1 // Copyright 2013 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
11 //! The "green scheduling" library
13 //! This library provides M:N threading for rust programs. Internally this has
14 //! the implementation of a green scheduler along with context switching and a
15 //! stack-allocation strategy. This can be optionally linked in to rust
16 //! programs in order to provide M:N functionality inside of 1:1 programs.
20 //! An M:N scheduling library implies that there are N OS thread upon which M
21 //! "green threads" are multiplexed. In other words, a set of green threads are
22 //! all run inside a pool of OS threads.
24 //! With this design, you can achieve _concurrency_ by spawning many green
25 //! threads, and you can achieve _parallelism_ by running the green threads
26 //! simultaneously on multiple OS threads. Each OS thread is a candidate for
27 //! being scheduled on a different core (the source of parallelism), and then
28 //! all of the green threads cooperatively schedule amongst one another (the
29 //! source of concurrency).
33 //! In order to coordinate among green threads, each OS thread is primarily
34 //! running something which we call a Scheduler. Whenever a reference to a
35 //! Scheduler is made, it is synonymous to referencing one OS thread. Each
36 //! scheduler is bound to one and exactly one OS thread, and the thread that it
37 //! is bound to never changes.
39 //! Each scheduler is connected to a pool of other schedulers (a `SchedPool`)
40 //! which is the thread pool term from above. A pool of schedulers all share the
41 //! work that they create. Furthermore, whenever a green thread is created (also
42 //! synonymously referred to as a green task), it is associated with a
43 //! `SchedPool` forevermore. A green thread cannot leave its scheduler pool.
45 //! Schedulers can have at most one green thread running on them at a time. When
46 //! a scheduler is asleep on its event loop, there are no green tasks running on
47 //! the OS thread or the scheduler. The term "context switch" is used for when
48 //! the running green thread is swapped out, but this simply changes the one
49 //! green thread which is running on the scheduler.
53 //! A green thread can largely be summarized by a stack and a register context.
54 //! Whenever a green thread is spawned, it allocates a stack, and then prepares
55 //! a register context for execution. The green task may be executed across
56 //! multiple OS threads, but it will always use the same stack and it will carry
57 //! its register context across OS threads.
59 //! Each green thread is cooperatively scheduled with other green threads.
60 //! Primarily, this means that there is no pre-emption of a green thread. The
61 //! major consequence of this design is that a green thread stuck in an infinite
62 //! loop will prevent all other green threads from running on that particular
65 //! Scheduling events for green threads occur on communication and I/O
66 //! boundaries. For example, if a green task blocks waiting for a message on a
67 //! channel some other green thread can now run on the scheduler. This also has
68 //! the consequence that until a green thread performs any form of scheduling
69 //! event, it will be running on the same OS thread (unconditionally).
73 //! With a pool of schedulers, a new green task has a number of options when
74 //! deciding where to run initially. The current implementation uses a concept
75 //! called work stealing in order to spread out work among schedulers.
77 //! In a work-stealing model, each scheduler maintains a local queue of tasks to
78 //! run, and this queue is stolen from by other schedulers. Implementation-wise,
79 //! work stealing has some hairy parts, but from a user-perspective, work
80 //! stealing simply implies what with M green threads and N schedulers where
81 //! M > N it is very likely that all schedulers will be busy executing work.
83 //! # Considerations when using libgreen
85 //! An M:N runtime has both pros and cons, and there is no one answer as to
86 //! whether M:N or 1:1 is appropriate to use. As always, there are many
87 //! advantages and disadvantages between the two. Regardless of the workload,
88 //! however, there are some aspects of using green thread which you should be
91 //! * The largest concern when using libgreen is interoperating with native
92 //! code. Care should be taken when calling native code that will block the OS
93 //! thread as it will prevent further green tasks from being scheduled on the
96 //! * Native code using thread-local-storage should be approached
97 //! with care. Green threads may migrate among OS threads at any time, so
98 //! native libraries using thread-local state may not always work.
100 //! * Native synchronization primitives (e.g. pthread mutexes) will also not
101 //! work for green threads. The reason for this is because native primitives
102 //! often operate on a _os thread_ granularity whereas green threads are
103 //! operating on a more granular unit of work.
105 //! * A green threading runtime is not fork-safe. If the process forks(), it
106 //! cannot expect to make reasonable progress by continuing to use green
109 //! Note that these concerns do not mean that operating with native code is a
110 //! lost cause. These are simply just concerns which should be considered when
111 //! invoking native code.
113 //! # Starting with libgreen
116 //! extern crate green;
119 //! fn start(argc: int, argv: *const *const u8) -> int {
120 //! green::start(argc, argv, green::basic::event_loop, main)
124 //! // this code is running in a pool of schedulers
128 //! > **Note**: This `main` function in this example does *not* have I/O
129 //! > support. The basic event loop does not provide any support
131 //! # Starting with I/O support in libgreen
134 //! extern crate green;
135 //! extern crate rustuv;
138 //! fn start(argc: int, argv: *const *const u8) -> int {
139 //! green::start(argc, argv, rustuv::event_loop, main)
143 //! // this code is running in a pool of schedulers all powered by libuv
147 //! The above code can also be shortened with a macro from libgreen.
150 //! #![feature(phase)]
151 //! #[phase(plugin)] extern crate green;
153 //! green_start!(main)
156 //! // run inside of a green pool
160 //! # Using a scheduler pool
162 //! This library adds a `GreenTaskBuilder` trait that extends the methods
163 //! available on `std::task::TaskBuilder` to allow spawning a green task,
164 //! possibly pinned to a particular scheduler thread:
167 //! use std::task::TaskBuilder;
168 //! use green::{SchedPool, PoolConfig, GreenTaskBuilder};
170 //! let config = PoolConfig::new();
171 //! let mut pool = SchedPool::new(config);
173 //! // Spawn tasks into the pool of schedulers
174 //! TaskBuilder::new().green(&mut pool).spawn(proc() {
175 //! // this code is running inside the pool of schedulers
178 //! // this code is also running inside the same scheduler pool
182 //! // Dynamically add a new scheduler to the scheduler pool. This adds another
183 //! // OS thread that green threads can be multiplexed on to.
184 //! let mut handle = pool.spawn_sched();
186 //! // Pin a task to the spawned scheduler
187 //! TaskBuilder::new().green_pinned(&mut pool, &mut handle).spawn(proc() {
191 //! // Handles keep schedulers alive, so be sure to drop all handles before
192 //! // destroying the sched pool
195 //! // Required to shut down this scheduler pool.
196 //! // The task will fail if `shutdown` is not called.
200 #![crate_id = "green#0.11.0-pre"]
202 #![license = "MIT/ASL2"]
203 #![crate_type = "rlib"]
204 #![crate_type = "dylib"]
205 #![doc(html_logo_url = "http://www.rust-lang.org/logos/rust-logo-128x128-blk-v2.png",
206 html_favicon_url = "http://www.rust-lang.org/favicon.ico",
207 html_root_url = "http://doc.rust-lang.org/",
208 html_playground_url = "http://play.rust-lang.org/")]
210 // NB this does *not* include globs, please keep it that way.
211 #![feature(macro_rules, phase)]
212 #![allow(visible_private_types)]
213 #![allow(deprecated)]
214 #![feature(default_type_params)]
216 #[cfg(test)] #[phase(plugin, link)] extern crate log;
217 #[cfg(test)] extern crate rustuv;
222 use std::mem::replace;
225 use std::rt::thread::Thread;
226 use std::rt::task::TaskOpts;
228 use std::sync::atomics::{SeqCst, AtomicUint, INIT_ATOMIC_UINT};
229 use std::sync::deque;
230 use std::task::{TaskBuilder, Spawner};
232 use sched::{Shutdown, Scheduler, SchedHandle, TaskFromFriend, PinnedTask, NewNeighbor};
233 use sleeper_list::SleeperList;
234 use stack::StackPool;
245 pub mod sleeper_list;
249 /// A helper macro for booting a program with libgreen
254 /// #![feature(phase)]
255 /// #[phase(plugin)] extern crate green;
257 /// green_start!(main)
260 /// // running with libgreen
264 macro_rules! green_start( ($f:ident) => (
270 fn start(argc: int, argv: *const *const u8) -> int {
271 green::start(argc, argv, rustuv::event_loop, super::$f)
276 /// Set up a default runtime configuration, given compiler-supplied arguments.
278 /// This function will block until the entire pool of M:N schedulers have
279 /// exited. This function also requires a local task to be available.
283 /// * `argc` & `argv` - The argument vector. On Unix this information is used
285 /// * `main` - The initial procedure to run inside of the M:N scheduling pool.
286 /// Once this procedure exits, the scheduling pool will begin to shut
287 /// down. The entire pool (and this function) will only return once
288 /// all child tasks have finished executing.
292 /// The return value is used as the process return code. 0 on success, 101 on
294 pub fn start(argc: int, argv: *const *const u8,
295 event_loop_factory: fn() -> Box<rtio::EventLoop + Send>,
296 main: proc():Send) -> int {
297 rt::init(argc, argv);
298 let mut main = Some(main);
300 simple::task().run(|| {
301 ret = Some(run(event_loop_factory, main.take_unwrap()));
303 // unsafe is ok b/c we're sure that the runtime is gone
304 unsafe { rt::cleanup() }
308 /// Execute the main function in a pool of M:N schedulers.
310 /// Configures the runtime according to the environment, by default using a task
311 /// scheduler with the same number of threads as cores. Returns a process exit
314 /// This function will not return until all schedulers in the associated pool
316 pub fn run(event_loop_factory: fn() -> Box<rtio::EventLoop + Send>,
317 main: proc():Send) -> int {
318 // Create a scheduler pool and spawn the main task into this pool. We will
319 // get notified over a channel when the main task exits.
320 let mut cfg = PoolConfig::new();
321 cfg.event_loop_factory = event_loop_factory;
322 let mut pool = SchedPool::new(cfg);
323 let (tx, rx) = channel();
324 let mut opts = TaskOpts::new();
325 opts.on_exit = Some(proc(r) tx.send(r));
326 opts.name = Some("<main>".into_maybe_owned());
327 pool.spawn(opts, main);
329 // Wait for the main task to return, and set the process error code
331 if rx.recv().is_err() {
332 os::set_exit_status(rt::DEFAULT_ERROR_CODE);
335 // Now that we're sure all tasks are dead, shut down the pool of schedulers,
336 // waiting for them all to return.
338 os::get_exit_status()
341 /// Configuration of how an M:N pool of schedulers is spawned.
342 pub struct PoolConfig {
343 /// The number of schedulers (OS threads) to spawn into this M:N pool.
345 /// A factory function used to create new event loops. If this is not
346 /// specified then the default event loop factory is used.
347 pub event_loop_factory: fn() -> Box<rtio::EventLoop + Send>,
351 /// Returns the default configuration, as determined the environment
352 /// variables of this process.
353 pub fn new() -> PoolConfig {
355 threads: rt::default_sched_threads(),
356 event_loop_factory: basic::event_loop,
361 /// A structure representing a handle to a pool of schedulers. This handle is
362 /// used to keep the pool alive and also reap the status from the pool.
363 pub struct SchedPool {
365 threads: Vec<Thread<()>>,
366 handles: Vec<SchedHandle>,
367 stealers: Vec<deque::Stealer<Box<task::GreenTask>>>,
369 stack_pool: StackPool,
370 deque_pool: deque::BufferPool<Box<task::GreenTask>>,
371 sleepers: SleeperList,
372 factory: fn() -> Box<rtio::EventLoop + Send>,
373 task_state: TaskState,
374 tasks_done: Receiver<()>,
377 /// This is an internal state shared among a pool of schedulers. This is used to
378 /// keep track of how many tasks are currently running in the pool and then
379 /// sending on a channel once the entire pool has been drained of all tasks.
382 cnt: Arc<AtomicUint>,
387 /// Execute the main function in a pool of M:N schedulers.
389 /// This will configure the pool according to the `config` parameter, and
390 /// initially run `main` inside the pool of schedulers.
391 pub fn new(config: PoolConfig) -> SchedPool {
392 static mut POOL_ID: AtomicUint = INIT_ATOMIC_UINT;
396 event_loop_factory: factory
398 assert!(nscheds > 0);
400 // The pool of schedulers that will be returned from this function
401 let (p, state) = TaskState::new();
402 let mut pool = SchedPool {
406 id: unsafe { POOL_ID.fetch_add(1, SeqCst) },
407 sleepers: SleeperList::new(),
408 stack_pool: StackPool::new(),
409 deque_pool: deque::BufferPool::new(),
416 // Create a work queue for each scheduler, ntimes. Create an extra
417 // for the main thread if that flag is set. We won't steal from it.
418 let mut workers = Vec::with_capacity(nscheds);
419 let mut stealers = Vec::with_capacity(nscheds);
421 for _ in range(0, nscheds) {
422 let (w, s) = pool.deque_pool.deque();
426 pool.stealers = stealers;
428 // Now that we've got all our work queues, create one scheduler per
429 // queue, spawn the scheduler into a thread, and be sure to keep a
430 // handle to the scheduler and the thread to keep them alive.
431 for worker in workers.move_iter() {
432 rtdebug!("inserting a regular scheduler");
434 let mut sched = box Scheduler::new(pool.id,
437 pool.stealers.clone(),
438 pool.sleepers.clone(),
439 pool.task_state.clone());
440 pool.handles.push(sched.make_handle());
441 pool.threads.push(Thread::start(proc() { sched.bootstrap(); }));
447 /// Creates a new task configured to run inside of this pool of schedulers.
448 /// This is useful to create a task which can then be sent to a specific
449 /// scheduler created by `spawn_sched` (and possibly pin it to that
451 #[deprecated = "use the green and green_pinned methods of GreenTaskBuilder instead"]
452 pub fn task(&mut self, opts: TaskOpts, f: proc():Send) -> Box<GreenTask> {
453 GreenTask::configure(&mut self.stack_pool, opts, f)
456 /// Spawns a new task into this pool of schedulers, using the specified
457 /// options to configure the new task which is spawned.
459 /// New tasks are spawned in a round-robin fashion to the schedulers in this
460 /// pool, but tasks can certainly migrate among schedulers once they're in
462 #[deprecated = "use the green and green_pinned methods of GreenTaskBuilder instead"]
463 pub fn spawn(&mut self, opts: TaskOpts, f: proc():Send) {
464 let task = self.task(opts, f);
466 // Figure out someone to send this task to
467 let idx = self.next_friend;
468 self.next_friend += 1;
469 if self.next_friend >= self.handles.len() {
470 self.next_friend = 0;
473 // Jettison the task away!
474 self.handles.get_mut(idx).send(TaskFromFriend(task));
477 /// Spawns a new scheduler into this M:N pool. A handle is returned to the
478 /// scheduler for use. The scheduler will not exit as long as this handle is
481 /// The scheduler spawned will participate in work stealing with all of the
482 /// other schedulers currently in the scheduler pool.
483 pub fn spawn_sched(&mut self) -> SchedHandle {
484 let (worker, stealer) = self.deque_pool.deque();
485 self.stealers.push(stealer.clone());
487 // Tell all existing schedulers about this new scheduler so they can all
488 // steal work from it
489 for handle in self.handles.mut_iter() {
490 handle.send(NewNeighbor(stealer.clone()));
493 // Create the new scheduler, using the same sleeper list as all the
494 // other schedulers as well as having a stealer handle to all other
496 let mut sched = box Scheduler::new(self.id,
499 self.stealers.clone(),
500 self.sleepers.clone(),
501 self.task_state.clone());
502 let ret = sched.make_handle();
503 self.handles.push(sched.make_handle());
504 self.threads.push(Thread::start(proc() { sched.bootstrap() }));
509 /// Consumes the pool of schedulers, waiting for all tasks to exit and all
510 /// schedulers to shut down.
512 /// This function is required to be called in order to drop a pool of
513 /// schedulers, it is considered an error to drop a pool without calling
516 /// This only waits for all tasks in *this pool* of schedulers to exit, any
517 /// native tasks or extern pools will not be waited on
518 pub fn shutdown(mut self) {
519 self.stealers = vec![];
521 // Wait for everyone to exit. We may have reached a 0-task count
522 // multiple times in the past, meaning there could be several buffered
523 // messages on the `tasks_done` port. We're guaranteed that after *some*
524 // message the current task count will be 0, so we just receive in a
525 // loop until everything is totally dead.
526 while self.task_state.active() {
527 self.tasks_done.recv();
530 // Now that everyone's gone, tell everything to shut down.
531 for mut handle in replace(&mut self.handles, vec![]).move_iter() {
532 handle.send(Shutdown);
534 for thread in replace(&mut self.threads, vec![]).move_iter() {
541 fn new() -> (Receiver<()>, TaskState) {
542 let (tx, rx) = channel();
544 cnt: Arc::new(AtomicUint::new(0)),
549 fn increment(&mut self) {
550 self.cnt.fetch_add(1, SeqCst);
553 fn active(&self) -> bool {
554 self.cnt.load(SeqCst) != 0
557 fn decrement(&mut self) {
558 let prev = self.cnt.fetch_sub(1, SeqCst);
565 impl Drop for SchedPool {
567 if self.threads.len() > 0 {
568 fail!("dropping a M:N scheduler pool that wasn't shut down");
573 /// A spawner for green tasks
574 pub struct GreenSpawner<'a>{
575 pool: &'a mut SchedPool,
576 handle: Option<&'a mut SchedHandle>
579 impl<'a> Spawner for GreenSpawner<'a> {
581 fn spawn(self, opts: TaskOpts, f: proc():Send) {
582 let GreenSpawner { pool, handle } = self;
584 None => pool.spawn(opts, f),
585 Some(h) => h.send(PinnedTask(pool.task(opts, f)))
590 /// An extension trait adding `green` configuration methods to `TaskBuilder`.
591 pub trait GreenTaskBuilder {
592 fn green<'a>(self, &'a mut SchedPool) -> TaskBuilder<GreenSpawner<'a>>;
593 fn green_pinned<'a>(self, &'a mut SchedPool, &'a mut SchedHandle)
594 -> TaskBuilder<GreenSpawner<'a>>;
597 impl<S: Spawner> GreenTaskBuilder for TaskBuilder<S> {
598 fn green<'a>(self, pool: &'a mut SchedPool) -> TaskBuilder<GreenSpawner<'a>> {
599 self.spawner(GreenSpawner {pool: pool, handle: None})
602 fn green_pinned<'a>(self, pool: &'a mut SchedPool, handle: &'a mut SchedHandle)
603 -> TaskBuilder<GreenSpawner<'a>> {
604 self.spawner(GreenSpawner {pool: pool, handle: Some(handle)})
610 use std::task::TaskBuilder;
611 use super::{SchedPool, PoolConfig, GreenTaskBuilder};
614 fn test_green_builder() {
615 let mut pool = SchedPool::new(PoolConfig::new());
616 let res = TaskBuilder::new().green(&mut pool).try(proc() {
617 "Success!".to_string()
619 assert_eq!(res.ok().unwrap(), "Success!".to_string());