8 |
|
|
9 |
|
import java.util.concurrent.*; |
10 |
|
|
11 |
+ |
import java.util.Random; |
12 |
|
import java.util.Collection; |
13 |
+ |
import java.util.concurrent.locks.LockSupport; |
14 |
|
|
15 |
|
/** |
16 |
|
* A thread managed by a {@link ForkJoinPool}. This class is |
17 |
|
* subclassable solely for the sake of adding functionality -- there |
18 |
< |
* are no overridable methods dealing with scheduling or |
19 |
< |
* execution. However, you can override initialization and termination |
20 |
< |
* methods surrounding the main task processing loop. If you do |
21 |
< |
* create such a subclass, you will also need to supply a custom |
22 |
< |
* ForkJoinWorkerThreadFactory to use it in a ForkJoinPool. |
18 |
> |
* are no overridable methods dealing with scheduling or execution. |
19 |
> |
* However, you can override initialization and termination methods |
20 |
> |
* surrounding the main task processing loop. If you do create such a |
21 |
> |
* subclass, you will also need to supply a custom {@link |
22 |
> |
* ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code |
23 |
> |
* ForkJoinPool}. |
24 |
|
* |
25 |
|
* @since 1.7 |
26 |
|
* @author Doug Lea |
27 |
|
*/ |
28 |
|
public class ForkJoinWorkerThread extends Thread { |
29 |
|
/* |
30 |
< |
* Algorithm overview: |
30 |
> |
* Overview: |
31 |
|
* |
32 |
< |
* 1. Work-Stealing: Work-stealing queues are special forms of |
33 |
< |
* Deques that support only three of the four possible |
34 |
< |
* end-operations -- push, pop, and deq (aka steal), and only do |
35 |
< |
* so under the constraints that push and pop are called only from |
36 |
< |
* the owning thread, while deq may be called from other threads. |
37 |
< |
* (If you are unfamiliar with them, you probably want to read |
38 |
< |
* Herlihy and Shavit's book "The Art of Multiprocessor |
39 |
< |
* programming", chapter 16 describing these in more detail before |
40 |
< |
* proceeding.) The main work-stealing queue design is roughly |
41 |
< |
* similar to "Dynamic Circular Work-Stealing Deque" by David |
42 |
< |
* Chase and Yossi Lev, SPAA 2005 |
43 |
< |
* (http://research.sun.com/scalable/pubs/index.html). The main |
44 |
< |
* difference ultimately stems from gc requirements that we null |
45 |
< |
* out taken slots as soon as we can, to maintain as small a |
46 |
< |
* footprint as possible even in programs generating huge numbers |
47 |
< |
* of tasks. To accomplish this, we shift the CAS arbitrating pop |
48 |
< |
* vs deq (steal) from being on the indices ("base" and "sp") to |
49 |
< |
* the slots themselves (mainly via method "casSlotNull()"). So, |
50 |
< |
* both a successful pop and deq mainly entail CAS'ing a non-null |
51 |
< |
* slot to null. Because we rely on CASes of references, we do |
52 |
< |
* not need tag bits on base or sp. They are simple ints as used |
53 |
< |
* in any circular array-based queue (see for example ArrayDeque). |
54 |
< |
* Updates to the indices must still be ordered in a way that |
55 |
< |
* guarantees that (sp - base) > 0 means the queue is empty, but |
56 |
< |
* otherwise may err on the side of possibly making the queue |
57 |
< |
* appear nonempty when a push, pop, or deq have not fully |
58 |
< |
* committed. Note that this means that the deq operation, |
59 |
< |
* considered individually, is not wait-free. One thief cannot |
60 |
< |
* successfully continue until another in-progress one (or, if |
61 |
< |
* previously empty, a push) completes. However, in the |
62 |
< |
* aggregate, we ensure at least probabilistic non-blockingness. If |
63 |
< |
* an attempted steal fails, a thief always chooses a different |
32 |
> |
* ForkJoinWorkerThreads are managed by ForkJoinPools and perform |
33 |
> |
* ForkJoinTasks. This class includes bookkeeping in support of |
34 |
> |
* worker activation, suspension, and lifecycle control described |
35 |
> |
* in more detail in the internal documentation of class |
36 |
> |
* ForkJoinPool. And as described further below, this class also |
37 |
> |
* includes special-cased support for some ForkJoinTask |
38 |
> |
* methods. But the main mechanics involve work-stealing: |
39 |
> |
* |
40 |
> |
* Work-stealing queues are special forms of Deques that support |
41 |
> |
* only three of the four possible end-operations -- push, pop, |
42 |
> |
* and deq (aka steal), under the further constraints that push |
43 |
> |
* and pop are called only from the owning thread, while deq may |
44 |
> |
* be called from other threads. (If you are unfamiliar with |
45 |
> |
* them, you probably want to read Herlihy and Shavit's book "The |
46 |
> |
* Art of Multiprocessor programming", chapter 16 describing these |
47 |
> |
* in more detail before proceeding.) The main work-stealing |
48 |
> |
* queue design is roughly similar to those in the papers "Dynamic |
49 |
> |
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
50 |
> |
* (http://research.sun.com/scalable/pubs/index.html) and |
51 |
> |
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
52 |
> |
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
53 |
> |
* The main differences ultimately stem from gc requirements that |
54 |
> |
* we null out taken slots as soon as we can, to maintain as small |
55 |
> |
* a footprint as possible even in programs generating huge |
56 |
> |
* numbers of tasks. To accomplish this, we shift the CAS |
57 |
> |
* arbitrating pop vs deq (steal) from being on the indices |
58 |
> |
* ("base" and "sp") to the slots themselves (mainly via method |
59 |
> |
* "casSlotNull()"). So, both a successful pop and deq mainly |
60 |
> |
* entail a CAS of a slot from non-null to null. Because we rely |
61 |
> |
* on CASes of references, we do not need tag bits on base or sp. |
62 |
> |
* They are simple ints as used in any circular array-based queue |
63 |
> |
* (see for example ArrayDeque). Updates to the indices must |
64 |
> |
* still be ordered in a way that guarantees that sp == base means |
65 |
> |
* the queue is empty, but otherwise may err on the side of |
66 |
> |
* possibly making the queue appear nonempty when a push, pop, or |
67 |
> |
* deq have not fully committed. Note that this means that the deq |
68 |
> |
* operation, considered individually, is not wait-free. One thief |
69 |
> |
* cannot successfully continue until another in-progress one (or, |
70 |
> |
* if previously empty, a push) completes. However, in the |
71 |
> |
* aggregate, we ensure at least probabilistic non-blockingness. |
72 |
> |
* If an attempted steal fails, a thief always chooses a different |
73 |
|
* random victim target to try next. So, in order for one thief to |
74 |
|
* progress, it suffices for any in-progress deq or new push on |
75 |
|
* any empty queue to complete. One reason this works well here is |
76 |
|
* that apparently-nonempty often means soon-to-be-stealable, |
77 |
< |
* which gives threads a chance to activate if necessary before |
78 |
< |
* stealing (see below). |
77 |
> |
* which gives threads a chance to set activation status if |
78 |
> |
* necessary before stealing. |
79 |
|
* |
80 |
|
* This approach also enables support for "async mode" where local |
81 |
|
* task processing is in FIFO, not LIFO order; simply by using a |
86 |
|
* Efficient implementation of this approach currently relies on |
87 |
|
* an uncomfortable amount of "Unsafe" mechanics. To maintain |
88 |
|
* correct orderings, reads and writes of variable base require |
89 |
< |
* volatile ordering. Variable sp does not require volatile write |
90 |
< |
* but needs cheaper store-ordering on writes. Because they are |
91 |
< |
* protected by volatile base reads, reads of the queue array and |
92 |
< |
* its slots do not need volatile load semantics, but writes (in |
93 |
< |
* push) require store order and CASes (in pop and deq) require |
94 |
< |
* (volatile) CAS semantics. Since these combinations aren't |
95 |
< |
* supported using ordinary volatiles, the only way to accomplish |
96 |
< |
* these efficiently is to use direct Unsafe calls. (Using external |
89 |
> |
* volatile ordering. Variable sp does not require volatile |
90 |
> |
* writes but still needs store-ordering, which we accomplish by |
91 |
> |
* pre-incrementing sp before filling the slot with an ordered |
92 |
> |
* store. (Pre-incrementing also enables backouts used in |
93 |
> |
* scanWhileJoining.) Because they are protected by volatile base |
94 |
> |
* reads, reads of the queue array and its slots by other threads |
95 |
> |
* do not need volatile load semantics, but writes (in push) |
96 |
> |
* require store order and CASes (in pop and deq) require |
97 |
> |
* (volatile) CAS semantics. (Michael, Saraswat, and Vechev's |
98 |
> |
* algorithm has similar properties, but without support for |
99 |
> |
* nulling slots.) Since these combinations aren't supported |
100 |
> |
* using ordinary volatiles, the only way to accomplish these |
101 |
> |
* efficiently is to use direct Unsafe calls. (Using external |
102 |
|
* AtomicIntegers and AtomicReferenceArrays for the indices and |
103 |
|
* array is significantly slower because of memory locality and |
104 |
< |
* indirection effects.) Further, performance on most platforms is |
105 |
< |
* very sensitive to placement and sizing of the (resizable) queue |
106 |
< |
* array. Even though these queues don't usually become all that |
107 |
< |
* big, the initial size must be large enough to counteract cache |
104 |
> |
* indirection effects.) |
105 |
> |
* |
106 |
> |
* Further, performance on most platforms is very sensitive to |
107 |
> |
* placement and sizing of the (resizable) queue array. Even |
108 |
> |
* though these queues don't usually become all that big, the |
109 |
> |
* initial size must be large enough to counteract cache |
110 |
|
* contention effects across multiple queues (especially in the |
111 |
|
* presence of GC cardmarking). Also, to improve thread-locality, |
112 |
< |
* queues are currently initialized immediately after the thread |
113 |
< |
* gets the initial signal to start processing tasks. However, |
114 |
< |
* all queue-related methods except pushTask are written in a way |
115 |
< |
* that allows them to instead be lazily allocated and/or disposed |
116 |
< |
* of when empty. All together, these low-level implementation |
117 |
< |
* choices produce as much as a factor of 4 performance |
118 |
< |
* improvement compared to naive implementations, and enable the |
119 |
< |
* processing of billions of tasks per second, sometimes at the |
120 |
< |
* expense of ugliness. |
121 |
< |
* |
122 |
< |
* 2. Run control: The primary run control is based on a global |
123 |
< |
* counter (activeCount) held by the pool. It uses an algorithm |
124 |
< |
* similar to that in Herlihy and Shavit section 17.6 to cause |
106 |
< |
* threads to eventually block when all threads declare they are |
107 |
< |
* inactive. (See variable "scans".) For this to work, threads |
108 |
< |
* must be declared active when executing tasks, and before |
109 |
< |
* stealing a task. They must be inactive before blocking on the |
110 |
< |
* Pool Barrier (awaiting a new submission or other Pool |
111 |
< |
* event). In between, there is some free play which we take |
112 |
< |
* advantage of to avoid contention and rapid flickering of the |
113 |
< |
* global activeCount: If inactive, we activate only if a victim |
114 |
< |
* queue appears to be nonempty (see above). Similarly, a thread |
115 |
< |
* tries to inactivate only after a full scan of other threads. |
116 |
< |
* The net effect is that contention on activeCount is rarely a |
117 |
< |
* measurable performance issue. (There are also a few other cases |
118 |
< |
* where we scan for work rather than retry/block upon |
119 |
< |
* contention.) |
120 |
< |
* |
121 |
< |
* 3. Selection control. We maintain policy of always choosing to |
122 |
< |
* run local tasks rather than stealing, and always trying to |
123 |
< |
* steal tasks before trying to run a new submission. All steals |
124 |
< |
* are currently performed in randomly-chosen deq-order. It may be |
125 |
< |
* worthwhile to bias these with locality / anti-locality |
126 |
< |
* information, but doing this well probably requires more |
127 |
< |
* lower-level information from JVMs than currently provided. |
112 |
> |
* queues are initialized after starting. All together, these |
113 |
> |
* low-level implementation choices produce as much as a factor of |
114 |
> |
* 4 performance improvement compared to naive implementations, |
115 |
> |
* and enable the processing of billions of tasks per second, |
116 |
> |
* sometimes at the expense of ugliness. |
117 |
> |
*/ |
118 |
> |
|
119 |
> |
/** |
120 |
> |
* Generator for initial random seeds for random victim |
121 |
> |
* selection. This is used only to create initial seeds. Random |
122 |
> |
* steals use a cheaper xorshift generator per steal attempt. We |
123 |
> |
* expect only rare contention on seedGenerator, so just use a |
124 |
> |
* plain Random. |
125 |
|
*/ |
126 |
+ |
private static final Random seedGenerator = new Random(); |
127 |
+ |
|
128 |
+ |
/** |
129 |
+ |
* The timeout value for suspending spares. Spare workers that |
130 |
+ |
* remain unsignalled for more than this time may be trimmed |
131 |
+ |
* (killed and removed from pool). Since our goal is to avoid |
132 |
+ |
* long-term thread buildup, the exact value of timeout does not |
133 |
+ |
* matter too much so long as it avoids most false-alarm timeouts |
134 |
+ |
* under GC stalls or momentarily high system load. |
135 |
+ |
*/ |
136 |
+ |
private static final long SPARE_KEEPALIVE_NANOS = |
137 |
+ |
5L * 1000L * 1000L * 1000L; // 5 secs |
138 |
|
|
139 |
|
/** |
140 |
|
* Capacity of work-stealing queue array upon initialization. |
158 |
|
|
159 |
|
/** |
160 |
|
* The work-stealing queue array. Size must be a power of two. |
161 |
< |
* Initialized when thread starts, to improve memory locality. |
161 |
> |
* Initialized in onStart, to improve memory locality. |
162 |
|
*/ |
163 |
|
private ForkJoinTask<?>[] queue; |
164 |
|
|
165 |
|
/** |
157 |
– |
* Index (mod queue.length) of next queue slot to push to or pop |
158 |
– |
* from. It is written only by owner thread, via ordered store. |
159 |
– |
* Both sp and base are allowed to wrap around on overflow, but |
160 |
– |
* (sp - base) still estimates size. |
161 |
– |
*/ |
162 |
– |
private volatile int sp; |
163 |
– |
|
164 |
– |
/** |
166 |
|
* Index (mod queue.length) of least valid queue slot, which is |
167 |
|
* always the next position to steal from if nonempty. |
168 |
|
*/ |
169 |
|
private volatile int base; |
170 |
|
|
171 |
|
/** |
172 |
< |
* Activity status. When true, this worker is considered active. |
173 |
< |
* Must be false upon construction. It must be true when executing |
174 |
< |
* tasks, and BEFORE stealing a task. It must be false before |
175 |
< |
* calling pool.sync. |
176 |
< |
*/ |
177 |
< |
private boolean active; |
172 |
> |
* Index (mod queue.length) of next queue slot to push to or pop |
173 |
> |
* from. It is written only by owner thread, and accessed by other |
174 |
> |
* threads only after reading (volatile) base. Both sp and base |
175 |
> |
* are allowed to wrap around on overflow, but (sp - base) still |
176 |
> |
* estimates size. |
177 |
> |
*/ |
178 |
> |
private int sp; |
179 |
|
|
180 |
|
/** |
181 |
< |
* Run state of this worker. Supports simple versions of the usual |
182 |
< |
* shutdown/shutdownNow control. |
181 |
> |
* Run state of this worker. In addition to the usual run levels, |
182 |
> |
* tracks if this worker is suspended as a spare, and if it was |
183 |
> |
* killed (trimmed) while suspended. However, "active" status is |
184 |
> |
* maintained separately. |
185 |
|
*/ |
186 |
|
private volatile int runState; |
187 |
|
|
188 |
+ |
private static final int TERMINATING = 0x01; |
189 |
+ |
private static final int TERMINATED = 0x02; |
190 |
+ |
private static final int SUSPENDED = 0x04; // inactive spare |
191 |
+ |
private static final int TRIMMED = 0x08; // killed while suspended |
192 |
+ |
|
193 |
+ |
/** |
194 |
+ |
* Number of LockSupport.park calls to block this thread for |
195 |
+ |
* suspension or event waits. Used for internal instrumention; |
196 |
+ |
* currently not exported but included because volatile write upon |
197 |
+ |
* park also provides a workaround for a JVM bug. |
198 |
+ |
*/ |
199 |
+ |
private volatile int parkCount; |
200 |
+ |
|
201 |
+ |
/** |
202 |
+ |
* Number of steals, transferred and reset in pool callbacks pool |
203 |
+ |
* when idle Accessed directly by pool. |
204 |
+ |
*/ |
205 |
+ |
int stealCount; |
206 |
+ |
|
207 |
|
/** |
208 |
|
* Seed for random number generator for choosing steal victims. |
209 |
< |
* Uses Marsaglia xorshift. Must be nonzero upon initialization. |
209 |
> |
* Uses Marsaglia xorshift. Must be initialized as nonzero. |
210 |
|
*/ |
211 |
|
private int seed; |
212 |
|
|
213 |
|
/** |
214 |
< |
* Number of steals, transferred to pool when idle |
214 |
> |
* Activity status. When true, this worker is considered active. |
215 |
> |
* Accessed directly by pool. Must be false upon construction. |
216 |
> |
*/ |
217 |
> |
boolean active; |
218 |
> |
|
219 |
> |
/** |
220 |
> |
* True if use local fifo, not default lifo, for local polling. |
221 |
> |
* Shadows value from ForkJoinPool, which resets it if changed |
222 |
> |
* pool-wide. |
223 |
|
*/ |
224 |
< |
private int stealCount; |
224 |
> |
private boolean locallyFifo; |
225 |
|
|
226 |
|
/** |
227 |
|
* Index of this worker in pool array. Set once by pool before |
228 |
< |
* running, and accessed directly by pool during cleanup etc. |
228 |
> |
* running, and accessed directly by pool to locate this worker in |
229 |
> |
* its workers array. |
230 |
|
*/ |
231 |
|
int poolIndex; |
232 |
|
|
233 |
|
/** |
234 |
< |
* The last barrier event waited for. Accessed in pool callback |
235 |
< |
* methods, but only by current thread. |
234 |
> |
* The last pool event waited for. Accessed only by pool in |
235 |
> |
* callback methods invoked within this thread. |
236 |
|
*/ |
237 |
< |
long lastEventCount; |
237 |
> |
int lastEventCount; |
238 |
|
|
239 |
|
/** |
240 |
< |
* True if use local fifo, not default lifo, for local polling |
240 |
> |
* Encoded index and event count of next event waiter. Used only |
241 |
> |
* by ForkJoinPool for managing event waiters. |
242 |
|
*/ |
243 |
< |
private boolean locallyFifo; |
243 |
> |
volatile long nextWaiter; |
244 |
|
|
245 |
|
/** |
246 |
|
* Creates a ForkJoinWorkerThread operating in the given pool. |
251 |
|
protected ForkJoinWorkerThread(ForkJoinPool pool) { |
252 |
|
if (pool == null) throw new NullPointerException(); |
253 |
|
this.pool = pool; |
254 |
< |
// Note: poolIndex is set by pool during construction |
255 |
< |
// Remaining initialization is deferred to onStart |
254 |
> |
// To avoid exposing construction details to subclasses, |
255 |
> |
// remaining initialization is in start() and onStart() |
256 |
|
} |
257 |
|
|
258 |
< |
// Public access methods |
258 |
> |
/** |
259 |
> |
* Performs additional initialization and starts this thread |
260 |
> |
*/ |
261 |
> |
final void start(int poolIndex, boolean locallyFifo, |
262 |
> |
UncaughtExceptionHandler ueh) { |
263 |
> |
this.poolIndex = poolIndex; |
264 |
> |
this.locallyFifo = locallyFifo; |
265 |
> |
if (ueh != null) |
266 |
> |
setUncaughtExceptionHandler(ueh); |
267 |
> |
setDaemon(true); |
268 |
> |
start(); |
269 |
> |
} |
270 |
> |
|
271 |
> |
// Public/protected methods |
272 |
|
|
273 |
|
/** |
274 |
|
* Returns the pool hosting this thread. |
293 |
|
} |
294 |
|
|
295 |
|
/** |
296 |
< |
* Establishes local first-in-first-out scheduling mode for forked |
297 |
< |
* tasks that are never joined. |
298 |
< |
* |
299 |
< |
* @param async if true, use locally FIFO scheduling |
296 |
> |
* Initializes internal state after construction but before |
297 |
> |
* processing any tasks. If you override this method, you must |
298 |
> |
* invoke super.onStart() at the beginning of the method. |
299 |
> |
* Initialization requires care: Most fields must have legal |
300 |
> |
* default values, to ensure that attempted accesses from other |
301 |
> |
* threads work correctly even before this thread starts |
302 |
> |
* processing tasks. |
303 |
|
*/ |
304 |
< |
void setAsyncMode(boolean async) { |
305 |
< |
locallyFifo = async; |
306 |
< |
} |
258 |
< |
|
259 |
< |
// Runstate management |
260 |
< |
|
261 |
< |
// Runstate values. Order matters |
262 |
< |
private static final int RUNNING = 0; |
263 |
< |
private static final int SHUTDOWN = 1; |
264 |
< |
private static final int TERMINATING = 2; |
265 |
< |
private static final int TERMINATED = 3; |
266 |
< |
|
267 |
< |
final boolean isShutdown() { return runState >= SHUTDOWN; } |
268 |
< |
final boolean isTerminating() { return runState >= TERMINATING; } |
269 |
< |
final boolean isTerminated() { return runState == TERMINATED; } |
270 |
< |
final boolean shutdown() { return transitionRunStateTo(SHUTDOWN); } |
271 |
< |
final boolean shutdownNow() { return transitionRunStateTo(TERMINATING); } |
304 |
> |
protected void onStart() { |
305 |
> |
int rs = seedGenerator.nextInt(); |
306 |
> |
seed = rs == 0? 1 : rs; // seed must be nonzero |
307 |
|
|
308 |
< |
/** |
309 |
< |
* Transitions to at least the given state. |
310 |
< |
* |
311 |
< |
* @return {@code true} if not already at least at given state |
277 |
< |
*/ |
278 |
< |
private boolean transitionRunStateTo(int state) { |
279 |
< |
for (;;) { |
280 |
< |
int s = runState; |
281 |
< |
if (s >= state) |
282 |
< |
return false; |
283 |
< |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, state)) |
284 |
< |
return true; |
285 |
< |
} |
286 |
< |
} |
308 |
> |
// Allocate name string and queue array in this thread |
309 |
> |
String pid = Integer.toString(pool.getPoolNumber()); |
310 |
> |
String wid = Integer.toString(poolIndex); |
311 |
> |
setName("ForkJoinPool-" + pid + "-worker-" + wid); |
312 |
|
|
313 |
< |
/** |
289 |
< |
* Tries to set status to active; fails on contention. |
290 |
< |
*/ |
291 |
< |
private boolean tryActivate() { |
292 |
< |
if (!active) { |
293 |
< |
if (!pool.tryIncrementActiveCount()) |
294 |
< |
return false; |
295 |
< |
active = true; |
296 |
< |
} |
297 |
< |
return true; |
313 |
> |
queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
314 |
|
} |
315 |
|
|
316 |
|
/** |
317 |
< |
* Tries to set status to inactive; fails on contention. |
317 |
> |
* Performs cleanup associated with termination of this worker |
318 |
> |
* thread. If you override this method, you must invoke |
319 |
> |
* {@code super.onTermination} at the end of the overridden method. |
320 |
> |
* |
321 |
> |
* @param exception the exception causing this thread to abort due |
322 |
> |
* to an unrecoverable error, or {@code null} if completed normally |
323 |
|
*/ |
324 |
< |
private boolean tryInactivate() { |
325 |
< |
if (active) { |
326 |
< |
if (!pool.tryDecrementActiveCount()) |
327 |
< |
return false; |
328 |
< |
active = false; |
324 |
> |
protected void onTermination(Throwable exception) { |
325 |
> |
try { |
326 |
> |
cancelTasks(); |
327 |
> |
setTerminated(); |
328 |
> |
pool.workerTerminated(this); |
329 |
> |
} catch (Throwable ex) { // Shouldn't ever happen |
330 |
> |
if (exception == null) // but if so, at least rethrown |
331 |
> |
exception = ex; |
332 |
> |
} finally { |
333 |
> |
if (exception != null) |
334 |
> |
UNSAFE.throwException(exception); |
335 |
|
} |
309 |
– |
return true; |
336 |
|
} |
337 |
|
|
338 |
|
/** |
313 |
– |
* Computes next value for random victim probe. Scans don't |
314 |
– |
* require a very high quality generator, but also not a crummy |
315 |
– |
* one. Marsaglia xor-shift is cheap and works well. |
316 |
– |
*/ |
317 |
– |
private static int xorShift(int r) { |
318 |
– |
r ^= (r << 13); |
319 |
– |
r ^= (r >>> 17); |
320 |
– |
return r ^ (r << 5); |
321 |
– |
} |
322 |
– |
|
323 |
– |
// Lifecycle methods |
324 |
– |
|
325 |
– |
/** |
339 |
|
* This method is required to be public, but should never be |
340 |
|
* called explicitly. It performs the main run loop to execute |
341 |
|
* ForkJoinTasks. |
344 |
|
Throwable exception = null; |
345 |
|
try { |
346 |
|
onStart(); |
334 |
– |
pool.sync(this); // await first pool event |
347 |
|
mainLoop(); |
348 |
|
} catch (Throwable ex) { |
349 |
|
exception = ex; |
352 |
|
} |
353 |
|
} |
354 |
|
|
355 |
+ |
// helpers for run() |
356 |
+ |
|
357 |
|
/** |
358 |
< |
* Executes tasks until shut down. |
358 |
> |
* Find and execute tasks and check status while running |
359 |
|
*/ |
360 |
|
private void mainLoop() { |
361 |
< |
while (!isShutdown()) { |
362 |
< |
ForkJoinTask<?> t = pollTask(); |
363 |
< |
if (t != null || (t = pollSubmission()) != null) |
364 |
< |
t.quietlyExec(); |
365 |
< |
else if (tryInactivate()) |
366 |
< |
pool.sync(this); |
361 |
> |
boolean ran = false; // true if ran task on previous step |
362 |
> |
ForkJoinPool p = pool; |
363 |
> |
for (;;) { |
364 |
> |
p.preStep(this, ran); |
365 |
> |
if (runState != 0) |
366 |
> |
return; |
367 |
> |
ForkJoinTask<?> t; // try to get and run stolen or submitted task |
368 |
> |
if (ran = (t = scan()) != null || (t = pollSubmission()) != null) { |
369 |
> |
t.tryExec(); |
370 |
> |
if (base != sp) |
371 |
> |
runLocalTasks(); |
372 |
> |
} |
373 |
|
} |
374 |
|
} |
375 |
|
|
376 |
|
/** |
377 |
< |
* Initializes internal state after construction but before |
378 |
< |
* processing any tasks. If you override this method, you must |
359 |
< |
* invoke super.onStart() at the beginning of the method. |
360 |
< |
* Initialization requires care: Most fields must have legal |
361 |
< |
* default values, to ensure that attempted accesses from other |
362 |
< |
* threads work correctly even before this thread starts |
363 |
< |
* processing tasks. |
377 |
> |
* Runs local tasks until queue is empty or shut down. Call only |
378 |
> |
* while active. |
379 |
|
*/ |
380 |
< |
protected void onStart() { |
381 |
< |
// Allocate while starting to improve chances of thread-local |
382 |
< |
// isolation |
383 |
< |
queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
384 |
< |
// Initial value of seed need not be especially random but |
385 |
< |
// should differ across workers and must be nonzero |
386 |
< |
int p = poolIndex + 1; |
387 |
< |
seed = p + (p << 8) + (p << 16) + (p << 24); // spread bits |
380 |
> |
private void runLocalTasks() { |
381 |
> |
while (runState == 0) { |
382 |
> |
ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask(); |
383 |
> |
if (t != null) |
384 |
> |
t.tryExec(); |
385 |
> |
else if (base == sp) |
386 |
> |
break; |
387 |
> |
} |
388 |
|
} |
389 |
|
|
390 |
|
/** |
391 |
< |
* Performs cleanup associated with termination of this worker |
377 |
< |
* thread. If you override this method, you must invoke |
378 |
< |
* {@code super.onTermination} at the end of the overridden method. |
391 |
> |
* If a submission exists, try to activate and take it |
392 |
|
* |
393 |
< |
* @param exception the exception causing this thread to abort due |
381 |
< |
* to an unrecoverable error, or {@code null} if completed normally |
393 |
> |
* @return a task, if available |
394 |
|
*/ |
395 |
< |
protected void onTermination(Throwable exception) { |
396 |
< |
// Execute remaining local tasks unless aborting or terminating |
397 |
< |
while (exception == null && !pool.isTerminating() && base != sp) { |
398 |
< |
try { |
399 |
< |
ForkJoinTask<?> t = popTask(); |
400 |
< |
if (t != null) |
389 |
< |
t.quietlyExec(); |
390 |
< |
} catch (Throwable ex) { |
391 |
< |
exception = ex; |
395 |
> |
private ForkJoinTask<?> pollSubmission() { |
396 |
> |
ForkJoinPool p = pool; |
397 |
> |
while (p.hasQueuedSubmissions()) { |
398 |
> |
if (active || (active = p.tryIncrementActiveCount())) { |
399 |
> |
ForkJoinTask<?> t = p.pollSubmission(); |
400 |
> |
return t != null ? t : scan(); // if missed, rescan |
401 |
|
} |
402 |
|
} |
403 |
< |
// Cancel other tasks, transition status, notify pool, and |
395 |
< |
// propagate exception to uncaught exception handler |
396 |
< |
try { |
397 |
< |
do {} while (!tryInactivate()); // ensure inactive |
398 |
< |
cancelTasks(); |
399 |
< |
runState = TERMINATED; |
400 |
< |
pool.workerTerminated(this); |
401 |
< |
} catch (Throwable ex) { // Shouldn't ever happen |
402 |
< |
if (exception == null) // but if so, at least rethrown |
403 |
< |
exception = ex; |
404 |
< |
} finally { |
405 |
< |
if (exception != null) |
406 |
< |
ForkJoinTask.rethrowException(exception); |
407 |
< |
} |
403 |
> |
return null; |
404 |
|
} |
405 |
|
|
406 |
< |
// Intrinsics-based support for queue operations. |
407 |
< |
|
408 |
< |
/** |
409 |
< |
* Adds in store-order the given task at given slot of q to null. |
410 |
< |
* Caller must ensure q is non-null and index is in range. |
406 |
> |
/* |
407 |
> |
* Intrinsics-based atomic writes for queue slots. These are |
408 |
> |
* basically the same as methods in AtomicObjectArray, but |
409 |
> |
* specialized for (1) ForkJoinTask elements (2) requirement that |
410 |
> |
* nullness and bounds checks have already been performed by |
411 |
> |
* callers and (3) effective offsets are known not to overflow |
412 |
> |
* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't |
413 |
> |
* need corresponding version for reads: plain array reads are OK |
414 |
> |
* because they protected by other volatile reads and are |
415 |
> |
* confirmed by CASes. |
416 |
> |
* |
417 |
> |
* Most uses don't actually call these methods, but instead contain |
418 |
> |
* inlined forms that enable more predictable optimization. We |
419 |
> |
* don't define the version of write used in pushTask at all, but |
420 |
> |
* instead inline there a store-fenced array slot write. |
421 |
|
*/ |
416 |
– |
private static void setSlot(ForkJoinTask<?>[] q, int i, |
417 |
– |
ForkJoinTask<?> t) { |
418 |
– |
UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t); |
419 |
– |
} |
422 |
|
|
423 |
|
/** |
424 |
< |
* CAS given slot of q to null. Caller must ensure q is non-null |
425 |
< |
* and index is in range. |
424 |
> |
* CASes slot i of array q from t to null. Caller must ensure q is |
425 |
> |
* non-null and index is in range. |
426 |
|
*/ |
427 |
< |
private static boolean casSlotNull(ForkJoinTask<?>[] q, int i, |
428 |
< |
ForkJoinTask<?> t) { |
427 |
> |
private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i, |
428 |
> |
ForkJoinTask<?> t) { |
429 |
|
return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null); |
430 |
|
} |
431 |
|
|
432 |
|
/** |
433 |
< |
* Sets sp in store-order. |
433 |
> |
* Performs a volatile write of the given task at given slot of |
434 |
> |
* array q. Caller must ensure q is non-null and index is in |
435 |
> |
* range. This method is used only during resets and backouts. |
436 |
|
*/ |
437 |
< |
private void storeSp(int s) { |
438 |
< |
UNSAFE.putOrderedInt(this, spOffset, s); |
437 |
> |
private static final void writeSlot(ForkJoinTask<?>[] q, int i, |
438 |
> |
ForkJoinTask<?> t) { |
439 |
> |
UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t); |
440 |
|
} |
441 |
|
|
442 |
< |
// Main queue methods |
442 |
> |
// queue methods |
443 |
|
|
444 |
|
/** |
445 |
< |
* Pushes a task. Called only by current thread. |
445 |
> |
* Pushes a task. Call only from this thread. |
446 |
|
* |
447 |
|
* @param t the task. Caller must ensure non-null. |
448 |
|
*/ |
449 |
|
final void pushTask(ForkJoinTask<?> t) { |
450 |
+ |
int s; |
451 |
|
ForkJoinTask<?>[] q = queue; |
452 |
< |
int mask = q.length - 1; |
453 |
< |
int s = sp; |
454 |
< |
setSlot(q, s & mask, t); |
449 |
< |
storeSp(++s); |
450 |
< |
if ((s -= base) == 1) |
452 |
> |
int mask = q.length - 1; // implicit assert q != null |
453 |
> |
UNSAFE.putOrderedObject(q, (((s = sp++) & mask) << qShift) + qBase, t); |
454 |
> |
if ((s -= base) <= 0) |
455 |
|
pool.signalWork(); |
456 |
< |
else if (s >= mask) |
456 |
> |
else if (s + 1 >= mask) |
457 |
|
growQueue(); |
458 |
|
} |
459 |
|
|
460 |
|
/** |
461 |
|
* Tries to take a task from the base of the queue, failing if |
462 |
< |
* either empty or contended. |
462 |
> |
* empty or contended. Note: Specializations of this code appear |
463 |
> |
* in scan and scanWhileJoining. |
464 |
|
* |
465 |
|
* @return a task, or null if none or contended |
466 |
|
*/ |
467 |
|
final ForkJoinTask<?> deqTask() { |
468 |
|
ForkJoinTask<?> t; |
469 |
|
ForkJoinTask<?>[] q; |
470 |
< |
int i; |
471 |
< |
int b; |
467 |
< |
if (sp != (b = base) && |
470 |
> |
int b, i; |
471 |
> |
if ((b = base) != sp && |
472 |
|
(q = queue) != null && // must read q after b |
473 |
|
(t = q[i = (q.length - 1) & b]) != null && |
474 |
< |
casSlotNull(q, i, t)) { |
474 |
> |
UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) { |
475 |
|
base = b + 1; |
476 |
|
return t; |
477 |
|
} |
479 |
|
} |
480 |
|
|
481 |
|
/** |
482 |
< |
* Tries to take a task from the base of own queue, activating if |
483 |
< |
* necessary, failing only if empty. Called only by current thread. |
482 |
> |
* Tries to take a task from the base of own queue. Assumes active |
483 |
> |
* status. Called only by current thread. |
484 |
|
* |
485 |
|
* @return a task, or null if none |
486 |
|
*/ |
487 |
|
final ForkJoinTask<?> locallyDeqTask() { |
488 |
< |
int b; |
489 |
< |
while (sp != (b = base)) { |
490 |
< |
if (tryActivate()) { |
491 |
< |
ForkJoinTask<?>[] q = queue; |
492 |
< |
int i = (q.length - 1) & b; |
493 |
< |
ForkJoinTask<?> t = q[i]; |
494 |
< |
if (t != null && casSlotNull(q, i, t)) { |
488 |
> |
ForkJoinTask<?>[] q = queue; |
489 |
> |
if (q != null) { |
490 |
> |
ForkJoinTask<?> t; |
491 |
> |
int b, i; |
492 |
> |
while (sp != (b = base)) { |
493 |
> |
if ((t = q[i = (q.length - 1) & b]) != null && |
494 |
> |
UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, |
495 |
> |
t, null)) { |
496 |
|
base = b + 1; |
497 |
|
return t; |
498 |
|
} |
502 |
|
} |
503 |
|
|
504 |
|
/** |
505 |
< |
* Returns a popped task, or null if empty. Ensures active status |
506 |
< |
* if non-null. Called only by current thread. |
505 |
> |
* Returns a popped task, or null if empty. Assumes active status. |
506 |
> |
* Called only by current thread. (Note: a specialization of this |
507 |
> |
* code appears in scanWhileJoining.) |
508 |
|
*/ |
509 |
|
final ForkJoinTask<?> popTask() { |
510 |
< |
int s = sp; |
511 |
< |
while (s != base) { |
512 |
< |
if (tryActivate()) { |
513 |
< |
ForkJoinTask<?>[] q = queue; |
514 |
< |
int mask = q.length - 1; |
515 |
< |
int i = (s - 1) & mask; |
516 |
< |
ForkJoinTask<?> t = q[i]; |
517 |
< |
if (t == null || !casSlotNull(q, i, t)) |
512 |
< |
break; |
513 |
< |
storeSp(s - 1); |
510 |
> |
int s; |
511 |
> |
ForkJoinTask<?>[] q = queue; |
512 |
> |
if (q != null && (s = sp) != base) { |
513 |
> |
int i = (q.length - 1) & --s; |
514 |
> |
ForkJoinTask<?> t = q[i]; |
515 |
> |
if (t != null && UNSAFE.compareAndSwapObject |
516 |
> |
(q, (i << qShift) + qBase, t, null)) { |
517 |
> |
sp = s; |
518 |
|
return t; |
519 |
|
} |
520 |
|
} |
529 |
|
* @param t the task. Caller must ensure non-null. |
530 |
|
*/ |
531 |
|
final boolean unpushTask(ForkJoinTask<?> t) { |
532 |
+ |
int s; |
533 |
|
ForkJoinTask<?>[] q = queue; |
534 |
< |
int mask = q.length - 1; |
535 |
< |
int s = sp - 1; |
536 |
< |
if (casSlotNull(q, s & mask, t)) { |
532 |
< |
storeSp(s); |
534 |
> |
if (q != null && UNSAFE.compareAndSwapObject |
535 |
> |
(q, (((q.length - 1) & (s = sp - 1)) << qShift) + qBase, t, null)){ |
536 |
> |
sp = s; |
537 |
|
return true; |
538 |
|
} |
539 |
|
return false; |
573 |
|
ForkJoinTask<?> t = oldQ[oldIndex]; |
574 |
|
if (t != null && !casSlotNull(oldQ, oldIndex, t)) |
575 |
|
t = null; |
576 |
< |
setSlot(newQ, b & newMask, t); |
576 |
> |
writeSlot(newQ, b & newMask, t); |
577 |
|
} while (++b != bf); |
578 |
|
pool.signalWork(); |
579 |
|
} |
580 |
|
|
581 |
|
/** |
582 |
+ |
* Computes next value for random victim probe in scan(). Scans |
583 |
+ |
* don't require a very high quality generator, but also not a |
584 |
+ |
* crummy one. Marsaglia xor-shift is cheap and works well enough. |
585 |
+ |
* Note: This is manually inlined in scan() |
586 |
+ |
*/ |
587 |
+ |
private static final int xorShift(int r) { |
588 |
+ |
r ^= r << 13; |
589 |
+ |
r ^= r >>> 17; |
590 |
+ |
return r ^ (r << 5); |
591 |
+ |
} |
592 |
+ |
|
593 |
+ |
/** |
594 |
|
* Tries to steal a task from another worker. Starts at a random |
595 |
|
* index of workers array, and probes workers until finding one |
596 |
|
* with non-empty queue or finding that all are empty. It |
597 |
|
* randomly selects the first n probes. If these are empty, it |
598 |
< |
* resorts to a full circular traversal, which is necessary to |
599 |
< |
* accurately set active status by caller. Also restarts if pool |
600 |
< |
* events occurred since last scan, which forces refresh of |
601 |
< |
* workers array, in case barrier was associated with resize. |
598 |
> |
* resorts to a circular sweep, which is necessary to accurately |
599 |
> |
* set active status. (The circular sweep uses steps of |
600 |
> |
* approximately half the array size plus 1, to avoid bias |
601 |
> |
* stemming from leftmost packing of the array in ForkJoinPool.) |
602 |
|
* |
603 |
|
* This method must be both fast and quiet -- usually avoiding |
604 |
|
* memory accesses that could disrupt cache sharing etc other than |
605 |
< |
* those needed to check for and take tasks. This accounts for, |
606 |
< |
* among other things, updating random seed in place without |
607 |
< |
* storing it until exit. |
605 |
> |
* those needed to check for and take tasks (or to activate if not |
606 |
> |
* already active). This accounts for, among other things, |
607 |
> |
* updating random seed in place without storing it until exit. |
608 |
|
* |
609 |
|
* @return a task, or null if none found |
610 |
|
*/ |
611 |
|
private ForkJoinTask<?> scan() { |
612 |
< |
ForkJoinTask<?> t = null; |
613 |
< |
int r = seed; // extract once to keep scan quiet |
614 |
< |
ForkJoinWorkerThread[] ws; // refreshed on outer loop |
615 |
< |
int mask; // must be power 2 minus 1 and > 0 |
616 |
< |
outer:do { |
617 |
< |
if ((ws = pool.workers) != null && (mask = ws.length - 1) > 0) { |
618 |
< |
int idx = r; |
619 |
< |
int probes = ~mask; // use random index while negative |
620 |
< |
for (;;) { |
621 |
< |
r = xorShift(r); // update random seed |
622 |
< |
ForkJoinWorkerThread v = ws[mask & idx]; |
623 |
< |
if (v == null || v.sp == v.base) { |
624 |
< |
if (probes <= mask) |
625 |
< |
idx = (probes++ < 0) ? r : (idx + 1); |
626 |
< |
else |
627 |
< |
break; |
612 |
> |
ForkJoinPool p = pool; |
613 |
> |
ForkJoinWorkerThread[] ws = p.workers; |
614 |
> |
int n = ws.length; // upper bound of #workers |
615 |
> |
boolean canSteal = active; // shadow active status |
616 |
> |
int r = seed; // extract seed once |
617 |
> |
int k = r; // index: random if j<0 else step |
618 |
> |
for (int j = -n; j < n; ++j) { |
619 |
> |
ForkJoinWorkerThread v = ws[k & (n - 1)]; |
620 |
> |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
621 |
> |
if (v != null && v.base != v.sp) { |
622 |
> |
if (canSteal || // ensure active status |
623 |
> |
(canSteal = active = p.tryIncrementActiveCount())) { |
624 |
> |
int b, i; // inlined specialization of deqTask |
625 |
> |
ForkJoinTask<?> t; |
626 |
> |
ForkJoinTask<?>[] q; |
627 |
> |
if ((b = v.base) != v.sp && // recheck |
628 |
> |
(q = v.queue) != null && |
629 |
> |
(t = q[i = (q.length - 1) & b]) != null && |
630 |
> |
UNSAFE.compareAndSwapObject |
631 |
> |
(q, (i << qShift) + qBase, t, null)) { |
632 |
> |
v.base = b + 1; |
633 |
> |
seed = r; |
634 |
> |
++stealCount; |
635 |
> |
return t; |
636 |
|
} |
613 |
– |
else if (!tryActivate() || (t = v.deqTask()) == null) |
614 |
– |
continue outer; // restart on contention |
615 |
– |
else |
616 |
– |
break outer; |
637 |
|
} |
638 |
+ |
j = -n; // reset on contention |
639 |
|
} |
640 |
< |
} while (pool.hasNewSyncEvent(this)); // retry on pool events |
641 |
< |
seed = r; |
642 |
< |
return t; |
640 |
> |
k = j >= 0? k + ((n >>> 1) | 1) : r; |
641 |
> |
} |
642 |
> |
return null; |
643 |
|
} |
644 |
|
|
645 |
+ |
// Run State management |
646 |
+ |
|
647 |
+ |
// status check methods used mainly by ForkJoinPool |
648 |
+ |
final boolean isTerminating() { return (runState & TERMINATING) != 0; } |
649 |
+ |
final boolean isTerminated() { return (runState & TERMINATED) != 0; } |
650 |
+ |
final boolean isSuspended() { return (runState & SUSPENDED) != 0; } |
651 |
+ |
final boolean isTrimmed() { return (runState & TRIMMED) != 0; } |
652 |
+ |
|
653 |
|
/** |
654 |
< |
* Gets and removes a local or stolen task. |
626 |
< |
* |
627 |
< |
* @return a task, if available |
654 |
> |
* Sets state to TERMINATING, also resuming if suspended. |
655 |
|
*/ |
656 |
< |
final ForkJoinTask<?> pollTask() { |
657 |
< |
ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask(); |
658 |
< |
if (t == null && (t = scan()) != null) |
659 |
< |
++stealCount; |
660 |
< |
return t; |
656 |
> |
final void shutdown() { |
657 |
> |
for (;;) { |
658 |
> |
int s = runState; |
659 |
> |
if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended |
660 |
> |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
661 |
> |
(s & ~SUSPENDED) | |
662 |
> |
(TRIMMED|TERMINATING))) { |
663 |
> |
LockSupport.unpark(this); |
664 |
> |
break; |
665 |
> |
} |
666 |
> |
} |
667 |
> |
else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
668 |
> |
s | TERMINATING)) |
669 |
> |
break; |
670 |
> |
} |
671 |
> |
} |
672 |
> |
|
673 |
> |
/** |
674 |
> |
* Sets state to TERMINATED. Called only by this thread. |
675 |
> |
*/ |
676 |
> |
private void setTerminated() { |
677 |
> |
int s; |
678 |
> |
do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset, |
679 |
> |
s = runState, |
680 |
> |
s | (TERMINATING|TERMINATED))); |
681 |
|
} |
682 |
|
|
683 |
|
/** |
684 |
< |
* Gets a local task. |
684 |
> |
* Instrumented version of park. Also used by ForkJoinPool.awaitEvent |
685 |
> |
*/ |
686 |
> |
final void doPark() { |
687 |
> |
++parkCount; |
688 |
> |
LockSupport.park(this); |
689 |
> |
} |
690 |
> |
|
691 |
> |
/** |
692 |
> |
* If suspended, tries to set status to unsuspended. |
693 |
> |
* Caller must unpark to actually resume |
694 |
|
* |
695 |
< |
* @return a task, if available |
695 |
> |
* @return true if successful |
696 |
|
*/ |
697 |
< |
final ForkJoinTask<?> pollLocalTask() { |
698 |
< |
return locallyFifo ? locallyDeqTask() : popTask(); |
697 |
> |
final boolean tryUnsuspend() { |
698 |
> |
int s; |
699 |
> |
return (((s = runState) & SUSPENDED) != 0 && |
700 |
> |
UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
701 |
> |
s & ~SUSPENDED)); |
702 |
|
} |
703 |
|
|
704 |
|
/** |
705 |
< |
* Returns a pool submission, if one exists, activating first. |
705 |
> |
* Sets suspended status and blocks as spare until resumed, |
706 |
> |
* shutdown, or timed out. |
707 |
|
* |
708 |
< |
* @return a submission, if available |
708 |
> |
* @return false if trimmed |
709 |
|
*/ |
710 |
< |
private ForkJoinTask<?> pollSubmission() { |
710 |
> |
final boolean suspendAsSpare() { |
711 |
> |
for (;;) { // set suspended unless terminating |
712 |
> |
int s = runState; |
713 |
> |
if ((s & TERMINATING) != 0) { // must kill |
714 |
> |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
715 |
> |
s | (TRIMMED | TERMINATING))) |
716 |
> |
return false; |
717 |
> |
} |
718 |
> |
else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
719 |
> |
s | SUSPENDED)) |
720 |
> |
break; |
721 |
> |
} |
722 |
> |
lastEventCount = 0; // reset upon resume |
723 |
|
ForkJoinPool p = pool; |
724 |
< |
while (p.hasQueuedSubmissions()) { |
725 |
< |
ForkJoinTask<?> t; |
726 |
< |
if (tryActivate() && (t = p.pollSubmission()) != null) |
727 |
< |
return t; |
724 |
> |
p.releaseWaiters(); // help others progress |
725 |
> |
p.accumulateStealCount(this); |
726 |
> |
interrupted(); // clear/ignore interrupts |
727 |
> |
if (poolIndex < p.getParallelism()) { // untimed wait |
728 |
> |
while ((runState & SUSPENDED) != 0) |
729 |
> |
doPark(); |
730 |
> |
return true; |
731 |
|
} |
732 |
< |
return null; |
732 |
> |
return timedSuspend(); // timed wait if apparently non-core |
733 |
> |
} |
734 |
> |
|
735 |
> |
/** |
736 |
> |
* Blocks as spare until resumed or timed out |
737 |
> |
* @return false if trimmed |
738 |
> |
*/ |
739 |
> |
private boolean timedSuspend() { |
740 |
> |
long nanos = SPARE_KEEPALIVE_NANOS; |
741 |
> |
long startTime = System.nanoTime(); |
742 |
> |
while ((runState & SUSPENDED) != 0) { |
743 |
> |
++parkCount; |
744 |
> |
if ((nanos -= (System.nanoTime() - startTime)) > 0) |
745 |
> |
LockSupport.parkNanos(this, nanos); |
746 |
> |
else { // try to trim on timeout |
747 |
> |
int s = runState; |
748 |
> |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
749 |
> |
(s & ~SUSPENDED) | |
750 |
> |
(TRIMMED|TERMINATING))) |
751 |
> |
return false; |
752 |
> |
} |
753 |
> |
} |
754 |
> |
return true; |
755 |
> |
} |
756 |
> |
|
757 |
> |
// Misc support methods for ForkJoinPool |
758 |
> |
|
759 |
> |
/** |
760 |
> |
* Returns an estimate of the number of tasks in the queue. Also |
761 |
> |
* used by ForkJoinTask. |
762 |
> |
*/ |
763 |
> |
final int getQueueSize() { |
764 |
> |
return -base + sp; |
765 |
|
} |
766 |
|
|
767 |
< |
// Methods accessed only by Pool |
767 |
> |
/** |
768 |
> |
* Set locallyFifo mode. Called only by ForkJoinPool |
769 |
> |
*/ |
770 |
> |
final void setAsyncMode(boolean async) { |
771 |
> |
locallyFifo = async; |
772 |
> |
} |
773 |
|
|
774 |
|
/** |
775 |
|
* Removes and cancels all tasks in queue. Can be called from any |
776 |
|
* thread. |
777 |
|
*/ |
778 |
|
final void cancelTasks() { |
779 |
< |
ForkJoinTask<?> t; |
780 |
< |
while (base != sp && (t = deqTask()) != null) |
781 |
< |
t.cancelIgnoringExceptions(); |
779 |
> |
while (base != sp) { |
780 |
> |
ForkJoinTask<?> t = deqTask(); |
781 |
> |
if (t != null) |
782 |
> |
t.cancelIgnoringExceptions(); |
783 |
> |
} |
784 |
|
} |
785 |
|
|
786 |
|
/** |
790 |
|
*/ |
791 |
|
final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
792 |
|
int n = 0; |
793 |
< |
ForkJoinTask<?> t; |
794 |
< |
while (base != sp && (t = deqTask()) != null) { |
795 |
< |
c.add(t); |
796 |
< |
++n; |
793 |
> |
while (base != sp) { |
794 |
> |
ForkJoinTask<?> t = deqTask(); |
795 |
> |
if (t != null) { |
796 |
> |
c.add(t); |
797 |
> |
++n; |
798 |
> |
} |
799 |
|
} |
800 |
|
return n; |
801 |
|
} |
802 |
|
|
803 |
< |
/** |
688 |
< |
* Gets and clears steal count for accumulation by pool. Called |
689 |
< |
* only when known to be idle (in pool.sync and termination). |
690 |
< |
*/ |
691 |
< |
final int getAndClearStealCount() { |
692 |
< |
int sc = stealCount; |
693 |
< |
stealCount = 0; |
694 |
< |
return sc; |
695 |
< |
} |
803 |
> |
// Support methods for ForkJoinTask |
804 |
|
|
805 |
|
/** |
806 |
< |
* Returns {@code true} if at least one worker in the given array |
807 |
< |
* appears to have at least one queued task. |
806 |
> |
* Returns an estimate of the number of tasks, offset by a |
807 |
> |
* function of number of idle workers. |
808 |
|
* |
809 |
< |
* @param ws array of workers |
809 |
> |
* This method provides a cheap heuristic guide for task |
810 |
> |
* partitioning when programmers, frameworks, tools, or languages |
811 |
> |
* have little or no idea about task granularity. In essence by |
812 |
> |
* offering this method, we ask users only about tradeoffs in |
813 |
> |
* overhead vs expected throughput and its variance, rather than |
814 |
> |
* how finely to partition tasks. |
815 |
> |
* |
816 |
> |
* In a steady state strict (tree-structured) computation, each |
817 |
> |
* thread makes available for stealing enough tasks for other |
818 |
> |
* threads to remain active. Inductively, if all threads play by |
819 |
> |
* the same rules, each thread should make available only a |
820 |
> |
* constant number of tasks. |
821 |
> |
* |
822 |
> |
* The minimum useful constant is just 1. But using a value of 1 |
823 |
> |
* would require immediate replenishment upon each steal to |
824 |
> |
* maintain enough tasks, which is infeasible. Further, |
825 |
> |
* partitionings/granularities of offered tasks should minimize |
826 |
> |
* steal rates, which in general means that threads nearer the top |
827 |
> |
* of computation tree should generate more than those nearer the |
828 |
> |
* bottom. In perfect steady state, each thread is at |
829 |
> |
* approximately the same level of computation tree. However, |
830 |
> |
* producing extra tasks amortizes the uncertainty of progress and |
831 |
> |
* diffusion assumptions. |
832 |
> |
* |
833 |
> |
* So, users will want to use values larger, but not much larger |
834 |
> |
* than 1 to both smooth over transient shortages and hedge |
835 |
> |
* against uneven progress; as traded off against the cost of |
836 |
> |
* extra task overhead. We leave the user to pick a threshold |
837 |
> |
* value to compare with the results of this call to guide |
838 |
> |
* decisions, but recommend values such as 3. |
839 |
> |
* |
840 |
> |
* When all threads are active, it is on average OK to estimate |
841 |
> |
* surplus strictly locally. In steady-state, if one thread is |
842 |
> |
* maintaining say 2 surplus tasks, then so are others. So we can |
843 |
> |
* just use estimated queue length (although note that (sp - base) |
844 |
> |
* can be an overestimate because of stealers lagging increments |
845 |
> |
* of base). However, this strategy alone leads to serious |
846 |
> |
* mis-estimates in some non-steady-state conditions (ramp-up, |
847 |
> |
* ramp-down, other stalls). We can detect many of these by |
848 |
> |
* further considering the number of "idle" threads, that are |
849 |
> |
* known to have zero queued tasks, so compensate by a factor of |
850 |
> |
* (#idle/#active) threads. |
851 |
|
*/ |
852 |
< |
static boolean hasQueuedTasks(ForkJoinWorkerThread[] ws) { |
853 |
< |
if (ws != null) { |
705 |
< |
int len = ws.length; |
706 |
< |
for (int j = 0; j < 2; ++j) { // need two passes for clean sweep |
707 |
< |
for (int i = 0; i < len; ++i) { |
708 |
< |
ForkJoinWorkerThread w = ws[i]; |
709 |
< |
if (w != null && w.sp != w.base) |
710 |
< |
return true; |
711 |
< |
} |
712 |
< |
} |
713 |
< |
} |
714 |
< |
return false; |
852 |
> |
final int getEstimatedSurplusTaskCount() { |
853 |
> |
return sp - base - pool.idlePerActive(); |
854 |
|
} |
855 |
|
|
717 |
– |
// Support methods for ForkJoinTask |
718 |
– |
|
856 |
|
/** |
857 |
< |
* Returns an estimate of the number of tasks in the queue. |
857 |
> |
* Gets and removes a local task. |
858 |
> |
* |
859 |
> |
* @return a task, if available |
860 |
|
*/ |
861 |
< |
final int getQueueSize() { |
862 |
< |
// suppress momentarily negative values |
863 |
< |
return Math.max(0, sp - base); |
861 |
> |
final ForkJoinTask<?> pollLocalTask() { |
862 |
> |
while (base != sp) { |
863 |
> |
if (active || (active = pool.tryIncrementActiveCount())) |
864 |
> |
return locallyFifo? locallyDeqTask() : popTask(); |
865 |
> |
} |
866 |
> |
return null; |
867 |
|
} |
868 |
|
|
869 |
|
/** |
870 |
< |
* Returns an estimate of the number of tasks, offset by a |
871 |
< |
* function of number of idle workers. |
870 |
> |
* Gets and removes a local or stolen task. |
871 |
> |
* |
872 |
> |
* @return a task, if available |
873 |
|
*/ |
874 |
< |
final int getEstimatedSurplusTaskCount() { |
875 |
< |
// The halving approximates weighting idle vs non-idle workers |
876 |
< |
return (sp - base) - (pool.getIdleThreadCount() >>> 1); |
874 |
> |
final ForkJoinTask<?> pollTask() { |
875 |
> |
ForkJoinTask<?> t; |
876 |
> |
return (t = pollLocalTask()) != null ? t : scan(); |
877 |
|
} |
878 |
|
|
879 |
|
/** |
880 |
< |
* Scans, returning early if joinMe done. |
880 |
> |
* Returns a stolen task, if available, unless joinMe is done |
881 |
> |
* |
882 |
> |
* This method is intrinsically nonmodular. To maintain the |
883 |
> |
* property that tasks are never stolen if the awaited task is |
884 |
> |
* ready, we must interleave mechanics of scan with status |
885 |
> |
* checks. We rely here on the commit points of deq that allow us |
886 |
> |
* to cancel a steal even after CASing slot to null, but before |
887 |
> |
* adjusting base index: If, after the CAS, we see that joinMe is |
888 |
> |
* ready, we can back out by placing the task back into the slot, |
889 |
> |
* without adjusting index. The scan loop is otherwise the same as |
890 |
> |
* in scan. |
891 |
> |
* |
892 |
> |
* The outer loop cannot be allowed to run forever, because it |
893 |
> |
* could lead to a form of deadlock if all threads are executing |
894 |
> |
* this method. However, we must also be patient before giving up, |
895 |
> |
* to cope with GC stalls, transient high loads, etc. The loop |
896 |
> |
* terminates (causing caller to possibly block this thread and |
897 |
> |
* create a replacement) only after #workers clean sweeps during |
898 |
> |
* which all running threads are active. |
899 |
|
*/ |
900 |
|
final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) { |
901 |
< |
ForkJoinTask<?> t = pollTask(); |
902 |
< |
if (t != null && joinMe.status < 0 && sp == base) { |
903 |
< |
pushTask(t); // unsteal if done and this task would be stealable |
904 |
< |
t = null; |
901 |
> |
int sweeps = 0; |
902 |
> |
int r = seed; |
903 |
> |
ForkJoinPool p = pool; |
904 |
> |
p.releaseWaiters(); // help other threads progress |
905 |
> |
while (joinMe.status >= 0) { |
906 |
> |
ForkJoinWorkerThread[] ws = p.workers; |
907 |
> |
int n = ws.length; |
908 |
> |
int k = r; |
909 |
> |
for (int j = -n; j < n; ++j) { |
910 |
> |
ForkJoinWorkerThread v = ws[k & (n - 1)]; |
911 |
> |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
912 |
> |
if (v != null) { |
913 |
> |
int b = v.base; |
914 |
> |
ForkJoinTask<?>[] q; |
915 |
> |
if (b != v.sp && (q = v.queue) != null) { |
916 |
> |
int i = (q.length - 1) & b; |
917 |
> |
ForkJoinTask<?> t = q[i]; |
918 |
> |
if (t != null) { |
919 |
> |
if (joinMe.status < 0) |
920 |
> |
return null; |
921 |
> |
if (UNSAFE.compareAndSwapObject |
922 |
> |
(q, (i << qShift) + qBase, t, null)) { |
923 |
> |
if (joinMe.status < 0) { |
924 |
> |
writeSlot(q, i, t); // back out |
925 |
> |
return null; |
926 |
> |
} |
927 |
> |
v.base = b + 1; |
928 |
> |
seed = r; |
929 |
> |
++stealCount; |
930 |
> |
return t; |
931 |
> |
} |
932 |
> |
} |
933 |
> |
sweeps = 0; // ensure rescan on contention |
934 |
> |
} |
935 |
> |
} |
936 |
> |
k = j >= 0? k + ((n >>> 1) | 1) : r; |
937 |
> |
if ((j & 7) == 0 && joinMe.status < 0) // periodically recheck |
938 |
> |
return null; |
939 |
> |
} |
940 |
> |
if ((sweeps = p.inactiveCount() == 0 ? sweeps + 1 : 0) > n) |
941 |
> |
return null; |
942 |
|
} |
943 |
< |
return t; |
943 |
> |
return null; |
944 |
|
} |
945 |
|
|
946 |
|
/** |
948 |
|
*/ |
949 |
|
final void helpQuiescePool() { |
950 |
|
for (;;) { |
951 |
< |
ForkJoinTask<?> t = pollTask(); |
952 |
< |
if (t != null) |
953 |
< |
t.quietlyExec(); |
954 |
< |
else if (tryInactivate() && pool.isQuiescent()) |
955 |
< |
break; |
951 |
> |
ForkJoinTask<?> t = pollLocalTask(); |
952 |
> |
if (t != null || (t = scan()) != null) |
953 |
> |
t.tryExec(); |
954 |
> |
else { |
955 |
> |
ForkJoinPool p = pool; |
956 |
> |
if (active) { |
957 |
> |
active = false; // inactivate |
958 |
> |
do {} while (!p.tryDecrementActiveCount()); |
959 |
> |
} |
960 |
> |
if (p.isQuiescent()) { |
961 |
> |
active = true; // re-activate |
962 |
> |
do {} while (!p.tryIncrementActiveCount()); |
963 |
> |
return; |
964 |
> |
} |
965 |
> |
} |
966 |
|
} |
759 |
– |
do {} while (!tryActivate()); // re-activate on exit |
967 |
|
} |
968 |
|
|
969 |
|
// Unsafe mechanics |
970 |
|
|
971 |
|
private static final sun.misc.Unsafe UNSAFE = getUnsafe(); |
765 |
– |
private static final long spOffset = |
766 |
– |
objectFieldOffset("sp", ForkJoinWorkerThread.class); |
972 |
|
private static final long runStateOffset = |
973 |
|
objectFieldOffset("runState", ForkJoinWorkerThread.class); |
974 |
< |
private static final long qBase; |
974 |
> |
private static final long qBase = |
975 |
> |
UNSAFE.arrayBaseOffset(ForkJoinTask[].class); |
976 |
|
private static final int qShift; |
977 |
|
|
978 |
|
static { |
773 |
– |
qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class); |
979 |
|
int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class); |
980 |
|
if ((s & (s-1)) != 0) |
981 |
|
throw new Error("data type scale not a power of two"); |