ViewVC Help

View File | Revision Log | Show Annotations | Download File | Root Listing

root/jsr166/jsr166/src/jsr166y/ForkJoinPool.java

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.64 by dl, Tue Aug 17 18:30:32 2010 UTC vs.
Revision 1.109 by jsr166, Fri Jul 1 18:20:11 2011 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
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.Random;
15	+	import java.util.concurrent.AbstractExecutorService;
16	+	import java.util.concurrent.Callable;
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.atomic.AtomicInteger;
23		import java.util.concurrent.locks.LockSupport;
24		import java.util.concurrent.locks.ReentrantLock;
25	<	import java.util.concurrent.atomic.AtomicInteger;
19	<	import java.util.concurrent.CountDownLatch;
25	>	import java.util.concurrent.locks.Condition;
26
27		/**
28		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 69 | Line 75 | import java.util.concurrent.CountDownLat
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>
#	Line 95 | Line 101 | import java.util.concurrent.CountDownLat
101		* daemon} mode, there is typically no need to explicitly {@link
102		* #shutdown} such a pool upon program exit.
103		*
104	<	* <pre>
104	>	*  <pre> {@code
105		* static final ForkJoinPool mainPool = new ForkJoinPool();
106		* ...
107		* public void sort(long[] array) {
108		*   mainPool.invoke(new SortTask(array, 0, array.length));
109	<	* }
104	<	* </pre>
109	>	* }}</pre>
110		*
111		* <p><b>Implementation notes</b>: This implementation restricts the
112		* maximum number of running threads to 32767. Attempts to create
#	Line 124 | Line 129 | public class ForkJoinPool extends Abstra
129		* set of worker threads: Submissions from non-FJ threads enter
130		* into a submission queue. Workers take these tasks and typically
131		* split them into subtasks that may be stolen by other workers.
132	<	* The main work-stealing mechanics implemented in class
133	<	* ForkJoinWorkerThread give first priority to processing tasks
134	<	* from their own queues (LIFO or FIFO, depending on mode), then
135	<	* to randomized FIFO steals of tasks in other worker queues, and
136	<	* lastly to new submissions. These mechanics do not consider
137	<	* affinities, loads, cache localities, etc, so rarely provide the
138	<	* best possible performance on a given machine, but portably
139	<	* provide good throughput by averaging over these factors.
140	<	* (Further, even if we did try to use such information, we do not
141	<	* usually have a basis for exploiting it. For example, some sets
142	<	* of tasks profit from cache affinities, but others are harmed by
143	<	* cache pollution effects.)
144	<	*
145	<	* Beyond work-stealing support and essential bookkeeping, the
146	<	* main responsibility of this framework is to take actions when
147	<	* one worker is waiting to join a task stolen (or always held by)
148	<	* another.  Becauae we are multiplexing many tasks on to a pool
149	<	* of workers, we can't just let them block (as in Thread.join).
150	<	* We also cannot just reassign the joiner's run-time stack with
151	<	* another and replace it later, which would be a form of
152	<	* "continuation", that even if possible is not necessarily a good
153	<	* idea. Given that the creation costs of most threads on most
154	<	* systems mainly surrounds setting up runtime stacks, thread
155	<	* creation and switching is usually not much more expensive than
156	<	* stack creation and switching, and is more flexible). Instead we
132	>	* Preference rules give first priority to processing tasks from
133	>	* their own queues (LIFO or FIFO, depending on mode), then to
134	>	* randomized FIFO steals of tasks in other worker queues, and
135	>	* lastly to new submissions.
136	>	*
137	>	* The main throughput advantages of work-stealing stem from
138	>	* decentralized control -- workers mostly take tasks from
139	>	* themselves or each other. We cannot negate this in the
140	>	* implementation of other management responsibilities. The main
141	>	* tactic for avoiding bottlenecks is packing nearly all
142	>	* essentially atomic control state into a single 64bit volatile
143	>	* variable ("ctl"). This variable is read on the order of 10-100
144	>	* times as often as it is modified (always via CAS). (There is
145	>	* some additional control state, for example variable "shutdown"
146	>	* for which we can cope with uncoordinated updates.)  This
147	>	* streamlines synchronization and control at the expense of messy
148	>	* constructions needed to repack status bits upon updates.
149	>	* Updates tend not to contend with each other except during
150	>	* bursts while submitted tasks begin or end.  In some cases when
151	>	* they do contend, threads can instead do something else
152	>	* (usually, scan for tasks) until contention subsides.
153	>	*
154	>	* To enable packing, we restrict maximum parallelism to (1<<15)-1
155	>	* (which is far in excess of normal operating range) to allow
156	>	* ids, counts, and their negations (used for thresholding) to fit
157	>	* into 16bit fields.
158	>	*
159	>	* Recording Workers.  Workers are recorded in the "workers" array
160	>	* that is created upon pool construction and expanded if (rarely)
161	>	* necessary.  This is an array as opposed to some other data
162	>	* structure to support index-based random steals by workers.
163	>	* Updates to the array recording new workers and unrecording
164	>	* terminated ones are protected from each other by a seqLock
165	>	* (scanGuard) but the array is otherwise concurrently readable,
166	>	* and accessed directly by workers. To simplify index-based
167	>	* operations, the array size is always a power of two, and all
168	>	* readers must tolerate null slots. To avoid flailing during
169	>	* start-up, the array is presized to hold twice #parallelism
170	>	* workers (which is unlikely to need further resizing during
171	>	* execution). But to avoid dealing with so many null slots,
172	>	* variable scanGuard includes a mask for the nearest power of two
173	>	* that contains all current workers.  All worker thread creation
174	>	* is on-demand, triggered by task submissions, replacement of
175	>	* terminated workers, and/or compensation for blocked
176	>	* workers. However, all other support code is set up to work with
177	>	* other policies.  To ensure that we do not hold on to worker
178	>	* references that would prevent GC, ALL accesses to workers are
179	>	* via indices into the workers array (which is one source of some
180	>	* of the messy code constructions here). In essence, the workers
181	>	* array serves as a weak reference mechanism. Thus for example
182	>	* the wait queue field of ctl stores worker indices, not worker
183	>	* references.  Access to the workers in associated methods (for
184	>	* example signalWork) must both index-check and null-check the
185	>	* IDs. All such accesses ignore bad IDs by returning out early
186	>	* from what they are doing, since this can only be associated
187	>	* with termination, in which case it is OK to give up.
188	>	*
189	>	* All uses of the workers array, as well as queue arrays, check
190	>	* that the array is non-null (even if previously non-null). This
191	>	* allows nulling during termination, which is currently not
192	>	* necessary, but remains an option for resource-revocation-based
193	>	* shutdown schemes.
194	>	*
195	>	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
196	>	* let workers spin indefinitely scanning for tasks when none can
197	>	* be found immediately, and we cannot start/resume workers unless
198	>	* there appear to be tasks available.  On the other hand, we must
199	>	* quickly prod them into action when new tasks are submitted or
200	>	* generated.  We park/unpark workers after placing in an event
201	>	* wait queue when they cannot find work. This "queue" is actually
202	>	* a simple Treiber stack, headed by the "id" field of ctl, plus a
203	>	* 15bit counter value to both wake up waiters (by advancing their
204	>	* count) and avoid ABA effects. Successors are held in worker
205	>	* field "nextWait".  Queuing deals with several intrinsic races,
206	>	* mainly that a task-producing thread can miss seeing (and
207	>	* signalling) another thread that gave up looking for work but
208	>	* has not yet entered the wait queue. We solve this by requiring
209	>	* a full sweep of all workers both before (in scan()) and after
210	>	* (in tryAwaitWork()) a newly waiting worker is added to the wait
211	>	* queue. During a rescan, the worker might release some other
212	>	* queued worker rather than itself, which has the same net
213	>	* effect. Because enqueued workers may actually be rescanning
214	>	* rather than waiting, we set and clear the "parked" field of
215	>	* ForkJoinWorkerThread to reduce unnecessary calls to unpark.
216	>	* (Use of the parked field requires a secondary recheck to avoid
217	>	* missed signals.)
218	>	*
219	>	* Signalling.  We create or wake up workers only when there
220	>	* appears to be at least one task they might be able to find and
221	>	* execute.  When a submission is added or another worker adds a
222	>	* task to a queue that previously had two or fewer tasks, they
223	>	* signal waiting workers (or trigger creation of new ones if
224	>	* fewer than the given parallelism level -- see signalWork).
225	>	* These primary signals are buttressed by signals during rescans
226	>	* as well as those performed when a worker steals a task and
227	>	* notices that there are more tasks too; together these cover the
228	>	* signals needed in cases when more than two tasks are pushed
229	>	* but untaken.
230	>	*
231	>	* Trimming workers. To release resources after periods of lack of
232	>	* use, a worker starting to wait when the pool is quiescent will
233	>	* time out and terminate if the pool has remained quiescent for
234	>	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
235	>	* terminating all workers after long periods of non-use.
236	>	*
237	>	* Submissions. External submissions are maintained in an
238	>	* array-based queue that is structured identically to
239	>	* ForkJoinWorkerThread queues except for the use of
240	>	* submissionLock in method addSubmission. Unlike the case for
241	>	* worker queues, multiple external threads can add new
242	>	* submissions, so adding requires a lock.
243	>	*
244	>	* Compensation. Beyond work-stealing support and lifecycle
245	>	* control, the main responsibility of this framework is to take
246	>	* actions when one worker is waiting to join a task stolen (or
247	>	* always held by) another.  Because we are multiplexing many
248	>	* tasks on to a pool of workers, we can't just let them block (as
249	>	* in Thread.join).  We also cannot just reassign the joiner's
250	>	* run-time stack with another and replace it later, which would
251	>	* be a form of "continuation", that even if possible is not
252	>	* necessarily a good idea since we sometimes need both an
253	>	* unblocked task and its continuation to progress. Instead we
254		* combine two tactics:
255		*
256		*   Helping: Arranging for the joiner to execute some task that it
257		*      would be running if the steal had not occurred.  Method
258	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
258	>	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
259		*      links to try to find such a task.
260		*
261		*   Compensating: Unless there are already enough live threads,
262	<	*      method helpMaintainParallelism() may create or or
263	<	*      re-activate a spare thread to compensate for blocked
264	<	*      joiners until they unblock.
163	<	*
164	<	* Because the determining existence of conservatively safe
165	<	* helping targets, the availability of already-created spares,
166	<	* and the apparent need to create new spares are all racy and
167	<	* require heuristic guidance, we rely on multiple retries of
168	<	* each. Further, because it is impossible to keep exactly the
169	<	* target (parallelism) number of threads running at any given
170	<	* time, we allow compensation during joins to fail, and enlist
171	<	* all other threads to help out whenever they are not otherwise
172	<	* occupied (i.e., mainly in method preStep).
262	>	*      method tryPreBlock() may create or re-activate a spare
263	>	*      thread to compensate for blocked joiners until they
264	>	*      unblock.
265		*
266		* The ManagedBlocker extension API can't use helping so relies
267		* only on compensation in method awaitBlocker.
268		*
269	<	* The main throughput advantages of work-stealing stem from
270	<	* decentralized control -- workers mostly steal tasks from each
271	<	* other. We do not want to negate this by creating bottlenecks
272	<	* implementing other management responsibilities. So we use a
273	<	* collection of techniques that avoid, reduce, or cope well with
274	<	* contention. These entail several instances of bit-packing into
275	<	* CASable fields to maintain only the minimally required
276	<	* atomicity. To enable such packing, we restrict maximum
277	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
278	<	* unbalanced increments and decrements) to fit into a 16 bit
279	<	* field, which is far in excess of normal operating range.  Even
280	<	* though updates to some of these bookkeeping fields do sometimes
281	<	* contend with each other, they don't normally cache-contend with
282	<	* updates to others enough to warrant memory padding or
283	<	* isolation. So they are all held as fields of ForkJoinPool
284	<	* objects.  The main capabilities are as follows:
285	<	*
286	<	* 1. Creating and removing workers. Workers are recorded in the
287	<	* "workers" array. This is an array as opposed to some other data
288	<	* structure to support index-based random steals by workers.
289	<	* Updates to the array recording new workers and unrecording
290	<	* terminated ones are protected from each other by a lock
291	<	* (workerLock) but the array is otherwise concurrently readable,
292	<	* and accessed directly by workers. To simplify index-based
293	<	* operations, the array size is always a power of two, and all
294	<	* readers must tolerate null slots. Currently, all worker thread
295	<	* creation is on-demand, triggered by task submissions,
296	<	* replacement of terminated workers, and/or compensation for
297	<	* blocked workers. However, all other support code is set up to
206	<	* work with other policies.
207	<	*
208	<	* To ensure that we do not hold on to worker references that
209	<	* would prevent GC, ALL accesses to workers are via indices into
210	<	* the workers array (which is one source of some of the unusual
211	<	* code constructions here). In essence, the workers array serves
212	<	* as a WeakReference mechanism. Thus for example the event queue
213	<	* stores worker indices, not worker references. Access to the
214	<	* workers in associated methods (for example releaseEventWaiters)
215	<	* must both index-check and null-check the IDs. All such accesses
216	<	* ignore bad IDs by returning out early from what they are doing,
217	<	* since this can only be associated with shutdown, in which case
218	<	* it is OK to give up. On termination, we just clobber these
219	<	* data structures without trying to use them.
220	<	*
221	<	* 2. Bookkeeping for dynamically adding and removing workers. We
222	<	* aim to approximately maintain the given level of parallelism.
223	<	* When some workers are known to be blocked (on joins or via
224	<	* ManagedBlocker), we may create or resume others to take their
225	<	* place until they unblock (see below). Implementing this
226	<	* requires counts of the number of "running" threads (i.e., those
227	<	* that are neither blocked nor artifically suspended) as well as
228	<	* the total number.  These two values are packed into one field,
229	<	* "workerCounts" because we need accurate snapshots when deciding
230	<	* to create, resume or suspend.  Note however that the
231	<	* correspondance of these counts to reality is not guaranteed. In
232	<	* particular updates for unblocked threads may lag until they
233	<	* actually wake up.
234	<	*
235	<	* 3. Maintaining global run state. The run state of the pool
236	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
237	<	* those in other Executor implementations, as well as a count of
238	<	* "active" workers -- those that are, or soon will be, or
239	<	* recently were executing tasks. The runLevel and active count
240	<	* are packed together in order to correctly trigger shutdown and
241	<	* termination. Without care, active counts can be subject to very
242	<	* high contention.  We substantially reduce this contention by
243	<	* relaxing update rules.  A worker must claim active status
244	<	* prospectively, by activating if it sees that a submitted or
245	<	* stealable task exists (it may find after activating that the
246	<	* task no longer exists). It stays active while processing this
247	<	* task (if it exists) and any other local subtasks it produces,
248	<	* until it cannot find any other tasks. It then tries
249	<	* inactivating (see method preStep), but upon update contention
250	<	* instead scans for more tasks, later retrying inactivation if it
251	<	* doesn't find any.
252	<	*
253	<	* 4. Managing idle workers waiting for tasks. We cannot let
254	<	* workers spin indefinitely scanning for tasks when none are
255	<	* available. On the other hand, we must quickly prod them into
256	<	* action when new tasks are submitted or generated.  We
257	<	* park/unpark these idle workers using an event-count scheme.
258	<	* Field eventCount is incremented upon events that may enable
259	<	* workers that previously could not find a task to now find one:
260	<	* Submission of a new task to the pool, or another worker pushing
261	<	* a task onto a previously empty queue.  (We also use this
262	<	* mechanism for configuration and termination actions that
263	<	* require wakeups of idle workers).  Each worker maintains its
264	<	* last known event count, and blocks when a scan for work did not
265	<	* find a task AND its lastEventCount matches the current
266	<	* eventCount. Waiting idle workers are recorded in a variant of
267	<	* Treiber stack headed by field eventWaiters which, when nonzero,
268	<	* encodes the thread index and count awaited for by the worker
269	<	* thread most recently calling eventSync. This thread in turn has
270	<	* a record (field nextEventWaiter) for the next waiting worker.
271	<	* In addition to allowing simpler decisions about need for
272	<	* wakeup, the event count bits in eventWaiters serve the role of
273	<	* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
274	<	* released threads also try to release others (but give up upon
275	<	* contention to reduce useless flailing).  The net effect is a
276	<	* tree-like diffusion of signals, where released threads (and
277	<	* possibly others) help with unparks.  To further reduce
278	<	* contention effects a bit, failed CASes to increment field
279	<	* eventCount are tolerated without retries in signalWork.
280	<	* Conceptually they are merged into the same event, which is OK
281	<	* when their only purpose is to enable workers to scan for work.
282	<	*
283	<	* 5. Managing suspension of extra workers. When a worker is about
284	<	* to block waiting for a join (or via ManagedBlockers), we may
285	<	* create a new thread to maintain parallelism level, or at least
286	<	* avoid starvation. Usually, extra threads are needed for only
287	<	* very short periods, yet join dependencies are such that we
288	<	* sometimes need them in bursts. Rather than create new threads
289	<	* each time this happens, we suspend no-longer-needed extra ones
290	<	* as "spares". For most purposes, we don't distinguish "extra"
291	<	* spare threads from normal "core" threads: On each call to
292	<	* preStep (the only point at which we can do this) a worker
293	<	* checks to see if there are now too many running workers, and if
294	<	* so, suspends itself.  Method helpMaintainParallelism looks for
295	<	* suspended threads to resume before considering creating a new
296	<	* replacement. The spares themselves are encoded on another
297	<	* variant of a Treiber Stack, headed at field "spareWaiters".
298	<	* Note that the use of spares is intrinsically racy.  One thread
299	<	* may become a spare at about the same time as another is
300	<	* needlessly being created. We counteract this and related slop
301	<	* in part by requiring resumed spares to immediately recheck (in
302	<	* preStep) to see whether they they should re-suspend.
303	<	*
304	<	* 6. Killing off unneeded workers. The Spare and Event queues use
305	<	* similar mechanisms to shed unused workers: The oldest (first)
306	<	* waiter uses a timed rather than hard wait. When this wait times
307	<	* out without a normal wakeup, it tries to shutdown any one (for
308	<	* convenience the newest) other waiter via tryShutdownSpare or
309	<	* tryShutdownWaiter, respectively. The wakeup rates for spares
310	<	* are much shorter than for waiters. Together, they will
311	<	* eventually reduce the number of worker threads to a minimum of
312	<	* one after a long enough period without use.
313	<	*
314	<	* 7. Deciding when to create new workers. The main dynamic
315	<	* control in this class is deciding when to create extra threads
316	<	* in method helpMaintainParallelism. We would like to keep
317	<	* exactly #parallelism threads running, which is an impossble
318	<	* task. We always need to create one when the number of running
319	<	* threads would become zero and all workers are busy. Beyond
320	<	* this, we must rely on heuristics that work well in the the
321	<	* presence of transients phenomena such as GC stalls, dynamic
322	<	* compilation, and wake-up lags. These transients are extremely
323	<	* common -- we are normally trying to fully saturate the CPUs on
324	<	* a machine, so almost any activity other than running tasks
325	<	* impedes accuracy. Our main defense is to allow some slack in
326	<	* creation thresholds, using rules that reflect the fact that the
327	<	* more threads we have running, the more likely that we are
328	<	* underestimating the number running threads. (We also include
329	<	* some heuristic use of Thread.yield when all workers appear to
330	<	* be busy, to improve likelihood of counts settling.) The rules
331	<	* also better cope with the fact that some of the methods in this
332	<	* class tend to never become compiled (but are interpreted), so
333	<	* some components of the entire set of controls might execute 100
334	<	* times faster than others. And similarly for cases where the
335	<	* apparent lack of work is just due to GC stalls and other
336	<	* transient system activity.
269	>	* It is impossible to keep exactly the target parallelism number
270	>	* of threads running at any given time.  Determining the
271	>	* existence of conservatively safe helping targets, the
272	>	* availability of already-created spares, and the apparent need
273	>	* to create new spares are all racy and require heuristic
274	>	* guidance, so we rely on multiple retries of each.  Currently,
275	>	* in keeping with on-demand signalling policy, we compensate only
276	>	* if blocking would leave less than one active (non-waiting,
277	>	* non-blocked) worker. Additionally, to avoid some false alarms
278	>	* due to GC, lagging counters, system activity, etc, compensated
279	>	* blocking for joins is only attempted after rechecks stabilize
280	>	* (retries are interspersed with Thread.yield, for good
281	>	* citizenship).  The variable blockedCount, incremented before
282	>	* blocking and decremented after, is sometimes needed to
283	>	* distinguish cases of waiting for work vs blocking on joins or
284	>	* other managed sync. Both cases are equivalent for most pool
285	>	* control, so we can update non-atomically. (Additionally,
286	>	* contention on blockedCount alleviates some contention on ctl).
287	>	*
288	>	* Shutdown and Termination. A call to shutdownNow atomically sets
289	>	* the ctl stop bit and then (non-atomically) sets each workers
290	>	* "terminate" status, cancels all unprocessed tasks, and wakes up
291	>	* all waiting workers.  Detecting whether termination should
292	>	* commence after a non-abrupt shutdown() call requires more work
293	>	* and bookkeeping. We need consensus about quiescence (i.e., that
294	>	* there is no more work) which is reflected in active counts so
295	>	* long as there are no current blockers, as well as possible
296	>	* re-evaluations during independent changes in blocking or
297	>	* quiescing workers.
298		*
299	<	* Beware that there is a lot of representation-level coupling
299	>	* Style notes: There is a lot of representation-level coupling
300		* among classes ForkJoinPool, ForkJoinWorkerThread, and
301	<	* ForkJoinTask.  For example, direct access to "workers" array by
301	>	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
302	>	* data structures managed by ForkJoinPool, so are directly
303	>	* accessed.  Conversely we allow access to "workers" array by
304		* workers, and direct access to ForkJoinTask.status by both
305		* ForkJoinPool and ForkJoinWorkerThread.  There is little point
306		* trying to reduce this, since any associated future changes in
307		* representations will need to be accompanied by algorithmic
308	<	* changes anyway.
309	<	*
310	<	* Style notes: There are lots of inline assignments (of form
311	<	* "while ((local = field) != 0)") which are usually the simplest
312	<	* way to ensure the required read orderings (which are sometimes
313	<	* critical). Also several occurrences of the unusual "do {}
314	<	* while(!cas...)" which is the simplest way to force an update of
315	<	* a CAS'ed variable. There are also other coding oddities that
316	<	* help some methods perform reasonably even when interpreted (not
317	<	* compiled), at the expense of some messy constructions that
318	<	* reduce byte code counts.
319	<	*
320	<	* The order of declarations in this file is: (1) statics (2)
321	<	* fields (along with constants used when unpacking some of them)
322	<	* (3) internal control methods (4) callbacks and other support
323	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
324	<	* methods (plus a few little helpers).
308	>	* changes anyway. All together, these low-level implementation
309	>	* choices produce as much as a factor of 4 performance
310	>	* improvement compared to naive implementations, and enable the
311	>	* processing of billions of tasks per second, at the expense of
312	>	* some ugliness.
313	>	*
314	>	* Methods signalWork() and scan() are the main bottlenecks so are
315	>	* especially heavily micro-optimized/mangled.  There are lots of
316	>	* inline assignments (of form "while ((local = field) != 0)")
317	>	* which are usually the simplest way to ensure the required read
318	>	* orderings (which are sometimes critical). This leads to a
319	>	* "C"-like style of listing declarations of these locals at the
320	>	* heads of methods or blocks.  There are several occurrences of
321	>	* the unusual "do {} while (!cas...)"  which is the simplest way
322	>	* to force an update of a CAS'ed variable. There are also other
323	>	* coding oddities that help some methods perform reasonably even
324	>	* when interpreted (not compiled).
325	>	*
326	>	* The order of declarations in this file is: (1) declarations of
327	>	* statics (2) fields (along with constants used when unpacking
328	>	* some of them), listed in an order that tends to reduce
329	>	* contention among them a bit under most JVMs.  (3) internal
330	>	* control methods (4) callbacks and other support for
331	>	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
332	>	* methods (plus a few little helpers). (6) static block
333	>	* initializing all statics in a minimally dependent order.
334		*/
335
336		/**
#	Line 393 | Line 365 | public class ForkJoinPool extends Abstra
365		* overridden in ForkJoinPool constructors.
366		*/
367		public static final ForkJoinWorkerThreadFactory
368	<	defaultForkJoinWorkerThreadFactory =
397	<	new DefaultForkJoinWorkerThreadFactory();
368	>	defaultForkJoinWorkerThreadFactory;
369
370		/**
371		* Permission required for callers of methods that may start or
372		* kill threads.
373		*/
374	<	private static final RuntimePermission modifyThreadPermission =
404	<	new RuntimePermission("modifyThread");
374	>	private static final RuntimePermission modifyThreadPermission;
375
376		/**
377		* If there is a security manager, makes sure caller has
#	Line 416 | Line 386 | public class ForkJoinPool extends Abstra
386		/**
387		* Generator for assigning sequence numbers as pool names.
388		*/
389	<	private static final AtomicInteger poolNumberGenerator =
420	<	new AtomicInteger();
389	>	private static final AtomicInteger poolNumberGenerator;
390
391		/**
392	<	* The wakeup interval (in nanoseconds) for the oldest worker
393	<	* worker waiting for an event invokes tryShutdownWaiter to shrink
394	<	* the number of workers.  The exact value does not matter too
395	<	* much, but should be long enough to slowly release resources
396	<	* during long periods without use without disrupting normal use.
392	>	* Generator for initial random seeds for worker victim
393	>	* selection. This is used only to create initial seeds. Random
394	>	* steals use a cheaper xorshift generator per steal attempt. We
395	>	* don't expect much contention on seedGenerator, so just use a
396	>	* plain Random.
397		*/
398	<	private static final long SHRINK_RATE_NANOS =
430	<	60L * 1000L * 1000L * 1000L; // one minute
398	>	static final Random workerSeedGenerator;
399
400		/**
401	<	* Absolute bound for parallelism level. Twice this number plus
402	<	* one (i.e., 0xfff) must fit into a 16bit field to enable
403	<	* word-packing for some counts and indices.
401	>	* Array holding all worker threads in the pool.  Initialized upon
402	>	* construction. Array size must be a power of two.  Updates and
403	>	* replacements are protected by scanGuard, but the array is
404	>	* always kept in a consistent enough state to be randomly
405	>	* accessed without locking by workers performing work-stealing,
406	>	* as well as other traversal-based methods in this class, so long
407	>	* as reads memory-acquire by first reading ctl. All readers must
408	>	* tolerate that some array slots may be null.
409		*/
410	<	private static final int MAX_WORKERS   = 0x7fff;
410	>	ForkJoinWorkerThread[] workers;
411
412		/**
413	<	* Array holding all worker threads in the pool.  Array size must
414	<	* be a power of two.  Updates and replacements are protected by
415	<	* workerLock, but the array is always kept in a consistent enough
443	<	* state to be randomly accessed without locking by workers
444	<	* performing work-stealing, as well as other traversal-based
445	<	* methods in this class. All readers must tolerate that some
446	<	* array slots may be null.
413	>	* Initial size for submission queue array. Must be a power of
414	>	* two.  In many applications, these always stay small so we use a
415	>	* small initial cap.
416		*/
417	<	volatile ForkJoinWorkerThread[] workers;
417	>	private static final int INITIAL_QUEUE_CAPACITY = 8;
418
419		/**
420	<	* Queue for external submissions.
420	>	* Maximum size for submission queue array. Must be a power of two
421	>	* less than or equal to 1 << (31 - width of array entry) to
422	>	* ensure lack of index wraparound, but is capped at a lower
423	>	* value to help users trap runaway computations.
424		*/
425	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
425	>	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
426
427		/**
428	<	* Lock protecting updates to workers array.
428	>	* Array serving as submission queue. Initialized upon construction.
429		*/
430	<	private final ReentrantLock workerLock;
430	>	private ForkJoinTask<?>[] submissionQueue;
431
432		/**
433	<	* Latch released upon termination.
433	>	* Lock protecting submissions array for addSubmission
434		*/
435	<	private final Phaser termination;
435	>	private final ReentrantLock submissionLock;
436
437		/**
438	<	* Creation factory for worker threads.
438	>	* Condition for awaitTermination, using submissionLock for
439	>	* convenience.
440		*/
441	<	private final ForkJoinWorkerThreadFactory factory;
441	>	private final Condition termination;
442
443		/**
444	<	* Sum of per-thread steal counts, updated only when threads are
472	<	* idle or terminating.
444	>	* Creation factory for worker threads.
445		*/
446	<	private volatile long stealCount;
446	>	private final ForkJoinWorkerThreadFactory factory;
447
448		/**
449	<	* Encoded record of top of treiber stack of threads waiting for
450	<	* events. The top 32 bits contain the count being waited for. The
479	<	* bottom 16 bits contains one plus the pool index of waiting
480	<	* worker thread. (Bits 16-31 are unused.)
449	>	* The uncaught exception handler used when any worker abruptly
450	>	* terminates.
451		*/
452	<	private volatile long eventWaiters;
483	<
484	<	private static final int  EVENT_COUNT_SHIFT = 32;
485	<	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
452	>	final Thread.UncaughtExceptionHandler ueh;
453
454		/**
455	<	* A counter for events that may wake up worker threads:
489	<	*   - Submission of a new task to the pool
490	<	*   - A worker pushing a task on an empty queue
491	<	*   - termination
455	>	* Prefix for assigning names to worker threads
456		*/
457	<	private volatile int eventCount;
457	>	private final String workerNamePrefix;
458
459		/**
460	<	* Encoded record of top of treiber stack of spare threads waiting
461	<	* for resumption. The top 16 bits contain an arbitrary count to
498	<	* avoid ABA effects. The bottom 16bits contains one plus the pool
499	<	* index of waiting worker thread.
500	<	*/
501	<	private volatile int spareWaiters;
502	<
503	<	private static final int SPARE_COUNT_SHIFT = 16;
504	<	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
505	<
506	<	/**
507	<	* Lifecycle control. The low word contains the number of workers
508	<	* that are (probably) executing tasks. This value is atomically
509	<	* incremented before a worker gets a task to run, and decremented
510	<	* when worker has no tasks and cannot find any.  Bits 16-18
511	<	* contain runLevel value. When all are zero, the pool is
512	<	* running. Level transitions are monotonic (running -> shutdown
513	<	* -> terminating -> terminated) so each transition adds a bit.
514	<	* These are bundled together to ensure consistent read for
515	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
516	<	* and active threads is zero).
517	<	*
518	<	* Notes: Most direct CASes are dependent on these bitfield
519	<	* positions.  Also, this field is non-private to enable direct
520	<	* performance-sensitive CASes in ForkJoinWorkerThread.
460	>	* Sum of per-thread steal counts, updated only when threads are
461	>	* idle or terminating.
462		*/
463	<	volatile int runState;
523	<
524	<	// Note: The order among run level values matters.
525	<	private static final int RUNLEVEL_SHIFT     = 16;
526	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
527	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
528	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
529	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
463	>	private volatile long stealCount;
464
465		/**
466	<	* Holds number of total (i.e., created and not yet terminated)
467	<	* and running (i.e., not blocked on joins or other managed sync)
468	<	* threads, packed together to ensure consistent snapshot when
469	<	* making decisions about creating and suspending spare
470	<	* threads. Updated only by CAS. Note that adding a new worker
471	<	* requires incrementing both counts, since workers start off in
472	<	* running state.
473	<	*/
474	<	private volatile int workerCounts;
475	<
476	<	private static final int TOTAL_COUNT_SHIFT  = 16;
477	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
478	<	private static final int ONE_RUNNING        = 1;
479	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
466	>	* Main pool control -- a long packed with:
467	>	* AC: Number of active running workers minus target parallelism (16 bits)
468	>	* TC: Number of total workers minus target parallelism (16 bits)
469	>	* ST: true if pool is terminating (1 bit)
470	>	* EC: the wait count of top waiting thread (15 bits)
471	>	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
472	>	*
473	>	* When convenient, we can extract the upper 32 bits of counts and
474	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
475	>	* (int)ctl.  The ec field is never accessed alone, but always
476	>	* together with id and st. The offsets of counts by the target
477	>	* parallelism and the positionings of fields makes it possible to
478	>	* perform the most common checks via sign tests of fields: When
479	>	* ac is negative, there are not enough active workers, when tc is
480	>	* negative, there are not enough total workers, when id is
481	>	* negative, there is at least one waiting worker, and when e is
482	>	* negative, the pool is terminating.  To deal with these possibly
483	>	* negative fields, we use casts in and out of "short" and/or
484	>	* signed shifts to maintain signedness.
485	>	*/
486	>	volatile long ctl;
487	>
488	>	// bit positions/shifts for fields
489	>	private static final int  AC_SHIFT   = 48;
490	>	private static final int  TC_SHIFT   = 32;
491	>	private static final int  ST_SHIFT   = 31;
492	>	private static final int  EC_SHIFT   = 16;
493	>
494	>	// bounds
495	>	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
496	>	private static final int  SMASK      = 0xffff;  // mask short bits
497	>	private static final int  SHORT_SIGN = 1 << 15;
498	>	private static final int  INT_SIGN   = 1 << 31;
499	>
500	>	// masks
501	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
502	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
503	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
504	>
505	>	// units for incrementing and decrementing
506	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
507	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
508	>
509	>	// masks and units for dealing with u = (int)(ctl >>> 32)
510	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
511	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
512	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
513	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
514	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
515	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
516	>
517	>	// masks and units for dealing with e = (int)ctl
518	>	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
519	>	private static final int  EC_UNIT    = 1 << EC_SHIFT;
520
521		/**
522		* The target parallelism level.
549	–	* Accessed directly by ForkJoinWorkerThreads.
523		*/
524		final int parallelism;
525
526		/**
527	<	* True if use local fifo, not default lifo, for local polling
528	<	* Read by, and replicated by ForkJoinWorkerThreads
527	>	* Index (mod submission queue length) of next element to take
528	>	* from submission queue. Usage is identical to that for
529	>	* per-worker queues -- see ForkJoinWorkerThread internal
530	>	* documentation.
531		*/
532	<	final boolean locallyFifo;
532	>	volatile int queueBase;
533
534		/**
535	<	* The uncaught exception handler used when any worker abruptly
536	<	* terminates.
535	>	* Index (mod submission queue length) of next element to add
536	>	* in submission queue. Usage is identical to that for
537	>	* per-worker queues -- see ForkJoinWorkerThread internal
538	>	* documentation.
539		*/
540	<	private final Thread.UncaughtExceptionHandler ueh;
540	>	int queueTop;
541
542		/**
543	<	* Pool number, just for assigning useful names to worker threads
543	>	* True when shutdown() has been called.
544		*/
545	<	private final int poolNumber;
569	<
545	>	volatile boolean shutdown;
546
547	<	// Utilities for CASing fields. Note that most of these
548	<	// are usually manually inlined by callers
547	>	/**
548	>	* True if use local fifo, not default lifo, for local polling.
549	>	* Read by, and replicated by ForkJoinWorkerThreads.
550	>	*/
551	>	final boolean locallyFifo;
552
553		/**
554	<	* Increments running count part of workerCounts
554	>	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
555	>	* When non-zero, suppresses automatic shutdown when active
556	>	* counts become zero.
557		*/
558	<	final void incrementRunningCount() {
578	<	int c;
579	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
580	<	c = workerCounts,
581	<	c + ONE_RUNNING));
582	<	}
558	>	volatile int quiescerCount;
559
560		/**
561	<	* Tries to decrement running count unless already zero
561	>	* The number of threads blocked in join.
562		*/
563	<	final boolean tryDecrementRunningCount() {
588	<	int wc = workerCounts;
589	<	if ((wc & RUNNING_COUNT_MASK) == 0)
590	<	return false;
591	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
592	<	wc, wc - ONE_RUNNING);
593	<	}
563	>	volatile int blockedCount;
564
565		/**
566	<	* Forces decrement of encoded workerCounts, awaiting nonzero if
597	<	* (rarely) necessary when other count updates lag.
598	<	*
599	<	* @param dr -- either zero or ONE_RUNNING
600	<	* @param dt == either zero or ONE_TOTAL
566	>	* Counter for worker Thread names (unrelated to their poolIndex)
567		*/
568	<	private void decrementWorkerCounts(int dr, int dt) {
603	<	for (;;) {
604	<	int wc = workerCounts;
605	<	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
606	<	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
607	<	if ((runState & TERMINATED) != 0)
608	<	return; // lagging termination on a backout
609	<	Thread.yield();
610	<	}
611	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
612	<	wc, wc - (dr + dt)))
613	<	return;
614	<	}
615	<	}
568	>	private volatile int nextWorkerNumber;
569
570		/**
571	<	* Increments event count
571	>	* The index for the next created worker. Accessed under scanGuard.
572		*/
573	<	private void advanceEventCount() {
621	<	int c;
622	<	do {} while(!UNSAFE.compareAndSwapInt(this, eventCountOffset,
623	<	c = eventCount, c+1));
624	<	}
573	>	private int nextWorkerIndex;
574
575		/**
576	<	* Tries incrementing active count; fails on contention.
577	<	* Called by workers before executing tasks.
578	<	*
579	<	* @return true on success
576	>	* SeqLock and index masking for updates to workers array.  Locked
577	>	* when SG_UNIT is set. Unlocking clears bit by adding
578	>	* SG_UNIT. Staleness of read-only operations can be checked by
579	>	* comparing scanGuard to value before the reads. The low 16 bits
580	>	* (i.e, anding with SMASK) hold (the smallest power of two
581	>	* covering all worker indices, minus one, and is used to avoid
582	>	* dealing with large numbers of null slots when the workers array
583	>	* is overallocated.
584		*/
585	<	final boolean tryIncrementActiveCount() {
586	<	int c;
587	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
635	<	c = runState, c + 1);
636	<	}
585	>	volatile int scanGuard;
586	>
587	>	private static final int SG_UNIT = 1 << 16;
588
589		/**
590	<	* Tries decrementing active count; fails on contention.
591	<	* Called when workers cannot find tasks to run.
590	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
591	>	* task when the pool is quiescent to instead try to shrink the
592	>	* number of workers.  The exact value does not matter too
593	>	* much. It must be short enough to release resources during
594	>	* sustained periods of idleness, but not so short that threads
595	>	* are continually re-created.
596		*/
597	<	final boolean tryDecrementActiveCount() {
598	<	int c;
644	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
645	<	c = runState, c - 1);
646	<	}
597	>	private static final long SHRINK_RATE =
598	>	4L * 1000L * 1000L * 1000L; // 4 seconds
599
600		/**
601	<	* Advances to at least the given level. Returns true if not
602	<	* already in at least the given level.
601	>	* Top-level loop for worker threads: On each step: if the
602	>	* previous step swept through all queues and found no tasks, or
603	>	* there are excess threads, then possibly blocks. Otherwise,
604	>	* scans for and, if found, executes a task. Returns when pool
605	>	* and/or worker terminate.
606	>	*
607	>	* @param w the worker
608		*/
609	<	private boolean advanceRunLevel(int level) {
610	<	for (;;) {
611	<	int s = runState;
612	<	if ((s & level) != 0)
613	<	return false;
614	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
615	<	return true;
609	>	final void work(ForkJoinWorkerThread w) {
610	>	boolean swept = false;                // true on empty scans
611	>	long c;
612	>	while (!w.terminate && (int)(c = ctl) >= 0) {
613	>	int a;                            // active count
614	>	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
615	>	swept = scan(w, a);
616	>	else if (tryAwaitWork(w, c))
617	>	swept = false;
618		}
619		}
620
621	<	// workers array maintenance
621	>	// Signalling
622
623		/**
624	<	* Records and returns a workers array index for new worker.
624	>	* Wakes up or creates a worker.
625		*/
626	<	private int recordWorker(ForkJoinWorkerThread w) {
627	<	// Try using slot totalCount-1. If not available, scan and/or resize
628	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
629	<	final ReentrantLock lock = this.workerLock;
630	<	lock.lock();
631	<	try {
632	<	ForkJoinWorkerThread[] ws = workers;
633	<	int n = ws.length;
634	<	if (k < 0 || k >= n || ws[k] != null) {
635	<	for (k = 0; k < n && ws[k] != null; ++k)
636	<	;
637	<	if (k == n)
638	<	ws = Arrays.copyOf(ws, n << 1);
626	>	final void signalWork() {
627	>	/*
628	>	* The while condition is true if: (there is are too few total
629	>	* workers OR there is at least one waiter) AND (there are too
630	>	* few active workers OR the pool is terminating).  The value
631	>	* of e distinguishes the remaining cases: zero (no waiters)
632	>	* for create, negative if terminating (in which case do
633	>	* nothing), else release a waiter. The secondary checks for
634	>	* release (non-null array etc) can fail if the pool begins
635	>	* terminating after the test, and don't impose any added cost
636	>	* because JVMs must perform null and bounds checks anyway.
637	>	*/
638	>	long c; int e, u;
639	>	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
640	>	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
641	>	if (e > 0) {                         // release a waiting worker
642	>	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
643	>	if ((ws = workers) == null ||
644	>	(i = ~e & SMASK) >= ws.length ||
645	>	(w = ws[i]) == null)
646	>	break;
647	>	long nc = (((long)(w.nextWait & E_MASK)) |
648	>	((long)(u + UAC_UNIT) << 32));
649	>	if (w.eventCount == e &&
650	>	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
651	>	w.eventCount = (e + EC_UNIT) & E_MASK;
652	>	if (w.parked)
653	>	UNSAFE.unpark(w);
654	>	break;
655	>	}
656	>	}
657	>	else if (UNSAFE.compareAndSwapLong
658	>	(this, ctlOffset, c,
659	>	(long)(((u + UTC_UNIT) & UTC_MASK) |
660	>	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
661	>	addWorker();
662	>	break;
663		}
681	–	ws[k] = w;
682	–	workers = ws; // volatile array write ensures slot visibility
683	–	} finally {
684	–	lock.unlock();
664		}
686	–	return k;
665		}
666
667		/**
668	<	* Nulls out record of worker in workers array
668	>	* Variant of signalWork to help release waiters on rescans.
669	>	* Tries once to release a waiter if active count < 0.
670	>	*
671	>	* @return false if failed due to contention, else true
672		*/
673	<	private void forgetWorker(ForkJoinWorkerThread w) {
674	<	int idx = w.poolIndex;
675	<	// Locking helps method recordWorker avoid unecessary expansion
676	<	final ReentrantLock lock = this.workerLock;
677	<	lock.lock();
678	<	try {
679	<	ForkJoinWorkerThread[] ws = workers;
680	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
681	<	ws[idx] = null;
682	<	} finally {
683	<	lock.unlock();
673	>	private boolean tryReleaseWaiter() {
674	>	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
675	>	if ((e = (int)(c = ctl)) > 0 &&
676	>	(int)(c >> AC_SHIFT) < 0 &&
677	>	(ws = workers) != null &&
678	>	(i = ~e & SMASK) < ws.length &&
679	>	(w = ws[i]) != null) {
680	>	long nc = ((long)(w.nextWait & E_MASK) |
681	>	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
682	>	if (w.eventCount != e ||
683	>	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
684	>	return false;
685	>	w.eventCount = (e + EC_UNIT) & E_MASK;
686	>	if (w.parked)
687	>	UNSAFE.unpark(w);
688		}
689	+	return true;
690		}
691
692	<	// adding and removing workers
692	>	// Scanning for tasks
693
694		/**
695	<	* Tries to create and add new worker. Assumes that worker counts
696	<	* are already updated to accommodate the worker, so adjusts on
697	<	* failure.
695	>	* Scans for and, if found, executes one task. Scans start at a
696	>	* random index of workers array, and randomly select the first
697	>	* (2*#workers)-1 probes, and then, if all empty, resort to 2
698	>	* circular sweeps, which is necessary to check quiescence. and
699	>	* taking a submission only if no stealable tasks were found.  The
700	>	* steal code inside the loop is a specialized form of
701	>	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
702	>	* helpJoinTask and signal propagation. The code for submission
703	>	* queues is almost identical. On each steal, the worker completes
704	>	* not only the task, but also all local tasks that this task may
705	>	* have generated. On detecting staleness or contention when
706	>	* trying to take a task, this method returns without finishing
707	>	* sweep, which allows global state rechecks before retry.
708		*
709	<	* @return the worker, or null on failure
709	>	* @param w the worker
710	>	* @param a the number of active workers
711	>	* @return true if swept all queues without finding a task
712		*/
713	<	private ForkJoinWorkerThread addWorker() {
714	<	ForkJoinWorkerThread w = null;
715	<	try {
716	<	w = factory.newThread(this);
717	<	} finally { // Adjust on either null or exceptional factory return
718	<	if (w == null) {
719	<	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
720	<	tryTerminate(false); // in case of failure during shutdown
713	>	private boolean scan(ForkJoinWorkerThread w, int a) {
714	>	int g = scanGuard; // mask 0 avoids useless scans if only one active
715	>	int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
716	>	ForkJoinWorkerThread[] ws = workers;
717	>	if (ws == null || ws.length <= m)         // staleness check
718	>	return false;
719	>	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
720	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
721	>	ForkJoinWorkerThread v = ws[k & m];
722	>	if (v != null && (b = v.queueBase) != v.queueTop &&
723	>	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
724	>	long u = (i << ASHIFT) + ABASE;
725	>	if ((t = q[i]) != null && v.queueBase == b &&
726	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
727	>	int d = (v.queueBase = b + 1) - v.queueTop;
728	>	v.stealHint = w.poolIndex;
729	>	if (d != 0)
730	>	signalWork();             // propagate if nonempty
731	>	w.execTask(t);
732	>	}
733	>	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
734	>	return false;                     // store next seed
735	>	}
736	>	else if (j < 0) {                     // xorshift
737	>	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
738		}
739	+	else
740	+	++k;
741		}
742	<	if (w != null) {
743	<	w.start(recordWorker(w), ueh);
744	<	advanceEventCount();
742	>	if (scanGuard != g)                       // staleness check
743	>	return false;
744	>	else {                                    // try to take submission
745	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
746	>	if ((b = queueBase) != queueTop &&
747	>	(q = submissionQueue) != null &&
748	>	(i = (q.length - 1) & b) >= 0) {
749	>	long u = (i << ASHIFT) + ABASE;
750	>	if ((t = q[i]) != null && queueBase == b &&
751	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
752	>	queueBase = b + 1;
753	>	w.execTask(t);
754	>	}
755	>	return false;
756	>	}
757	>	return true;                         // all queues empty
758		}
729	–	return w;
759		}
760
761		/**
762	<	* Final callback from terminating worker.  Removes record of
763	<	* worker from array, and adjusts counts. If pool is shutting
764	<	* down, tries to complete terminatation.
765	<	*
766	<	* @param w the worker
767	<	*/
768	<	final void workerTerminated(ForkJoinWorkerThread w) {
769	<	forgetWorker(w);
770	<	decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
771	<	while (w.stealCount != 0) // collect final count
772	<	tryAccumulateStealCount(w);
773	<	tryTerminate(false);
762	>	* Tries to enqueue worker w in wait queue and await change in
763	>	* worker's eventCount.  If the pool is quiescent and there is
764	>	* more than one worker, possibly terminates worker upon exit.
765	>	* Otherwise, before blocking, rescans queues to avoid missed
766	>	* signals.  Upon finding work, releases at least one worker
767	>	* (which may be the current worker). Rescans restart upon
768	>	* detected staleness or failure to release due to
769	>	* contention. Note the unusual conventions about Thread.interrupt
770	>	* here and elsewhere: Because interrupts are used solely to alert
771	>	* threads to check termination, which is checked here anyway, we
772	>	* clear status (using Thread.interrupted) before any call to
773	>	* park, so that park does not immediately return due to status
774	>	* being set via some other unrelated call to interrupt in user
775	>	* code.
776	>	*
777	>	* @param w the calling worker
778	>	* @param c the ctl value on entry
779	>	* @return true if waited or another thread was released upon enq
780	>	*/
781	>	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
782	>	int v = w.eventCount;
783	>	w.nextWait = (int)c;                      // w's successor record
784	>	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
785	>	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
786	>	long d = ctl; // return true if lost to a deq, to force scan
787	>	return (int)d != (int)c && (d & AC_MASK) >= (c & AC_MASK);
788	>	}
789	>	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
790	>	long s = stealCount;
791	>	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
792	>	sc = w.stealCount = 0;
793	>	else if (w.eventCount != v)
794	>	return true;                      // update next time
795	>	}
796	>	if ((!shutdown || !tryTerminate(false)) &&
797	>	(int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 &&
798	>	blockedCount == 0 && quiescerCount == 0)
799	>	idleAwaitWork(w, nc, c, v);           // quiescent
800	>	for (boolean rescanned = false;;) {
801	>	if (w.eventCount != v)
802	>	return true;
803	>	if (!rescanned) {
804	>	int g = scanGuard, m = g & SMASK;
805	>	ForkJoinWorkerThread[] ws = workers;
806	>	if (ws != null && m < ws.length) {
807	>	rescanned = true;
808	>	for (int i = 0; i <= m; ++i) {
809	>	ForkJoinWorkerThread u = ws[i];
810	>	if (u != null) {
811	>	if (u.queueBase != u.queueTop &&
812	>	!tryReleaseWaiter())
813	>	rescanned = false; // contended
814	>	if (w.eventCount != v)
815	>	return true;
816	>	}
817	>	}
818	>	}
819	>	if (scanGuard != g ||              // stale
820	>	(queueBase != queueTop && !tryReleaseWaiter()))
821	>	rescanned = false;
822	>	if (!rescanned)
823	>	Thread.yield();                // reduce contention
824	>	else
825	>	Thread.interrupted();          // clear before park
826	>	}
827	>	else {
828	>	w.parked = true;                   // must recheck
829	>	if (w.eventCount != v) {
830	>	w.parked = false;
831	>	return true;
832	>	}
833	>	LockSupport.park(this);
834	>	rescanned = w.parked = false;
835	>	}
836	>	}
837		}
838
747	–	// Waiting for and signalling events
748	–
839		/**
840	<	* Releases workers blocked on a count not equal to current count.
841	<	* Normally called after precheck that eventWaiters isn't zero to
842	<	* avoid wasted array checks. Gives up upon a change in count or
843	<	* contention, letting other workers take over.
844	<	*/
845	<	private void releaseEventWaiters() {
846	<	ForkJoinWorkerThread[] ws = workers;
847	<	int n = ws.length;
848	<	long h = eventWaiters;
849	<	int ec = eventCount;
850	<	ForkJoinWorkerThread w; int id;
851	<	while ((int)(h >>> EVENT_COUNT_SHIFT) != ec &&
852	<	(id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
853	<	id < n && (w = ws[id]) != null &&
854	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
855	<	h,  h = w.nextWaiter)) {
856	<	LockSupport.unpark(w);
857	<	if (eventWaiters != h || eventCount != ec)
858	<	break;
840	>	* If inactivating worker w has caused pool to become
841	>	* quiescent, check for pool termination, and wait for event
842	>	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
843	>	* this case because quiescence reflects consensus about lack
844	>	* of work). On timeout, if ctl has not changed, terminate the
845	>	* worker. Upon its termination (see deregisterWorker), it may
846	>	* wake up another worker to possibly repeat this process.
847	>	*
848	>	* @param w the calling worker
849	>	* @param currentCtl the ctl value after enqueuing w
850	>	* @param prevCtl the ctl value if w terminated
851	>	* @param v the eventCount w awaits change
852	>	*/
853	>	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
854	>	long prevCtl, int v) {
855	>	if (w.eventCount == v) {
856	>	if (shutdown)
857	>	tryTerminate(false);
858	>	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
859	>	while (ctl == currentCtl) {
860	>	long startTime = System.nanoTime();
861	>	w.parked = true;
862	>	if (w.eventCount == v)             // must recheck
863	>	LockSupport.parkNanos(this, SHRINK_RATE);
864	>	w.parked = false;
865	>	if (w.eventCount != v)
866	>	break;
867	>	else if (System.nanoTime() - startTime <
868	>	SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
869	>	Thread.interrupted();          // spurious wakeup
870	>	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
871	>	currentCtl, prevCtl)) {
872	>	w.terminate = true;            // restore previous
873	>	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
874	>	break;
875	>	}
876	>	}
877		}
878		}
879
880	<	/**
773	<	* Tries to advance eventCount and releases waiters. Called only
774	<	* from workers.
775	<	*/
776	<	final void signalWork() {
777	<	int c; // try to increment event count -- CAS failure OK
778	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
779	<	if (eventWaiters != 0L)
780	<	releaseEventWaiters();
781	<	}
880	>	// Submissions
881
882		/**
883	<	* Adds the given worker to event queue and blocks until
884	<	* terminating or event count advances from the workers
885	<	* lastEventCount value
886	<	*
887	<	* @param w the calling worker thread
888	<	*/
889	<	private void eventSync(ForkJoinWorkerThread w) {
890	<	int ec = w.lastEventCount;
891	<	long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
892	<	long h;
893	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
894	<	(((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
895	<	(int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
896	<	eventCount == ec) {
897	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
898	<	w.nextWaiter = h, nh)) {
800	<	awaitEvent(w, ec);
801	<	break;
883	>	* Enqueues the given task in the submissionQueue.  Same idea as
884	>	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
885	>	*
886	>	* @param t the task
887	>	*/
888	>	private void addSubmission(ForkJoinTask<?> t) {
889	>	final ReentrantLock lock = this.submissionLock;
890	>	lock.lock();
891	>	try {
892	>	ForkJoinTask<?>[] q; int s, m;
893	>	if ((q = submissionQueue) != null) {    // ignore if queue removed
894	>	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
895	>	UNSAFE.putOrderedObject(q, u, t);
896	>	queueTop = s + 1;
897	>	if (s - queueBase == m)
898	>	growSubmissionQueue();
899		}
900	+	} finally {
901	+	lock.unlock();
902		}
903	+	signalWork();
904		}
905
906	+	//  (pollSubmission is defined below with exported methods)
907	+
908		/**
909	<	* Blocks the given worker (that has already been entered as an
910	<	* event waiter) until terminating or event count advances from
911	<	* the given value. The oldest (first) waiter uses a timed wait to
912	<	* occasionally one-by-one shrink the number of workers (to a
913	<	* minumum of one) if the pool has not been used for extended
914	<	* periods.
915	<	*
916	<	* @param w the calling worker thread
917	<	* @param ec the count
918	<	*/
919	<	private void awaitEvent(ForkJoinWorkerThread w, int ec) {
920	<	while (eventCount == ec) {
921	<	if (tryAccumulateStealCount(w)) { // transfer while idle
922	<	boolean untimed = (w.nextWaiter != 0L ||
923	<	(workerCounts & RUNNING_COUNT_MASK) <= 1);
924	<	long startTime = untimed? 0 : System.nanoTime();
925	<	Thread.interrupted();         // clear/ignore interrupt
926	<	if (eventCount != ec || !w.isRunning() ||
927	<	runState >= TERMINATING)  // recheck after clear
928	<	break;
929	<	if (untimed)
828	<	LockSupport.park(w);
829	<	else {
830	<	LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
831	<	if (eventCount != ec || !w.isRunning() ||
832	<	runState >= TERMINATING)
833	<	break;
834	<	if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
835	<	tryShutdownWaiter(ec);
836	<	}
909	>	* Creates or doubles submissionQueue array.
910	>	* Basically identical to ForkJoinWorkerThread version.
911	>	*/
912	>	private void growSubmissionQueue() {
913	>	ForkJoinTask<?>[] oldQ = submissionQueue;
914	>	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
915	>	if (size > MAXIMUM_QUEUE_CAPACITY)
916	>	throw new RejectedExecutionException("Queue capacity exceeded");
917	>	if (size < INITIAL_QUEUE_CAPACITY)
918	>	size = INITIAL_QUEUE_CAPACITY;
919	>	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
920	>	int mask = size - 1;
921	>	int top = queueTop;
922	>	int oldMask;
923	>	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
924	>	for (int b = queueBase; b != top; ++b) {
925	>	long u = ((b & oldMask) << ASHIFT) + ABASE;
926	>	Object x = UNSAFE.getObjectVolatile(oldQ, u);
927	>	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
928	>	UNSAFE.putObjectVolatile
929	>	(q, ((b & mask) << ASHIFT) + ABASE, x);
930		}
931		}
932		}
933
934	+	// Blocking support
935	+
936		/**
937	<	* Callback from the oldest waiter in awaitEvent waking up after a
938	<	* period of non-use. Tries (once) to shutdown an event waiter (or
939	<	* a spare, if one exists). Note that we don't need CAS or locks
940	<	* here because the method is called only from one thread
941	<	* occasionally waking (and even misfires are OK). Note that
942	<	* until the shutdown worker fully terminates, workerCounts
943	<	* will overestimate total count, which is tolerable.
944	<	*
945	<	* @param ec the event count waited on by caller (to abort
946	<	* attempt if count has since changed).
947	<	*/
948	<	private void tryShutdownWaiter(int ec) {
949	<	if (spareWaiters != 0) { // prefer killing spares
950	<	tryShutdownSpare();
951	<	return;
952	<	}
953	<	ForkJoinWorkerThread[] ws = workers;
954	<	int n = ws.length;
955	<	long h = eventWaiters;
956	<	ForkJoinWorkerThread w; int id; long nh;
957	<	if (runState == 0 &&
958	<	submissionQueue.isEmpty() &&
959	<	eventCount == ec &&
960	<	(id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
961	<	id < n && (w = ws[id]) != null &&
962	<	(nh = w.nextWaiter) != 0L && // keep at least one worker
963	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh)) {
964	<	w.shutdown();
965	<	LockSupport.unpark(w);
937	>	* Tries to increment blockedCount, decrement active count
938	>	* (sometimes implicitly) and possibly release or create a
939	>	* compensating worker in preparation for blocking. Fails
940	>	* on contention or termination.
941	>	*
942	>	* @return true if the caller can block, else should recheck and retry
943	>	*/
944	>	private boolean tryPreBlock() {
945	>	int b = blockedCount;
946	>	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
947	>	int pc = parallelism;
948	>	do {
949	>	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
950	>	int e, ac, tc, i;
951	>	long c = ctl;
952	>	int u = (int)(c >>> 32);
953	>	if ((e = (int)c) < 0) {
954	>	// skip -- terminating
955	>	}
956	>	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
957	>	(ws = workers) != null &&
958	>	(i = ~e & SMASK) < ws.length &&
959	>	(w = ws[i]) != null) {
960	>	long nc = ((long)(w.nextWait & E_MASK) |
961	>	(c & (AC_MASK|TC_MASK)));
962	>	if (w.eventCount == e &&
963	>	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
964	>	w.eventCount = (e + EC_UNIT) & E_MASK;
965	>	if (w.parked)
966	>	UNSAFE.unpark(w);
967	>	return true;             // release an idle worker
968	>	}
969	>	}
970	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
971	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
972	>	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
973	>	return true;             // no compensation needed
974	>	}
975	>	else if (tc + pc < MAX_ID) {
976	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
977	>	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
978	>	addWorker();
979	>	return true;            // create a replacement
980	>	}
981	>	}
982	>	// try to back out on any failure and let caller retry
983	>	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
984	>	b = blockedCount, b - 1));
985		}
986	<	releaseEventWaiters();
986	>	return false;
987		}
988
875	–	// Maintaining spares
876	–
989		/**
990	<	* Pushes worker onto the spare stack
990	>	* Decrements blockedCount and increments active count.
991		*/
992	<	final void pushSpare(ForkJoinWorkerThread w) {
993	<	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
994	<	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
995	<	w.nextSpare = spareWaiters,ns));
992	>	private void postBlock() {
993	>	long c;
994	>	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
995	>	c = ctl, c + AC_UNIT));
996	>	int b;
997	>	do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
998	>	b = blockedCount, b - 1));
999		}
1000
1001		/**
1002	<	* Callback from oldest spare occasionally waking up.  Tries
1003	<	* (once) to shutdown a spare. Same idea as tryShutdownWaiter.
1002	>	* Possibly blocks waiting for the given task to complete, or
1003	>	* cancels the task if terminating.  Fails to wait if contended.
1004	>	*
1005	>	* @param joinMe the task
1006		*/
1007	<	final void tryShutdownSpare() {
1008	<	int sw, id;
1009	<	ForkJoinWorkerThread w;
1010	<	ForkJoinWorkerThread[] ws;
1011	<	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
1012	<	id < (ws = workers).length && (w = ws[id]) != null &&
1013	<	(workerCounts & RUNNING_COUNT_MASK) >= parallelism &&
1014	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1015	<	sw, w.nextSpare)) {
899	<	w.shutdown();
900	<	LockSupport.unpark(w);
901	<	advanceEventCount();
1007	>	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1008	>	Thread.interrupted(); // clear interrupts before checking termination
1009	>	if (joinMe.status >= 0) {
1010	>	if (tryPreBlock()) {
1011	>	joinMe.tryAwaitDone(0L);
1012	>	postBlock();
1013	>	}
1014	>	else if ((ctl & STOP_BIT) != 0L)
1015	>	joinMe.cancelIgnoringExceptions();
1016		}
1017		}
1018
1019		/**
1020	<	* Tries (once) to resume a spare if worker counts match
1021	<	* the given count.
1020	>	* Possibly blocks the given worker waiting for joinMe to
1021	>	* complete or timeout.
1022		*
1023	<	* @param wc workerCounts value on invocation of this method
1023	>	* @param joinMe the task
1024	>	* @param millis the wait time for underlying Object.wait
1025		*/
1026	<	private void tryResumeSpare(int wc) {
1027	<	ForkJoinWorkerThread[] ws = workers;
1028	<	int n = ws.length;
1029	<	int sw, id, rs;  ForkJoinWorkerThread w;
1030	<	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
1031	<	id < n && (w = ws[id]) != null &&
1032	<	(rs = runState) < TERMINATING &&
1033	<	eventWaiters == 0L && workerCounts == wc) {
1034	<	// In case all workers busy, heuristically back off to let settle
1035	<	Thread.yield();
1036	<	if (eventWaiters == 0L && runState == rs && // recheck
1037	<	workerCounts == wc && spareWaiters == sw &&
1038	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1039	<	sw, w.nextSpare)) {
1040	<	int c;              // increment running count before resume
1041	<	do {} while(!UNSAFE.compareAndSwapInt
1042	<	(this, workerCountsOffset,
1043	<	c = workerCounts, c + ONE_RUNNING));
1044	<	if (w.tryUnsuspend())
1045	<	LockSupport.unpark(w);
1046	<	else               // back out if w was shutdown
1047	<	decrementWorkerCounts(ONE_RUNNING, 0);
1026	>	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1027	>	while (joinMe.status >= 0) {
1028	>	Thread.interrupted();
1029	>	if ((ctl & STOP_BIT) != 0L) {
1030	>	joinMe.cancelIgnoringExceptions();
1031	>	break;
1032	>	}
1033	>	if (tryPreBlock()) {
1034	>	long last = System.nanoTime();
1035	>	while (joinMe.status >= 0) {
1036	>	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1037	>	if (millis <= 0)
1038	>	break;
1039	>	joinMe.tryAwaitDone(millis);
1040	>	if (joinMe.status < 0)
1041	>	break;
1042	>	if ((ctl & STOP_BIT) != 0L) {
1043	>	joinMe.cancelIgnoringExceptions();
1044	>	break;
1045	>	}
1046	>	long now = System.nanoTime();
1047	>	nanos -= now - last;
1048	>	last = now;
1049	>	}
1050	>	postBlock();
1051	>	break;
1052		}
1053		}
1054		}
1055
937	–	// adding workers on demand
938	–
1056		/**
1057	<	* Adds one or more workers if needed to establish target parallelism.
941	<	* Retries upon contention.
1057	>	* If necessary, compensates for blocker, and blocks.
1058		*/
1059	<	private void addWorkerIfBelowTarget() {
1060	<	int pc = parallelism;
1061	<	int wc;
1062	<	while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < pc &&
1063	<	runState < TERMINATING) {
1064	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1065	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
1066	<	if (addWorker() == null)
1067	<	break;
1059	>	private void awaitBlocker(ManagedBlocker blocker)
1060	>	throws InterruptedException {
1061	>	while (!blocker.isReleasable()) {
1062	>	if (tryPreBlock()) {
1063	>	try {
1064	>	do {} while (!blocker.isReleasable() && !blocker.block());
1065	>	} finally {
1066	>	postBlock();
1067	>	}
1068	>	break;
1069		}
1070		}
1071		}
1072
1073	+	// Creating, registering and deregistring workers
1074	+
1075		/**
1076	<	* Tries (once) to add a new worker if all existing workers are
1077	<	* busy, and there are either no running workers or the deficit is
959	<	* at least twice the surplus.
960	<	*
961	<	* @param wc workerCounts value on invocation of this method
1076	>	* Tries to create and start a worker; minimally rolls back counts
1077	>	* on failure.
1078		*/
1079	<	private void tryAddWorkerIfBusy(int wc) {
1080	<	int tc, rc, rs;
1081	<	int pc = parallelism;
1082	<	if ((tc = wc >>> TOTAL_COUNT_SHIFT) < MAX_WORKERS &&
1083	<	((rc = wc & RUNNING_COUNT_MASK) == 0 ||
1084	<	rc < pc - ((tc - pc) << 1)) &&
1085	<	(rs = runState) < TERMINATING &&
1086	<	(rs & ACTIVE_COUNT_MASK) == tc) {
1087	<	// Since all workers busy, heuristically back off to let settle
1088	<	Thread.yield();
1089	<	if (eventWaiters == 0L && spareWaiters == 0 && // recheck
1090	<	runState == rs && workerCounts == wc &&
1091	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1092	<	wc + (ONE_RUNNING|ONE_TOTAL)))
1093	<	addWorker();
1079	>	private void addWorker() {
1080	>	Throwable ex = null;
1081	>	ForkJoinWorkerThread t = null;
1082	>	try {
1083	>	t = factory.newThread(this);
1084	>	} catch (Throwable e) {
1085	>	ex = e;
1086	>	}
1087	>	if (t == null) {  // null or exceptional factory return
1088	>	long c;       // adjust counts
1089	>	do {} while (!UNSAFE.compareAndSwapLong
1090	>	(this, ctlOffset, c = ctl,
1091	>	(((c - AC_UNIT) & AC_MASK) |
1092	>	((c - TC_UNIT) & TC_MASK) |
1093	>	(c & ~(AC_MASK|TC_MASK)))));
1094	>	// Propagate exception if originating from an external caller
1095	>	if (!tryTerminate(false) && ex != null &&
1096	>	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1097	>	UNSAFE.throwException(ex);
1098		}
1099	+	else
1100	+	t.start();
1101		}
1102
1103		/**
1104	<	* Does at most one of:
1105	<	*
1106	<	* 1. Help wake up existing workers waiting for work via
1107	<	*    releaseEventWaiters. (If any exist, then it doesn't
1108	<	*    matter right now if under target parallelism level.)
1109	<	*
1110	<	* 2. If a spare exists, try (once) to resume it via tryResumeSpare.
1111	<	*
990	<	* 3. If there are not enough total workers, add some
991	<	*    via addWorkerIfBelowTarget;
992	<	*
993	<	* 4. Try (once) to add a new worker if all existing workers
994	<	*     are busy, via tryAddWorkerIfBusy
995	<	*/
996	<	private void helpMaintainParallelism() {
997	<	long h; int pc, wc;
998	<	if (((int)((h = eventWaiters) & WAITER_ID_MASK)) != 0) {
999	<	if ((int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1000	<	releaseEventWaiters(); // avoid useless call
1001	<	}
1002	<	else if ((pc = parallelism) >
1003	<	((wc = workerCounts) & RUNNING_COUNT_MASK)) {
1004	<	if (spareWaiters != 0)
1005	<	tryResumeSpare(wc);
1006	<	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
1007	<	addWorkerIfBelowTarget();
1008	<	else
1009	<	tryAddWorkerIfBusy(wc);
1104	>	* Callback from ForkJoinWorkerThread constructor to assign a
1105	>	* public name
1106	>	*/
1107	>	final String nextWorkerName() {
1108	>	for (int n;;) {
1109	>	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
1110	>	n = nextWorkerNumber, ++n))
1111	>	return workerNamePrefix + n;
1112		}
1113		}
1114
1115		/**
1116	<	* Callback from workers invoked upon each top-level action (i.e.,
1117	<	* stealing a task or taking a submission and running it).
1016	<	* Performs one or more of the following:
1017	<	*
1018	<	* 1. If the worker is active, try to set its active status to
1019	<	*    inactive and update activeCount. On contention, we may try
1020	<	*    again on this or subsequent call.
1021	<	*
1022	<	* 2. Release any existing event waiters that are now relesable
1023	<	*
1024	<	* 3. If there are too many running threads, suspend this worker
1025	<	*    (first forcing inactive if necessary).  If it is not
1026	<	*    needed, it may be killed while suspended via
1027	<	*    tryShutdownSpare. Otherwise, upon resume it rechecks to make
1028	<	*    sure that it is still needed.
1029	<	*
1030	<	* 4. If more than 1 miss, await the next task event via
1031	<	*    eventSync (first forcing inactivation if necessary), upon
1032	<	*    which worker may also be killed, via tryShutdownWaiter.
1033	<	*
1034	<	* 5. Help reactivate other workers via helpMaintainParallelism
1116	>	* Callback from ForkJoinWorkerThread constructor to
1117	>	* determine its poolIndex and record in workers array.
1118		*
1119		* @param w the worker
1120	<	* @param misses the number of scans by caller failing to find work
1038	<	* (saturating at 2 to avoid wraparound)
1120	>	* @return the worker's pool index
1121		*/
1122	<	final void preStep(ForkJoinWorkerThread w, int misses) {
1123	<	boolean active = w.active;
1124	<	int pc = parallelism;
1125	<	for (;;) {
1126	<	int rs, wc, rc, ec; long h;
1127	<	if (active && UNSAFE.compareAndSwapInt(this, runStateOffset,
1128	<	rs = runState, rs - 1))
1129	<	active = w.active = false;
1130	<	if (((int)((h = eventWaiters) & WAITER_ID_MASK)) != 0 &&
1131	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount) {
1132	<	releaseEventWaiters();
1133	<	if (misses > 1)
1134	<	continue;                  // clear before sync below
1135	<	}
1136	<	if ((rc = ((wc = workerCounts) & RUNNING_COUNT_MASK)) > pc) {
1137	<	if (!active &&                 // must inactivate to suspend
1138	<	workerCounts == wc &&      // try to suspend as spare
1139	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1140	<	wc, wc - ONE_RUNNING)) {
1141	<	w.suspendAsSpare();
1142	<	if (!w.isRunning())
1143	<	break;                 // was killed while spare
1122	>	final int registerWorker(ForkJoinWorkerThread w) {
1123	>	/*
1124	>	* In the typical case, a new worker acquires the lock, uses
1125	>	* next available index and returns quickly.  Since we should
1126	>	* not block callers (ultimately from signalWork or
1127	>	* tryPreBlock) waiting for the lock needed to do this, we
1128	>	* instead help release other workers while waiting for the
1129	>	* lock.
1130	>	*/
1131	>	for (int g;;) {
1132	>	ForkJoinWorkerThread[] ws;
1133	>	if (((g = scanGuard) & SG_UNIT) == 0 &&
1134	>	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1135	>	g, g | SG_UNIT)) {
1136	>	int k = nextWorkerIndex;
1137	>	try {
1138	>	if ((ws = workers) != null) { // ignore on shutdown
1139	>	int n = ws.length;
1140	>	if (k < 0 || k >= n || ws[k] != null) {
1141	>	for (k = 0; k < n && ws[k] != null; ++k)
1142	>	;
1143	>	if (k == n)
1144	>	ws = workers = Arrays.copyOf(ws, n << 1);
1145	>	}
1146	>	ws[k] = w;
1147	>	nextWorkerIndex = k + 1;
1148	>	int m = g & SMASK;
1149	>	g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
1150	>	}
1151	>	} finally {
1152	>	scanGuard = g;
1153		}
1154	<	continue;
1154	>	return k;
1155		}
1156	<	if (misses > 0) {
1157	<	if ((ec = eventCount) == w.lastEventCount && misses > 1) {
1158	<	if (!active) {             // must inactivate to sync
1159	<	eventSync(w);
1160	<	if (w.isRunning())
1070	<	misses = 1;        // don't re-sync
1071	<	else
1072	<	break;             // was killed while waiting
1156	>	else if ((ws = workers) != null) { // help release others
1157	>	for (ForkJoinWorkerThread u : ws) {
1158	>	if (u != null && u.queueBase != u.queueTop) {
1159	>	if (tryReleaseWaiter())
1160	>	break;
1161		}
1074	–	continue;
1162		}
1076	–	w.lastEventCount = ec;
1163		}
1078	–	if (rc < pc)
1079	–	helpMaintainParallelism();
1080	–	break;
1164		}
1165		}
1166
1167		/**
1168	<	* Helps and/or blocks awaiting join of the given task.
1169	<	* Alternates between helpJoinTask() and helpMaintainParallelism()
1170	<	* as many times as there is a deficit in running count (or longer
1088	<	* if running count would become zero), then blocks if task still
1089	<	* not done.
1168	>	* Final callback from terminating worker.  Removes record of
1169	>	* worker from array, and adjusts counts. If pool is shutting
1170	>	* down, tries to complete termination.
1171		*
1172	<	* @param joinMe the task to join
1172	>	* @param w the worker
1173		*/
1174	<	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1175	<	int threshold = parallelism;         // descend blocking thresholds
1176	<	while (joinMe.status >= 0) {
1177	<	boolean block; int wc;
1178	<	worker.helpJoinTask(joinMe);
1179	<	if (joinMe.status < 0)
1180	<	break;
1181	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1182	<	if (threshold > 0)
1183	<	--threshold;
1184	<	else
1185	<	advanceEventCount(); // force release
1186	<	block = false;
1174	>	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
1175	>	int idx = w.poolIndex;
1176	>	int sc = w.stealCount;
1177	>	int steps = 0;
1178	>	// Remove from array, adjust worker counts and collect steal count.
1179	>	// We can intermix failed removes or adjusts with steal updates
1180	>	do {
1181	>	long s, c;
1182	>	int g;
1183	>	if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1184	>	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1185	>	g, g |= SG_UNIT)) {
1186	>	ForkJoinWorkerThread[] ws = workers;
1187	>	if (ws != null && idx >= 0 &&
1188	>	idx < ws.length && ws[idx] == w)
1189	>	ws[idx] = null;    // verify
1190	>	nextWorkerIndex = idx;
1191	>	scanGuard = g + SG_UNIT;
1192	>	steps = 1;
1193		}
1194	+	if (steps == 1 &&
1195	+	UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1196	+	(((c - AC_UNIT) & AC_MASK) |
1197	+	((c - TC_UNIT) & TC_MASK) |
1198	+	(c & ~(AC_MASK|TC_MASK)))))
1199	+	steps = 2;
1200	+	if (sc != 0 &&
1201	+	UNSAFE.compareAndSwapLong(this, stealCountOffset,
1202	+	s = stealCount, s + sc))
1203	+	sc = 0;
1204	+	} while (steps != 2 || sc != 0);
1205	+	if (!tryTerminate(false)) {
1206	+	if (ex != null)   // possibly replace if died abnormally
1207	+	signalWork();
1208		else
1209	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1109	<	wc, wc - ONE_RUNNING);
1110	<	helpMaintainParallelism();
1111	<	if (block) {
1112	<	int c;
1113	<	joinMe.internalAwaitDone();
1114	<	do {} while (!UNSAFE.compareAndSwapInt
1115	<	(this, workerCountsOffset,
1116	<	c = workerCounts, c + ONE_RUNNING));
1117	<	break;
1118	<	}
1209	>	tryReleaseWaiter();
1210		}
1211		}
1212
1213	<	/**
1123	<	* Same idea as awaitJoin, but no helping
1124	<	*/
1125	<	final void awaitBlocker(ManagedBlocker blocker)
1126	<	throws InterruptedException {
1127	<	int threshold = parallelism;
1128	<	while (!blocker.isReleasable()) {
1129	<	boolean block; int wc;
1130	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1131	<	if (threshold > 0)
1132	<	--threshold;
1133	<	else
1134	<	advanceEventCount();
1135	<	block = false;
1136	<	}
1137	<	else
1138	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1139	<	wc, wc - ONE_RUNNING);
1140	<	helpMaintainParallelism();
1141	<	if (block) {
1142	<	try {
1143	<	do {} while (!blocker.isReleasable() && !blocker.block());
1144	<	} finally {
1145	<	int c;
1146	<	do {} while (!UNSAFE.compareAndSwapInt
1147	<	(this, workerCountsOffset,
1148	<	c = workerCounts, c + ONE_RUNNING));
1149	<	}
1150	<	break;
1151	<	}
1152	<	}
1153	<	}
1213	>	// Shutdown and termination
1214
1215		/**
1216		* Possibly initiates and/or completes termination.
#	Line 1160 | Line 1220 | public class ForkJoinPool extends Abstra
1220		* @return true if now terminating or terminated
1221		*/
1222		private boolean tryTerminate(boolean now) {
1223	<	if (now)
1224	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1225	<	else if (runState < SHUTDOWN ||
1226	<	!submissionQueue.isEmpty() ||
1227	<	(runState & ACTIVE_COUNT_MASK) != 0)
1228	<	return false;
1229	<
1230	<	if (advanceRunLevel(TERMINATING))
1231	<	startTerminating();
1232	<
1233	<	// Finish now if all threads terminated; else in some subsequent call
1234	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1235	<	advanceRunLevel(TERMINATED);
1236	<	termination.arrive();
1223	>	long c;
1224	>	while (((c = ctl) & STOP_BIT) == 0) {
1225	>	if (!now) {
1226	>	if ((int)(c >> AC_SHIFT) != -parallelism)
1227	>	return false;
1228	>	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
1229	>	queueBase != queueTop) {
1230	>	if (ctl == c) // staleness check
1231	>	return false;
1232	>	continue;
1233	>	}
1234	>	}
1235	>	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
1236	>	startTerminating();
1237	>	}
1238	>	if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
1239	>	final ReentrantLock lock = this.submissionLock;
1240	>	lock.lock();
1241	>	try {
1242	>	termination.signalAll();
1243	>	} finally {
1244	>	lock.unlock();
1245	>	}
1246		}
1247		return true;
1248		}
1249
1250		/**
1251	<	* Actions on transition to TERMINATING
1252	<	*
1253	<	* Runs up to four passes through workers: (0) shutting down each
1254	<	* (without waking up if parked) to quickly spread notifications
1255	<	* without unnecessary bouncing around event queues etc (1) wake
1256	<	* up and help cancel tasks (2) interrupt (3) mop up races with
1188	<	* interrupted workers
1251	>	* Runs up to three passes through workers: (0) Setting
1252	>	* termination status for each worker, followed by wakeups up to
1253	>	* queued workers; (1) helping cancel tasks; (2) interrupting
1254	>	* lagging threads (likely in external tasks, but possibly also
1255	>	* blocked in joins).  Each pass repeats previous steps because of
1256	>	* potential lagging thread creation.
1257		*/
1258		private void startTerminating() {
1259		cancelSubmissions();
1260	<	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1193	<	advanceEventCount();
1194	<	eventWaiters = 0L; // clobber lists
1195	<	spareWaiters = 0;
1260	>	for (int pass = 0; pass < 3; ++pass) {
1261		ForkJoinWorkerThread[] ws = workers;
1262	<	int n = ws.length;
1263	<	for (int i = 0; i < n; ++i) {
1264	<	ForkJoinWorkerThread w = ws[i];
1265	<	if (w != null) {
1266	<	w.shutdown();
1267	<	if (passes > 0 && !w.isTerminated()) {
1268	<	w.cancelTasks();
1269	<	LockSupport.unpark(w);
1270	<	if (passes > 1) {
1271	<	try {
1272	<	w.interrupt();
1208	<	} catch (SecurityException ignore) {
1262	>	if (ws != null) {
1263	>	for (ForkJoinWorkerThread w : ws) {
1264	>	if (w != null) {
1265	>	w.terminate = true;
1266	>	if (pass > 0) {
1267	>	w.cancelTasks();
1268	>	if (pass > 1 && !w.isInterrupted()) {
1269	>	try {
1270	>	w.interrupt();
1271	>	} catch (SecurityException ignore) {
1272	>	}
1273		}
1274		}
1275		}
1276		}
1277	+	terminateWaiters();
1278		}
1279		}
1280		}
1281
1282		/**
1283	<	* Clear out and cancel submissions, ignoring exceptions
1283	>	* Polls and cancels all submissions. Called only during termination.
1284		*/
1285		private void cancelSubmissions() {
1286	<	ForkJoinTask<?> task;
1287	<	while ((task = submissionQueue.poll()) != null) {
1288	<	try {
1289	<	task.cancel(false);
1290	<	} catch (Throwable ignore) {
1286	>	while (queueBase != queueTop) {
1287	>	ForkJoinTask<?> task = pollSubmission();
1288	>	if (task != null) {
1289	>	try {
1290	>	task.cancel(false);
1291	>	} catch (Throwable ignore) {
1292	>	}
1293		}
1294		}
1295		}
1296
1297	<	// misc support for ForkJoinWorkerThread
1297	>	/**
1298	>	* Tries to set the termination status of waiting workers, and
1299	>	* then wakes them up (after which they will terminate).
1300	>	*/
1301	>	private void terminateWaiters() {
1302	>	ForkJoinWorkerThread[] ws = workers;
1303	>	if (ws != null) {
1304	>	ForkJoinWorkerThread w; long c; int i, e;
1305	>	int n = ws.length;
1306	>	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
1307	>	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
1308	>	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
1309	>	(long)(w.nextWait & E_MASK) |
1310	>	((c + AC_UNIT) & AC_MASK) |
1311	>	(c & (TC_MASK|STOP_BIT)))) {
1312	>	w.terminate = true;
1313	>	w.eventCount = e + EC_UNIT;
1314	>	if (w.parked)
1315	>	UNSAFE.unpark(w);
1316	>	}
1317	>	}
1318	>	}
1319	>	}
1320	>
1321	>	// misc ForkJoinWorkerThread support
1322
1323		/**
1324	<	* Returns pool number
1324	>	* Increments or decrements quiescerCount. Needed only to prevent
1325	>	* triggering shutdown if a worker is transiently inactive while
1326	>	* checking quiescence.
1327	>	*
1328	>	* @param delta 1 for increment, -1 for decrement
1329		*/
1330	<	final int getPoolNumber() {
1331	<	return poolNumber;
1330	>	final void addQuiescerCount(int delta) {
1331	>	int c;
1332	>	do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
1333	>	c = quiescerCount, c + delta));
1334		}
1335
1336		/**
1337	<	* Tries to accumulates steal count from a worker, clearing
1338	<	* the worker's value.
1337	>	* Directly increments or decrements active count without queuing.
1338	>	* This method is used to transiently assert inactivation while
1339	>	* checking quiescence.
1340		*
1341	<	* @return true if worker steal count now zero
1341	>	* @param delta 1 for increment, -1 for decrement
1342		*/
1343	<	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1344	<	int sc = w.stealCount;
1345	<	long c = stealCount;
1346	<	// CAS even if zero, for fence effects
1347	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1250	<	if (sc != 0)
1251	<	w.stealCount = 0;
1252	<	return true;
1253	<	}
1254	<	return sc == 0;
1343	>	final void addActiveCount(int delta) {
1344	>	long d = (long)delta << AC_SHIFT;
1345	>	long c;
1346	>	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,
1347	>	c = ctl, c + d));
1348		}
1349
1350		/**
#	Line 1259 | Line 1352 | public class ForkJoinPool extends Abstra
1352		* active thread.
1353		*/
1354		final int idlePerActive() {
1355	<	int pc = parallelism; // use parallelism, not rc
1356	<	int ac = runState;    // no mask -- artifically boosts during shutdown
1357	<	// Use exact results for small values, saturate past 4
1358	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1355	>	// Approximate at powers of two for small values, saturate past 4
1356	>	int p = parallelism;
1357	>	int a = p + (int)(ctl >> AC_SHIFT);
1358	>	return (a > (p >>>= 1) ? 0 :
1359	>	a > (p >>>= 1) ? 1 :
1360	>	a > (p >>>= 1) ? 2 :
1361	>	a > (p >>>= 1) ? 4 :
1362	>	8);
1363		}
1364
1365	<	// Public and protected methods
1365	>	// Exported methods
1366
1367		// Constructors
1368
#	Line 1312 | Line 1409 | public class ForkJoinPool extends Abstra
1409		* use {@link #defaultForkJoinWorkerThreadFactory}.
1410		* @param handler the handler for internal worker threads that
1411		* terminate due to unrecoverable errors encountered while executing
1412	<	* tasks. For default value, use <code>null</code>.
1412	>	* tasks. For default value, use {@code null}.
1413		* @param asyncMode if true,
1414		* establishes local first-in-first-out scheduling mode for forked
1415		* tasks that are never joined. This mode may be more appropriate
1416		* than default locally stack-based mode in applications in which
1417		* worker threads only process event-style asynchronous tasks.
1418	<	* For default value, use <code>false</code>.
1418	>	* For default value, use {@code false}.
1419		* @throws IllegalArgumentException if parallelism less than or
1420		*         equal to zero, or greater than implementation limit
1421		* @throws NullPointerException if the factory is null
#	Line 1334 | Line 1431 | public class ForkJoinPool extends Abstra
1431		checkPermission();
1432		if (factory == null)
1433		throw new NullPointerException();
1434	<	if (parallelism <= 0 || parallelism > MAX_WORKERS)
1434	>	if (parallelism <= 0 || parallelism > MAX_ID)
1435		throw new IllegalArgumentException();
1436		this.parallelism = parallelism;
1437		this.factory = factory;
1438		this.ueh = handler;
1439		this.locallyFifo = asyncMode;
1440	<	int arraySize = initialArraySizeFor(parallelism);
1441	<	this.workers = new ForkJoinWorkerThread[arraySize];
1442	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1443	<	this.workerLock = new ReentrantLock();
1444	<	this.termination = new Phaser(1);
1445	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
1446	<	}
1447	<
1448	<	/**
1449	<	* Returns initial power of two size for workers array.
1450	<	* @param pc the initial parallelism level
1451	<	*/
1452	<	private static int initialArraySizeFor(int pc) {
1453	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1454	<	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1455	<	size |= size >>> 1;
1456	<	size |= size >>> 2;
1360	<	size |= size >>> 4;
1361	<	size |= size >>> 8;
1362	<	return size + 1;
1440	>	long np = (long)(-parallelism); // offset ctl counts
1441	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1442	>	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1443	>	// initialize workers array with room for 2*parallelism if possible
1444	>	int n = parallelism << 1;
1445	>	if (n >= MAX_ID)
1446	>	n = MAX_ID;
1447	>	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1448	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1449	>	}
1450	>	workers = new ForkJoinWorkerThread[n + 1];
1451	>	this.submissionLock = new ReentrantLock();
1452	>	this.termination = submissionLock.newCondition();
1453	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
1454	>	sb.append(poolNumberGenerator.incrementAndGet());
1455	>	sb.append("-worker-");
1456	>	this.workerNamePrefix = sb.toString();
1457		}
1458
1459		// Execution methods
1460
1461		/**
1368	–	* Common code for execute, invoke and submit
1369	–	*/
1370	–	private <T> void doSubmit(ForkJoinTask<T> task) {
1371	–	if (task == null)
1372	–	throw new NullPointerException();
1373	–	if (runState >= SHUTDOWN)
1374	–	throw new RejectedExecutionException();
1375	–	submissionQueue.offer(task);
1376	–	advanceEventCount();
1377	–	if (eventWaiters != 0L)
1378	–	releaseEventWaiters();
1379	–	if ((workerCounts >>> TOTAL_COUNT_SHIFT) < parallelism)
1380	–	addWorkerIfBelowTarget();
1381	–	}
1382	–
1383	–	/**
1462		* Performs the given task, returning its result upon completion.
1463	+	* If the computation encounters an unchecked Exception or Error,
1464	+	* it is rethrown as the outcome of this invocation.  Rethrown
1465	+	* exceptions behave in the same way as regular exceptions, but,
1466	+	* when possible, contain stack traces (as displayed for example
1467	+	* using {@code ex.printStackTrace()}) of both the current thread
1468	+	* as well as the thread actually encountering the exception;
1469	+	* minimally only the latter.
1470		*
1471		* @param task the task
1472		* @return the task's result
#	Line 1390 | Line 1475 | public class ForkJoinPool extends Abstra
1475		*         scheduled for execution
1476		*/
1477		public <T> T invoke(ForkJoinTask<T> task) {
1478	<	doSubmit(task);
1479	<	return task.join();
1478	>	Thread t = Thread.currentThread();
1479	>	if (task == null)
1480	>	throw new NullPointerException();
1481	>	if (shutdown)
1482	>	throw new RejectedExecutionException();
1483	>	if ((t instanceof ForkJoinWorkerThread) &&
1484	>	((ForkJoinWorkerThread)t).pool == this)
1485	>	return task.invoke();  // bypass submit if in same pool
1486	>	else {
1487	>	addSubmission(task);
1488	>	return task.join();
1489	>	}
1490	>	}
1491	>
1492	>	/**
1493	>	* Unless terminating, forks task if within an ongoing FJ
1494	>	* computation in the current pool, else submits as external task.
1495	>	*/
1496	>	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1497	>	ForkJoinWorkerThread w;
1498	>	Thread t = Thread.currentThread();
1499	>	if (shutdown)
1500	>	throw new RejectedExecutionException();
1501	>	if ((t instanceof ForkJoinWorkerThread) &&
1502	>	(w = (ForkJoinWorkerThread)t).pool == this)
1503	>	w.pushTask(task);
1504	>	else
1505	>	addSubmission(task);
1506		}
1507
1508		/**
#	Line 1403 | Line 1514 | public class ForkJoinPool extends Abstra
1514		*         scheduled for execution
1515		*/
1516		public void execute(ForkJoinTask<?> task) {
1517	<	doSubmit(task);
1517	>	if (task == null)
1518	>	throw new NullPointerException();
1519	>	forkOrSubmit(task);
1520		}
1521
1522		// AbstractExecutorService methods
#	Line 1414 | Line 1527 | public class ForkJoinPool extends Abstra
1527		*         scheduled for execution
1528		*/
1529		public void execute(Runnable task) {
1530	+	if (task == null)
1531	+	throw new NullPointerException();
1532		ForkJoinTask<?> job;
1533		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1534		job = (ForkJoinTask<?>) task;
1535		else
1536		job = ForkJoinTask.adapt(task, null);
1537	<	doSubmit(job);
1537	>	forkOrSubmit(job);
1538		}
1539
1540		/**
#	Line 1432 | Line 1547 | public class ForkJoinPool extends Abstra
1547		*         scheduled for execution
1548		*/
1549		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1550	<	doSubmit(task);
1550	>	if (task == null)
1551	>	throw new NullPointerException();
1552	>	forkOrSubmit(task);
1553		return task;
1554		}
1555
#	Line 1442 | Line 1559 | public class ForkJoinPool extends Abstra
1559		*         scheduled for execution
1560		*/
1561		public <T> ForkJoinTask<T> submit(Callable<T> task) {
1562	+	if (task == null)
1563	+	throw new NullPointerException();
1564		ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1565	<	doSubmit(job);
1565	>	forkOrSubmit(job);
1566		return job;
1567		}
1568
#	Line 1453 | Line 1572 | public class ForkJoinPool extends Abstra
1572		*         scheduled for execution
1573		*/
1574		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
1575	+	if (task == null)
1576	+	throw new NullPointerException();
1577		ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1578	<	doSubmit(job);
1578	>	forkOrSubmit(job);
1579		return job;
1580		}
1581
#	Line 1464 | Line 1585 | public class ForkJoinPool extends Abstra
1585		*         scheduled for execution
1586		*/
1587		public ForkJoinTask<?> submit(Runnable task) {
1588	+	if (task == null)
1589	+	throw new NullPointerException();
1590		ForkJoinTask<?> job;
1591		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1592		job = (ForkJoinTask<?>) task;
1593		else
1594		job = ForkJoinTask.adapt(task, null);
1595	<	doSubmit(job);
1595	>	forkOrSubmit(job);
1596		return job;
1597		}
1598
#	Line 1529 | Line 1652 | public class ForkJoinPool extends Abstra
1652
1653		/**
1654		* Returns the number of worker threads that have started but not
1655	<	* yet terminated.  This result returned by this method may differ
1655	>	* yet terminated.  The result returned by this method may differ
1656		* from {@link #getParallelism} when threads are created to
1657		* maintain parallelism when others are cooperatively blocked.
1658		*
1659		* @return the number of worker threads
1660		*/
1661		public int getPoolSize() {
1662	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
1662	>	return parallelism + (short)(ctl >>> TC_SHIFT);
1663		}
1664
1665		/**
#	Line 1558 | Line 1681 | public class ForkJoinPool extends Abstra
1681		* @return the number of worker threads
1682		*/
1683		public int getRunningThreadCount() {
1684	<	return workerCounts & RUNNING_COUNT_MASK;
1684	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
1685	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
1686		}
1687
1688		/**
#	Line 1569 | Line 1693 | public class ForkJoinPool extends Abstra
1693		* @return the number of active threads
1694		*/
1695		public int getActiveThreadCount() {
1696	<	return runState & ACTIVE_COUNT_MASK;
1696	>	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
1697	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
1698		}
1699
1700		/**
#	Line 1584 | Line 1709 | public class ForkJoinPool extends Abstra
1709		* @return {@code true} if all threads are currently idle
1710		*/
1711		public boolean isQuiescent() {
1712	<	return (runState & ACTIVE_COUNT_MASK) == 0;
1712	>	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
1713		}
1714
1715		/**
#	Line 1614 | Line 1739 | public class ForkJoinPool extends Abstra
1739		*/
1740		public long getQueuedTaskCount() {
1741		long count = 0;
1742	<	ForkJoinWorkerThread[] ws = workers;
1743	<	int n = ws.length;
1744	<	for (int i = 0; i < n; ++i) {
1745	<	ForkJoinWorkerThread w = ws[i];
1746	<	if (w != null)
1747	<	count += w.getQueueSize();
1742	>	ForkJoinWorkerThread[] ws;
1743	>	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1744	>	(ws = workers) != null) {
1745	>	for (ForkJoinWorkerThread w : ws)
1746	>	if (w != null)
1747	>	count -= w.queueBase - w.queueTop; // must read base first
1748		}
1749		return count;
1750		}
1751
1752		/**
1753		* Returns an estimate of the number of tasks submitted to this
1754	<	* pool that have not yet begun executing.  This method takes time
1755	<	* proportional to the number of submissions.
1754	>	* pool that have not yet begun executing.  This method may take
1755	>	* time proportional to the number of submissions.
1756		*
1757		* @return the number of queued submissions
1758		*/
1759		public int getQueuedSubmissionCount() {
1760	<	return submissionQueue.size();
1760	>	return -queueBase + queueTop;
1761		}
1762
1763		/**
#	Line 1642 | Line 1767 | public class ForkJoinPool extends Abstra
1767		* @return {@code true} if there are any queued submissions
1768		*/
1769		public boolean hasQueuedSubmissions() {
1770	<	return !submissionQueue.isEmpty();
1770	>	return queueBase != queueTop;
1771		}
1772
1773		/**
#	Line 1653 | Line 1778 | public class ForkJoinPool extends Abstra
1778		* @return the next submission, or {@code null} if none
1779		*/
1780		protected ForkJoinTask<?> pollSubmission() {
1781	<	return submissionQueue.poll();
1781	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1782	>	while ((b = queueBase) != queueTop &&
1783	>	(q = submissionQueue) != null &&
1784	>	(i = (q.length - 1) & b) >= 0) {
1785	>	long u = (i << ASHIFT) + ABASE;
1786	>	if ((t = q[i]) != null &&
1787	>	queueBase == b &&
1788	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1789	>	queueBase = b + 1;
1790	>	return t;
1791	>	}
1792	>	}
1793	>	return null;
1794		}
1795
1796		/**
#	Line 1674 | Line 1811 | public class ForkJoinPool extends Abstra
1811		* @return the number of elements transferred
1812		*/
1813		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1814	<	int count = submissionQueue.drainTo(c);
1815	<	ForkJoinWorkerThread[] ws = workers;
1816	<	int n = ws.length;
1817	<	for (int i = 0; i < n; ++i) {
1818	<	ForkJoinWorkerThread w = ws[i];
1819	<	if (w != null)
1820	<	count += w.drainTasksTo(c);
1814	>	int count = 0;
1815	>	while (queueBase != queueTop) {
1816	>	ForkJoinTask<?> t = pollSubmission();
1817	>	if (t != null) {
1818	>	c.add(t);
1819	>	++count;
1820	>	}
1821	>	}
1822	>	ForkJoinWorkerThread[] ws;
1823	>	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1824	>	(ws = workers) != null) {
1825	>	for (ForkJoinWorkerThread w : ws)
1826	>	if (w != null)
1827	>	count += w.drainTasksTo(c);
1828		}
1829		return count;
1830		}
#	Line 1696 | Line 1840 | public class ForkJoinPool extends Abstra
1840		long st = getStealCount();
1841		long qt = getQueuedTaskCount();
1842		long qs = getQueuedSubmissionCount();
1699	–	int wc = workerCounts;
1700	–	int tc = wc >>> TOTAL_COUNT_SHIFT;
1701	–	int rc = wc & RUNNING_COUNT_MASK;
1843		int pc = parallelism;
1844	<	int rs = runState;
1845	<	int ac = rs & ACTIVE_COUNT_MASK;
1844	>	long c = ctl;
1845	>	int tc = pc + (short)(c >>> TC_SHIFT);
1846	>	int rc = pc + (int)(c >> AC_SHIFT);
1847	>	if (rc < 0) // ignore transient negative
1848	>	rc = 0;
1849	>	int ac = rc + blockedCount;
1850	>	String level;
1851	>	if ((c & STOP_BIT) != 0)
1852	>	level = (tc == 0) ? "Terminated" : "Terminating";
1853	>	else
1854	>	level = shutdown ? "Shutting down" : "Running";
1855		return super.toString() +
1856	<	"[" + runLevelToString(rs) +
1856	>	"[" + level +
1857		", parallelism = " + pc +
1858		", size = " + tc +
1859		", active = " + ac +
#	Line 1714 | Line 1864 | public class ForkJoinPool extends Abstra
1864		"]";
1865		}
1866
1717	–	private static String runLevelToString(int s) {
1718	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1719	–	((s & TERMINATING) != 0 ? "Terminating" :
1720	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1721	–	"Running")));
1722	–	}
1723	–
1867		/**
1868		* Initiates an orderly shutdown in which previously submitted
1869		* tasks are executed, but no new tasks will be accepted.
#	Line 1735 | Line 1878 | public class ForkJoinPool extends Abstra
1878		*/
1879		public void shutdown() {
1880		checkPermission();
1881	<	advanceRunLevel(SHUTDOWN);
1881	>	shutdown = true;
1882		tryTerminate(false);
1883		}
1884
#	Line 1757 | Line 1900 | public class ForkJoinPool extends Abstra
1900		*/
1901		public List<Runnable> shutdownNow() {
1902		checkPermission();
1903	+	shutdown = true;
1904		tryTerminate(true);
1905		return Collections.emptyList();
1906		}
#	Line 1767 | Line 1911 | public class ForkJoinPool extends Abstra
1911		* @return {@code true} if all tasks have completed following shut down
1912		*/
1913		public boolean isTerminated() {
1914	<	return runState >= TERMINATED;
1914	>	long c = ctl;
1915	>	return ((c & STOP_BIT) != 0L &&
1916	>	(short)(c >>> TC_SHIFT) == -parallelism);
1917		}
1918
1919		/**
#	Line 1775 | Line 1921 | public class ForkJoinPool extends Abstra
1921		* commenced but not yet completed.  This method may be useful for
1922		* debugging. A return of {@code true} reported a sufficient
1923		* period after shutdown may indicate that submitted tasks have
1924	<	* ignored or suppressed interruption, causing this executor not
1925	<	* to properly terminate.
1924	>	* ignored or suppressed interruption, or are waiting for IO,
1925	>	* causing this executor not to properly terminate. (See the
1926	>	* advisory notes for class {@link ForkJoinTask} stating that
1927	>	* tasks should not normally entail blocking operations.  But if
1928	>	* they do, they must abort them on interrupt.)
1929		*
1930		* @return {@code true} if terminating but not yet terminated
1931		*/
1932		public boolean isTerminating() {
1933	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
1933	>	long c = ctl;
1934	>	return ((c & STOP_BIT) != 0L &&
1935	>	(short)(c >>> TC_SHIFT) != -parallelism);
1936	>	}
1937	>
1938	>	/**
1939	>	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1940	>	*/
1941	>	final boolean isAtLeastTerminating() {
1942	>	return (ctl & STOP_BIT) != 0L;
1943		}
1944
1945		/**
#	Line 1790 | Line 1948 | public class ForkJoinPool extends Abstra
1948		* @return {@code true} if this pool has been shut down
1949		*/
1950		public boolean isShutdown() {
1951	<	return runState >= SHUTDOWN;
1951	>	return shutdown;
1952		}
1953
1954		/**
#	Line 1806 | Line 1964 | public class ForkJoinPool extends Abstra
1964		*/
1965		public boolean awaitTermination(long timeout, TimeUnit unit)
1966		throws InterruptedException {
1967	+	long nanos = unit.toNanos(timeout);
1968	+	final ReentrantLock lock = this.submissionLock;
1969	+	lock.lock();
1970		try {
1971	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
1972	<	} catch(TimeoutException ex) {
1973	<	return false;
1971	>	for (;;) {
1972	>	if (isTerminated())
1973	>	return true;
1974	>	if (nanos <= 0)
1975	>	return false;
1976	>	nanos = termination.awaitNanos(nanos);
1977	>	}
1978	>	} finally {
1979	>	lock.unlock();
1980		}
1981		}
1982
#	Line 1821 | Line 1988 | public class ForkJoinPool extends Abstra
1988		* {@code isReleasable} must return {@code true} if blocking is
1989		* not necessary. Method {@code block} blocks the current thread
1990		* if necessary (perhaps internally invoking {@code isReleasable}
1991	<	* before actually blocking). The unusual methods in this API
1992	<	* accommodate synchronizers that may, but don't usually, block
1993	<	* for long periods. Similarly, they allow more efficient internal
1994	<	* handling of cases in which additional workers may be, but
1995	<	* usually are not, needed to ensure sufficient parallelism.
1996	<	* Toward this end, implementations of method {@code isReleasable}
1997	<	* must be amenable to repeated invocation.
1991	>	* before actually blocking). These actions are performed by any
1992	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
1993	>	* unusual methods in this API accommodate synchronizers that may,
1994	>	* but don't usually, block for long periods. Similarly, they
1995	>	* allow more efficient internal handling of cases in which
1996	>	* additional workers may be, but usually are not, needed to
1997	>	* ensure sufficient parallelism.  Toward this end,
1998	>	* implementations of method {@code isReleasable} must be amenable
1999	>	* to repeated invocation.
2000		*
2001		* <p>For example, here is a ManagedBlocker based on a
2002		* ReentrantLock:
#	Line 1855 | Line 2024 | public class ForkJoinPool extends Abstra
2024		*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
2025		*   public boolean block() throws InterruptedException {
2026		*     if (item == null)
2027	<	*       item = queue.take
2027	>	*       item = queue.take();
2028		*     return true;
2029		*   }
2030		*   public boolean isReleasable() {
2031	<	*     return item != null || (item = queue.poll) != null;
2031	>	*     return item != null || (item = queue.poll()) != null;
2032		*   }
2033		*   public E getItem() { // call after pool.managedBlock completes
2034		*     return item;
#	Line 1929 | Line 2098 | public class ForkJoinPool extends Abstra
2098		}
2099
2100		// Unsafe mechanics
2101	<
2102	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
2103	<	private static final long workerCountsOffset =
2104	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
2105	<	private static final long runStateOffset =
2106	<	objectFieldOffset("runState", ForkJoinPool.class);
2107	<	private static final long eventCountOffset =
2108	<	objectFieldOffset("eventCount", ForkJoinPool.class);
2109	<	private static final long eventWaitersOffset =
2110	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
2111	<	private static final long stealCountOffset =
2112	<	objectFieldOffset("stealCount",ForkJoinPool.class);
2113	<	private static final long spareWaitersOffset =
2114	<	objectFieldOffset("spareWaiters",ForkJoinPool.class);
2115	<
2116	<	private static long objectFieldOffset(String field, Class<?> klazz) {
2101	>	private static final sun.misc.Unsafe UNSAFE;
2102	>	private static final long ctlOffset;
2103	>	private static final long stealCountOffset;
2104	>	private static final long blockedCountOffset;
2105	>	private static final long quiescerCountOffset;
2106	>	private static final long scanGuardOffset;
2107	>	private static final long nextWorkerNumberOffset;
2108	>	private static final long ABASE;
2109	>	private static final int ASHIFT;
2110	>
2111	>	static {
2112	>	poolNumberGenerator = new AtomicInteger();
2113	>	workerSeedGenerator = new Random();
2114	>	modifyThreadPermission = new RuntimePermission("modifyThread");
2115	>	defaultForkJoinWorkerThreadFactory =
2116	>	new DefaultForkJoinWorkerThreadFactory();
2117		try {
2118	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
2119	<	} catch (NoSuchFieldException e) {
2120	<	// Convert Exception to corresponding Error
2121	<	NoSuchFieldError error = new NoSuchFieldError(field);
2122	<	error.initCause(e);
2123	<	throw error;
2124	<	}
2118	>	UNSAFE = getUnsafe();
2119	>	Class<?> k = ForkJoinPool.class;
2120	>	ctlOffset = UNSAFE.objectFieldOffset
2121	>	(k.getDeclaredField("ctl"));
2122	>	stealCountOffset = UNSAFE.objectFieldOffset
2123	>	(k.getDeclaredField("stealCount"));
2124	>	blockedCountOffset = UNSAFE.objectFieldOffset
2125	>	(k.getDeclaredField("blockedCount"));
2126	>	quiescerCountOffset = UNSAFE.objectFieldOffset
2127	>	(k.getDeclaredField("quiescerCount"));
2128	>	scanGuardOffset = UNSAFE.objectFieldOffset
2129	>	(k.getDeclaredField("scanGuard"));
2130	>	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
2131	>	(k.getDeclaredField("nextWorkerNumber"));
2132	>	} catch (Exception e) {
2133	>	throw new Error(e);
2134	>	}
2135	>	Class<?> a = ForkJoinTask[].class;
2136	>	ABASE = UNSAFE.arrayBaseOffset(a);
2137	>	int s = UNSAFE.arrayIndexScale(a);
2138	>	if ((s & (s-1)) != 0)
2139	>	throw new Error("data type scale not a power of two");
2140	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
2141		}
2142
2143		/**

Diff Legend

–	Removed lines
+	Added lines
<	Changed lines
>	Changed lines