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; |
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; |
23 |
|
import java.util.concurrent.atomic.AtomicInteger; |
24 |
|
import java.util.concurrent.locks.LockSupport; |
25 |
|
import java.util.concurrent.locks.ReentrantLock; |
26 |
+ |
import java.util.concurrent.locks.Condition; |
27 |
|
|
28 |
|
/** |
29 |
|
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
131 |
|
* set of worker threads: Submissions from non-FJ threads enter |
132 |
|
* into a submission queue. Workers take these tasks and typically |
133 |
|
* split them into subtasks that may be stolen by other workers. |
134 |
< |
* The main work-stealing mechanics implemented in class |
135 |
< |
* ForkJoinWorkerThread give first priority to processing tasks |
136 |
< |
* from their own queues (LIFO or FIFO, depending on mode), then |
137 |
< |
* to randomized FIFO steals of tasks in other worker queues, and |
138 |
< |
* lastly to new submissions. These mechanics do not consider |
139 |
< |
* affinities, loads, cache localities, etc, so rarely provide the |
140 |
< |
* best possible performance on a given machine, but portably |
141 |
< |
* provide good throughput by averaging over these factors. |
142 |
< |
* (Further, even if we did try to use such information, we do not |
143 |
< |
* usually have a basis for exploiting it. For example, some sets |
144 |
< |
* of tasks profit from cache affinities, but others are harmed by |
145 |
< |
* cache pollution effects.) |
146 |
< |
* |
147 |
< |
* Beyond work-stealing support and essential bookkeeping, the |
148 |
< |
* main responsibility of this framework is to take actions when |
149 |
< |
* one worker is waiting to join a task stolen (or always held by) |
150 |
< |
* another. Because we are multiplexing many tasks on to a pool |
151 |
< |
* of workers, we can't just let them block (as in Thread.join). |
152 |
< |
* We also cannot just reassign the joiner's run-time stack with |
153 |
< |
* another and replace it later, which would be a form of |
154 |
< |
* "continuation", that even if possible is not necessarily a good |
155 |
< |
* idea. Given that the creation costs of most threads on most |
156 |
< |
* systems mainly surrounds setting up runtime stacks, thread |
157 |
< |
* creation and switching is usually not much more expensive than |
158 |
< |
* stack creation and switching, and is more flexible). Instead we |
134 |
> |
* Preference rules give first priority to processing tasks from |
135 |
> |
* their own queues (LIFO or FIFO, depending on mode), then to |
136 |
> |
* randomized FIFO steals of tasks in other worker queues, and |
137 |
> |
* lastly to new submissions. |
138 |
> |
* |
139 |
> |
* The main throughput advantages of work-stealing stem from |
140 |
> |
* decentralized control -- workers mostly take tasks from |
141 |
> |
* themselves or each other. We cannot negate this in the |
142 |
> |
* implementation of other management responsibilities. The main |
143 |
> |
* tactic for avoiding bottlenecks is packing nearly all |
144 |
> |
* essentially atomic control state into a single 64bit volatile |
145 |
> |
* variable ("ctl"). This variable is read on the order of 10-100 |
146 |
> |
* times as often as it is modified (always via CAS). (There is |
147 |
> |
* some additional control state, for example variable "shutdown" |
148 |
> |
* for which we can cope with uncoordinated updates.) This |
149 |
> |
* streamlines synchronization and control at the expense of messy |
150 |
> |
* constructions needed to repack status bits upon updates. |
151 |
> |
* Updates tend not to contend with each other except during |
152 |
> |
* bursts while submitted tasks begin or end. In some cases when |
153 |
> |
* they do contend, threads can instead do something else |
154 |
> |
* (usually, scan for tasks) until contention subsides. |
155 |
> |
* |
156 |
> |
* To enable packing, we restrict maximum parallelism to (1<<15)-1 |
157 |
> |
* (which is far in excess of normal operating range) to allow |
158 |
> |
* ids, counts, and their negations (used for thresholding) to fit |
159 |
> |
* into 16bit fields. |
160 |
> |
* |
161 |
> |
* Recording Workers. Workers are recorded in the "workers" array |
162 |
> |
* that is created upon pool construction and expanded if (rarely) |
163 |
> |
* necessary. This is an array as opposed to some other data |
164 |
> |
* structure to support index-based random steals by workers. |
165 |
> |
* Updates to the array recording new workers and unrecording |
166 |
> |
* terminated ones are protected from each other by a seqLock |
167 |
> |
* (scanGuard) but the array is otherwise concurrently readable, |
168 |
> |
* and accessed directly by workers. To simplify index-based |
169 |
> |
* operations, the array size is always a power of two, and all |
170 |
> |
* readers must tolerate null slots. To avoid flailing during |
171 |
> |
* start-up, the array is presized to hold twice #parallelism |
172 |
> |
* workers (which is unlikely to need further resizing during |
173 |
> |
* execution). But to avoid dealing with so many null slots, |
174 |
> |
* variable scanGuard includes a mask for the nearest power of two |
175 |
> |
* that contains all current workers. All worker thread creation |
176 |
> |
* is on-demand, triggered by task submissions, replacement of |
177 |
> |
* terminated workers, and/or compensation for blocked |
178 |
> |
* workers. However, all other support code is set up to work with |
179 |
> |
* other policies. To ensure that we do not hold on to worker |
180 |
> |
* references that would prevent GC, ALL accesses to workers are |
181 |
> |
* via indices into the workers array (which is one source of some |
182 |
> |
* of the messy code constructions here). In essence, the workers |
183 |
> |
* array serves as a weak reference mechanism. Thus for example |
184 |
> |
* the wait queue field of ctl stores worker indices, not worker |
185 |
> |
* references. Access to the workers in associated methods (for |
186 |
> |
* example signalWork) must both index-check and null-check the |
187 |
> |
* IDs. All such accesses ignore bad IDs by returning out early |
188 |
> |
* from what they are doing, since this can only be associated |
189 |
> |
* with termination, in which case it is OK to give up. |
190 |
> |
* |
191 |
> |
* All uses of the workers array, as well as queue arrays, check |
192 |
> |
* that the array is non-null (even if previously non-null). This |
193 |
> |
* allows nulling during termination, which is currently not |
194 |
> |
* necessary, but remains an option for resource-revocation-based |
195 |
> |
* shutdown schemes. |
196 |
> |
* |
197 |
> |
* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot |
198 |
> |
* let workers spin indefinitely scanning for tasks when none can |
199 |
> |
* be found immediately, and we cannot start/resume workers unless |
200 |
> |
* there appear to be tasks available. On the other hand, we must |
201 |
> |
* quickly prod them into action when new tasks are submitted or |
202 |
> |
* generated. We park/unpark workers after placing in an event |
203 |
> |
* wait queue when they cannot find work. This "queue" is actually |
204 |
> |
* a simple Treiber stack, headed by the "id" field of ctl, plus a |
205 |
> |
* 15bit counter value to both wake up waiters (by advancing their |
206 |
> |
* count) and avoid ABA effects. Successors are held in worker |
207 |
> |
* field "nextWait". Queuing deals with several intrinsic races, |
208 |
> |
* mainly that a task-producing thread can miss seeing (and |
209 |
> |
* signalling) another thread that gave up looking for work but |
210 |
> |
* has not yet entered the wait queue. We solve this by requiring |
211 |
> |
* a full sweep of all workers both before (in scan()) and after |
212 |
> |
* (in tryAwaitWork()) a newly waiting worker is added to the wait |
213 |
> |
* queue. During a rescan, the worker might release some other |
214 |
> |
* queued worker rather than itself, which has the same net |
215 |
> |
* effect. Because enqueued workers may actually be rescanning |
216 |
> |
* rather than waiting, we set and clear the "parked" field of |
217 |
> |
* ForkJoinWorkerThread to reduce unnecessary calls to unpark. |
218 |
> |
* (Use of the parked field requires a secondary recheck to avoid |
219 |
> |
* missed signals.) |
220 |
> |
* |
221 |
> |
* Signalling. We create or wake up workers only when there |
222 |
> |
* appears to be at least one task they might be able to find and |
223 |
> |
* execute. When a submission is added or another worker adds a |
224 |
> |
* task to a queue that previously had two or fewer tasks, they |
225 |
> |
* signal waiting workers (or trigger creation of new ones if |
226 |
> |
* fewer than the given parallelism level -- see signalWork). |
227 |
> |
* These primary signals are buttressed by signals during rescans |
228 |
> |
* as well as those performed when a worker steals a task and |
229 |
> |
* notices that there are more tasks too; together these cover the |
230 |
> |
* signals needed in cases when more than two tasks are pushed |
231 |
> |
* but untaken. |
232 |
> |
* |
233 |
> |
* Trimming workers. To release resources after periods of lack of |
234 |
> |
* use, a worker starting to wait when the pool is quiescent will |
235 |
> |
* time out and terminate if the pool has remained quiescent for |
236 |
> |
* SHRINK_RATE nanosecs. This will slowly propagate, eventually |
237 |
> |
* terminating all workers after long periods of non-use. |
238 |
> |
* |
239 |
> |
* Submissions. External submissions are maintained in an |
240 |
> |
* array-based queue that is structured identically to |
241 |
> |
* ForkJoinWorkerThread queues except for the use of |
242 |
> |
* submissionLock in method addSubmission. Unlike the case for |
243 |
> |
* worker queues, multiple external threads can add new |
244 |
> |
* submissions, so adding requires a lock. |
245 |
> |
* |
246 |
> |
* Compensation. Beyond work-stealing support and lifecycle |
247 |
> |
* control, the main responsibility of this framework is to take |
248 |
> |
* actions when one worker is waiting to join a task stolen (or |
249 |
> |
* always held by) another. Because we are multiplexing many |
250 |
> |
* tasks on to a pool of workers, we can't just let them block (as |
251 |
> |
* in Thread.join). We also cannot just reassign the joiner's |
252 |
> |
* run-time stack with another and replace it later, which would |
253 |
> |
* be a form of "continuation", that even if possible is not |
254 |
> |
* necessarily a good idea since we sometimes need both an |
255 |
> |
* unblocked task and its continuation to progress. Instead we |
256 |
|
* combine two tactics: |
257 |
|
* |
258 |
|
* Helping: Arranging for the joiner to execute some task that it |
259 |
|
* would be running if the steal had not occurred. Method |
260 |
< |
* ForkJoinWorkerThread.helpJoinTask tracks joining->stealing |
260 |
> |
* ForkJoinWorkerThread.joinTask tracks joining->stealing |
261 |
|
* links to try to find such a task. |
262 |
|
* |
263 |
|
* Compensating: Unless there are already enough live threads, |
264 |
< |
* method helpMaintainParallelism() may create or |
265 |
< |
* re-activate a spare thread to compensate for blocked |
266 |
< |
* joiners until they unblock. |
168 |
< |
* |
169 |
< |
* It is impossible to keep exactly the target (parallelism) |
170 |
< |
* number of threads running at any given time. Determining |
171 |
< |
* existence of conservatively safe helping targets, the |
172 |
< |
* availability of already-created spares, and the apparent need |
173 |
< |
* to create new spares are all racy and require heuristic |
174 |
< |
* guidance, so we rely on multiple retries of each. Compensation |
175 |
< |
* occurs in slow-motion. It is triggered only upon timeouts of |
176 |
< |
* Object.wait used for joins. This reduces poor decisions that |
177 |
< |
* would otherwise be made when threads are waiting for others |
178 |
< |
* that are stalled because of unrelated activities such as |
179 |
< |
* garbage collection. |
264 |
> |
* method tryPreBlock() may create or re-activate a spare |
265 |
> |
* thread to compensate for blocked joiners until they |
266 |
> |
* unblock. |
267 |
|
* |
268 |
|
* The ManagedBlocker extension API can't use helping so relies |
269 |
|
* only on compensation in method awaitBlocker. |
270 |
|
* |
271 |
< |
* The main throughput advantages of work-stealing stem from |
272 |
< |
* decentralized control -- workers mostly steal tasks from each |
273 |
< |
* other. We do not want to negate this by creating bottlenecks |
274 |
< |
* implementing other management responsibilities. So we use a |
275 |
< |
* collection of techniques that avoid, reduce, or cope well with |
276 |
< |
* contention. These entail several instances of bit-packing into |
277 |
< |
* CASable fields to maintain only the minimally required |
278 |
< |
* atomicity. To enable such packing, we restrict maximum |
279 |
< |
* parallelism to (1<<15)-1 (enabling twice this (to accommodate |
280 |
< |
* unbalanced increments and decrements) to fit into a 16 bit |
281 |
< |
* field, which is far in excess of normal operating range. Even |
282 |
< |
* though updates to some of these bookkeeping fields do sometimes |
283 |
< |
* contend with each other, they don't normally cache-contend with |
284 |
< |
* updates to others enough to warrant memory padding or |
285 |
< |
* isolation. So they are all held as fields of ForkJoinPool |
286 |
< |
* objects. The main capabilities are as follows: |
287 |
< |
* |
288 |
< |
* 1. Creating and removing workers. Workers are recorded in the |
289 |
< |
* "workers" array. This is an array as opposed to some other data |
290 |
< |
* structure to support index-based random steals by workers. |
291 |
< |
* Updates to the array recording new workers and unrecording |
292 |
< |
* terminated ones are protected from each other by a lock |
293 |
< |
* (workerLock) but the array is otherwise concurrently readable, |
294 |
< |
* and accessed directly by workers. To simplify index-based |
295 |
< |
* operations, the array size is always a power of two, and all |
296 |
< |
* readers must tolerate null slots. Currently, all worker thread |
297 |
< |
* creation is on-demand, triggered by task submissions, |
298 |
< |
* replacement of terminated workers, and/or compensation for |
299 |
< |
* blocked workers. However, all other support code is set up to |
213 |
< |
* work with other policies. |
214 |
< |
* |
215 |
< |
* To ensure that we do not hold on to worker references that |
216 |
< |
* would prevent GC, ALL accesses to workers are via indices into |
217 |
< |
* the workers array (which is one source of some of the unusual |
218 |
< |
* code constructions here). In essence, the workers array serves |
219 |
< |
* as a WeakReference mechanism. Thus for example the event queue |
220 |
< |
* stores worker indices, not worker references. Access to the |
221 |
< |
* workers in associated methods (for example releaseEventWaiters) |
222 |
< |
* must both index-check and null-check the IDs. All such accesses |
223 |
< |
* ignore bad IDs by returning out early from what they are doing, |
224 |
< |
* since this can only be associated with shutdown, in which case |
225 |
< |
* it is OK to give up. On termination, we just clobber these |
226 |
< |
* data structures without trying to use them. |
227 |
< |
* |
228 |
< |
* 2. Bookkeeping for dynamically adding and removing workers. We |
229 |
< |
* aim to approximately maintain the given level of parallelism. |
230 |
< |
* When some workers are known to be blocked (on joins or via |
231 |
< |
* ManagedBlocker), we may create or resume others to take their |
232 |
< |
* place until they unblock (see below). Implementing this |
233 |
< |
* requires counts of the number of "running" threads (i.e., those |
234 |
< |
* that are neither blocked nor artificially suspended) as well as |
235 |
< |
* the total number. These two values are packed into one field, |
236 |
< |
* "workerCounts" because we need accurate snapshots when deciding |
237 |
< |
* to create, resume or suspend. Note however that the |
238 |
< |
* correspondence of these counts to reality is not guaranteed. In |
239 |
< |
* particular updates for unblocked threads may lag until they |
240 |
< |
* actually wake up. |
241 |
< |
* |
242 |
< |
* 3. Maintaining global run state. The run state of the pool |
243 |
< |
* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to |
244 |
< |
* those in other Executor implementations, as well as a count of |
245 |
< |
* "active" workers -- those that are, or soon will be, or |
246 |
< |
* recently were executing tasks. The runLevel and active count |
247 |
< |
* are packed together in order to correctly trigger shutdown and |
248 |
< |
* termination. Without care, active counts can be subject to very |
249 |
< |
* high contention. We substantially reduce this contention by |
250 |
< |
* relaxing update rules. A worker must claim active status |
251 |
< |
* prospectively, by activating if it sees that a submitted or |
252 |
< |
* stealable task exists (it may find after activating that the |
253 |
< |
* task no longer exists). It stays active while processing this |
254 |
< |
* task (if it exists) and any other local subtasks it produces, |
255 |
< |
* until it cannot find any other tasks. It then tries |
256 |
< |
* inactivating (see method preStep), but upon update contention |
257 |
< |
* instead scans for more tasks, later retrying inactivation if it |
258 |
< |
* doesn't find any. |
259 |
< |
* |
260 |
< |
* 4. Managing idle workers waiting for tasks. We cannot let |
261 |
< |
* workers spin indefinitely scanning for tasks when none are |
262 |
< |
* available. On the other hand, we must quickly prod them into |
263 |
< |
* action when new tasks are submitted or generated. We |
264 |
< |
* park/unpark these idle workers using an event-count scheme. |
265 |
< |
* Field eventCount is incremented upon events that may enable |
266 |
< |
* workers that previously could not find a task to now find one: |
267 |
< |
* Submission of a new task to the pool, or another worker pushing |
268 |
< |
* a task onto a previously empty queue. (We also use this |
269 |
< |
* mechanism for configuration and termination actions that |
270 |
< |
* require wakeups of idle workers). Each worker maintains its |
271 |
< |
* last known event count, and blocks when a scan for work did not |
272 |
< |
* find a task AND its lastEventCount matches the current |
273 |
< |
* eventCount. Waiting idle workers are recorded in a variant of |
274 |
< |
* Treiber stack headed by field eventWaiters which, when nonzero, |
275 |
< |
* encodes the thread index and count awaited for by the worker |
276 |
< |
* thread most recently calling eventSync. This thread in turn has |
277 |
< |
* a record (field nextEventWaiter) for the next waiting worker. |
278 |
< |
* In addition to allowing simpler decisions about need for |
279 |
< |
* wakeup, the event count bits in eventWaiters serve the role of |
280 |
< |
* tags to avoid ABA errors in Treiber stacks. Upon any wakeup, |
281 |
< |
* released threads also try to release at most two others. The |
282 |
< |
* net effect is a tree-like diffusion of signals, where released |
283 |
< |
* threads (and possibly others) help with unparks. To further |
284 |
< |
* reduce contention effects a bit, failed CASes to increment |
285 |
< |
* field eventCount are tolerated without retries in signalWork. |
286 |
< |
* Conceptually they are merged into the same event, which is OK |
287 |
< |
* when their only purpose is to enable workers to scan for work. |
288 |
< |
* |
289 |
< |
* 5. Managing suspension of extra workers. When a worker notices |
290 |
< |
* (usually upon timeout of a wait()) that there are too few |
291 |
< |
* running threads, we may create a new thread to maintain |
292 |
< |
* parallelism level, or at least avoid starvation. Usually, extra |
293 |
< |
* threads are needed for only very short periods, yet join |
294 |
< |
* dependencies are such that we sometimes need them in |
295 |
< |
* bursts. Rather than create new threads each time this happens, |
296 |
< |
* we suspend no-longer-needed extra ones as "spares". For most |
297 |
< |
* purposes, we don't distinguish "extra" spare threads from |
298 |
< |
* normal "core" threads: On each call to preStep (the only point |
299 |
< |
* at which we can do this) a worker checks to see if there are |
300 |
< |
* now too many running workers, and if so, suspends itself. |
301 |
< |
* Method helpMaintainParallelism looks for suspended threads to |
302 |
< |
* resume before considering creating a new replacement. The |
303 |
< |
* spares themselves are encoded on another variant of a Treiber |
304 |
< |
* Stack, headed at field "spareWaiters". Note that the use of |
305 |
< |
* spares is intrinsically racy. One thread may become a spare at |
306 |
< |
* about the same time as another is needlessly being created. We |
307 |
< |
* counteract this and related slop in part by requiring resumed |
308 |
< |
* spares to immediately recheck (in preStep) to see whether they |
309 |
< |
* should re-suspend. |
310 |
< |
* |
311 |
< |
* 6. Killing off unneeded workers. A timeout mechanism is used to |
312 |
< |
* shed unused workers: The oldest (first) event queue waiter uses |
313 |
< |
* a timed rather than hard wait. When this wait times out without |
314 |
< |
* a normal wakeup, it tries to shutdown any one (for convenience |
315 |
< |
* the newest) other spare or event waiter via |
316 |
< |
* tryShutdownUnusedWorker. This eventually reduces the number of |
317 |
< |
* worker threads to a minimum of one after a long enough period |
318 |
< |
* without use. |
319 |
< |
* |
320 |
< |
* 7. Deciding when to create new workers. The main dynamic |
321 |
< |
* control in this class is deciding when to create extra threads |
322 |
< |
* in method helpMaintainParallelism. We would like to keep |
323 |
< |
* exactly #parallelism threads running, which is an impossible |
324 |
< |
* task. We always need to create one when the number of running |
325 |
< |
* threads would become zero and all workers are busy. Beyond |
326 |
< |
* this, we must rely on heuristics that work well in the |
327 |
< |
* presence of transient phenomena such as GC stalls, dynamic |
328 |
< |
* compilation, and wake-up lags. These transients are extremely |
329 |
< |
* common -- we are normally trying to fully saturate the CPUs on |
330 |
< |
* a machine, so almost any activity other than running tasks |
331 |
< |
* impedes accuracy. Our main defense is to allow parallelism to |
332 |
< |
* lapse for a while during joins, and use a timeout to see if, |
333 |
< |
* after the resulting settling, there is still a need for |
334 |
< |
* additional workers. This also better copes with the fact that |
335 |
< |
* some of the methods in this class tend to never become compiled |
336 |
< |
* (but are interpreted), so some components of the entire set of |
337 |
< |
* controls might execute 100 times faster than others. And |
338 |
< |
* similarly for cases where the apparent lack of work is just due |
339 |
< |
* to GC stalls and other transient system activity. |
271 |
> |
* It is impossible to keep exactly the target parallelism number |
272 |
> |
* of threads running at any given time. Determining the |
273 |
> |
* existence of conservatively safe helping targets, the |
274 |
> |
* availability of already-created spares, and the apparent need |
275 |
> |
* to create new spares are all racy and require heuristic |
276 |
> |
* guidance, so we rely on multiple retries of each. Currently, |
277 |
> |
* in keeping with on-demand signalling policy, we compensate only |
278 |
> |
* if blocking would leave less than one active (non-waiting, |
279 |
> |
* non-blocked) worker. Additionally, to avoid some false alarms |
280 |
> |
* due to GC, lagging counters, system activity, etc, compensated |
281 |
> |
* blocking for joins is only attempted after rechecks stabilize |
282 |
> |
* (retries are interspersed with Thread.yield, for good |
283 |
> |
* citizenship). The variable blockedCount, incremented before |
284 |
> |
* blocking and decremented after, is sometimes needed to |
285 |
> |
* distinguish cases of waiting for work vs blocking on joins or |
286 |
> |
* other managed sync. Both cases are equivalent for most pool |
287 |
> |
* control, so we can update non-atomically. (Additionally, |
288 |
> |
* contention on blockedCount alleviates some contention on ctl). |
289 |
> |
* |
290 |
> |
* Shutdown and Termination. A call to shutdownNow atomically sets |
291 |
> |
* the ctl stop bit and then (non-atomically) sets each workers |
292 |
> |
* "terminate" status, cancels all unprocessed tasks, and wakes up |
293 |
> |
* all waiting workers. Detecting whether termination should |
294 |
> |
* commence after a non-abrupt shutdown() call requires more work |
295 |
> |
* and bookkeeping. We need consensus about quiesence (i.e., that |
296 |
> |
* there is no more work) which is reflected in active counts so |
297 |
> |
* long as there are no current blockers, as well as possible |
298 |
> |
* re-evaluations during independent changes in blocking or |
299 |
> |
* quiescing workers. |
300 |
|
* |
301 |
< |
* Beware that there is a lot of representation-level coupling |
301 |
> |
* Style notes: There is a lot of representation-level coupling |
302 |
|
* among classes ForkJoinPool, ForkJoinWorkerThread, and |
303 |
< |
* ForkJoinTask. For example, direct access to "workers" array by |
303 |
> |
* ForkJoinTask. Most fields of ForkJoinWorkerThread maintain |
304 |
> |
* data structures managed by ForkJoinPool, so are directly |
305 |
> |
* accessed. Conversely we allow access to "workers" array by |
306 |
|
* workers, and direct access to ForkJoinTask.status by both |
307 |
|
* ForkJoinPool and ForkJoinWorkerThread. There is little point |
308 |
|
* trying to reduce this, since any associated future changes in |
309 |
|
* representations will need to be accompanied by algorithmic |
310 |
< |
* changes anyway. |
311 |
< |
* |
312 |
< |
* Style notes: There are lots of inline assignments (of form |
313 |
< |
* "while ((local = field) != 0)") which are usually the simplest |
314 |
< |
* way to ensure the required read orderings (which are sometimes |
315 |
< |
* critical). Also several occurrences of the unusual "do {} |
316 |
< |
* while (!cas...)" which is the simplest way to force an update of |
317 |
< |
* a CAS'ed variable. There are also other coding oddities that |
318 |
< |
* help some methods perform reasonably even when interpreted (not |
319 |
< |
* compiled), at the expense of some messy constructions that |
320 |
< |
* reduce byte code counts. |
321 |
< |
* |
322 |
< |
* The order of declarations in this file is: (1) statics (2) |
323 |
< |
* fields (along with constants used when unpacking some of them) |
324 |
< |
* (3) internal control methods (4) callbacks and other support |
325 |
< |
* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported |
326 |
< |
* methods (plus a few little helpers). |
310 |
> |
* changes anyway. All together, these low-level implementation |
311 |
> |
* choices produce as much as a factor of 4 performance |
312 |
> |
* improvement compared to naive implementations, and enable the |
313 |
> |
* processing of billions of tasks per second, at the expense of |
314 |
> |
* some ugliness. |
315 |
> |
* |
316 |
> |
* Methods signalWork() and scan() are the main bottlenecks so are |
317 |
> |
* especially heavily micro-optimized/mangled. There are lots of |
318 |
> |
* inline assignments (of form "while ((local = field) != 0)") |
319 |
> |
* which are usually the simplest way to ensure the required read |
320 |
> |
* orderings (which are sometimes critical). This leads to a |
321 |
> |
* "C"-like style of listing declarations of these locals at the |
322 |
> |
* heads of methods or blocks. There are several occurrences of |
323 |
> |
* the unusual "do {} while (!cas...)" which is the simplest way |
324 |
> |
* to force an update of a CAS'ed variable. There are also other |
325 |
> |
* coding oddities that help some methods perform reasonably even |
326 |
> |
* when interpreted (not compiled). |
327 |
> |
* |
328 |
> |
* The order of declarations in this file is: (1) declarations of |
329 |
> |
* statics (2) fields (along with constants used when unpacking |
330 |
> |
* some of them), listed in an order that tends to reduce |
331 |
> |
* contention among them a bit under most JVMs. (3) internal |
332 |
> |
* control methods (4) callbacks and other support for |
333 |
> |
* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported |
334 |
> |
* methods (plus a few little helpers). (6) static block |
335 |
> |
* initializing all statics in a minimally dependent order. |
336 |
|
*/ |
337 |
|
|
338 |
|
/** |
367 |
|
* overridden in ForkJoinPool constructors. |
368 |
|
*/ |
369 |
|
public static final ForkJoinWorkerThreadFactory |
370 |
< |
defaultForkJoinWorkerThreadFactory = |
400 |
< |
new DefaultForkJoinWorkerThreadFactory(); |
370 |
> |
defaultForkJoinWorkerThreadFactory; |
371 |
|
|
372 |
|
/** |
373 |
|
* Permission required for callers of methods that may start or |
374 |
|
* kill threads. |
375 |
|
*/ |
376 |
< |
private static final RuntimePermission modifyThreadPermission = |
407 |
< |
new RuntimePermission("modifyThread"); |
376 |
> |
private static final RuntimePermission modifyThreadPermission; |
377 |
|
|
378 |
|
/** |
379 |
|
* If there is a security manager, makes sure caller has |
388 |
|
/** |
389 |
|
* Generator for assigning sequence numbers as pool names. |
390 |
|
*/ |
391 |
< |
private static final AtomicInteger poolNumberGenerator = |
423 |
< |
new AtomicInteger(); |
391 |
> |
private static final AtomicInteger poolNumberGenerator; |
392 |
|
|
393 |
|
/** |
394 |
< |
* The time to block in a join (see awaitJoin) before checking if |
395 |
< |
* a new worker should be (re)started to maintain parallelism |
396 |
< |
* level. The value should be short enough to maintain global |
397 |
< |
* responsiveness and progress but long enough to avoid |
398 |
< |
* counterproductive firings during GC stalls or unrelated system |
431 |
< |
* activity, and to not bog down systems with continual re-firings |
432 |
< |
* on GCs or legitimately long waits. |
394 |
> |
* Generator for initial random seeds for worker victim |
395 |
> |
* selection. This is used only to create initial seeds. Random |
396 |
> |
* steals use a cheaper xorshift generator per steal attempt. We |
397 |
> |
* don't expect much contention on seedGenerator, so just use a |
398 |
> |
* plain Random. |
399 |
|
*/ |
400 |
< |
private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second |
400 |
> |
static final Random workerSeedGenerator; |
401 |
|
|
402 |
|
/** |
403 |
< |
* The wakeup interval (in nanoseconds) for the oldest worker |
404 |
< |
* waiting for an event to invoke tryShutdownUnusedWorker to |
405 |
< |
* shrink the number of workers. The exact value does not matter |
406 |
< |
* too much. It must be short enough to release resources during |
407 |
< |
* sustained periods of idleness, but not so short that threads |
408 |
< |
* are continually re-created. |
403 |
> |
* Array holding all worker threads in the pool. Initialized upon |
404 |
> |
* construction. Array size must be a power of two. Updates and |
405 |
> |
* replacements are protected by scanGuard, but the array is |
406 |
> |
* always kept in a consistent enough state to be randomly |
407 |
> |
* accessed without locking by workers performing work-stealing, |
408 |
> |
* as well as other traversal-based methods in this class, so long |
409 |
> |
* as reads memory-acquire by first reading ctl. All readers must |
410 |
> |
* tolerate that some array slots may be null. |
411 |
|
*/ |
412 |
< |
private static final long SHRINK_RATE_NANOS = |
445 |
< |
30L * 1000L * 1000L * 1000L; // 2 per minute |
412 |
> |
ForkJoinWorkerThread[] workers; |
413 |
|
|
414 |
|
/** |
415 |
< |
* Absolute bound for parallelism level. Twice this number plus |
416 |
< |
* one (i.e., 0xfff) must fit into a 16bit field to enable |
417 |
< |
* word-packing for some counts and indices. |
415 |
> |
* Initial size for submission queue array. Must be a power of |
416 |
> |
* two. In many applications, these always stay small so we use a |
417 |
> |
* small initial cap. |
418 |
|
*/ |
419 |
< |
private static final int MAX_WORKERS = 0x7fff; |
419 |
> |
private static final int INITIAL_QUEUE_CAPACITY = 8; |
420 |
|
|
421 |
|
/** |
422 |
< |
* Array holding all worker threads in the pool. Array size must |
423 |
< |
* be a power of two. Updates and replacements are protected by |
424 |
< |
* workerLock, but the array is always kept in a consistent enough |
425 |
< |
* state to be randomly accessed without locking by workers |
459 |
< |
* performing work-stealing, as well as other traversal-based |
460 |
< |
* methods in this class. All readers must tolerate that some |
461 |
< |
* array slots may be null. |
422 |
> |
* Maximum size for submission queue array. Must be a power of two |
423 |
> |
* less than or equal to 1 << (31 - width of array entry) to |
424 |
> |
* ensure lack of index wraparound, but is capped at a lower |
425 |
> |
* value to help users trap runaway computations. |
426 |
|
*/ |
427 |
< |
volatile ForkJoinWorkerThread[] workers; |
427 |
> |
private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M |
428 |
|
|
429 |
|
/** |
430 |
< |
* Queue for external submissions. |
430 |
> |
* Array serving as submission queue. Initialized upon construction. |
431 |
|
*/ |
432 |
< |
private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue; |
432 |
> |
private ForkJoinTask<?>[] submissionQueue; |
433 |
|
|
434 |
|
/** |
435 |
< |
* Lock protecting updates to workers array. |
435 |
> |
* Lock protecting submissions array for addSubmission |
436 |
|
*/ |
437 |
< |
private final ReentrantLock workerLock; |
437 |
> |
private final ReentrantLock submissionLock; |
438 |
|
|
439 |
|
/** |
440 |
< |
* Latch released upon termination. |
440 |
> |
* Condition for awaitTermination, using submissionLock for |
441 |
> |
* convenience. |
442 |
|
*/ |
443 |
< |
private final Phaser termination; |
443 |
> |
private final Condition termination; |
444 |
|
|
445 |
|
/** |
446 |
|
* Creation factory for worker threads. |
448 |
|
private final ForkJoinWorkerThreadFactory factory; |
449 |
|
|
450 |
|
/** |
451 |
< |
* Sum of per-thread steal counts, updated only when threads are |
452 |
< |
* idle or terminating. |
451 |
> |
* The uncaught exception handler used when any worker abruptly |
452 |
> |
* terminates. |
453 |
|
*/ |
454 |
< |
private volatile long stealCount; |
454 |
> |
final Thread.UncaughtExceptionHandler ueh; |
455 |
|
|
456 |
|
/** |
457 |
< |
* Encoded record of top of Treiber stack of threads waiting for |
493 |
< |
* events. The top 32 bits contain the count being waited for. The |
494 |
< |
* bottom 16 bits contains one plus the pool index of waiting |
495 |
< |
* worker thread. (Bits 16-31 are unused.) |
457 |
> |
* Prefix for assigning names to worker threads |
458 |
|
*/ |
459 |
< |
private volatile long eventWaiters; |
498 |
< |
|
499 |
< |
private static final int EVENT_COUNT_SHIFT = 32; |
500 |
< |
private static final int WAITER_ID_MASK = (1 << 16) - 1; |
459 |
> |
private final String workerNamePrefix; |
460 |
|
|
461 |
|
/** |
462 |
< |
* A counter for events that may wake up worker threads: |
463 |
< |
* - Submission of a new task to the pool |
505 |
< |
* - A worker pushing a task on an empty queue |
506 |
< |
* - termination |
462 |
> |
* Sum of per-thread steal counts, updated only when threads are |
463 |
> |
* idle or terminating. |
464 |
|
*/ |
465 |
< |
private volatile int eventCount; |
465 |
> |
private volatile long stealCount; |
466 |
|
|
467 |
|
/** |
468 |
< |
* Encoded record of top of Treiber stack of spare threads waiting |
469 |
< |
* for resumption. The top 16 bits contain an arbitrary count to |
470 |
< |
* avoid ABA effects. The bottom 16bits contains one plus the pool |
471 |
< |
* index of waiting worker thread. |
472 |
< |
*/ |
473 |
< |
private volatile int spareWaiters; |
474 |
< |
|
475 |
< |
private static final int SPARE_COUNT_SHIFT = 16; |
476 |
< |
private static final int SPARE_ID_MASK = (1 << 16) - 1; |
468 |
> |
* Main pool control -- a long packed with: |
469 |
> |
* AC: Number of active running workers minus target parallelism (16 bits) |
470 |
> |
* TC: Number of total workers minus target parallelism (16bits) |
471 |
> |
* ST: true if pool is terminating (1 bit) |
472 |
> |
* EC: the wait count of top waiting thread (15 bits) |
473 |
> |
* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits) |
474 |
> |
* |
475 |
> |
* When convenient, we can extract the upper 32 bits of counts and |
476 |
> |
* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e = |
477 |
> |
* (int)ctl. The ec field is never accessed alone, but always |
478 |
> |
* together with id and st. The offsets of counts by the target |
479 |
> |
* parallelism and the positionings of fields makes it possible to |
480 |
> |
* perform the most common checks via sign tests of fields: When |
481 |
> |
* ac is negative, there are not enough active workers, when tc is |
482 |
> |
* negative, there are not enough total workers, when id is |
483 |
> |
* negative, there is at least one waiting worker, and when e is |
484 |
> |
* negative, the pool is terminating. To deal with these possibly |
485 |
> |
* negative fields, we use casts in and out of "short" and/or |
486 |
> |
* signed shifts to maintain signedness. |
487 |
> |
*/ |
488 |
> |
volatile long ctl; |
489 |
> |
|
490 |
> |
// bit positions/shifts for fields |
491 |
> |
private static final int AC_SHIFT = 48; |
492 |
> |
private static final int TC_SHIFT = 32; |
493 |
> |
private static final int ST_SHIFT = 31; |
494 |
> |
private static final int EC_SHIFT = 16; |
495 |
> |
|
496 |
> |
// bounds |
497 |
> |
private static final int MAX_ID = 0x7fff; // max poolIndex |
498 |
> |
private static final int SMASK = 0xffff; // mask short bits |
499 |
> |
private static final int SHORT_SIGN = 1 << 15; |
500 |
> |
private static final int INT_SIGN = 1 << 31; |
501 |
> |
|
502 |
> |
// masks |
503 |
> |
private static final long STOP_BIT = 0x0001L << ST_SHIFT; |
504 |
> |
private static final long AC_MASK = ((long)SMASK) << AC_SHIFT; |
505 |
> |
private static final long TC_MASK = ((long)SMASK) << TC_SHIFT; |
506 |
> |
|
507 |
> |
// units for incrementing and decrementing |
508 |
> |
private static final long TC_UNIT = 1L << TC_SHIFT; |
509 |
> |
private static final long AC_UNIT = 1L << AC_SHIFT; |
510 |
> |
|
511 |
> |
// masks and units for dealing with u = (int)(ctl >>> 32) |
512 |
> |
private static final int UAC_SHIFT = AC_SHIFT - 32; |
513 |
> |
private static final int UTC_SHIFT = TC_SHIFT - 32; |
514 |
> |
private static final int UAC_MASK = SMASK << UAC_SHIFT; |
515 |
> |
private static final int UTC_MASK = SMASK << UTC_SHIFT; |
516 |
> |
private static final int UAC_UNIT = 1 << UAC_SHIFT; |
517 |
> |
private static final int UTC_UNIT = 1 << UTC_SHIFT; |
518 |
> |
|
519 |
> |
// masks and units for dealing with e = (int)ctl |
520 |
> |
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
521 |
> |
private static final int EC_UNIT = 1 << EC_SHIFT; |
522 |
|
|
523 |
|
/** |
524 |
< |
* Lifecycle control. The low word contains the number of workers |
523 |
< |
* that are (probably) executing tasks. This value is atomically |
524 |
< |
* incremented before a worker gets a task to run, and decremented |
525 |
< |
* when a worker has no tasks and cannot find any. Bits 16-18 |
526 |
< |
* contain runLevel value. When all are zero, the pool is |
527 |
< |
* running. Level transitions are monotonic (running -> shutdown |
528 |
< |
* -> terminating -> terminated) so each transition adds a bit. |
529 |
< |
* These are bundled together to ensure consistent read for |
530 |
< |
* termination checks (i.e., that runLevel is at least SHUTDOWN |
531 |
< |
* and active threads is zero). |
532 |
< |
* |
533 |
< |
* Notes: Most direct CASes are dependent on these bitfield |
534 |
< |
* positions. Also, this field is non-private to enable direct |
535 |
< |
* performance-sensitive CASes in ForkJoinWorkerThread. |
524 |
> |
* The target parallelism level. |
525 |
|
*/ |
526 |
< |
volatile int runState; |
538 |
< |
|
539 |
< |
// Note: The order among run level values matters. |
540 |
< |
private static final int RUNLEVEL_SHIFT = 16; |
541 |
< |
private static final int SHUTDOWN = 1 << RUNLEVEL_SHIFT; |
542 |
< |
private static final int TERMINATING = 1 << (RUNLEVEL_SHIFT + 1); |
543 |
< |
private static final int TERMINATED = 1 << (RUNLEVEL_SHIFT + 2); |
544 |
< |
private static final int ACTIVE_COUNT_MASK = (1 << RUNLEVEL_SHIFT) - 1; |
526 |
> |
final int parallelism; |
527 |
|
|
528 |
|
/** |
529 |
< |
* Holds number of total (i.e., created and not yet terminated) |
530 |
< |
* and running (i.e., not blocked on joins or other managed sync) |
531 |
< |
* threads, packed together to ensure consistent snapshot when |
532 |
< |
* making decisions about creating and suspending spare |
551 |
< |
* threads. Updated only by CAS. Note that adding a new worker |
552 |
< |
* requires incrementing both counts, since workers start off in |
553 |
< |
* running state. |
529 |
> |
* Index (mod submission queue length) of next element to take |
530 |
> |
* from submission queue. Usage is identical to that for |
531 |
> |
* per-worker queues -- see ForkJoinWorkerThread internal |
532 |
> |
* documentation. |
533 |
|
*/ |
534 |
< |
private volatile int workerCounts; |
534 |
> |
volatile int queueBase; |
535 |
|
|
536 |
< |
private static final int TOTAL_COUNT_SHIFT = 16; |
537 |
< |
private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1; |
538 |
< |
private static final int ONE_RUNNING = 1; |
539 |
< |
private static final int ONE_TOTAL = 1 << TOTAL_COUNT_SHIFT; |
536 |
> |
/** |
537 |
> |
* Index (mod submission queue length) of next element to add |
538 |
> |
* in submission queue. Usage is identical to that for |
539 |
> |
* per-worker queues -- see ForkJoinWorkerThread internal |
540 |
> |
* documentation. |
541 |
> |
*/ |
542 |
> |
int queueTop; |
543 |
|
|
544 |
|
/** |
545 |
< |
* The target parallelism level. |
564 |
< |
* Accessed directly by ForkJoinWorkerThreads. |
545 |
> |
* True when shutdown() has been called. |
546 |
|
*/ |
547 |
< |
final int parallelism; |
547 |
> |
volatile boolean shutdown; |
548 |
|
|
549 |
|
/** |
550 |
|
* True if use local fifo, not default lifo, for local polling |
553 |
|
final boolean locallyFifo; |
554 |
|
|
555 |
|
/** |
556 |
< |
* The uncaught exception handler used when any worker abruptly |
557 |
< |
* terminates. |
556 |
> |
* The number of threads in ForkJoinWorkerThreads.helpQuiescePool. |
557 |
> |
* When non-zero, suppresses automatic shutdown when active |
558 |
> |
* counts become zero. |
559 |
|
*/ |
560 |
< |
private final Thread.UncaughtExceptionHandler ueh; |
560 |
> |
volatile int quiescerCount; |
561 |
|
|
562 |
|
/** |
563 |
< |
* Pool number, just for assigning useful names to worker threads |
563 |
> |
* The number of threads blocked in join. |
564 |
|
*/ |
565 |
< |
private final int poolNumber; |
584 |
< |
|
585 |
< |
// Utilities for CASing fields. Note that most of these |
586 |
< |
// are usually manually inlined by callers |
565 |
> |
volatile int blockedCount; |
566 |
|
|
567 |
|
/** |
568 |
< |
* Increments running count part of workerCounts. |
568 |
> |
* Counter for worker Thread names (unrelated to their poolIndex) |
569 |
|
*/ |
570 |
< |
final void incrementRunningCount() { |
592 |
< |
int c; |
593 |
< |
do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
594 |
< |
c = workerCounts, |
595 |
< |
c + ONE_RUNNING)); |
596 |
< |
} |
570 |
> |
private volatile int nextWorkerNumber; |
571 |
|
|
572 |
|
/** |
573 |
< |
* Tries to increment running count part of workerCounts. |
573 |
> |
* The index for the next created worker. Accessed under scanGuard. |
574 |
|
*/ |
575 |
< |
final boolean tryIncrementRunningCount() { |
602 |
< |
int c; |
603 |
< |
return UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
604 |
< |
c = workerCounts, |
605 |
< |
c + ONE_RUNNING); |
606 |
< |
} |
575 |
> |
private int nextWorkerIndex; |
576 |
|
|
577 |
|
/** |
578 |
< |
* Tries to decrement running count unless already zero. |
578 |
> |
* SeqLock and index masking for updates to workers array. Locked |
579 |
> |
* when SG_UNIT is set. Unlocking clears bit by adding |
580 |
> |
* SG_UNIT. Staleness of read-only operations can be checked by |
581 |
> |
* comparing scanGuard to value before the reads. The low 16 bits |
582 |
> |
* (i.e, anding with SMASK) hold (the smallest power of two |
583 |
> |
* covering all worker indices, minus one, and is used to avoid |
584 |
> |
* dealing with large numbers of null slots when the workers array |
585 |
> |
* is overallocated. |
586 |
|
*/ |
587 |
< |
final boolean tryDecrementRunningCount() { |
612 |
< |
int wc = workerCounts; |
613 |
< |
if ((wc & RUNNING_COUNT_MASK) == 0) |
614 |
< |
return false; |
615 |
< |
return UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
616 |
< |
wc, wc - ONE_RUNNING); |
617 |
< |
} |
587 |
> |
volatile int scanGuard; |
588 |
|
|
589 |
< |
/** |
620 |
< |
* Forces decrement of encoded workerCounts, awaiting nonzero if |
621 |
< |
* (rarely) necessary when other count updates lag. |
622 |
< |
* |
623 |
< |
* @param dr -- either zero or ONE_RUNNING |
624 |
< |
* @param dt -- either zero or ONE_TOTAL |
625 |
< |
*/ |
626 |
< |
private void decrementWorkerCounts(int dr, int dt) { |
627 |
< |
for (;;) { |
628 |
< |
int wc = workerCounts; |
629 |
< |
if ((wc & RUNNING_COUNT_MASK) - dr < 0 || |
630 |
< |
(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) { |
631 |
< |
if ((runState & TERMINATED) != 0) |
632 |
< |
return; // lagging termination on a backout |
633 |
< |
Thread.yield(); |
634 |
< |
} |
635 |
< |
if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
636 |
< |
wc, wc - (dr + dt))) |
637 |
< |
return; |
638 |
< |
} |
639 |
< |
} |
589 |
> |
private static final int SG_UNIT = 1 << 16; |
590 |
|
|
591 |
|
/** |
592 |
< |
* Tries decrementing active count; fails on contention. |
593 |
< |
* Called when workers cannot find tasks to run. |
592 |
> |
* The wakeup interval (in nanoseconds) for a worker waiting for a |
593 |
> |
* task when the pool is quiescent to instead try to shrink the |
594 |
> |
* number of workers. The exact value does not matter too |
595 |
> |
* much. It must be short enough to release resources during |
596 |
> |
* sustained periods of idleness, but not so short that threads |
597 |
> |
* are continually re-created. |
598 |
|
*/ |
599 |
< |
final boolean tryDecrementActiveCount() { |
600 |
< |
int c; |
647 |
< |
return UNSAFE.compareAndSwapInt(this, runStateOffset, |
648 |
< |
c = runState, c - 1); |
649 |
< |
} |
599 |
> |
private static final long SHRINK_RATE = |
600 |
> |
4L * 1000L * 1000L * 1000L; // 4 seconds |
601 |
|
|
602 |
|
/** |
603 |
< |
* Advances to at least the given level. Returns true if not |
604 |
< |
* already in at least the given level. |
603 |
> |
* Top-level loop for worker threads: On each step: if the |
604 |
> |
* previous step swept through all queues and found no tasks, or |
605 |
> |
* there are excess threads, then possibly blocks. Otherwise, |
606 |
> |
* scans for and, if found, executes a task. Returns when pool |
607 |
> |
* and/or worker terminate. |
608 |
> |
* |
609 |
> |
* @param w the worker |
610 |
|
*/ |
611 |
< |
private boolean advanceRunLevel(int level) { |
612 |
< |
for (;;) { |
613 |
< |
int s = runState; |
614 |
< |
if ((s & level) != 0) |
615 |
< |
return false; |
616 |
< |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level)) |
617 |
< |
return true; |
611 |
> |
final void work(ForkJoinWorkerThread w) { |
612 |
> |
boolean swept = false; // true on empty scans |
613 |
> |
long c; |
614 |
> |
while (!w.terminate && (int)(c = ctl) >= 0) { |
615 |
> |
int a; // active count |
616 |
> |
if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0) |
617 |
> |
swept = scan(w, a); |
618 |
> |
else if (tryAwaitWork(w, c)) |
619 |
> |
swept = false; |
620 |
|
} |
621 |
|
} |
622 |
|
|
623 |
< |
// workers array maintenance |
623 |
> |
// Signalling |
624 |
|
|
625 |
|
/** |
626 |
< |
* Records and returns a workers array index for new worker. |
626 |
> |
* Wakes up or creates a worker. |
627 |
|
*/ |
628 |
< |
private int recordWorker(ForkJoinWorkerThread w) { |
629 |
< |
// Try using slot totalCount-1. If not available, scan and/or resize |
630 |
< |
int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1; |
631 |
< |
final ReentrantLock lock = this.workerLock; |
632 |
< |
lock.lock(); |
633 |
< |
try { |
634 |
< |
ForkJoinWorkerThread[] ws = workers; |
635 |
< |
int n = ws.length; |
636 |
< |
if (k < 0 || k >= n || ws[k] != null) { |
637 |
< |
for (k = 0; k < n && ws[k] != null; ++k) |
638 |
< |
; |
639 |
< |
if (k == n) |
640 |
< |
ws = workers = Arrays.copyOf(ws, n << 1); |
641 |
< |
} |
642 |
< |
ws[k] = w; |
643 |
< |
int c = eventCount; // advance event count to ensure visibility |
644 |
< |
UNSAFE.compareAndSwapInt(this, eventCountOffset, c, c+1); |
645 |
< |
} finally { |
646 |
< |
lock.unlock(); |
628 |
> |
final void signalWork() { |
629 |
> |
/* |
630 |
> |
* The while condition is true if: (there is are too few total |
631 |
> |
* workers OR there is at least one waiter) AND (there are too |
632 |
> |
* few active workers OR the pool is terminating). The value |
633 |
> |
* of e distinguishes the remaining cases: zero (no waiters) |
634 |
> |
* for create, negative if terminating (in which case do |
635 |
> |
* nothing), else release a waiter. The secondary checks for |
636 |
> |
* release (non-null array etc) can fail if the pool begins |
637 |
> |
* terminating after the test, and don't impose any added cost |
638 |
> |
* because JVMs must perform null and bounds checks anyway. |
639 |
> |
*/ |
640 |
> |
long c; int e, u; |
641 |
> |
while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) & |
642 |
> |
(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) { |
643 |
> |
if (e > 0) { // release a waiting worker |
644 |
> |
int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws; |
645 |
> |
if ((ws = workers) == null || |
646 |
> |
(i = ~e & SMASK) >= ws.length || |
647 |
> |
(w = ws[i]) == null) |
648 |
> |
break; |
649 |
> |
long nc = (((long)(w.nextWait & E_MASK)) | |
650 |
> |
((long)(u + UAC_UNIT) << 32)); |
651 |
> |
if (w.eventCount == e && |
652 |
> |
UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { |
653 |
> |
w.eventCount = (e + EC_UNIT) & E_MASK; |
654 |
> |
if (w.parked) |
655 |
> |
UNSAFE.unpark(w); |
656 |
> |
break; |
657 |
> |
} |
658 |
> |
} |
659 |
> |
else if (UNSAFE.compareAndSwapLong |
660 |
> |
(this, ctlOffset, c, |
661 |
> |
(long)(((u + UTC_UNIT) & UTC_MASK) | |
662 |
> |
((u + UAC_UNIT) & UAC_MASK)) << 32)) { |
663 |
> |
addWorker(); |
664 |
> |
break; |
665 |
> |
} |
666 |
|
} |
690 |
– |
return k; |
667 |
|
} |
668 |
|
|
669 |
|
/** |
670 |
< |
* Nulls out record of worker in workers array. |
670 |
> |
* Variant of signalWork to help release waiters on rescans. |
671 |
> |
* Tries once to release a waiter if active count < 0. |
672 |
> |
* |
673 |
> |
* @return false if failed due to contention, else true |
674 |
|
*/ |
675 |
< |
private void forgetWorker(ForkJoinWorkerThread w) { |
676 |
< |
int idx = w.poolIndex; |
677 |
< |
// Locking helps method recordWorker avoid unnecessary expansion |
678 |
< |
final ReentrantLock lock = this.workerLock; |
679 |
< |
lock.lock(); |
680 |
< |
try { |
681 |
< |
ForkJoinWorkerThread[] ws = workers; |
682 |
< |
if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify |
683 |
< |
ws[idx] = null; |
684 |
< |
} finally { |
685 |
< |
lock.unlock(); |
675 |
> |
private boolean tryReleaseWaiter() { |
676 |
> |
long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws; |
677 |
> |
if ((e = (int)(c = ctl)) > 0 && |
678 |
> |
(int)(c >> AC_SHIFT) < 0 && |
679 |
> |
(ws = workers) != null && |
680 |
> |
(i = ~e & SMASK) < ws.length && |
681 |
> |
(w = ws[i]) != null) { |
682 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
683 |
> |
((c + AC_UNIT) & (AC_MASK|TC_MASK))); |
684 |
> |
if (w.eventCount != e || |
685 |
> |
!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) |
686 |
> |
return false; |
687 |
> |
w.eventCount = (e + EC_UNIT) & E_MASK; |
688 |
> |
if (w.parked) |
689 |
> |
UNSAFE.unpark(w); |
690 |
|
} |
691 |
+ |
return true; |
692 |
|
} |
693 |
|
|
694 |
+ |
// Scanning for tasks |
695 |
+ |
|
696 |
|
/** |
697 |
< |
* Final callback from terminating worker. Removes record of |
698 |
< |
* worker from array, and adjusts counts. If pool is shutting |
699 |
< |
* down, tries to complete termination. |
697 |
> |
* Scans for and, if found, executes one task. Scans start at a |
698 |
> |
* random index of workers array, and randomly select the first |
699 |
> |
* (2*#workers)-1 probes, and then, if all empty, resort to 2 |
700 |
> |
* circular sweeps, which is necessary to check quiescence. and |
701 |
> |
* taking a submission only if no stealable tasks were found. The |
702 |
> |
* steal code inside the loop is a specialized form of |
703 |
> |
* ForkJoinWorkerThread.deqTask, followed bookkeeping to support |
704 |
> |
* helpJoinTask and signal propagation. The code for submission |
705 |
> |
* queues is almost identical. On each steal, the worker completes |
706 |
> |
* not only the task, but also all local tasks that this task may |
707 |
> |
* have generated. On detecting staleness or contention when |
708 |
> |
* trying to take a task, this method returns without finishing |
709 |
> |
* sweep, which allows global state rechecks before retry. |
710 |
|
* |
711 |
|
* @param w the worker |
712 |
+ |
* @param a the number of active workers |
713 |
+ |
* @return true if swept all queues without finding a task |
714 |
|
*/ |
715 |
< |
final void workerTerminated(ForkJoinWorkerThread w) { |
716 |
< |
forgetWorker(w); |
717 |
< |
decrementWorkerCounts(w.isTrimmed() ? 0 : ONE_RUNNING, ONE_TOTAL); |
720 |
< |
while (w.stealCount != 0) // collect final count |
721 |
< |
tryAccumulateStealCount(w); |
722 |
< |
tryTerminate(false); |
723 |
< |
} |
724 |
< |
|
725 |
< |
// Waiting for and signalling events |
726 |
< |
|
727 |
< |
/** |
728 |
< |
* Releases workers blocked on a count not equal to current count. |
729 |
< |
* Normally called after precheck that eventWaiters isn't zero to |
730 |
< |
* avoid wasted array checks. Gives up upon a change in count or |
731 |
< |
* upon releasing four workers, letting others take over. |
732 |
< |
*/ |
733 |
< |
private void releaseEventWaiters() { |
715 |
> |
private boolean scan(ForkJoinWorkerThread w, int a) { |
716 |
> |
int g = scanGuard; // mask 0 avoids useless scans if only one active |
717 |
> |
int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK; |
718 |
|
ForkJoinWorkerThread[] ws = workers; |
719 |
< |
int n = ws.length; |
720 |
< |
long h = eventWaiters; |
721 |
< |
int ec = eventCount; |
722 |
< |
int releases = 4; |
723 |
< |
ForkJoinWorkerThread w; int id; |
724 |
< |
while ((id = (((int)h) & WAITER_ID_MASK) - 1) >= 0 && |
725 |
< |
(int)(h >>> EVENT_COUNT_SHIFT) != ec && |
726 |
< |
id < n && (w = ws[id]) != null) { |
727 |
< |
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
728 |
< |
h, w.nextWaiter)) { |
729 |
< |
LockSupport.unpark(w); |
730 |
< |
if (--releases == 0) |
731 |
< |
break; |
719 |
> |
if (ws == null || ws.length <= m) // staleness check |
720 |
> |
return false; |
721 |
> |
for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) { |
722 |
> |
ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i; |
723 |
> |
ForkJoinWorkerThread v = ws[k & m]; |
724 |
> |
if (v != null && (b = v.queueBase) != v.queueTop && |
725 |
> |
(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) { |
726 |
> |
long u = (i << ASHIFT) + ABASE; |
727 |
> |
if ((t = q[i]) != null && v.queueBase == b && |
728 |
> |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
729 |
> |
int d = (v.queueBase = b + 1) - v.queueTop; |
730 |
> |
v.stealHint = w.poolIndex; |
731 |
> |
if (d != 0) |
732 |
> |
signalWork(); // propagate if nonempty |
733 |
> |
w.execTask(t); |
734 |
> |
} |
735 |
> |
r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5); |
736 |
> |
return false; // store next seed |
737 |
|
} |
738 |
< |
if (eventCount != ec) |
739 |
< |
break; |
740 |
< |
h = eventWaiters; |
738 |
> |
else if (j < 0) { // xorshift |
739 |
> |
r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5; |
740 |
> |
} |
741 |
> |
else |
742 |
> |
++k; |
743 |
> |
} |
744 |
> |
if (scanGuard != g) // staleness check |
745 |
> |
return false; |
746 |
> |
else { // try to take submission |
747 |
> |
ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i; |
748 |
> |
if ((b = queueBase) != queueTop && |
749 |
> |
(q = submissionQueue) != null && |
750 |
> |
(i = (q.length - 1) & b) >= 0) { |
751 |
> |
long u = (i << ASHIFT) + ABASE; |
752 |
> |
if ((t = q[i]) != null && queueBase == b && |
753 |
> |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
754 |
> |
queueBase = b + 1; |
755 |
> |
w.execTask(t); |
756 |
> |
} |
757 |
> |
return false; |
758 |
> |
} |
759 |
> |
return true; // all queues empty |
760 |
|
} |
761 |
|
} |
762 |
|
|
763 |
|
/** |
764 |
< |
* Tries to advance eventCount and releases waiters. Called only |
765 |
< |
* from workers. |
766 |
< |
*/ |
767 |
< |
final void signalWork() { |
768 |
< |
int c; // try to increment event count -- CAS failure OK |
769 |
< |
UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1); |
770 |
< |
if (eventWaiters != 0L) |
771 |
< |
releaseEventWaiters(); |
772 |
< |
} |
773 |
< |
|
774 |
< |
/** |
775 |
< |
* Adds the given worker to event queue and blocks until |
776 |
< |
* terminating or event count advances from the given value |
777 |
< |
* |
778 |
< |
* @param w the calling worker thread |
779 |
< |
* @param ec the count |
780 |
< |
*/ |
781 |
< |
private void eventSync(ForkJoinWorkerThread w, int ec) { |
782 |
< |
long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1)); |
783 |
< |
long h; |
784 |
< |
while ((runState < SHUTDOWN || !tryTerminate(false)) && |
785 |
< |
(((int)(h = eventWaiters) & WAITER_ID_MASK) == 0 || |
786 |
< |
(int)(h >>> EVENT_COUNT_SHIFT) == ec) && |
787 |
< |
eventCount == ec) { |
788 |
< |
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset, |
789 |
< |
w.nextWaiter = h, nh)) { |
790 |
< |
awaitEvent(w, ec); |
791 |
< |
break; |
764 |
> |
* Tries to enqueue worker w in wait queue and await change in |
765 |
> |
* worker's eventCount. If the pool is quiescent and there is |
766 |
> |
* more than one worker, possibly terminates worker upon exit. |
767 |
> |
* Otherwise, before blocking, rescans queues to avoid missed |
768 |
> |
* signals. Upon finding work, releases at least one worker |
769 |
> |
* (which may be the current worker). Rescans restart upon |
770 |
> |
* detected staleness or failure to release due to |
771 |
> |
* contention. Note the unusual conventions about Thread.interrupt |
772 |
> |
* here and elsewhere: Because interrupts are used solely to alert |
773 |
> |
* threads to check termination, which is checked here anyway, we |
774 |
> |
* clear status (using Thread.interrupted) before any call to |
775 |
> |
* park, so that park does not immediately return due to status |
776 |
> |
* being set via some other unrelated call to interrupt in user |
777 |
> |
* code. |
778 |
> |
* |
779 |
> |
* @param w the calling worker |
780 |
> |
* @param c the ctl value on entry |
781 |
> |
* @return true if waited or another thread was released upon enq |
782 |
> |
*/ |
783 |
> |
private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) { |
784 |
> |
int v = w.eventCount; |
785 |
> |
w.nextWait = (int)c; // w's successor record |
786 |
> |
long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); |
787 |
> |
if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { |
788 |
> |
long d = ctl; // return true if lost to a deq, to force scan |
789 |
> |
return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L; |
790 |
> |
} |
791 |
> |
for (int sc = w.stealCount; sc != 0;) { // accumulate stealCount |
792 |
> |
long s = stealCount; |
793 |
> |
if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc)) |
794 |
> |
sc = w.stealCount = 0; |
795 |
> |
else if (w.eventCount != v) |
796 |
> |
return true; // update next time |
797 |
> |
} |
798 |
> |
if ((!shutdown || !tryTerminate(false)) && |
799 |
> |
(int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 && |
800 |
> |
blockedCount == 0 && quiescerCount == 0) |
801 |
> |
idleAwaitWork(w, nc, c, v); // quiescent |
802 |
> |
for (boolean rescanned = false;;) { |
803 |
> |
if (w.eventCount != v) |
804 |
> |
return true; |
805 |
> |
if (!rescanned) { |
806 |
> |
int g = scanGuard, m = g & SMASK; |
807 |
> |
ForkJoinWorkerThread[] ws = workers; |
808 |
> |
if (ws != null && m < ws.length) { |
809 |
> |
rescanned = true; |
810 |
> |
for (int i = 0; i <= m; ++i) { |
811 |
> |
ForkJoinWorkerThread u = ws[i]; |
812 |
> |
if (u != null) { |
813 |
> |
if (u.queueBase != u.queueTop && |
814 |
> |
!tryReleaseWaiter()) |
815 |
> |
rescanned = false; // contended |
816 |
> |
if (w.eventCount != v) |
817 |
> |
return true; |
818 |
> |
} |
819 |
> |
} |
820 |
> |
} |
821 |
> |
if (scanGuard != g || // stale |
822 |
> |
(queueBase != queueTop && !tryReleaseWaiter())) |
823 |
> |
rescanned = false; |
824 |
> |
if (!rescanned) |
825 |
> |
Thread.yield(); // reduce contention |
826 |
> |
else |
827 |
> |
Thread.interrupted(); // clear before park |
828 |
> |
} |
829 |
> |
else { |
830 |
> |
w.parked = true; // must recheck |
831 |
> |
if (w.eventCount != v) { |
832 |
> |
w.parked = false; |
833 |
> |
return true; |
834 |
> |
} |
835 |
> |
LockSupport.park(this); |
836 |
> |
rescanned = w.parked = false; |
837 |
|
} |
838 |
|
} |
839 |
|
} |
840 |
|
|
841 |
|
/** |
842 |
< |
* Blocks the given worker (that has already been entered as an |
843 |
< |
* event waiter) until terminating or event count advances from |
844 |
< |
* the given value. The oldest (first) waiter uses a timed wait to |
845 |
< |
* occasionally one-by-one shrink the number of workers (to a |
846 |
< |
* minimum of one) if the pool has not been used for extended |
847 |
< |
* periods. |
848 |
< |
* |
849 |
< |
* @param w the calling worker thread |
850 |
< |
* @param ec the count |
851 |
< |
*/ |
852 |
< |
private void awaitEvent(ForkJoinWorkerThread w, int ec) { |
853 |
< |
while (eventCount == ec) { |
854 |
< |
if (tryAccumulateStealCount(w)) { // transfer while idle |
855 |
< |
boolean untimed = (w.nextWaiter != 0L || |
856 |
< |
(workerCounts & RUNNING_COUNT_MASK) <= 1); |
857 |
< |
long startTime = untimed ? 0 : System.nanoTime(); |
858 |
< |
Thread.interrupted(); // clear/ignore interrupt |
859 |
< |
if (w.isTerminating() || eventCount != ec) |
860 |
< |
break; // recheck after clear |
861 |
< |
if (untimed) |
862 |
< |
LockSupport.park(w); |
863 |
< |
else { |
864 |
< |
LockSupport.parkNanos(w, SHRINK_RATE_NANOS); |
865 |
< |
if (eventCount != ec || w.isTerminating()) |
866 |
< |
break; |
867 |
< |
if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS) |
868 |
< |
tryShutdownUnusedWorker(ec); |
842 |
> |
* If inactivating worker w has caused pool to become |
843 |
> |
* quiescent, check for pool termination, and wait for event |
844 |
> |
* for up to SHRINK_RATE nanosecs (rescans are unnecessary in |
845 |
> |
* this case because quiescence reflects consensus about lack |
846 |
> |
* of work). On timeout, if ctl has not changed, terminate the |
847 |
> |
* worker. Upon its termination (see deregisterWorker), it may |
848 |
> |
* wake up another worker to possibly repeat this process. |
849 |
> |
* |
850 |
> |
* @param w the calling worker |
851 |
> |
* @param currentCtl the ctl value after enqueuing w |
852 |
> |
* @param prevCtl the ctl value if w terminated |
853 |
> |
* @param v the eventCount w awaits change |
854 |
> |
*/ |
855 |
> |
private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl, |
856 |
> |
long prevCtl, int v) { |
857 |
> |
if (w.eventCount == v) { |
858 |
> |
if (shutdown) |
859 |
> |
tryTerminate(false); |
860 |
> |
ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs |
861 |
> |
while (ctl == currentCtl) { |
862 |
> |
long startTime = System.nanoTime(); |
863 |
> |
w.parked = true; |
864 |
> |
if (w.eventCount == v) // must recheck |
865 |
> |
LockSupport.parkNanos(this, SHRINK_RATE); |
866 |
> |
w.parked = false; |
867 |
> |
if (w.eventCount != v) |
868 |
> |
break; |
869 |
> |
else if (System.nanoTime() - startTime < |
870 |
> |
SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop |
871 |
> |
Thread.interrupted(); // spurious wakeup |
872 |
> |
else if (UNSAFE.compareAndSwapLong(this, ctlOffset, |
873 |
> |
currentCtl, prevCtl)) { |
874 |
> |
w.terminate = true; // restore previous |
875 |
> |
w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK; |
876 |
> |
break; |
877 |
|
} |
878 |
|
} |
879 |
|
} |
880 |
|
} |
881 |
|
|
882 |
< |
// Maintaining parallelism |
882 |
> |
// Submissions |
883 |
|
|
884 |
|
/** |
885 |
< |
* Pushes worker onto the spare stack. |
885 |
> |
* Enqueues the given task in the submissionQueue. Same idea as |
886 |
> |
* ForkJoinWorkerThread.pushTask except for use of submissionLock. |
887 |
> |
* |
888 |
> |
* @param t the task |
889 |
|
*/ |
890 |
< |
final void pushSpare(ForkJoinWorkerThread w) { |
891 |
< |
int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1); |
892 |
< |
do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
893 |
< |
w.nextSpare = spareWaiters,ns)); |
890 |
> |
private void addSubmission(ForkJoinTask<?> t) { |
891 |
> |
final ReentrantLock lock = this.submissionLock; |
892 |
> |
lock.lock(); |
893 |
> |
try { |
894 |
> |
ForkJoinTask<?>[] q; int s, m; |
895 |
> |
if ((q = submissionQueue) != null) { // ignore if queue removed |
896 |
> |
long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE; |
897 |
> |
UNSAFE.putOrderedObject(q, u, t); |
898 |
> |
queueTop = s + 1; |
899 |
> |
if (s - queueBase == m) |
900 |
> |
growSubmissionQueue(); |
901 |
> |
} |
902 |
> |
} finally { |
903 |
> |
lock.unlock(); |
904 |
> |
} |
905 |
> |
signalWork(); |
906 |
|
} |
907 |
|
|
908 |
+ |
// (pollSubmission is defined below with exported methods) |
909 |
+ |
|
910 |
|
/** |
911 |
< |
* Tries (once) to resume a spare if the number of running |
912 |
< |
* threads is less than target. |
911 |
> |
* Creates or doubles submissionQueue array. |
912 |
> |
* Basically identical to ForkJoinWorkerThread version. |
913 |
|
*/ |
914 |
< |
private void tryResumeSpare() { |
915 |
< |
int sw, id; |
916 |
< |
ForkJoinWorkerThread[] ws = workers; |
917 |
< |
int n = ws.length; |
918 |
< |
ForkJoinWorkerThread w; |
919 |
< |
if ((sw = spareWaiters) != 0 && |
920 |
< |
(id = (sw & SPARE_ID_MASK) - 1) >= 0 && |
921 |
< |
id < n && (w = ws[id]) != null && |
922 |
< |
(runState >= TERMINATING || |
923 |
< |
(workerCounts & RUNNING_COUNT_MASK) < parallelism) && |
924 |
< |
spareWaiters == sw && |
925 |
< |
UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
926 |
< |
sw, w.nextSpare)) { |
927 |
< |
int c; // increment running count before resume |
928 |
< |
do {} while (!UNSAFE.compareAndSwapInt |
929 |
< |
(this, workerCountsOffset, |
930 |
< |
c = workerCounts, c + ONE_RUNNING)); |
931 |
< |
if (w.tryUnsuspend()) |
932 |
< |
LockSupport.unpark(w); |
855 |
< |
else // back out if w was shutdown |
856 |
< |
decrementWorkerCounts(ONE_RUNNING, 0); |
914 |
> |
private void growSubmissionQueue() { |
915 |
> |
ForkJoinTask<?>[] oldQ = submissionQueue; |
916 |
> |
int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY; |
917 |
> |
if (size > MAXIMUM_QUEUE_CAPACITY) |
918 |
> |
throw new RejectedExecutionException("Queue capacity exceeded"); |
919 |
> |
if (size < INITIAL_QUEUE_CAPACITY) |
920 |
> |
size = INITIAL_QUEUE_CAPACITY; |
921 |
> |
ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size]; |
922 |
> |
int mask = size - 1; |
923 |
> |
int top = queueTop; |
924 |
> |
int oldMask; |
925 |
> |
if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) { |
926 |
> |
for (int b = queueBase; b != top; ++b) { |
927 |
> |
long u = ((b & oldMask) << ASHIFT) + ABASE; |
928 |
> |
Object x = UNSAFE.getObjectVolatile(oldQ, u); |
929 |
> |
if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null)) |
930 |
> |
UNSAFE.putObjectVolatile |
931 |
> |
(q, ((b & mask) << ASHIFT) + ABASE, x); |
932 |
> |
} |
933 |
|
} |
934 |
|
} |
935 |
|
|
936 |
+ |
// Blocking support |
937 |
+ |
|
938 |
|
/** |
939 |
< |
* Tries to increase the number of running workers if below target |
940 |
< |
* parallelism: If a spare exists tries to resume it via |
941 |
< |
* tryResumeSpare. Otherwise, if not enough total workers or all |
942 |
< |
* existing workers are busy, adds a new worker. In all cases also |
943 |
< |
* helps wake up releasable workers waiting for work. |
944 |
< |
*/ |
945 |
< |
private void helpMaintainParallelism() { |
946 |
< |
int pc = parallelism; |
947 |
< |
int wc, rs, tc; |
948 |
< |
while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc && |
949 |
< |
(rs = runState) < TERMINATING) { |
950 |
< |
if (spareWaiters != 0) |
951 |
< |
tryResumeSpare(); |
952 |
< |
else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS || |
953 |
< |
(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc)) |
954 |
< |
break; // enough total |
955 |
< |
else if (runState == rs && workerCounts == wc && |
956 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc, |
879 |
< |
wc + (ONE_RUNNING|ONE_TOTAL))) { |
880 |
< |
ForkJoinWorkerThread w = null; |
881 |
< |
Throwable fail = null; |
882 |
< |
try { |
883 |
< |
w = factory.newThread(this); |
884 |
< |
} catch (Throwable ex) { |
885 |
< |
fail = ex; |
939 |
> |
* Tries to increment blockedCount, decrement active count |
940 |
> |
* (sometimes implicitly) and possibly release or create a |
941 |
> |
* compensating worker in preparation for blocking. Fails |
942 |
> |
* on contention or termination. |
943 |
> |
* |
944 |
> |
* @return true if the caller can block, else should recheck and retry |
945 |
> |
*/ |
946 |
> |
private boolean tryPreBlock() { |
947 |
> |
int b = blockedCount; |
948 |
> |
if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) { |
949 |
> |
int pc = parallelism; |
950 |
> |
do { |
951 |
> |
ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w; |
952 |
> |
int e, ac, tc, rc, i; |
953 |
> |
long c = ctl; |
954 |
> |
int u = (int)(c >>> 32); |
955 |
> |
if ((e = (int)c) < 0) { |
956 |
> |
// skip -- terminating |
957 |
|
} |
958 |
< |
if (w == null) { // null or exceptional factory return |
959 |
< |
decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL); |
960 |
< |
tryTerminate(false); // handle failure during shutdown |
961 |
< |
// If originating from an external caller, |
962 |
< |
// propagate exception, else ignore |
963 |
< |
if (fail != null && runState < TERMINATING && |
964 |
< |
!(Thread.currentThread() instanceof |
965 |
< |
ForkJoinWorkerThread)) |
966 |
< |
UNSAFE.throwException(fail); |
967 |
< |
break; |
958 |
> |
else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 && |
959 |
> |
(ws = workers) != null && |
960 |
> |
(i = ~e & SMASK) < ws.length && |
961 |
> |
(w = ws[i]) != null) { |
962 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
963 |
> |
(c & (AC_MASK|TC_MASK))); |
964 |
> |
if (w.eventCount == e && |
965 |
> |
UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { |
966 |
> |
w.eventCount = (e + EC_UNIT) & E_MASK; |
967 |
> |
if (w.parked) |
968 |
> |
UNSAFE.unpark(w); |
969 |
> |
return true; // release an idle worker |
970 |
> |
} |
971 |
|
} |
972 |
< |
w.start(recordWorker(w), ueh); |
973 |
< |
if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) |
974 |
< |
break; // add at most one unless total below target |
975 |
< |
} |
972 |
> |
else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) { |
973 |
> |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
974 |
> |
if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) |
975 |
> |
return true; // no compensation needed |
976 |
> |
} |
977 |
> |
else if (tc + pc < MAX_ID) { |
978 |
> |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
979 |
> |
if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { |
980 |
> |
addWorker(); |
981 |
> |
return true; // create a replacement |
982 |
> |
} |
983 |
> |
} |
984 |
> |
// try to back out on any failure and let caller retry |
985 |
> |
} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset, |
986 |
> |
b = blockedCount, b - 1)); |
987 |
|
} |
988 |
< |
if (eventWaiters != 0L) |
904 |
< |
releaseEventWaiters(); |
988 |
> |
return false; |
989 |
|
} |
990 |
|
|
991 |
|
/** |
992 |
< |
* Callback from the oldest waiter in awaitEvent waking up after a |
909 |
< |
* period of non-use. If all workers are idle, tries (once) to |
910 |
< |
* shutdown an event waiter or a spare, if one exists. Note that |
911 |
< |
* we don't need CAS or locks here because the method is called |
912 |
< |
* only from one thread occasionally waking (and even misfires are |
913 |
< |
* OK). Note that until the shutdown worker fully terminates, |
914 |
< |
* workerCounts will overestimate total count, which is tolerable. |
915 |
< |
* |
916 |
< |
* @param ec the event count waited on by caller (to abort |
917 |
< |
* attempt if count has since changed). |
992 |
> |
* Decrements blockedCount and increments active count |
993 |
|
*/ |
994 |
< |
private void tryShutdownUnusedWorker(int ec) { |
995 |
< |
if (runState == 0 && eventCount == ec) { // only trigger if all idle |
996 |
< |
ForkJoinWorkerThread[] ws = workers; |
997 |
< |
int n = ws.length; |
998 |
< |
ForkJoinWorkerThread w = null; |
999 |
< |
boolean shutdown = false; |
1000 |
< |
int sw; |
926 |
< |
long h; |
927 |
< |
if ((sw = spareWaiters) != 0) { // prefer killing spares |
928 |
< |
int id = (sw & SPARE_ID_MASK) - 1; |
929 |
< |
if (id >= 0 && id < n && (w = ws[id]) != null && |
930 |
< |
UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
931 |
< |
sw, w.nextSpare)) |
932 |
< |
shutdown = true; |
933 |
< |
} |
934 |
< |
else if ((h = eventWaiters) != 0L) { |
935 |
< |
long nh; |
936 |
< |
int id = (((int)h) & WAITER_ID_MASK) - 1; |
937 |
< |
if (id >= 0 && id < n && (w = ws[id]) != null && |
938 |
< |
(nh = w.nextWaiter) != 0L && // keep at least one worker |
939 |
< |
UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh)) |
940 |
< |
shutdown = true; |
941 |
< |
} |
942 |
< |
if (w != null && shutdown) { |
943 |
< |
w.shutdown(); |
944 |
< |
LockSupport.unpark(w); |
945 |
< |
} |
946 |
< |
} |
947 |
< |
releaseEventWaiters(); // in case of interference |
994 |
> |
private void postBlock() { |
995 |
> |
long c; |
996 |
> |
do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, // no mask |
997 |
> |
c = ctl, c + AC_UNIT)); |
998 |
> |
int b; |
999 |
> |
do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset, |
1000 |
> |
b = blockedCount, b - 1)); |
1001 |
|
} |
1002 |
|
|
1003 |
|
/** |
1004 |
< |
* Callback from workers invoked upon each top-level action (i.e., |
1005 |
< |
* stealing a task or taking a submission and running it). |
953 |
< |
* Performs one or more of the following: |
954 |
< |
* |
955 |
< |
* 1. If the worker is active and either did not run a task |
956 |
< |
* or there are too many workers, try to set its active status |
957 |
< |
* to inactive and update activeCount. On contention, we may |
958 |
< |
* try again in this or a subsequent call. |
959 |
< |
* |
960 |
< |
* 2. If not enough total workers, help create some. |
961 |
< |
* |
962 |
< |
* 3. If there are too many running workers, suspend this worker |
963 |
< |
* (first forcing inactive if necessary). If it is not needed, |
964 |
< |
* it may be shutdown while suspended (via |
965 |
< |
* tryShutdownUnusedWorker). Otherwise, upon resume it |
966 |
< |
* rechecks running thread count and need for event sync. |
967 |
< |
* |
968 |
< |
* 4. If worker did not run a task, await the next task event via |
969 |
< |
* eventSync if necessary (first forcing inactivation), upon |
970 |
< |
* which the worker may be shutdown via |
971 |
< |
* tryShutdownUnusedWorker. Otherwise, help release any |
972 |
< |
* existing event waiters that are now releasable, |
1004 |
> |
* Possibly blocks waiting for the given task to complete, or |
1005 |
> |
* cancels the task if terminating. Fails to wait if contended. |
1006 |
|
* |
1007 |
< |
* @param w the worker |
975 |
< |
* @param ran true if worker ran a task since last call to this method |
1007 |
> |
* @param joinMe the task |
1008 |
|
*/ |
1009 |
< |
final void preStep(ForkJoinWorkerThread w, boolean ran) { |
1010 |
< |
int wec = w.lastEventCount; |
1011 |
< |
boolean active = w.active; |
1012 |
< |
boolean inactivate = false; |
1013 |
< |
int pc = parallelism; |
1014 |
< |
while (w.runState == 0) { |
1015 |
< |
int rs = runState; |
984 |
< |
if (rs >= TERMINATING) { // propagate shutdown |
985 |
< |
w.shutdown(); |
986 |
< |
break; |
987 |
< |
} |
988 |
< |
if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) && |
989 |
< |
UNSAFE.compareAndSwapInt(this, runStateOffset, rs, --rs)) { |
990 |
< |
inactivate = active = w.active = false; |
991 |
< |
if (rs == SHUTDOWN) { // all inactive and shut down |
992 |
< |
tryTerminate(false); |
993 |
< |
continue; |
994 |
< |
} |
995 |
< |
} |
996 |
< |
int wc = workerCounts; // try to suspend as spare |
997 |
< |
if ((wc & RUNNING_COUNT_MASK) > pc) { |
998 |
< |
if (!(inactivate |= active) && // must inactivate to suspend |
999 |
< |
workerCounts == wc && |
1000 |
< |
UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
1001 |
< |
wc, wc - ONE_RUNNING)) |
1002 |
< |
w.suspendAsSpare(); |
1003 |
< |
} |
1004 |
< |
else if ((wc >>> TOTAL_COUNT_SHIFT) < pc) |
1005 |
< |
helpMaintainParallelism(); // not enough workers |
1006 |
< |
else if (ran) |
1007 |
< |
break; |
1008 |
< |
else { |
1009 |
< |
long h = eventWaiters; |
1010 |
< |
int ec = eventCount; |
1011 |
< |
if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec) |
1012 |
< |
releaseEventWaiters(); // release others before waiting |
1013 |
< |
else if (ec != wec) { |
1014 |
< |
w.lastEventCount = ec; // no need to wait |
1015 |
< |
break; |
1016 |
< |
} |
1017 |
< |
else if (!(inactivate |= active)) |
1018 |
< |
eventSync(w, wec); // must inactivate before sync |
1009 |
> |
final void tryAwaitJoin(ForkJoinTask<?> joinMe) { |
1010 |
> |
int s; |
1011 |
> |
Thread.interrupted(); // clear interrupts before checking termination |
1012 |
> |
if (joinMe.status >= 0) { |
1013 |
> |
if (tryPreBlock()) { |
1014 |
> |
joinMe.tryAwaitDone(0L); |
1015 |
> |
postBlock(); |
1016 |
|
} |
1017 |
+ |
else if ((ctl & STOP_BIT) != 0L) |
1018 |
+ |
joinMe.cancelIgnoringExceptions(); |
1019 |
|
} |
1020 |
|
} |
1021 |
|
|
1022 |
|
/** |
1023 |
< |
* Helps and/or blocks awaiting join of the given task. |
1024 |
< |
* See above for explanation. |
1023 |
> |
* Possibly blocks the given worker waiting for joinMe to |
1024 |
> |
* complete or timeout |
1025 |
|
* |
1026 |
< |
* @param joinMe the task to join |
1027 |
< |
* @param worker the current worker thread |
1029 |
< |
* @param timed true if wait should time out |
1030 |
< |
* @param nanos timeout value if timed |
1026 |
> |
* @param joinMe the task |
1027 |
> |
* @param millis the wait time for underlying Object.wait |
1028 |
|
*/ |
1029 |
< |
final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker, |
1033 |
< |
boolean timed, long nanos) { |
1034 |
< |
long startTime = timed ? System.nanoTime() : 0L; |
1035 |
< |
int retries = 2 + (parallelism >> 2); // #helpJoins before blocking |
1036 |
< |
boolean running = true; // false when count decremented |
1029 |
> |
final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) { |
1030 |
|
while (joinMe.status >= 0) { |
1031 |
< |
if (runState >= TERMINATING) { |
1031 |
> |
Thread.interrupted(); |
1032 |
> |
if ((ctl & STOP_BIT) != 0L) { |
1033 |
|
joinMe.cancelIgnoringExceptions(); |
1034 |
|
break; |
1035 |
|
} |
1036 |
< |
running = worker.helpJoinTask(joinMe, running); |
1037 |
< |
if (joinMe.status < 0) |
1038 |
< |
break; |
1039 |
< |
if (retries > 0) { |
1040 |
< |
--retries; |
1041 |
< |
continue; |
1042 |
< |
} |
1043 |
< |
int wc = workerCounts; |
1044 |
< |
if ((wc & RUNNING_COUNT_MASK) != 0) { |
1045 |
< |
if (running) { |
1046 |
< |
if (!UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
1047 |
< |
wc, wc - ONE_RUNNING)) |
1054 |
< |
continue; |
1055 |
< |
running = false; |
1056 |
< |
} |
1057 |
< |
long h = eventWaiters; |
1058 |
< |
if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != eventCount) |
1059 |
< |
releaseEventWaiters(); |
1060 |
< |
if ((workerCounts & RUNNING_COUNT_MASK) != 0) { |
1061 |
< |
long ms; int ns; |
1062 |
< |
if (!timed) { |
1063 |
< |
ms = JOIN_TIMEOUT_MILLIS; |
1064 |
< |
ns = 0; |
1065 |
< |
} |
1066 |
< |
else { // at most JOIN_TIMEOUT_MILLIS per wait |
1067 |
< |
long nt = nanos - (System.nanoTime() - startTime); |
1068 |
< |
if (nt <= 0L) |
1069 |
< |
break; |
1070 |
< |
ms = nt / 1000000; |
1071 |
< |
if (ms > JOIN_TIMEOUT_MILLIS) { |
1072 |
< |
ms = JOIN_TIMEOUT_MILLIS; |
1073 |
< |
ns = 0; |
1074 |
< |
} |
1075 |
< |
else |
1076 |
< |
ns = (int) (nt % 1000000); |
1036 |
> |
if (tryPreBlock()) { |
1037 |
> |
long last = System.nanoTime(); |
1038 |
> |
while (joinMe.status >= 0) { |
1039 |
> |
long millis = TimeUnit.NANOSECONDS.toMillis(nanos); |
1040 |
> |
if (millis <= 0) |
1041 |
> |
break; |
1042 |
> |
joinMe.tryAwaitDone(millis); |
1043 |
> |
if (joinMe.status < 0) |
1044 |
> |
break; |
1045 |
> |
if ((ctl & STOP_BIT) != 0L) { |
1046 |
> |
joinMe.cancelIgnoringExceptions(); |
1047 |
> |
break; |
1048 |
|
} |
1049 |
< |
joinMe.internalAwaitDone(ms, ns); |
1049 |
> |
long now = System.nanoTime(); |
1050 |
> |
nanos -= now - last; |
1051 |
> |
last = now; |
1052 |
|
} |
1053 |
< |
if (joinMe.status < 0) |
1054 |
< |
break; |
1053 |
> |
postBlock(); |
1054 |
> |
break; |
1055 |
|
} |
1083 |
– |
helpMaintainParallelism(); |
1084 |
– |
} |
1085 |
– |
if (!running) { |
1086 |
– |
int c; |
1087 |
– |
do {} while (!UNSAFE.compareAndSwapInt |
1088 |
– |
(this, workerCountsOffset, |
1089 |
– |
c = workerCounts, c + ONE_RUNNING)); |
1056 |
|
} |
1057 |
|
} |
1058 |
|
|
1059 |
|
/** |
1060 |
< |
* Same idea as awaitJoin, but no helping, retries, or timeouts. |
1060 |
> |
* If necessary, compensates for blocker, and blocks |
1061 |
|
*/ |
1062 |
< |
final void awaitBlocker(ManagedBlocker blocker) |
1062 |
> |
private void awaitBlocker(ManagedBlocker blocker) |
1063 |
|
throws InterruptedException { |
1064 |
|
while (!blocker.isReleasable()) { |
1065 |
< |
int wc = workerCounts; |
1100 |
< |
if ((wc & RUNNING_COUNT_MASK) == 0) |
1101 |
< |
helpMaintainParallelism(); |
1102 |
< |
else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, |
1103 |
< |
wc, wc - ONE_RUNNING)) { |
1065 |
> |
if (tryPreBlock()) { |
1066 |
|
try { |
1067 |
< |
while (!blocker.isReleasable()) { |
1106 |
< |
long h = eventWaiters; |
1107 |
< |
if (h != 0L && |
1108 |
< |
(int)(h >>> EVENT_COUNT_SHIFT) != eventCount) |
1109 |
< |
releaseEventWaiters(); |
1110 |
< |
else if ((workerCounts & RUNNING_COUNT_MASK) == 0 && |
1111 |
< |
runState < TERMINATING) |
1112 |
< |
helpMaintainParallelism(); |
1113 |
< |
else if (blocker.block()) |
1114 |
< |
break; |
1115 |
< |
} |
1067 |
> |
do {} while (!blocker.isReleasable() && !blocker.block()); |
1068 |
|
} finally { |
1069 |
< |
int c; |
1118 |
< |
do {} while (!UNSAFE.compareAndSwapInt |
1119 |
< |
(this, workerCountsOffset, |
1120 |
< |
c = workerCounts, c + ONE_RUNNING)); |
1069 |
> |
postBlock(); |
1070 |
|
} |
1071 |
|
break; |
1072 |
|
} |
1073 |
|
} |
1074 |
|
} |
1075 |
|
|
1076 |
+ |
// Creating, registering and deregistring workers |
1077 |
+ |
|
1078 |
+ |
/** |
1079 |
+ |
* Tries to create and start a worker; minimally rolls back counts |
1080 |
+ |
* on failure. |
1081 |
+ |
*/ |
1082 |
+ |
private void addWorker() { |
1083 |
+ |
Throwable ex = null; |
1084 |
+ |
ForkJoinWorkerThread t = null; |
1085 |
+ |
try { |
1086 |
+ |
t = factory.newThread(this); |
1087 |
+ |
} catch (Throwable e) { |
1088 |
+ |
ex = e; |
1089 |
+ |
} |
1090 |
+ |
if (t == null) { // null or exceptional factory return |
1091 |
+ |
long c; // adjust counts |
1092 |
+ |
do {} while (!UNSAFE.compareAndSwapLong |
1093 |
+ |
(this, ctlOffset, c = ctl, |
1094 |
+ |
(((c - AC_UNIT) & AC_MASK) | |
1095 |
+ |
((c - TC_UNIT) & TC_MASK) | |
1096 |
+ |
(c & ~(AC_MASK|TC_MASK))))); |
1097 |
+ |
// Propagate exception if originating from an external caller |
1098 |
+ |
if (!tryTerminate(false) && ex != null && |
1099 |
+ |
!(Thread.currentThread() instanceof ForkJoinWorkerThread)) |
1100 |
+ |
UNSAFE.throwException(ex); |
1101 |
+ |
} |
1102 |
+ |
else |
1103 |
+ |
t.start(); |
1104 |
+ |
} |
1105 |
+ |
|
1106 |
+ |
/** |
1107 |
+ |
* Callback from ForkJoinWorkerThread constructor to assign a |
1108 |
+ |
* public name |
1109 |
+ |
*/ |
1110 |
+ |
final String nextWorkerName() { |
1111 |
+ |
for (int n;;) { |
1112 |
+ |
if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset, |
1113 |
+ |
n = nextWorkerNumber, ++n)) |
1114 |
+ |
return workerNamePrefix + n; |
1115 |
+ |
} |
1116 |
+ |
} |
1117 |
+ |
|
1118 |
+ |
/** |
1119 |
+ |
* Callback from ForkJoinWorkerThread constructor to |
1120 |
+ |
* determine its poolIndex and record in workers array. |
1121 |
+ |
* |
1122 |
+ |
* @param w the worker |
1123 |
+ |
* @return the worker's pool index |
1124 |
+ |
*/ |
1125 |
+ |
final int registerWorker(ForkJoinWorkerThread w) { |
1126 |
+ |
/* |
1127 |
+ |
* In the typical case, a new worker acquires the lock, uses |
1128 |
+ |
* next available index and returns quickly. Since we should |
1129 |
+ |
* not block callers (ultimately from signalWork or |
1130 |
+ |
* tryPreBlock) waiting for the lock needed to do this, we |
1131 |
+ |
* instead help release other workers while waiting for the |
1132 |
+ |
* lock. |
1133 |
+ |
*/ |
1134 |
+ |
for (int g;;) { |
1135 |
+ |
ForkJoinWorkerThread[] ws; |
1136 |
+ |
if (((g = scanGuard) & SG_UNIT) == 0 && |
1137 |
+ |
UNSAFE.compareAndSwapInt(this, scanGuardOffset, |
1138 |
+ |
g, g | SG_UNIT)) { |
1139 |
+ |
int k = nextWorkerIndex; |
1140 |
+ |
try { |
1141 |
+ |
if ((ws = workers) != null) { // ignore on shutdown |
1142 |
+ |
int n = ws.length; |
1143 |
+ |
if (k < 0 || k >= n || ws[k] != null) { |
1144 |
+ |
for (k = 0; k < n && ws[k] != null; ++k) |
1145 |
+ |
; |
1146 |
+ |
if (k == n) |
1147 |
+ |
ws = workers = Arrays.copyOf(ws, n << 1); |
1148 |
+ |
} |
1149 |
+ |
ws[k] = w; |
1150 |
+ |
nextWorkerIndex = k + 1; |
1151 |
+ |
int m = g & SMASK; |
1152 |
+ |
g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1); |
1153 |
+ |
} |
1154 |
+ |
} finally { |
1155 |
+ |
scanGuard = g; |
1156 |
+ |
} |
1157 |
+ |
return k; |
1158 |
+ |
} |
1159 |
+ |
else if ((ws = workers) != null) { // help release others |
1160 |
+ |
for (ForkJoinWorkerThread u : ws) { |
1161 |
+ |
if (u != null && u.queueBase != u.queueTop) { |
1162 |
+ |
if (tryReleaseWaiter()) |
1163 |
+ |
break; |
1164 |
+ |
} |
1165 |
+ |
} |
1166 |
+ |
} |
1167 |
+ |
} |
1168 |
+ |
} |
1169 |
+ |
|
1170 |
+ |
/** |
1171 |
+ |
* Final callback from terminating worker. Removes record of |
1172 |
+ |
* worker from array, and adjusts counts. If pool is shutting |
1173 |
+ |
* down, tries to complete termination. |
1174 |
+ |
* |
1175 |
+ |
* @param w the worker |
1176 |
+ |
*/ |
1177 |
+ |
final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) { |
1178 |
+ |
int idx = w.poolIndex; |
1179 |
+ |
int sc = w.stealCount; |
1180 |
+ |
int steps = 0; |
1181 |
+ |
// Remove from array, adjust worker counts and collect steal count. |
1182 |
+ |
// We can intermix failed removes or adjusts with steal updates |
1183 |
+ |
do { |
1184 |
+ |
long s, c; |
1185 |
+ |
int g; |
1186 |
+ |
if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 && |
1187 |
+ |
UNSAFE.compareAndSwapInt(this, scanGuardOffset, |
1188 |
+ |
g, g |= SG_UNIT)) { |
1189 |
+ |
ForkJoinWorkerThread[] ws = workers; |
1190 |
+ |
if (ws != null && idx >= 0 && |
1191 |
+ |
idx < ws.length && ws[idx] == w) |
1192 |
+ |
ws[idx] = null; // verify |
1193 |
+ |
nextWorkerIndex = idx; |
1194 |
+ |
scanGuard = g + SG_UNIT; |
1195 |
+ |
steps = 1; |
1196 |
+ |
} |
1197 |
+ |
if (steps == 1 && |
1198 |
+ |
UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl, |
1199 |
+ |
(((c - AC_UNIT) & AC_MASK) | |
1200 |
+ |
((c - TC_UNIT) & TC_MASK) | |
1201 |
+ |
(c & ~(AC_MASK|TC_MASK))))) |
1202 |
+ |
steps = 2; |
1203 |
+ |
if (sc != 0 && |
1204 |
+ |
UNSAFE.compareAndSwapLong(this, stealCountOffset, |
1205 |
+ |
s = stealCount, s + sc)) |
1206 |
+ |
sc = 0; |
1207 |
+ |
} while (steps != 2 || sc != 0); |
1208 |
+ |
if (!tryTerminate(false)) { |
1209 |
+ |
if (ex != null) // possibly replace if died abnormally |
1210 |
+ |
signalWork(); |
1211 |
+ |
else |
1212 |
+ |
tryReleaseWaiter(); |
1213 |
+ |
} |
1214 |
+ |
} |
1215 |
+ |
|
1216 |
+ |
// Shutdown and termination |
1217 |
+ |
|
1218 |
|
/** |
1219 |
|
* Possibly initiates and/or completes termination. |
1220 |
|
* |
1223 |
|
* @return true if now terminating or terminated |
1224 |
|
*/ |
1225 |
|
private boolean tryTerminate(boolean now) { |
1226 |
< |
if (now) |
1227 |
< |
advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN |
1228 |
< |
else if (runState < SHUTDOWN || |
1229 |
< |
!submissionQueue.isEmpty() || |
1230 |
< |
(runState & ACTIVE_COUNT_MASK) != 0) |
1231 |
< |
return false; |
1232 |
< |
|
1233 |
< |
if (advanceRunLevel(TERMINATING)) |
1234 |
< |
startTerminating(); |
1235 |
< |
|
1236 |
< |
// Finish now if all threads terminated; else in some subsequent call |
1237 |
< |
if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) { |
1238 |
< |
advanceRunLevel(TERMINATED); |
1239 |
< |
termination.forceTermination(); |
1226 |
> |
long c; |
1227 |
> |
while (((c = ctl) & STOP_BIT) == 0) { |
1228 |
> |
if (!now) { |
1229 |
> |
if ((int)(c >> AC_SHIFT) != -parallelism) |
1230 |
> |
return false; |
1231 |
> |
if (!shutdown || blockedCount != 0 || quiescerCount != 0 || |
1232 |
> |
queueBase != queueTop) { |
1233 |
> |
if (ctl == c) // staleness check |
1234 |
> |
return false; |
1235 |
> |
continue; |
1236 |
> |
} |
1237 |
> |
} |
1238 |
> |
if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT)) |
1239 |
> |
startTerminating(); |
1240 |
> |
} |
1241 |
> |
if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers |
1242 |
> |
final ReentrantLock lock = this.submissionLock; |
1243 |
> |
lock.lock(); |
1244 |
> |
try { |
1245 |
> |
termination.signalAll(); |
1246 |
> |
} finally { |
1247 |
> |
lock.unlock(); |
1248 |
> |
} |
1249 |
|
} |
1250 |
|
return true; |
1251 |
|
} |
1252 |
|
|
1253 |
|
/** |
1254 |
< |
* Actions on transition to TERMINATING |
1255 |
< |
* |
1256 |
< |
* Runs up to four passes through workers: (0) shutting down each |
1257 |
< |
* (without waking up if parked) to quickly spread notifications |
1258 |
< |
* without unnecessary bouncing around event queues etc (1) wake |
1259 |
< |
* up and help cancel tasks (2) interrupt (3) mop up races with |
1160 |
< |
* interrupted workers |
1254 |
> |
* Runs up to three passes through workers: (0) Setting |
1255 |
> |
* termination status for each worker, followed by wakeups up to |
1256 |
> |
* queued workers; (1) helping cancel tasks; (2) interrupting |
1257 |
> |
* lagging threads (likely in external tasks, but possibly also |
1258 |
> |
* blocked in joins). Each pass repeats previous steps because of |
1259 |
> |
* potential lagging thread creation. |
1260 |
|
*/ |
1261 |
|
private void startTerminating() { |
1262 |
|
cancelSubmissions(); |
1263 |
< |
for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) { |
1264 |
< |
int c; // advance event count |
1265 |
< |
UNSAFE.compareAndSwapInt(this, eventCountOffset, |
1266 |
< |
c = eventCount, c+1); |
1267 |
< |
eventWaiters = 0L; // clobber lists |
1268 |
< |
spareWaiters = 0; |
1269 |
< |
for (ForkJoinWorkerThread w : workers) { |
1270 |
< |
if (w != null) { |
1271 |
< |
w.shutdown(); |
1272 |
< |
if (passes > 0 && !w.isTerminated()) { |
1273 |
< |
w.cancelTasks(); |
1274 |
< |
LockSupport.unpark(w); |
1275 |
< |
if (passes > 1 && !w.isInterrupted()) { |
1177 |
< |
try { |
1178 |
< |
w.interrupt(); |
1179 |
< |
} catch (SecurityException ignore) { |
1263 |
> |
for (int pass = 0; pass < 3; ++pass) { |
1264 |
> |
ForkJoinWorkerThread[] ws = workers; |
1265 |
> |
if (ws != null) { |
1266 |
> |
for (ForkJoinWorkerThread w : ws) { |
1267 |
> |
if (w != null) { |
1268 |
> |
w.terminate = true; |
1269 |
> |
if (pass > 0) { |
1270 |
> |
w.cancelTasks(); |
1271 |
> |
if (pass > 1 && !w.isInterrupted()) { |
1272 |
> |
try { |
1273 |
> |
w.interrupt(); |
1274 |
> |
} catch (SecurityException ignore) { |
1275 |
> |
} |
1276 |
|
} |
1277 |
|
} |
1278 |
|
} |
1279 |
|
} |
1280 |
+ |
terminateWaiters(); |
1281 |
|
} |
1282 |
|
} |
1283 |
|
} |
1284 |
|
|
1285 |
|
/** |
1286 |
< |
* Clears out and cancels submissions, ignoring exceptions. |
1286 |
> |
* Polls and cancels all submissions. Called only during termination. |
1287 |
|
*/ |
1288 |
|
private void cancelSubmissions() { |
1289 |
< |
ForkJoinTask<?> task; |
1290 |
< |
while ((task = submissionQueue.poll()) != null) { |
1291 |
< |
try { |
1292 |
< |
task.cancel(false); |
1293 |
< |
} catch (Throwable ignore) { |
1289 |
> |
while (queueBase != queueTop) { |
1290 |
> |
ForkJoinTask<?> task = pollSubmission(); |
1291 |
> |
if (task != null) { |
1292 |
> |
try { |
1293 |
> |
task.cancel(false); |
1294 |
> |
} catch (Throwable ignore) { |
1295 |
> |
} |
1296 |
|
} |
1297 |
|
} |
1298 |
|
} |
1299 |
|
|
1300 |
< |
// misc support for ForkJoinWorkerThread |
1300 |
> |
/** |
1301 |
> |
* Tries to set the termination status of waiting workers, and |
1302 |
> |
* then wakes them up (after which they will terminate). |
1303 |
> |
*/ |
1304 |
> |
private void terminateWaiters() { |
1305 |
> |
ForkJoinWorkerThread[] ws = workers; |
1306 |
> |
if (ws != null) { |
1307 |
> |
ForkJoinWorkerThread w; long c; int i, e; |
1308 |
> |
int n = ws.length; |
1309 |
> |
while ((i = ~(e = (int)(c = ctl)) & SMASK) < n && |
1310 |
> |
(w = ws[i]) != null && w.eventCount == (e & E_MASK)) { |
1311 |
> |
if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, |
1312 |
> |
(long)(w.nextWait & E_MASK) | |
1313 |
> |
((c + AC_UNIT) & AC_MASK) | |
1314 |
> |
(c & (TC_MASK|STOP_BIT)))) { |
1315 |
> |
w.terminate = true; |
1316 |
> |
w.eventCount = e + EC_UNIT; |
1317 |
> |
if (w.parked) |
1318 |
> |
UNSAFE.unpark(w); |
1319 |
> |
} |
1320 |
> |
} |
1321 |
> |
} |
1322 |
> |
} |
1323 |
> |
|
1324 |
> |
// misc ForkJoinWorkerThread support |
1325 |
|
|
1326 |
|
/** |
1327 |
< |
* Returns pool number. |
1327 |
> |
* Increment or decrement quiescerCount. Needed only to prevent |
1328 |
> |
* triggering shutdown if a worker is transiently inactive while |
1329 |
> |
* checking quiescence. |
1330 |
> |
* |
1331 |
> |
* @param delta 1 for increment, -1 for decrement |
1332 |
|
*/ |
1333 |
< |
final int getPoolNumber() { |
1334 |
< |
return poolNumber; |
1333 |
> |
final void addQuiescerCount(int delta) { |
1334 |
> |
int c; |
1335 |
> |
do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset, |
1336 |
> |
c = quiescerCount, c + delta)); |
1337 |
|
} |
1338 |
|
|
1339 |
|
/** |
1340 |
< |
* Tries to accumulate steal count from a worker, clearing |
1341 |
< |
* the worker's value if successful. |
1340 |
> |
* Directly increment or decrement active count without |
1341 |
> |
* queuing. This method is used to transiently assert inactivation |
1342 |
> |
* while checking quiescence. |
1343 |
|
* |
1344 |
< |
* @return true if worker steal count now zero |
1344 |
> |
* @param delta 1 for increment, -1 for decrement |
1345 |
|
*/ |
1346 |
< |
final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) { |
1347 |
< |
int sc = w.stealCount; |
1348 |
< |
long c = stealCount; |
1349 |
< |
// CAS even if zero, for fence effects |
1350 |
< |
if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) { |
1351 |
< |
if (sc != 0) |
1222 |
< |
w.stealCount = 0; |
1223 |
< |
return true; |
1224 |
< |
} |
1225 |
< |
return sc == 0; |
1346 |
> |
final void addActiveCount(int delta) { |
1347 |
> |
long d = delta < 0 ? -AC_UNIT : AC_UNIT; |
1348 |
> |
long c; |
1349 |
> |
do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl, |
1350 |
> |
((c + d) & AC_MASK) | |
1351 |
> |
(c & ~AC_MASK))); |
1352 |
|
} |
1353 |
|
|
1354 |
|
/** |
1356 |
|
* active thread. |
1357 |
|
*/ |
1358 |
|
final int idlePerActive() { |
1359 |
< |
int pc = parallelism; // use parallelism, not rc |
1360 |
< |
int ac = runState; // no mask -- artificially boosts during shutdown |
1361 |
< |
// Use exact results for small values, saturate past 4 |
1362 |
< |
return ((pc <= ac) ? 0 : |
1363 |
< |
(pc >>> 1 <= ac) ? 1 : |
1364 |
< |
(pc >>> 2 <= ac) ? 3 : |
1365 |
< |
pc >>> 3); |
1359 |
> |
// Approximate at powers of two for small values, saturate past 4 |
1360 |
> |
int p = parallelism; |
1361 |
> |
int a = p + (int)(ctl >> AC_SHIFT); |
1362 |
> |
return (a > (p >>>= 1) ? 0 : |
1363 |
> |
a > (p >>>= 1) ? 1 : |
1364 |
> |
a > (p >>>= 1) ? 2 : |
1365 |
> |
a > (p >>>= 1) ? 4 : |
1366 |
> |
8); |
1367 |
|
} |
1368 |
|
|
1369 |
< |
// Public and protected methods |
1369 |
> |
// Exported methods |
1370 |
|
|
1371 |
|
// Constructors |
1372 |
|
|
1435 |
|
checkPermission(); |
1436 |
|
if (factory == null) |
1437 |
|
throw new NullPointerException(); |
1438 |
< |
if (parallelism <= 0 || parallelism > MAX_WORKERS) |
1438 |
> |
if (parallelism <= 0 || parallelism > MAX_ID) |
1439 |
|
throw new IllegalArgumentException(); |
1440 |
|
this.parallelism = parallelism; |
1441 |
|
this.factory = factory; |
1442 |
|
this.ueh = handler; |
1443 |
|
this.locallyFifo = asyncMode; |
1444 |
< |
int arraySize = initialArraySizeFor(parallelism); |
1445 |
< |
this.workers = new ForkJoinWorkerThread[arraySize]; |
1446 |
< |
this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>(); |
1447 |
< |
this.workerLock = new ReentrantLock(); |
1448 |
< |
this.termination = new Phaser(1); |
1449 |
< |
this.poolNumber = poolNumberGenerator.incrementAndGet(); |
1450 |
< |
} |
1451 |
< |
|
1452 |
< |
/** |
1453 |
< |
* Returns initial power of two size for workers array. |
1454 |
< |
* @param pc the initial parallelism level |
1455 |
< |
*/ |
1456 |
< |
private static int initialArraySizeFor(int pc) { |
1457 |
< |
// If possible, initially allocate enough space for one spare |
1458 |
< |
int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS; |
1459 |
< |
// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16) |
1460 |
< |
size |= size >>> 1; |
1334 |
< |
size |= size >>> 2; |
1335 |
< |
size |= size >>> 4; |
1336 |
< |
size |= size >>> 8; |
1337 |
< |
return size + 1; |
1444 |
> |
long np = (long)(-parallelism); // offset ctl counts |
1445 |
> |
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
1446 |
> |
this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
1447 |
> |
// initialize workers array with room for 2*parallelism if possible |
1448 |
> |
int n = parallelism << 1; |
1449 |
> |
if (n >= MAX_ID) |
1450 |
> |
n = MAX_ID; |
1451 |
> |
else { // See Hackers Delight, sec 3.2, where n < (1 << 16) |
1452 |
> |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; |
1453 |
> |
} |
1454 |
> |
workers = new ForkJoinWorkerThread[n + 1]; |
1455 |
> |
this.submissionLock = new ReentrantLock(); |
1456 |
> |
this.termination = submissionLock.newCondition(); |
1457 |
> |
StringBuilder sb = new StringBuilder("ForkJoinPool-"); |
1458 |
> |
sb.append(poolNumberGenerator.incrementAndGet()); |
1459 |
> |
sb.append("-worker-"); |
1460 |
> |
this.workerNamePrefix = sb.toString(); |
1461 |
|
} |
1462 |
|
|
1463 |
|
// Execution methods |
1464 |
|
|
1465 |
|
/** |
1343 |
– |
* Submits task and creates, starts, or resumes some workers if necessary |
1344 |
– |
*/ |
1345 |
– |
private <T> void doSubmit(ForkJoinTask<T> task) { |
1346 |
– |
submissionQueue.offer(task); |
1347 |
– |
int c; // try to increment event count -- CAS failure OK |
1348 |
– |
UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1); |
1349 |
– |
helpMaintainParallelism(); |
1350 |
– |
} |
1351 |
– |
|
1352 |
– |
/** |
1466 |
|
* Performs the given task, returning its result upon completion. |
1467 |
+ |
* If the computation encounters an unchecked Exception or Error, |
1468 |
+ |
* it is rethrown as the outcome of this invocation. Rethrown |
1469 |
+ |
* exceptions behave in the same way as regular exceptions, but, |
1470 |
+ |
* when possible, contain stack traces (as displayed for example |
1471 |
+ |
* using {@code ex.printStackTrace()}) of both the current thread |
1472 |
+ |
* as well as the thread actually encountering the exception; |
1473 |
+ |
* minimally only the latter. |
1474 |
|
* |
1475 |
|
* @param task the task |
1476 |
|
* @return the task's result |
1479 |
|
* scheduled for execution |
1480 |
|
*/ |
1481 |
|
public <T> T invoke(ForkJoinTask<T> task) { |
1482 |
+ |
Thread t = Thread.currentThread(); |
1483 |
|
if (task == null) |
1484 |
|
throw new NullPointerException(); |
1485 |
< |
if (runState >= SHUTDOWN) |
1485 |
> |
if (shutdown) |
1486 |
|
throw new RejectedExecutionException(); |
1366 |
– |
Thread t = Thread.currentThread(); |
1487 |
|
if ((t instanceof ForkJoinWorkerThread) && |
1488 |
|
((ForkJoinWorkerThread)t).pool == this) |
1489 |
|
return task.invoke(); // bypass submit if in same pool |
1490 |
|
else { |
1491 |
< |
doSubmit(task); |
1491 |
> |
addSubmission(task); |
1492 |
|
return task.join(); |
1493 |
|
} |
1494 |
|
} |
1498 |
|
* computation in the current pool, else submits as external task. |
1499 |
|
*/ |
1500 |
|
private <T> void forkOrSubmit(ForkJoinTask<T> task) { |
1501 |
< |
if (runState >= SHUTDOWN) |
1382 |
< |
throw new RejectedExecutionException(); |
1501 |
> |
ForkJoinWorkerThread w; |
1502 |
|
Thread t = Thread.currentThread(); |
1503 |
+ |
if (shutdown) |
1504 |
+ |
throw new RejectedExecutionException(); |
1505 |
|
if ((t instanceof ForkJoinWorkerThread) && |
1506 |
< |
((ForkJoinWorkerThread)t).pool == this) |
1507 |
< |
task.fork(); |
1506 |
> |
(w = (ForkJoinWorkerThread)t).pool == this) |
1507 |
> |
w.pushTask(task); |
1508 |
|
else |
1509 |
< |
doSubmit(task); |
1509 |
> |
addSubmission(task); |
1510 |
|
} |
1511 |
|
|
1512 |
|
/** |
1663 |
|
* @return the number of worker threads |
1664 |
|
*/ |
1665 |
|
public int getPoolSize() { |
1666 |
< |
return workerCounts >>> TOTAL_COUNT_SHIFT; |
1666 |
> |
return parallelism + (short)(ctl >>> TC_SHIFT); |
1667 |
|
} |
1668 |
|
|
1669 |
|
/** |
1685 |
|
* @return the number of worker threads |
1686 |
|
*/ |
1687 |
|
public int getRunningThreadCount() { |
1688 |
< |
return workerCounts & RUNNING_COUNT_MASK; |
1688 |
> |
int r = parallelism + (int)(ctl >> AC_SHIFT); |
1689 |
> |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
1690 |
|
} |
1691 |
|
|
1692 |
|
/** |
1697 |
|
* @return the number of active threads |
1698 |
|
*/ |
1699 |
|
public int getActiveThreadCount() { |
1700 |
< |
return runState & ACTIVE_COUNT_MASK; |
1700 |
> |
int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount; |
1701 |
> |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
1702 |
|
} |
1703 |
|
|
1704 |
|
/** |
1713 |
|
* @return {@code true} if all threads are currently idle |
1714 |
|
*/ |
1715 |
|
public boolean isQuiescent() { |
1716 |
< |
return (runState & ACTIVE_COUNT_MASK) == 0; |
1716 |
> |
return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0; |
1717 |
|
} |
1718 |
|
|
1719 |
|
/** |
1743 |
|
*/ |
1744 |
|
public long getQueuedTaskCount() { |
1745 |
|
long count = 0; |
1746 |
< |
for (ForkJoinWorkerThread w : workers) |
1747 |
< |
if (w != null) |
1748 |
< |
count += w.getQueueSize(); |
1746 |
> |
ForkJoinWorkerThread[] ws; |
1747 |
> |
if ((short)(ctl >>> TC_SHIFT) > -parallelism && |
1748 |
> |
(ws = workers) != null) { |
1749 |
> |
for (ForkJoinWorkerThread w : ws) |
1750 |
> |
if (w != null) |
1751 |
> |
count -= w.queueBase - w.queueTop; // must read base first |
1752 |
> |
} |
1753 |
|
return count; |
1754 |
|
} |
1755 |
|
|
1756 |
|
/** |
1757 |
|
* Returns an estimate of the number of tasks submitted to this |
1758 |
< |
* pool that have not yet begun executing. This method takes time |
1759 |
< |
* proportional to the number of submissions. |
1758 |
> |
* pool that have not yet begun executing. This method may take |
1759 |
> |
* time proportional to the number of submissions. |
1760 |
|
* |
1761 |
|
* @return the number of queued submissions |
1762 |
|
*/ |
1763 |
|
public int getQueuedSubmissionCount() { |
1764 |
< |
return submissionQueue.size(); |
1764 |
> |
return -queueBase + queueTop; |
1765 |
|
} |
1766 |
|
|
1767 |
|
/** |
1771 |
|
* @return {@code true} if there are any queued submissions |
1772 |
|
*/ |
1773 |
|
public boolean hasQueuedSubmissions() { |
1774 |
< |
return !submissionQueue.isEmpty(); |
1774 |
> |
return queueBase != queueTop; |
1775 |
|
} |
1776 |
|
|
1777 |
|
/** |
1782 |
|
* @return the next submission, or {@code null} if none |
1783 |
|
*/ |
1784 |
|
protected ForkJoinTask<?> pollSubmission() { |
1785 |
< |
return submissionQueue.poll(); |
1785 |
> |
ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i; |
1786 |
> |
while ((b = queueBase) != queueTop && |
1787 |
> |
(q = submissionQueue) != null && |
1788 |
> |
(i = (q.length - 1) & b) >= 0) { |
1789 |
> |
long u = (i << ASHIFT) + ABASE; |
1790 |
> |
if ((t = q[i]) != null && |
1791 |
> |
queueBase == b && |
1792 |
> |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
1793 |
> |
queueBase = b + 1; |
1794 |
> |
return t; |
1795 |
> |
} |
1796 |
> |
} |
1797 |
> |
return null; |
1798 |
|
} |
1799 |
|
|
1800 |
|
/** |
1815 |
|
* @return the number of elements transferred |
1816 |
|
*/ |
1817 |
|
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
1818 |
< |
int count = submissionQueue.drainTo(c); |
1819 |
< |
for (ForkJoinWorkerThread w : workers) |
1820 |
< |
if (w != null) |
1821 |
< |
count += w.drainTasksTo(c); |
1818 |
> |
int count = 0; |
1819 |
> |
while (queueBase != queueTop) { |
1820 |
> |
ForkJoinTask<?> t = pollSubmission(); |
1821 |
> |
if (t != null) { |
1822 |
> |
c.add(t); |
1823 |
> |
++count; |
1824 |
> |
} |
1825 |
> |
} |
1826 |
> |
ForkJoinWorkerThread[] ws; |
1827 |
> |
if ((short)(ctl >>> TC_SHIFT) > -parallelism && |
1828 |
> |
(ws = workers) != null) { |
1829 |
> |
for (ForkJoinWorkerThread w : ws) |
1830 |
> |
if (w != null) |
1831 |
> |
count += w.drainTasksTo(c); |
1832 |
> |
} |
1833 |
|
return count; |
1834 |
|
} |
1835 |
|
|
1844 |
|
long st = getStealCount(); |
1845 |
|
long qt = getQueuedTaskCount(); |
1846 |
|
long qs = getQueuedSubmissionCount(); |
1697 |
– |
int wc = workerCounts; |
1698 |
– |
int tc = wc >>> TOTAL_COUNT_SHIFT; |
1699 |
– |
int rc = wc & RUNNING_COUNT_MASK; |
1847 |
|
int pc = parallelism; |
1848 |
< |
int rs = runState; |
1849 |
< |
int ac = rs & ACTIVE_COUNT_MASK; |
1848 |
> |
long c = ctl; |
1849 |
> |
int tc = pc + (short)(c >>> TC_SHIFT); |
1850 |
> |
int rc = pc + (int)(c >> AC_SHIFT); |
1851 |
> |
if (rc < 0) // ignore transient negative |
1852 |
> |
rc = 0; |
1853 |
> |
int ac = rc + blockedCount; |
1854 |
> |
String level; |
1855 |
> |
if ((c & STOP_BIT) != 0) |
1856 |
> |
level = (tc == 0) ? "Terminated" : "Terminating"; |
1857 |
> |
else |
1858 |
> |
level = shutdown ? "Shutting down" : "Running"; |
1859 |
|
return super.toString() + |
1860 |
< |
"[" + runLevelToString(rs) + |
1860 |
> |
"[" + level + |
1861 |
|
", parallelism = " + pc + |
1862 |
|
", size = " + tc + |
1863 |
|
", active = " + ac + |
1868 |
|
"]"; |
1869 |
|
} |
1870 |
|
|
1715 |
– |
private static String runLevelToString(int s) { |
1716 |
– |
return ((s & TERMINATED) != 0 ? "Terminated" : |
1717 |
– |
((s & TERMINATING) != 0 ? "Terminating" : |
1718 |
– |
((s & SHUTDOWN) != 0 ? "Shutting down" : |
1719 |
– |
"Running"))); |
1720 |
– |
} |
1721 |
– |
|
1871 |
|
/** |
1872 |
|
* Initiates an orderly shutdown in which previously submitted |
1873 |
|
* tasks are executed, but no new tasks will be accepted. |
1882 |
|
*/ |
1883 |
|
public void shutdown() { |
1884 |
|
checkPermission(); |
1885 |
< |
advanceRunLevel(SHUTDOWN); |
1885 |
> |
shutdown = true; |
1886 |
|
tryTerminate(false); |
1887 |
|
} |
1888 |
|
|
1904 |
|
*/ |
1905 |
|
public List<Runnable> shutdownNow() { |
1906 |
|
checkPermission(); |
1907 |
+ |
shutdown = true; |
1908 |
|
tryTerminate(true); |
1909 |
|
return Collections.emptyList(); |
1910 |
|
} |
1915 |
|
* @return {@code true} if all tasks have completed following shut down |
1916 |
|
*/ |
1917 |
|
public boolean isTerminated() { |
1918 |
< |
return runState >= TERMINATED; |
1918 |
> |
long c = ctl; |
1919 |
> |
return ((c & STOP_BIT) != 0L && |
1920 |
> |
(short)(c >>> TC_SHIFT) == -parallelism); |
1921 |
|
} |
1922 |
|
|
1923 |
|
/** |
1934 |
|
* @return {@code true} if terminating but not yet terminated |
1935 |
|
*/ |
1936 |
|
public boolean isTerminating() { |
1937 |
< |
return (runState & (TERMINATING|TERMINATED)) == TERMINATING; |
1937 |
> |
long c = ctl; |
1938 |
> |
return ((c & STOP_BIT) != 0L && |
1939 |
> |
(short)(c >>> TC_SHIFT) != -parallelism); |
1940 |
|
} |
1941 |
|
|
1942 |
|
/** |
1943 |
|
* Returns true if terminating or terminated. Used by ForkJoinWorkerThread. |
1944 |
|
*/ |
1945 |
|
final boolean isAtLeastTerminating() { |
1946 |
< |
return runState >= TERMINATING; |
1946 |
> |
return (ctl & STOP_BIT) != 0L; |
1947 |
|
} |
1948 |
|
|
1949 |
|
/** |
1952 |
|
* @return {@code true} if this pool has been shut down |
1953 |
|
*/ |
1954 |
|
public boolean isShutdown() { |
1955 |
< |
return runState >= SHUTDOWN; |
1955 |
> |
return shutdown; |
1956 |
|
} |
1957 |
|
|
1958 |
|
/** |
1968 |
|
*/ |
1969 |
|
public boolean awaitTermination(long timeout, TimeUnit unit) |
1970 |
|
throws InterruptedException { |
1971 |
+ |
long nanos = unit.toNanos(timeout); |
1972 |
+ |
final ReentrantLock lock = this.submissionLock; |
1973 |
+ |
lock.lock(); |
1974 |
|
try { |
1975 |
< |
termination.awaitAdvanceInterruptibly(0, timeout, unit); |
1976 |
< |
} catch (TimeoutException ex) { |
1977 |
< |
return false; |
1975 |
> |
for (;;) { |
1976 |
> |
if (isTerminated()) |
1977 |
> |
return true; |
1978 |
> |
if (nanos <= 0) |
1979 |
> |
return false; |
1980 |
> |
nanos = termination.awaitNanos(nanos); |
1981 |
> |
} |
1982 |
> |
} finally { |
1983 |
> |
lock.unlock(); |
1984 |
|
} |
1822 |
– |
return true; |
1985 |
|
} |
1986 |
|
|
1987 |
|
/** |
1992 |
|
* {@code isReleasable} must return {@code true} if blocking is |
1993 |
|
* not necessary. Method {@code block} blocks the current thread |
1994 |
|
* if necessary (perhaps internally invoking {@code isReleasable} |
1995 |
< |
* before actually blocking). The unusual methods in this API |
1996 |
< |
* accommodate synchronizers that may, but don't usually, block |
1997 |
< |
* for long periods. Similarly, they allow more efficient internal |
1998 |
< |
* handling of cases in which additional workers may be, but |
1999 |
< |
* usually are not, needed to ensure sufficient parallelism. |
2000 |
< |
* Toward this end, implementations of method {@code isReleasable} |
2001 |
< |
* must be amenable to repeated invocation. |
1995 |
> |
* before actually blocking). These actions are performed by any |
1996 |
> |
* thread invoking {@link ForkJoinPool#managedBlock}. The |
1997 |
> |
* unusual methods in this API accommodate synchronizers that may, |
1998 |
> |
* but don't usually, block for long periods. Similarly, they |
1999 |
> |
* allow more efficient internal handling of cases in which |
2000 |
> |
* additional workers may be, but usually are not, needed to |
2001 |
> |
* ensure sufficient parallelism. Toward this end, |
2002 |
> |
* implementations of method {@code isReleasable} must be amenable |
2003 |
> |
* to repeated invocation. |
2004 |
|
* |
2005 |
|
* <p>For example, here is a ManagedBlocker based on a |
2006 |
|
* ReentrantLock: |
2102 |
|
} |
2103 |
|
|
2104 |
|
// Unsafe mechanics |
2105 |
< |
|
2106 |
< |
private static final sun.misc.Unsafe UNSAFE = getUnsafe(); |
2107 |
< |
private static final long workerCountsOffset = |
2108 |
< |
objectFieldOffset("workerCounts", ForkJoinPool.class); |
2109 |
< |
private static final long runStateOffset = |
2110 |
< |
objectFieldOffset("runState", ForkJoinPool.class); |
2111 |
< |
private static final long eventCountOffset = |
2112 |
< |
objectFieldOffset("eventCount", ForkJoinPool.class); |
2113 |
< |
private static final long eventWaitersOffset = |
2114 |
< |
objectFieldOffset("eventWaiters", ForkJoinPool.class); |
2115 |
< |
private static final long stealCountOffset = |
2116 |
< |
objectFieldOffset("stealCount", ForkJoinPool.class); |
2117 |
< |
private static final long spareWaitersOffset = |
2118 |
< |
objectFieldOffset("spareWaiters", ForkJoinPool.class); |
2119 |
< |
|
2120 |
< |
private static long objectFieldOffset(String field, Class<?> klazz) { |
2105 |
> |
private static final sun.misc.Unsafe UNSAFE; |
2106 |
> |
private static final long ctlOffset; |
2107 |
> |
private static final long stealCountOffset; |
2108 |
> |
private static final long blockedCountOffset; |
2109 |
> |
private static final long quiescerCountOffset; |
2110 |
> |
private static final long scanGuardOffset; |
2111 |
> |
private static final long nextWorkerNumberOffset; |
2112 |
> |
private static final long ABASE; |
2113 |
> |
private static final int ASHIFT; |
2114 |
> |
|
2115 |
> |
static { |
2116 |
> |
poolNumberGenerator = new AtomicInteger(); |
2117 |
> |
workerSeedGenerator = new Random(); |
2118 |
> |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
2119 |
> |
defaultForkJoinWorkerThreadFactory = |
2120 |
> |
new DefaultForkJoinWorkerThreadFactory(); |
2121 |
> |
int s; |
2122 |
|
try { |
2123 |
< |
return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field)); |
2124 |
< |
} catch (NoSuchFieldException e) { |
2125 |
< |
// Convert Exception to corresponding Error |
2126 |
< |
NoSuchFieldError error = new NoSuchFieldError(field); |
2127 |
< |
error.initCause(e); |
2128 |
< |
throw error; |
2129 |
< |
} |
2123 |
> |
UNSAFE = getUnsafe(); |
2124 |
> |
Class<?> k = ForkJoinPool.class; |
2125 |
> |
ctlOffset = UNSAFE.objectFieldOffset |
2126 |
> |
(k.getDeclaredField("ctl")); |
2127 |
> |
stealCountOffset = UNSAFE.objectFieldOffset |
2128 |
> |
(k.getDeclaredField("stealCount")); |
2129 |
> |
blockedCountOffset = UNSAFE.objectFieldOffset |
2130 |
> |
(k.getDeclaredField("blockedCount")); |
2131 |
> |
quiescerCountOffset = UNSAFE.objectFieldOffset |
2132 |
> |
(k.getDeclaredField("quiescerCount")); |
2133 |
> |
scanGuardOffset = UNSAFE.objectFieldOffset |
2134 |
> |
(k.getDeclaredField("scanGuard")); |
2135 |
> |
nextWorkerNumberOffset = UNSAFE.objectFieldOffset |
2136 |
> |
(k.getDeclaredField("nextWorkerNumber")); |
2137 |
> |
Class<?> a = ForkJoinTask[].class; |
2138 |
> |
ABASE = UNSAFE.arrayBaseOffset(a); |
2139 |
> |
s = UNSAFE.arrayIndexScale(a); |
2140 |
> |
} catch (Exception e) { |
2141 |
> |
throw new Error(e); |
2142 |
> |
} |
2143 |
> |
if ((s & (s-1)) != 0) |
2144 |
> |
throw new Error("data type scale not a power of two"); |
2145 |
> |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
2146 |
|
} |
2147 |
|
|
2148 |
|
/** |