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 1.85 by dl, Sun Nov 21 13:55:04 2010 UTC vs.
Revision 1.93 by dl, Wed Feb 23 12:48:43 2011 UTC

#	Line 11 | Line 11 | 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;
#	Line 22 | Line 23 | import java.util.concurrent.TimeoutExcep
23		import java.util.concurrent.atomic.AtomicInteger;
24		import java.util.concurrent.locks.LockSupport;
25		import java.util.concurrent.locks.ReentrantLock;
26	+	import java.util.concurrent.locks.Condition;
27
28		/**
29		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 129 | Line 131 | public class ForkJoinPool extends Abstra
131		* set of worker threads: Submissions from non-FJ threads enter
132		* into a submission queue. Workers take these tasks and typically
133		* split them into subtasks that may be stolen by other workers.
134	<	* The main work-stealing mechanics implemented in class
135	<	* ForkJoinWorkerThread give first priority to processing tasks
136	<	* from their own queues (LIFO or FIFO, depending on mode), then
137	<	* to randomized FIFO steals of tasks in other worker queues, and
138	<	* lastly to new submissions. These mechanics do not consider
139	<	* affinities, loads, cache localities, etc, so rarely provide the
140	<	* best possible performance on a given machine, but portably
141	<	* provide good throughput by averaging over these factors.
142	<	* (Further, even if we did try to use such information, we do not
143	<	* usually have a basis for exploiting it. For example, some sets
144	<	* of tasks profit from cache affinities, but others are harmed by
145	<	* cache pollution effects.)
146	<	*
147	<	* Beyond work-stealing support and essential bookkeeping, the
148	<	* main responsibility of this framework is to take actions when
149	<	* one worker is waiting to join a task stolen (or always held by)
150	<	* another.  Because we are multiplexing many tasks on to a pool
151	<	* of workers, we can't just let them block (as in Thread.join).
152	<	* We also cannot just reassign the joiner's run-time stack with
153	<	* another and replace it later, which would be a form of
154	<	* "continuation", that even if possible is not necessarily a good
155	<	* idea. Given that the creation costs of most threads on most
156	<	* systems mainly surrounds setting up runtime stacks, thread
157	<	* creation and switching is usually not much more expensive than
158	<	* stack creation and switching, and is more flexible). Instead we
134	>	* Preference rules give first priority to processing tasks from
135	>	* their own queues (LIFO or FIFO, depending on mode), then to
136	>	* randomized FIFO steals of tasks in other worker queues, and
137	>	* lastly to new submissions.
138	>	*
139	>	* The main throughput advantages of work-stealing stem from
140	>	* decentralized control -- workers mostly take tasks from
141	>	* themselves or each other. We cannot negate this in the
142	>	* implementation of other management responsibilities. The main
143	>	* tactic for avoiding bottlenecks is packing nearly all
144	>	* essentially atomic control state into a single 64bit volatile
145	>	* variable ("ctl"). This variable is read on the order of 10-100
146	>	* times as often as it is modified (always via CAS). (There is
147	>	* some additional control state, for example variable "shutdown"
148	>	* for which we can cope with uncoordinated updates.)  This
149	>	* streamlines synchronization and control at the expense of messy
150	>	* constructions needed to repack status bits upon updates.
151	>	* Updates tend not to contend with each other except during
152	>	* bursts while submitted tasks begin or end.  In some cases when
153	>	* they do contend, threads can instead do something else
154	>	* (usually, scan for tesks) until contention subsides.
155	>	*
156	>	* To enable packing, we restrict maximum parallelism to (1<<15)-1
157	>	* (which is far in excess of normal operating range) to allow
158	>	* ids, counts, and their negations (used for thresholding) to fit
159	>	* into 16bit fields.
160	>	*
161	>	* Recording Workers.  Workers are recorded in the "workers" array
162	>	* that is created upon pool construction and expanded if (rarely)
163	>	* necessary.  This is an array as opposed to some other data
164	>	* structure to support index-based random steals by workers.
165	>	* Updates to the array recording new workers and unrecording
166	>	* terminated ones are protected from each other by a seqLock
167	>	* (scanGuard) but the array is otherwise concurrently readable,
168	>	* and accessed directly by workers. To simplify index-based
169	>	* operations, the array size is always a power of two, and all
170	>	* readers must tolerate null slots. To avoid flailing during
171	>	* start-up, the array is presized to hold twice #parallelism
172	>	* workers (which is unlikely to need further resizing during
173	>	* execution). But to avoid dealing with so many null slots,
174	>	* variable scanGuard includes a mask for the nearest power of two
175	>	* that contains all current workers.  All worker thread creation
176	>	* is on-demand, triggered by task submissions, replacement of
177	>	* terminated workers, and/or compensation for blocked
178	>	* workers. However, all other support code is set up to work with
179	>	* other policies.  To ensure that we do not hold on to worker
180	>	* references that would prevent GC, ALL accesses to workers are
181	>	* via indices into the workers array (which is one source of some
182	>	* of the messy code constructions here). In essence, the workers
183	>	* array serves as a weak reference mechanism. Thus for example
184	>	* the wait queue field of ctl stores worker indices, not worker
185	>	* references.  Access to the workers in associated methods (for
186	>	* example signalWork) must both index-check and null-check the
187	>	* IDs. All such accesses ignore bad IDs by returning out early
188	>	* from what they are doing, since this can only be associated
189	>	* with termination, in which case it is OK to give up.
190	>	*
191	>	* All uses of the workers array, as well as queue arrays, check
192	>	* that the array is non-null (even if previously non-null). This
193	>	* allows nulling during termination, which is currently not
194	>	* necessary, but remains an option for resource-revocation-based
195	>	* shutdown schemes.
196	>	*
197	>	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
198	>	* let workers spin indefinitely scanning for tasks when none are
199	>	* can be immediately found, and we cannot start/resume workers
200	>	* unless there appear to be tasks available.  On the other hand,
201	>	* we must quickly prod them into action when new tasks are
202	>	* submitted or generated.  We park/unpark workers after placing
203	>	* in an event wait queue when they cannot find work. This "queue"
204	>	* is actually a simple Treiber stack, headed by the "id" field of
205	>	* ctl, plus a 15bit counter value to both wake up waiters (by
206	>	* advancing their count) and avoid ABA effects. Successors are
207	>	* held in worker field "nextWait".  Queuing deals with several
208	>	* intrinsic races, mainly that a task-producing thread can miss
209	>	* seeing (and signalling) another thread that gave up looking for
210	>	* work but has not yet entered the wait queue. We solve this by
211	>	* requiring a full sweep of all workers both before (in scan())
212	>	* and after (in awaitWork()) a newly waiting worker is added to
213	>	* the wait queue. During a rescan, the worker might release some
214	>	* other queued worker rather than itself, which has the same net
215	>	* effect.
216	>	*
217	>	* Signalling.  We create or wake up workers only when there
218	>	* appears to be at least one task they might be able to find and
219	>	* execute.  When a submission is added or another worker adds a
220	>	* task to a queue that previously had two or fewer tasks, they
221	>	* signal waiting workers (or trigger creation of new ones if
222	>	* fewer than the given parallelism level -- see signalWork).
223	>	* These primary signals are buttressed by signals during rescans
224	>	* as well as those performed when a worker steals a task and
225	>	* notices that there are more tasks too; together these cover the
226	>	* signals needed in cases when more than two tasks are pushed
227	>	* but untaken.
228	>	*
229	>	* Trimming workers. To release resources after periods of lack of
230	>	* use, a worker starting to wait when the pool is quiescent will
231	>	* time out and terminate if the pool has remained quiescent for
232	>	* SHRINK_RATE nanosecs.
233	>	*
234	>	* Submissions. External submissions are maintained in an
235	>	* array-based queue that is structured identically to
236	>	* ForkJoinWorkerThread queues (which see) except for the use of
237	>	* submissionLock in method addSubmission. Unlike worker queues,
238	>	* multiple external threads can add new submissions.
239	>	*
240	>	* Compensation. Beyond work-stealing support and lifecycle
241	>	* control, the main responsibility of this framework is to take
242	>	* actions when one worker is waiting to join a task stolen (or
243	>	* always held by) another.  Because we are multiplexing many
244	>	* tasks on to a pool of workers, we can't just let them block (as
245	>	* in Thread.join).  We also cannot just reassign the joiner's
246	>	* run-time stack with another and replace it later, which would
247	>	* be a form of "continuation", that even if possible is not
248	>	* necessarily a good idea since we sometimes need both an
249	>	* unblocked task and its continuation to progress. Instead we
250		* combine two tactics:
251		*
252		*   Helping: Arranging for the joiner to execute some task that it
253		*      would be running if the steal had not occurred.  Method
254	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
254	>	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
255		*      links to try to find such a task.
256		*
257		*   Compensating: Unless there are already enough live threads,
258	<	*      method helpMaintainParallelism() may create or
259	<	*      re-activate a spare thread to compensate for blocked
260	<	*      joiners until they unblock.
168	<	*
169	<	* It is impossible to keep exactly the target (parallelism)
170	<	* number of threads running at any given time.  Determining
171	<	* existence of conservatively safe helping targets, the
172	<	* availability of already-created spares, and the apparent need
173	<	* to create new spares are all racy and require heuristic
174	<	* guidance, so we rely on multiple retries of each.  Compensation
175	<	* occurs in slow-motion. It is triggered only upon timeouts of
176	<	* Object.wait used for joins. This reduces poor decisions that
177	<	* would otherwise be made when threads are waiting for others
178	<	* that are stalled because of unrelated activities such as
179	<	* garbage collection.
258	>	*      method tryPreBlock() may create or re-activate a spare
259	>	*      thread to compensate for blocked joiners until they
260	>	*      unblock.
261		*
262		* The ManagedBlocker extension API can't use helping so relies
263		* only on compensation in method awaitBlocker.
264		*
265	<	* The main throughput advantages of work-stealing stem from
266	<	* decentralized control -- workers mostly steal tasks from each
267	<	* other. We do not want to negate this by creating bottlenecks
268	<	* implementing other management responsibilities. So we use a
269	<	* collection of techniques that avoid, reduce, or cope well with
270	<	* contention. These entail several instances of bit-packing into
271	<	* CASable fields to maintain only the minimally required
272	<	* atomicity. To enable such packing, we restrict maximum
273	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
274	<	* unbalanced increments and decrements) to fit into a 16 bit
275	<	* field, which is far in excess of normal operating range.  Even
276	<	* though updates to some of these bookkeeping fields do sometimes
277	<	* contend with each other, they don't normally cache-contend with
278	<	* updates to others enough to warrant memory padding or
279	<	* isolation. So they are all held as fields of ForkJoinPool
280	<	* objects.  The main capabilities are as follows:
281	<	*
282	<	* 1. Creating and removing workers. Workers are recorded in the
283	<	* "workers" array. This is an array as opposed to some other data
284	<	* structure to support index-based random steals by workers.
285	<	* Updates to the array recording new workers and unrecording
286	<	* terminated ones are protected from each other by a lock
287	<	* (workerLock) but the array is otherwise concurrently readable,
288	<	* and accessed directly by workers. To simplify index-based
289	<	* operations, the array size is always a power of two, and all
290	<	* readers must tolerate null slots. Currently, all worker thread
291	<	* creation is on-demand, triggered by task submissions,
292	<	* replacement of terminated workers, and/or compensation for
293	<	* blocked workers. However, all other support code is set up to
294	<	* work with other policies.
214	<	*
215	<	* To ensure that we do not hold on to worker references that
216	<	* would prevent GC, ALL accesses to workers are via indices into
217	<	* the workers array (which is one source of some of the unusual
218	<	* code constructions here). In essence, the workers array serves
219	<	* as a WeakReference mechanism. Thus for example the event queue
220	<	* stores worker indices, not worker references. Access to the
221	<	* workers in associated methods (for example releaseEventWaiters)
222	<	* must both index-check and null-check the IDs. All such accesses
223	<	* ignore bad IDs by returning out early from what they are doing,
224	<	* since this can only be associated with shutdown, in which case
225	<	* it is OK to give up. On termination, we just clobber these
226	<	* data structures without trying to use them.
227	<	*
228	<	* 2. Bookkeeping for dynamically adding and removing workers. We
229	<	* aim to approximately maintain the given level of parallelism.
230	<	* When some workers are known to be blocked (on joins or via
231	<	* ManagedBlocker), we may create or resume others to take their
232	<	* place until they unblock (see below). Implementing this
233	<	* requires counts of the number of "running" threads (i.e., those
234	<	* that are neither blocked nor artificially suspended) as well as
235	<	* the total number.  These two values are packed into one field,
236	<	* "workerCounts" because we need accurate snapshots when deciding
237	<	* to create, resume or suspend.  Note however that the
238	<	* correspondence of these counts to reality is not guaranteed. In
239	<	* particular updates for unblocked threads may lag until they
240	<	* actually wake up.
241	<	*
242	<	* 3. Maintaining global run state. The run state of the pool
243	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
244	<	* those in other Executor implementations, as well as a count of
245	<	* "active" workers -- those that are, or soon will be, or
246	<	* recently were executing tasks. The runLevel and active count
247	<	* are packed together in order to correctly trigger shutdown and
248	<	* termination. Without care, active counts can be subject to very
249	<	* high contention.  We substantially reduce this contention by
250	<	* relaxing update rules.  A worker must claim active status
251	<	* prospectively, by activating if it sees that a submitted or
252	<	* stealable task exists (it may find after activating that the
253	<	* task no longer exists). It stays active while processing this
254	<	* task (if it exists) and any other local subtasks it produces,
255	<	* until it cannot find any other tasks. It then tries
256	<	* inactivating (see method preStep), but upon update contention
257	<	* instead scans for more tasks, later retrying inactivation if it
258	<	* doesn't find any.
259	<	*
260	<	* 4. Managing idle workers waiting for tasks. We cannot let
261	<	* workers spin indefinitely scanning for tasks when none are
262	<	* available. On the other hand, we must quickly prod them into
263	<	* action when new tasks are submitted or generated.  We
264	<	* park/unpark these idle workers using an event-count scheme.
265	<	* Field eventCount is incremented upon events that may enable
266	<	* workers that previously could not find a task to now find one:
267	<	* Submission of a new task to the pool, or another worker pushing
268	<	* a task onto a previously empty queue.  (We also use this
269	<	* mechanism for configuration and termination actions that
270	<	* require wakeups of idle workers).  Each worker maintains its
271	<	* last known event count, and blocks when a scan for work did not
272	<	* find a task AND its lastEventCount matches the current
273	<	* eventCount. Waiting idle workers are recorded in a variant of
274	<	* Treiber stack headed by field eventWaiters which, when nonzero,
275	<	* encodes the thread index and count awaited for by the worker
276	<	* thread most recently calling eventSync. This thread in turn has
277	<	* a record (field nextEventWaiter) for the next waiting worker.
278	<	* In addition to allowing simpler decisions about need for
279	<	* wakeup, the event count bits in eventWaiters serve the role of
280	<	* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
281	<	* released threads also try to release at most two others.  The
282	<	* net effect is a tree-like diffusion of signals, where released
283	<	* threads (and possibly others) help with unparks.  To further
284	<	* reduce contention effects a bit, failed CASes to increment
285	<	* field eventCount are tolerated without retries in signalWork.
286	<	* Conceptually they are merged into the same event, which is OK
287	<	* when their only purpose is to enable workers to scan for work.
288	<	*
289	<	* 5. Managing suspension of extra workers. When a worker notices
290	<	* (usually upon timeout of a wait()) that there are too few
291	<	* running threads, we may create a new thread to maintain
292	<	* parallelism level, or at least avoid starvation. Usually, extra
293	<	* threads are needed for only very short periods, yet join
294	<	* dependencies are such that we sometimes need them in
295	<	* bursts. Rather than create new threads each time this happens,
296	<	* we suspend no-longer-needed extra ones as "spares". For most
297	<	* purposes, we don't distinguish "extra" spare threads from
298	<	* normal "core" threads: On each call to preStep (the only point
299	<	* at which we can do this) a worker checks to see if there are
300	<	* now too many running workers, and if so, suspends itself.
301	<	* Method helpMaintainParallelism looks for suspended threads to
302	<	* resume before considering creating a new replacement. The
303	<	* spares themselves are encoded on another variant of a Treiber
304	<	* Stack, headed at field "spareWaiters".  Note that the use of
305	<	* spares is intrinsically racy.  One thread may become a spare at
306	<	* about the same time as another is needlessly being created. We
307	<	* counteract this and related slop in part by requiring resumed
308	<	* spares to immediately recheck (in preStep) to see whether they
309	<	* should re-suspend.
310	<	*
311	<	* 6. Killing off unneeded workers. A timeout mechanism is used to
312	<	* shed unused workers: The oldest (first) event queue waiter uses
313	<	* a timed rather than hard wait. When this wait times out without
314	<	* a normal wakeup, it tries to shutdown any one (for convenience
315	<	* the newest) other spare or event waiter via
316	<	* tryShutdownUnusedWorker. This eventually reduces the number of
317	<	* worker threads to a minimum of one after a long enough period
318	<	* without use.
319	<	*
320	<	* 7. Deciding when to create new workers. The main dynamic
321	<	* control in this class is deciding when to create extra threads
322	<	* in method helpMaintainParallelism. We would like to keep
323	<	* exactly #parallelism threads running, which is an impossible
324	<	* task. We always need to create one when the number of running
325	<	* threads would become zero and all workers are busy. Beyond
326	<	* this, we must rely on heuristics that work well in the
327	<	* presence of transient phenomena such as GC stalls, dynamic
328	<	* compilation, and wake-up lags. These transients are extremely
329	<	* common -- we are normally trying to fully saturate the CPUs on
330	<	* a machine, so almost any activity other than running tasks
331	<	* impedes accuracy. Our main defense is to allow parallelism to
332	<	* lapse for a while during joins, and use a timeout to see if,
333	<	* after the resulting settling, there is still a need for
334	<	* additional workers.  This also better copes with the fact that
335	<	* some of the methods in this class tend to never become compiled
336	<	* (but are interpreted), so some components of the entire set of
337	<	* controls might execute 100 times faster than others. And
338	<	* similarly for cases where the apparent lack of work is just due
339	<	* to GC stalls and other transient system activity.
265	>	* It is impossible to keep exactly the target parallelism number
266	>	* of threads running at any given time.  Determining the
267	>	* existence of conservatively safe helping targets, the
268	>	* availability of already-created spares, and the apparent need
269	>	* to create new spares are all racy and require heuristic
270	>	* guidance, so we rely on multiple retries of each.  Currently,
271	>	* in keeping with on-demand signalling policy, we compensate only
272	>	* if blocking would leave less than one active (non-waiting,
273	>	* non-blocked) worker. Additionally, to avoid some false alarms
274	>	* due to GC, lagging counters, system activity, etc, compensated
275	>	* blocking for joins is only attempted after a number of rechecks
276	>	* proportional to the current apparent deficit (where retries are
277	>	* interspersed with Thread.yield, for good citizenship).  The
278	>	* variable blockedCount, incremented before blocking and
279	>	* decremented after, is sometimes needed to distinguish cases of
280	>	* waiting for work vs blocking on joins or other managed sync,
281	>	* but both the cases are equivalent for most pool control, so we
282	>	* can update non-atomically. (Additionally, contention on
283	>	* blockedCount alleviates some contention on ctl).
284	>	*
285	>	* Shutdown and Termination. A call to shutdownNow atomically sets
286	>	* the ctl stop bit and then (non-atomically) sets each workers
287	>	* "terminate" status, cancels all unprocessed tasks, and wakes up
288	>	* all waiting workers.  Detecting whether termination should
289	>	* commence after a non-abrupt shutdown() call requires more work
290	>	* and bookkeeping. We need consensus about quiesence (i.e., that
291	>	* there is no more work) which is reflected in active counts so
292	>	* long as there are no current blockers, as well as possible
293	>	* re-evaluations during independent changes in blocking or
294	>	* quiescing workers.
295		*
296	<	* Beware that there is a lot of representation-level coupling
296	>	* Style notes: There is a lot of representation-level coupling
297		* among classes ForkJoinPool, ForkJoinWorkerThread, and
298	<	* ForkJoinTask.  For example, direct access to "workers" array by
298	>	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
299	>	* data structures managed by ForkJoinPool, so are directly
300	>	* accessed.  Conversely we allow access to "workers" array by
301		* workers, and direct access to ForkJoinTask.status by both
302		* ForkJoinPool and ForkJoinWorkerThread.  There is little point
303		* trying to reduce this, since any associated future changes in
304		* representations will need to be accompanied by algorithmic
305	<	* changes anyway.
306	<	*
307	<	* Style notes: There are lots of inline assignments (of form
308	<	* "while ((local = field) != 0)") which are usually the simplest
309	<	* way to ensure the required read orderings (which are sometimes
310	<	* critical). Also several occurrences of the unusual "do {}
311	<	* while (!cas...)" which is the simplest way to force an update of
312	<	* a CAS'ed variable. There are also other coding oddities that
313	<	* help some methods perform reasonably even when interpreted (not
314	<	* compiled), at the expense of some messy constructions that
315	<	* reduce byte code counts.
316	<	*
317	<	* The order of declarations in this file is: (1) statics (2)
318	<	* fields (along with constants used when unpacking some of them)
319	<	* (3) internal control methods (4) callbacks and other support
320	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
321	<	* methods (plus a few little helpers).
305	>	* changes anyway. All together, these low-level implementation
306	>	* choices produce as much as a factor of 4 performance
307	>	* improvement compared to naive implementations, and enable the
308	>	* processing of billions of tasks per second, at the expense of
309	>	* some ugliness.
310	>	*
311	>	* Methods signalWork() and scan() are the main bottlenecks so are
312	>	* especially heavily micro-optimized/mangled.  There are lots of
313	>	* inline assignments (of form "while ((local = field) != 0)")
314	>	* which are usually the simplest way to ensure the required read
315	>	* orderings (which are sometimes critical). This leads to a
316	>	* "C"-like style of listing declarations of these locals at the
317	>	* heads of methods or blocks.  There are several occurrences of
318	>	* the unusual "do {} while (!cas...)"  which is the simplest way
319	>	* to force an update of a CAS'ed variable. There are also other
320	>	* coding oddities that help some methods perform reasonably even
321	>	* when interpreted (not compiled).
322	>	*
323	>	* The order of declarations in this file is: (1) declarations of
324	>	* statics (2) fields (along with constants used when unpacking
325	>	* some of them), listed in an order that tends to reduce
326	>	* contention among them a bit under most JVMs.  (3) internal
327	>	* control methods (4) callbacks and other support for
328	>	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
329	>	* methods (plus a few little helpers). (6) static block
330	>	* initializing all statics in a minimally dependent order.
331		*/
332
333		/**
#	Line 396 | Line 362 | public class ForkJoinPool extends Abstra
362		* overridden in ForkJoinPool constructors.
363		*/
364		public static final ForkJoinWorkerThreadFactory
365	<	defaultForkJoinWorkerThreadFactory =
400	<	new DefaultForkJoinWorkerThreadFactory();
365	>	defaultForkJoinWorkerThreadFactory;
366
367		/**
368		* Permission required for callers of methods that may start or
369		* kill threads.
370		*/
371	<	private static final RuntimePermission modifyThreadPermission =
407	<	new RuntimePermission("modifyThread");
371	>	private static final RuntimePermission modifyThreadPermission;
372
373		/**
374		* If there is a security manager, makes sure caller has
#	Line 419 | Line 383 | public class ForkJoinPool extends Abstra
383		/**
384		* Generator for assigning sequence numbers as pool names.
385		*/
386	<	private static final AtomicInteger poolNumberGenerator =
423	<	new AtomicInteger();
386	>	private static final AtomicInteger poolNumberGenerator;
387
388		/**
389	<	* The time to block in a join (see awaitJoin) before checking if
390	<	* a new worker should be (re)started to maintain parallelism
391	<	* level. The value should be short enough to maintain global
392	<	* responsiveness and progress but long enough to avoid
393	<	* counterproductive firings during GC stalls or unrelated system
431	<	* activity, and to not bog down systems with continual re-firings
432	<	* on GCs or legitimately long waits.
389	>	* Generator for initial random seeds for worker victim
390	>	* selection. This is used only to create initial seeds. Random
391	>	* steals use a cheaper xorshift generator per steal attempt. We
392	>	* don't expect much contention on seedGenerator, so just use a
393	>	* plain Random.
394		*/
395	<	private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
395	>	static final Random workerSeedGenerator;
396
397		/**
398	<	* The wakeup interval (in nanoseconds) for the oldest worker
399	<	* waiting for an event to invoke tryShutdownUnusedWorker to
400	<	* shrink the number of workers.  The exact value does not matter
401	<	* too much. It must be short enough to release resources during
402	<	* sustained periods of idleness, but not so short that threads
403	<	* are continually re-created.
398	>	* Array holding all worker threads in the pool.  Initialized upon
399	>	* construction. Array size must be a power of two.  Updates and
400	>	* replacements are protected by scanGuard, but the array is
401	>	* always kept in a consistent enough state to be randomly
402	>	* accessed without locking by workers performing work-stealing,
403	>	* as well as other traversal-based methods in this class, so long
404	>	* as reads memory-acquire by first reading ctl. All readers must
405	>	* tolerate that some array slots may be null.
406		*/
407	<	private static final long SHRINK_RATE_NANOS =
445	<	30L * 1000L * 1000L * 1000L; // 2 per minute
407	>	ForkJoinWorkerThread[] workers;
408
409		/**
410	<	* Absolute bound for parallelism level. Twice this number plus
411	<	* one (i.e., 0xfff) must fit into a 16bit field to enable
412	<	* word-packing for some counts and indices.
410	>	* Initial size for submission queue array. Must be a power of
411	>	* two.  In many applications, these always stay small so we use a
412	>	* small initial cap.
413		*/
414	<	private static final int MAX_WORKERS   = 0x7fff;
414	>	private static final int INITIAL_QUEUE_CAPACITY = 8;
415
416		/**
417	<	* Array holding all worker threads in the pool.  Array size must
418	<	* be a power of two.  Updates and replacements are protected by
419	<	* workerLock, but the array is always kept in a consistent enough
420	<	* state to be randomly accessed without locking by workers
459	<	* performing work-stealing, as well as other traversal-based
460	<	* methods in this class. All readers must tolerate that some
461	<	* array slots may be null.
417	>	* Maximum size for submission queue array. Must be a power of two
418	>	* less than or equal to 1 << (31 - width of array entry) to
419	>	* ensure lack of index wraparound, but is capped at a lower
420	>	* value to help users trap runaway computations.
421		*/
422	<	volatile ForkJoinWorkerThread[] workers;
422	>	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
423
424		/**
425	<	* Queue for external submissions.
425	>	* Array serving as submission queue. Initialized upon construction.
426		*/
427	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
427	>	private ForkJoinTask<?>[] submissionQueue;
428
429		/**
430	<	* Lock protecting updates to workers array.
430	>	* Lock protecting submissions array for addSubmission
431		*/
432	<	private final ReentrantLock workerLock;
432	>	private final ReentrantLock submissionLock;
433
434		/**
435	<	* Latch released upon termination.
435	>	* Condition for awaitTermination, using submissionLock for
436	>	* convenience.
437		*/
438	<	private final Phaser termination;
438	>	private final Condition termination;
439
440		/**
441		* Creation factory for worker threads.
#	Line 483 | Line 443 | public class ForkJoinPool extends Abstra
443		private final ForkJoinWorkerThreadFactory factory;
444
445		/**
446	<	* Sum of per-thread steal counts, updated only when threads are
447	<	* idle or terminating.
446	>	* The uncaught exception handler used when any worker abruptly
447	>	* terminates.
448		*/
449	<	private volatile long stealCount;
449	>	final Thread.UncaughtExceptionHandler ueh;
450
451		/**
452	<	* Encoded record of top of Treiber stack of threads waiting for
493	<	* events. The top 32 bits contain the count being waited for. The
494	<	* bottom 16 bits contains one plus the pool index of waiting
495	<	* worker thread. (Bits 16-31 are unused.)
452	>	* Prefix for assigning names to worker threads
453		*/
454	<	private volatile long eventWaiters;
498	<
499	<	private static final int  EVENT_COUNT_SHIFT = 32;
500	<	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
454	>	private final String workerNamePrefix;
455
456		/**
457	<	* A counter for events that may wake up worker threads:
458	<	*   - Submission of a new task to the pool
505	<	*   - A worker pushing a task on an empty queue
506	<	*   - termination
457	>	* Sum of per-thread steal counts, updated only when threads are
458	>	* idle or terminating.
459		*/
460	<	private volatile int eventCount;
460	>	private volatile long stealCount;
461
462		/**
463	<	* Encoded record of top of Treiber stack of spare threads waiting
464	<	* for resumption. The top 16 bits contain an arbitrary count to
465	<	* avoid ABA effects. The bottom 16bits contains one plus the pool
466	<	* index of waiting worker thread.
467	<	*/
468	<	private volatile int spareWaiters;
469	<
470	<	private static final int SPARE_COUNT_SHIFT = 16;
471	<	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
463	>	* Main pool control -- a long packed with:
464	>	* AC: Number of active running workers minus target parallelism (16 bits)
465	>	* TC: Number of total workers minus target parallelism (16bits)
466	>	* ST: true if pool is terminating (1 bit)
467	>	* EC: the wait count of top waiting thread (15 bits)
468	>	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
469	>	*
470	>	* When convenient, we can extract the upper 32 bits of counts and
471	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
472	>	* (int)ctl.  The ec field is never accessed alone, but always
473	>	* together with id and st. The offsets of counts by the target
474	>	* parallelism and the positionings of fields makes it possible to
475	>	* perform the most common checks via sign tests of fields: When
476	>	* ac is negative, there are not enough active workers, when tc is
477	>	* negative, there are not enough total workers, when id is
478	>	* negative, there is at least one waiting worker, and when e is
479	>	* negative, the pool is terminating.  To deal with these possibly
480	>	* negative fields, we use casts in and out of "short" and/or
481	>	* signed shifts to maintain signedness.  Note: AC_SHIFT is
482	>	* redundantly declared in ForkJoinWorkerThread in order to
483	>	* integrate a surplus-threads check.
484	>	*/
485	>	volatile long ctl;
486	>
487	>	// bit positions/shifts for fields
488	>	private static final int  AC_SHIFT   = 48;
489	>	private static final int  TC_SHIFT   = 32;
490	>	private static final int  ST_SHIFT   = 31;
491	>	private static final int  EC_SHIFT   = 16;
492	>
493	>	// bounds
494	>	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
495	>	private static final int  SMASK      = 0xffff;  // mask short bits
496	>	private static final int  SHORT_SIGN = 1 << 15;
497	>	private static final int  INT_SIGN   = 1 << 31;
498	>
499	>	// masks
500	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
501	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
502	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
503	>
504	>	// units for incrementing and decrementing
505	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
506	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
507	>
508	>	// masks and units for dealing with u = (int)(ctl >>> 32)
509	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
510	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
511	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
512	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
513	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
514	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
515	>
516	>	// masks and units for dealing with e = (int)ctl
517	>	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
518	>	private static final int  EC_UNIT    = 1 << EC_SHIFT;
519
520		/**
521	<	* Lifecycle control. The low word contains the number of workers
523	<	* that are (probably) executing tasks. This value is atomically
524	<	* incremented before a worker gets a task to run, and decremented
525	<	* when a worker has no tasks and cannot find any.  Bits 16-18
526	<	* contain runLevel value. When all are zero, the pool is
527	<	* running. Level transitions are monotonic (running -> shutdown
528	<	* -> terminating -> terminated) so each transition adds a bit.
529	<	* These are bundled together to ensure consistent read for
530	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
531	<	* and active threads is zero).
532	<	*
533	<	* Notes: Most direct CASes are dependent on these bitfield
534	<	* positions.  Also, this field is non-private to enable direct
535	<	* performance-sensitive CASes in ForkJoinWorkerThread.
521	>	* The target parallelism level.
522		*/
523	<	volatile int runState;
538	<
539	<	// Note: The order among run level values matters.
540	<	private static final int RUNLEVEL_SHIFT     = 16;
541	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
542	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
543	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
544	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
523	>	final int parallelism;
524
525		/**
526	<	* Holds number of total (i.e., created and not yet terminated)
527	<	* and running (i.e., not blocked on joins or other managed sync)
549	<	* threads, packed together to ensure consistent snapshot when
550	<	* making decisions about creating and suspending spare
551	<	* threads. Updated only by CAS. Note that adding a new worker
552	<	* requires incrementing both counts, since workers start off in
553	<	* running state.
526	>	* Index (mod submission queue length) of next element to take
527	>	* from submission queue.
528		*/
529	<	private volatile int workerCounts;
529	>	volatile int queueBase;
530
531	<	private static final int TOTAL_COUNT_SHIFT  = 16;
532	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
533	<	private static final int ONE_RUNNING        = 1;
534	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
531	>	/**
532	>	* Index (mod submission queue length) of next element to add
533	>	* in submission queue.
534	>	*/
535	>	int queueTop;
536
537		/**
538	<	* The target parallelism level.
564	<	* Accessed directly by ForkJoinWorkerThreads.
538	>	* True when shutdown() has been called.
539		*/
540	<	final int parallelism;
540	>	volatile boolean shutdown;
541
542		/**
543		* True if use local fifo, not default lifo, for local polling
#	Line 572 | Line 546 | public class ForkJoinPool extends Abstra
546		final boolean locallyFifo;
547
548		/**
549	<	* The uncaught exception handler used when any worker abruptly
550	<	* terminates.
549	>	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
550	>	* When non-zero, suppresses automatic shutdown when active
551	>	* counts become zero.
552		*/
553	<	private final Thread.UncaughtExceptionHandler ueh;
553	>	volatile int quiescerCount;
554
555		/**
556	<	* Pool number, just for assigning useful names to worker threads
556	>	* The number of threads blocked in join.
557		*/
558	<	private final int poolNumber;
584	<
585	<	// Utilities for CASing fields. Note that most of these
586	<	// are usually manually inlined by callers
558	>	volatile int blockedCount;
559
560		/**
561	<	* Increments running count part of workerCounts
561	>	* Counter for worker Thread names (unrelated to their poolIndex)
562		*/
563	<	final void incrementRunningCount() {
592	<	int c;
593	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
594	<	c = workerCounts,
595	<	c + ONE_RUNNING));
596	<	}
563	>	private volatile int nextWorkerNumber;
564
565		/**
566	<	* Tries to increment running count part of workerCounts
566	>	* The index for the next created worker. Accessed under scanGuard.
567		*/
568	<	final boolean tryIncrementRunningCount() {
602	<	int c;
603	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
604	<	c = workerCounts,
605	<	c + ONE_RUNNING);
606	<	}
568	>	private int nextWorkerIndex;
569
570		/**
571	<	* Tries to decrement running count unless already zero
571	>	* SeqLock and index masking for for updates to workers array.
572	>	* Locked when SG_UNIT is set. Unlocking clears bit by adding
573	>	* SG_UNIT. Staleness of read-only operations can be checked by
574	>	* comparing scanGuard to value before the reads. The low 16 bits
575	>	* (i.e, anding with SMASK) hold (the smallest power of two
576	>	* covering all worker indices, minus one, and is used to avoid
577	>	* dealing with large numbers of null slots when the workers array
578	>	* is overallocated.
579		*/
580	<	final boolean tryDecrementRunningCount() {
612	<	int wc = workerCounts;
613	<	if ((wc & RUNNING_COUNT_MASK) == 0)
614	<	return false;
615	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
616	<	wc, wc - ONE_RUNNING);
617	<	}
580	>	volatile int scanGuard;
581
582	<	/**
620	<	* Forces decrement of encoded workerCounts, awaiting nonzero if
621	<	* (rarely) necessary when other count updates lag.
622	<	*
623	<	* @param dr -- either zero or ONE_RUNNING
624	<	* @param dt -- either zero or ONE_TOTAL
625	<	*/
626	<	private void decrementWorkerCounts(int dr, int dt) {
627	<	for (;;) {
628	<	int wc = workerCounts;
629	<	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
630	<	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
631	<	if ((runState & TERMINATED) != 0)
632	<	return; // lagging termination on a backout
633	<	Thread.yield();
634	<	}
635	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
636	<	wc, wc - (dr + dt)))
637	<	return;
638	<	}
639	<	}
582	>	private static final int SG_UNIT = 1 << 16;
583
584		/**
585	<	* Tries decrementing active count; fails on contention.
586	<	* Called when workers cannot find tasks to run.
585	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
586	>	* task when the pool is quiescent to instead try to shrink the
587	>	* number of workers.  The exact value does not matter too
588	>	* much. It must be short enough to release resources during
589	>	* sustained periods of idleness, but not so short that threads
590	>	* are continually re-created.
591		*/
592	<	final boolean tryDecrementActiveCount() {
593	<	int c;
647	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
648	<	c = runState, c - 1);
649	<	}
592	>	private static final long SHRINK_RATE =
593	>	4L * 1000L * 1000L * 1000L; // 4 seconds
594
595		/**
596	<	* Advances to at least the given level. Returns true if not
597	<	* already in at least the given level.
596	>	* Top-level loop for worker threads: On each step: if the
597	>	* previous step swept through all queues and found no tasks, or
598	>	* there are excess threads, then possibly blocks. Otherwise,
599	>	* scans for and, if found, executes a task. Returns when pool
600	>	* and/or worker terminate.
601	>	*
602	>	* @param w the worker
603		*/
604	<	private boolean advanceRunLevel(int level) {
605	<	for (;;) {
606	<	int s = runState;
607	<	if ((s & level) != 0)
608	<	return false;
609	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
610	<	return true;
604	>	final void work(ForkJoinWorkerThread w) {
605	>	boolean swept = false;                // true on empty scans
606	>	long c;
607	>	while (!w.terminate && (int)(c = ctl) >= 0) {
608	>	int a;                            // active count
609	>	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
610	>	swept = scan(w, a);
611	>	else if (tryAwaitWork(w, c))
612	>	swept = false;
613		}
614		}
615
616	<	// workers array maintenance
616	>	// Signalling
617
618		/**
619	<	* Records and returns a workers array index for new worker.
619	>	* Wakes up or creates a worker.
620		*/
621	<	private int recordWorker(ForkJoinWorkerThread w) {
622	<	// Try using slot totalCount-1. If not available, scan and/or resize
623	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
624	<	final ReentrantLock lock = this.workerLock;
625	<	lock.lock();
626	<	try {
627	<	ForkJoinWorkerThread[] ws = workers;
628	<	int n = ws.length;
629	<	if (k < 0 || k >= n || ws[k] != null) {
630	<	for (k = 0; k < n && ws[k] != null; ++k)
631	<	;
632	<	if (k == n)
633	<	ws = workers = Arrays.copyOf(ws, n << 1);
634	<	}
635	<	ws[k] = w;
636	<	int c = eventCount; // advance event count to ensure visibility
637	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c, c+1);
638	<	} finally {
639	<	lock.unlock();
621	>	final void signalWork() {
622	>	/*
623	>	* The while condition is true if: (there is are too few total
624	>	* workers OR there is at least one waiter) AND (there are too
625	>	* few active workers OR the pool is terminating).  The value
626	>	* of e distinguishes the remaining cases: zero (no waiters)
627	>	* for create, negative if terminating (in which case do
628	>	* nothing), else release a waiter. The secondary checks for
629	>	* release (non-null array etc) can fail if the pool begins
630	>	* terminating after the test, and don't impose any added cost
631	>	* because JVMs must perform null and bounds checks anyway.
632	>	*/
633	>	long c; int e, u;
634	>	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
635	>	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
636	>	if (e > 0) {                         // release a waiting worker
637	>	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
638	>	if ((ws = workers) == null ||
639	>	(i = ~e & SMASK) >= ws.length ||
640	>	(w = ws[i]) == null)
641	>	break;
642	>	long nc = (((long)(w.nextWait & E_MASK)) |
643	>	((long)(u + UAC_UNIT) << 32));
644	>	if (w.eventCount == e &&
645	>	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
646	>	w.eventCount = (e + EC_UNIT) & E_MASK;
647	>	if (w.parked)
648	>	UNSAFE.unpark(w);
649	>	break;
650	>	}
651	>	}
652	>	else if (UNSAFE.compareAndSwapLong
653	>	(this, ctlOffset, c,
654	>	(long)(((u + UTC_UNIT) & UTC_MASK) |
655	>	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
656	>	addWorker();
657	>	break;
658	>	}
659		}
690	–	return k;
660		}
661
662		/**
663	<	* Nulls out record of worker in workers array.
663	>	* Variant of signalWork to help release waiters on rescans.
664	>	* Tries once to release a waiter if active count < 0.
665	>	*
666	>	* @return false if failed due to contention, else true
667		*/
668	<	private void forgetWorker(ForkJoinWorkerThread w) {
669	<	int idx = w.poolIndex;
670	<	// Locking helps method recordWorker avoid unnecessary expansion
671	<	final ReentrantLock lock = this.workerLock;
672	<	lock.lock();
673	<	try {
674	<	ForkJoinWorkerThread[] ws = workers;
675	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
676	<	ws[idx] = null;
677	<	} finally {
678	<	lock.unlock();
668	>	private boolean tryReleaseWaiter() {
669	>	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
670	>	if ((e = (int)(c = ctl)) > 0 &&
671	>	(int)(c >> AC_SHIFT) < 0 &&
672	>	(ws = workers) != null &&
673	>	(i = ~e & SMASK) < ws.length &&
674	>	(w = ws[i]) != null) {
675	>	long nc = ((long)(w.nextWait & E_MASK) |
676	>	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
677	>	if (w.eventCount != e ||
678	>	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
679	>	return false;
680	>	w.eventCount = (e + EC_UNIT) & E_MASK;
681	>	if (w.parked)
682	>	UNSAFE.unpark(w);
683		}
684	+	return true;
685		}
686
687	+	// Scanning for tasks
688	+
689		/**
690	<	* Final callback from terminating worker.  Removes record of
691	<	* worker from array, and adjusts counts. If pool is shutting
692	<	* down, tries to complete termination.
690	>	* Scans for and, if found, executes one task. Scans start at a
691	>	* random index of workers array, and randomly select the first
692	>	* (2*#workers)-1 probes, and then, if all empty, resort to 2
693	>	* circular sweeps, which is necessary to check quiescence. and
694	>	* taking a submission only if no stealable tasks were found.  The
695	>	* steal code inside the loop is a specialized form of
696	>	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
697	>	* helpJoinTask and signal propagation. The code for submission
698	>	* queues is almost identical. On each steal, the worker completes
699	>	* not only the task, but also all local tasks that this task may
700	>	* have generated. On detecting staleness or contention when
701	>	* trying to take a task, this method returns without finishing
702	>	* sweep, which allows global state rechecks before retry.
703		*
704		* @param w the worker
705	+	* @param a the number of active workers
706	+	* @return true if swept all queues without finding a task
707		*/
708	<	final void workerTerminated(ForkJoinWorkerThread w) {
709	<	forgetWorker(w);
710	<	decrementWorkerCounts(w.isTrimmed() ? 0 : ONE_RUNNING, ONE_TOTAL);
720	<	while (w.stealCount != 0) // collect final count
721	<	tryAccumulateStealCount(w);
722	<	tryTerminate(false);
723	<	}
724	<
725	<	// Waiting for and signalling events
726	<
727	<	/**
728	<	* Releases workers blocked on a count not equal to current count.
729	<	* Normally called after precheck that eventWaiters isn't zero to
730	<	* avoid wasted array checks. Gives up upon a change in count or
731	<	* upon releasing two workers, letting others take over.
732	<	*/
733	<	private void releaseEventWaiters() {
708	>	private boolean scan(ForkJoinWorkerThread w, int a) {
709	>	int g = scanGuard; // mask 0 avoids useless scans if only one active
710	>	int m = parallelism == 1 - a? 0 : g & SMASK;
711		ForkJoinWorkerThread[] ws = workers;
712	<	int n = ws.length;
713	<	long h = eventWaiters;
714	<	int ec = eventCount;
715	<	boolean releasedOne = false;
716	<	ForkJoinWorkerThread w; int id;
717	<	while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
718	<	(int)(h >>> EVENT_COUNT_SHIFT) != ec &&
719	<	id < n && (w = ws[id]) != null) {
720	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
721	<	h,  w.nextWaiter)) {
722	<	LockSupport.unpark(w);
723	<	if (releasedOne) // exit on second release
724	<	break;
725	<	releasedOne = true;
712	>	if (ws == null || ws.length <= m)         // staleness check
713	>	return false;
714	>	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
715	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
716	>	ForkJoinWorkerThread v = ws[k & m];
717	>	if (v != null && (b = v.queueBase) != v.queueTop &&
718	>	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
719	>	long u = (i << ASHIFT) + ABASE;
720	>	if ((t = q[i]) != null && v.queueBase == b &&
721	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
722	>	int d = (v.queueBase = b + 1) - v.queueTop;
723	>	v.stealHint = w.poolIndex;
724	>	if (d != 0)
725	>	signalWork();             // propagate if nonempty
726	>	w.execTask(t);
727	>	}
728	>	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
729	>	return false;                     // store next seed
730		}
731	<	if (eventCount != ec)
732	<	break;
733	<	h = eventWaiters;
731	>	else if (j < 0) {                     // xorshift
732	>	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
733	>	}
734	>	else
735	>	++k;
736	>	}
737	>	if (scanGuard != g)                       // staleness check
738	>	return false;
739	>	else {                                    // try to take submission
740	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
741	>	if ((b = queueBase) != queueTop &&
742	>	(q = submissionQueue) != null &&
743	>	(i = (q.length - 1) & b) >= 0) {
744	>	long u = (i << ASHIFT) + ABASE;
745	>	if ((t = q[i]) != null && queueBase == b &&
746	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
747	>	queueBase = b + 1;
748	>	w.execTask(t);
749	>	}
750	>	return false;
751	>	}
752	>	return true;                         // all queues empty
753		}
754		}
755
756		/**
757	<	* Tries to advance eventCount and releases waiters. Called only
758	<	* from workers.
759	<	*/
760	<	final void signalWork() {
761	<	int c; // try to increment event count -- CAS failure OK
762	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
763	<	if (eventWaiters != 0L)
764	<	releaseEventWaiters();
765	<	}
766	<
767	<	/**
768	<	* Adds the given worker to event queue and blocks until
769	<	* terminating or event count advances from the given value
770	<	*
771	<	* @param w the calling worker thread
772	<	* @param ec the count
773	<	*/
774	<	private void eventSync(ForkJoinWorkerThread w, int ec) {
775	<	long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
776	<	long h;
777	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
778	<	(((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
779	<	(int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
780	<	eventCount == ec) {
781	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
782	<	w.nextWaiter = h, nh)) {
783	<	awaitEvent(w, ec);
784	<	break;
757	>	* Tries to enqueue worker w in wait queue and await change in
758	>	* worker's eventCount.  If the pool is quiescent, possibly
759	>	* terminates worker upon exit.  Otherwise, before blocking,
760	>	* rescans queues to avoid missed signals.  Upon finding work,
761	>	* releases at least one worker (which may be the current
762	>	* worker). Rescans restart upon detected staleness or failure to
763	>	* release due to contention.
764	>	*
765	>	* @param w the calling worker
766	>	* @param c the ctl value on entry
767	>	* @return true if waited or another thread was released upon enq
768	>	*/
769	>	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
770	>	int v = w.eventCount;
771	>	w.nextWait = (int)c;                      // w's successor record
772	>	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
773	>	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
774	>	long d = ctl; // return true if lost to a deq, to force scan
775	>	return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
776	>	}
777	>	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
778	>	long s = stealCount;
779	>	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
780	>	sc = w.stealCount = 0;
781	>	else if (w.eventCount != v)
782	>	return true;                      // update next time
783	>	}
784	>	if (parallelism + (int)(nc >> AC_SHIFT) == 0 &&
785	>	blockedCount == 0 && quiescerCount == 0)
786	>	idleAwaitWork(w, nc, c, v);           // quiescent
787	>	for (boolean rescanned = false;;) {
788	>	if (w.eventCount != v)
789	>	return true;
790	>	if (!rescanned) {
791	>	int g = scanGuard, m = g & SMASK;
792	>	ForkJoinWorkerThread[] ws = workers;
793	>	if (ws != null && m < ws.length) {
794	>	rescanned = true;
795	>	for (int i = 0; i <= m; ++i) {
796	>	ForkJoinWorkerThread u = ws[i];
797	>	if (u != null) {
798	>	if (u.queueBase != u.queueTop &&
799	>	!tryReleaseWaiter())
800	>	rescanned = false; // contended
801	>	if (w.eventCount != v)
802	>	return true;
803	>	}
804	>	}
805	>	}
806	>	if (scanGuard != g ||              // stale
807	>	(queueBase != queueTop && !tryReleaseWaiter()))
808	>	rescanned = false;
809	>	if (!rescanned)
810	>	Thread.yield();                // reduce contention
811	>	else
812	>	Thread.interrupted();          // clear before park
813	>	}
814	>	else {
815	>	w.parked = true;                   // must recheck
816	>	if (w.eventCount != v) {
817	>	w.parked = false;
818	>	return true;
819	>	}
820	>	LockSupport.park(this);
821	>	rescanned = w.parked = false;
822		}
823		}
824		}
825
826		/**
827	<	* Blocks the given worker (that has already been entered as an
828	<	* event waiter) until terminating or event count advances from
829	<	* the given value. The oldest (first) waiter uses a timed wait to
830	<	* occasionally one-by-one shrink the number of workers (to a
831	<	* minimum of one) if the pool has not been used for extended
832	<	* periods.
833	<	*
834	<	* @param w the calling worker thread
835	<	* @param ec the count
836	<	*/
837	<	private void awaitEvent(ForkJoinWorkerThread w, int ec) {
838	<	while (eventCount == ec) {
839	<	if (tryAccumulateStealCount(w)) { // transfer while idle
840	<	boolean untimed = (w.nextWaiter != 0L ||
841	<	(workerCounts & RUNNING_COUNT_MASK) <= 1);
842	<	long startTime = untimed ? 0 : System.nanoTime();
843	<	Thread.interrupted();         // clear/ignore interrupt
844	<	if (w.isTerminating() || eventCount != ec)
845	<	break;                    // recheck after clear
846	<	if (untimed)
847	<	LockSupport.park(w);
848	<	else {
849	<	LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
850	<	if (eventCount != ec || w.isTerminating())
851	<	break;
852	<	if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
853	<	tryShutdownUnusedWorker(ec);
827	>	* If inactivating worker w has caused pool to become
828	>	* quiescent, check for pool termination, and wait for event
829	>	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
830	>	* this case because quiescence reflects consensus about lack
831	>	* of work). On timeout, if ctl has not changed, terminate the
832	>	* worker. Upon its termination (see deregisterWorker), it may
833	>	* wake up another worker to possibly repeat this process.
834	>	*
835	>	* @param w the calling worker
836	>	* @param currentCtl the ctl value after enqueuing w
837	>	* @param prevCtl the ctl value if w terminated
838	>	* @param v the eventCount w awaits change
839	>	*/
840	>	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
841	>	long prevCtl, int v) {
842	>	if (w.eventCount == v) {
843	>	if (shutdown)
844	>	tryTerminate(false);
845	>	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
846	>	while (ctl == currentCtl) {
847	>	long startTime = System.nanoTime();
848	>	w.parked = true;
849	>	if (w.eventCount == v)             // must recheck
850	>	LockSupport.parkNanos(this, SHRINK_RATE);
851	>	w.parked = false;
852	>	if (w.eventCount != v)
853	>	break;
854	>	else if (System.nanoTime() - startTime < SHRINK_RATE)
855	>	Thread.interrupted();          // spurious wakeup
856	>	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
857	>	currentCtl, prevCtl)) {
858	>	w.terminate = true;            // restore previous
859	>	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
860	>	break;
861		}
862		}
863		}
864		}
865
866	<	// Maintaining parallelism
866	>	// Submissions
867
868		/**
869	<	* Pushes worker onto the spare stack.
869	>	* Enqueues the given task in the submissionQueue.  Same idea as
870	>	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
871	>	*
872	>	* @param t the task
873		*/
874	<	final void pushSpare(ForkJoinWorkerThread w) {
875	<	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
876	<	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
877	<	w.nextSpare = spareWaiters,ns));
874	>	private void addSubmission(ForkJoinTask<?> t) {
875	>	final ReentrantLock lock = this.submissionLock;
876	>	lock.lock();
877	>	try {
878	>	ForkJoinTask<?>[] q; int s, m;
879	>	if ((q = submissionQueue) != null) {    // ignore if queue removed
880	>	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
881	>	UNSAFE.putOrderedObject(q, u, t);
882	>	queueTop = s + 1;
883	>	if (s - queueBase == m)
884	>	growSubmissionQueue();
885	>	}
886	>	} finally {
887	>	lock.unlock();
888	>	}
889	>	signalWork();
890		}
891
892	+	//  (pollSubmission is defined below with exported methods)
893	+
894		/**
895	<	* Tries (once) to resume a spare if the number of running
896	<	* threads is less than target.
895	>	* Creates or doubles submissionQueue array.
896	>	* Basically identical to ForkJoinWorkerThread version
897		*/
898	<	private void tryResumeSpare() {
899	<	int sw, id;
900	<	ForkJoinWorkerThread[] ws = workers;
901	<	int n = ws.length;
902	<	ForkJoinWorkerThread w;
903	<	if ((sw = spareWaiters) != 0 &&
904	<	(id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
905	<	id < n && (w = ws[id]) != null &&
906	<	(runState >= TERMINATING ||
907	<	(workerCounts & RUNNING_COUNT_MASK) < parallelism) &&
908	<	spareWaiters == sw &&
909	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
910	<	sw, w.nextSpare)) {
911	<	int c; // increment running count before resume
912	<	do {} while (!UNSAFE.compareAndSwapInt
913	<	(this, workerCountsOffset,
914	<	c = workerCounts, c + ONE_RUNNING));
915	<	if (w.tryUnsuspend())
916	<	LockSupport.unpark(w);
856	<	else   // back out if w was shutdown
857	<	decrementWorkerCounts(ONE_RUNNING, 0);
898	>	private void growSubmissionQueue() {
899	>	ForkJoinTask<?>[] oldQ = submissionQueue;
900	>	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
901	>	if (size > MAXIMUM_QUEUE_CAPACITY)
902	>	throw new RejectedExecutionException("Queue capacity exceeded");
903	>	if (size < INITIAL_QUEUE_CAPACITY)
904	>	size = INITIAL_QUEUE_CAPACITY;
905	>	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
906	>	int mask = size - 1;
907	>	int top = queueTop;
908	>	int oldMask;
909	>	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
910	>	for (int b = queueBase; b != top; ++b) {
911	>	long u = ((b & oldMask) << ASHIFT) + ABASE;
912	>	Object x = UNSAFE.getObjectVolatile(oldQ, u);
913	>	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
914	>	UNSAFE.putObjectVolatile
915	>	(q, ((b & mask) << ASHIFT) + ABASE, x);
916	>	}
917		}
918		}
919
920	+	// Blocking support
921	+
922		/**
923	<	* Tries to increase the number of running workers if below target
924	<	* parallelism: If a spare exists tries to resume it via
925	<	* tryResumeSpare.  Otherwise, if not enough total workers or all
926	<	* existing workers are busy, adds a new worker. In all cases also
927	<	* helps wake up releasable workers waiting for work.
928	<	*/
929	<	private void helpMaintainParallelism() {
930	<	int pc = parallelism;
931	<	int wc, rs, tc;
932	<	while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
933	<	(rs = runState) < TERMINATING) {
934	<	if (spareWaiters != 0)
935	<	tryResumeSpare();
936	<	else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
937	<	(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
938	<	break;   // enough total
939	<	else if (runState == rs && workerCounts == wc &&
940	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
880	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
881	<	ForkJoinWorkerThread w = null;
882	<	Throwable fail = null;
883	<	try {
884	<	w = factory.newThread(this);
885	<	} catch (Throwable ex) {
886	<	fail = ex;
923	>	* Tries to increment blockedCount, decrement active count
924	>	* (sometimes implicitly) and possibly release or create a
925	>	* compensating worker in preparation for blocking. Fails
926	>	* on contention or termination.
927	>	*
928	>	* @return true if the caller can block, else should recheck and retry
929	>	*/
930	>	private boolean tryPreBlock() {
931	>	int b = blockedCount;
932	>	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
933	>	int pc = parallelism;
934	>	do {
935	>	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
936	>	int e, ac, tc, rc, i;
937	>	long c = ctl;
938	>	int u = (int)(c >>> 32);
939	>	if ((e = (int)c) < 0) {
940	>	// skip -- terminating
941		}
942	<	if (w == null) { // null or exceptional factory return
943	<	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
944	<	tryTerminate(false); // handle failure during shutdown
945	<	// If originating from an external caller,
946	<	// propagate exception, else ignore
947	<	if (fail != null && runState < TERMINATING &&
948	<	!(Thread.currentThread() instanceof
949	<	ForkJoinWorkerThread))
950	<	UNSAFE.throwException(fail);
951	<	break;
942	>	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
943	>	(ws = workers) != null &&
944	>	(i = ~e & SMASK) < ws.length &&
945	>	(w = ws[i]) != null) {
946	>	long nc = ((long)(w.nextWait & E_MASK) |
947	>	(c & (AC_MASK|TC_MASK)));
948	>	if (w.eventCount == e &&
949	>	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
950	>	w.eventCount = (e + EC_UNIT) & E_MASK;
951	>	if (w.parked)
952	>	UNSAFE.unpark(w);
953	>	return true;             // release an idle worker
954	>	}
955		}
956	<	w.start(recordWorker(w), ueh);
957	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc)
958	<	break; // add at most one unless total below target
959	<	}
956	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
957	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
958	>	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
959	>	return true;             // no compensation needed
960	>	}
961	>	else if (tc + pc < MAX_ID) {
962	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
963	>	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
964	>	addWorker();
965	>	return true;            // create a replacement
966	>	}
967	>	}
968	>	// try to back out on any failure and let caller retry
969	>	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
970	>	b = blockedCount, b - 1));
971		}
972	<	if (eventWaiters != 0L)
905	<	releaseEventWaiters();
972	>	return false;
973		}
974
975		/**
976	<	* Callback from the oldest waiter in awaitEvent waking up after a
910	<	* period of non-use. If all workers are idle, tries (once) to
911	<	* shutdown an event waiter or a spare, if one exists. Note that
912	<	* we don't need CAS or locks here because the method is called
913	<	* only from one thread occasionally waking (and even misfires are
914	<	* OK). Note that until the shutdown worker fully terminates,
915	<	* workerCounts will overestimate total count, which is tolerable.
916	<	*
917	<	* @param ec the event count waited on by caller (to abort
918	<	* attempt if count has since changed).
976	>	* Decrements blockedCount and increments active count
977		*/
978	<	private void tryShutdownUnusedWorker(int ec) {
979	<	if (runState == 0 && eventCount == ec) { // only trigger if all idle
980	<	ForkJoinWorkerThread[] ws = workers;
981	<	int n = ws.length;
982	<	ForkJoinWorkerThread w = null;
983	<	boolean shutdown = false;
984	<	int sw;
927	<	long h;
928	<	if ((sw = spareWaiters) != 0) { // prefer killing spares
929	<	int id = (sw & SPARE_ID_MASK) - 1;
930	<	if (id >= 0 && id < n && (w = ws[id]) != null &&
931	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
932	<	sw, w.nextSpare))
933	<	shutdown = true;
934	<	}
935	<	else if ((h = eventWaiters) != 0L) {
936	<	long nh;
937	<	int id = ((int)(h & WAITER_ID_MASK)) - 1;
938	<	if (id >= 0 && id < n && (w = ws[id]) != null &&
939	<	(nh = w.nextWaiter) != 0L && // keep at least one worker
940	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
941	<	shutdown = true;
942	<	}
943	<	if (w != null && shutdown) {
944	<	w.shutdown();
945	<	LockSupport.unpark(w);
946	<	}
947	<	}
948	<	releaseEventWaiters(); // in case of interference
978	>	private void postBlock() {
979	>	long c;
980	>	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
981	>	c = ctl, c + AC_UNIT));
982	>	int b;
983	>	do {} while(!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
984	>	b = blockedCount, b - 1));
985		}
986
987		/**
988	<	* Callback from workers invoked upon each top-level action (i.e.,
989	<	* stealing a task or taking a submission and running it).
954	<	* Performs one or more of the following:
955	<	*
956	<	* 1. If the worker is active and either did not run a task
957	<	*    or there are too many workers, try to set its active status
958	<	*    to inactive and update activeCount. On contention, we may
959	<	*    try again in this or a subsequent call.
960	<	*
961	<	* 2. If not enough total workers, help create some.
962	<	*
963	<	* 3. If there are too many running workers, suspend this worker
964	<	*    (first forcing inactive if necessary).  If it is not needed,
965	<	*    it may be shutdown while suspended (via
966	<	*    tryShutdownUnusedWorker).  Otherwise, upon resume it
967	<	*    rechecks running thread count and need for event sync.
968	<	*
969	<	* 4. If worker did not run a task, await the next task event via
970	<	*    eventSync if necessary (first forcing inactivation), upon
971	<	*    which the worker may be shutdown via
972	<	*    tryShutdownUnusedWorker.  Otherwise, help release any
973	<	*    existing event waiters that are now releasable,
988	>	* Possibly blocks waiting for the given task to complete, or
989	>	* cancels the task if terminating.  Fails to wait if contended.
990		*
991	<	* @param w the worker
976	<	* @param ran true if worker ran a task since last call to this method
991	>	* @param joinMe the task
992		*/
993	<	final void preStep(ForkJoinWorkerThread w, boolean ran) {
994	<	int wec = w.lastEventCount;
995	<	boolean active = w.active;
996	<	boolean inactivate = false;
997	<	int pc = parallelism;
998	<	while (w.runState == 0) {
999	<	int rs = runState;
985	<	if (rs >= TERMINATING) {           // propagate shutdown
986	<	w.shutdown();
987	<	break;
988	<	}
989	<	if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
990	<	UNSAFE.compareAndSwapInt(this, runStateOffset, rs, --rs)) {
991	<	inactivate = active = w.active = false;
992	<	if (rs == SHUTDOWN) {          // all inactive and shut down
993	<	tryTerminate(false);
994	<	continue;
995	<	}
996	<	}
997	<	int wc = workerCounts;             // try to suspend as spare
998	<	if ((wc & RUNNING_COUNT_MASK) > pc) {
999	<	if (!(inactivate |= active) && // must inactivate to suspend
1000	<	workerCounts == wc &&
1001	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1002	<	wc, wc - ONE_RUNNING))
1003	<	w.suspendAsSpare();
1004	<	}
1005	<	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
1006	<	helpMaintainParallelism();     // not enough workers
1007	<	else if (ran)
1008	<	break;
1009	<	else {
1010	<	long h = eventWaiters;
1011	<	int ec = eventCount;
1012	<	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
1013	<	releaseEventWaiters();     // release others before waiting
1014	<	else if (ec != wec) {
1015	<	w.lastEventCount = ec;     // no need to wait
1016	<	break;
1017	<	}
1018	<	else if (!(inactivate |= active))
1019	<	eventSync(w, wec);         // must inactivate before sync
993	>	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
994	>	int s;
995	>	Thread.interrupted(); // clear interrupts before checking termination
996	>	if (joinMe.status >= 0) {
997	>	if (tryPreBlock()) {
998	>	joinMe.tryAwaitDone(0L);
999	>	postBlock();
1000		}
1001	+	if ((ctl & STOP_BIT) != 0L)
1002	+	joinMe.cancelIgnoringExceptions();
1003		}
1004		}
1005
1006		/**
1007	<	* Helps and/or blocks awaiting join of the given task.
1008	<	* See above for explanation.
1007	>	* Possibly blocks the given worker waiting for joinMe to
1008	>	* complete or timeout
1009		*
1010	<	* @param joinMe the task to join
1011	<	* @param worker the current worker thread
1030	<	* @param timed true if wait should time out
1031	<	* @param nanos timeout value if timed
1010	>	* @param joinMe the task
1011	>	* @param millis the wait time for underlying Object.wait
1012		*/
1013	<	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker,
1034	<	boolean timed, long nanos) {
1035	<	long startTime = timed? System.nanoTime() : 0L;
1036	<	int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1037	<	boolean running = true;               // false when count decremented
1013	>	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1014		while (joinMe.status >= 0) {
1015	<	if (runState >= TERMINATING) {
1015	>	Thread.interrupted();
1016	>	if ((ctl & STOP_BIT) != 0L) {
1017		joinMe.cancelIgnoringExceptions();
1018		break;
1019		}
1020	<	running = worker.helpJoinTask(joinMe, running);
1021	<	if (joinMe.status < 0)
1022	<	break;
1023	<	if (retries > 0) {
1024	<	--retries;
1025	<	continue;
1026	<	}
1027	<	int wc = workerCounts;
1028	<	if ((wc & RUNNING_COUNT_MASK) != 0) {
1029	<	if (running) {
1030	<	if (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1054	<	wc, wc - ONE_RUNNING))
1055	<	continue;
1056	<	running = false;
1057	<	}
1058	<	long h = eventWaiters;
1059	<	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1060	<	releaseEventWaiters();
1061	<	if (joinMe.status < 0)
1062	<	break;
1063	<	if ((workerCounts & RUNNING_COUNT_MASK) != 0) {
1064	<	long ms; int ns;
1065	<	if (!timed) {
1066	<	ms = JOIN_TIMEOUT_MILLIS;
1067	<	ns = 0;
1068	<	}
1069	<	else { // at most JOIN_TIMEOUT_MILLIS per wait
1070	<	long nt = nanos - (System.nanoTime() - startTime);
1071	<	if (nt <= 0L)
1072	<	break;
1073	<	ms = nt / 1000000;
1074	<	if (ms > JOIN_TIMEOUT_MILLIS) {
1075	<	ms = JOIN_TIMEOUT_MILLIS;
1076	<	ns = 0;
1077	<	}
1078	<	else
1079	<	ns = (int) (nt % 1000000);
1080	<	}
1081	<	if (joinMe.internalAwaitDone(ms, ns) < 0)
1020	>	if (tryPreBlock()) {
1021	>	long last = System.nanoTime();
1022	>	while (joinMe.status >= 0) {
1023	>	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1024	>	if (millis <= 0)
1025	>	break;
1026	>	joinMe.tryAwaitDone(millis);
1027	>	if (joinMe.status < 0)
1028	>	break;
1029	>	if ((ctl & STOP_BIT) != 0L) {
1030	>	joinMe.cancelIgnoringExceptions();
1031		break;
1032	+	}
1033	+	long now = System.nanoTime();
1034	+	nanos -= now - last;
1035	+	last = now;
1036		}
1037	+	postBlock();
1038	+	break;
1039		}
1085	–	helpMaintainParallelism();
1086	–	}
1087	–	if (!running) {
1088	–	int c;
1089	–	do {} while (!UNSAFE.compareAndSwapInt
1090	–	(this, workerCountsOffset,
1091	–	c = workerCounts, c + ONE_RUNNING));
1040		}
1041		}
1042
1043		/**
1044	<	* Same idea as awaitJoin, but no helping, retries, or timeouts.
1044	>	* If necessary, compensates for blocker, and blocks
1045		*/
1046	<	final void awaitBlocker(ManagedBlocker blocker)
1046	>	private void awaitBlocker(ManagedBlocker blocker)
1047		throws InterruptedException {
1048		while (!blocker.isReleasable()) {
1049	<	int wc = workerCounts;
1102	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1103	<	helpMaintainParallelism();
1104	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1105	<	wc, wc - ONE_RUNNING)) {
1049	>	if (tryPreBlock()) {
1050		try {
1051	<	while (!blocker.isReleasable()) {
1108	<	long h = eventWaiters;
1109	<	if (h != 0L &&
1110	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1111	<	releaseEventWaiters();
1112	<	else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1113	<	runState < TERMINATING)
1114	<	helpMaintainParallelism();
1115	<	else if (blocker.block())
1116	<	break;
1117	<	}
1051	>	do {} while (!blocker.isReleasable() && !blocker.block());
1052		} finally {
1053	<	int c;
1120	<	do {} while (!UNSAFE.compareAndSwapInt
1121	<	(this, workerCountsOffset,
1122	<	c = workerCounts, c + ONE_RUNNING));
1053	>	postBlock();
1054		}
1055		break;
1056		}
1057		}
1058		}
1059
1060	+	// Creating, registering and deregistring workers
1061	+
1062	+	/**
1063	+	* Tries to create and start a worker; minimally rolls back counts
1064	+	* on failure.
1065	+	*/
1066	+	private void addWorker() {
1067	+	Throwable ex = null;
1068	+	ForkJoinWorkerThread t = null;
1069	+	try {
1070	+	t = factory.newThread(this);
1071	+	} catch (Throwable e) {
1072	+	ex = e;
1073	+	}
1074	+	if (t == null) {  // null or exceptional factory return
1075	+	long c;       // adjust counts
1076	+	do {} while (!UNSAFE.compareAndSwapLong
1077	+	(this, ctlOffset, c = ctl,
1078	+	(((c - AC_UNIT) & AC_MASK) |
1079	+	((c - TC_UNIT) & TC_MASK) |
1080	+	(c & ~(AC_MASK|TC_MASK)))));
1081	+	// Propagate exception if originating from an external caller
1082	+	if (!tryTerminate(false) && ex != null &&
1083	+	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1084	+	UNSAFE.throwException(ex);
1085	+	}
1086	+	else
1087	+	t.start();
1088	+	}
1089	+
1090	+	/**
1091	+	* Callback from ForkJoinWorkerThread constructor to assign a
1092	+	* public name
1093	+	*/
1094	+	final String nextWorkerName() {
1095	+	for (int n;;) {
1096	+	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
1097	+	n = nextWorkerNumber, ++n))
1098	+	return workerNamePrefix + n;
1099	+	}
1100	+	}
1101	+
1102	+	/**
1103	+	* Callback from ForkJoinWorkerThread constructor to
1104	+	* determine its poolIndex and record in workers array.
1105	+	*
1106	+	* @param w the worker
1107	+	* @return the worker's pool index
1108	+	*/
1109	+	final int registerWorker(ForkJoinWorkerThread w) {
1110	+	/*
1111	+	* In the typical case, a new worker acquires the lock, uses
1112	+	* next available index and returns quickly.  Since we should
1113	+	* not block callers (ultimately from signalWork or
1114	+	* tryPreBlock) waiting for the lock needed to do this, we
1115	+	* instead help release other workers while waiting for the
1116	+	* lock.
1117	+	*/
1118	+	for (int g;;) {
1119	+	ForkJoinWorkerThread[] ws;
1120	+	if (((g = scanGuard) & SG_UNIT) == 0 &&
1121	+	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1122	+	g, g | SG_UNIT)) {
1123	+	int k = nextWorkerIndex;
1124	+	try {
1125	+	if ((ws = workers) != null) { // ignore on shutdown
1126	+	int n = ws.length;
1127	+	if (k < 0 || k >= n || ws[k] != null) {
1128	+	for (k = 0; k < n && ws[k] != null; ++k)
1129	+	;
1130	+	if (k == n)
1131	+	ws = workers = Arrays.copyOf(ws, n << 1);
1132	+	}
1133	+	ws[k] = w;
1134	+	nextWorkerIndex = k + 1;
1135	+	int m = g & SMASK;
1136	+	g = k >= m? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
1137	+	}
1138	+	} finally {
1139	+	scanGuard = g;
1140	+	}
1141	+	return k;
1142	+	}
1143	+	else if ((ws = workers) != null) { // help release others
1144	+	for (ForkJoinWorkerThread u : ws) {
1145	+	if (u != null && u.queueBase != u.queueTop) {
1146	+	if (tryReleaseWaiter())
1147	+	break;
1148	+	}
1149	+	}
1150	+	}
1151	+	}
1152	+	}
1153	+
1154	+	/**
1155	+	* Final callback from terminating worker.  Removes record of
1156	+	* worker from array, and adjusts counts. If pool is shutting
1157	+	* down, tries to complete termination.
1158	+	*
1159	+	* @param w the worker
1160	+	*/
1161	+	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
1162	+	int idx = w.poolIndex;
1163	+	int sc = w.stealCount;
1164	+	int steps = 0;
1165	+	// Remove from array, adjust worker counts and collect steal count.
1166	+	// We can intermix failed removes or adjusts with steal updates
1167	+	do {
1168	+	long s, c;
1169	+	int g;
1170	+	if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1171	+	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1172	+	g, g |= SG_UNIT)) {
1173	+	ForkJoinWorkerThread[] ws = workers;
1174	+	if (ws != null && idx >= 0 &&
1175	+	idx < ws.length && ws[idx] == w)
1176	+	ws[idx] = null;    // verify
1177	+	nextWorkerIndex = idx;
1178	+	scanGuard = g + SG_UNIT;
1179	+	steps = 1;
1180	+	}
1181	+	if (steps == 1 &&
1182	+	UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1183	+	(((c - AC_UNIT) & AC_MASK) |
1184	+	((c - TC_UNIT) & TC_MASK) |
1185	+	(c & ~(AC_MASK|TC_MASK)))))
1186	+	steps = 2;
1187	+	if (sc != 0 &&
1188	+	UNSAFE.compareAndSwapLong(this, stealCountOffset,
1189	+	s = stealCount, s + sc))
1190	+	sc = 0;
1191	+	} while (steps != 2 || sc != 0);
1192	+	if (!tryTerminate(false)) {
1193	+	if (ex != null)   // possibly replace if died abnormally
1194	+	signalWork();
1195	+	else
1196	+	tryReleaseWaiter();
1197	+	}
1198	+	}
1199	+
1200	+	// Shutdown and termination
1201	+
1202		/**
1203		* Possibly initiates and/or completes termination.
1204		*
#	Line 1134 | Line 1207 | public class ForkJoinPool extends Abstra
1207		* @return true if now terminating or terminated
1208		*/
1209		private boolean tryTerminate(boolean now) {
1210	<	if (now)
1211	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1212	<	else if (runState < SHUTDOWN ||
1213	<	!submissionQueue.isEmpty() ||
1214	<	(runState & ACTIVE_COUNT_MASK) != 0)
1215	<	return false;
1216	<
1217	<	if (advanceRunLevel(TERMINATING))
1218	<	startTerminating();
1219	<
1220	<	// Finish now if all threads terminated; else in some subsequent call
1221	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1222	<	advanceRunLevel(TERMINATED);
1223	<	termination.forceTermination();
1210	>	long c;
1211	>	while (((c = ctl) & STOP_BIT) == 0) {
1212	>	if (!now) {
1213	>	if ((int)(c >> AC_SHIFT) != -parallelism)
1214	>	return false;
1215	>	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
1216	>	queueTop - queueBase > 0) {
1217	>	if (ctl == c) // staleness check
1218	>	return false;
1219	>	continue;
1220	>	}
1221	>	}
1222	>	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
1223	>	startTerminating();
1224	>	}
1225	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1226	>	submissionLock.lock();
1227	>	termination.signalAll();
1228	>	submissionLock.unlock();
1229		}
1230		return true;
1231		}
1232
1233		/**
1234	<	* Actions on transition to TERMINATING
1235	<	*
1236	<	* Runs up to four passes through workers: (0) shutting down each
1237	<	* (without waking up if parked) to quickly spread notifications
1238	<	* without unnecessary bouncing around event queues etc (1) wake
1239	<	* up and help cancel tasks (2) interrupt (3) mop up races with
1162	<	* interrupted workers
1234	>	* Runs up to three passes through workers: (0) Setting
1235	>	* termination status for each worker, followed by wakeups up
1236	>	* queued workers (1) helping cancel tasks (2) interrupting
1237	>	* lagging threads (likely in external tasks, but possibly also
1238	>	* blocked in joins).  Each pass repeats previous steps because of
1239	>	* potential lagging thread creation.
1240		*/
1241		private void startTerminating() {
1242		cancelSubmissions();
1243	<	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1244	<	int c; // advance event count
1245	<	UNSAFE.compareAndSwapInt(this, eventCountOffset,
1246	<	c = eventCount, c+1);
1247	<	eventWaiters = 0L; // clobber lists
1248	<	spareWaiters = 0;
1249	<	for (ForkJoinWorkerThread w : workers) {
1250	<	if (w != null) {
1251	<	w.shutdown();
1252	<	if (passes > 0 && !w.isTerminated()) {
1253	<	w.cancelTasks();
1254	<	LockSupport.unpark(w);
1255	<	if (passes > 1 && !w.isInterrupted()) {
1179	<	try {
1180	<	w.interrupt();
1181	<	} catch (SecurityException ignore) {
1243	>	for (int pass = 0; pass < 3; ++pass) {
1244	>	ForkJoinWorkerThread[] ws = workers;
1245	>	if (ws != null) {
1246	>	for (ForkJoinWorkerThread w : ws) {
1247	>	if (w != null) {
1248	>	w.terminate = true;
1249	>	if (pass > 0) {
1250	>	w.cancelTasks();
1251	>	if (pass > 1 && !w.isInterrupted()) {
1252	>	try {
1253	>	w.interrupt();
1254	>	} catch (SecurityException ignore) {
1255	>	}
1256		}
1257		}
1258		}
1259		}
1260	+	terminateWaiters();
1261		}
1262		}
1263		}
1264
1265		/**
1266	<	* Clears out and cancels submissions, ignoring exceptions.
1266	>	* Polls and cancels all submissions. Called only during termination.
1267		*/
1268		private void cancelSubmissions() {
1269	<	ForkJoinTask<?> task;
1270	<	while ((task = submissionQueue.poll()) != null) {
1271	<	try {
1272	<	task.cancel(false);
1273	<	} catch (Throwable ignore) {
1269	>	while (queueBase != queueTop) {
1270	>	ForkJoinTask<?> task = pollSubmission();
1271	>	if (task != null) {
1272	>	try {
1273	>	task.cancel(false);
1274	>	} catch (Throwable ignore) {
1275	>	}
1276		}
1277		}
1278		}
1279
1280	<	// misc support for ForkJoinWorkerThread
1280	>	/**
1281	>	* Tries to set the termination status of waiting workers, and
1282	>	* then wake them up (after which they will terminate).
1283	>	*/
1284	>	private void terminateWaiters() {
1285	>	ForkJoinWorkerThread[] ws = workers;
1286	>	if (ws != null) {
1287	>	ForkJoinWorkerThread w; long c; int i, e;
1288	>	int n = ws.length;
1289	>	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
1290	>	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
1291	>	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
1292	>	(long)(w.nextWait & E_MASK) |
1293	>	((c + AC_UNIT) & AC_MASK) |
1294	>	(c & (TC_MASK|STOP_BIT)))) {
1295	>	w.terminate = true;
1296	>	w.eventCount = e + EC_UNIT;
1297	>	if (w.parked)
1298	>	UNSAFE.unpark(w);
1299	>	}
1300	>	}
1301	>	}
1302	>	}
1303	>
1304	>	// misc ForkJoinWorkerThread support
1305
1306		/**
1307	<	* Returns pool number.
1307	>	* Increment or decrement quiescerCount. Needed only to prevent
1308	>	* triggering shutdown if a worker is transiently inactive while
1309	>	* checking quiescence.
1310	>	*
1311	>	* @param delta 1 for increment, -1 for decrement
1312		*/
1313	<	final int getPoolNumber() {
1314	<	return poolNumber;
1313	>	final void addQuiescerCount(int delta) {
1314	>	int c;
1315	>	do {} while(!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
1316	>	c = quiescerCount, c + delta));
1317		}
1318
1319		/**
1320	<	* Tries to accumulate steal count from a worker, clearing
1321	<	* the worker's value if successful.
1320	>	* Directly increment or decrement active count without
1321	>	* queuing. This method is used to transiently assert inactivation
1322	>	* while checking quiescence.
1323		*
1324	<	* @return true if worker steal count now zero
1324	>	* @param delta 1 for increment, -1 for decrement
1325		*/
1326	<	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1327	<	int sc = w.stealCount;
1328	<	long c = stealCount;
1329	<	// CAS even if zero, for fence effects
1330	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1331	<	if (sc != 0)
1224	<	w.stealCount = 0;
1225	<	return true;
1226	<	}
1227	<	return sc == 0;
1326	>	final void addActiveCount(int delta) {
1327	>	long d = delta < 0 ? -AC_UNIT : AC_UNIT;
1328	>	long c;
1329	>	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1330	>	((c + d) & AC_MASK) |
1331	>	(c & ~AC_MASK)));
1332		}
1333
1334		/**
#	Line 1232 | Line 1336 | public class ForkJoinPool extends Abstra
1336		* active thread.
1337		*/
1338		final int idlePerActive() {
1339	<	int pc = parallelism; // use parallelism, not rc
1340	<	int ac = runState;    // no mask -- artificially boosts during shutdown
1341	<	// Use exact results for small values, saturate past 4
1342	<	return ((pc <= ac) ? 0 :
1343	<	(pc >>> 1 <= ac) ? 1 :
1344	<	(pc >>> 2 <= ac) ? 3 :
1345	<	pc >>> 3);
1339	>	// Approximate at powers of two for small values, saturate past 4
1340	>	int p = parallelism;
1341	>	int a = p + (int)(ctl >> AC_SHIFT);
1342	>	return (a > (p >>>= 1) ? 0 :
1343	>	a > (p >>>= 1) ? 1 :
1344	>	a > (p >>>= 1) ? 2 :
1345	>	a > (p >>>= 1) ? 4 :
1346	>	8);
1347		}
1348
1349	<	// Public and protected methods
1349	>	// Exported methods
1350
1351		// Constructors
1352
#	Line 1310 | Line 1415 | public class ForkJoinPool extends Abstra
1415		checkPermission();
1416		if (factory == null)
1417		throw new NullPointerException();
1418	<	if (parallelism <= 0 || parallelism > MAX_WORKERS)
1418	>	if (parallelism <= 0 || parallelism > MAX_ID)
1419		throw new IllegalArgumentException();
1420		this.parallelism = parallelism;
1421		this.factory = factory;
1422		this.ueh = handler;
1423		this.locallyFifo = asyncMode;
1424	<	int arraySize = initialArraySizeFor(parallelism);
1425	<	this.workers = new ForkJoinWorkerThread[arraySize];
1426	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1427	<	this.workerLock = new ReentrantLock();
1428	<	this.termination = new Phaser(1);
1429	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
1430	<	}
1431	<
1432	<	/**
1433	<	* Returns initial power of two size for workers array.
1434	<	* @param pc the initial parallelism level
1435	<	*/
1436	<	private static int initialArraySizeFor(int pc) {
1437	<	// If possible, initially allocate enough space for one spare
1438	<	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1439	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1440	<	size |= size >>> 1;
1336	<	size |= size >>> 2;
1337	<	size |= size >>> 4;
1338	<	size |= size >>> 8;
1339	<	return size + 1;
1424	>	long np = (long)(-parallelism); // offset ctl counts
1425	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1426	>	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1427	>	// initialize workers array with room for 2*parallelism if possible
1428	>	int n = parallelism << 1;
1429	>	if (n >= MAX_ID)
1430	>	n = MAX_ID;
1431	>	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1432	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1433	>	}
1434	>	workers = new ForkJoinWorkerThread[n + 1];
1435	>	this.submissionLock = new ReentrantLock();
1436	>	this.termination = submissionLock.newCondition();
1437	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
1438	>	sb.append(poolNumberGenerator.incrementAndGet());
1439	>	sb.append("-worker-");
1440	>	this.workerNamePrefix = sb.toString();
1441		}
1442
1443		// Execution methods
1444
1445		/**
1345	–	* Submits task and creates, starts, or resumes some workers if necessary
1346	–	*/
1347	–	private <T> void doSubmit(ForkJoinTask<T> task) {
1348	–	submissionQueue.offer(task);
1349	–	int c; // try to increment event count -- CAS failure OK
1350	–	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1351	–	helpMaintainParallelism();
1352	–	}
1353	–
1354	–	/**
1446		* Performs the given task, returning its result upon completion.
1447	+	* If the computation encounters an unchecked Exception or Error,
1448	+	* it is rethrown as the outcome of this invocation.  Rethrown
1449	+	* exceptions behave in the same way as regular exceptions, but,
1450	+	* when possible, contain stack traces (as displayed for example
1451	+	* using {@code ex.printStackTrace()}) of both the current thread
1452	+	* as well as the thread actually encountering the exception;
1453	+	* minimally only the latter.
1454		*
1455		* @param task the task
1456		* @return the task's result
#	Line 1361 | Line 1459 | public class ForkJoinPool extends Abstra
1459		*         scheduled for execution
1460		*/
1461		public <T> T invoke(ForkJoinTask<T> task) {
1462	+	Thread t = Thread.currentThread();
1463		if (task == null)
1464		throw new NullPointerException();
1465	<	if (runState >= SHUTDOWN)
1465	>	if (shutdown)
1466		throw new RejectedExecutionException();
1368	–	Thread t = Thread.currentThread();
1467		if ((t instanceof ForkJoinWorkerThread) &&
1468		((ForkJoinWorkerThread)t).pool == this)
1469		return task.invoke();  // bypass submit if in same pool
1470		else {
1471	<	doSubmit(task);
1471	>	addSubmission(task);
1472		return task.join();
1473		}
1474		}
#	Line 1380 | Line 1478 | public class ForkJoinPool extends Abstra
1478		* computation in the current pool, else submits as external task.
1479		*/
1480		private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1481	<	if (runState >= SHUTDOWN)
1384	<	throw new RejectedExecutionException();
1481	>	ForkJoinWorkerThread w;
1482		Thread t = Thread.currentThread();
1483	+	if (shutdown)
1484	+	throw new RejectedExecutionException();
1485		if ((t instanceof ForkJoinWorkerThread) &&
1486	<	((ForkJoinWorkerThread)t).pool == this)
1487	<	task.fork();
1486	>	(w = (ForkJoinWorkerThread)t).pool == this)
1487	>	w.pushTask(task);
1488		else
1489	<	doSubmit(task);
1489	>	addSubmission(task);
1490		}
1491
1492		/**
#	Line 1544 | Line 1643 | public class ForkJoinPool extends Abstra
1643		* @return the number of worker threads
1644		*/
1645		public int getPoolSize() {
1646	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
1646	>	return parallelism + (short)(ctl >>> TC_SHIFT);
1647		}
1648
1649		/**
#	Line 1566 | Line 1665 | public class ForkJoinPool extends Abstra
1665		* @return the number of worker threads
1666		*/
1667		public int getRunningThreadCount() {
1668	<	return workerCounts & RUNNING_COUNT_MASK;
1668	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
1669	>	return r <= 0? 0 : r; // suppress momentarily negative values
1670		}
1671
1672		/**
#	Line 1577 | Line 1677 | public class ForkJoinPool extends Abstra
1677		* @return the number of active threads
1678		*/
1679		public int getActiveThreadCount() {
1680	<	return runState & ACTIVE_COUNT_MASK;
1680	>	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
1681	>	return r <= 0? 0 : r; // suppress momentarily negative values
1682		}
1683
1684		/**
#	Line 1592 | Line 1693 | public class ForkJoinPool extends Abstra
1693		* @return {@code true} if all threads are currently idle
1694		*/
1695		public boolean isQuiescent() {
1696	<	return (runState & ACTIVE_COUNT_MASK) == 0;
1696	>	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
1697		}
1698
1699		/**
#	Line 1622 | Line 1723 | public class ForkJoinPool extends Abstra
1723		*/
1724		public long getQueuedTaskCount() {
1725		long count = 0;
1726	<	for (ForkJoinWorkerThread w : workers)
1727	<	if (w != null)
1728	<	count += w.getQueueSize();
1726	>	ForkJoinWorkerThread[] ws;
1727	>	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1728	>	(ws = workers) != null) {
1729	>	for (ForkJoinWorkerThread w : ws)
1730	>	if (w != null)
1731	>	count -= w.queueBase - w.queueTop; // must read base first
1732	>	}
1733		return count;
1734		}
1735
1736		/**
1737		* Returns an estimate of the number of tasks submitted to this
1738	<	* pool that have not yet begun executing.  This method takes time
1739	<	* proportional to the number of submissions.
1738	>	* pool that have not yet begun executing.  This meThod may take
1739	>	* time proportional to the number of submissions.
1740		*
1741		* @return the number of queued submissions
1742		*/
1743		public int getQueuedSubmissionCount() {
1744	<	return submissionQueue.size();
1744	>	return -queueBase + queueTop;
1745		}
1746
1747		/**
#	Line 1646 | Line 1751 | public class ForkJoinPool extends Abstra
1751		* @return {@code true} if there are any queued submissions
1752		*/
1753		public boolean hasQueuedSubmissions() {
1754	<	return !submissionQueue.isEmpty();
1754	>	return queueBase != queueTop;
1755		}
1756
1757		/**
#	Line 1657 | Line 1762 | public class ForkJoinPool extends Abstra
1762		* @return the next submission, or {@code null} if none
1763		*/
1764		protected ForkJoinTask<?> pollSubmission() {
1765	<	return submissionQueue.poll();
1765	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1766	>	while ((b = queueBase) != queueTop &&
1767	>	(q = submissionQueue) != null &&
1768	>	(i = (q.length - 1) & b) >= 0) {
1769	>	long u = (i << ASHIFT) + ABASE;
1770	>	if ((t = q[i]) != null &&
1771	>	queueBase == b &&
1772	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1773	>	queueBase = b + 1;
1774	>	return t;
1775	>	}
1776	>	}
1777	>	return null;
1778		}
1779
1780		/**
#	Line 1678 | Line 1795 | public class ForkJoinPool extends Abstra
1795		* @return the number of elements transferred
1796		*/
1797		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1798	<	int count = submissionQueue.drainTo(c);
1799	<	for (ForkJoinWorkerThread w : workers)
1800	<	if (w != null)
1801	<	count += w.drainTasksTo(c);
1798	>	int count = 0;
1799	>	while (queueBase != queueTop) {
1800	>	ForkJoinTask<?> t = pollSubmission();
1801	>	if (t != null) {
1802	>	c.add(t);
1803	>	++count;
1804	>	}
1805	>	}
1806	>	ForkJoinWorkerThread[] ws;
1807	>	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1808	>	(ws = workers) != null) {
1809	>	for (ForkJoinWorkerThread w : ws)
1810	>	if (w != null)
1811	>	count += w.drainTasksTo(c);
1812	>	}
1813		return count;
1814		}
1815
#	Line 1696 | Line 1824 | public class ForkJoinPool extends Abstra
1824		long st = getStealCount();
1825		long qt = getQueuedTaskCount();
1826		long qs = getQueuedSubmissionCount();
1699	–	int wc = workerCounts;
1700	–	int tc = wc >>> TOTAL_COUNT_SHIFT;
1701	–	int rc = wc & RUNNING_COUNT_MASK;
1827		int pc = parallelism;
1828	<	int rs = runState;
1829	<	int ac = rs & ACTIVE_COUNT_MASK;
1828	>	long c = ctl;
1829	>	int tc = pc + (short)(c >>> TC_SHIFT);
1830	>	int rc = pc + (int)(c >> AC_SHIFT);
1831	>	if (rc < 0) // ignore transient negative
1832	>	rc = 0;
1833	>	int ac = rc + blockedCount;
1834	>	String level;
1835	>	if ((c & STOP_BIT) != 0)
1836	>	level = (tc == 0)? "Terminated" : "Terminating";
1837	>	else
1838	>	level = shutdown? "Shutting down" : "Running";
1839		return super.toString() +
1840	<	"[" + runLevelToString(rs) +
1840	>	"[" + level +
1841		", parallelism = " + pc +
1842		", size = " + tc +
1843		", active = " + ac +
#	Line 1714 | Line 1848 | public class ForkJoinPool extends Abstra
1848		"]";
1849		}
1850
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	–
1851		/**
1852		* Initiates an orderly shutdown in which previously submitted
1853		* tasks are executed, but no new tasks will be accepted.
#	Line 1735 | Line 1862 | public class ForkJoinPool extends Abstra
1862		*/
1863		public void shutdown() {
1864		checkPermission();
1865	<	advanceRunLevel(SHUTDOWN);
1865	>	shutdown = true;
1866		tryTerminate(false);
1867		}
1868
#	Line 1757 | Line 1884 | public class ForkJoinPool extends Abstra
1884		*/
1885		public List<Runnable> shutdownNow() {
1886		checkPermission();
1887	+	shutdown = true;
1888		tryTerminate(true);
1889		return Collections.emptyList();
1890		}
#	Line 1767 | Line 1895 | public class ForkJoinPool extends Abstra
1895		* @return {@code true} if all tasks have completed following shut down
1896		*/
1897		public boolean isTerminated() {
1898	<	return runState >= TERMINATED;
1898	>	long c = ctl;
1899	>	return ((c & STOP_BIT) != 0L &&
1900	>	(short)(c >>> TC_SHIFT) == -parallelism);
1901		}
1902
1903		/**
#	Line 1775 | Line 1905 | public class ForkJoinPool extends Abstra
1905		* commenced but not yet completed.  This method may be useful for
1906		* debugging. A return of {@code true} reported a sufficient
1907		* period after shutdown may indicate that submitted tasks have
1908	<	* ignored or suppressed interruption, causing this executor not
1909	<	* to properly terminate.
1908	>	* ignored or suppressed interruption, or are waiting for IO,
1909	>	* causing this executor not to properly terminate. (See the
1910	>	* advisory notes for class {@link ForkJoinTask} stating that
1911	>	* tasks should not normally entail blocking operations.  But if
1912	>	* they do, they must abort them on interrupt.)
1913		*
1914		* @return {@code true} if terminating but not yet terminated
1915		*/
1916		public boolean isTerminating() {
1917	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
1917	>	long c = ctl;
1918	>	return ((c & STOP_BIT) != 0L &&
1919	>	(short)(c >>> TC_SHIFT) != -parallelism);
1920		}
1921
1922		/**
1923		* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1924		*/
1925		final boolean isAtLeastTerminating() {
1926	<	return runState >= TERMINATING;
1926	>	return (ctl & STOP_BIT) != 0L;
1927		}
1928
1929		/**
#	Line 1797 | Line 1932 | public class ForkJoinPool extends Abstra
1932		* @return {@code true} if this pool has been shut down
1933		*/
1934		public boolean isShutdown() {
1935	<	return runState >= SHUTDOWN;
1935	>	return shutdown;
1936		}
1937
1938		/**
#	Line 1813 | Line 1948 | public class ForkJoinPool extends Abstra
1948		*/
1949		public boolean awaitTermination(long timeout, TimeUnit unit)
1950		throws InterruptedException {
1951	+	long nanos = unit.toNanos(timeout);
1952	+	final ReentrantLock lock = this.submissionLock;
1953	+	lock.lock();
1954		try {
1955	<	termination.awaitAdvanceInterruptibly(0, timeout, unit);
1956	<	} catch (TimeoutException ex) {
1957	<	return false;
1955	>	for (;;) {
1956	>	if (isTerminated())
1957	>	return true;
1958	>	if (nanos <= 0)
1959	>	return false;
1960	>	nanos = termination.awaitNanos(nanos);
1961	>	}
1962	>	} finally {
1963	>	lock.unlock();
1964		}
1821	–	return true;
1965		}
1966
1967		/**
#	Line 1829 | Line 1972 | public class ForkJoinPool extends Abstra
1972		* {@code isReleasable} must return {@code true} if blocking is
1973		* not necessary. Method {@code block} blocks the current thread
1974		* if necessary (perhaps internally invoking {@code isReleasable}
1975	<	* before actually blocking). The unusual methods in this API
1976	<	* accommodate synchronizers that may, but don't usually, block
1977	<	* for long periods. Similarly, they allow more efficient internal
1978	<	* handling of cases in which additional workers may be, but
1979	<	* usually are not, needed to ensure sufficient parallelism.
1980	<	* Toward this end, implementations of method {@code isReleasable}
1981	<	* must be amenable to repeated invocation.
1975	>	* before actually blocking). These actions are performed by any
1976	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
1977	>	* unusual methods in this API accommodate synchronizers that may,
1978	>	* but don't usually, block for long periods. Similarly, they
1979	>	* allow more efficient internal handling of cases in which
1980	>	* additional workers may be, but usually are not, needed to
1981	>	* ensure sufficient parallelism.  Toward this end,
1982	>	* implementations of method {@code isReleasable} must be amenable
1983	>	* to repeated invocation.
1984		*
1985		* <p>For example, here is a ManagedBlocker based on a
1986		* ReentrantLock:
#	Line 1937 | Line 2082 | public class ForkJoinPool extends Abstra
2082		}
2083
2084		// Unsafe mechanics
2085	<
2086	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
2087	<	private static final long workerCountsOffset =
2088	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
2089	<	private static final long runStateOffset =
2090	<	objectFieldOffset("runState", ForkJoinPool.class);
2091	<	private static final long eventCountOffset =
2092	<	objectFieldOffset("eventCount", ForkJoinPool.class);
2093	<	private static final long eventWaitersOffset =
2094	<	objectFieldOffset("eventWaiters", ForkJoinPool.class);
2095	<	private static final long stealCountOffset =
2096	<	objectFieldOffset("stealCount", ForkJoinPool.class);
2097	<	private static final long spareWaitersOffset =
2098	<	objectFieldOffset("spareWaiters", ForkJoinPool.class);
2099	<
2100	<	private static long objectFieldOffset(String field, Class<?> klazz) {
2085	>	private static final sun.misc.Unsafe UNSAFE;
2086	>	private static final long ctlOffset;
2087	>	private static final long stealCountOffset;
2088	>	private static final long blockedCountOffset;
2089	>	private static final long quiescerCountOffset;
2090	>	private static final long scanGuardOffset;
2091	>	private static final long nextWorkerNumberOffset;
2092	>	private static final long ABASE;
2093	>	private static final int ASHIFT;
2094	>
2095	>	static {
2096	>	poolNumberGenerator = new AtomicInteger();
2097	>	workerSeedGenerator = new Random();
2098	>	modifyThreadPermission = new RuntimePermission("modifyThread");
2099	>	defaultForkJoinWorkerThreadFactory =
2100	>	new DefaultForkJoinWorkerThreadFactory();
2101	>	int s;
2102		try {
2103	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
2104	<	} catch (NoSuchFieldException e) {
2105	<	// Convert Exception to corresponding Error
2106	<	NoSuchFieldError error = new NoSuchFieldError(field);
2107	<	error.initCause(e);
2108	<	throw error;
2109	<	}
2103	>	UNSAFE = getUnsafe();
2104	>	Class k = ForkJoinPool.class;
2105	>	ctlOffset = UNSAFE.objectFieldOffset
2106	>	(k.getDeclaredField("ctl"));
2107	>	stealCountOffset = UNSAFE.objectFieldOffset
2108	>	(k.getDeclaredField("stealCount"));
2109	>	blockedCountOffset = UNSAFE.objectFieldOffset
2110	>	(k.getDeclaredField("blockedCount"));
2111	>	quiescerCountOffset = UNSAFE.objectFieldOffset
2112	>	(k.getDeclaredField("quiescerCount"));
2113	>	scanGuardOffset = UNSAFE.objectFieldOffset
2114	>	(k.getDeclaredField("scanGuard"));
2115	>	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
2116	>	(k.getDeclaredField("nextWorkerNumber"));
2117	>	Class a = ForkJoinTask[].class;
2118	>	ABASE = UNSAFE.arrayBaseOffset(a);
2119	>	s = UNSAFE.arrayIndexScale(a);
2120	>	} catch (Exception e) {
2121	>	throw new Error(e);
2122	>	}
2123	>	if ((s & (s-1)) != 0)
2124	>	throw new Error("data type scale not a power of two");
2125	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
2126		}
2127
2128		/**

Diff Legend

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