1 |
|
/* |
2 |
|
* Written by Doug Lea with assistance from members of JCP JSR-166 |
3 |
|
* Expert Group and released to the public domain, as explained at |
4 |
< |
* http://creativecommons.org/licenses/publicdomain |
4 |
> |
* http://creativecommons.org/publicdomain/zero/1.0/ |
5 |
|
*/ |
6 |
|
|
7 |
|
package jsr166y; |
8 |
|
|
9 |
– |
import java.util.concurrent.*; |
10 |
– |
|
9 |
|
import java.util.ArrayList; |
10 |
|
import java.util.Arrays; |
11 |
|
import java.util.Collection; |
12 |
|
import java.util.Collections; |
13 |
|
import java.util.List; |
14 |
+ |
import java.util.Random; |
15 |
+ |
import java.util.concurrent.AbstractExecutorService; |
16 |
+ |
import java.util.concurrent.Callable; |
17 |
+ |
import java.util.concurrent.ExecutorService; |
18 |
+ |
import java.util.concurrent.Future; |
19 |
+ |
import java.util.concurrent.RejectedExecutionException; |
20 |
+ |
import java.util.concurrent.RunnableFuture; |
21 |
+ |
import java.util.concurrent.TimeUnit; |
22 |
+ |
import java.util.concurrent.atomic.AtomicInteger; |
23 |
|
import java.util.concurrent.locks.LockSupport; |
24 |
|
import java.util.concurrent.locks.ReentrantLock; |
25 |
< |
import java.util.concurrent.atomic.AtomicInteger; |
19 |
< |
import java.util.concurrent.CountDownLatch; |
25 |
> |
import java.util.concurrent.locks.Condition; |
26 |
|
|
27 |
|
/** |
28 |
|
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
75 |
|
* <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td> |
76 |
|
* </tr> |
77 |
|
* <tr> |
78 |
< |
* <td> <b>Arange async execution</td> |
78 |
> |
* <td> <b>Arrange async execution</td> |
79 |
|
* <td> {@link #execute(ForkJoinTask)}</td> |
80 |
|
* <td> {@link ForkJoinTask#fork}</td> |
81 |
|
* </tr> |
101 |
|
* daemon} mode, there is typically no need to explicitly {@link |
102 |
|
* #shutdown} such a pool upon program exit. |
103 |
|
* |
104 |
< |
* <pre> |
104 |
> |
* <pre> {@code |
105 |
|
* static final ForkJoinPool mainPool = new ForkJoinPool(); |
106 |
|
* ... |
107 |
|
* public void sort(long[] array) { |
108 |
|
* mainPool.invoke(new SortTask(array, 0, array.length)); |
109 |
< |
* } |
104 |
< |
* </pre> |
109 |
> |
* }}</pre> |
110 |
|
* |
111 |
|
* <p><b>Implementation notes</b>: This implementation restricts the |
112 |
|
* maximum number of running threads to 32767. Attempts to create |
129 |
|
* set of worker threads: Submissions from non-FJ threads enter |
130 |
|
* into a submission queue. Workers take these tasks and typically |
131 |
|
* split them into subtasks that may be stolen by other workers. |
132 |
< |
* The main work-stealing mechanics implemented in class |
133 |
< |
* ForkJoinWorkerThread give first priority to processing tasks |
134 |
< |
* from their own queues (LIFO or FIFO, depending on mode), then |
135 |
< |
* to randomized FIFO steals of tasks in other worker queues, and |
136 |
< |
* lastly to new submissions. These mechanics do not consider |
137 |
< |
* affinities, loads, cache localities, etc, so rarely provide the |
138 |
< |
* best possible performance on a given machine, but portably |
139 |
< |
* provide good throughput by averaging over these factors. |
140 |
< |
* (Further, even if we did try to use such information, we do not |
141 |
< |
* usually have a basis for exploiting it. For example, some sets |
142 |
< |
* of tasks profit from cache affinities, but others are harmed by |
143 |
< |
* cache pollution effects.) |
144 |
< |
* |
145 |
< |
* Beyond work-stealing support and essential bookkeeping, the |
146 |
< |
* main responsibility of this framework is to take actions when |
147 |
< |
* one worker is waiting to join a task stolen (or always held by) |
148 |
< |
* another. Becauae we are multiplexing many tasks on to a pool |
149 |
< |
* of workers, we can't just let them block (as in Thread.join). |
150 |
< |
* We also cannot just reassign the joiner's run-time stack with |
151 |
< |
* another and replace it later, which would be a form of |
152 |
< |
* "continuation", that even if possible is not necessarily a good |
153 |
< |
* idea. Given that the creation costs of most threads on most |
154 |
< |
* systems mainly surrounds setting up runtime stacks, thread |
155 |
< |
* creation and switching is usually not much more expensive than |
156 |
< |
* stack creation and switching, and is more flexible). Instead we |
132 |
> |
* Preference rules give first priority to processing tasks from |
133 |
> |
* their own queues (LIFO or FIFO, depending on mode), then to |
134 |
> |
* randomized FIFO steals of tasks in other worker queues, and |
135 |
> |
* lastly to new submissions. |
136 |
> |
* |
137 |
> |
* The main throughput advantages of work-stealing stem from |
138 |
> |
* decentralized control -- workers mostly take tasks from |
139 |
> |
* themselves or each other. We cannot negate this in the |
140 |
> |
* implementation of other management responsibilities. The main |
141 |
> |
* tactic for avoiding bottlenecks is packing nearly all |
142 |
> |
* essentially atomic control state into a single 64bit volatile |
143 |
> |
* variable ("ctl"). This variable is read on the order of 10-100 |
144 |
> |
* times as often as it is modified (always via CAS). (There is |
145 |
> |
* some additional control state, for example variable "shutdown" |
146 |
> |
* for which we can cope with uncoordinated updates.) This |
147 |
> |
* streamlines synchronization and control at the expense of messy |
148 |
> |
* constructions needed to repack status bits upon updates. |
149 |
> |
* Updates tend not to contend with each other except during |
150 |
> |
* bursts while submitted tasks begin or end. In some cases when |
151 |
> |
* they do contend, threads can instead do something else |
152 |
> |
* (usually, scan for tasks) until contention subsides. |
153 |
> |
* |
154 |
> |
* To enable packing, we restrict maximum parallelism to (1<<15)-1 |
155 |
> |
* (which is far in excess of normal operating range) to allow |
156 |
> |
* ids, counts, and their negations (used for thresholding) to fit |
157 |
> |
* into 16bit fields. |
158 |
> |
* |
159 |
> |
* Recording Workers. Workers are recorded in the "workers" array |
160 |
> |
* that is created upon pool construction and expanded if (rarely) |
161 |
> |
* necessary. This is an array as opposed to some other data |
162 |
> |
* structure to support index-based random steals by workers. |
163 |
> |
* Updates to the array recording new workers and unrecording |
164 |
> |
* terminated ones are protected from each other by a seqLock |
165 |
> |
* (scanGuard) but the array is otherwise concurrently readable, |
166 |
> |
* and accessed directly by workers. To simplify index-based |
167 |
> |
* operations, the array size is always a power of two, and all |
168 |
> |
* readers must tolerate null slots. To avoid flailing during |
169 |
> |
* start-up, the array is presized to hold twice #parallelism |
170 |
> |
* workers (which is unlikely to need further resizing during |
171 |
> |
* execution). But to avoid dealing with so many null slots, |
172 |
> |
* variable scanGuard includes a mask for the nearest power of two |
173 |
> |
* that contains all current workers. All worker thread creation |
174 |
> |
* is on-demand, triggered by task submissions, replacement of |
175 |
> |
* terminated workers, and/or compensation for blocked |
176 |
> |
* workers. However, all other support code is set up to work with |
177 |
> |
* other policies. To ensure that we do not hold on to worker |
178 |
> |
* references that would prevent GC, ALL accesses to workers are |
179 |
> |
* via indices into the workers array (which is one source of some |
180 |
> |
* of the messy code constructions here). In essence, the workers |
181 |
> |
* array serves as a weak reference mechanism. Thus for example |
182 |
> |
* the wait queue field of ctl stores worker indices, not worker |
183 |
> |
* references. Access to the workers in associated methods (for |
184 |
> |
* example signalWork) must both index-check and null-check the |
185 |
> |
* IDs. All such accesses ignore bad IDs by returning out early |
186 |
> |
* from what they are doing, since this can only be associated |
187 |
> |
* with termination, in which case it is OK to give up. |
188 |
> |
* |
189 |
> |
* All uses of the workers array, as well as queue arrays, check |
190 |
> |
* that the array is non-null (even if previously non-null). This |
191 |
> |
* allows nulling during termination, which is currently not |
192 |
> |
* necessary, but remains an option for resource-revocation-based |
193 |
> |
* shutdown schemes. |
194 |
> |
* |
195 |
> |
* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot |
196 |
> |
* let workers spin indefinitely scanning for tasks when none can |
197 |
> |
* be found immediately, and we cannot start/resume workers unless |
198 |
> |
* there appear to be tasks available. On the other hand, we must |
199 |
> |
* quickly prod them into action when new tasks are submitted or |
200 |
> |
* generated. We park/unpark workers after placing in an event |
201 |
> |
* wait queue when they cannot find work. This "queue" is actually |
202 |
> |
* a simple Treiber stack, headed by the "id" field of ctl, plus a |
203 |
> |
* 15bit counter value to both wake up waiters (by advancing their |
204 |
> |
* count) and avoid ABA effects. Successors are held in worker |
205 |
> |
* field "nextWait". Queuing deals with several intrinsic races, |
206 |
> |
* mainly that a task-producing thread can miss seeing (and |
207 |
> |
* signalling) another thread that gave up looking for work but |
208 |
> |
* has not yet entered the wait queue. We solve this by requiring |
209 |
> |
* a full sweep of all workers both before (in scan()) and after |
210 |
> |
* (in tryAwaitWork()) a newly waiting worker is added to the wait |
211 |
> |
* queue. During a rescan, the worker might release some other |
212 |
> |
* queued worker rather than itself, which has the same net |
213 |
> |
* effect. Because enqueued workers may actually be rescanning |
214 |
> |
* rather than waiting, we set and clear the "parked" field of |
215 |
> |
* ForkJoinWorkerThread to reduce unnecessary calls to unpark. |
216 |
> |
* (Use of the parked field requires a secondary recheck to avoid |
217 |
> |
* missed signals.) |
218 |
> |
* |
219 |
> |
* Signalling. We create or wake up workers only when there |
220 |
> |
* appears to be at least one task they might be able to find and |
221 |
> |
* execute. When a submission is added or another worker adds a |
222 |
> |
* task to a queue that previously had two or fewer tasks, they |
223 |
> |
* signal waiting workers (or trigger creation of new ones if |
224 |
> |
* fewer than the given parallelism level -- see signalWork). |
225 |
> |
* These primary signals are buttressed by signals during rescans |
226 |
> |
* as well as those performed when a worker steals a task and |
227 |
> |
* notices that there are more tasks too; together these cover the |
228 |
> |
* signals needed in cases when more than two tasks are pushed |
229 |
> |
* but untaken. |
230 |
> |
* |
231 |
> |
* Trimming workers. To release resources after periods of lack of |
232 |
> |
* use, a worker starting to wait when the pool is quiescent will |
233 |
> |
* time out and terminate if the pool has remained quiescent for |
234 |
> |
* SHRINK_RATE nanosecs. This will slowly propagate, eventually |
235 |
> |
* terminating all workers after long periods of non-use. |
236 |
> |
* |
237 |
> |
* Submissions. External submissions are maintained in an |
238 |
> |
* array-based queue that is structured identically to |
239 |
> |
* ForkJoinWorkerThread queues except for the use of |
240 |
> |
* submissionLock in method addSubmission. Unlike the case for |
241 |
> |
* worker queues, multiple external threads can add new |
242 |
> |
* submissions, so adding requires a lock. |
243 |
> |
* |
244 |
> |
* Compensation. Beyond work-stealing support and lifecycle |
245 |
> |
* control, the main responsibility of this framework is to take |
246 |
> |
* actions when one worker is waiting to join a task stolen (or |
247 |
> |
* always held by) another. Because we are multiplexing many |
248 |
> |
* tasks on to a pool of workers, we can't just let them block (as |
249 |
> |
* in Thread.join). We also cannot just reassign the joiner's |
250 |
> |
* run-time stack with another and replace it later, which would |
251 |
> |
* be a form of "continuation", that even if possible is not |
252 |
> |
* necessarily a good idea since we sometimes need both an |
253 |
> |
* unblocked task and its continuation to progress. Instead we |
254 |
|
* combine two tactics: |
255 |
|
* |
256 |
|
* Helping: Arranging for the joiner to execute some task that it |
257 |
|
* would be running if the steal had not occurred. Method |
258 |
< |
* ForkJoinWorkerThread.helpJoinTask tracks joining->stealing |
258 |
> |
* ForkJoinWorkerThread.joinTask tracks joining->stealing |
259 |
|
* links to try to find such a task. |
260 |
|
* |
261 |
|
* Compensating: Unless there are already enough live threads, |
262 |
< |
* method helpMaintainParallelism() may create or or |
263 |
< |
* re-activate a spare thread to compensate for blocked |
264 |
< |
* joiners until they unblock. |
163 |
< |
* |
164 |
< |
* Because the determining existence of conservatively safe |
165 |
< |
* helping targets, the availability of already-created spares, |
166 |
< |
* and the apparent need to create new spares are all racy and |
167 |
< |
* require heuristic guidance, we rely on multiple retries of |
168 |
< |
* each. Further, because it is impossible to keep exactly the |
169 |
< |
* target (parallelism) number of threads running at any given |
170 |
< |
* time, we allow compensation during joins to fail, and enlist |
171 |
< |
* all other threads to help out whenever they are not otherwise |
172 |
< |
* occupied (i.e., mainly in method preStep). |
262 |
> |
* method tryPreBlock() may create or re-activate a spare |
263 |
> |
* thread to compensate for blocked joiners until they |
264 |
> |
* unblock. |
265 |
|
* |
266 |
|
* The ManagedBlocker extension API can't use helping so relies |
267 |
|
* only on compensation in method awaitBlocker. |
268 |
|
* |
269 |
< |
* The main throughput advantages of work-stealing stem from |
270 |
< |
* decentralized control -- workers mostly steal tasks from each |
271 |
< |
* other. We do not want to negate this by creating bottlenecks |
272 |
< |
* implementing other management responsibilities. So we use a |
273 |
< |
* collection of techniques that avoid, reduce, or cope well with |
274 |
< |
* contention. These entail several instances of bit-packing into |
275 |
< |
* CASable fields to maintain only the minimally required |
276 |
< |
* atomicity. To enable such packing, we restrict maximum |
277 |
< |
* parallelism to (1<<15)-1 (enabling twice this (to accommodate |
278 |
< |
* unbalanced increments and decrements) to fit into a 16 bit |
279 |
< |
* field, which is far in excess of normal operating range. Even |
280 |
< |
* though updates to some of these bookkeeping fields do sometimes |
281 |
< |
* contend with each other, they don't normally cache-contend with |
282 |
< |
* updates to others enough to warrant memory padding or |
283 |
< |
* isolation. So they are all held as fields of ForkJoinPool |
284 |
< |
* objects. The main capabilities are as follows: |
269 |
> |
* It is impossible to keep exactly the target parallelism number |
270 |
> |
* of threads running at any given time. Determining the |
271 |
> |
* existence of conservatively safe helping targets, the |
272 |
> |
* availability of already-created spares, and the apparent need |
273 |
> |
* to create new spares are all racy and require heuristic |
274 |
> |
* guidance, so we rely on multiple retries of each. Currently, |
275 |
> |
* in keeping with on-demand signalling policy, we compensate only |
276 |
> |
* if blocking would leave less than one active (non-waiting, |
277 |
> |
* non-blocked) worker. Additionally, to avoid some false alarms |
278 |
> |
* due to GC, lagging counters, system activity, etc, compensated |
279 |
> |
* blocking for joins is only attempted after rechecks stabilize |
280 |
> |
* (retries are interspersed with Thread.yield, for good |
281 |
> |
* citizenship). The variable blockedCount, incremented before |
282 |
> |
* blocking and decremented after, is sometimes needed to |
283 |
> |
* distinguish cases of waiting for work vs blocking on joins or |
284 |
> |
* other managed sync. Both cases are equivalent for most pool |
285 |
> |
* control, so we can update non-atomically. (Additionally, |
286 |
> |
* contention on blockedCount alleviates some contention on ctl). |
287 |
> |
* |
288 |
> |
* Shutdown and Termination. A call to shutdownNow atomically sets |
289 |
> |
* the ctl stop bit and then (non-atomically) sets each workers |
290 |
> |
* "terminate" status, cancels all unprocessed tasks, and wakes up |
291 |
> |
* all waiting workers. Detecting whether termination should |
292 |
> |
* commence after a non-abrupt shutdown() call requires more work |
293 |
> |
* and bookkeeping. We need consensus about quiescence (i.e., that |
294 |
> |
* there is no more work) which is reflected in active counts so |
295 |
> |
* long as there are no current blockers, as well as possible |
296 |
> |
* re-evaluations during independent changes in blocking or |
297 |
> |
* quiescing workers. |
298 |
|
* |
299 |
< |
* 1. Creating and removing workers. Workers are recorded in the |
195 |
< |
* "workers" array. This is an array as opposed to some other data |
196 |
< |
* structure to support index-based random steals by workers. |
197 |
< |
* Updates to the array recording new workers and unrecording |
198 |
< |
* terminated ones are protected from each other by a lock |
199 |
< |
* (workerLock) but the array is otherwise concurrently readable, |
200 |
< |
* and accessed directly by workers. To simplify index-based |
201 |
< |
* operations, the array size is always a power of two, and all |
202 |
< |
* readers must tolerate null slots. Currently, all worker thread |
203 |
< |
* creation is on-demand, triggered by task submissions, |
204 |
< |
* replacement of terminated workers, and/or compensation for |
205 |
< |
* blocked workers. However, all other support code is set up to |
206 |
< |
* work with other policies. |
207 |
< |
* |
208 |
< |
* To ensure that we do not hold on to worker references that |
209 |
< |
* would prevent GC, ALL accesses to workers are via indices into |
210 |
< |
* the workers array (which is one source of some of the unusual |
211 |
< |
* code constructions here). In essence, the workers array serves |
212 |
< |
* as a WeakReference mechanism. Thus for example the event queue |
213 |
< |
* stores worker indices, not worker references. Access to the |
214 |
< |
* workers in associated methods (for example releaseEventWaiters) |
215 |
< |
* must both index-check and null-check the IDs. All such accesses |
216 |
< |
* ignore bad IDs by returning out early from what they are doing, |
217 |
< |
* since this can only be associated with shutdown, in which case |
218 |
< |
* it is OK to give up. On termination, we just clobber these |
219 |
< |
* data structures without trying to use them. |
220 |
< |
* |
221 |
< |
* 2. Bookkeeping for dynamically adding and removing workers. We |
222 |
< |
* aim to approximately maintain the given level of parallelism. |
223 |
< |
* When some workers are known to be blocked (on joins or via |
224 |
< |
* ManagedBlocker), we may create or resume others to take their |
225 |
< |
* place until they unblock (see below). Implementing this |
226 |
< |
* requires counts of the number of "running" threads (i.e., those |
227 |
< |
* that are neither blocked nor artifically suspended) as well as |
228 |
< |
* the total number. These two values are packed into one field, |
229 |
< |
* "workerCounts" because we need accurate snapshots when deciding |
230 |
< |
* to create, resume or suspend. Note however that the |
231 |
< |
* correspondance of these counts to reality is not guaranteed. In |
232 |
< |
* particular updates for unblocked threads may lag until they |
233 |
< |
* actually wake up. |
234 |
< |
* |
235 |
< |
* 3. Maintaining global run state. The run state of the pool |
236 |
< |
* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to |
237 |
< |
* those in other Executor implementations, as well as a count of |
238 |
< |
* "active" workers -- those that are, or soon will be, or |
239 |
< |
* recently were executing tasks. The runLevel and active count |
240 |
< |
* are packed together in order to correctly trigger shutdown and |
241 |
< |
* termination. Without care, active counts can be subject to very |
242 |
< |
* high contention. We substantially reduce this contention by |
243 |
< |
* relaxing update rules. A worker must claim active status |
244 |
< |
* prospectively, by activating if it sees that a submitted or |
245 |
< |
* stealable task exists (it may find after activating that the |
246 |
< |
* task no longer exists). It stays active while processing this |
247 |
< |
* task (if it exists) and any other local subtasks it produces, |
248 |
< |
* until it cannot find any other tasks. It then tries |
249 |
< |
* inactivating (see method preStep), but upon update contention |
250 |
< |
* instead scans for more tasks, later retrying inactivation if it |
251 |
< |
* doesn't find any. |
252 |
< |
* |
253 |
< |
* 4. Managing idle workers waiting for tasks. We cannot let |
254 |
< |
* workers spin indefinitely scanning for tasks when none are |
255 |
< |
* available. On the other hand, we must quickly prod them into |
256 |
< |
* action when new tasks are submitted or generated. We |
257 |
< |
* park/unpark these idle workers using an event-count scheme. |
258 |
< |
* Field eventCount is incremented upon events that may enable |
259 |
< |
* workers that previously could not find a task to now find one: |
260 |
< |
* Submission of a new task to the pool, or another worker pushing |
261 |
< |
* a task onto a previously empty queue. (We also use this |
262 |
< |
* mechanism for termination actions that require wakeups of idle |
263 |
< |
* workers). Each worker maintains its last known event count, |
264 |
< |
* and blocks when a scan for work did not find a task AND its |
265 |
< |
* lastEventCount matches the current eventCount. Waiting idle |
266 |
< |
* workers are recorded in a variant of Treiber stack headed by |
267 |
< |
* field eventWaiters which, when nonzero, encodes the thread |
268 |
< |
* index and count awaited for by the worker thread most recently |
269 |
< |
* calling eventSync. This thread in turn has a record (field |
270 |
< |
* nextEventWaiter) for the next waiting worker. In addition to |
271 |
< |
* allowing simpler decisions about need for wakeup, the event |
272 |
< |
* count bits in eventWaiters serve the role of tags to avoid ABA |
273 |
< |
* errors in Treiber stacks. To reduce delays in task diffusion, |
274 |
< |
* workers not otherwise occupied may invoke method |
275 |
< |
* releaseEventWaiters, that removes and signals (unparks) workers |
276 |
< |
* not waiting on current count. To reduce stalls, To minimize |
277 |
< |
* task production stalls associate with signalling, any worker |
278 |
< |
* pushing a task on an empty queue invokes the weaker method |
279 |
< |
* signalWork, that only releases idle workers until it detects |
280 |
< |
* interference by other threads trying to release, and lets them |
281 |
< |
* take over. The net effect is a tree-like diffusion of signals, |
282 |
< |
* where released threads (and possibly others) help with unparks. |
283 |
< |
* To further reduce contention effects a bit, failed CASes to |
284 |
< |
* increment field eventCount are tolerated without retries. |
285 |
< |
* Conceptually they are merged into the same event, which is OK |
286 |
< |
* when their only purpose is to enable workers to scan for work. |
287 |
< |
* |
288 |
< |
* 5. Managing suspension of extra workers. When a worker is about |
289 |
< |
* to block waiting for a join (or via ManagedBlockers), we may |
290 |
< |
* create a new thread to maintain parallelism level, or at least |
291 |
< |
* avoid starvation. Usually, extra threads are needed for only |
292 |
< |
* very short periods, yet join dependencies are such that we |
293 |
< |
* sometimes need them in bursts. Rather than create new threads |
294 |
< |
* each time this happens, we suspend no-longer-needed extra ones |
295 |
< |
* as "spares". For most purposes, we don't distinguish "extra" |
296 |
< |
* spare threads from normal "core" threads: On each call to |
297 |
< |
* preStep (the only point at which we can do this) a worker |
298 |
< |
* checks to see if there are now too many running workers, and if |
299 |
< |
* so, suspends itself. Method helpMaintainParallelism looks for |
300 |
< |
* suspended threads to resume before considering creating a new |
301 |
< |
* replacement. The spares themselves are encoded on another |
302 |
< |
* variant of a Treiber Stack, headed at field "spareWaiters". |
303 |
< |
* Note that the use of spares is intrinsically racy. One thread |
304 |
< |
* may become a spare at about the same time as another is |
305 |
< |
* needlessly being created. We counteract this and related slop |
306 |
< |
* in part by requiring resumed spares to immediately recheck (in |
307 |
< |
* preStep) to see whether they they should re-suspend. To avoid |
308 |
< |
* long-term build-up of spares, the oldest spare (see |
309 |
< |
* ForkJoinWorkerThread.suspendAsSpare) occasionally wakes up if |
310 |
< |
* not signalled and calls tryTrimSpare, which uses two different |
311 |
< |
* thresholds: Always killing if the number of spares is greater |
312 |
< |
* that 25% of total, and killing others only at a slower rate |
313 |
< |
* (UNUSED_SPARE_TRIM_RATE_NANOS). |
314 |
< |
* |
315 |
< |
* 6. Deciding when to create new workers. The main dynamic |
316 |
< |
* control in this class is deciding when to create extra threads |
317 |
< |
* in method helpMaintainParallelism. We would like to keep |
318 |
< |
* exactly #parallelism threads running, which is an impossble |
319 |
< |
* task. We always need to create one when the number of running |
320 |
< |
* threads would become zero and all workers are busy. Beyond |
321 |
< |
* this, we must rely on heuristics that work well in the the |
322 |
< |
* presence of transients phenomena such as GC stalls, dynamic |
323 |
< |
* compilation, and wake-up lags. These transients are extremely |
324 |
< |
* common -- we are normally trying to fully saturate the CPUs on |
325 |
< |
* a machine, so almost any activity other than running tasks |
326 |
< |
* impedes accuracy. Our main defense is to allow some slack in |
327 |
< |
* creation thresholds, using rules that reflect the fact that the |
328 |
< |
* more threads we have running, the more likely that we are |
329 |
< |
* underestimating the number running threads. The rules also |
330 |
< |
* better cope with the fact that some of the methods in this |
331 |
< |
* class tend to never become compiled (but are interpreted), so |
332 |
< |
* some components of the entire set of controls might execute 100 |
333 |
< |
* times faster than others. And similarly for cases where the |
334 |
< |
* apparent lack of work is just due to GC stalls and other |
335 |
< |
* transient system activity. |
336 |
< |
* |
337 |
< |
* Beware that there is a lot of representation-level coupling |
299 |
> |
* Style notes: There is a lot of representation-level coupling |
300 |
|
* among classes ForkJoinPool, ForkJoinWorkerThread, and |
301 |
< |
* ForkJoinTask. For example, direct access to "workers" array by |
301 |
> |
* ForkJoinTask. Most fields of ForkJoinWorkerThread maintain |
302 |
> |
* data structures managed by ForkJoinPool, so are directly |
303 |
> |
* accessed. Conversely we allow access to "workers" array by |
304 |
|
* workers, and direct access to ForkJoinTask.status by both |
305 |
|
* ForkJoinPool and ForkJoinWorkerThread. There is little point |
306 |
|
* trying to reduce this, since any associated future changes in |
307 |
|
* representations will need to be accompanied by algorithmic |
308 |
< |
* changes anyway. |
309 |
< |
* |
310 |
< |
* Style notes: There are lots of inline assignments (of form |
311 |
< |
* "while ((local = field) != 0)") which are usually the simplest |
312 |
< |
* way to ensure the required read orderings (which are sometimes |
313 |
< |
* critical). Also several occurrences of the unusual "do {} |
314 |
< |
* while(!cas...)" which is the simplest way to force an update of |
315 |
< |
* a CAS'ed variable. There are also other coding oddities that |
316 |
< |
* help some methods perform reasonably even when interpreted (not |
317 |
< |
* compiled), at the expense of some messy constructions that |
318 |
< |
* reduce byte code counts. |
319 |
< |
* |
320 |
< |
* The order of declarations in this file is: (1) statics (2) |
321 |
< |
* fields (along with constants used when unpacking some of them) |
322 |
< |
* (3) internal control methods (4) callbacks and other support |
323 |
< |
* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported |
324 |
< |
* methods (plus a few little helpers). |
308 |
> |
* changes anyway. All together, these low-level implementation |
309 |
> |
* choices produce as much as a factor of 4 performance |
310 |
> |
* improvement compared to naive implementations, and enable the |
311 |
> |
* processing of billions of tasks per second, at the expense of |
312 |
> |
* some ugliness. |
313 |
> |
* |
314 |
> |
* Methods signalWork() and scan() are the main bottlenecks so are |
315 |
> |
* especially heavily micro-optimized/mangled. There are lots of |
316 |
> |
* inline assignments (of form "while ((local = field) != 0)") |
317 |
> |
* which are usually the simplest way to ensure the required read |
318 |
> |
* orderings (which are sometimes critical). This leads to a |
319 |
> |
* "C"-like style of listing declarations of these locals at the |
320 |
> |
* heads of methods or blocks. There are several occurrences of |
321 |
> |
* the unusual "do {} while (!cas...)" which is the simplest way |
322 |
> |
* to force an update of a CAS'ed variable. There are also other |
323 |
> |
* coding oddities that help some methods perform reasonably even |
324 |
> |
* when interpreted (not compiled). |
325 |
> |
* |
326 |
> |
* The order of declarations in this file is: (1) declarations of |
327 |
> |
* statics (2) fields (along with constants used when unpacking |
328 |
> |
* some of them), listed in an order that tends to reduce |
329 |
> |
* contention among them a bit under most JVMs. (3) internal |
330 |
> |
* control methods (4) callbacks and other support for |
331 |
> |
* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported |
332 |
> |
* methods (plus a few little helpers). (6) static block |
333 |
> |
* initializing all statics in a minimally dependent order. |
334 |
|
*/ |
335 |
|
|
336 |
|
/** |
365 |
|
* overridden in ForkJoinPool constructors. |
366 |
|
*/ |
367 |
|
public static final ForkJoinWorkerThreadFactory |
368 |
< |
defaultForkJoinWorkerThreadFactory = |
396 |
< |
new DefaultForkJoinWorkerThreadFactory(); |
368 |
> |
defaultForkJoinWorkerThreadFactory; |
369 |
|
|
370 |
|
/** |
371 |
|
* Permission required for callers of methods that may start or |
372 |
|
* kill threads. |
373 |
|
*/ |
374 |
< |
private static final RuntimePermission modifyThreadPermission = |
403 |
< |
new RuntimePermission("modifyThread"); |
374 |
> |
private static final RuntimePermission modifyThreadPermission; |
375 |
|
|
376 |
|
/** |
377 |
|
* If there is a security manager, makes sure caller has |
386 |
|
/** |
387 |
|
* Generator for assigning sequence numbers as pool names. |
388 |
|
*/ |
389 |
< |
private static final AtomicInteger poolNumberGenerator = |
419 |
< |
new AtomicInteger(); |
389 |
> |
private static final AtomicInteger poolNumberGenerator; |
390 |
|
|
391 |
|
/** |
392 |
< |
* Absolute bound for parallelism level. Twice this number plus |
393 |
< |
* one (i.e., 0xfff) must fit into a 16bit field to enable |
394 |
< |
* word-packing for some counts and indices. |
392 |
> |
* Generator for initial random seeds for worker victim |
393 |
> |
* selection. This is used only to create initial seeds. Random |
394 |
> |
* steals use a cheaper xorshift generator per steal attempt. We |
395 |
> |
* don't expect much contention on seedGenerator, so just use a |
396 |
> |
* plain Random. |
397 |
|
*/ |
398 |
< |
private static final int MAX_WORKERS = 0x7fff; |
398 |
> |
static final Random workerSeedGenerator; |
399 |
|
|
400 |
|
/** |
401 |
< |
* Array holding all worker threads in the pool. Array size must |
402 |
< |
* be a power of two. Updates and replacements are protected by |
403 |
< |
* workerLock, but the array is always kept in a consistent enough |
404 |
< |
* state to be randomly accessed without locking by workers |
405 |
< |
* performing work-stealing, as well as other traversal-based |
406 |
< |
* methods in this class. All readers must tolerate that some |
407 |
< |
* array slots may be null. |
401 |
> |
* Array holding all worker threads in the pool. Initialized upon |
402 |
> |
* construction. Array size must be a power of two. Updates and |
403 |
> |
* replacements are protected by scanGuard, but the array is |
404 |
> |
* always kept in a consistent enough state to be randomly |
405 |
> |
* accessed without locking by workers performing work-stealing, |
406 |
> |
* as well as other traversal-based methods in this class, so long |
407 |
> |
* as reads memory-acquire by first reading ctl. All readers must |
408 |
> |
* tolerate that some array slots may be null. |
409 |
|
*/ |
410 |
< |
volatile ForkJoinWorkerThread[] workers; |
410 |
> |
ForkJoinWorkerThread[] workers; |
411 |
|
|
412 |
|
/** |
413 |
< |
* Queue for external submissions. |
413 |
> |
* Initial size for submission queue array. Must be a power of |
414 |
> |
* two. In many applications, these always stay small so we use a |
415 |
> |
* small initial cap. |
416 |
|
*/ |
417 |
< |
private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue; |
417 |
> |
private static final int INITIAL_QUEUE_CAPACITY = 8; |
418 |
|
|
419 |
|
/** |
420 |
< |
* Lock protecting updates to workers array. |
420 |
> |
* Maximum size for submission queue array. Must be a power of two |
421 |
> |
* less than or equal to 1 << (31 - width of array entry) to |
422 |
> |
* ensure lack of index wraparound, but is capped at a lower |
423 |
> |
* value to help users trap runaway computations. |
424 |
|
*/ |
425 |
< |
private final ReentrantLock workerLock; |
425 |
> |
private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M |
426 |
|
|
427 |
|
/** |
428 |
< |
* Latch released upon termination. |
428 |
> |
* Array serving as submission queue. Initialized upon construction. |
429 |
|
*/ |
430 |
< |
private final Phaser termination; |
430 |
> |
private ForkJoinTask<?>[] submissionQueue; |
431 |
|
|
432 |
|
/** |
433 |
< |
* Creation factory for worker threads. |
433 |
> |
* Lock protecting submissions array for addSubmission |
434 |
|
*/ |
435 |
< |
private final ForkJoinWorkerThreadFactory factory; |
435 |
> |
private final ReentrantLock submissionLock; |
436 |
|
|
437 |
|
/** |
438 |
< |
* Sum of per-thread steal counts, updated only when threads are |
439 |
< |
* idle or terminating. |
438 |
> |
* Condition for awaitTermination, using submissionLock for |
439 |
> |
* convenience. |
440 |
|
*/ |
441 |
< |
private volatile long stealCount; |
441 |
> |
private final Condition termination; |
442 |
|
|
443 |
|
/** |
444 |
< |
* The last nanoTime that a spare thread was trimmed |
444 |
> |
* Creation factory for worker threads. |
445 |
|
*/ |
446 |
< |
private volatile long trimTime; |
446 |
> |
private final ForkJoinWorkerThreadFactory factory; |
447 |
|
|
448 |
|
/** |
449 |
< |
* The rate at which to trim unused spares |
449 |
> |
* The uncaught exception handler used when any worker abruptly |
450 |
> |
* terminates. |
451 |
|
*/ |
452 |
< |
static final long UNUSED_SPARE_TRIM_RATE_NANOS = |
474 |
< |
1000L * 1000L * 1000L; // 1 sec |
452 |
> |
final Thread.UncaughtExceptionHandler ueh; |
453 |
|
|
454 |
|
/** |
455 |
< |
* Encoded record of top of treiber stack of threads waiting for |
478 |
< |
* events. The top 32 bits contain the count being waited for. The |
479 |
< |
* bottom 16 bits contains one plus the pool index of waiting |
480 |
< |
* worker thread. (Bits 16-31 are unused.) |
455 |
> |
* Prefix for assigning names to worker threads |
456 |
|
*/ |
457 |
< |
private volatile long eventWaiters; |
483 |
< |
|
484 |
< |
private static final int EVENT_COUNT_SHIFT = 32; |
485 |
< |
private static final long WAITER_ID_MASK = (1L << 16) - 1L; |
457 |
> |
private final String workerNamePrefix; |
458 |
|
|
459 |
|
/** |
460 |
< |
* A counter for events that may wake up worker threads: |
461 |
< |
* - Submission of a new task to the pool |
490 |
< |
* - A worker pushing a task on an empty queue |
491 |
< |
* - termination |
460 |
> |
* Sum of per-thread steal counts, updated only when threads are |
461 |
> |
* idle or terminating. |
462 |
|
*/ |
463 |
< |
private volatile int eventCount; |
463 |
> |
private volatile long stealCount; |
464 |
|
|
465 |
|
/** |
466 |
< |
* Encoded record of top of treiber stack of spare threads waiting |
467 |
< |
* for resumption. The top 16 bits contain an arbitrary count to |
468 |
< |
* avoid ABA effects. The bottom 16bits contains one plus the pool |
469 |
< |
* index of waiting worker thread. |
470 |
< |
*/ |
471 |
< |
private volatile int spareWaiters; |
472 |
< |
|
473 |
< |
private static final int SPARE_COUNT_SHIFT = 16; |
474 |
< |
private static final int SPARE_ID_MASK = (1 << 16) - 1; |
466 |
> |
* Main pool control -- a long packed with: |
467 |
> |
* AC: Number of active running workers minus target parallelism (16 bits) |
468 |
> |
* TC: Number of total workers minus target parallelism (16 bits) |
469 |
> |
* ST: true if pool is terminating (1 bit) |
470 |
> |
* EC: the wait count of top waiting thread (15 bits) |
471 |
> |
* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits) |
472 |
> |
* |
473 |
> |
* When convenient, we can extract the upper 32 bits of counts and |
474 |
> |
* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e = |
475 |
> |
* (int)ctl. The ec field is never accessed alone, but always |
476 |
> |
* together with id and st. The offsets of counts by the target |
477 |
> |
* parallelism and the positionings of fields makes it possible to |
478 |
> |
* perform the most common checks via sign tests of fields: When |
479 |
> |
* ac is negative, there are not enough active workers, when tc is |
480 |
> |
* negative, there are not enough total workers, when id is |
481 |
> |
* negative, there is at least one waiting worker, and when e is |
482 |
> |
* negative, the pool is terminating. To deal with these possibly |
483 |
> |
* negative fields, we use casts in and out of "short" and/or |
484 |
> |
* signed shifts to maintain signedness. |
485 |
> |
*/ |
486 |
> |
volatile long ctl; |
487 |
> |
|
488 |
> |
// bit positions/shifts for fields |
489 |
> |
private static final int AC_SHIFT = 48; |
490 |
> |
private static final int TC_SHIFT = 32; |
491 |
> |
private static final int ST_SHIFT = 31; |
492 |
> |
private static final int EC_SHIFT = 16; |
493 |
> |
|
494 |
> |
// bounds |
495 |
> |
private static final int MAX_ID = 0x7fff; // max poolIndex |
496 |
> |
private static final int SMASK = 0xffff; // mask short bits |
497 |
> |
private static final int SHORT_SIGN = 1 << 15; |
498 |
> |
private static final int INT_SIGN = 1 << 31; |
499 |
> |
|
500 |
> |
// masks |
501 |
> |
private static final long STOP_BIT = 0x0001L << ST_SHIFT; |
502 |
> |
private static final long AC_MASK = ((long)SMASK) << AC_SHIFT; |
503 |
> |
private static final long TC_MASK = ((long)SMASK) << TC_SHIFT; |
504 |
> |
|
505 |
> |
// units for incrementing and decrementing |
506 |
> |
private static final long TC_UNIT = 1L << TC_SHIFT; |
507 |
> |
private static final long AC_UNIT = 1L << AC_SHIFT; |
508 |
> |
|
509 |
> |
// masks and units for dealing with u = (int)(ctl >>> 32) |
510 |
> |
private static final int UAC_SHIFT = AC_SHIFT - 32; |
511 |
> |
private static final int UTC_SHIFT = TC_SHIFT - 32; |
512 |
> |
private static final int UAC_MASK = SMASK << UAC_SHIFT; |
513 |
> |
private static final int UTC_MASK = SMASK << UTC_SHIFT; |
514 |
> |
private static final int UAC_UNIT = 1 << UAC_SHIFT; |
515 |
> |
private static final int UTC_UNIT = 1 << UTC_SHIFT; |
516 |
> |
|
517 |
> |
// masks and units for dealing with e = (int)ctl |
518 |
> |
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
519 |
> |
private static final int EC_UNIT = 1 << EC_SHIFT; |
520 |
|
|
521 |
|
/** |
522 |
< |
* Lifecycle control. The low word contains the number of workers |
508 |
< |
* that are (probably) executing tasks. This value is atomically |
509 |
< |
* incremented before a worker gets a task to run, and decremented |
510 |
< |
* when worker has no tasks and cannot find any. Bits 16-18 |
511 |
< |
* contain runLevel value. When all are zero, the pool is |
512 |
< |
* running. Level transitions are monotonic (running -> shutdown |
513 |
< |
* -> terminating -> terminated) so each transition adds a bit. |
514 |
< |
* These are bundled together to ensure consistent read for |
515 |
< |
* termination checks (i.e., that runLevel is at least SHUTDOWN |
516 |
< |
* and active threads is zero). |
522 |
> |
* The target parallelism level. |
523 |
|
*/ |
524 |
< |
private volatile int runState; |
519 |
< |
|
520 |
< |
// Note: The order among run level values matters. |
521 |
< |
private static final int RUNLEVEL_SHIFT = 16; |
522 |
< |
private static final int SHUTDOWN = 1 << RUNLEVEL_SHIFT; |
523 |
< |
private static final int TERMINATING = 1 << (RUNLEVEL_SHIFT + 1); |
524 |
< |
private static final int TERMINATED = 1 << (RUNLEVEL_SHIFT + 2); |
525 |
< |
private static final int ACTIVE_COUNT_MASK = (1 << RUNLEVEL_SHIFT) - 1; |
526 |
< |
private static final int ONE_ACTIVE = 1; // active update delta |
524 |
> |
final int parallelism; |
525 |
|
|
526 |
|
/** |
527 |
< |
* Holds number of total (i.e., created and not yet terminated) |
528 |
< |
* and running (i.e., not blocked on joins or other managed sync) |
529 |
< |
* threads, packed together to ensure consistent snapshot when |
530 |
< |
* making decisions about creating and suspending spare |
533 |
< |
* threads. Updated only by CAS. Note that adding a new worker |
534 |
< |
* requires incrementing both counts, since workers start off in |
535 |
< |
* running state. |
527 |
> |
* Index (mod submission queue length) of next element to take |
528 |
> |
* from submission queue. Usage is identical to that for |
529 |
> |
* per-worker queues -- see ForkJoinWorkerThread internal |
530 |
> |
* documentation. |
531 |
|
*/ |
532 |
< |
private volatile int workerCounts; |
532 |
> |
volatile int queueBase; |
533 |
|
|
534 |
< |
private static final int TOTAL_COUNT_SHIFT = 16; |
535 |
< |
private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1; |
536 |
< |
private static final int ONE_RUNNING = 1; |
537 |
< |
private static final int ONE_TOTAL = 1 << TOTAL_COUNT_SHIFT; |
534 |
> |
/** |
535 |
> |
* Index (mod submission queue length) of next element to add |
536 |
> |
* in submission queue. Usage is identical to that for |
537 |
> |
* per-worker queues -- see ForkJoinWorkerThread internal |
538 |
> |
* documentation. |
539 |
> |
*/ |
540 |
> |
int queueTop; |
541 |
|
|
542 |
|
/** |
543 |
< |
* The target parallelism level. |
546 |
< |
* Accessed directly by ForkJoinWorkerThreads. |
543 |
> |
* True when shutdown() has been called. |
544 |
|
*/ |
545 |
< |
final int parallelism; |
545 |
> |
volatile boolean shutdown; |
546 |
|
|
547 |
|
/** |
548 |
< |
* True if use local fifo, not default lifo, for local polling |
549 |
< |
* Read by, and replicated by ForkJoinWorkerThreads |
548 |
> |
* True if use local fifo, not default lifo, for local polling. |
549 |
> |
* Read by, and replicated by ForkJoinWorkerThreads. |
550 |
|
*/ |
551 |
|
final boolean locallyFifo; |
552 |
|
|
553 |
|
/** |
554 |
< |
* The uncaught exception handler used when any worker abruptly |
555 |
< |
* terminates. |
554 |
> |
* The number of threads in ForkJoinWorkerThreads.helpQuiescePool. |
555 |
> |
* When non-zero, suppresses automatic shutdown when active |
556 |
> |
* counts become zero. |
557 |
|
*/ |
558 |
< |
private final Thread.UncaughtExceptionHandler ueh; |
558 |
> |
volatile int quiescerCount; |
559 |
|
|
560 |
|
/** |
561 |
< |
* Pool number, just for assigning useful names to worker threads |
561 |
> |
* The number of threads blocked in join. |
562 |
|
*/ |
563 |
< |
private final int poolNumber; |
563 |
> |
volatile int blockedCount; |
564 |
|
|
565 |
+ |
/** |
566 |
+ |
* Counter for worker Thread names (unrelated to their poolIndex) |
567 |
+ |
*/ |
568 |
+ |
private volatile int nextWorkerNumber; |
569 |
|
|
570 |
< |
// Utilities for CASing fields. Note that several of these |
571 |
< |
// are manually inlined by callers |
570 |
> |
/** |
571 |
> |
* The index for the next created worker. Accessed under scanGuard. |
572 |
> |
*/ |
573 |
> |
private int nextWorkerIndex; |
574 |
|
|
575 |
|
/** |
576 |
< |
* Increments running count part of workerCounts |
576 |
> |
* SeqLock and index masking for updates to workers array. Locked |
577 |
> |
* when SG_UNIT is set. Unlocking clears bit by adding |
578 |
> |
* SG_UNIT. Staleness of read-only operations can be checked by |
579 |
> |
* comparing scanGuard to value before the reads. The low 16 bits |
580 |
> |
* (i.e, anding with SMASK) hold (the smallest power of two |
581 |
> |
* covering all worker indices, minus one, and is used to avoid |
582 |
> |
* dealing with large numbers of null slots when the workers array |
583 |
> |
* is overallocated. |
584 |
|
*/ |
585 |
< |
final void incrementRunningCount() { |
586 |
< |
int c; |
587 |
< |
do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
577 |
< |
c = workerCounts, |
578 |
< |
c + ONE_RUNNING)); |
579 |
< |
} |
585 |
> |
volatile int scanGuard; |
586 |
> |
|
587 |
> |
private static final int SG_UNIT = 1 << 16; |
588 |
|
|
589 |
|
/** |
590 |
< |
* Tries to decrement running count unless already zero |
590 |
> |
* The wakeup interval (in nanoseconds) for a worker waiting for a |
591 |
> |
* task when the pool is quiescent to instead try to shrink the |
592 |
> |
* number of workers. The exact value does not matter too |
593 |
> |
* much. It must be short enough to release resources during |
594 |
> |
* sustained periods of idleness, but not so short that threads |
595 |
> |
* are continually re-created. |
596 |
|
*/ |
597 |
< |
final boolean tryDecrementRunningCount() { |
598 |
< |
int wc = workerCounts; |
586 |
< |
if ((wc & RUNNING_COUNT_MASK) == 0) |
587 |
< |
return false; |
588 |
< |
return UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
589 |
< |
wc, wc - ONE_RUNNING); |
590 |
< |
} |
597 |
> |
private static final long SHRINK_RATE = |
598 |
> |
4L * 1000L * 1000L * 1000L; // 4 seconds |
599 |
|
|
600 |
|
/** |
601 |
< |
* Forces decrement of encoded workerCounts, awaiting nonzero if |
602 |
< |
* (rarely) necessary when other count updates lag. |
601 |
> |
* Top-level loop for worker threads: On each step: if the |
602 |
> |
* previous step swept through all queues and found no tasks, or |
603 |
> |
* there are excess threads, then possibly blocks. Otherwise, |
604 |
> |
* scans for and, if found, executes a task. Returns when pool |
605 |
> |
* and/or worker terminate. |
606 |
|
* |
607 |
< |
* @param dr -- either zero or ONE_RUNNING |
597 |
< |
* @param dt == either zero or ONE_TOTAL |
607 |
> |
* @param w the worker |
608 |
|
*/ |
609 |
< |
private void decrementWorkerCounts(int dr, int dt) { |
610 |
< |
for (;;) { |
611 |
< |
int wc = workerCounts; |
612 |
< |
if (wc == 0 && (runState & TERMINATED) != 0) |
613 |
< |
return; // lagging termination on a backout |
614 |
< |
if ((wc & RUNNING_COUNT_MASK) - dr < 0 || |
615 |
< |
(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) |
616 |
< |
Thread.yield(); |
617 |
< |
if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
608 |
< |
wc, wc - (dr + dt))) |
609 |
< |
return; |
609 |
> |
final void work(ForkJoinWorkerThread w) { |
610 |
> |
boolean swept = false; // true on empty scans |
611 |
> |
long c; |
612 |
> |
while (!w.terminate && (int)(c = ctl) >= 0) { |
613 |
> |
int a; // active count |
614 |
> |
if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0) |
615 |
> |
swept = scan(w, a); |
616 |
> |
else if (tryAwaitWork(w, c)) |
617 |
> |
swept = false; |
618 |
|
} |
619 |
|
} |
620 |
|
|
621 |
+ |
// Signalling |
622 |
+ |
|
623 |
|
/** |
624 |
< |
* Increments event count |
624 |
> |
* Wakes up or creates a worker. |
625 |
|
*/ |
626 |
< |
private void advanceEventCount() { |
627 |
< |
int c; |
628 |
< |
do {} while(!UNSAFE.compareAndSwapInt(this, eventCountOffset, |
629 |
< |
c = eventCount, c+1)); |
626 |
> |
final void signalWork() { |
627 |
> |
/* |
628 |
> |
* The while condition is true if: (there is are too few total |
629 |
> |
* workers OR there is at least one waiter) AND (there are too |
630 |
> |
* few active workers OR the pool is terminating). The value |
631 |
> |
* of e distinguishes the remaining cases: zero (no waiters) |
632 |
> |
* for create, negative if terminating (in which case do |
633 |
> |
* nothing), else release a waiter. The secondary checks for |
634 |
> |
* release (non-null array etc) can fail if the pool begins |
635 |
> |
* terminating after the test, and don't impose any added cost |
636 |
> |
* because JVMs must perform null and bounds checks anyway. |
637 |
> |
*/ |
638 |
> |
long c; int e, u; |
639 |
> |
while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) & |
640 |
> |
(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) { |
641 |
> |
if (e > 0) { // release a waiting worker |
642 |
> |
int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws; |
643 |
> |
if ((ws = workers) == null || |
644 |
> |
(i = ~e & SMASK) >= ws.length || |
645 |
> |
(w = ws[i]) == null) |
646 |
> |
break; |
647 |
> |
long nc = (((long)(w.nextWait & E_MASK)) | |
648 |
> |
((long)(u + UAC_UNIT) << 32)); |
649 |
> |
if (w.eventCount == e && |
650 |
> |
UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { |
651 |
> |
w.eventCount = (e + EC_UNIT) & E_MASK; |
652 |
> |
if (w.parked) |
653 |
> |
UNSAFE.unpark(w); |
654 |
> |
break; |
655 |
> |
} |
656 |
> |
} |
657 |
> |
else if (UNSAFE.compareAndSwapLong |
658 |
> |
(this, ctlOffset, c, |
659 |
> |
(long)(((u + UTC_UNIT) & UTC_MASK) | |
660 |
> |
((u + UAC_UNIT) & UAC_MASK)) << 32)) { |
661 |
> |
addWorker(); |
662 |
> |
break; |
663 |
> |
} |
664 |
> |
} |
665 |
|
} |
666 |
|
|
667 |
|
/** |
668 |
< |
* Tries incrementing active count; fails on contention. |
669 |
< |
* Called by workers before executing tasks. |
668 |
> |
* Variant of signalWork to help release waiters on rescans. |
669 |
> |
* Tries once to release a waiter if active count < 0. |
670 |
|
* |
671 |
< |
* @return true on success |
671 |
> |
* @return false if failed due to contention, else true |
672 |
|
*/ |
673 |
< |
final boolean tryIncrementActiveCount() { |
674 |
< |
int c; |
675 |
< |
return UNSAFE.compareAndSwapInt(this, runStateOffset, |
676 |
< |
c = runState, c + ONE_ACTIVE); |
673 |
> |
private boolean tryReleaseWaiter() { |
674 |
> |
long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws; |
675 |
> |
if ((e = (int)(c = ctl)) > 0 && |
676 |
> |
(int)(c >> AC_SHIFT) < 0 && |
677 |
> |
(ws = workers) != null && |
678 |
> |
(i = ~e & SMASK) < ws.length && |
679 |
> |
(w = ws[i]) != null) { |
680 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
681 |
> |
((c + AC_UNIT) & (AC_MASK|TC_MASK))); |
682 |
> |
if (w.eventCount != e || |
683 |
> |
!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) |
684 |
> |
return false; |
685 |
> |
w.eventCount = (e + EC_UNIT) & E_MASK; |
686 |
> |
if (w.parked) |
687 |
> |
UNSAFE.unpark(w); |
688 |
> |
} |
689 |
> |
return true; |
690 |
|
} |
691 |
|
|
692 |
+ |
// Scanning for tasks |
693 |
+ |
|
694 |
|
/** |
695 |
< |
* Tries decrementing active count; fails on contention. |
696 |
< |
* Called when workers cannot find tasks to run. |
695 |
> |
* Scans for and, if found, executes one task. Scans start at a |
696 |
> |
* random index of workers array, and randomly select the first |
697 |
> |
* (2*#workers)-1 probes, and then, if all empty, resort to 2 |
698 |
> |
* circular sweeps, which is necessary to check quiescence. and |
699 |
> |
* taking a submission only if no stealable tasks were found. The |
700 |
> |
* steal code inside the loop is a specialized form of |
701 |
> |
* ForkJoinWorkerThread.deqTask, followed bookkeeping to support |
702 |
> |
* helpJoinTask and signal propagation. The code for submission |
703 |
> |
* queues is almost identical. On each steal, the worker completes |
704 |
> |
* not only the task, but also all local tasks that this task may |
705 |
> |
* have generated. On detecting staleness or contention when |
706 |
> |
* trying to take a task, this method returns without finishing |
707 |
> |
* sweep, which allows global state rechecks before retry. |
708 |
> |
* |
709 |
> |
* @param w the worker |
710 |
> |
* @param a the number of active workers |
711 |
> |
* @return true if swept all queues without finding a task |
712 |
|
*/ |
713 |
< |
final boolean tryDecrementActiveCount() { |
714 |
< |
int c; |
715 |
< |
return UNSAFE.compareAndSwapInt(this, runStateOffset, |
716 |
< |
c = runState, c - ONE_ACTIVE); |
713 |
> |
private boolean scan(ForkJoinWorkerThread w, int a) { |
714 |
> |
int g = scanGuard; // mask 0 avoids useless scans if only one active |
715 |
> |
int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK; |
716 |
> |
ForkJoinWorkerThread[] ws = workers; |
717 |
> |
if (ws == null || ws.length <= m) // staleness check |
718 |
> |
return false; |
719 |
> |
for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) { |
720 |
> |
ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i; |
721 |
> |
ForkJoinWorkerThread v = ws[k & m]; |
722 |
> |
if (v != null && (b = v.queueBase) != v.queueTop && |
723 |
> |
(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) { |
724 |
> |
long u = (i << ASHIFT) + ABASE; |
725 |
> |
if ((t = q[i]) != null && v.queueBase == b && |
726 |
> |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
727 |
> |
int d = (v.queueBase = b + 1) - v.queueTop; |
728 |
> |
v.stealHint = w.poolIndex; |
729 |
> |
if (d != 0) |
730 |
> |
signalWork(); // propagate if nonempty |
731 |
> |
w.execTask(t); |
732 |
> |
} |
733 |
> |
r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5); |
734 |
> |
return false; // store next seed |
735 |
> |
} |
736 |
> |
else if (j < 0) { // xorshift |
737 |
> |
r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5; |
738 |
> |
} |
739 |
> |
else |
740 |
> |
++k; |
741 |
> |
} |
742 |
> |
if (scanGuard != g) // staleness check |
743 |
> |
return false; |
744 |
> |
else { // try to take submission |
745 |
> |
ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i; |
746 |
> |
if ((b = queueBase) != queueTop && |
747 |
> |
(q = submissionQueue) != null && |
748 |
> |
(i = (q.length - 1) & b) >= 0) { |
749 |
> |
long u = (i << ASHIFT) + ABASE; |
750 |
> |
if ((t = q[i]) != null && queueBase == b && |
751 |
> |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
752 |
> |
queueBase = b + 1; |
753 |
> |
w.execTask(t); |
754 |
> |
} |
755 |
> |
return false; |
756 |
> |
} |
757 |
> |
return true; // all queues empty |
758 |
> |
} |
759 |
|
} |
760 |
|
|
761 |
|
/** |
762 |
< |
* Advances to at least the given level. Returns true if not |
763 |
< |
* already in at least the given level. |
764 |
< |
*/ |
765 |
< |
private boolean advanceRunLevel(int level) { |
766 |
< |
for (;;) { |
767 |
< |
int s = runState; |
768 |
< |
if ((s & level) != 0) |
769 |
< |
return false; |
770 |
< |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level)) |
762 |
> |
* Tries to enqueue worker w in wait queue and await change in |
763 |
> |
* worker's eventCount. If the pool is quiescent and there is |
764 |
> |
* more than one worker, possibly terminates worker upon exit. |
765 |
> |
* Otherwise, before blocking, rescans queues to avoid missed |
766 |
> |
* signals. Upon finding work, releases at least one worker |
767 |
> |
* (which may be the current worker). Rescans restart upon |
768 |
> |
* detected staleness or failure to release due to |
769 |
> |
* contention. Note the unusual conventions about Thread.interrupt |
770 |
> |
* here and elsewhere: Because interrupts are used solely to alert |
771 |
> |
* threads to check termination, which is checked here anyway, we |
772 |
> |
* clear status (using Thread.interrupted) before any call to |
773 |
> |
* park, so that park does not immediately return due to status |
774 |
> |
* being set via some other unrelated call to interrupt in user |
775 |
> |
* code. |
776 |
> |
* |
777 |
> |
* @param w the calling worker |
778 |
> |
* @param c the ctl value on entry |
779 |
> |
* @return true if waited or another thread was released upon enq |
780 |
> |
*/ |
781 |
> |
private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) { |
782 |
> |
int v = w.eventCount; |
783 |
> |
w.nextWait = (int)c; // w's successor record |
784 |
> |
long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); |
785 |
> |
if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { |
786 |
> |
long d = ctl; // return true if lost to a deq, to force scan |
787 |
> |
return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L; |
788 |
> |
} |
789 |
> |
for (int sc = w.stealCount; sc != 0;) { // accumulate stealCount |
790 |
> |
long s = stealCount; |
791 |
> |
if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc)) |
792 |
> |
sc = w.stealCount = 0; |
793 |
> |
else if (w.eventCount != v) |
794 |
> |
return true; // update next time |
795 |
> |
} |
796 |
> |
if ((!shutdown || !tryTerminate(false)) && |
797 |
> |
(int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 && |
798 |
> |
blockedCount == 0 && quiescerCount == 0) |
799 |
> |
idleAwaitWork(w, nc, c, v); // quiescent |
800 |
> |
for (boolean rescanned = false;;) { |
801 |
> |
if (w.eventCount != v) |
802 |
|
return true; |
803 |
+ |
if (!rescanned) { |
804 |
+ |
int g = scanGuard, m = g & SMASK; |
805 |
+ |
ForkJoinWorkerThread[] ws = workers; |
806 |
+ |
if (ws != null && m < ws.length) { |
807 |
+ |
rescanned = true; |
808 |
+ |
for (int i = 0; i <= m; ++i) { |
809 |
+ |
ForkJoinWorkerThread u = ws[i]; |
810 |
+ |
if (u != null) { |
811 |
+ |
if (u.queueBase != u.queueTop && |
812 |
+ |
!tryReleaseWaiter()) |
813 |
+ |
rescanned = false; // contended |
814 |
+ |
if (w.eventCount != v) |
815 |
+ |
return true; |
816 |
+ |
} |
817 |
+ |
} |
818 |
+ |
} |
819 |
+ |
if (scanGuard != g || // stale |
820 |
+ |
(queueBase != queueTop && !tryReleaseWaiter())) |
821 |
+ |
rescanned = false; |
822 |
+ |
if (!rescanned) |
823 |
+ |
Thread.yield(); // reduce contention |
824 |
+ |
else |
825 |
+ |
Thread.interrupted(); // clear before park |
826 |
+ |
} |
827 |
+ |
else { |
828 |
+ |
w.parked = true; // must recheck |
829 |
+ |
if (w.eventCount != v) { |
830 |
+ |
w.parked = false; |
831 |
+ |
return true; |
832 |
+ |
} |
833 |
+ |
LockSupport.park(this); |
834 |
+ |
rescanned = w.parked = false; |
835 |
+ |
} |
836 |
|
} |
837 |
|
} |
838 |
|
|
658 |
– |
// workers array maintenance |
659 |
– |
|
839 |
|
/** |
840 |
< |
* Records and returns a workers array index for new worker. |
841 |
< |
*/ |
842 |
< |
private int recordWorker(ForkJoinWorkerThread w) { |
843 |
< |
// Try using slot totalCount-1. If not available, scan and/or resize |
844 |
< |
int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1; |
845 |
< |
final ReentrantLock lock = this.workerLock; |
846 |
< |
lock.lock(); |
847 |
< |
try { |
848 |
< |
ForkJoinWorkerThread[] ws = workers; |
849 |
< |
int n = ws.length; |
850 |
< |
if (k < 0 || k >= n || ws[k] != null) { |
851 |
< |
for (k = 0; k < n && ws[k] != null; ++k) |
852 |
< |
; |
853 |
< |
if (k == n) |
854 |
< |
ws = Arrays.copyOf(ws, n << 1); |
840 |
> |
* If inactivating worker w has caused pool to become |
841 |
> |
* quiescent, check for pool termination, and wait for event |
842 |
> |
* for up to SHRINK_RATE nanosecs (rescans are unnecessary in |
843 |
> |
* this case because quiescence reflects consensus about lack |
844 |
> |
* of work). On timeout, if ctl has not changed, terminate the |
845 |
> |
* worker. Upon its termination (see deregisterWorker), it may |
846 |
> |
* wake up another worker to possibly repeat this process. |
847 |
> |
* |
848 |
> |
* @param w the calling worker |
849 |
> |
* @param currentCtl the ctl value after enqueuing w |
850 |
> |
* @param prevCtl the ctl value if w terminated |
851 |
> |
* @param v the eventCount w awaits change |
852 |
> |
*/ |
853 |
> |
private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl, |
854 |
> |
long prevCtl, int v) { |
855 |
> |
if (w.eventCount == v) { |
856 |
> |
if (shutdown) |
857 |
> |
tryTerminate(false); |
858 |
> |
ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs |
859 |
> |
while (ctl == currentCtl) { |
860 |
> |
long startTime = System.nanoTime(); |
861 |
> |
w.parked = true; |
862 |
> |
if (w.eventCount == v) // must recheck |
863 |
> |
LockSupport.parkNanos(this, SHRINK_RATE); |
864 |
> |
w.parked = false; |
865 |
> |
if (w.eventCount != v) |
866 |
> |
break; |
867 |
> |
else if (System.nanoTime() - startTime < |
868 |
> |
SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop |
869 |
> |
Thread.interrupted(); // spurious wakeup |
870 |
> |
else if (UNSAFE.compareAndSwapLong(this, ctlOffset, |
871 |
> |
currentCtl, prevCtl)) { |
872 |
> |
w.terminate = true; // restore previous |
873 |
> |
w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK; |
874 |
> |
break; |
875 |
> |
} |
876 |
|
} |
677 |
– |
ws[k] = w; |
678 |
– |
workers = ws; // volatile array write ensures slot visibility |
679 |
– |
} finally { |
680 |
– |
lock.unlock(); |
877 |
|
} |
682 |
– |
return k; |
878 |
|
} |
879 |
|
|
880 |
+ |
// Submissions |
881 |
+ |
|
882 |
|
/** |
883 |
< |
* Nulls out record of worker in workers array |
883 |
> |
* Enqueues the given task in the submissionQueue. Same idea as |
884 |
> |
* ForkJoinWorkerThread.pushTask except for use of submissionLock. |
885 |
> |
* |
886 |
> |
* @param t the task |
887 |
|
*/ |
888 |
< |
private void forgetWorker(ForkJoinWorkerThread w) { |
889 |
< |
int idx = w.poolIndex; |
690 |
< |
// Locking helps method recordWorker avoid unecessary expansion |
691 |
< |
final ReentrantLock lock = this.workerLock; |
888 |
> |
private void addSubmission(ForkJoinTask<?> t) { |
889 |
> |
final ReentrantLock lock = this.submissionLock; |
890 |
|
lock.lock(); |
891 |
|
try { |
892 |
< |
ForkJoinWorkerThread[] ws = workers; |
893 |
< |
if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify |
894 |
< |
ws[idx] = null; |
892 |
> |
ForkJoinTask<?>[] q; int s, m; |
893 |
> |
if ((q = submissionQueue) != null) { // ignore if queue removed |
894 |
> |
long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE; |
895 |
> |
UNSAFE.putOrderedObject(q, u, t); |
896 |
> |
queueTop = s + 1; |
897 |
> |
if (s - queueBase == m) |
898 |
> |
growSubmissionQueue(); |
899 |
> |
} |
900 |
|
} finally { |
901 |
|
lock.unlock(); |
902 |
|
} |
903 |
+ |
signalWork(); |
904 |
|
} |
905 |
|
|
906 |
< |
// adding and removing workers |
906 |
> |
// (pollSubmission is defined below with exported methods) |
907 |
|
|
908 |
|
/** |
909 |
< |
* Tries to create and add new worker. Assumes that worker counts |
910 |
< |
* are already updated to accommodate the worker, so adjusts on |
707 |
< |
* failure. |
909 |
> |
* Creates or doubles submissionQueue array. |
910 |
> |
* Basically identical to ForkJoinWorkerThread version. |
911 |
|
*/ |
912 |
< |
private void addWorker() { |
913 |
< |
ForkJoinWorkerThread w = null; |
914 |
< |
try { |
915 |
< |
w = factory.newThread(this); |
916 |
< |
} finally { // Adjust on either null or exceptional factory return |
917 |
< |
if (w == null) { |
918 |
< |
decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL); |
919 |
< |
tryTerminate(false); // in case of failure during shutdown |
912 |
> |
private void growSubmissionQueue() { |
913 |
> |
ForkJoinTask<?>[] oldQ = submissionQueue; |
914 |
> |
int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY; |
915 |
> |
if (size > MAXIMUM_QUEUE_CAPACITY) |
916 |
> |
throw new RejectedExecutionException("Queue capacity exceeded"); |
917 |
> |
if (size < INITIAL_QUEUE_CAPACITY) |
918 |
> |
size = INITIAL_QUEUE_CAPACITY; |
919 |
> |
ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size]; |
920 |
> |
int mask = size - 1; |
921 |
> |
int top = queueTop; |
922 |
> |
int oldMask; |
923 |
> |
if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) { |
924 |
> |
for (int b = queueBase; b != top; ++b) { |
925 |
> |
long u = ((b & oldMask) << ASHIFT) + ABASE; |
926 |
> |
Object x = UNSAFE.getObjectVolatile(oldQ, u); |
927 |
> |
if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null)) |
928 |
> |
UNSAFE.putObjectVolatile |
929 |
> |
(q, ((b & mask) << ASHIFT) + ABASE, x); |
930 |
|
} |
931 |
|
} |
719 |
– |
if (w != null) |
720 |
– |
w.start(recordWorker(w), ueh); |
932 |
|
} |
933 |
|
|
934 |
+ |
// Blocking support |
935 |
+ |
|
936 |
|
/** |
937 |
< |
* Final callback from terminating worker. Removes record of |
938 |
< |
* worker from array, and adjusts counts. If pool is shutting |
939 |
< |
* down, tries to complete terminatation. |
940 |
< |
* |
941 |
< |
* @param w the worker |
942 |
< |
*/ |
943 |
< |
final void workerTerminated(ForkJoinWorkerThread w) { |
944 |
< |
forgetWorker(w); |
945 |
< |
decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL); |
946 |
< |
while (w.stealCount != 0) // collect final count |
947 |
< |
tryAccumulateStealCount(w); |
948 |
< |
tryTerminate(false); |
937 |
> |
* Tries to increment blockedCount, decrement active count |
938 |
> |
* (sometimes implicitly) and possibly release or create a |
939 |
> |
* compensating worker in preparation for blocking. Fails |
940 |
> |
* on contention or termination. |
941 |
> |
* |
942 |
> |
* @return true if the caller can block, else should recheck and retry |
943 |
> |
*/ |
944 |
> |
private boolean tryPreBlock() { |
945 |
> |
int b = blockedCount; |
946 |
> |
if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) { |
947 |
> |
int pc = parallelism; |
948 |
> |
do { |
949 |
> |
ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w; |
950 |
> |
int e, ac, tc, i; |
951 |
> |
long c = ctl; |
952 |
> |
int u = (int)(c >>> 32); |
953 |
> |
if ((e = (int)c) < 0) { |
954 |
> |
// skip -- terminating |
955 |
> |
} |
956 |
> |
else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 && |
957 |
> |
(ws = workers) != null && |
958 |
> |
(i = ~e & SMASK) < ws.length && |
959 |
> |
(w = ws[i]) != null) { |
960 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
961 |
> |
(c & (AC_MASK|TC_MASK))); |
962 |
> |
if (w.eventCount == e && |
963 |
> |
UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { |
964 |
> |
w.eventCount = (e + EC_UNIT) & E_MASK; |
965 |
> |
if (w.parked) |
966 |
> |
UNSAFE.unpark(w); |
967 |
> |
return true; // release an idle worker |
968 |
> |
} |
969 |
> |
} |
970 |
> |
else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) { |
971 |
> |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
972 |
> |
if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) |
973 |
> |
return true; // no compensation needed |
974 |
> |
} |
975 |
> |
else if (tc + pc < MAX_ID) { |
976 |
> |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
977 |
> |
if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { |
978 |
> |
addWorker(); |
979 |
> |
return true; // create a replacement |
980 |
> |
} |
981 |
> |
} |
982 |
> |
// try to back out on any failure and let caller retry |
983 |
> |
} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset, |
984 |
> |
b = blockedCount, b - 1)); |
985 |
> |
} |
986 |
> |
return false; |
987 |
|
} |
988 |
|
|
738 |
– |
// Waiting for and signalling events |
739 |
– |
|
989 |
|
/** |
990 |
< |
* Releases workers blocked on a count not equal to current count. |
742 |
< |
* Normally called after precheck that eventWaiters isn't zero to |
743 |
< |
* avoid wasted array checks. |
744 |
< |
* |
745 |
< |
* @param signalling true if caller is a signalling worker so can |
746 |
< |
* exit upon (conservatively) detected contention by other threads |
747 |
< |
* who will continue to release |
990 |
> |
* Decrements blockedCount and increments active count. |
991 |
|
*/ |
992 |
< |
private void releaseEventWaiters(boolean signalling) { |
993 |
< |
ForkJoinWorkerThread[] ws = workers; |
994 |
< |
int n = ws.length; |
995 |
< |
long h; // head of stack |
996 |
< |
ForkJoinWorkerThread w; int id, ec; |
997 |
< |
while ((id = ((int)((h = eventWaiters) & WAITER_ID_MASK)) - 1) >= 0 && |
998 |
< |
(int)(h >>> EVENT_COUNT_SHIFT) != (ec = eventCount) && |
756 |
< |
id < n && (w = ws[id]) != null) { |
757 |
< |
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
758 |
< |
h, h = w.nextWaiter)) |
759 |
< |
LockSupport.unpark(w); |
760 |
< |
if (signalling && (eventCount != ec || eventWaiters != h)) |
761 |
< |
break; |
762 |
< |
} |
992 |
> |
private void postBlock() { |
993 |
> |
long c; |
994 |
> |
do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, // no mask |
995 |
> |
c = ctl, c + AC_UNIT)); |
996 |
> |
int b; |
997 |
> |
do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset, |
998 |
> |
b = blockedCount, b - 1)); |
999 |
|
} |
1000 |
|
|
1001 |
|
/** |
1002 |
< |
* Tries to advance eventCount and releases waiters. Called only |
1003 |
< |
* from workers. |
1002 |
> |
* Possibly blocks waiting for the given task to complete, or |
1003 |
> |
* cancels the task if terminating. Fails to wait if contended. |
1004 |
> |
* |
1005 |
> |
* @param joinMe the task |
1006 |
|
*/ |
1007 |
< |
final void signalWork() { |
1008 |
< |
int c; // try to increment event count -- CAS failure OK |
1009 |
< |
UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1); |
1010 |
< |
if (eventWaiters != 0L) |
1011 |
< |
releaseEventWaiters(true); |
1007 |
> |
final void tryAwaitJoin(ForkJoinTask<?> joinMe) { |
1008 |
> |
Thread.interrupted(); // clear interrupts before checking termination |
1009 |
> |
if (joinMe.status >= 0) { |
1010 |
> |
if (tryPreBlock()) { |
1011 |
> |
joinMe.tryAwaitDone(0L); |
1012 |
> |
postBlock(); |
1013 |
> |
} |
1014 |
> |
else if ((ctl & STOP_BIT) != 0L) |
1015 |
> |
joinMe.cancelIgnoringExceptions(); |
1016 |
> |
} |
1017 |
|
} |
1018 |
|
|
1019 |
|
/** |
1020 |
< |
* Blocks worker until terminating or event count |
1021 |
< |
* advances from last value held by worker |
1022 |
< |
* |
1023 |
< |
* @param w the calling worker thread |
1024 |
< |
*/ |
1025 |
< |
private void eventSync(ForkJoinWorkerThread w) { |
1026 |
< |
int wec = w.lastEventCount; |
1027 |
< |
long nh = (((long)wec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1)); |
1028 |
< |
long h; |
1029 |
< |
while ((runState < SHUTDOWN || !tryTerminate(false)) && |
1030 |
< |
((h = eventWaiters) == 0L || |
1031 |
< |
(int)(h >>> EVENT_COUNT_SHIFT) == wec) && |
1032 |
< |
eventCount == wec) { |
1033 |
< |
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
1034 |
< |
w.nextWaiter = h, nh)) { |
1035 |
< |
while (runState < TERMINATING && eventCount == wec) { |
1036 |
< |
if (!tryAccumulateStealCount(w)) // transfer while idle |
1037 |
< |
continue; |
1038 |
< |
Thread.interrupted(); // clear/ignore interrupt |
1039 |
< |
if (eventCount != wec) |
1020 |
> |
* Possibly blocks the given worker waiting for joinMe to |
1021 |
> |
* complete or timeout. |
1022 |
> |
* |
1023 |
> |
* @param joinMe the task |
1024 |
> |
* @param millis the wait time for underlying Object.wait |
1025 |
> |
*/ |
1026 |
> |
final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) { |
1027 |
> |
while (joinMe.status >= 0) { |
1028 |
> |
Thread.interrupted(); |
1029 |
> |
if ((ctl & STOP_BIT) != 0L) { |
1030 |
> |
joinMe.cancelIgnoringExceptions(); |
1031 |
> |
break; |
1032 |
> |
} |
1033 |
> |
if (tryPreBlock()) { |
1034 |
> |
long last = System.nanoTime(); |
1035 |
> |
while (joinMe.status >= 0) { |
1036 |
> |
long millis = TimeUnit.NANOSECONDS.toMillis(nanos); |
1037 |
> |
if (millis <= 0) |
1038 |
> |
break; |
1039 |
> |
joinMe.tryAwaitDone(millis); |
1040 |
> |
if (joinMe.status < 0) |
1041 |
|
break; |
1042 |
< |
LockSupport.park(w); |
1042 |
> |
if ((ctl & STOP_BIT) != 0L) { |
1043 |
> |
joinMe.cancelIgnoringExceptions(); |
1044 |
> |
break; |
1045 |
> |
} |
1046 |
> |
long now = System.nanoTime(); |
1047 |
> |
nanos -= now - last; |
1048 |
> |
last = now; |
1049 |
|
} |
1050 |
+ |
postBlock(); |
1051 |
|
break; |
1052 |
|
} |
1053 |
|
} |
803 |
– |
w.lastEventCount = eventCount; |
1054 |
|
} |
1055 |
|
|
806 |
– |
// Maintaining spares |
807 |
– |
|
1056 |
|
/** |
1057 |
< |
* Pushes worker onto the spare stack |
1057 |
> |
* If necessary, compensates for blocker, and blocks. |
1058 |
|
*/ |
1059 |
< |
final void pushSpare(ForkJoinWorkerThread w) { |
1060 |
< |
int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex+1); |
1061 |
< |
do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
1062 |
< |
w.nextSpare = spareWaiters,ns)); |
1063 |
< |
} |
1064 |
< |
|
1065 |
< |
/** |
1066 |
< |
* Tries (once) to resume a spare if running count is less than |
1067 |
< |
* target parallelism. Fails on contention or stale workers. |
1068 |
< |
*/ |
821 |
< |
private void tryResumeSpare() { |
822 |
< |
int sw, id; |
823 |
< |
ForkJoinWorkerThread w; |
824 |
< |
ForkJoinWorkerThread[] ws; |
825 |
< |
if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 && |
826 |
< |
id < (ws = workers).length && (w = ws[id]) != null && |
827 |
< |
(workerCounts & RUNNING_COUNT_MASK) < parallelism && |
828 |
< |
eventWaiters == 0L && |
829 |
< |
spareWaiters == sw && |
830 |
< |
UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
831 |
< |
sw, w.nextSpare) && |
832 |
< |
w.tryUnsuspend()) { |
833 |
< |
int c; // try increment; if contended, finish after unpark |
834 |
< |
boolean inc = UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
835 |
< |
c = workerCounts, |
836 |
< |
c + ONE_RUNNING); |
837 |
< |
LockSupport.unpark(w); |
838 |
< |
if (!inc) { |
839 |
< |
do {} while(!UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
840 |
< |
c = workerCounts, |
841 |
< |
c + ONE_RUNNING)); |
1059 |
> |
private void awaitBlocker(ManagedBlocker blocker) |
1060 |
> |
throws InterruptedException { |
1061 |
> |
while (!blocker.isReleasable()) { |
1062 |
> |
if (tryPreBlock()) { |
1063 |
> |
try { |
1064 |
> |
do {} while (!blocker.isReleasable() && !blocker.block()); |
1065 |
> |
} finally { |
1066 |
> |
postBlock(); |
1067 |
> |
} |
1068 |
> |
break; |
1069 |
|
} |
1070 |
|
} |
1071 |
|
} |
1072 |
|
|
1073 |
< |
/** |
847 |
< |
* Callback from oldest spare occasionally waking up. Tries |
848 |
< |
* (once) to shutdown a spare if more than 25% spare overage, or |
849 |
< |
* if UNUSED_SPARE_TRIM_RATE_NANOS have elapsed and there are at |
850 |
< |
* least #parallelism running threads. Note that we don't need CAS |
851 |
< |
* or locks here because the method is called only from the oldest |
852 |
< |
* suspended spare occasionally waking (and even misfires are OK). |
853 |
< |
* |
854 |
< |
* @param now the wake up nanoTime of caller |
855 |
< |
*/ |
856 |
< |
final void tryTrimSpare(long now) { |
857 |
< |
long lastTrim = trimTime; |
858 |
< |
trimTime = now; |
859 |
< |
helpMaintainParallelism(); // first, help wake up any needed spares |
860 |
< |
int sw, id; |
861 |
< |
ForkJoinWorkerThread w; |
862 |
< |
ForkJoinWorkerThread[] ws; |
863 |
< |
int pc = parallelism; |
864 |
< |
int wc = workerCounts; |
865 |
< |
if ((wc & RUNNING_COUNT_MASK) >= pc && |
866 |
< |
(((wc >>> TOTAL_COUNT_SHIFT) - pc) > (pc >>> 2) + 1 ||// approx 25% |
867 |
< |
now - lastTrim >= UNUSED_SPARE_TRIM_RATE_NANOS) && |
868 |
< |
(id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 && |
869 |
< |
id < (ws = workers).length && (w = ws[id]) != null && |
870 |
< |
UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
871 |
< |
sw, w.nextSpare)) |
872 |
< |
w.shutdown(false); |
873 |
< |
} |
1073 |
> |
// Creating, registering and deregistring workers |
1074 |
|
|
1075 |
|
/** |
1076 |
< |
* Does at most one of: |
1077 |
< |
* |
1078 |
< |
* 1. Help wake up existing workers waiting for work via |
1079 |
< |
* releaseEventWaiters. (If any exist, then it probably doesn't |
1080 |
< |
* matter right now if under target parallelism level.) |
1081 |
< |
* |
1082 |
< |
* 2. If below parallelism level and a spare exists, try (once) |
1083 |
< |
* to resume it via tryResumeSpare. |
1084 |
< |
* |
1085 |
< |
* 3. If neither of the above, tries (once) to add a new |
1086 |
< |
* worker if either there are not enough total, or if all |
1087 |
< |
* existing workers are busy, there are either no running |
1088 |
< |
* workers or the deficit is at least twice the surplus. |
1089 |
< |
*/ |
1090 |
< |
private void helpMaintainParallelism() { |
1091 |
< |
// uglified to work better when not compiled |
1092 |
< |
int pc, wc, rc, tc, rs; long h; |
1093 |
< |
if ((h = eventWaiters) != 0L) { |
1094 |
< |
if ((int)(h >>> EVENT_COUNT_SHIFT) != eventCount) |
1095 |
< |
releaseEventWaiters(false); // avoid useless call |
1096 |
< |
} |
1097 |
< |
else if ((pc = parallelism) > |
898 |
< |
(rc = ((wc = workerCounts) & RUNNING_COUNT_MASK))) { |
899 |
< |
if (spareWaiters != 0) |
900 |
< |
tryResumeSpare(); |
901 |
< |
else if ((rs = runState) < TERMINATING && |
902 |
< |
((tc = wc >>> TOTAL_COUNT_SHIFT) < pc || |
903 |
< |
(tc == (rs & ACTIVE_COUNT_MASK) && // all busy |
904 |
< |
(rc == 0 || // must add |
905 |
< |
rc < pc - ((tc - pc) << 1)) && // within slack |
906 |
< |
tc < MAX_WORKERS && runState == rs)) && // recheck busy |
907 |
< |
workerCounts == wc && |
908 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc, |
909 |
< |
wc + (ONE_RUNNING|ONE_TOTAL))) |
910 |
< |
addWorker(); |
1076 |
> |
* Tries to create and start a worker; minimally rolls back counts |
1077 |
> |
* on failure. |
1078 |
> |
*/ |
1079 |
> |
private void addWorker() { |
1080 |
> |
Throwable ex = null; |
1081 |
> |
ForkJoinWorkerThread t = null; |
1082 |
> |
try { |
1083 |
> |
t = factory.newThread(this); |
1084 |
> |
} catch (Throwable e) { |
1085 |
> |
ex = e; |
1086 |
> |
} |
1087 |
> |
if (t == null) { // null or exceptional factory return |
1088 |
> |
long c; // adjust counts |
1089 |
> |
do {} while (!UNSAFE.compareAndSwapLong |
1090 |
> |
(this, ctlOffset, c = ctl, |
1091 |
> |
(((c - AC_UNIT) & AC_MASK) | |
1092 |
> |
((c - TC_UNIT) & TC_MASK) | |
1093 |
> |
(c & ~(AC_MASK|TC_MASK))))); |
1094 |
> |
// Propagate exception if originating from an external caller |
1095 |
> |
if (!tryTerminate(false) && ex != null && |
1096 |
> |
!(Thread.currentThread() instanceof ForkJoinWorkerThread)) |
1097 |
> |
UNSAFE.throwException(ex); |
1098 |
|
} |
1099 |
+ |
else |
1100 |
+ |
t.start(); |
1101 |
|
} |
1102 |
|
|
1103 |
|
/** |
1104 |
< |
* Callback from workers invoked upon each top-level action (i.e., |
1105 |
< |
* stealing a task or taking a submission and running |
917 |
< |
* it). Performs one or more of the following: |
918 |
< |
* |
919 |
< |
* 1. If the worker cannot find work (misses > 0), updates its |
920 |
< |
* active status to inactive and updates activeCount unless |
921 |
< |
* this is the first miss and there is contention, in which |
922 |
< |
* case it may try again (either in this or a subsequent |
923 |
< |
* call). |
924 |
< |
* |
925 |
< |
* 2. If there are at least 2 misses, awaits the next task event |
926 |
< |
* via eventSync |
927 |
< |
* |
928 |
< |
* 3. If there are too many running threads, suspends this worker |
929 |
< |
* (first forcing inactivation if necessary). If it is not |
930 |
< |
* needed, it may be killed while suspended via |
931 |
< |
* tryTrimSpare. Otherwise, upon resume it rechecks to make |
932 |
< |
* sure that it is still needed. |
933 |
< |
* |
934 |
< |
* 4. Helps release and/or reactivate other workers via |
935 |
< |
* helpMaintainParallelism |
936 |
< |
* |
937 |
< |
* @param w the worker |
938 |
< |
* @param misses the number of scans by caller failing to find work |
939 |
< |
* (saturating at 2 just to avoid wraparound) |
1104 |
> |
* Callback from ForkJoinWorkerThread constructor to assign a |
1105 |
> |
* public name |
1106 |
|
*/ |
1107 |
< |
final void preStep(ForkJoinWorkerThread w, int misses) { |
1108 |
< |
boolean active = w.active; |
1109 |
< |
int pc = parallelism; |
1110 |
< |
for (;;) { |
1111 |
< |
int wc = workerCounts; |
946 |
< |
int rc = wc & RUNNING_COUNT_MASK; |
947 |
< |
if (active && (misses > 0 || rc > pc)) { |
948 |
< |
int rs; // try inactivate |
949 |
< |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, |
950 |
< |
rs = runState, rs - ONE_ACTIVE)) |
951 |
< |
active = w.active = false; |
952 |
< |
else if (misses > 1 || rc > pc || |
953 |
< |
(rs & ACTIVE_COUNT_MASK) >= pc) |
954 |
< |
continue; // force inactivate |
955 |
< |
} |
956 |
< |
if (misses > 1) { |
957 |
< |
misses = 0; // don't re-sync |
958 |
< |
eventSync(w); // continue loop to recheck rc |
959 |
< |
} |
960 |
< |
else if (rc > pc) { |
961 |
< |
if (workerCounts == wc && // try to suspend as spare |
962 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
963 |
< |
wc, wc - ONE_RUNNING) && |
964 |
< |
!w.suspendAsSpare()) // false if killed |
965 |
< |
break; |
966 |
< |
} |
967 |
< |
else { |
968 |
< |
if (rc < pc || eventWaiters != 0L) |
969 |
< |
helpMaintainParallelism(); |
970 |
< |
break; |
971 |
< |
} |
1107 |
> |
final String nextWorkerName() { |
1108 |
> |
for (int n;;) { |
1109 |
> |
if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset, |
1110 |
> |
n = nextWorkerNumber, ++n)) |
1111 |
> |
return workerNamePrefix + n; |
1112 |
|
} |
1113 |
|
} |
1114 |
|
|
1115 |
|
/** |
1116 |
< |
* Helps and/or blocks awaiting join of the given task. |
1117 |
< |
* Alternates between helpJoinTask() and helpMaintainParallelism() |
978 |
< |
* as many times as there is a deficit in running count (or longer |
979 |
< |
* if running count would become zero), then blocks if task still |
980 |
< |
* not done. |
1116 |
> |
* Callback from ForkJoinWorkerThread constructor to |
1117 |
> |
* determine its poolIndex and record in workers array. |
1118 |
|
* |
1119 |
< |
* @param joinMe the task to join |
1119 |
> |
* @param w the worker |
1120 |
> |
* @return the worker's pool index |
1121 |
|
*/ |
1122 |
< |
final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) { |
1123 |
< |
int threshold = parallelism; // descend blocking thresholds |
1124 |
< |
while (joinMe.status >= 0) { |
1125 |
< |
boolean block; int wc; |
1126 |
< |
worker.helpJoinTask(joinMe); |
1127 |
< |
if (joinMe.status < 0) |
1128 |
< |
break; |
1129 |
< |
if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) { |
1130 |
< |
if (threshold > 0) |
1131 |
< |
--threshold; |
1132 |
< |
else |
1133 |
< |
advanceEventCount(); // force release |
1134 |
< |
block = false; |
1122 |
> |
final int registerWorker(ForkJoinWorkerThread w) { |
1123 |
> |
/* |
1124 |
> |
* In the typical case, a new worker acquires the lock, uses |
1125 |
> |
* next available index and returns quickly. Since we should |
1126 |
> |
* not block callers (ultimately from signalWork or |
1127 |
> |
* tryPreBlock) waiting for the lock needed to do this, we |
1128 |
> |
* instead help release other workers while waiting for the |
1129 |
> |
* lock. |
1130 |
> |
*/ |
1131 |
> |
for (int g;;) { |
1132 |
> |
ForkJoinWorkerThread[] ws; |
1133 |
> |
if (((g = scanGuard) & SG_UNIT) == 0 && |
1134 |
> |
UNSAFE.compareAndSwapInt(this, scanGuardOffset, |
1135 |
> |
g, g | SG_UNIT)) { |
1136 |
> |
int k = nextWorkerIndex; |
1137 |
> |
try { |
1138 |
> |
if ((ws = workers) != null) { // ignore on shutdown |
1139 |
> |
int n = ws.length; |
1140 |
> |
if (k < 0 || k >= n || ws[k] != null) { |
1141 |
> |
for (k = 0; k < n && ws[k] != null; ++k) |
1142 |
> |
; |
1143 |
> |
if (k == n) |
1144 |
> |
ws = workers = Arrays.copyOf(ws, n << 1); |
1145 |
> |
} |
1146 |
> |
ws[k] = w; |
1147 |
> |
nextWorkerIndex = k + 1; |
1148 |
> |
int m = g & SMASK; |
1149 |
> |
g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1); |
1150 |
> |
} |
1151 |
> |
} finally { |
1152 |
> |
scanGuard = g; |
1153 |
> |
} |
1154 |
> |
return k; |
1155 |
|
} |
1156 |
< |
else |
1157 |
< |
block = UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
1158 |
< |
wc, wc - ONE_RUNNING); |
1159 |
< |
helpMaintainParallelism(); |
1160 |
< |
if (block) { |
1161 |
< |
int c; |
1162 |
< |
joinMe.internalAwaitDone(); |
1005 |
< |
do {} while (!UNSAFE.compareAndSwapInt |
1006 |
< |
(this, workerCountsOffset, |
1007 |
< |
c = workerCounts, c + ONE_RUNNING)); |
1008 |
< |
break; |
1156 |
> |
else if ((ws = workers) != null) { // help release others |
1157 |
> |
for (ForkJoinWorkerThread u : ws) { |
1158 |
> |
if (u != null && u.queueBase != u.queueTop) { |
1159 |
> |
if (tryReleaseWaiter()) |
1160 |
> |
break; |
1161 |
> |
} |
1162 |
> |
} |
1163 |
|
} |
1164 |
|
} |
1165 |
|
} |
1166 |
|
|
1167 |
|
/** |
1168 |
< |
* Same idea as awaitJoin, but no helping |
1168 |
> |
* Final callback from terminating worker. Removes record of |
1169 |
> |
* worker from array, and adjusts counts. If pool is shutting |
1170 |
> |
* down, tries to complete termination. |
1171 |
> |
* |
1172 |
> |
* @param w the worker |
1173 |
|
*/ |
1174 |
< |
final void awaitBlocker(ManagedBlocker blocker) |
1175 |
< |
throws InterruptedException { |
1176 |
< |
int threshold = parallelism; |
1177 |
< |
while (!blocker.isReleasable()) { |
1178 |
< |
boolean block; int wc; |
1179 |
< |
if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) { |
1180 |
< |
if (threshold > 0) |
1181 |
< |
--threshold; |
1182 |
< |
else |
1183 |
< |
advanceEventCount(); |
1184 |
< |
block = false; |
1174 |
> |
final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) { |
1175 |
> |
int idx = w.poolIndex; |
1176 |
> |
int sc = w.stealCount; |
1177 |
> |
int steps = 0; |
1178 |
> |
// Remove from array, adjust worker counts and collect steal count. |
1179 |
> |
// We can intermix failed removes or adjusts with steal updates |
1180 |
> |
do { |
1181 |
> |
long s, c; |
1182 |
> |
int g; |
1183 |
> |
if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 && |
1184 |
> |
UNSAFE.compareAndSwapInt(this, scanGuardOffset, |
1185 |
> |
g, g |= SG_UNIT)) { |
1186 |
> |
ForkJoinWorkerThread[] ws = workers; |
1187 |
> |
if (ws != null && idx >= 0 && |
1188 |
> |
idx < ws.length && ws[idx] == w) |
1189 |
> |
ws[idx] = null; // verify |
1190 |
> |
nextWorkerIndex = idx; |
1191 |
> |
scanGuard = g + SG_UNIT; |
1192 |
> |
steps = 1; |
1193 |
|
} |
1194 |
+ |
if (steps == 1 && |
1195 |
+ |
UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl, |
1196 |
+ |
(((c - AC_UNIT) & AC_MASK) | |
1197 |
+ |
((c - TC_UNIT) & TC_MASK) | |
1198 |
+ |
(c & ~(AC_MASK|TC_MASK))))) |
1199 |
+ |
steps = 2; |
1200 |
+ |
if (sc != 0 && |
1201 |
+ |
UNSAFE.compareAndSwapLong(this, stealCountOffset, |
1202 |
+ |
s = stealCount, s + sc)) |
1203 |
+ |
sc = 0; |
1204 |
+ |
} while (steps != 2 || sc != 0); |
1205 |
+ |
if (!tryTerminate(false)) { |
1206 |
+ |
if (ex != null) // possibly replace if died abnormally |
1207 |
+ |
signalWork(); |
1208 |
|
else |
1209 |
< |
block = UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
1030 |
< |
wc, wc - ONE_RUNNING); |
1031 |
< |
helpMaintainParallelism(); |
1032 |
< |
if (block) { |
1033 |
< |
try { |
1034 |
< |
do {} while (!blocker.isReleasable() && !blocker.block()); |
1035 |
< |
} finally { |
1036 |
< |
int c; |
1037 |
< |
do {} while (!UNSAFE.compareAndSwapInt |
1038 |
< |
(this, workerCountsOffset, |
1039 |
< |
c = workerCounts, c + ONE_RUNNING)); |
1040 |
< |
} |
1041 |
< |
break; |
1042 |
< |
} |
1209 |
> |
tryReleaseWaiter(); |
1210 |
|
} |
1211 |
|
} |
1212 |
|
|
1213 |
+ |
// Shutdown and termination |
1214 |
+ |
|
1215 |
|
/** |
1216 |
|
* Possibly initiates and/or completes termination. |
1217 |
|
* |
1220 |
|
* @return true if now terminating or terminated |
1221 |
|
*/ |
1222 |
|
private boolean tryTerminate(boolean now) { |
1223 |
< |
if (now) |
1224 |
< |
advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN |
1225 |
< |
else if (runState < SHUTDOWN || |
1226 |
< |
!submissionQueue.isEmpty() || |
1227 |
< |
(runState & ACTIVE_COUNT_MASK) != 0) |
1228 |
< |
return false; |
1229 |
< |
|
1230 |
< |
if (advanceRunLevel(TERMINATING)) |
1231 |
< |
startTerminating(); |
1232 |
< |
|
1233 |
< |
// Finish now if all threads terminated; else in some subsequent call |
1234 |
< |
if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) { |
1235 |
< |
advanceRunLevel(TERMINATED); |
1236 |
< |
termination.arrive(); |
1223 |
> |
long c; |
1224 |
> |
while (((c = ctl) & STOP_BIT) == 0) { |
1225 |
> |
if (!now) { |
1226 |
> |
if ((int)(c >> AC_SHIFT) != -parallelism) |
1227 |
> |
return false; |
1228 |
> |
if (!shutdown || blockedCount != 0 || quiescerCount != 0 || |
1229 |
> |
queueBase != queueTop) { |
1230 |
> |
if (ctl == c) // staleness check |
1231 |
> |
return false; |
1232 |
> |
continue; |
1233 |
> |
} |
1234 |
> |
} |
1235 |
> |
if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT)) |
1236 |
> |
startTerminating(); |
1237 |
> |
} |
1238 |
> |
if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers |
1239 |
> |
final ReentrantLock lock = this.submissionLock; |
1240 |
> |
lock.lock(); |
1241 |
> |
try { |
1242 |
> |
termination.signalAll(); |
1243 |
> |
} finally { |
1244 |
> |
lock.unlock(); |
1245 |
> |
} |
1246 |
|
} |
1247 |
|
return true; |
1248 |
|
} |
1249 |
|
|
1250 |
|
/** |
1251 |
< |
* Actions on transition to TERMINATING |
1252 |
< |
* |
1253 |
< |
* Runs up to four passes through workers: (0) shutting down each |
1254 |
< |
* quietly (without waking up if parked) to quickly spread |
1255 |
< |
* notifications without unnecessary bouncing around event queues |
1256 |
< |
* etc (1) wake up and help cancel tasks (2) interrupt (3) mop up |
1079 |
< |
* races with interrupted workers |
1251 |
> |
* Runs up to three passes through workers: (0) Setting |
1252 |
> |
* termination status for each worker, followed by wakeups up to |
1253 |
> |
* queued workers; (1) helping cancel tasks; (2) interrupting |
1254 |
> |
* lagging threads (likely in external tasks, but possibly also |
1255 |
> |
* blocked in joins). Each pass repeats previous steps because of |
1256 |
> |
* potential lagging thread creation. |
1257 |
|
*/ |
1258 |
|
private void startTerminating() { |
1259 |
|
cancelSubmissions(); |
1260 |
< |
for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) { |
1084 |
< |
advanceEventCount(); |
1085 |
< |
eventWaiters = 0L; // clobber lists |
1086 |
< |
spareWaiters = 0; |
1260 |
> |
for (int pass = 0; pass < 3; ++pass) { |
1261 |
|
ForkJoinWorkerThread[] ws = workers; |
1262 |
< |
int n = ws.length; |
1263 |
< |
for (int i = 0; i < n; ++i) { |
1264 |
< |
ForkJoinWorkerThread w = ws[i]; |
1265 |
< |
if (w != null) { |
1266 |
< |
w.shutdown(true); |
1267 |
< |
if (passes > 0 && !w.isTerminated()) { |
1268 |
< |
w.cancelTasks(); |
1269 |
< |
LockSupport.unpark(w); |
1270 |
< |
if (passes > 1) { |
1271 |
< |
try { |
1272 |
< |
w.interrupt(); |
1099 |
< |
} catch (SecurityException ignore) { |
1262 |
> |
if (ws != null) { |
1263 |
> |
for (ForkJoinWorkerThread w : ws) { |
1264 |
> |
if (w != null) { |
1265 |
> |
w.terminate = true; |
1266 |
> |
if (pass > 0) { |
1267 |
> |
w.cancelTasks(); |
1268 |
> |
if (pass > 1 && !w.isInterrupted()) { |
1269 |
> |
try { |
1270 |
> |
w.interrupt(); |
1271 |
> |
} catch (SecurityException ignore) { |
1272 |
> |
} |
1273 |
|
} |
1274 |
|
} |
1275 |
|
} |
1276 |
|
} |
1277 |
+ |
terminateWaiters(); |
1278 |
|
} |
1279 |
|
} |
1280 |
|
} |
1281 |
|
|
1282 |
|
/** |
1283 |
< |
* Clear out and cancel submissions, ignoring exceptions |
1283 |
> |
* Polls and cancels all submissions. Called only during termination. |
1284 |
|
*/ |
1285 |
|
private void cancelSubmissions() { |
1286 |
< |
ForkJoinTask<?> task; |
1287 |
< |
while ((task = submissionQueue.poll()) != null) { |
1288 |
< |
try { |
1289 |
< |
task.cancel(false); |
1290 |
< |
} catch (Throwable ignore) { |
1286 |
> |
while (queueBase != queueTop) { |
1287 |
> |
ForkJoinTask<?> task = pollSubmission(); |
1288 |
> |
if (task != null) { |
1289 |
> |
try { |
1290 |
> |
task.cancel(false); |
1291 |
> |
} catch (Throwable ignore) { |
1292 |
> |
} |
1293 |
|
} |
1294 |
|
} |
1295 |
|
} |
1296 |
|
|
1297 |
< |
// misc support for ForkJoinWorkerThread |
1297 |
> |
/** |
1298 |
> |
* Tries to set the termination status of waiting workers, and |
1299 |
> |
* then wakes them up (after which they will terminate). |
1300 |
> |
*/ |
1301 |
> |
private void terminateWaiters() { |
1302 |
> |
ForkJoinWorkerThread[] ws = workers; |
1303 |
> |
if (ws != null) { |
1304 |
> |
ForkJoinWorkerThread w; long c; int i, e; |
1305 |
> |
int n = ws.length; |
1306 |
> |
while ((i = ~(e = (int)(c = ctl)) & SMASK) < n && |
1307 |
> |
(w = ws[i]) != null && w.eventCount == (e & E_MASK)) { |
1308 |
> |
if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, |
1309 |
> |
(long)(w.nextWait & E_MASK) | |
1310 |
> |
((c + AC_UNIT) & AC_MASK) | |
1311 |
> |
(c & (TC_MASK|STOP_BIT)))) { |
1312 |
> |
w.terminate = true; |
1313 |
> |
w.eventCount = e + EC_UNIT; |
1314 |
> |
if (w.parked) |
1315 |
> |
UNSAFE.unpark(w); |
1316 |
> |
} |
1317 |
> |
} |
1318 |
> |
} |
1319 |
> |
} |
1320 |
> |
|
1321 |
> |
// misc ForkJoinWorkerThread support |
1322 |
|
|
1323 |
|
/** |
1324 |
< |
* Returns pool number |
1324 |
> |
* Increments or decrements quiescerCount. Needed only to prevent |
1325 |
> |
* triggering shutdown if a worker is transiently inactive while |
1326 |
> |
* checking quiescence. |
1327 |
> |
* |
1328 |
> |
* @param delta 1 for increment, -1 for decrement |
1329 |
|
*/ |
1330 |
< |
final int getPoolNumber() { |
1331 |
< |
return poolNumber; |
1330 |
> |
final void addQuiescerCount(int delta) { |
1331 |
> |
int c; |
1332 |
> |
do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset, |
1333 |
> |
c = quiescerCount, c + delta)); |
1334 |
|
} |
1335 |
|
|
1336 |
|
/** |
1337 |
< |
* Tries to accumulates steal count from a worker, clearing |
1338 |
< |
* the worker's value. |
1337 |
> |
* Directly increments or decrements active count without queuing. |
1338 |
> |
* This method is used to transiently assert inactivation while |
1339 |
> |
* checking quiescence. |
1340 |
|
* |
1341 |
< |
* @return true if worker steal count now zero |
1341 |
> |
* @param delta 1 for increment, -1 for decrement |
1342 |
|
*/ |
1343 |
< |
final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) { |
1344 |
< |
int sc = w.stealCount; |
1345 |
< |
long c = stealCount; |
1346 |
< |
// CAS even if zero, for fence effects |
1347 |
< |
if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) { |
1348 |
< |
if (sc != 0) |
1142 |
< |
w.stealCount = 0; |
1143 |
< |
return true; |
1144 |
< |
} |
1145 |
< |
return sc == 0; |
1343 |
> |
final void addActiveCount(int delta) { |
1344 |
> |
long d = delta < 0 ? -AC_UNIT : AC_UNIT; |
1345 |
> |
long c; |
1346 |
> |
do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl, |
1347 |
> |
((c + d) & AC_MASK) | |
1348 |
> |
(c & ~AC_MASK))); |
1349 |
|
} |
1350 |
|
|
1351 |
|
/** |
1353 |
|
* active thread. |
1354 |
|
*/ |
1355 |
|
final int idlePerActive() { |
1356 |
< |
int pc = parallelism; // use parallelism, not rc |
1357 |
< |
int ac = runState; // no mask -- artifically boosts during shutdown |
1358 |
< |
// Use exact results for small values, saturate past 4 |
1359 |
< |
return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3; |
1356 |
> |
// Approximate at powers of two for small values, saturate past 4 |
1357 |
> |
int p = parallelism; |
1358 |
> |
int a = p + (int)(ctl >> AC_SHIFT); |
1359 |
> |
return (a > (p >>>= 1) ? 0 : |
1360 |
> |
a > (p >>>= 1) ? 1 : |
1361 |
> |
a > (p >>>= 1) ? 2 : |
1362 |
> |
a > (p >>>= 1) ? 4 : |
1363 |
> |
8); |
1364 |
|
} |
1365 |
|
|
1366 |
< |
// Public and protected methods |
1366 |
> |
// Exported methods |
1367 |
|
|
1368 |
|
// Constructors |
1369 |
|
|
1410 |
|
* use {@link #defaultForkJoinWorkerThreadFactory}. |
1411 |
|
* @param handler the handler for internal worker threads that |
1412 |
|
* terminate due to unrecoverable errors encountered while executing |
1413 |
< |
* tasks. For default value, use <code>null</code>. |
1413 |
> |
* tasks. For default value, use {@code null}. |
1414 |
|
* @param asyncMode if true, |
1415 |
|
* establishes local first-in-first-out scheduling mode for forked |
1416 |
|
* tasks that are never joined. This mode may be more appropriate |
1417 |
|
* than default locally stack-based mode in applications in which |
1418 |
|
* worker threads only process event-style asynchronous tasks. |
1419 |
< |
* For default value, use <code>false</code>. |
1419 |
> |
* For default value, use {@code false}. |
1420 |
|
* @throws IllegalArgumentException if parallelism less than or |
1421 |
|
* equal to zero, or greater than implementation limit |
1422 |
|
* @throws NullPointerException if the factory is null |
1432 |
|
checkPermission(); |
1433 |
|
if (factory == null) |
1434 |
|
throw new NullPointerException(); |
1435 |
< |
if (parallelism <= 0 || parallelism > MAX_WORKERS) |
1435 |
> |
if (parallelism <= 0 || parallelism > MAX_ID) |
1436 |
|
throw new IllegalArgumentException(); |
1437 |
|
this.parallelism = parallelism; |
1438 |
|
this.factory = factory; |
1439 |
|
this.ueh = handler; |
1440 |
|
this.locallyFifo = asyncMode; |
1441 |
< |
int arraySize = initialArraySizeFor(parallelism); |
1442 |
< |
this.workers = new ForkJoinWorkerThread[arraySize]; |
1443 |
< |
this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>(); |
1444 |
< |
this.workerLock = new ReentrantLock(); |
1445 |
< |
this.termination = new Phaser(1); |
1446 |
< |
this.poolNumber = poolNumberGenerator.incrementAndGet(); |
1447 |
< |
this.trimTime = System.nanoTime(); |
1448 |
< |
} |
1449 |
< |
|
1450 |
< |
/** |
1451 |
< |
* Returns initial power of two size for workers array. |
1452 |
< |
* @param pc the initial parallelism level |
1453 |
< |
*/ |
1454 |
< |
private static int initialArraySizeFor(int pc) { |
1455 |
< |
// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16) |
1456 |
< |
int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS; |
1457 |
< |
size |= size >>> 1; |
1251 |
< |
size |= size >>> 2; |
1252 |
< |
size |= size >>> 4; |
1253 |
< |
size |= size >>> 8; |
1254 |
< |
return size + 1; |
1441 |
> |
long np = (long)(-parallelism); // offset ctl counts |
1442 |
> |
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
1443 |
> |
this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
1444 |
> |
// initialize workers array with room for 2*parallelism if possible |
1445 |
> |
int n = parallelism << 1; |
1446 |
> |
if (n >= MAX_ID) |
1447 |
> |
n = MAX_ID; |
1448 |
> |
else { // See Hackers Delight, sec 3.2, where n < (1 << 16) |
1449 |
> |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; |
1450 |
> |
} |
1451 |
> |
workers = new ForkJoinWorkerThread[n + 1]; |
1452 |
> |
this.submissionLock = new ReentrantLock(); |
1453 |
> |
this.termination = submissionLock.newCondition(); |
1454 |
> |
StringBuilder sb = new StringBuilder("ForkJoinPool-"); |
1455 |
> |
sb.append(poolNumberGenerator.incrementAndGet()); |
1456 |
> |
sb.append("-worker-"); |
1457 |
> |
this.workerNamePrefix = sb.toString(); |
1458 |
|
} |
1459 |
|
|
1460 |
|
// Execution methods |
1461 |
|
|
1462 |
|
/** |
1260 |
– |
* Common code for execute, invoke and submit |
1261 |
– |
*/ |
1262 |
– |
private <T> void doSubmit(ForkJoinTask<T> task) { |
1263 |
– |
if (task == null) |
1264 |
– |
throw new NullPointerException(); |
1265 |
– |
if (runState >= SHUTDOWN) |
1266 |
– |
throw new RejectedExecutionException(); |
1267 |
– |
submissionQueue.offer(task); |
1268 |
– |
advanceEventCount(); |
1269 |
– |
helpMaintainParallelism(); // start or wake up workers |
1270 |
– |
} |
1271 |
– |
|
1272 |
– |
/** |
1463 |
|
* Performs the given task, returning its result upon completion. |
1464 |
< |
* If the caller is already engaged in a fork/join computation in |
1465 |
< |
* the current pool, this method is equivalent in effect to |
1466 |
< |
* {@link ForkJoinTask#invoke}. |
1464 |
> |
* If the computation encounters an unchecked Exception or Error, |
1465 |
> |
* it is rethrown as the outcome of this invocation. Rethrown |
1466 |
> |
* exceptions behave in the same way as regular exceptions, but, |
1467 |
> |
* when possible, contain stack traces (as displayed for example |
1468 |
> |
* using {@code ex.printStackTrace()}) of both the current thread |
1469 |
> |
* as well as the thread actually encountering the exception; |
1470 |
> |
* minimally only the latter. |
1471 |
|
* |
1472 |
|
* @param task the task |
1473 |
|
* @return the task's result |
1476 |
|
* scheduled for execution |
1477 |
|
*/ |
1478 |
|
public <T> T invoke(ForkJoinTask<T> task) { |
1479 |
< |
doSubmit(task); |
1480 |
< |
return task.join(); |
1479 |
> |
Thread t = Thread.currentThread(); |
1480 |
> |
if (task == null) |
1481 |
> |
throw new NullPointerException(); |
1482 |
> |
if (shutdown) |
1483 |
> |
throw new RejectedExecutionException(); |
1484 |
> |
if ((t instanceof ForkJoinWorkerThread) && |
1485 |
> |
((ForkJoinWorkerThread)t).pool == this) |
1486 |
> |
return task.invoke(); // bypass submit if in same pool |
1487 |
> |
else { |
1488 |
> |
addSubmission(task); |
1489 |
> |
return task.join(); |
1490 |
> |
} |
1491 |
> |
} |
1492 |
> |
|
1493 |
> |
/** |
1494 |
> |
* Unless terminating, forks task if within an ongoing FJ |
1495 |
> |
* computation in the current pool, else submits as external task. |
1496 |
> |
*/ |
1497 |
> |
private <T> void forkOrSubmit(ForkJoinTask<T> task) { |
1498 |
> |
ForkJoinWorkerThread w; |
1499 |
> |
Thread t = Thread.currentThread(); |
1500 |
> |
if (shutdown) |
1501 |
> |
throw new RejectedExecutionException(); |
1502 |
> |
if ((t instanceof ForkJoinWorkerThread) && |
1503 |
> |
(w = (ForkJoinWorkerThread)t).pool == this) |
1504 |
> |
w.pushTask(task); |
1505 |
> |
else |
1506 |
> |
addSubmission(task); |
1507 |
|
} |
1508 |
|
|
1509 |
|
/** |
1510 |
|
* Arranges for (asynchronous) execution of the given task. |
1291 |
– |
* If the caller is already engaged in a fork/join computation in |
1292 |
– |
* the current pool, this method is equivalent in effect to |
1293 |
– |
* {@link ForkJoinTask#fork}. |
1511 |
|
* |
1512 |
|
* @param task the task |
1513 |
|
* @throws NullPointerException if the task is null |
1515 |
|
* scheduled for execution |
1516 |
|
*/ |
1517 |
|
public void execute(ForkJoinTask<?> task) { |
1518 |
< |
doSubmit(task); |
1518 |
> |
if (task == null) |
1519 |
> |
throw new NullPointerException(); |
1520 |
> |
forkOrSubmit(task); |
1521 |
|
} |
1522 |
|
|
1523 |
|
// AbstractExecutorService methods |
1528 |
|
* scheduled for execution |
1529 |
|
*/ |
1530 |
|
public void execute(Runnable task) { |
1531 |
+ |
if (task == null) |
1532 |
+ |
throw new NullPointerException(); |
1533 |
|
ForkJoinTask<?> job; |
1534 |
|
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
1535 |
|
job = (ForkJoinTask<?>) task; |
1536 |
|
else |
1537 |
|
job = ForkJoinTask.adapt(task, null); |
1538 |
< |
doSubmit(job); |
1538 |
> |
forkOrSubmit(job); |
1539 |
|
} |
1540 |
|
|
1541 |
|
/** |
1542 |
|
* Submits a ForkJoinTask for execution. |
1322 |
– |
* If the caller is already engaged in a fork/join computation in |
1323 |
– |
* the current pool, this method is equivalent in effect to |
1324 |
– |
* {@link ForkJoinTask#fork}. |
1543 |
|
* |
1544 |
|
* @param task the task to submit |
1545 |
|
* @return the task |
1548 |
|
* scheduled for execution |
1549 |
|
*/ |
1550 |
|
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
1551 |
< |
doSubmit(task); |
1551 |
> |
if (task == null) |
1552 |
> |
throw new NullPointerException(); |
1553 |
> |
forkOrSubmit(task); |
1554 |
|
return task; |
1555 |
|
} |
1556 |
|
|
1560 |
|
* scheduled for execution |
1561 |
|
*/ |
1562 |
|
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
1563 |
+ |
if (task == null) |
1564 |
+ |
throw new NullPointerException(); |
1565 |
|
ForkJoinTask<T> job = ForkJoinTask.adapt(task); |
1566 |
< |
doSubmit(job); |
1566 |
> |
forkOrSubmit(job); |
1567 |
|
return job; |
1568 |
|
} |
1569 |
|
|
1573 |
|
* scheduled for execution |
1574 |
|
*/ |
1575 |
|
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
1576 |
+ |
if (task == null) |
1577 |
+ |
throw new NullPointerException(); |
1578 |
|
ForkJoinTask<T> job = ForkJoinTask.adapt(task, result); |
1579 |
< |
doSubmit(job); |
1579 |
> |
forkOrSubmit(job); |
1580 |
|
return job; |
1581 |
|
} |
1582 |
|
|
1586 |
|
* scheduled for execution |
1587 |
|
*/ |
1588 |
|
public ForkJoinTask<?> submit(Runnable task) { |
1589 |
+ |
if (task == null) |
1590 |
+ |
throw new NullPointerException(); |
1591 |
|
ForkJoinTask<?> job; |
1592 |
|
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
1593 |
|
job = (ForkJoinTask<?>) task; |
1594 |
|
else |
1595 |
|
job = ForkJoinTask.adapt(task, null); |
1596 |
< |
doSubmit(job); |
1596 |
> |
forkOrSubmit(job); |
1597 |
|
return job; |
1598 |
|
} |
1599 |
|
|
1653 |
|
|
1654 |
|
/** |
1655 |
|
* Returns the number of worker threads that have started but not |
1656 |
< |
* yet terminated. This result returned by this method may differ |
1656 |
> |
* yet terminated. The result returned by this method may differ |
1657 |
|
* from {@link #getParallelism} when threads are created to |
1658 |
|
* maintain parallelism when others are cooperatively blocked. |
1659 |
|
* |
1660 |
|
* @return the number of worker threads |
1661 |
|
*/ |
1662 |
|
public int getPoolSize() { |
1663 |
< |
return workerCounts >>> TOTAL_COUNT_SHIFT; |
1663 |
> |
return parallelism + (short)(ctl >>> TC_SHIFT); |
1664 |
|
} |
1665 |
|
|
1666 |
|
/** |
1682 |
|
* @return the number of worker threads |
1683 |
|
*/ |
1684 |
|
public int getRunningThreadCount() { |
1685 |
< |
return workerCounts & RUNNING_COUNT_MASK; |
1685 |
> |
int r = parallelism + (int)(ctl >> AC_SHIFT); |
1686 |
> |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
1687 |
|
} |
1688 |
|
|
1689 |
|
/** |
1694 |
|
* @return the number of active threads |
1695 |
|
*/ |
1696 |
|
public int getActiveThreadCount() { |
1697 |
< |
return runState & ACTIVE_COUNT_MASK; |
1697 |
> |
int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount; |
1698 |
> |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
1699 |
|
} |
1700 |
|
|
1701 |
|
/** |
1710 |
|
* @return {@code true} if all threads are currently idle |
1711 |
|
*/ |
1712 |
|
public boolean isQuiescent() { |
1713 |
< |
return (runState & ACTIVE_COUNT_MASK) == 0; |
1713 |
> |
return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0; |
1714 |
|
} |
1715 |
|
|
1716 |
|
/** |
1740 |
|
*/ |
1741 |
|
public long getQueuedTaskCount() { |
1742 |
|
long count = 0; |
1743 |
< |
ForkJoinWorkerThread[] ws = workers; |
1744 |
< |
int n = ws.length; |
1745 |
< |
for (int i = 0; i < n; ++i) { |
1746 |
< |
ForkJoinWorkerThread w = ws[i]; |
1747 |
< |
if (w != null) |
1748 |
< |
count += w.getQueueSize(); |
1743 |
> |
ForkJoinWorkerThread[] ws; |
1744 |
> |
if ((short)(ctl >>> TC_SHIFT) > -parallelism && |
1745 |
> |
(ws = workers) != null) { |
1746 |
> |
for (ForkJoinWorkerThread w : ws) |
1747 |
> |
if (w != null) |
1748 |
> |
count -= w.queueBase - w.queueTop; // must read base first |
1749 |
|
} |
1750 |
|
return count; |
1751 |
|
} |
1752 |
|
|
1753 |
|
/** |
1754 |
|
* Returns an estimate of the number of tasks submitted to this |
1755 |
< |
* pool that have not yet begun executing. This method takes time |
1756 |
< |
* proportional to the number of submissions. |
1755 |
> |
* pool that have not yet begun executing. This method may take |
1756 |
> |
* time proportional to the number of submissions. |
1757 |
|
* |
1758 |
|
* @return the number of queued submissions |
1759 |
|
*/ |
1760 |
|
public int getQueuedSubmissionCount() { |
1761 |
< |
return submissionQueue.size(); |
1761 |
> |
return -queueBase + queueTop; |
1762 |
|
} |
1763 |
|
|
1764 |
|
/** |
1768 |
|
* @return {@code true} if there are any queued submissions |
1769 |
|
*/ |
1770 |
|
public boolean hasQueuedSubmissions() { |
1771 |
< |
return !submissionQueue.isEmpty(); |
1771 |
> |
return queueBase != queueTop; |
1772 |
|
} |
1773 |
|
|
1774 |
|
/** |
1779 |
|
* @return the next submission, or {@code null} if none |
1780 |
|
*/ |
1781 |
|
protected ForkJoinTask<?> pollSubmission() { |
1782 |
< |
return submissionQueue.poll(); |
1782 |
> |
ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i; |
1783 |
> |
while ((b = queueBase) != queueTop && |
1784 |
> |
(q = submissionQueue) != null && |
1785 |
> |
(i = (q.length - 1) & b) >= 0) { |
1786 |
> |
long u = (i << ASHIFT) + ABASE; |
1787 |
> |
if ((t = q[i]) != null && |
1788 |
> |
queueBase == b && |
1789 |
> |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
1790 |
> |
queueBase = b + 1; |
1791 |
> |
return t; |
1792 |
> |
} |
1793 |
> |
} |
1794 |
> |
return null; |
1795 |
|
} |
1796 |
|
|
1797 |
|
/** |
1812 |
|
* @return the number of elements transferred |
1813 |
|
*/ |
1814 |
|
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
1815 |
< |
int count = submissionQueue.drainTo(c); |
1816 |
< |
ForkJoinWorkerThread[] ws = workers; |
1817 |
< |
int n = ws.length; |
1818 |
< |
for (int i = 0; i < n; ++i) { |
1819 |
< |
ForkJoinWorkerThread w = ws[i]; |
1820 |
< |
if (w != null) |
1821 |
< |
count += w.drainTasksTo(c); |
1815 |
> |
int count = 0; |
1816 |
> |
while (queueBase != queueTop) { |
1817 |
> |
ForkJoinTask<?> t = pollSubmission(); |
1818 |
> |
if (t != null) { |
1819 |
> |
c.add(t); |
1820 |
> |
++count; |
1821 |
> |
} |
1822 |
> |
} |
1823 |
> |
ForkJoinWorkerThread[] ws; |
1824 |
> |
if ((short)(ctl >>> TC_SHIFT) > -parallelism && |
1825 |
> |
(ws = workers) != null) { |
1826 |
> |
for (ForkJoinWorkerThread w : ws) |
1827 |
> |
if (w != null) |
1828 |
> |
count += w.drainTasksTo(c); |
1829 |
|
} |
1830 |
|
return count; |
1831 |
|
} |
1841 |
|
long st = getStealCount(); |
1842 |
|
long qt = getQueuedTaskCount(); |
1843 |
|
long qs = getQueuedSubmissionCount(); |
1597 |
– |
int wc = workerCounts; |
1598 |
– |
int tc = wc >>> TOTAL_COUNT_SHIFT; |
1599 |
– |
int rc = wc & RUNNING_COUNT_MASK; |
1844 |
|
int pc = parallelism; |
1845 |
< |
int rs = runState; |
1846 |
< |
int ac = rs & ACTIVE_COUNT_MASK; |
1845 |
> |
long c = ctl; |
1846 |
> |
int tc = pc + (short)(c >>> TC_SHIFT); |
1847 |
> |
int rc = pc + (int)(c >> AC_SHIFT); |
1848 |
> |
if (rc < 0) // ignore transient negative |
1849 |
> |
rc = 0; |
1850 |
> |
int ac = rc + blockedCount; |
1851 |
> |
String level; |
1852 |
> |
if ((c & STOP_BIT) != 0) |
1853 |
> |
level = (tc == 0) ? "Terminated" : "Terminating"; |
1854 |
> |
else |
1855 |
> |
level = shutdown ? "Shutting down" : "Running"; |
1856 |
|
return super.toString() + |
1857 |
< |
"[" + runLevelToString(rs) + |
1857 |
> |
"[" + level + |
1858 |
|
", parallelism = " + pc + |
1859 |
|
", size = " + tc + |
1860 |
|
", active = " + ac + |
1865 |
|
"]"; |
1866 |
|
} |
1867 |
|
|
1615 |
– |
private static String runLevelToString(int s) { |
1616 |
– |
return ((s & TERMINATED) != 0 ? "Terminated" : |
1617 |
– |
((s & TERMINATING) != 0 ? "Terminating" : |
1618 |
– |
((s & SHUTDOWN) != 0 ? "Shutting down" : |
1619 |
– |
"Running"))); |
1620 |
– |
} |
1621 |
– |
|
1868 |
|
/** |
1869 |
|
* Initiates an orderly shutdown in which previously submitted |
1870 |
|
* tasks are executed, but no new tasks will be accepted. |
1879 |
|
*/ |
1880 |
|
public void shutdown() { |
1881 |
|
checkPermission(); |
1882 |
< |
advanceRunLevel(SHUTDOWN); |
1882 |
> |
shutdown = true; |
1883 |
|
tryTerminate(false); |
1884 |
|
} |
1885 |
|
|
1901 |
|
*/ |
1902 |
|
public List<Runnable> shutdownNow() { |
1903 |
|
checkPermission(); |
1904 |
+ |
shutdown = true; |
1905 |
|
tryTerminate(true); |
1906 |
|
return Collections.emptyList(); |
1907 |
|
} |
1912 |
|
* @return {@code true} if all tasks have completed following shut down |
1913 |
|
*/ |
1914 |
|
public boolean isTerminated() { |
1915 |
< |
return runState >= TERMINATED; |
1915 |
> |
long c = ctl; |
1916 |
> |
return ((c & STOP_BIT) != 0L && |
1917 |
> |
(short)(c >>> TC_SHIFT) == -parallelism); |
1918 |
|
} |
1919 |
|
|
1920 |
|
/** |
1922 |
|
* commenced but not yet completed. This method may be useful for |
1923 |
|
* debugging. A return of {@code true} reported a sufficient |
1924 |
|
* period after shutdown may indicate that submitted tasks have |
1925 |
< |
* ignored or suppressed interruption, causing this executor not |
1926 |
< |
* to properly terminate. |
1925 |
> |
* ignored or suppressed interruption, or are waiting for IO, |
1926 |
> |
* causing this executor not to properly terminate. (See the |
1927 |
> |
* advisory notes for class {@link ForkJoinTask} stating that |
1928 |
> |
* tasks should not normally entail blocking operations. But if |
1929 |
> |
* they do, they must abort them on interrupt.) |
1930 |
|
* |
1931 |
|
* @return {@code true} if terminating but not yet terminated |
1932 |
|
*/ |
1933 |
|
public boolean isTerminating() { |
1934 |
< |
return (runState & (TERMINATING|TERMINATED)) == TERMINATING; |
1934 |
> |
long c = ctl; |
1935 |
> |
return ((c & STOP_BIT) != 0L && |
1936 |
> |
(short)(c >>> TC_SHIFT) != -parallelism); |
1937 |
> |
} |
1938 |
> |
|
1939 |
> |
/** |
1940 |
> |
* Returns true if terminating or terminated. Used by ForkJoinWorkerThread. |
1941 |
> |
*/ |
1942 |
> |
final boolean isAtLeastTerminating() { |
1943 |
> |
return (ctl & STOP_BIT) != 0L; |
1944 |
|
} |
1945 |
|
|
1946 |
|
/** |
1949 |
|
* @return {@code true} if this pool has been shut down |
1950 |
|
*/ |
1951 |
|
public boolean isShutdown() { |
1952 |
< |
return runState >= SHUTDOWN; |
1952 |
> |
return shutdown; |
1953 |
|
} |
1954 |
|
|
1955 |
|
/** |
1965 |
|
*/ |
1966 |
|
public boolean awaitTermination(long timeout, TimeUnit unit) |
1967 |
|
throws InterruptedException { |
1968 |
+ |
long nanos = unit.toNanos(timeout); |
1969 |
+ |
final ReentrantLock lock = this.submissionLock; |
1970 |
+ |
lock.lock(); |
1971 |
|
try { |
1972 |
< |
return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0; |
1973 |
< |
} catch(TimeoutException ex) { |
1974 |
< |
return false; |
1972 |
> |
for (;;) { |
1973 |
> |
if (isTerminated()) |
1974 |
> |
return true; |
1975 |
> |
if (nanos <= 0) |
1976 |
> |
return false; |
1977 |
> |
nanos = termination.awaitNanos(nanos); |
1978 |
> |
} |
1979 |
> |
} finally { |
1980 |
> |
lock.unlock(); |
1981 |
|
} |
1982 |
|
} |
1983 |
|
|
1989 |
|
* {@code isReleasable} must return {@code true} if blocking is |
1990 |
|
* not necessary. Method {@code block} blocks the current thread |
1991 |
|
* if necessary (perhaps internally invoking {@code isReleasable} |
1992 |
< |
* before actually blocking). The unusual methods in this API |
1993 |
< |
* accommodate synchronizers that may, but don't usually, block |
1994 |
< |
* for long periods. Similarly, they allow more efficient internal |
1995 |
< |
* handling of cases in which additional workers may be, but |
1996 |
< |
* usually are not, needed to ensure sufficient parallelism. |
1997 |
< |
* Toward this end, implementations of method {@code isReleasable} |
1998 |
< |
* must be amenable to repeated invocation. |
1992 |
> |
* before actually blocking). These actions are performed by any |
1993 |
> |
* thread invoking {@link ForkJoinPool#managedBlock}. The |
1994 |
> |
* unusual methods in this API accommodate synchronizers that may, |
1995 |
> |
* but don't usually, block for long periods. Similarly, they |
1996 |
> |
* allow more efficient internal handling of cases in which |
1997 |
> |
* additional workers may be, but usually are not, needed to |
1998 |
> |
* ensure sufficient parallelism. Toward this end, |
1999 |
> |
* implementations of method {@code isReleasable} must be amenable |
2000 |
> |
* to repeated invocation. |
2001 |
|
* |
2002 |
|
* <p>For example, here is a ManagedBlocker based on a |
2003 |
|
* ReentrantLock: |
2025 |
|
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
2026 |
|
* public boolean block() throws InterruptedException { |
2027 |
|
* if (item == null) |
2028 |
< |
* item = queue.take |
2028 |
> |
* item = queue.take(); |
2029 |
|
* return true; |
2030 |
|
* } |
2031 |
|
* public boolean isReleasable() { |
2032 |
< |
* return item != null || (item = queue.poll) != null; |
2032 |
> |
* return item != null || (item = queue.poll()) != null; |
2033 |
|
* } |
2034 |
|
* public E getItem() { // call after pool.managedBlock completes |
2035 |
|
* return item; |
2099 |
|
} |
2100 |
|
|
2101 |
|
// Unsafe mechanics |
2102 |
< |
|
2103 |
< |
private static final sun.misc.Unsafe UNSAFE = getUnsafe(); |
2104 |
< |
private static final long workerCountsOffset = |
2105 |
< |
objectFieldOffset("workerCounts", ForkJoinPool.class); |
2106 |
< |
private static final long runStateOffset = |
2107 |
< |
objectFieldOffset("runState", ForkJoinPool.class); |
2108 |
< |
private static final long eventCountOffset = |
2109 |
< |
objectFieldOffset("eventCount", ForkJoinPool.class); |
2110 |
< |
private static final long eventWaitersOffset = |
2111 |
< |
objectFieldOffset("eventWaiters",ForkJoinPool.class); |
2112 |
< |
private static final long stealCountOffset = |
2113 |
< |
objectFieldOffset("stealCount",ForkJoinPool.class); |
2114 |
< |
private static final long spareWaitersOffset = |
2115 |
< |
objectFieldOffset("spareWaiters",ForkJoinPool.class); |
2116 |
< |
|
2117 |
< |
private static long objectFieldOffset(String field, Class<?> klazz) { |
2102 |
> |
private static final sun.misc.Unsafe UNSAFE; |
2103 |
> |
private static final long ctlOffset; |
2104 |
> |
private static final long stealCountOffset; |
2105 |
> |
private static final long blockedCountOffset; |
2106 |
> |
private static final long quiescerCountOffset; |
2107 |
> |
private static final long scanGuardOffset; |
2108 |
> |
private static final long nextWorkerNumberOffset; |
2109 |
> |
private static final long ABASE; |
2110 |
> |
private static final int ASHIFT; |
2111 |
> |
|
2112 |
> |
static { |
2113 |
> |
poolNumberGenerator = new AtomicInteger(); |
2114 |
> |
workerSeedGenerator = new Random(); |
2115 |
> |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
2116 |
> |
defaultForkJoinWorkerThreadFactory = |
2117 |
> |
new DefaultForkJoinWorkerThreadFactory(); |
2118 |
|
try { |
2119 |
< |
return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field)); |
2120 |
< |
} catch (NoSuchFieldException e) { |
2121 |
< |
// Convert Exception to corresponding Error |
2122 |
< |
NoSuchFieldError error = new NoSuchFieldError(field); |
2123 |
< |
error.initCause(e); |
2124 |
< |
throw error; |
2125 |
< |
} |
2119 |
> |
UNSAFE = getUnsafe(); |
2120 |
> |
Class<?> k = ForkJoinPool.class; |
2121 |
> |
ctlOffset = UNSAFE.objectFieldOffset |
2122 |
> |
(k.getDeclaredField("ctl")); |
2123 |
> |
stealCountOffset = UNSAFE.objectFieldOffset |
2124 |
> |
(k.getDeclaredField("stealCount")); |
2125 |
> |
blockedCountOffset = UNSAFE.objectFieldOffset |
2126 |
> |
(k.getDeclaredField("blockedCount")); |
2127 |
> |
quiescerCountOffset = UNSAFE.objectFieldOffset |
2128 |
> |
(k.getDeclaredField("quiescerCount")); |
2129 |
> |
scanGuardOffset = UNSAFE.objectFieldOffset |
2130 |
> |
(k.getDeclaredField("scanGuard")); |
2131 |
> |
nextWorkerNumberOffset = UNSAFE.objectFieldOffset |
2132 |
> |
(k.getDeclaredField("nextWorkerNumber")); |
2133 |
> |
} catch (Exception e) { |
2134 |
> |
throw new Error(e); |
2135 |
> |
} |
2136 |
> |
Class<?> a = ForkJoinTask[].class; |
2137 |
> |
ABASE = UNSAFE.arrayBaseOffset(a); |
2138 |
> |
int s = UNSAFE.arrayIndexScale(a); |
2139 |
> |
if ((s & (s-1)) != 0) |
2140 |
> |
throw new Error("data type scale not a power of two"); |
2141 |
> |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
2142 |
|
} |
2143 |
|
|
2144 |
|
/** |