    }
+                    None => break,
+                }
+            }
         }
+    }
 
-        loop {
-            let task = match self.inner.lock().list.pop_back() {
-                Some(task) => task,
-                None => return,
-            };
-
-            task.shutdown();
-        }
+    #[inline]
+    pub(crate) fn get_shard_size(&self) -> usize {
+        self.list.shard_size()
     }
 
     pub(crate) fn active_tasks_count(&self) -> usize {
-        self.inner.lock().list.count()
+        self.list.len()
     }
 
     pub(crate) fn remove(&self, task: &Task<S>) -> Option<Task<S>> {
@@ -179,11 +179,27 @@ impl<S: 'static> OwnedTasks<S> {
 
         // safety: We just checked that the provided task is not in some other
         // linked list.
-        unsafe { self.inner.lock().list.remove(task.header_ptr()) }
+        unsafe { self.list.remove(task.header_ptr()) }
     }
 
     pub(crate) fn is_empty(&self) -> bool {
-        self.inner.lock().list.is_empty()
+        self.list.is_empty()
+    }
+
+    /// Generates the size of the sharded list based on the number of worker threads.
+    ///
+    /// The sharded lock design can effectively alleviate
+    /// lock contention performance problems caused by high concurrency.
+    ///
+    /// However, as the number of shards increases, the memory continuity between
+    /// nodes in the intrusive linked list will diminish. Furthermore,
+    /// the construction time of the sharded list will also increase with a higher number of shards.
+    ///
+    /// Due to the above reasons, we set a maximum value for the shared list size,
+    /// denoted as `MAX_SHARED_LIST_SIZE`.
+    fn gen_shared_list_size(num_cores: usize) -> usize {
+        const MAX_SHARED_LIST_SIZE: usize = 1 << 16;
+        usize::min(MAX_SHARED_LIST_SIZE, num_cores.next_power_of_two() * 4)
     }
 }
 
@@ -192,9 +208,9 @@ cfg_taskdump! {
         /// Locks the tasks, and calls `f` on an iterator over them.
         pub(crate) fn for_each<F>(&self, f: F)
         where
-            F: FnMut(&Task<S>)
+            F: FnMut(&Task<S>),
         {
-            self.inner.lock().list.for_each(f)
+            self.list.for_each(f);
         }
     }
 }

tokio/src/runtime/task/mod.rs

@@ -208,6 +208,7 @@ cfg_taskdump! {
 
 use crate::future::Future;
 use crate::util::linked_list;
+use crate::util::sharded_list;
 
 use std::marker::PhantomData;
 use std::ptr::NonNull;
@@ -503,3 +504,16 @@ unsafe impl<S> linked_list::Link for Task<S> {
         self::core::Trailer::addr_of_owned(Header::get_trailer(target))
     }
 }
+
+/// # Safety
+///
+/// The id of a task is never changed after creation of the task, so the return value of
+/// `get_shard_id` will not change. (The cast may throw away the upper 32 bits of the task id, but
+/// the shard id still won't change from call to call.)
+unsafe impl<S> sharded_list::ShardedListItem for Task<S> {
+    unsafe fn get_shard_id(target: NonNull<Self::Target>) -> usize {
+        // SAFETY: The caller guarantees that `target` points at a valid task.
+        let task_id = unsafe { Header::get_id(target) };
+        task_id.0 as usize
+    }
+}