Struct rayon_core::registry::Registry

pub(crate) struct Registry {
    thread_infos: Vec<ThreadInfo>,
    sleep: Sleep,
    injected_jobs: Injector<JobRef>,
    broadcasts: Mutex<Vec<Worker<JobRef>>>,
    panic_handler: Option<Box<dyn Fn(Box<dyn Any + Send>) + Send + Sync>>,
    start_handler: Option<Box<dyn Fn(usize) + Send + Sync>>,
    exit_handler: Option<Box<dyn Fn(usize) + Send + Sync>>,
    terminate_count: AtomicUsize,
}

Fields

thread_infos: Vec<ThreadInfo>

sleep: Sleep

injected_jobs: Injector<JobRef>

broadcasts: Mutex<Vec<Worker<JobRef>>>

panic_handler: Option<Box<dyn Fn(Box<dyn Any + Send>) + Send + Sync>>

start_handler: Option<Box<dyn Fn(usize) + Send + Sync>>

exit_handler: Option<Box<dyn Fn(usize) + Send + Sync>>

terminate_count: AtomicUsize

Implementations

impl Registry

pub(crate) fn new<S>( builder: ThreadPoolBuilder<S>, ) -> Result<Arc<Self>, ThreadPoolBuildError>

where S: ThreadSpawn,

pub(crate) fn current() -> Arc<Registry>

pub(crate) fn current_num_threads() -> usize

Returns the number of threads in the current registry. This is better than Registry::current().num_threads() because it avoids incrementing the Arc.

pub(crate) fn current_thread(&self) -> Option<&WorkerThread>

Returns the current WorkerThread if it’s part of this Registry.

pub(crate) fn id(&self) -> RegistryId

Returns an opaque identifier for this registry.

pub(crate) fn num_threads(&self) -> usize

pub(crate) fn catch_unwind(&self, f: impl FnOnce())

pub(crate) fn wait_until_primed(&self)

Waits for the worker threads to get up and running. This is meant to be used for benchmarking purposes, primarily, so that you can get more consistent numbers by having everything “ready to go”.

pub(crate) fn inject_or_push(&self, job_ref: JobRef)

//////////////////////////////////////////////////////////////////////// MAIN LOOP

So long as all of the worker threads are hanging out in their top-level loop, there is no work to be done. Push a job into the given registry. If we are running on a worker thread for the registry, this will push onto the deque. Else, it will inject from the outside (which is slower).

pub(crate) fn inject(&self, injected_job: JobRef)

Push a job into the “external jobs” queue; it will be taken by whatever worker has nothing to do. Use this if you know that you are not on a worker of this registry.

fn has_injected_job(&self) -> bool

fn pop_injected_job(&self) -> Option<JobRef>

pub(crate) fn inject_broadcast( &self, injected_jobs: impl ExactSizeIterator<Item = JobRef>, )

Push a job into each thread’s own “external jobs” queue; it will be executed only on that thread, when it has nothing else to do locally, before it tries to steal other work.

Panics if not given exactly as many jobs as there are threads.

pub(crate) fn in_worker<OP, R>(&self, op: OP) -> R

where OP: FnOnce(&WorkerThread, bool) -> R + Send, R: Send,

If already in a worker-thread of this registry, just execute op. Otherwise, inject op in this thread-pool. Either way, block until op completes and return its return value. If op panics, that panic will be propagated as well. The second argument indicates true if injection was performed, false if executed directly.

unsafe fn in_worker_cold<OP, R>(&self, op: OP) -> R

where OP: FnOnce(&WorkerThread, bool) -> R + Send, R: Send,

unsafe fn in_worker_cross<OP, R>( &self, current_thread: &WorkerThread, op: OP, ) -> R

where OP: FnOnce(&WorkerThread, bool) -> R + Send, R: Send,

pub(crate) fn increment_terminate_count(&self)

Increments the terminate counter. This increment should be balanced by a call to terminate, which will decrement. This is used when spawning asynchronous work, which needs to prevent the registry from terminating so long as it is active.

Note that blocking functions such as join and scope do not need to concern themselves with this fn; their context is responsible for ensuring the current thread-pool will not terminate until they return.

The global thread-pool always has an outstanding reference (the initial one). Custom thread-pools have one outstanding reference that is dropped when the ThreadPool is dropped: since installing the thread-pool blocks until any joins/scopes complete, this ensures that joins/scopes are covered.

The exception is ::spawn(), which can create a job outside of any blocking scope. In that case, the job itself holds a terminate count and is responsible for invoking terminate() when finished.

pub(crate) fn terminate(&self)

Signals that the thread-pool which owns this registry has been dropped. The worker threads will gradually terminate, once any extant work is completed.

pub(crate) fn notify_worker_latch_is_set(&self, target_worker_index: usize)

Notify the worker that the latch they are sleeping on has been “set”.