6 |
|
|
7 |
|
package jsr166y; |
8 |
|
|
9 |
– |
import java.util.concurrent.*; |
10 |
– |
|
9 |
|
import java.util.ArrayList; |
10 |
|
import java.util.Arrays; |
11 |
|
import java.util.Collection; |
12 |
|
import java.util.Collections; |
13 |
|
import java.util.List; |
14 |
+ |
import java.util.concurrent.AbstractExecutorService; |
15 |
+ |
import java.util.concurrent.Callable; |
16 |
+ |
import java.util.concurrent.CountDownLatch; |
17 |
+ |
import java.util.concurrent.ExecutorService; |
18 |
+ |
import java.util.concurrent.Future; |
19 |
+ |
import java.util.concurrent.RejectedExecutionException; |
20 |
+ |
import java.util.concurrent.RunnableFuture; |
21 |
+ |
import java.util.concurrent.TimeUnit; |
22 |
+ |
import java.util.concurrent.TimeoutException; |
23 |
+ |
import java.util.concurrent.atomic.AtomicInteger; |
24 |
|
import java.util.concurrent.locks.LockSupport; |
25 |
|
import java.util.concurrent.locks.ReentrantLock; |
18 |
– |
import java.util.concurrent.atomic.AtomicInteger; |
19 |
– |
import java.util.concurrent.CountDownLatch; |
26 |
|
|
27 |
|
/** |
28 |
|
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
58 |
|
* convenient form for informal monitoring. |
59 |
|
* |
60 |
|
* <p> As is the case with other ExecutorServices, there are three |
61 |
< |
* main task execution methods summarized in the follwoing |
61 |
> |
* main task execution methods summarized in the following |
62 |
|
* table. These are designed to be used by clients not already engaged |
63 |
|
* in fork/join computations in the current pool. The main forms of |
64 |
|
* these methods accept instances of {@code ForkJoinTask}, but |
66 |
|
* Runnable}- or {@code Callable}- based activities as well. However, |
67 |
|
* tasks that are already executing in a pool should normally |
68 |
|
* <em>NOT</em> use these pool execution methods, but instead use the |
69 |
< |
* within-computation forms listed in the table. To avoid inadvertant |
64 |
< |
* cyclic task dependencies and to improve performance, task |
65 |
< |
* submissions to the current pool by an ongoing fork/join |
66 |
< |
* computations may be implicitly translated to the corresponding |
67 |
< |
* ForkJoinTask forms. |
69 |
> |
* within-computation forms listed in the table. |
70 |
|
* |
71 |
|
* <table BORDER CELLPADDING=3 CELLSPACING=1> |
72 |
|
* <tr> |
75 |
|
* <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td> |
76 |
|
* </tr> |
77 |
|
* <tr> |
78 |
< |
* <td> <b>Arange async execution</td> |
78 |
> |
* <td> <b>Arrange async execution</td> |
79 |
|
* <td> {@link #execute(ForkJoinTask)}</td> |
80 |
|
* <td> {@link ForkJoinTask#fork}</td> |
81 |
|
* </tr> |
90 |
|
* <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td> |
91 |
|
* </tr> |
92 |
|
* </table> |
93 |
< |
* |
93 |
> |
* |
94 |
|
* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is |
95 |
|
* used for all parallel task execution in a program or subsystem. |
96 |
|
* Otherwise, use would not usually outweigh the construction and |
115 |
|
* {@code IllegalArgumentException}. |
116 |
|
* |
117 |
|
* <p>This implementation rejects submitted tasks (that is, by throwing |
118 |
< |
* {@link RejectedExecutionException}) only when the pool is shut down. |
118 |
> |
* {@link RejectedExecutionException}) only when the pool is shut down |
119 |
> |
* or internal resources have been exhausted. |
120 |
|
* |
121 |
|
* @since 1.7 |
122 |
|
* @author Doug Lea |
143 |
|
* of tasks profit from cache affinities, but others are harmed by |
144 |
|
* cache pollution effects.) |
145 |
|
* |
146 |
+ |
* Beyond work-stealing support and essential bookkeeping, the |
147 |
+ |
* main responsibility of this framework is to take actions when |
148 |
+ |
* one worker is waiting to join a task stolen (or always held by) |
149 |
+ |
* another. Because we are multiplexing many tasks on to a pool |
150 |
+ |
* of workers, we can't just let them block (as in Thread.join). |
151 |
+ |
* We also cannot just reassign the joiner's run-time stack with |
152 |
+ |
* another and replace it later, which would be a form of |
153 |
+ |
* "continuation", that even if possible is not necessarily a good |
154 |
+ |
* idea. Given that the creation costs of most threads on most |
155 |
+ |
* systems mainly surrounds setting up runtime stacks, thread |
156 |
+ |
* creation and switching is usually not much more expensive than |
157 |
+ |
* stack creation and switching, and is more flexible). Instead we |
158 |
+ |
* combine two tactics: |
159 |
+ |
* |
160 |
+ |
* Helping: Arranging for the joiner to execute some task that it |
161 |
+ |
* would be running if the steal had not occurred. Method |
162 |
+ |
* ForkJoinWorkerThread.helpJoinTask tracks joining->stealing |
163 |
+ |
* links to try to find such a task. |
164 |
+ |
* |
165 |
+ |
* Compensating: Unless there are already enough live threads, |
166 |
+ |
* method helpMaintainParallelism() may create or |
167 |
+ |
* re-activate a spare thread to compensate for blocked |
168 |
+ |
* joiners until they unblock. |
169 |
+ |
* |
170 |
+ |
* It is impossible to keep exactly the target (parallelism) |
171 |
+ |
* number of threads running at any given time. Determining |
172 |
+ |
* existence of conservatively safe helping targets, the |
173 |
+ |
* availability of already-created spares, and the apparent need |
174 |
+ |
* to create new spares are all racy and require heuristic |
175 |
+ |
* guidance, so we rely on multiple retries of each. Compensation |
176 |
+ |
* occurs in slow-motion. It is triggered only upon timeouts of |
177 |
+ |
* Object.wait used for joins. This reduces poor decisions that |
178 |
+ |
* would otherwise be made when threads are waiting for others |
179 |
+ |
* that are stalled because of unrelated activities such as |
180 |
+ |
* garbage collection. |
181 |
+ |
* |
182 |
+ |
* The ManagedBlocker extension API can't use helping so relies |
183 |
+ |
* only on compensation in method awaitBlocker. |
184 |
+ |
* |
185 |
|
* The main throughput advantages of work-stealing stem from |
186 |
|
* decentralized control -- workers mostly steal tasks from each |
187 |
|
* other. We do not want to negate this by creating bottlenecks |
188 |
< |
* implementing the management responsibilities of this class. So |
189 |
< |
* we use a collection of techniques that avoid, reduce, or cope |
190 |
< |
* well with contention. These entail several instances of |
191 |
< |
* bit-packing into CASable fields to maintain only the minimally |
192 |
< |
* required atomicity. To enable such packing, we restrict maximum |
193 |
< |
* parallelism to (1<<15)-1 (enabling twice this to fit into a 16 |
194 |
< |
* bit field), which is far in excess of normal operating range. |
195 |
< |
* Even though updates to some of these bookkeeping fields do |
196 |
< |
* sometimes contend with each other, they don't normally |
197 |
< |
* cache-contend with updates to others enough to warrant memory |
198 |
< |
* padding or isolation. So they are all held as fields of |
199 |
< |
* ForkJoinPool objects. The main capabilities are as follows: |
188 |
> |
* implementing other management responsibilities. So we use a |
189 |
> |
* collection of techniques that avoid, reduce, or cope well with |
190 |
> |
* contention. These entail several instances of bit-packing into |
191 |
> |
* CASable fields to maintain only the minimally required |
192 |
> |
* atomicity. To enable such packing, we restrict maximum |
193 |
> |
* parallelism to (1<<15)-1 (enabling twice this (to accommodate |
194 |
> |
* unbalanced increments and decrements) to fit into a 16 bit |
195 |
> |
* field, which is far in excess of normal operating range. Even |
196 |
> |
* though updates to some of these bookkeeping fields do sometimes |
197 |
> |
* contend with each other, they don't normally cache-contend with |
198 |
> |
* updates to others enough to warrant memory padding or |
199 |
> |
* isolation. So they are all held as fields of ForkJoinPool |
200 |
> |
* objects. The main capabilities are as follows: |
201 |
|
* |
202 |
|
* 1. Creating and removing workers. Workers are recorded in the |
203 |
|
* "workers" array. This is an array as opposed to some other data |
213 |
|
* blocked workers. However, all other support code is set up to |
214 |
|
* work with other policies. |
215 |
|
* |
216 |
+ |
* To ensure that we do not hold on to worker references that |
217 |
+ |
* would prevent GC, ALL accesses to workers are via indices into |
218 |
+ |
* the workers array (which is one source of some of the unusual |
219 |
+ |
* code constructions here). In essence, the workers array serves |
220 |
+ |
* as a WeakReference mechanism. Thus for example the event queue |
221 |
+ |
* stores worker indices, not worker references. Access to the |
222 |
+ |
* workers in associated methods (for example releaseEventWaiters) |
223 |
+ |
* must both index-check and null-check the IDs. All such accesses |
224 |
+ |
* ignore bad IDs by returning out early from what they are doing, |
225 |
+ |
* since this can only be associated with shutdown, in which case |
226 |
+ |
* it is OK to give up. On termination, we just clobber these |
227 |
+ |
* data structures without trying to use them. |
228 |
+ |
* |
229 |
|
* 2. Bookkeeping for dynamically adding and removing workers. We |
230 |
|
* aim to approximately maintain the given level of parallelism. |
231 |
|
* When some workers are known to be blocked (on joins or via |
232 |
|
* ManagedBlocker), we may create or resume others to take their |
233 |
|
* place until they unblock (see below). Implementing this |
234 |
|
* requires counts of the number of "running" threads (i.e., those |
235 |
< |
* that are neither blocked nor artifically suspended) as well as |
235 |
> |
* that are neither blocked nor artificially suspended) as well as |
236 |
|
* the total number. These two values are packed into one field, |
237 |
|
* "workerCounts" because we need accurate snapshots when deciding |
238 |
< |
* to create, resume or suspend. To support these decisions, |
239 |
< |
* updates to spare counts must be prospective (not |
240 |
< |
* retrospective). For example, the running count is decremented |
241 |
< |
* before blocking by a thread about to block as a spare, but |
186 |
< |
* incremented by the thread about to unblock it. Updates upon |
187 |
< |
* resumption ofr threads blocking in awaitJoin or awaitBlocker |
188 |
< |
* cannot usually be prospective, so the running count is in |
189 |
< |
* general an upper bound of the number of productively running |
190 |
< |
* threads Updates to the workerCounts field sometimes transiently |
191 |
< |
* encounter a fair amount of contention when join dependencies |
192 |
< |
* are such that many threads block or unblock at about the same |
193 |
< |
* time. We alleviate this by sometimes performing an alternative |
194 |
< |
* action on contention like releasing waiters or locating spares. |
238 |
> |
* to create, resume or suspend. Note however that the |
239 |
> |
* correspondence of these counts to reality is not guaranteed. In |
240 |
> |
* particular updates for unblocked threads may lag until they |
241 |
> |
* actually wake up. |
242 |
|
* |
243 |
|
* 3. Maintaining global run state. The run state of the pool |
244 |
|
* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to |
267 |
|
* workers that previously could not find a task to now find one: |
268 |
|
* Submission of a new task to the pool, or another worker pushing |
269 |
|
* a task onto a previously empty queue. (We also use this |
270 |
< |
* mechanism for termination and reconfiguration actions that |
270 |
> |
* mechanism for configuration and termination actions that |
271 |
|
* require wakeups of idle workers). Each worker maintains its |
272 |
|
* last known event count, and blocks when a scan for work did not |
273 |
|
* find a task AND its lastEventCount matches the current |
278 |
|
* a record (field nextEventWaiter) for the next waiting worker. |
279 |
|
* In addition to allowing simpler decisions about need for |
280 |
|
* wakeup, the event count bits in eventWaiters serve the role of |
281 |
< |
* tags to avoid ABA errors in Treiber stacks. To reduce delays |
282 |
< |
* in task diffusion, workers not otherwise occupied may invoke |
283 |
< |
* method releaseWaiters, that removes and signals (unparks) |
284 |
< |
* workers not waiting on current count. To minimize task |
285 |
< |
* production stalls associate with signalling, any worker pushing |
286 |
< |
* a task on an empty queue invokes the weaker method signalWork, |
240 |
< |
* that only releases idle workers until it detects interference |
241 |
< |
* by other threads trying to release, and lets them take |
242 |
< |
* over. The net effect is a tree-like diffusion of signals, where |
243 |
< |
* released threads (and possibly others) help with unparks. To |
244 |
< |
* further reduce contention effects a bit, failed CASes to |
245 |
< |
* increment field eventCount are tolerated without retries. |
281 |
> |
* tags to avoid ABA errors in Treiber stacks. Upon any wakeup, |
282 |
> |
* released threads also try to release at most two others. The |
283 |
> |
* net effect is a tree-like diffusion of signals, where released |
284 |
> |
* threads (and possibly others) help with unparks. To further |
285 |
> |
* reduce contention effects a bit, failed CASes to increment |
286 |
> |
* field eventCount are tolerated without retries in signalWork. |
287 |
|
* Conceptually they are merged into the same event, which is OK |
288 |
|
* when their only purpose is to enable workers to scan for work. |
289 |
|
* |
290 |
< |
* 5. Managing suspension of extra workers. When a worker is about |
291 |
< |
* to block waiting for a join (or via ManagedBlockers), we may |
292 |
< |
* create a new thread to maintain parallelism level, or at least |
293 |
< |
* avoid starvation (see below). Usually, extra threads are needed |
294 |
< |
* for only very short periods, yet join dependencies are such |
295 |
< |
* that we sometimes need them in bursts. Rather than create new |
296 |
< |
* threads each time this happens, we suspend no-longer-needed |
297 |
< |
* extra ones as "spares". For most purposes, we don't distinguish |
298 |
< |
* "extra" spare threads from normal "core" threads: On each call |
299 |
< |
* to preStep (the only point at which we can do this) a worker |
300 |
< |
* checks to see if there are now too many running workers, and if |
301 |
< |
* so, suspends itself. Methods awaitJoin and awaitBlocker look |
302 |
< |
* for suspended threads to resume before considering creating a |
303 |
< |
* new replacement. We don't need a special data structure to |
304 |
< |
* maintain spares; simply scanning the workers array looking for |
305 |
< |
* worker.isSuspended() is fine because the calling thread is |
306 |
< |
* otherwise not doing anything useful anyway; we are at least as |
307 |
< |
* happy if after locating a spare, the caller doesn't actually |
308 |
< |
* block because the join is ready before we try to adjust and |
309 |
< |
* compensate. Note that this is intrinsically racy. One thread |
310 |
< |
* may become a spare at about the same time as another is |
311 |
< |
* needlessly being created. We counteract this and related slop |
312 |
< |
* in part by requiring resumed spares to immediately recheck (in |
313 |
< |
* preStep) to see whether they they should re-suspend. The only |
314 |
< |
* effective difference between "extra" and "core" threads is that |
315 |
< |
* we allow the "extra" ones to time out and die if they are not |
316 |
< |
* resumed within a keep-alive interval of a few seconds. This is |
317 |
< |
* implemented mainly within ForkJoinWorkerThread, but requires |
318 |
< |
* some coordination (isTrimmed() -- meaning killed while |
319 |
< |
* suspended) to correctly maintain pool counts. |
320 |
< |
* |
321 |
< |
* 6. Deciding when to create new workers. The main dynamic |
322 |
< |
* control in this class is deciding when to create extra threads, |
323 |
< |
* in methods awaitJoin and awaitBlocker. We always need to create |
324 |
< |
* one when the number of running threads becomes zero. But |
325 |
< |
* because blocked joins are typically dependent, we don't |
326 |
< |
* necessarily need or want one-to-one replacement. Instead, we |
327 |
< |
* use a combination of heuristics that adds threads only when the |
328 |
< |
* pool appears to be approaching starvation. These effectively |
329 |
< |
* reduce churn at the price of systematically undershooting |
330 |
< |
* target parallelism when many threads are blocked. However, |
331 |
< |
* biasing toward undeshooting partially compensates for the above |
332 |
< |
* mechanics to suspend extra threads, that normally lead to |
333 |
< |
* overshoot because we can only suspend workers in-between |
334 |
< |
* top-level actions. It also better copes with the fact that some |
335 |
< |
* of the methods in this class tend to never become compiled (but |
336 |
< |
* are interpreted), so some components of the entire set of |
337 |
< |
* controls might execute many times faster than others. And |
290 |
> |
* 5. Managing suspension of extra workers. When a worker notices |
291 |
> |
* (usually upon timeout of a wait()) that there are too few |
292 |
> |
* running threads, we may create a new thread to maintain |
293 |
> |
* parallelism level, or at least avoid starvation. Usually, extra |
294 |
> |
* threads are needed for only very short periods, yet join |
295 |
> |
* dependencies are such that we sometimes need them in |
296 |
> |
* bursts. Rather than create new threads each time this happens, |
297 |
> |
* we suspend no-longer-needed extra ones as "spares". For most |
298 |
> |
* purposes, we don't distinguish "extra" spare threads from |
299 |
> |
* normal "core" threads: On each call to preStep (the only point |
300 |
> |
* at which we can do this) a worker checks to see if there are |
301 |
> |
* now too many running workers, and if so, suspends itself. |
302 |
> |
* Method helpMaintainParallelism looks for suspended threads to |
303 |
> |
* resume before considering creating a new replacement. The |
304 |
> |
* spares themselves are encoded on another variant of a Treiber |
305 |
> |
* Stack, headed at field "spareWaiters". Note that the use of |
306 |
> |
* spares is intrinsically racy. One thread may become a spare at |
307 |
> |
* about the same time as another is needlessly being created. We |
308 |
> |
* counteract this and related slop in part by requiring resumed |
309 |
> |
* spares to immediately recheck (in preStep) to see whether they |
310 |
> |
* should re-suspend. |
311 |
> |
* |
312 |
> |
* 6. Killing off unneeded workers. A timeout mechanism is used to |
313 |
> |
* shed unused workers: The oldest (first) event queue waiter uses |
314 |
> |
* a timed rather than hard wait. When this wait times out without |
315 |
> |
* a normal wakeup, it tries to shutdown any one (for convenience |
316 |
> |
* the newest) other spare or event waiter via |
317 |
> |
* tryShutdownUnusedWorker. This eventually reduces the number of |
318 |
> |
* worker threads to a minimum of one after a long enough period |
319 |
> |
* without use. |
320 |
> |
* |
321 |
> |
* 7. Deciding when to create new workers. The main dynamic |
322 |
> |
* control in this class is deciding when to create extra threads |
323 |
> |
* in method helpMaintainParallelism. We would like to keep |
324 |
> |
* exactly #parallelism threads running, which is an impossible |
325 |
> |
* task. We always need to create one when the number of running |
326 |
> |
* threads would become zero and all workers are busy. Beyond |
327 |
> |
* this, we must rely on heuristics that work well in the |
328 |
> |
* presence of transient phenomena such as GC stalls, dynamic |
329 |
> |
* compilation, and wake-up lags. These transients are extremely |
330 |
> |
* common -- we are normally trying to fully saturate the CPUs on |
331 |
> |
* a machine, so almost any activity other than running tasks |
332 |
> |
* impedes accuracy. Our main defense is to allow parallelism to |
333 |
> |
* lapse for a while during joins, and use a timeout to see if, |
334 |
> |
* after the resulting settling, there is still a need for |
335 |
> |
* additional workers. This also better copes with the fact that |
336 |
> |
* some of the methods in this class tend to never become compiled |
337 |
> |
* (but are interpreted), so some components of the entire set of |
338 |
> |
* controls might execute 100 times faster than others. And |
339 |
|
* similarly for cases where the apparent lack of work is just due |
340 |
|
* to GC stalls and other transient system activity. |
341 |
|
* |
350 |
|
* |
351 |
|
* Style notes: There are lots of inline assignments (of form |
352 |
|
* "while ((local = field) != 0)") which are usually the simplest |
353 |
< |
* way to ensure read orderings. Also several occurrences of the |
354 |
< |
* unusual "do {} while(!cas...)" which is the simplest way to |
355 |
< |
* force an update of a CAS'ed variable. There are also a few |
356 |
< |
* other coding oddities that help some methods perform reasonably |
357 |
< |
* even when interpreted (not compiled). |
353 |
> |
* way to ensure the required read orderings (which are sometimes |
354 |
> |
* critical). Also several occurrences of the unusual "do {} |
355 |
> |
* while (!cas...)" which is the simplest way to force an update of |
356 |
> |
* a CAS'ed variable. There are also other coding oddities that |
357 |
> |
* help some methods perform reasonably even when interpreted (not |
358 |
> |
* compiled), at the expense of some messy constructions that |
359 |
> |
* reduce byte code counts. |
360 |
|
* |
361 |
|
* The order of declarations in this file is: (1) statics (2) |
362 |
|
* fields (along with constants used when unpacking some of them) |
424 |
|
new AtomicInteger(); |
425 |
|
|
426 |
|
/** |
427 |
< |
* Absolute bound for parallelism level. Twice this number must |
428 |
< |
* fit into a 16bit field to enable word-packing for some counts. |
427 |
> |
* The time to block in a join (see awaitJoin) before checking if |
428 |
> |
* a new worker should be (re)started to maintain parallelism |
429 |
> |
* level. The value should be short enough to maintain global |
430 |
> |
* responsiveness and progress but long enough to avoid |
431 |
> |
* counterproductive firings during GC stalls or unrelated system |
432 |
> |
* activity, and to not bog down systems with continual re-firings |
433 |
> |
* on GCs or legitimately long waits. |
434 |
> |
*/ |
435 |
> |
private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second |
436 |
> |
|
437 |
> |
/** |
438 |
> |
* The wakeup interval (in nanoseconds) for the oldest worker |
439 |
> |
* waiting for an event to invoke tryShutdownUnusedWorker to |
440 |
> |
* shrink the number of workers. The exact value does not matter |
441 |
> |
* too much. It must be short enough to release resources during |
442 |
> |
* sustained periods of idleness, but not so short that threads |
443 |
> |
* are continually re-created. |
444 |
> |
*/ |
445 |
> |
private static final long SHRINK_RATE_NANOS = |
446 |
> |
30L * 1000L * 1000L * 1000L; // 2 per minute |
447 |
> |
|
448 |
> |
/** |
449 |
> |
* Absolute bound for parallelism level. Twice this number plus |
450 |
> |
* one (i.e., 0xfff) must fit into a 16bit field to enable |
451 |
> |
* word-packing for some counts and indices. |
452 |
|
*/ |
453 |
< |
private static final int MAX_THREADS = 0x7fff; |
453 |
> |
private static final int MAX_WORKERS = 0x7fff; |
454 |
|
|
455 |
|
/** |
456 |
|
* Array holding all worker threads in the pool. Array size must |
490 |
|
private volatile long stealCount; |
491 |
|
|
492 |
|
/** |
493 |
< |
* Encoded record of top of treiber stack of threads waiting for |
493 |
> |
* Encoded record of top of Treiber stack of threads waiting for |
494 |
|
* events. The top 32 bits contain the count being waited for. The |
495 |
< |
* bottom word contains one plus the pool index of waiting worker |
496 |
< |
* thread. |
495 |
> |
* bottom 16 bits contains one plus the pool index of waiting |
496 |
> |
* worker thread. (Bits 16-31 are unused.) |
497 |
|
*/ |
498 |
|
private volatile long eventWaiters; |
499 |
|
|
500 |
|
private static final int EVENT_COUNT_SHIFT = 32; |
501 |
< |
private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L; |
501 |
> |
private static final long WAITER_ID_MASK = (1L << 16) - 1L; |
502 |
|
|
503 |
|
/** |
504 |
|
* A counter for events that may wake up worker threads: |
505 |
|
* - Submission of a new task to the pool |
506 |
|
* - A worker pushing a task on an empty queue |
507 |
< |
* - termination and reconfiguration |
507 |
> |
* - termination |
508 |
|
*/ |
509 |
|
private volatile int eventCount; |
510 |
|
|
511 |
|
/** |
512 |
+ |
* Encoded record of top of Treiber stack of spare threads waiting |
513 |
+ |
* for resumption. The top 16 bits contain an arbitrary count to |
514 |
+ |
* avoid ABA effects. The bottom 16bits contains one plus the pool |
515 |
+ |
* index of waiting worker thread. |
516 |
+ |
*/ |
517 |
+ |
private volatile int spareWaiters; |
518 |
+ |
|
519 |
+ |
private static final int SPARE_COUNT_SHIFT = 16; |
520 |
+ |
private static final int SPARE_ID_MASK = (1 << 16) - 1; |
521 |
+ |
|
522 |
+ |
/** |
523 |
|
* Lifecycle control. The low word contains the number of workers |
524 |
|
* that are (probably) executing tasks. This value is atomically |
525 |
|
* incremented before a worker gets a task to run, and decremented |
530 |
|
* These are bundled together to ensure consistent read for |
531 |
|
* termination checks (i.e., that runLevel is at least SHUTDOWN |
532 |
|
* and active threads is zero). |
533 |
+ |
* |
534 |
+ |
* Notes: Most direct CASes are dependent on these bitfield |
535 |
+ |
* positions. Also, this field is non-private to enable direct |
536 |
+ |
* performance-sensitive CASes in ForkJoinWorkerThread. |
537 |
|
*/ |
538 |
< |
private volatile int runState; |
538 |
> |
volatile int runState; |
539 |
|
|
540 |
|
// Note: The order among run level values matters. |
541 |
|
private static final int RUNLEVEL_SHIFT = 16; |
543 |
|
private static final int TERMINATING = 1 << (RUNLEVEL_SHIFT + 1); |
544 |
|
private static final int TERMINATED = 1 << (RUNLEVEL_SHIFT + 2); |
545 |
|
private static final int ACTIVE_COUNT_MASK = (1 << RUNLEVEL_SHIFT) - 1; |
464 |
– |
private static final int ONE_ACTIVE = 1; // active update delta |
546 |
|
|
547 |
|
/** |
548 |
|
* Holds number of total (i.e., created and not yet terminated) |
551 |
|
* making decisions about creating and suspending spare |
552 |
|
* threads. Updated only by CAS. Note that adding a new worker |
553 |
|
* requires incrementing both counts, since workers start off in |
554 |
< |
* running state. This field is also used for memory-fencing |
474 |
< |
* configuration parameters. |
554 |
> |
* running state. |
555 |
|
*/ |
556 |
|
private volatile int workerCounts; |
557 |
|
|
583 |
|
*/ |
584 |
|
private final int poolNumber; |
585 |
|
|
586 |
< |
// utilities for updating fields |
586 |
> |
// Utilities for CASing fields. Note that most of these |
587 |
> |
// are usually manually inlined by callers |
588 |
|
|
589 |
|
/** |
590 |
< |
* Increments running count. Also used by ForkJoinTask. |
590 |
> |
* Increments running count part of workerCounts |
591 |
|
*/ |
592 |
|
final void incrementRunningCount() { |
593 |
|
int c; |
594 |
|
do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
595 |
< |
c = workerCounts, |
595 |
> |
c = workerCounts, |
596 |
|
c + ONE_RUNNING)); |
597 |
|
} |
598 |
< |
|
598 |
> |
|
599 |
|
/** |
600 |
|
* Tries to decrement running count unless already zero |
601 |
|
*/ |
608 |
|
} |
609 |
|
|
610 |
|
/** |
611 |
< |
* Tries incrementing active count; fails on contention. |
612 |
< |
* Called by workers before executing tasks. |
611 |
> |
* Forces decrement of encoded workerCounts, awaiting nonzero if |
612 |
> |
* (rarely) necessary when other count updates lag. |
613 |
|
* |
614 |
< |
* @return true on success |
614 |
> |
* @param dr -- either zero or ONE_RUNNING |
615 |
> |
* @param dt -- either zero or ONE_TOTAL |
616 |
|
*/ |
617 |
< |
final boolean tryIncrementActiveCount() { |
618 |
< |
int c; |
619 |
< |
return UNSAFE.compareAndSwapInt(this, runStateOffset, |
620 |
< |
c = runState, c + ONE_ACTIVE); |
617 |
> |
private void decrementWorkerCounts(int dr, int dt) { |
618 |
> |
for (;;) { |
619 |
> |
int wc = workerCounts; |
620 |
> |
if ((wc & RUNNING_COUNT_MASK) - dr < 0 || |
621 |
> |
(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) { |
622 |
> |
if ((runState & TERMINATED) != 0) |
623 |
> |
return; // lagging termination on a backout |
624 |
> |
Thread.yield(); |
625 |
> |
} |
626 |
> |
if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
627 |
> |
wc, wc - (dr + dt))) |
628 |
> |
return; |
629 |
> |
} |
630 |
|
} |
631 |
|
|
632 |
|
/** |
636 |
|
final boolean tryDecrementActiveCount() { |
637 |
|
int c; |
638 |
|
return UNSAFE.compareAndSwapInt(this, runStateOffset, |
639 |
< |
c = runState, c - ONE_ACTIVE); |
639 |
> |
c = runState, c - 1); |
640 |
|
} |
641 |
|
|
642 |
|
/** |
665 |
|
lock.lock(); |
666 |
|
try { |
667 |
|
ForkJoinWorkerThread[] ws = workers; |
668 |
< |
int nws = ws.length; |
669 |
< |
if (k < 0 || k >= nws || ws[k] != null) { |
670 |
< |
for (k = 0; k < nws && ws[k] != null; ++k) |
668 |
> |
int n = ws.length; |
669 |
> |
if (k < 0 || k >= n || ws[k] != null) { |
670 |
> |
for (k = 0; k < n && ws[k] != null; ++k) |
671 |
|
; |
672 |
< |
if (k == nws) |
673 |
< |
ws = Arrays.copyOf(ws, nws << 1); |
672 |
> |
if (k == n) |
673 |
> |
ws = Arrays.copyOf(ws, n << 1); |
674 |
|
} |
675 |
|
ws[k] = w; |
676 |
|
workers = ws; // volatile array write ensures slot visibility |
681 |
|
} |
682 |
|
|
683 |
|
/** |
684 |
< |
* Nulls out record of worker in workers array |
684 |
> |
* Nulls out record of worker in workers array. |
685 |
|
*/ |
686 |
|
private void forgetWorker(ForkJoinWorkerThread w) { |
687 |
|
int idx = w.poolIndex; |
688 |
< |
// Locking helps method recordWorker avoid unecessary expansion |
688 |
> |
// Locking helps method recordWorker avoid unnecessary expansion |
689 |
|
final ReentrantLock lock = this.workerLock; |
690 |
|
lock.lock(); |
691 |
|
try { |
697 |
|
} |
698 |
|
} |
699 |
|
|
609 |
– |
// adding and removing workers |
610 |
– |
|
700 |
|
/** |
701 |
< |
* Tries to create and add new worker. Assumes that worker counts |
702 |
< |
* are already updated to accommodate the worker, so adjusts on |
703 |
< |
* failure. |
701 |
> |
* Final callback from terminating worker. Removes record of |
702 |
> |
* worker from array, and adjusts counts. If pool is shutting |
703 |
> |
* down, tries to complete termination. |
704 |
|
* |
705 |
< |
* @return new worker or null if creation failed |
705 |
> |
* @param w the worker |
706 |
|
*/ |
707 |
< |
private ForkJoinWorkerThread addWorker() { |
708 |
< |
ForkJoinWorkerThread w = null; |
709 |
< |
try { |
710 |
< |
w = factory.newThread(this); |
711 |
< |
} finally { // Adjust on either null or exceptional factory return |
712 |
< |
if (w == null) { |
624 |
< |
onWorkerCreationFailure(); |
625 |
< |
return null; |
626 |
< |
} |
627 |
< |
} |
628 |
< |
w.start(recordWorker(w), ueh); |
629 |
< |
return w; |
707 |
> |
final void workerTerminated(ForkJoinWorkerThread w) { |
708 |
> |
forgetWorker(w); |
709 |
> |
decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL); |
710 |
> |
while (w.stealCount != 0) // collect final count |
711 |
> |
tryAccumulateStealCount(w); |
712 |
> |
tryTerminate(false); |
713 |
|
} |
714 |
|
|
715 |
+ |
// Waiting for and signalling events |
716 |
+ |
|
717 |
|
/** |
718 |
< |
* Adjusts counts upon failure to create worker |
718 |
> |
* Releases workers blocked on a count not equal to current count. |
719 |
> |
* Normally called after precheck that eventWaiters isn't zero to |
720 |
> |
* avoid wasted array checks. Gives up upon a change in count or |
721 |
> |
* upon releasing two workers, letting others take over. |
722 |
|
*/ |
723 |
< |
private void onWorkerCreationFailure() { |
724 |
< |
for (;;) { |
725 |
< |
int wc = workerCounts; |
726 |
< |
if ((wc >>> TOTAL_COUNT_SHIFT) > 0 && |
727 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
728 |
< |
wc, wc - (ONE_RUNNING|ONE_TOTAL))) |
723 |
> |
private void releaseEventWaiters() { |
724 |
> |
ForkJoinWorkerThread[] ws = workers; |
725 |
> |
int n = ws.length; |
726 |
> |
long h = eventWaiters; |
727 |
> |
int ec = eventCount; |
728 |
> |
boolean releasedOne = false; |
729 |
> |
ForkJoinWorkerThread w; int id; |
730 |
> |
while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 && |
731 |
> |
(int)(h >>> EVENT_COUNT_SHIFT) != ec && |
732 |
> |
id < n && (w = ws[id]) != null) { |
733 |
> |
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
734 |
> |
h, w.nextWaiter)) { |
735 |
> |
LockSupport.unpark(w); |
736 |
> |
if (releasedOne) // exit on second release |
737 |
> |
break; |
738 |
> |
releasedOne = true; |
739 |
> |
} |
740 |
> |
if (eventCount != ec) |
741 |
|
break; |
742 |
+ |
h = eventWaiters; |
743 |
|
} |
643 |
– |
tryTerminate(false); // in case of failure during shutdown |
744 |
|
} |
745 |
|
|
746 |
|
/** |
747 |
< |
* Create enough total workers to establish target parallelism, |
748 |
< |
* giving up if terminating or addWorker fails |
747 |
> |
* Tries to advance eventCount and releases waiters. Called only |
748 |
> |
* from workers. |
749 |
|
*/ |
750 |
< |
private void ensureEnoughTotalWorkers() { |
751 |
< |
int wc; |
752 |
< |
while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism && |
753 |
< |
runState < TERMINATING) { |
754 |
< |
if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
655 |
< |
wc, wc + (ONE_RUNNING|ONE_TOTAL)) && |
656 |
< |
addWorker() == null)) |
657 |
< |
break; |
658 |
< |
} |
750 |
> |
final void signalWork() { |
751 |
> |
int c; // try to increment event count -- CAS failure OK |
752 |
> |
UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1); |
753 |
> |
if (eventWaiters != 0L) |
754 |
> |
releaseEventWaiters(); |
755 |
|
} |
756 |
|
|
757 |
|
/** |
758 |
< |
* Final callback from terminating worker. Removes record of |
759 |
< |
* worker from array, and adjusts counts. If pool is shutting |
664 |
< |
* down, tries to complete terminatation, else possibly replaces |
665 |
< |
* the worker. |
758 |
> |
* Adds the given worker to event queue and blocks until |
759 |
> |
* terminating or event count advances from the given value |
760 |
|
* |
761 |
< |
* @param w the worker |
761 |
> |
* @param w the calling worker thread |
762 |
> |
* @param ec the count |
763 |
|
*/ |
764 |
< |
final void workerTerminated(ForkJoinWorkerThread w) { |
765 |
< |
if (w.active) { // force inactive |
766 |
< |
w.active = false; |
767 |
< |
do {} while (!tryDecrementActiveCount()); |
768 |
< |
} |
769 |
< |
forgetWorker(w); |
770 |
< |
|
771 |
< |
// Decrement total count, and if was running, running count |
772 |
< |
// Spin (waiting for other updates) if either would be negative |
773 |
< |
int nr = w.isTrimmed() ? 0 : ONE_RUNNING; |
679 |
< |
int unit = ONE_TOTAL + nr; |
680 |
< |
for (;;) { |
681 |
< |
int wc = workerCounts; |
682 |
< |
int rc = wc & RUNNING_COUNT_MASK; |
683 |
< |
if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0) |
684 |
< |
Thread.yield(); // back off if waiting for other updates |
685 |
< |
else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
686 |
< |
wc, wc - unit)) |
764 |
> |
private void eventSync(ForkJoinWorkerThread w, int ec) { |
765 |
> |
long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1)); |
766 |
> |
long h; |
767 |
> |
while ((runState < SHUTDOWN || !tryTerminate(false)) && |
768 |
> |
(((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 || |
769 |
> |
(int)(h >>> EVENT_COUNT_SHIFT) == ec) && |
770 |
> |
eventCount == ec) { |
771 |
> |
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
772 |
> |
w.nextWaiter = h, nh)) { |
773 |
> |
awaitEvent(w, ec); |
774 |
|
break; |
775 |
+ |
} |
776 |
|
} |
689 |
– |
|
690 |
– |
accumulateStealCount(w); // collect final count |
691 |
– |
if (!tryTerminate(false)) |
692 |
– |
ensureEnoughTotalWorkers(); |
777 |
|
} |
778 |
|
|
695 |
– |
// Waiting for and signalling events |
696 |
– |
|
779 |
|
/** |
780 |
< |
* Releases workers blocked on a count not equal to current count. |
780 |
> |
* Blocks the given worker (that has already been entered as an |
781 |
> |
* event waiter) until terminating or event count advances from |
782 |
> |
* the given value. The oldest (first) waiter uses a timed wait to |
783 |
> |
* occasionally one-by-one shrink the number of workers (to a |
784 |
> |
* minimum of one) if the pool has not been used for extended |
785 |
> |
* periods. |
786 |
> |
* |
787 |
> |
* @param w the calling worker thread |
788 |
> |
* @param ec the count |
789 |
|
*/ |
790 |
< |
private void releaseWaiters() { |
791 |
< |
long top; |
792 |
< |
int id; |
793 |
< |
while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 && |
794 |
< |
(int)(top >>> EVENT_COUNT_SHIFT) != eventCount) { |
795 |
< |
ForkJoinWorkerThread[] ws = workers; |
796 |
< |
ForkJoinWorkerThread w; |
797 |
< |
if (ws.length >= id && (w = ws[id - 1]) != null && |
798 |
< |
UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
799 |
< |
top, w.nextWaiter)) |
800 |
< |
LockSupport.unpark(w); |
790 |
> |
private void awaitEvent(ForkJoinWorkerThread w, int ec) { |
791 |
> |
while (eventCount == ec) { |
792 |
> |
if (tryAccumulateStealCount(w)) { // transfer while idle |
793 |
> |
boolean untimed = (w.nextWaiter != 0L || |
794 |
> |
(workerCounts & RUNNING_COUNT_MASK) <= 1); |
795 |
> |
long startTime = untimed? 0 : System.nanoTime(); |
796 |
> |
Thread.interrupted(); // clear/ignore interrupt |
797 |
> |
if (eventCount != ec || w.runState != 0 || |
798 |
> |
runState >= TERMINATING) // recheck after clear |
799 |
> |
break; |
800 |
> |
if (untimed) |
801 |
> |
LockSupport.park(w); |
802 |
> |
else { |
803 |
> |
LockSupport.parkNanos(w, SHRINK_RATE_NANOS); |
804 |
> |
if (eventCount != ec || w.runState != 0 || |
805 |
> |
runState >= TERMINATING) |
806 |
> |
break; |
807 |
> |
if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS) |
808 |
> |
tryShutdownUnusedWorker(ec); |
809 |
> |
} |
810 |
> |
} |
811 |
|
} |
812 |
|
} |
813 |
|
|
814 |
+ |
// Maintaining parallelism |
815 |
+ |
|
816 |
|
/** |
817 |
< |
* Ensures eventCount on exit is different (mod 2^32) than on |
716 |
< |
* entry and wakes up all waiters |
817 |
> |
* Pushes worker onto the spare stack. |
818 |
|
*/ |
819 |
< |
private void signalEvent() { |
820 |
< |
int c; |
821 |
< |
do {} while (!UNSAFE.compareAndSwapInt(this, eventCountOffset, |
822 |
< |
c = eventCount, c+1)); |
722 |
< |
releaseWaiters(); |
819 |
> |
final void pushSpare(ForkJoinWorkerThread w) { |
820 |
> |
int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1); |
821 |
> |
do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
822 |
> |
w.nextSpare = spareWaiters,ns)); |
823 |
|
} |
824 |
|
|
825 |
|
/** |
826 |
< |
* Advances eventCount and releases waiters until interference by |
827 |
< |
* other releasing threads is detected. |
826 |
> |
* Tries (once) to resume a spare if the number of running |
827 |
> |
* threads is less than target. |
828 |
|
*/ |
829 |
< |
final void signalWork() { |
830 |
< |
// EventCount CAS failures are OK -- any change in count suffices. |
831 |
< |
int ec; |
832 |
< |
UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1); |
833 |
< |
outer:for (;;) { |
834 |
< |
long top = eventWaiters; |
835 |
< |
ec = eventCount; |
836 |
< |
for (;;) { |
837 |
< |
ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w; |
838 |
< |
int id = (int)(top & WAITER_INDEX_MASK); |
839 |
< |
if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec) |
840 |
< |
return; |
841 |
< |
if ((ws = workers).length < id || (w = ws[id - 1]) == null || |
842 |
< |
!UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
843 |
< |
top, top = w.nextWaiter)) |
844 |
< |
continue outer; // possibly stale; reread |
829 |
> |
private void tryResumeSpare() { |
830 |
> |
int sw, id; |
831 |
> |
ForkJoinWorkerThread[] ws = workers; |
832 |
> |
int n = ws.length; |
833 |
> |
ForkJoinWorkerThread w; |
834 |
> |
if ((sw = spareWaiters) != 0 && |
835 |
> |
(id = (sw & SPARE_ID_MASK) - 1) >= 0 && |
836 |
> |
id < n && (w = ws[id]) != null && |
837 |
> |
(workerCounts & RUNNING_COUNT_MASK) < parallelism && |
838 |
> |
spareWaiters == sw && |
839 |
> |
UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
840 |
> |
sw, w.nextSpare)) { |
841 |
> |
int c; // increment running count before resume |
842 |
> |
do {} while (!UNSAFE.compareAndSwapInt |
843 |
> |
(this, workerCountsOffset, |
844 |
> |
c = workerCounts, c + ONE_RUNNING)); |
845 |
> |
if (w.tryUnsuspend()) |
846 |
|
LockSupport.unpark(w); |
847 |
< |
if (top != eventWaiters) // let someone else take over |
848 |
< |
return; |
748 |
< |
} |
847 |
> |
else // back out if w was shutdown |
848 |
> |
decrementWorkerCounts(ONE_RUNNING, 0); |
849 |
|
} |
850 |
|
} |
851 |
|
|
852 |
|
/** |
853 |
< |
* If worker is inactive, blocks until terminating or event count |
854 |
< |
* advances from last value held by worker; in any case helps |
855 |
< |
* release others. |
856 |
< |
* |
857 |
< |
* @param w the calling worker thread |
853 |
> |
* Tries to increase the number of running workers if below target |
854 |
> |
* parallelism: If a spare exists tries to resume it via |
855 |
> |
* tryResumeSpare. Otherwise, if not enough total workers or all |
856 |
> |
* existing workers are busy, adds a new worker. In all cases also |
857 |
> |
* helps wake up releasable workers waiting for work. |
858 |
|
*/ |
859 |
< |
private void eventSync(ForkJoinWorkerThread w) { |
860 |
< |
if (!w.active) { |
861 |
< |
int prev = w.lastEventCount; |
862 |
< |
long nextTop = (((long)prev << EVENT_COUNT_SHIFT) | |
863 |
< |
((long)(w.poolIndex + 1))); |
864 |
< |
long top; |
865 |
< |
while ((runState < SHUTDOWN || !tryTerminate(false)) && |
866 |
< |
(((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 || |
867 |
< |
(int)(top >>> EVENT_COUNT_SHIFT) == prev) && |
868 |
< |
eventCount == prev) { |
869 |
< |
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
870 |
< |
w.nextWaiter = top, nextTop)) { |
871 |
< |
accumulateStealCount(w); // transfer steals while idle |
872 |
< |
Thread.interrupted(); // clear/ignore interrupt |
873 |
< |
while (eventCount == prev) |
874 |
< |
w.doPark(); |
859 |
> |
private void helpMaintainParallelism() { |
860 |
> |
int pc = parallelism; |
861 |
> |
int wc, rs, tc; |
862 |
> |
while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc && |
863 |
> |
(rs = runState) < TERMINATING) { |
864 |
> |
if (spareWaiters != 0) |
865 |
> |
tryResumeSpare(); |
866 |
> |
else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS || |
867 |
> |
(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc)) |
868 |
> |
break; // enough total |
869 |
> |
else if (runState == rs && workerCounts == wc && |
870 |
> |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc, |
871 |
> |
wc + (ONE_RUNNING|ONE_TOTAL))) { |
872 |
> |
ForkJoinWorkerThread w = null; |
873 |
> |
try { |
874 |
> |
w = factory.newThread(this); |
875 |
> |
} finally { // adjust on null or exceptional factory return |
876 |
> |
if (w == null) { |
877 |
> |
decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL); |
878 |
> |
tryTerminate(false); // handle failure during shutdown |
879 |
> |
} |
880 |
> |
} |
881 |
> |
if (w == null) |
882 |
|
break; |
883 |
+ |
w.start(recordWorker(w), ueh); |
884 |
+ |
if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) { |
885 |
+ |
int c; // advance event count |
886 |
+ |
UNSAFE.compareAndSwapInt(this, eventCountOffset, |
887 |
+ |
c = eventCount, c+1); |
888 |
+ |
break; // add at most one unless total below target |
889 |
|
} |
890 |
|
} |
778 |
– |
w.lastEventCount = eventCount; |
891 |
|
} |
892 |
< |
releaseWaiters(); |
892 |
> |
if (eventWaiters != 0L) |
893 |
> |
releaseEventWaiters(); |
894 |
> |
} |
895 |
> |
|
896 |
> |
/** |
897 |
> |
* Callback from the oldest waiter in awaitEvent waking up after a |
898 |
> |
* period of non-use. If all workers are idle, tries (once) to |
899 |
> |
* shutdown an event waiter or a spare, if one exists. Note that |
900 |
> |
* we don't need CAS or locks here because the method is called |
901 |
> |
* only from one thread occasionally waking (and even misfires are |
902 |
> |
* OK). Note that until the shutdown worker fully terminates, |
903 |
> |
* workerCounts will overestimate total count, which is tolerable. |
904 |
> |
* |
905 |
> |
* @param ec the event count waited on by caller (to abort |
906 |
> |
* attempt if count has since changed). |
907 |
> |
*/ |
908 |
> |
private void tryShutdownUnusedWorker(int ec) { |
909 |
> |
if (runState == 0 && eventCount == ec) { // only trigger if all idle |
910 |
> |
ForkJoinWorkerThread[] ws = workers; |
911 |
> |
int n = ws.length; |
912 |
> |
ForkJoinWorkerThread w = null; |
913 |
> |
boolean shutdown = false; |
914 |
> |
int sw; |
915 |
> |
long h; |
916 |
> |
if ((sw = spareWaiters) != 0) { // prefer killing spares |
917 |
> |
int id = (sw & SPARE_ID_MASK) - 1; |
918 |
> |
if (id >= 0 && id < n && (w = ws[id]) != null && |
919 |
> |
UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
920 |
> |
sw, w.nextSpare)) |
921 |
> |
shutdown = true; |
922 |
> |
} |
923 |
> |
else if ((h = eventWaiters) != 0L) { |
924 |
> |
long nh; |
925 |
> |
int id = ((int)(h & WAITER_ID_MASK)) - 1; |
926 |
> |
if (id >= 0 && id < n && (w = ws[id]) != null && |
927 |
> |
(nh = w.nextWaiter) != 0L && // keep at least one worker |
928 |
> |
UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh)) |
929 |
> |
shutdown = true; |
930 |
> |
} |
931 |
> |
if (w != null && shutdown) { |
932 |
> |
w.shutdown(); |
933 |
> |
LockSupport.unpark(w); |
934 |
> |
} |
935 |
> |
} |
936 |
> |
releaseEventWaiters(); // in case of interference |
937 |
|
} |
938 |
|
|
939 |
|
/** |
940 |
|
* Callback from workers invoked upon each top-level action (i.e., |
941 |
< |
* stealing a task or taking a submission and running |
942 |
< |
* it). Performs one or both of the following: |
941 |
> |
* stealing a task or taking a submission and running it). |
942 |
> |
* Performs one or more of the following: |
943 |
|
* |
944 |
< |
* * If the worker cannot find work, updates its active status to |
945 |
< |
* inactive and updates activeCount unless there is contention, in |
946 |
< |
* which case it may try again (either in this or a subsequent |
947 |
< |
* call). Additionally, awaits the next task event and/or helps |
948 |
< |
* wake up other releasable waiters. |
949 |
< |
* |
950 |
< |
* * If there are too many running threads, suspends this worker |
951 |
< |
* (first forcing inactivation if necessary). If it is not |
952 |
< |
* resumed before a keepAlive elapses, the worker may be "trimmed" |
953 |
< |
* -- killed while suspended within suspendAsSpare. Otherwise, |
954 |
< |
* upon resume it rechecks to make sure that it is still needed. |
944 |
> |
* 1. If the worker is active and either did not run a task |
945 |
> |
* or there are too many workers, try to set its active status |
946 |
> |
* to inactive and update activeCount. On contention, we may |
947 |
> |
* try again in this or a subsequent call. |
948 |
> |
* |
949 |
> |
* 2. If not enough total workers, help create some. |
950 |
> |
* |
951 |
> |
* 3. If there are too many running workers, suspend this worker |
952 |
> |
* (first forcing inactive if necessary). If it is not needed, |
953 |
> |
* it may be shutdown while suspended (via |
954 |
> |
* tryShutdownUnusedWorker). Otherwise, upon resume it |
955 |
> |
* rechecks running thread count and need for event sync. |
956 |
> |
* |
957 |
> |
* 4. If worker did not run a task, await the next task event via |
958 |
> |
* eventSync if necessary (first forcing inactivation), upon |
959 |
> |
* which the worker may be shutdown via |
960 |
> |
* tryShutdownUnusedWorker. Otherwise, help release any |
961 |
> |
* existing event waiters that are now releasable, |
962 |
|
* |
963 |
|
* @param w the worker |
964 |
< |
* @param worked false if the worker scanned for work but didn't |
802 |
< |
* find any (in which case it may block waiting for work). |
964 |
> |
* @param ran true if worker ran a task since last call to this method |
965 |
|
*/ |
966 |
< |
final void preStep(ForkJoinWorkerThread w, boolean worked) { |
966 |
> |
final void preStep(ForkJoinWorkerThread w, boolean ran) { |
967 |
> |
int wec = w.lastEventCount; |
968 |
|
boolean active = w.active; |
969 |
< |
boolean inactivate = !worked & active; |
970 |
< |
for (;;) { |
971 |
< |
if (inactivate) { |
972 |
< |
int rs = runState; |
973 |
< |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, |
974 |
< |
rs, rs - ONE_ACTIVE)) |
975 |
< |
inactivate = active = w.active = false; |
813 |
< |
} |
969 |
> |
boolean inactivate = false; |
970 |
> |
int pc = parallelism; |
971 |
> |
int rs; |
972 |
> |
while (w.runState == 0 && (rs = runState) < TERMINATING) { |
973 |
> |
if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) && |
974 |
> |
UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1)) |
975 |
> |
inactivate = active = w.active = false; |
976 |
|
int wc = workerCounts; |
977 |
< |
if ((wc & RUNNING_COUNT_MASK) <= parallelism) { |
978 |
< |
if (!worked) |
979 |
< |
eventSync(w); |
980 |
< |
return; |
977 |
> |
if ((wc & RUNNING_COUNT_MASK) > pc) { |
978 |
> |
if (!(inactivate |= active) && // must inactivate to suspend |
979 |
> |
workerCounts == wc && // try to suspend as spare |
980 |
> |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
981 |
> |
wc, wc - ONE_RUNNING)) |
982 |
> |
w.suspendAsSpare(); |
983 |
|
} |
984 |
< |
if (!(inactivate |= active) && // must inactivate to suspend |
985 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
986 |
< |
wc, wc - ONE_RUNNING) && |
987 |
< |
!w.suspendAsSpare()) // false if trimmed |
988 |
< |
return; |
984 |
> |
else if ((wc >>> TOTAL_COUNT_SHIFT) < pc) |
985 |
> |
helpMaintainParallelism(); // not enough workers |
986 |
> |
else if (!ran) { |
987 |
> |
long h = eventWaiters; |
988 |
> |
int ec = eventCount; |
989 |
> |
if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec) |
990 |
> |
releaseEventWaiters(); // release others before waiting |
991 |
> |
else if (ec != wec) { |
992 |
> |
w.lastEventCount = ec; // no need to wait |
993 |
> |
break; |
994 |
> |
} |
995 |
> |
else if (!(inactivate |= active)) |
996 |
> |
eventSync(w, wec); // must inactivate before sync |
997 |
> |
} |
998 |
> |
else |
999 |
> |
break; |
1000 |
|
} |
1001 |
|
} |
1002 |
|
|
1003 |
|
/** |
1004 |
< |
* Tries to decrement running count, and if so, possibly creates |
1005 |
< |
* or resumes compensating threads before blocking on task joinMe. |
831 |
< |
* This code is sprawled out with manual inlining to evade some |
832 |
< |
* JIT oddities. |
1004 |
> |
* Helps and/or blocks awaiting join of the given task. |
1005 |
> |
* See above for explanation. |
1006 |
|
* |
1007 |
|
* @param joinMe the task to join |
1008 |
< |
* @return task status on exit |
1008 |
> |
* @param worker the current worker thread |
1009 |
|
*/ |
1010 |
< |
final int tryAwaitJoin(ForkJoinTask<?> joinMe) { |
1011 |
< |
int cw = workerCounts; // read now to spoil CAS if counts change as ... |
1012 |
< |
releaseWaiters(); // ... a byproduct of releaseWaiters |
1013 |
< |
int stat = joinMe.status; |
1014 |
< |
if (stat >= 0 && // inline variant of tryDecrementRunningCount |
1015 |
< |
(cw & RUNNING_COUNT_MASK) > 0 && |
1016 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
1017 |
< |
cw, cw - ONE_RUNNING)) { |
1018 |
< |
int pc = parallelism; |
1019 |
< |
int scans = 0; // to require confirming passes to add threads |
1020 |
< |
outer: while ((workerCounts & RUNNING_COUNT_MASK) < pc) { |
1021 |
< |
if ((stat = joinMe.status) < 0) |
1022 |
< |
break; |
1023 |
< |
ForkJoinWorkerThread spare = null; |
1024 |
< |
ForkJoinWorkerThread[] ws = workers; |
1025 |
< |
int nws = ws.length; |
1026 |
< |
for (int i = 0; i < nws; ++i) { |
1027 |
< |
ForkJoinWorkerThread w = ws[i]; |
1028 |
< |
if (w != null && w.isSuspended()) { |
1029 |
< |
spare = w; |
1030 |
< |
break; |
1031 |
< |
} |
1032 |
< |
} |
1033 |
< |
if ((stat = joinMe.status) < 0) // recheck to narrow race |
1034 |
< |
break; |
1035 |
< |
int wc = workerCounts; |
1036 |
< |
int rc = wc & RUNNING_COUNT_MASK; |
864 |
< |
if (rc >= pc) |
865 |
< |
break; |
866 |
< |
if (spare != null) { |
867 |
< |
if (spare.tryUnsuspend()) { |
868 |
< |
int c; // inline incrementRunningCount |
869 |
< |
do {} while (!UNSAFE.compareAndSwapInt |
870 |
< |
(this, workerCountsOffset, |
871 |
< |
c = workerCounts, c + ONE_RUNNING)); |
872 |
< |
LockSupport.unpark(spare); |
873 |
< |
break; |
874 |
< |
} |
875 |
< |
continue; |
876 |
< |
} |
877 |
< |
int tc = wc >>> TOTAL_COUNT_SHIFT; |
878 |
< |
int sc = tc - pc; |
879 |
< |
if (rc > 0) { |
880 |
< |
int p = pc; |
881 |
< |
int s = sc; |
882 |
< |
while (s-- >= 0) { // try keeping 3/4 live |
883 |
< |
if (rc > (p -= (p >>> 2) + 1)) |
884 |
< |
break outer; |
885 |
< |
} |
886 |
< |
} |
887 |
< |
if (scans++ > sc && tc < MAX_THREADS && |
888 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc, |
889 |
< |
wc + (ONE_RUNNING|ONE_TOTAL))) { |
890 |
< |
addWorker(); |
891 |
< |
break; |
892 |
< |
} |
1010 |
> |
final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) { |
1011 |
> |
int retries = 2 + (parallelism >> 2); // #helpJoins before blocking |
1012 |
> |
while (joinMe.status >= 0) { |
1013 |
> |
int wc; |
1014 |
> |
worker.helpJoinTask(joinMe); |
1015 |
> |
if (joinMe.status < 0) |
1016 |
> |
break; |
1017 |
> |
else if (retries > 0) |
1018 |
> |
--retries; |
1019 |
> |
else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 && |
1020 |
> |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
1021 |
> |
wc, wc - ONE_RUNNING)) { |
1022 |
> |
int stat, c; long h; |
1023 |
> |
while ((stat = joinMe.status) >= 0 && |
1024 |
> |
(h = eventWaiters) != 0L && // help release others |
1025 |
> |
(int)(h >>> EVENT_COUNT_SHIFT) != eventCount) |
1026 |
> |
releaseEventWaiters(); |
1027 |
> |
if (stat >= 0 && |
1028 |
> |
((workerCounts & RUNNING_COUNT_MASK) == 0 || |
1029 |
> |
(stat = |
1030 |
> |
joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0)) |
1031 |
> |
helpMaintainParallelism(); // timeout or no running workers |
1032 |
> |
do {} while (!UNSAFE.compareAndSwapInt |
1033 |
> |
(this, workerCountsOffset, |
1034 |
> |
c = workerCounts, c + ONE_RUNNING)); |
1035 |
> |
if (stat < 0) |
1036 |
> |
break; // else restart |
1037 |
|
} |
894 |
– |
if (stat >= 0) |
895 |
– |
stat = joinMe.internalAwaitDone(); |
896 |
– |
int c; // inline incrementRunningCount |
897 |
– |
do {} while (!UNSAFE.compareAndSwapInt |
898 |
– |
(this, workerCountsOffset, |
899 |
– |
c = workerCounts, c + ONE_RUNNING)); |
1038 |
|
} |
901 |
– |
return stat; |
1039 |
|
} |
1040 |
|
|
1041 |
|
/** |
1042 |
< |
* Same idea as (and mostly pasted from) tryAwaitJoin, but |
906 |
< |
* self-contained |
1042 |
> |
* Same idea as awaitJoin, but no helping, retries, or timeouts. |
1043 |
|
*/ |
1044 |
|
final void awaitBlocker(ManagedBlocker blocker) |
1045 |
|
throws InterruptedException { |
1046 |
< |
for (;;) { |
911 |
< |
if (blocker.isReleasable()) |
912 |
< |
return; |
913 |
< |
int cw = workerCounts; |
914 |
< |
releaseWaiters(); |
915 |
< |
if ((cw & RUNNING_COUNT_MASK) > 0 && |
916 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
917 |
< |
cw, cw - ONE_RUNNING)) |
918 |
< |
break; |
919 |
< |
} |
920 |
< |
boolean done = false; |
921 |
< |
int pc = parallelism; |
922 |
< |
int scans = 0; |
923 |
< |
outer: while ((workerCounts & RUNNING_COUNT_MASK) < pc) { |
924 |
< |
if (done = blocker.isReleasable()) |
925 |
< |
break; |
926 |
< |
ForkJoinWorkerThread spare = null; |
927 |
< |
ForkJoinWorkerThread[] ws = workers; |
928 |
< |
int nws = ws.length; |
929 |
< |
for (int i = 0; i < nws; ++i) { |
930 |
< |
ForkJoinWorkerThread w = ws[i]; |
931 |
< |
if (w != null && w.isSuspended()) { |
932 |
< |
spare = w; |
933 |
< |
break; |
934 |
< |
} |
935 |
< |
} |
936 |
< |
if (done = blocker.isReleasable()) |
937 |
< |
break; |
1046 |
> |
while (!blocker.isReleasable()) { |
1047 |
|
int wc = workerCounts; |
1048 |
< |
int rc = wc & RUNNING_COUNT_MASK; |
1049 |
< |
if (rc >= pc) |
1050 |
< |
break; |
1051 |
< |
if (spare != null) { |
1052 |
< |
if (spare.tryUnsuspend()) { |
1048 |
> |
if ((wc & RUNNING_COUNT_MASK) != 0 && |
1049 |
> |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
1050 |
> |
wc, wc - ONE_RUNNING)) { |
1051 |
> |
try { |
1052 |
> |
while (!blocker.isReleasable()) { |
1053 |
> |
long h = eventWaiters; |
1054 |
> |
if (h != 0L && |
1055 |
> |
(int)(h >>> EVENT_COUNT_SHIFT) != eventCount) |
1056 |
> |
releaseEventWaiters(); |
1057 |
> |
else if ((workerCounts & RUNNING_COUNT_MASK) == 0 && |
1058 |
> |
runState < TERMINATING) |
1059 |
> |
helpMaintainParallelism(); |
1060 |
> |
else if (blocker.block()) |
1061 |
> |
break; |
1062 |
> |
} |
1063 |
> |
} finally { |
1064 |
|
int c; |
1065 |
|
do {} while (!UNSAFE.compareAndSwapInt |
1066 |
|
(this, workerCountsOffset, |
1067 |
|
c = workerCounts, c + ONE_RUNNING)); |
948 |
– |
LockSupport.unpark(spare); |
949 |
– |
break; |
950 |
– |
} |
951 |
– |
continue; |
952 |
– |
} |
953 |
– |
int tc = wc >>> TOTAL_COUNT_SHIFT; |
954 |
– |
int sc = tc - pc; |
955 |
– |
if (rc > 0) { |
956 |
– |
int p = pc; |
957 |
– |
int s = sc; |
958 |
– |
while (s-- >= 0) { |
959 |
– |
if (rc > (p -= (p >>> 2) + 1)) |
960 |
– |
break outer; |
1068 |
|
} |
962 |
– |
} |
963 |
– |
if (scans++ > sc && tc < MAX_THREADS && |
964 |
– |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc, |
965 |
– |
wc + (ONE_RUNNING|ONE_TOTAL))) { |
966 |
– |
addWorker(); |
1069 |
|
break; |
1070 |
|
} |
1071 |
|
} |
1072 |
< |
try { |
971 |
< |
if (!done) |
972 |
< |
do {} while (!blocker.isReleasable() && |
973 |
< |
!blocker.block()); |
974 |
< |
} finally { |
975 |
< |
int c; |
976 |
< |
do {} while (!UNSAFE.compareAndSwapInt |
977 |
< |
(this, workerCountsOffset, |
978 |
< |
c = workerCounts, c + ONE_RUNNING)); |
979 |
< |
} |
980 |
< |
} |
1072 |
> |
} |
1073 |
|
|
1074 |
|
/** |
1075 |
|
* Possibly initiates and/or completes termination. |
1099 |
|
|
1100 |
|
/** |
1101 |
|
* Actions on transition to TERMINATING |
1102 |
+ |
* |
1103 |
+ |
* Runs up to four passes through workers: (0) shutting down each |
1104 |
+ |
* (without waking up if parked) to quickly spread notifications |
1105 |
+ |
* without unnecessary bouncing around event queues etc (1) wake |
1106 |
+ |
* up and help cancel tasks (2) interrupt (3) mop up races with |
1107 |
+ |
* interrupted workers |
1108 |
|
*/ |
1109 |
|
private void startTerminating() { |
1110 |
< |
for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers |
1111 |
< |
cancelSubmissions(); |
1112 |
< |
shutdownWorkers(); |
1113 |
< |
cancelWorkerTasks(); |
1114 |
< |
signalEvent(); |
1115 |
< |
interruptWorkers(); |
1110 |
> |
cancelSubmissions(); |
1111 |
> |
for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) { |
1112 |
> |
int c; // advance event count |
1113 |
> |
UNSAFE.compareAndSwapInt(this, eventCountOffset, |
1114 |
> |
c = eventCount, c+1); |
1115 |
> |
eventWaiters = 0L; // clobber lists |
1116 |
> |
spareWaiters = 0; |
1117 |
> |
for (ForkJoinWorkerThread w : workers) { |
1118 |
> |
if (w != null) { |
1119 |
> |
w.shutdown(); |
1120 |
> |
if (passes > 0 && !w.isTerminated()) { |
1121 |
> |
w.cancelTasks(); |
1122 |
> |
LockSupport.unpark(w); |
1123 |
> |
if (passes > 1) { |
1124 |
> |
try { |
1125 |
> |
w.interrupt(); |
1126 |
> |
} catch (SecurityException ignore) { |
1127 |
> |
} |
1128 |
> |
} |
1129 |
> |
} |
1130 |
> |
} |
1131 |
> |
} |
1132 |
|
} |
1133 |
|
} |
1134 |
|
|
1135 |
|
/** |
1136 |
< |
* Clear out and cancel submissions, ignoring exceptions |
1136 |
> |
* Clears out and cancels submissions, ignoring exceptions. |
1137 |
|
*/ |
1138 |
|
private void cancelSubmissions() { |
1139 |
|
ForkJoinTask<?> task; |
1145 |
|
} |
1146 |
|
} |
1147 |
|
|
1034 |
– |
/** |
1035 |
– |
* Sets all worker run states to at least shutdown, |
1036 |
– |
* also resuming suspended workers |
1037 |
– |
*/ |
1038 |
– |
private void shutdownWorkers() { |
1039 |
– |
ForkJoinWorkerThread[] ws = workers; |
1040 |
– |
int nws = ws.length; |
1041 |
– |
for (int i = 0; i < nws; ++i) { |
1042 |
– |
ForkJoinWorkerThread w = ws[i]; |
1043 |
– |
if (w != null) |
1044 |
– |
w.shutdown(); |
1045 |
– |
} |
1046 |
– |
} |
1047 |
– |
|
1048 |
– |
/** |
1049 |
– |
* Clears out and cancels all locally queued tasks |
1050 |
– |
*/ |
1051 |
– |
private void cancelWorkerTasks() { |
1052 |
– |
ForkJoinWorkerThread[] ws = workers; |
1053 |
– |
int nws = ws.length; |
1054 |
– |
for (int i = 0; i < nws; ++i) { |
1055 |
– |
ForkJoinWorkerThread w = ws[i]; |
1056 |
– |
if (w != null) |
1057 |
– |
w.cancelTasks(); |
1058 |
– |
} |
1059 |
– |
} |
1060 |
– |
|
1061 |
– |
/** |
1062 |
– |
* Unsticks all workers blocked on joins etc |
1063 |
– |
*/ |
1064 |
– |
private void interruptWorkers() { |
1065 |
– |
ForkJoinWorkerThread[] ws = workers; |
1066 |
– |
int nws = ws.length; |
1067 |
– |
for (int i = 0; i < nws; ++i) { |
1068 |
– |
ForkJoinWorkerThread w = ws[i]; |
1069 |
– |
if (w != null && !w.isTerminated()) { |
1070 |
– |
try { |
1071 |
– |
w.interrupt(); |
1072 |
– |
} catch (SecurityException ignore) { |
1073 |
– |
} |
1074 |
– |
} |
1075 |
– |
} |
1076 |
– |
} |
1077 |
– |
|
1148 |
|
// misc support for ForkJoinWorkerThread |
1149 |
|
|
1150 |
|
/** |
1151 |
< |
* Returns pool number |
1151 |
> |
* Returns pool number. |
1152 |
|
*/ |
1153 |
|
final int getPoolNumber() { |
1154 |
|
return poolNumber; |
1155 |
|
} |
1156 |
|
|
1157 |
|
/** |
1158 |
< |
* Accumulates steal count from a worker, clearing |
1159 |
< |
* the worker's value |
1158 |
> |
* Tries to accumulate steal count from a worker, clearing |
1159 |
> |
* the worker's value if successful. |
1160 |
> |
* |
1161 |
> |
* @return true if worker steal count now zero |
1162 |
|
*/ |
1163 |
< |
final void accumulateStealCount(ForkJoinWorkerThread w) { |
1163 |
> |
final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) { |
1164 |
|
int sc = w.stealCount; |
1165 |
< |
if (sc != 0) { |
1166 |
< |
long c; |
1167 |
< |
w.stealCount = 0; |
1168 |
< |
do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset, |
1169 |
< |
c = stealCount, c + sc)); |
1165 |
> |
long c = stealCount; |
1166 |
> |
// CAS even if zero, for fence effects |
1167 |
> |
if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) { |
1168 |
> |
if (sc != 0) |
1169 |
> |
w.stealCount = 0; |
1170 |
> |
return true; |
1171 |
|
} |
1172 |
+ |
return sc == 0; |
1173 |
|
} |
1174 |
|
|
1175 |
|
/** |
1177 |
|
* active thread. |
1178 |
|
*/ |
1179 |
|
final int idlePerActive() { |
1180 |
< |
int pc = parallelism; // use targeted parallelism, not rc |
1181 |
< |
int ac = runState; // no mask -- artifically boosts during shutdown |
1180 |
> |
int pc = parallelism; // use parallelism, not rc |
1181 |
> |
int ac = runState; // no mask -- artificially boosts during shutdown |
1182 |
|
// Use exact results for small values, saturate past 4 |
1183 |
< |
return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3; |
1183 |
> |
return ((pc <= ac) ? 0 : |
1184 |
> |
(pc >>> 1 <= ac) ? 1 : |
1185 |
> |
(pc >>> 2 <= ac) ? 3 : |
1186 |
> |
pc >>> 3); |
1187 |
|
} |
1188 |
|
|
1189 |
|
// Public and protected methods |
1231 |
|
* use {@link java.lang.Runtime#availableProcessors}. |
1232 |
|
* @param factory the factory for creating new threads. For default value, |
1233 |
|
* use {@link #defaultForkJoinWorkerThreadFactory}. |
1234 |
< |
* @param handler the handler for internal worker threads that |
1235 |
< |
* terminate due to unrecoverable errors encountered while executing |
1236 |
< |
* tasks. For default value, use <code>null</code>. |
1237 |
< |
* @param asyncMode if true, |
1234 |
> |
* @param handler the handler for internal worker threads that |
1235 |
> |
* terminate due to unrecoverable errors encountered while executing |
1236 |
> |
* tasks. For default value, use {@code null}. |
1237 |
> |
* @param asyncMode if true, |
1238 |
|
* establishes local first-in-first-out scheduling mode for forked |
1239 |
|
* tasks that are never joined. This mode may be more appropriate |
1240 |
|
* than default locally stack-based mode in applications in which |
1241 |
|
* worker threads only process event-style asynchronous tasks. |
1242 |
< |
* For default value, use <code>false</code>. |
1242 |
> |
* For default value, use {@code false}. |
1243 |
|
* @throws IllegalArgumentException if parallelism less than or |
1244 |
|
* equal to zero, or greater than implementation limit |
1245 |
|
* @throws NullPointerException if the factory is null |
1248 |
|
* because it does not hold {@link |
1249 |
|
* java.lang.RuntimePermission}{@code ("modifyThread")} |
1250 |
|
*/ |
1251 |
< |
public ForkJoinPool(int parallelism, |
1251 |
> |
public ForkJoinPool(int parallelism, |
1252 |
|
ForkJoinWorkerThreadFactory factory, |
1253 |
|
Thread.UncaughtExceptionHandler handler, |
1254 |
|
boolean asyncMode) { |
1255 |
|
checkPermission(); |
1256 |
|
if (factory == null) |
1257 |
|
throw new NullPointerException(); |
1258 |
< |
if (parallelism <= 0 || parallelism > MAX_THREADS) |
1258 |
> |
if (parallelism <= 0 || parallelism > MAX_WORKERS) |
1259 |
|
throw new IllegalArgumentException(); |
1260 |
|
this.parallelism = parallelism; |
1261 |
|
this.factory = factory; |
1274 |
|
* @param pc the initial parallelism level |
1275 |
|
*/ |
1276 |
|
private static int initialArraySizeFor(int pc) { |
1277 |
< |
// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16) |
1278 |
< |
int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS; |
1277 |
> |
// If possible, initially allocate enough space for one spare |
1278 |
> |
int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS; |
1279 |
> |
// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16) |
1280 |
|
size |= size >>> 1; |
1281 |
|
size |= size >>> 2; |
1282 |
|
size |= size >>> 4; |
1294 |
|
throw new NullPointerException(); |
1295 |
|
if (runState >= SHUTDOWN) |
1296 |
|
throw new RejectedExecutionException(); |
1297 |
< |
// Convert submissions to current pool into forks |
1298 |
< |
Thread t = Thread.currentThread(); |
1299 |
< |
ForkJoinWorkerThread w; |
1300 |
< |
if ((t instanceof ForkJoinWorkerThread) && |
1223 |
< |
(w = (ForkJoinWorkerThread) t).pool == this) |
1224 |
< |
w.pushTask(task); |
1225 |
< |
else { |
1226 |
< |
submissionQueue.offer(task); |
1227 |
< |
signalEvent(); |
1228 |
< |
ensureEnoughTotalWorkers(); |
1229 |
< |
} |
1297 |
> |
submissionQueue.offer(task); |
1298 |
> |
int c; // try to increment event count -- CAS failure OK |
1299 |
> |
UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1); |
1300 |
> |
helpMaintainParallelism(); // create, start, or resume some workers |
1301 |
|
} |
1302 |
|
|
1303 |
|
/** |
1304 |
|
* Performs the given task, returning its result upon completion. |
1234 |
– |
* If the caller is already engaged in a fork/join computation in |
1235 |
– |
* the current pool, this method is equivalent in effect to |
1236 |
– |
* {@link ForkJoinTask#invoke}. |
1305 |
|
* |
1306 |
|
* @param task the task |
1307 |
|
* @return the task's result |
1316 |
|
|
1317 |
|
/** |
1318 |
|
* Arranges for (asynchronous) execution of the given task. |
1251 |
– |
* If the caller is already engaged in a fork/join computation in |
1252 |
– |
* the current pool, this method is equivalent in effect to |
1253 |
– |
* {@link ForkJoinTask#fork}. |
1319 |
|
* |
1320 |
|
* @param task the task |
1321 |
|
* @throws NullPointerException if the task is null |
1344 |
|
|
1345 |
|
/** |
1346 |
|
* Submits a ForkJoinTask for execution. |
1282 |
– |
* If the caller is already engaged in a fork/join computation in |
1283 |
– |
* the current pool, this method is equivalent in effect to |
1284 |
– |
* {@link ForkJoinTask#fork}. |
1347 |
|
* |
1348 |
|
* @param task the task to submit |
1349 |
|
* @return the task |
1449 |
|
|
1450 |
|
/** |
1451 |
|
* Returns the number of worker threads that have started but not |
1452 |
< |
* yet terminated. This result returned by this method may differ |
1452 |
> |
* yet terminated. The result returned by this method may differ |
1453 |
|
* from {@link #getParallelism} when threads are created to |
1454 |
|
* maintain parallelism when others are cooperatively blocked. |
1455 |
|
* |
1534 |
|
*/ |
1535 |
|
public long getQueuedTaskCount() { |
1536 |
|
long count = 0; |
1537 |
< |
ForkJoinWorkerThread[] ws = workers; |
1476 |
< |
int nws = ws.length; |
1477 |
< |
for (int i = 0; i < nws; ++i) { |
1478 |
< |
ForkJoinWorkerThread w = ws[i]; |
1537 |
> |
for (ForkJoinWorkerThread w : workers) |
1538 |
|
if (w != null) |
1539 |
|
count += w.getQueueSize(); |
1481 |
– |
} |
1540 |
|
return count; |
1541 |
|
} |
1542 |
|
|
1590 |
|
* @return the number of elements transferred |
1591 |
|
*/ |
1592 |
|
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
1593 |
< |
int n = submissionQueue.drainTo(c); |
1594 |
< |
ForkJoinWorkerThread[] ws = workers; |
1537 |
< |
int nws = ws.length; |
1538 |
< |
for (int i = 0; i < nws; ++i) { |
1539 |
< |
ForkJoinWorkerThread w = ws[i]; |
1540 |
< |
if (w != null) |
1541 |
< |
n += w.drainTasksTo(c); |
1542 |
< |
} |
1543 |
< |
return n; |
1544 |
< |
} |
1545 |
< |
|
1546 |
< |
/** |
1547 |
< |
* Returns count of total parks by existing workers. |
1548 |
< |
* Used during development only since not meaningful to users. |
1549 |
< |
*/ |
1550 |
< |
private int collectParkCount() { |
1551 |
< |
int count = 0; |
1552 |
< |
ForkJoinWorkerThread[] ws = workers; |
1553 |
< |
int nws = ws.length; |
1554 |
< |
for (int i = 0; i < nws; ++i) { |
1555 |
< |
ForkJoinWorkerThread w = ws[i]; |
1593 |
> |
int count = submissionQueue.drainTo(c); |
1594 |
> |
for (ForkJoinWorkerThread w : workers) |
1595 |
|
if (w != null) |
1596 |
< |
count += w.parkCount; |
1558 |
< |
} |
1596 |
> |
count += w.drainTasksTo(c); |
1597 |
|
return count; |
1598 |
|
} |
1599 |
|
|
1614 |
|
int pc = parallelism; |
1615 |
|
int rs = runState; |
1616 |
|
int ac = rs & ACTIVE_COUNT_MASK; |
1579 |
– |
// int pk = collectParkCount(); |
1617 |
|
return super.toString() + |
1618 |
|
"[" + runLevelToString(rs) + |
1619 |
|
", parallelism = " + pc + |
1623 |
|
", steals = " + st + |
1624 |
|
", tasks = " + qt + |
1625 |
|
", submissions = " + qs + |
1589 |
– |
// ", parks = " + pk + |
1626 |
|
"]"; |
1627 |
|
} |
1628 |
|
|
1720 |
|
throws InterruptedException { |
1721 |
|
try { |
1722 |
|
return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0; |
1723 |
< |
} catch(TimeoutException ex) { |
1723 |
> |
} catch (TimeoutException ex) { |
1724 |
|
return false; |
1725 |
|
} |
1726 |
|
} |
1729 |
|
* Interface for extending managed parallelism for tasks running |
1730 |
|
* in {@link ForkJoinPool}s. |
1731 |
|
* |
1732 |
< |
* <p>A {@code ManagedBlocker} provides two methods. |
1733 |
< |
* Method {@code isReleasable} must return {@code true} if |
1734 |
< |
* blocking is not necessary. Method {@code block} blocks the |
1735 |
< |
* current thread if necessary (perhaps internally invoking |
1736 |
< |
* {@code isReleasable} before actually blocking). |
1732 |
> |
* <p>A {@code ManagedBlocker} provides two methods. Method |
1733 |
> |
* {@code isReleasable} must return {@code true} if blocking is |
1734 |
> |
* not necessary. Method {@code block} blocks the current thread |
1735 |
> |
* if necessary (perhaps internally invoking {@code isReleasable} |
1736 |
> |
* before actually blocking). The unusual methods in this API |
1737 |
> |
* accommodate synchronizers that may, but don't usually, block |
1738 |
> |
* for long periods. Similarly, they allow more efficient internal |
1739 |
> |
* handling of cases in which additional workers may be, but |
1740 |
> |
* usually are not, needed to ensure sufficient parallelism. |
1741 |
> |
* Toward this end, implementations of method {@code isReleasable} |
1742 |
> |
* must be amenable to repeated invocation. |
1743 |
|
* |
1744 |
|
* <p>For example, here is a ManagedBlocker based on a |
1745 |
|
* ReentrantLock: |
1757 |
|
* return hasLock || (hasLock = lock.tryLock()); |
1758 |
|
* } |
1759 |
|
* }}</pre> |
1760 |
+ |
* |
1761 |
+ |
* <p>Here is a class that possibly blocks waiting for an |
1762 |
+ |
* item on a given queue: |
1763 |
+ |
* <pre> {@code |
1764 |
+ |
* class QueueTaker<E> implements ManagedBlocker { |
1765 |
+ |
* final BlockingQueue<E> queue; |
1766 |
+ |
* volatile E item = null; |
1767 |
+ |
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
1768 |
+ |
* public boolean block() throws InterruptedException { |
1769 |
+ |
* if (item == null) |
1770 |
+ |
* item = queue.take(); |
1771 |
+ |
* return true; |
1772 |
+ |
* } |
1773 |
+ |
* public boolean isReleasable() { |
1774 |
+ |
* return item != null || (item = queue.poll()) != null; |
1775 |
+ |
* } |
1776 |
+ |
* public E getItem() { // call after pool.managedBlock completes |
1777 |
+ |
* return item; |
1778 |
+ |
* } |
1779 |
+ |
* }}</pre> |
1780 |
|
*/ |
1781 |
|
public static interface ManagedBlocker { |
1782 |
|
/** |
1819 |
|
public static void managedBlock(ManagedBlocker blocker) |
1820 |
|
throws InterruptedException { |
1821 |
|
Thread t = Thread.currentThread(); |
1822 |
< |
if (t instanceof ForkJoinWorkerThread) |
1823 |
< |
((ForkJoinWorkerThread) t).pool.awaitBlocker(blocker); |
1822 |
> |
if (t instanceof ForkJoinWorkerThread) { |
1823 |
> |
ForkJoinWorkerThread w = (ForkJoinWorkerThread) t; |
1824 |
> |
w.pool.awaitBlocker(blocker); |
1825 |
> |
} |
1826 |
|
else { |
1827 |
|
do {} while (!blocker.isReleasable() && !blocker.block()); |
1828 |
|
} |
1850 |
|
private static final long eventCountOffset = |
1851 |
|
objectFieldOffset("eventCount", ForkJoinPool.class); |
1852 |
|
private static final long eventWaitersOffset = |
1853 |
< |
objectFieldOffset("eventWaiters",ForkJoinPool.class); |
1853 |
> |
objectFieldOffset("eventWaiters", ForkJoinPool.class); |
1854 |
|
private static final long stealCountOffset = |
1855 |
< |
objectFieldOffset("stealCount",ForkJoinPool.class); |
1855 |
> |
objectFieldOffset("stealCount", ForkJoinPool.class); |
1856 |
> |
private static final long spareWaitersOffset = |
1857 |
> |
objectFieldOffset("spareWaiters", ForkJoinPool.class); |
1858 |
|
|
1859 |
|
private static long objectFieldOffset(String field, Class<?> klazz) { |
1860 |
|
try { |