| 1 |
|
/* |
| 2 |
|
* Written by Doug Lea with assistance from members of JCP JSR-166 |
| 3 |
|
* Expert Group and released to the public domain, as explained at |
| 4 |
< |
* http://creativecommons.org/licenses/publicdomain |
| 4 |
> |
* http://creativecommons.org/publicdomain/zero/1.0/ |
| 5 |
|
*/ |
| 6 |
|
|
| 7 |
|
package jsr166y; |
| 8 |
|
|
| 9 |
– |
import java.util.concurrent.*; |
| 10 |
– |
|
| 11 |
– |
import java.util.Random; |
| 9 |
|
import java.util.Collection; |
| 10 |
< |
import java.util.concurrent.locks.LockSupport; |
| 10 |
> |
import java.util.concurrent.RejectedExecutionException; |
| 11 |
|
|
| 12 |
|
/** |
| 13 |
< |
* A thread managed by a {@link ForkJoinPool}. This class is |
| 14 |
< |
* subclassable solely for the sake of adding functionality -- there |
| 15 |
< |
* are no overridable methods dealing with scheduling or execution. |
| 16 |
< |
* However, you can override initialization and termination methods |
| 17 |
< |
* surrounding the main task processing loop. If you do create such a |
| 18 |
< |
* subclass, you will also need to supply a custom {@link |
| 19 |
< |
* ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code |
| 20 |
< |
* ForkJoinPool}. |
| 13 |
> |
* A thread managed by a {@link ForkJoinPool}, which executes |
| 14 |
> |
* {@link ForkJoinTask}s. |
| 15 |
> |
* This class is subclassable solely for the sake of adding |
| 16 |
> |
* functionality -- there are no overridable methods dealing with |
| 17 |
> |
* scheduling or execution. However, you can override initialization |
| 18 |
> |
* and termination methods surrounding the main task processing loop. |
| 19 |
> |
* If you do create such a subclass, you will also need to supply a |
| 20 |
> |
* custom {@link ForkJoinPool.ForkJoinWorkerThreadFactory} to use it |
| 21 |
> |
* in a {@code ForkJoinPool}. |
| 22 |
|
* |
| 23 |
|
* @since 1.7 |
| 24 |
|
* @author Doug Lea |
| 53 |
|
* a footprint as possible even in programs generating huge |
| 54 |
|
* numbers of tasks. To accomplish this, we shift the CAS |
| 55 |
|
* arbitrating pop vs deq (steal) from being on the indices |
| 56 |
< |
* ("base" and "sp") to the slots themselves (mainly via method |
| 57 |
< |
* "casSlotNull()"). So, both a successful pop and deq mainly |
| 58 |
< |
* entail a CAS of a slot from non-null to null. Because we rely |
| 59 |
< |
* on CASes of references, we do not need tag bits on base or sp. |
| 60 |
< |
* They are simple ints as used in any circular array-based queue |
| 61 |
< |
* (see for example ArrayDeque). Updates to the indices must |
| 62 |
< |
* still be ordered in a way that guarantees that sp == base means |
| 63 |
< |
* the queue is empty, but otherwise may err on the side of |
| 64 |
< |
* possibly making the queue appear nonempty when a push, pop, or |
| 65 |
< |
* deq have not fully committed. Note that this means that the deq |
| 66 |
< |
* operation, considered individually, is not wait-free. One thief |
| 67 |
< |
* cannot successfully continue until another in-progress one (or, |
| 68 |
< |
* if previously empty, a push) completes. However, in the |
| 56 |
> |
* ("queueBase" and "queueTop") to the slots themselves (mainly |
| 57 |
> |
* via method "casSlotNull()"). So, both a successful pop and deq |
| 58 |
> |
* mainly entail a CAS of a slot from non-null to null. Because |
| 59 |
> |
* we rely on CASes of references, we do not need tag bits on |
| 60 |
> |
* queueBase or queueTop. They are simple ints as used in any |
| 61 |
> |
* circular array-based queue (see for example ArrayDeque). |
| 62 |
> |
* Updates to the indices must still be ordered in a way that |
| 63 |
> |
* guarantees that queueTop == queueBase means the queue is empty, |
| 64 |
> |
* but otherwise may err on the side of possibly making the queue |
| 65 |
> |
* appear nonempty when a push, pop, or deq have not fully |
| 66 |
> |
* committed. Note that this means that the deq operation, |
| 67 |
> |
* considered individually, is not wait-free. One thief cannot |
| 68 |
> |
* successfully continue until another in-progress one (or, if |
| 69 |
> |
* previously empty, a push) completes. However, in the |
| 70 |
|
* aggregate, we ensure at least probabilistic non-blockingness. |
| 71 |
|
* If an attempted steal fails, a thief always chooses a different |
| 72 |
|
* random victim target to try next. So, in order for one thief to |
| 73 |
|
* progress, it suffices for any in-progress deq or new push on |
| 74 |
< |
* 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 set activation status if |
| 78 |
< |
* necessary before stealing. |
| 74 |
> |
* any empty queue to complete. |
| 75 |
|
* |
| 76 |
|
* This approach also enables support for "async mode" where local |
| 77 |
|
* task processing is in FIFO, not LIFO order; simply by using a |
| 78 |
|
* version of deq rather than pop when locallyFifo is true (as set |
| 79 |
|
* by the ForkJoinPool). This allows use in message-passing |
| 80 |
< |
* frameworks in which tasks are never joined. |
| 80 |
> |
* frameworks in which tasks are never joined. However neither |
| 81 |
> |
* mode considers affinities, loads, cache localities, etc, so |
| 82 |
> |
* rarely provide the best possible performance on a given |
| 83 |
> |
* machine, but portably provide good throughput by averaging over |
| 84 |
> |
* these factors. (Further, even if we did try to use such |
| 85 |
> |
* information, we do not usually have a basis for exploiting |
| 86 |
> |
* it. For example, some sets of tasks profit from cache |
| 87 |
> |
* affinities, but others are harmed by cache pollution effects.) |
| 88 |
|
* |
| 89 |
|
* When a worker would otherwise be blocked waiting to join a |
| 90 |
|
* task, it first tries a form of linear helping: Each worker |
| 100 |
|
* technique for implementing efficient futures" SIGPLAN Notices, |
| 101 |
|
* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs |
| 102 |
|
* in that: (1) We only maintain dependency links across workers |
| 103 |
< |
* upon steals, rather than maintain per-task bookkeeping. This |
| 104 |
< |
* may require a linear scan of workers array to locate stealers, |
| 105 |
< |
* but usually doesn't because stealers leave hints (that may |
| 106 |
< |
* become stale/wrong) of where to locate the kathem. This |
| 107 |
< |
* isolates cost to when it is needed, rather than adding to |
| 108 |
< |
* per-task overhead. (2) It is "shallow", ignoring nesting and |
| 109 |
< |
* potentially cyclic mutual steals. (3) It is intentionally |
| 110 |
< |
* racy: field currentJoin is updated only while actively joining, |
| 111 |
< |
* which means that we could miss links in the chain during |
| 112 |
< |
* long-lived tasks, GC stalls etc. (4) We bound the number of |
| 113 |
< |
* attempts to find work (see MAX_HELP_DEPTH) and fall back to |
| 114 |
< |
* suspending the worker and if necessary replacing it with a |
| 115 |
< |
* spare (see ForkJoinPool.tryAwaitJoin). |
| 103 |
> |
* upon steals, rather than use per-task bookkeeping. This may |
| 104 |
> |
* require a linear scan of workers array to locate stealers, but |
| 105 |
> |
* usually doesn't because stealers leave hints (that may become |
| 106 |
> |
* stale/wrong) of where to locate them. This isolates cost to |
| 107 |
> |
* when it is needed, rather than adding to per-task overhead. |
| 108 |
> |
* (2) It is "shallow", ignoring nesting and potentially cyclic |
| 109 |
> |
* mutual steals. (3) It is intentionally racy: field currentJoin |
| 110 |
> |
* is updated only while actively joining, which means that we |
| 111 |
> |
* miss links in the chain during long-lived tasks, GC stalls etc |
| 112 |
> |
* (which is OK since blocking in such cases is usually a good |
| 113 |
> |
* idea). (4) We bound the number of attempts to find work (see |
| 114 |
> |
* MAX_HELP) and fall back to suspending the worker and if |
| 115 |
> |
* necessary replacing it with another. |
| 116 |
|
* |
| 117 |
|
* Efficient implementation of these algorithms currently relies |
| 118 |
|
* on an uncomfortable amount of "Unsafe" mechanics. To maintain |
| 119 |
< |
* correct orderings, reads and writes of variable base require |
| 120 |
< |
* volatile ordering. Variable sp does not require volatile |
| 121 |
< |
* writes but still needs store-ordering, which we accomplish by |
| 122 |
< |
* pre-incrementing sp before filling the slot with an ordered |
| 123 |
< |
* store. (Pre-incrementing also enables backouts used in |
| 124 |
< |
* joinTask.) Because they are protected by volatile base reads, |
| 125 |
< |
* reads of the queue array and its slots by other threads do not |
| 126 |
< |
* need volatile load semantics, but writes (in push) require |
| 127 |
< |
* store order and CASes (in pop and deq) require (volatile) CAS |
| 128 |
< |
* semantics. (Michael, Saraswat, and Vechev's algorithm has |
| 129 |
< |
* similar properties, but without support for nulling slots.) |
| 130 |
< |
* Since these combinations aren't supported using ordinary |
| 131 |
< |
* volatiles, the only way to accomplish these efficiently is to |
| 132 |
< |
* use direct Unsafe calls. (Using external AtomicIntegers and |
| 133 |
< |
* AtomicReferenceArrays for the indices and array is |
| 131 |
< |
* significantly slower because of memory locality and indirection |
| 132 |
< |
* effects.) |
| 119 |
> |
* correct orderings, reads and writes of variable queueBase |
| 120 |
> |
* require volatile ordering. Variable queueTop need not be |
| 121 |
> |
* volatile because non-local reads always follow those of |
| 122 |
> |
* queueBase. Similarly, because they are protected by volatile |
| 123 |
> |
* queueBase reads, reads of the queue array and its slots by |
| 124 |
> |
* other threads do not need volatile load semantics, but writes |
| 125 |
> |
* (in push) require store order and CASes (in pop and deq) |
| 126 |
> |
* require (volatile) CAS semantics. (Michael, Saraswat, and |
| 127 |
> |
* Vechev's algorithm has similar properties, but without support |
| 128 |
> |
* for nulling slots.) Since these combinations aren't supported |
| 129 |
> |
* using ordinary volatiles, the only way to accomplish these |
| 130 |
> |
* efficiently is to use direct Unsafe calls. (Using external |
| 131 |
> |
* AtomicIntegers and AtomicReferenceArrays for the indices and |
| 132 |
> |
* array is significantly slower because of memory locality and |
| 133 |
> |
* indirection effects.) |
| 134 |
|
* |
| 135 |
|
* Further, performance on most platforms is very sensitive to |
| 136 |
|
* placement and sizing of the (resizable) queue array. Even |
| 138 |
|
* initial size must be large enough to counteract cache |
| 139 |
|
* contention effects across multiple queues (especially in the |
| 140 |
|
* presence of GC cardmarking). Also, to improve thread-locality, |
| 141 |
< |
* queues are initialized after starting. All together, these |
| 141 |
< |
* low-level implementation choices produce as much as a factor of |
| 142 |
< |
* 4 performance improvement compared to naive implementations, |
| 143 |
< |
* and enable the processing of billions of tasks per second, |
| 144 |
< |
* sometimes at the expense of ugliness. |
| 145 |
< |
*/ |
| 146 |
< |
|
| 147 |
< |
/** |
| 148 |
< |
* Generator for initial random seeds for random victim |
| 149 |
< |
* selection. This is used only to create initial seeds. Random |
| 150 |
< |
* steals use a cheaper xorshift generator per steal attempt. We |
| 151 |
< |
* expect only rare contention on seedGenerator, so just use a |
| 152 |
< |
* plain Random. |
| 141 |
> |
* queues are initialized after starting. |
| 142 |
|
*/ |
| 154 |
– |
private static final Random seedGenerator = new Random(); |
| 143 |
|
|
| 144 |
|
/** |
| 145 |
< |
* The timeout value for suspending spares. Spare workers that |
| 158 |
< |
* remain unsignalled for more than this time may be trimmed |
| 159 |
< |
* (killed and removed from pool). Since our goal is to avoid |
| 160 |
< |
* long-term thread buildup, the exact value of timeout does not |
| 161 |
< |
* matter too much so long as it avoids most false-alarm timeouts |
| 162 |
< |
* under GC stalls or momentarily high system load. |
| 145 |
> |
* Mask for pool indices encoded as shorts |
| 146 |
|
*/ |
| 147 |
< |
private static final long SPARE_KEEPALIVE_NANOS = |
| 165 |
< |
5L * 1000L * 1000L * 1000L; // 5 secs |
| 166 |
< |
|
| 167 |
< |
/** |
| 168 |
< |
* The maximum stolen->joining link depth allowed in helpJoinTask. |
| 169 |
< |
* Depths for legitimate chains are unbounded, but we use a fixed |
| 170 |
< |
* constant to avoid (otherwise unchecked) cycles and bound |
| 171 |
< |
* staleness of traversal parameters at the expense of sometimes |
| 172 |
< |
* blocking when we could be helping. |
| 173 |
< |
*/ |
| 174 |
< |
private static final int MAX_HELP_DEPTH = 8; |
| 147 |
> |
private static final int SMASK = 0xffff; |
| 148 |
|
|
| 149 |
|
/** |
| 150 |
|
* Capacity of work-stealing queue array upon initialization. |
| 154 |
|
private static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
| 155 |
|
|
| 156 |
|
/** |
| 157 |
< |
* Maximum work-stealing queue array size. Must be less than or |
| 158 |
< |
* equal to 1 << 28 to ensure lack of index wraparound. (This |
| 159 |
< |
* is less than usual bounds, because we need leftshift by 3 |
| 160 |
< |
* to be in int range). |
| 157 |
> |
* Maximum size for queue array. Must be a power of two |
| 158 |
> |
* less than or equal to 1 << (31 - width of array entry) to |
| 159 |
> |
* ensure lack of index wraparound, but is capped at a lower |
| 160 |
> |
* value to help users trap runaway computations. |
| 161 |
|
*/ |
| 162 |
< |
private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28; |
| 162 |
> |
private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M |
| 163 |
> |
|
| 164 |
> |
/** |
| 165 |
> |
* The work-stealing queue array. Size must be a power of two. |
| 166 |
> |
* Initialized when started (as oposed to when constructed), to |
| 167 |
> |
* improve memory locality. |
| 168 |
> |
*/ |
| 169 |
> |
ForkJoinTask<?>[] queue; |
| 170 |
|
|
| 171 |
|
/** |
| 172 |
|
* The pool this thread works in. Accessed directly by ForkJoinTask. |
| 174 |
|
final ForkJoinPool pool; |
| 175 |
|
|
| 176 |
|
/** |
| 177 |
< |
* The work-stealing queue array. Size must be a power of two. |
| 178 |
< |
* Initialized in onStart, to improve memory locality. |
| 177 |
> |
* Index (mod queue.length) of next queue slot to push to or pop |
| 178 |
> |
* from. It is written only by owner thread, and accessed by other |
| 179 |
> |
* threads only after reading (volatile) queueBase. Both queueTop |
| 180 |
> |
* and queueBase are allowed to wrap around on overflow, but |
| 181 |
> |
* (queueTop - queueBase) still estimates size. |
| 182 |
|
*/ |
| 183 |
< |
private ForkJoinTask<?>[] queue; |
| 183 |
> |
int queueTop; |
| 184 |
|
|
| 185 |
|
/** |
| 186 |
|
* Index (mod queue.length) of least valid queue slot, which is |
| 187 |
|
* always the next position to steal from if nonempty. |
| 188 |
|
*/ |
| 189 |
< |
private volatile int base; |
| 207 |
< |
|
| 208 |
< |
/** |
| 209 |
< |
* Index (mod queue.length) of next queue slot to push to or pop |
| 210 |
< |
* from. It is written only by owner thread, and accessed by other |
| 211 |
< |
* threads only after reading (volatile) base. Both sp and base |
| 212 |
< |
* are allowed to wrap around on overflow, but (sp - base) still |
| 213 |
< |
* estimates size. |
| 214 |
< |
*/ |
| 215 |
< |
private int sp; |
| 189 |
> |
volatile int queueBase; |
| 190 |
|
|
| 191 |
|
/** |
| 192 |
|
* The index of most recent stealer, used as a hint to avoid |
| 195 |
|
* of them (usually the most current). Declared non-volatile, |
| 196 |
|
* relying on other prevailing sync to keep reasonably current. |
| 197 |
|
*/ |
| 198 |
< |
private int stealHint; |
| 198 |
> |
int stealHint; |
| 199 |
|
|
| 200 |
|
/** |
| 201 |
< |
* Run state of this worker. In addition to the usual run levels, |
| 202 |
< |
* tracks if this worker is suspended as a spare, and if it was |
| 203 |
< |
* killed (trimmed) while suspended. However, "active" status is |
| 230 |
< |
* maintained separately. |
| 201 |
> |
* Index of this worker in pool array. Set once by pool before |
| 202 |
> |
* running, and accessed directly by pool to locate this worker in |
| 203 |
> |
* its workers array. |
| 204 |
|
*/ |
| 205 |
< |
private volatile int runState; |
| 233 |
< |
|
| 234 |
< |
private static final int TERMINATING = 0x01; |
| 235 |
< |
private static final int TERMINATED = 0x02; |
| 236 |
< |
private static final int SUSPENDED = 0x04; // inactive spare |
| 237 |
< |
private static final int TRIMMED = 0x08; // killed while suspended |
| 205 |
> |
final int poolIndex; |
| 206 |
|
|
| 207 |
|
/** |
| 208 |
< |
* Number of LockSupport.park calls to block this thread for |
| 209 |
< |
* suspension or event waits. Used for internal instrumention; |
| 242 |
< |
* currently not exported but included because volatile write upon |
| 243 |
< |
* park also provides a workaround for a JVM bug. |
| 208 |
> |
* Encoded record for pool task waits. Usages are always |
| 209 |
> |
* surrounded by volatile reads/writes |
| 210 |
|
*/ |
| 211 |
< |
volatile int parkCount; |
| 211 |
> |
int nextWait; |
| 212 |
|
|
| 213 |
|
/** |
| 214 |
< |
* Number of steals, transferred and reset in pool callbacks pool |
| 215 |
< |
* when idle Accessed directly by pool. |
| 214 |
> |
* Complement of poolIndex, offset by count of entries of task |
| 215 |
> |
* waits. Accessed by ForkJoinPool to manage event waiters. |
| 216 |
|
*/ |
| 217 |
< |
int stealCount; |
| 217 |
> |
volatile int eventCount; |
| 218 |
|
|
| 219 |
|
/** |
| 220 |
|
* Seed for random number generator for choosing steal victims. |
| 221 |
|
* Uses Marsaglia xorshift. Must be initialized as nonzero. |
| 222 |
|
*/ |
| 223 |
< |
private int seed; |
| 258 |
< |
|
| 223 |
> |
int seed; |
| 224 |
|
|
| 225 |
|
/** |
| 226 |
< |
* Activity status. When true, this worker is considered active. |
| 227 |
< |
* Accessed directly by pool. Must be false upon construction. |
| 226 |
> |
* Number of steals. Directly accessed (and reset) by pool when |
| 227 |
> |
* idle. |
| 228 |
|
*/ |
| 229 |
< |
boolean active; |
| 229 |
> |
int stealCount; |
| 230 |
|
|
| 231 |
|
/** |
| 232 |
< |
* True if use local fifo, not default lifo, for local polling. |
| 268 |
< |
* Shadows value from ForkJoinPool, which resets it if changed |
| 269 |
< |
* pool-wide. |
| 232 |
> |
* True if this worker should or did terminate |
| 233 |
|
*/ |
| 234 |
< |
private final boolean locallyFifo; |
| 234 |
> |
volatile boolean terminate; |
| 235 |
|
|
| 236 |
|
/** |
| 237 |
< |
* Index of this worker in pool array. Set once by pool before |
| 275 |
< |
* running, and accessed directly by pool to locate this worker in |
| 276 |
< |
* its workers array. |
| 237 |
> |
* Set to true before LockSupport.park; false on return |
| 238 |
|
*/ |
| 239 |
< |
int poolIndex; |
| 239 |
> |
volatile boolean parked; |
| 240 |
|
|
| 241 |
|
/** |
| 242 |
< |
* The last pool event waited for. Accessed only by pool in |
| 243 |
< |
* callback methods invoked within this thread. |
| 242 |
> |
* True if use local fifo, not default lifo, for local polling. |
| 243 |
> |
* Shadows value from ForkJoinPool. |
| 244 |
|
*/ |
| 245 |
< |
int lastEventCount; |
| 245 |
> |
final boolean locallyFifo; |
| 246 |
|
|
| 247 |
|
/** |
| 248 |
< |
* Encoded index and event count of next event waiter. Used only |
| 249 |
< |
* by ForkJoinPool for managing event waiters. |
| 248 |
> |
* The task most recently stolen from another worker (or |
| 249 |
> |
* submission queue). All uses are surrounded by enough volatile |
| 250 |
> |
* reads/writes to maintain as non-volatile. |
| 251 |
|
*/ |
| 252 |
< |
volatile long nextWaiter; |
| 252 |
> |
ForkJoinTask<?> currentSteal; |
| 253 |
|
|
| 254 |
|
/** |
| 255 |
|
* The task currently being joined, set only when actively trying |
| 256 |
< |
* to helpStealer. Written only by current thread, but read by |
| 257 |
< |
* others. |
| 296 |
< |
*/ |
| 297 |
< |
private volatile ForkJoinTask<?> currentJoin; |
| 298 |
< |
|
| 299 |
< |
/** |
| 300 |
< |
* The task most recently stolen from another worker (or |
| 301 |
< |
* submission queue). Not volatile because always read/written in |
| 302 |
< |
* presence of related volatiles in those cases where it matters. |
| 256 |
> |
* to help other stealers in helpJoinTask. All uses are surrounded |
| 257 |
> |
* by enough volatile reads/writes to maintain as non-volatile. |
| 258 |
|
*/ |
| 259 |
< |
private ForkJoinTask<?> currentSteal; |
| 259 |
> |
ForkJoinTask<?> currentJoin; |
| 260 |
|
|
| 261 |
|
/** |
| 262 |
|
* Creates a ForkJoinWorkerThread operating in the given pool. |
| 265 |
|
* @throws NullPointerException if pool is null |
| 266 |
|
*/ |
| 267 |
|
protected ForkJoinWorkerThread(ForkJoinPool pool) { |
| 268 |
+ |
super(pool.nextWorkerName()); |
| 269 |
|
this.pool = pool; |
| 270 |
< |
this.locallyFifo = pool.locallyFifo; |
| 271 |
< |
// To avoid exposing construction details to subclasses, |
| 272 |
< |
// remaining initialization is in start() and onStart() |
| 273 |
< |
} |
| 274 |
< |
|
| 319 |
< |
/** |
| 320 |
< |
* Performs additional initialization and starts this thread |
| 321 |
< |
*/ |
| 322 |
< |
final void start(int poolIndex, UncaughtExceptionHandler ueh) { |
| 323 |
< |
this.poolIndex = poolIndex; |
| 270 |
> |
int k = pool.registerWorker(this); |
| 271 |
> |
poolIndex = k; |
| 272 |
> |
eventCount = ~k & SMASK; // clear wait count |
| 273 |
> |
locallyFifo = pool.locallyFifo; |
| 274 |
> |
Thread.UncaughtExceptionHandler ueh = pool.ueh; |
| 275 |
|
if (ueh != null) |
| 276 |
|
setUncaughtExceptionHandler(ueh); |
| 277 |
|
setDaemon(true); |
| 327 |
– |
start(); |
| 278 |
|
} |
| 279 |
|
|
| 280 |
< |
// Public/protected methods |
| 280 |
> |
// Public methods |
| 281 |
|
|
| 282 |
|
/** |
| 283 |
|
* Returns the pool hosting this thread. |
| 301 |
|
return poolIndex; |
| 302 |
|
} |
| 303 |
|
|
| 304 |
+ |
// Randomization |
| 305 |
+ |
|
| 306 |
+ |
/** |
| 307 |
+ |
* Computes next value for random victim probes and backoffs. |
| 308 |
+ |
* Scans don't require a very high quality generator, but also not |
| 309 |
+ |
* a crummy one. Marsaglia xor-shift is cheap and works well |
| 310 |
+ |
* enough. Note: This is manually inlined in FJP.scan() to avoid |
| 311 |
+ |
* writes inside busy loops. |
| 312 |
+ |
*/ |
| 313 |
+ |
private int nextSeed() { |
| 314 |
+ |
int r = seed; |
| 315 |
+ |
r ^= r << 13; |
| 316 |
+ |
r ^= r >>> 17; |
| 317 |
+ |
r ^= r << 5; |
| 318 |
+ |
return seed = r; |
| 319 |
+ |
} |
| 320 |
+ |
|
| 321 |
+ |
// Run State management |
| 322 |
+ |
|
| 323 |
|
/** |
| 324 |
|
* Initializes internal state after construction but before |
| 325 |
|
* processing any tasks. If you override this method, you must |
| 326 |
< |
* invoke super.onStart() at the beginning of the method. |
| 326 |
> |
* invoke {@code super.onStart()} at the beginning of the method. |
| 327 |
|
* Initialization requires care: Most fields must have legal |
| 328 |
|
* default values, to ensure that attempted accesses from other |
| 329 |
|
* threads work correctly even before this thread starts |
| 330 |
|
* processing tasks. |
| 331 |
|
*/ |
| 332 |
|
protected void onStart() { |
| 364 |
– |
int rs = seedGenerator.nextInt(); |
| 365 |
– |
seed = rs == 0? 1 : rs; // seed must be nonzero |
| 366 |
– |
|
| 367 |
– |
// Allocate name string and arrays in this thread |
| 368 |
– |
String pid = Integer.toString(pool.getPoolNumber()); |
| 369 |
– |
String wid = Integer.toString(poolIndex); |
| 370 |
– |
setName("ForkJoinPool-" + pid + "-worker-" + wid); |
| 371 |
– |
|
| 333 |
|
queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
| 334 |
+ |
int r = pool.workerSeedGenerator.nextInt(); |
| 335 |
+ |
seed = (r == 0)? 1 : r; // must be nonzero |
| 336 |
|
} |
| 337 |
|
|
| 338 |
|
/** |
| 345 |
|
*/ |
| 346 |
|
protected void onTermination(Throwable exception) { |
| 347 |
|
try { |
| 348 |
+ |
terminate = true; |
| 349 |
|
cancelTasks(); |
| 350 |
< |
setTerminated(); |
| 387 |
< |
pool.workerTerminated(this); |
| 350 |
> |
pool.deregisterWorker(this, exception); |
| 351 |
|
} catch (Throwable ex) { // Shouldn't ever happen |
| 352 |
|
if (exception == null) // but if so, at least rethrown |
| 353 |
|
exception = ex; |
| 360 |
|
/** |
| 361 |
|
* This method is required to be public, but should never be |
| 362 |
|
* called explicitly. It performs the main run loop to execute |
| 363 |
< |
* ForkJoinTasks. |
| 363 |
> |
* {@link ForkJoinTask}s. |
| 364 |
|
*/ |
| 365 |
|
public void run() { |
| 366 |
|
Throwable exception = null; |
| 367 |
|
try { |
| 368 |
|
onStart(); |
| 369 |
< |
mainLoop(); |
| 369 |
> |
pool.work(this); |
| 370 |
|
} catch (Throwable ex) { |
| 371 |
|
exception = ex; |
| 372 |
|
} finally { |
| 374 |
|
} |
| 375 |
|
} |
| 376 |
|
|
| 414 |
– |
// helpers for run() |
| 415 |
– |
|
| 416 |
– |
/** |
| 417 |
– |
* Find and execute tasks and check status while running |
| 418 |
– |
*/ |
| 419 |
– |
private void mainLoop() { |
| 420 |
– |
int emptyScans = 0; // consecutive times failed to find work |
| 421 |
– |
ForkJoinPool p = pool; |
| 422 |
– |
for (;;) { |
| 423 |
– |
p.preStep(this, emptyScans); |
| 424 |
– |
if (runState != 0) |
| 425 |
– |
return; |
| 426 |
– |
ForkJoinTask<?> t; // try to get and run stolen or submitted task |
| 427 |
– |
if ((t = scan()) != null || (t = pollSubmission()) != null) { |
| 428 |
– |
t.tryExec(); |
| 429 |
– |
if (base != sp) |
| 430 |
– |
runLocalTasks(); |
| 431 |
– |
currentSteal = null; |
| 432 |
– |
emptyScans = 0; |
| 433 |
– |
} |
| 434 |
– |
else |
| 435 |
– |
++emptyScans; |
| 436 |
– |
} |
| 437 |
– |
} |
| 438 |
– |
|
| 439 |
– |
/** |
| 440 |
– |
* Runs local tasks until queue is empty or shut down. Call only |
| 441 |
– |
* while active. |
| 442 |
– |
*/ |
| 443 |
– |
private void runLocalTasks() { |
| 444 |
– |
while (runState == 0) { |
| 445 |
– |
ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask(); |
| 446 |
– |
if (t != null) |
| 447 |
– |
t.tryExec(); |
| 448 |
– |
else if (base == sp) |
| 449 |
– |
break; |
| 450 |
– |
} |
| 451 |
– |
} |
| 452 |
– |
|
| 453 |
– |
/** |
| 454 |
– |
* If a submission exists, try to activate and take it |
| 455 |
– |
* |
| 456 |
– |
* @return a task, if available |
| 457 |
– |
*/ |
| 458 |
– |
private ForkJoinTask<?> pollSubmission() { |
| 459 |
– |
ForkJoinPool p = pool; |
| 460 |
– |
while (p.hasQueuedSubmissions()) { |
| 461 |
– |
if (active || (active = p.tryIncrementActiveCount())) { |
| 462 |
– |
ForkJoinTask<?> t = p.pollSubmission(); |
| 463 |
– |
if (t != null) { |
| 464 |
– |
currentSteal = t; |
| 465 |
– |
return t; |
| 466 |
– |
} |
| 467 |
– |
return scan(); // if missed, rescan |
| 468 |
– |
} |
| 469 |
– |
} |
| 470 |
– |
return null; |
| 471 |
– |
} |
| 472 |
– |
|
| 377 |
|
/* |
| 378 |
|
* Intrinsics-based atomic writes for queue slots. These are |
| 379 |
< |
* basically the same as methods in AtomicObjectArray, but |
| 379 |
> |
* basically the same as methods in AtomicReferenceArray, but |
| 380 |
|
* specialized for (1) ForkJoinTask elements (2) requirement that |
| 381 |
|
* nullness and bounds checks have already been performed by |
| 382 |
|
* callers and (3) effective offsets are known not to overflow |
| 383 |
|
* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't |
| 384 |
|
* need corresponding version for reads: plain array reads are OK |
| 385 |
< |
* because they protected by other volatile reads and are |
| 385 |
> |
* because they are protected by other volatile reads and are |
| 386 |
|
* confirmed by CASes. |
| 387 |
|
* |
| 388 |
< |
* Most uses don't actually call these methods, but instead contain |
| 389 |
< |
* inlined forms that enable more predictable optimization. We |
| 390 |
< |
* don't define the version of write used in pushTask at all, but |
| 391 |
< |
* instead inline there a store-fenced array slot write. |
| 388 |
> |
* Most uses don't actually call these methods, but instead |
| 389 |
> |
* contain inlined forms that enable more predictable |
| 390 |
> |
* optimization. We don't define the version of write used in |
| 391 |
> |
* pushTask at all, but instead inline there a store-fenced array |
| 392 |
> |
* slot write. |
| 393 |
> |
* |
| 394 |
> |
* Also in most methods, as a performance (not correctness) issue, |
| 395 |
> |
* we'd like to encourage compilers not to arbitrarily postpone |
| 396 |
> |
* setting queueTop after writing slot. Currently there is no |
| 397 |
> |
* intrinsic for arranging this, but using Unsafe putOrderedInt |
| 398 |
> |
* may be a preferable strategy on some compilers even though its |
| 399 |
> |
* main effect is a pre-, not post- fence. To simplify possible |
| 400 |
> |
* changes, the option is left in comments next to the associated |
| 401 |
> |
* assignments. |
| 402 |
|
*/ |
| 403 |
|
|
| 404 |
|
/** |
| 407 |
|
*/ |
| 408 |
|
private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i, |
| 409 |
|
ForkJoinTask<?> t) { |
| 410 |
< |
return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null); |
| 410 |
> |
return UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, t, null); |
| 411 |
|
} |
| 412 |
|
|
| 413 |
|
/** |
| 416 |
|
* range. This method is used only during resets and backouts. |
| 417 |
|
*/ |
| 418 |
|
private static final void writeSlot(ForkJoinTask<?>[] q, int i, |
| 419 |
< |
ForkJoinTask<?> t) { |
| 420 |
< |
UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t); |
| 419 |
> |
ForkJoinTask<?> t) { |
| 420 |
> |
UNSAFE.putObjectVolatile(q, (i << ASHIFT) + ABASE, t); |
| 421 |
|
} |
| 422 |
|
|
| 423 |
|
// queue methods |
| 428 |
|
* @param t the task. Caller must ensure non-null. |
| 429 |
|
*/ |
| 430 |
|
final void pushTask(ForkJoinTask<?> t) { |
| 431 |
< |
ForkJoinTask<?>[] q = queue; |
| 432 |
< |
int mask = q.length - 1; // implicit assert q != null |
| 433 |
< |
int s = sp++; // ok to increment sp before slot write |
| 434 |
< |
UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t); |
| 435 |
< |
if ((s -= base) == 0) |
| 436 |
< |
pool.signalWork(); // was empty |
| 437 |
< |
else if (s == mask) |
| 438 |
< |
growQueue(); // is full |
| 431 |
> |
ForkJoinTask<?>[] q; int s, m; |
| 432 |
> |
if ((q = queue) != null) { // ignore if queue removed |
| 433 |
> |
long u = (((s = queueTop) & (m = q.length - 1)) << ASHIFT) + ABASE; |
| 434 |
> |
UNSAFE.putOrderedObject(q, u, t); |
| 435 |
> |
queueTop = s + 1; // or use putOrderedInt |
| 436 |
> |
if ((s -= queueBase) <= 2) |
| 437 |
> |
pool.signalWork(); |
| 438 |
> |
else if (s == m) |
| 439 |
> |
growQueue(); |
| 440 |
> |
} |
| 441 |
> |
} |
| 442 |
> |
|
| 443 |
> |
/** |
| 444 |
> |
* Creates or doubles queue array. Transfers elements by |
| 445 |
> |
* emulating steals (deqs) from old array and placing, oldest |
| 446 |
> |
* first, into new array. |
| 447 |
> |
*/ |
| 448 |
> |
private void growQueue() { |
| 449 |
> |
ForkJoinTask<?>[] oldQ = queue; |
| 450 |
> |
int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY; |
| 451 |
> |
if (size > MAXIMUM_QUEUE_CAPACITY) |
| 452 |
> |
throw new RejectedExecutionException("Queue capacity exceeded"); |
| 453 |
> |
if (size < INITIAL_QUEUE_CAPACITY) |
| 454 |
> |
size = INITIAL_QUEUE_CAPACITY; |
| 455 |
> |
ForkJoinTask<?>[] q = queue = new ForkJoinTask<?>[size]; |
| 456 |
> |
int mask = size - 1; |
| 457 |
> |
int top = queueTop; |
| 458 |
> |
int oldMask; |
| 459 |
> |
if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) { |
| 460 |
> |
for (int b = queueBase; b != top; ++b) { |
| 461 |
> |
long u = ((b & oldMask) << ASHIFT) + ABASE; |
| 462 |
> |
Object x = UNSAFE.getObjectVolatile(oldQ, u); |
| 463 |
> |
if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null)) |
| 464 |
> |
UNSAFE.putObjectVolatile |
| 465 |
> |
(q, ((b & mask) << ASHIFT) + ABASE, x); |
| 466 |
> |
} |
| 467 |
> |
} |
| 468 |
|
} |
| 469 |
|
|
| 470 |
|
/** |
| 475 |
|
* @return a task, or null if none or contended |
| 476 |
|
*/ |
| 477 |
|
final ForkJoinTask<?> deqTask() { |
| 478 |
< |
ForkJoinTask<?> t; |
| 479 |
< |
ForkJoinTask<?>[] q; |
| 537 |
< |
int b, i; |
| 538 |
< |
if ((b = base) != sp && |
| 478 |
> |
ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i; |
| 479 |
> |
if (queueTop != (b = queueBase) && |
| 480 |
|
(q = queue) != null && // must read q after b |
| 481 |
< |
(t = q[i = (q.length - 1) & b]) != null && base == b && |
| 482 |
< |
UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) { |
| 483 |
< |
base = b + 1; |
| 481 |
> |
(i = (q.length - 1) & b) >= 0 && |
| 482 |
> |
(t = q[i]) != null && queueBase == b && |
| 483 |
> |
UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, t, null)) { |
| 484 |
> |
queueBase = b + 1; |
| 485 |
|
return t; |
| 486 |
|
} |
| 487 |
|
return null; |
| 488 |
|
} |
| 489 |
|
|
| 490 |
|
/** |
| 491 |
< |
* Tries to take a task from the base of own queue. Assumes active |
| 492 |
< |
* status. Called only by current thread. |
| 491 |
> |
* Tries to take a task from the base of own queue. Called only |
| 492 |
> |
* by this thread. |
| 493 |
|
* |
| 494 |
|
* @return a task, or null if none |
| 495 |
|
*/ |
| 496 |
|
final ForkJoinTask<?> locallyDeqTask() { |
| 497 |
+ |
ForkJoinTask<?> t; int m, b, i; |
| 498 |
|
ForkJoinTask<?>[] q = queue; |
| 499 |
< |
if (q != null) { |
| 500 |
< |
ForkJoinTask<?> t; |
| 501 |
< |
int b, i; |
| 502 |
< |
while (sp != (b = base)) { |
| 503 |
< |
if ((t = q[i = (q.length - 1) & b]) != null && base == b && |
| 561 |
< |
UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, |
| 499 |
> |
if (q != null && (m = q.length - 1) >= 0) { |
| 500 |
> |
while (queueTop != (b = queueBase)) { |
| 501 |
> |
if ((t = q[i = m & b]) != null && |
| 502 |
> |
queueBase == b && |
| 503 |
> |
UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, |
| 504 |
|
t, null)) { |
| 505 |
< |
base = b + 1; |
| 505 |
> |
queueBase = b + 1; |
| 506 |
|
return t; |
| 507 |
|
} |
| 508 |
|
} |
| 511 |
|
} |
| 512 |
|
|
| 513 |
|
/** |
| 514 |
< |
* Returns a popped task, or null if empty. Assumes active status. |
| 515 |
< |
* Called only by current thread. |
| 514 |
> |
* Returns a popped task, or null if empty. |
| 515 |
> |
* Called only by this thread. |
| 516 |
|
*/ |
| 517 |
< |
final ForkJoinTask<?> popTask() { |
| 518 |
< |
int s; |
| 519 |
< |
ForkJoinTask<?>[] q; |
| 520 |
< |
if (base != (s = sp) && (q = queue) != null) { |
| 521 |
< |
int i = (q.length - 1) & --s; |
| 522 |
< |
ForkJoinTask<?> t = q[i]; |
| 523 |
< |
if (t != null && UNSAFE.compareAndSwapObject |
| 524 |
< |
(q, (i << qShift) + qBase, t, null)) { |
| 525 |
< |
sp = s; |
| 526 |
< |
return t; |
| 517 |
> |
private ForkJoinTask<?> popTask() { |
| 518 |
> |
int m; |
| 519 |
> |
ForkJoinTask<?>[] q = queue; |
| 520 |
> |
if (q != null && (m = q.length - 1) >= 0) { |
| 521 |
> |
for (int s; (s = queueTop) != queueBase;) { |
| 522 |
> |
int i = m & --s; |
| 523 |
> |
long u = (i << ASHIFT) + ABASE; // raw offset |
| 524 |
> |
ForkJoinTask<?> t = q[i]; |
| 525 |
> |
if (t == null) // lost to stealer |
| 526 |
> |
break; |
| 527 |
> |
if (UNSAFE.compareAndSwapObject(q, u, t, null)) { |
| 528 |
> |
queueTop = s; // or putOrderedInt |
| 529 |
> |
return t; |
| 530 |
> |
} |
| 531 |
|
} |
| 532 |
|
} |
| 533 |
|
return null; |
| 535 |
|
|
| 536 |
|
/** |
| 537 |
|
* Specialized version of popTask to pop only if topmost element |
| 538 |
< |
* is the given task. Called only by current thread while |
| 593 |
< |
* active. |
| 538 |
> |
* is the given task. Called only by this thread. |
| 539 |
|
* |
| 540 |
|
* @param t the task. Caller must ensure non-null. |
| 541 |
|
*/ |
| 542 |
|
final boolean unpushTask(ForkJoinTask<?> t) { |
| 598 |
– |
int s; |
| 543 |
|
ForkJoinTask<?>[] q; |
| 544 |
< |
if (base != (s = sp) && (q = queue) != null && |
| 544 |
> |
int s; |
| 545 |
> |
if ((q = queue) != null && (s = queueTop) != queueBase && |
| 546 |
|
UNSAFE.compareAndSwapObject |
| 547 |
< |
(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) { |
| 548 |
< |
sp = s; |
| 547 |
> |
(q, (((q.length - 1) & --s) << ASHIFT) + ABASE, t, null)) { |
| 548 |
> |
queueTop = s; // or putOrderedInt |
| 549 |
|
return true; |
| 550 |
|
} |
| 551 |
|
return false; |
| 552 |
|
} |
| 553 |
|
|
| 554 |
|
/** |
| 555 |
< |
* Returns next task or null if empty or contended |
| 555 |
> |
* Returns next task, or null if empty or contended. |
| 556 |
|
*/ |
| 557 |
|
final ForkJoinTask<?> peekTask() { |
| 558 |
+ |
int m; |
| 559 |
|
ForkJoinTask<?>[] q = queue; |
| 560 |
< |
if (q == null) |
| 560 |
> |
if (q == null || (m = q.length - 1) < 0) |
| 561 |
|
return null; |
| 562 |
< |
int mask = q.length - 1; |
| 563 |
< |
int i = locallyFifo ? base : (sp - 1); |
| 618 |
< |
return q[i & mask]; |
| 619 |
< |
} |
| 620 |
< |
|
| 621 |
< |
/** |
| 622 |
< |
* Doubles queue array size. Transfers elements by emulating |
| 623 |
< |
* steals (deqs) from old array and placing, oldest first, into |
| 624 |
< |
* new array. |
| 625 |
< |
*/ |
| 626 |
< |
private void growQueue() { |
| 627 |
< |
ForkJoinTask<?>[] oldQ = queue; |
| 628 |
< |
int oldSize = oldQ.length; |
| 629 |
< |
int newSize = oldSize << 1; |
| 630 |
< |
if (newSize > MAXIMUM_QUEUE_CAPACITY) |
| 631 |
< |
throw new RejectedExecutionException("Queue capacity exceeded"); |
| 632 |
< |
ForkJoinTask<?>[] newQ = queue = new ForkJoinTask<?>[newSize]; |
| 633 |
< |
|
| 634 |
< |
int b = base; |
| 635 |
< |
int bf = b + oldSize; |
| 636 |
< |
int oldMask = oldSize - 1; |
| 637 |
< |
int newMask = newSize - 1; |
| 638 |
< |
do { |
| 639 |
< |
int oldIndex = b & oldMask; |
| 640 |
< |
ForkJoinTask<?> t = oldQ[oldIndex]; |
| 641 |
< |
if (t != null && !casSlotNull(oldQ, oldIndex, t)) |
| 642 |
< |
t = null; |
| 643 |
< |
writeSlot(newQ, b & newMask, t); |
| 644 |
< |
} while (++b != bf); |
| 645 |
< |
pool.signalWork(); |
| 646 |
< |
} |
| 647 |
< |
|
| 648 |
< |
/** |
| 649 |
< |
* Computes next value for random victim probe in scan(). Scans |
| 650 |
< |
* don't require a very high quality generator, but also not a |
| 651 |
< |
* crummy one. Marsaglia xor-shift is cheap and works well enough. |
| 652 |
< |
* Note: This is manually inlined in scan() |
| 653 |
< |
*/ |
| 654 |
< |
private static final int xorShift(int r) { |
| 655 |
< |
r ^= r << 13; |
| 656 |
< |
r ^= r >>> 17; |
| 657 |
< |
return r ^ (r << 5); |
| 658 |
< |
} |
| 659 |
< |
|
| 660 |
< |
/** |
| 661 |
< |
* Tries to steal a task from another worker. Starts at a random |
| 662 |
< |
* index of workers array, and probes workers until finding one |
| 663 |
< |
* with non-empty queue or finding that all are empty. It |
| 664 |
< |
* randomly selects the first n probes. If these are empty, it |
| 665 |
< |
* resorts to a circular sweep, which is necessary to accurately |
| 666 |
< |
* set active status. (The circular sweep uses steps of |
| 667 |
< |
* approximately half the array size plus 1, to avoid bias |
| 668 |
< |
* stemming from leftmost packing of the array in ForkJoinPool.) |
| 669 |
< |
* |
| 670 |
< |
* This method must be both fast and quiet -- usually avoiding |
| 671 |
< |
* memory accesses that could disrupt cache sharing etc other than |
| 672 |
< |
* those needed to check for and take tasks (or to activate if not |
| 673 |
< |
* already active). This accounts for, among other things, |
| 674 |
< |
* updating random seed in place without storing it until exit. |
| 675 |
< |
* |
| 676 |
< |
* @return a task, or null if none found |
| 677 |
< |
*/ |
| 678 |
< |
private ForkJoinTask<?> scan() { |
| 679 |
< |
ForkJoinPool p = pool; |
| 680 |
< |
ForkJoinWorkerThread[] ws; // worker array |
| 681 |
< |
int n; // upper bound of #workers |
| 682 |
< |
if ((ws = p.workers) != null && (n = ws.length) > 1) { |
| 683 |
< |
boolean canSteal = active; // shadow active status |
| 684 |
< |
int r = seed; // extract seed once |
| 685 |
< |
int mask = n - 1; |
| 686 |
< |
int j = -n; // loop counter |
| 687 |
< |
int k = r; // worker index, random if j < 0 |
| 688 |
< |
for (;;) { |
| 689 |
< |
ForkJoinWorkerThread v = ws[k & mask]; |
| 690 |
< |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift |
| 691 |
< |
if (v != null && v.base != v.sp) { |
| 692 |
< |
if (canSteal || // ensure active status |
| 693 |
< |
(canSteal = active = p.tryIncrementActiveCount())) { |
| 694 |
< |
int b = v.base; // inline specialized deqTask |
| 695 |
< |
ForkJoinTask<?>[] q; |
| 696 |
< |
if (b != v.sp && (q = v.queue) != null) { |
| 697 |
< |
ForkJoinTask<?> t; |
| 698 |
< |
int i = (q.length - 1) & b; |
| 699 |
< |
long u = (i << qShift) + qBase; // raw offset |
| 700 |
< |
if ((t = q[i]) != null && v.base == b && |
| 701 |
< |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
| 702 |
< |
currentSteal = t; |
| 703 |
< |
v.stealHint = poolIndex; |
| 704 |
< |
v.base = b + 1; |
| 705 |
< |
seed = r; |
| 706 |
< |
++stealCount; |
| 707 |
< |
return t; |
| 708 |
< |
} |
| 709 |
< |
} |
| 710 |
< |
} |
| 711 |
< |
j = -n; |
| 712 |
< |
k = r; // restart on contention |
| 713 |
< |
} |
| 714 |
< |
else if (++j <= 0) |
| 715 |
< |
k = r; |
| 716 |
< |
else if (j <= n) |
| 717 |
< |
k += (n >>> 1) | 1; |
| 718 |
< |
else |
| 719 |
< |
break; |
| 720 |
< |
} |
| 721 |
< |
} |
| 722 |
< |
return null; |
| 562 |
> |
int i = locallyFifo ? queueBase : (queueTop - 1); |
| 563 |
> |
return q[i & m]; |
| 564 |
|
} |
| 565 |
|
|
| 566 |
< |
// Run State management |
| 726 |
< |
|
| 727 |
< |
// status check methods used mainly by ForkJoinPool |
| 728 |
< |
final boolean isTerminating() { return (runState & TERMINATING) != 0; } |
| 729 |
< |
final boolean isTerminated() { return (runState & TERMINATED) != 0; } |
| 730 |
< |
final boolean isSuspended() { return (runState & SUSPENDED) != 0; } |
| 731 |
< |
final boolean isTrimmed() { return (runState & TRIMMED) != 0; } |
| 566 |
> |
// Support methods for ForkJoinPool |
| 567 |
|
|
| 568 |
|
/** |
| 569 |
< |
* Sets state to TERMINATING, also resuming if suspended. |
| 569 |
> |
* Runs the given task, plus any local tasks until queue is empty |
| 570 |
|
*/ |
| 571 |
< |
final void shutdown() { |
| 571 |
> |
final void execTask(ForkJoinTask<?> t) { |
| 572 |
> |
currentSteal = t; |
| 573 |
|
for (;;) { |
| 574 |
< |
int s = runState; |
| 575 |
< |
if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended |
| 576 |
< |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
| 741 |
< |
(s & ~SUSPENDED) | |
| 742 |
< |
(TRIMMED|TERMINATING))) { |
| 743 |
< |
LockSupport.unpark(this); |
| 744 |
< |
break; |
| 745 |
< |
} |
| 746 |
< |
} |
| 747 |
< |
else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
| 748 |
< |
s | TERMINATING)) |
| 749 |
< |
break; |
| 750 |
< |
} |
| 751 |
< |
} |
| 752 |
< |
|
| 753 |
< |
/** |
| 754 |
< |
* Sets state to TERMINATED. Called only by this thread. |
| 755 |
< |
*/ |
| 756 |
< |
private void setTerminated() { |
| 757 |
< |
int s; |
| 758 |
< |
do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset, |
| 759 |
< |
s = runState, |
| 760 |
< |
s | (TERMINATING|TERMINATED))); |
| 761 |
< |
} |
| 762 |
< |
|
| 763 |
< |
/** |
| 764 |
< |
* Instrumented version of park used by ForkJoinPool.awaitEvent |
| 765 |
< |
*/ |
| 766 |
< |
final void doPark() { |
| 767 |
< |
++parkCount; |
| 768 |
< |
LockSupport.park(this); |
| 769 |
< |
} |
| 770 |
< |
|
| 771 |
< |
/** |
| 772 |
< |
* If suspended, tries to set status to unsuspended and unparks. |
| 773 |
< |
* |
| 774 |
< |
* @return true if successful |
| 775 |
< |
*/ |
| 776 |
< |
final boolean tryResumeSpare() { |
| 777 |
< |
int s = runState; |
| 778 |
< |
if ((s & SUSPENDED) != 0 && |
| 779 |
< |
UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
| 780 |
< |
s & ~SUSPENDED)) { |
| 781 |
< |
LockSupport.unpark(this); |
| 782 |
< |
return true; |
| 783 |
< |
} |
| 784 |
< |
return false; |
| 785 |
< |
} |
| 786 |
< |
|
| 787 |
< |
/** |
| 788 |
< |
* Sets suspended status and blocks as spare until resumed, |
| 789 |
< |
* shutdown, or timed out. |
| 790 |
< |
* |
| 791 |
< |
* @return false if trimmed |
| 792 |
< |
*/ |
| 793 |
< |
final boolean suspendAsSpare() { |
| 794 |
< |
for (;;) { // set suspended unless terminating |
| 795 |
< |
int s = runState; |
| 796 |
< |
if ((s & TERMINATING) != 0) { // must kill |
| 797 |
< |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
| 798 |
< |
s | (TRIMMED | TERMINATING))) |
| 799 |
< |
return false; |
| 800 |
< |
} |
| 801 |
< |
else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
| 802 |
< |
s | SUSPENDED)) |
| 574 |
> |
if (t != null) |
| 575 |
> |
t.doExec(); |
| 576 |
> |
if (queueTop == queueBase) |
| 577 |
|
break; |
| 578 |
+ |
t = locallyFifo ? locallyDeqTask() : popTask(); |
| 579 |
|
} |
| 580 |
< |
boolean timed; |
| 581 |
< |
long nanos; |
| 807 |
< |
long startTime; |
| 808 |
< |
if (poolIndex < pool.parallelism) { |
| 809 |
< |
timed = false; |
| 810 |
< |
nanos = 0L; |
| 811 |
< |
startTime = 0L; |
| 812 |
< |
} |
| 813 |
< |
else { |
| 814 |
< |
timed = true; |
| 815 |
< |
nanos = SPARE_KEEPALIVE_NANOS; |
| 816 |
< |
startTime = System.nanoTime(); |
| 817 |
< |
} |
| 818 |
< |
pool.accumulateStealCount(this); |
| 819 |
< |
lastEventCount = 0; // reset upon resume |
| 820 |
< |
interrupted(); // clear/ignore interrupts |
| 821 |
< |
while ((runState & SUSPENDED) != 0) { |
| 822 |
< |
++parkCount; |
| 823 |
< |
if (!timed) |
| 824 |
< |
LockSupport.park(this); |
| 825 |
< |
else if ((nanos -= (System.nanoTime() - startTime)) > 0) |
| 826 |
< |
LockSupport.parkNanos(this, nanos); |
| 827 |
< |
else { // try to trim on timeout |
| 828 |
< |
int s = runState; |
| 829 |
< |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
| 830 |
< |
(s & ~SUSPENDED) | |
| 831 |
< |
(TRIMMED|TERMINATING))) |
| 832 |
< |
return false; |
| 833 |
< |
} |
| 834 |
< |
} |
| 835 |
< |
return true; |
| 836 |
< |
} |
| 837 |
< |
|
| 838 |
< |
// Misc support methods for ForkJoinPool |
| 839 |
< |
|
| 840 |
< |
/** |
| 841 |
< |
* Returns an estimate of the number of tasks in the queue. Also |
| 842 |
< |
* used by ForkJoinTask. |
| 843 |
< |
*/ |
| 844 |
< |
final int getQueueSize() { |
| 845 |
< |
return -base + sp; |
| 580 |
> |
++stealCount; |
| 581 |
> |
currentSteal = null; |
| 582 |
|
} |
| 583 |
|
|
| 584 |
|
/** |
| 586 |
|
* thread. |
| 587 |
|
*/ |
| 588 |
|
final void cancelTasks() { |
| 589 |
< |
ForkJoinTask<?> cj = currentJoin; // try to kill live tasks |
| 590 |
< |
if (cj != null) { |
| 855 |
< |
currentJoin = null; |
| 589 |
> |
ForkJoinTask<?> cj = currentJoin; // try to cancel ongoing tasks |
| 590 |
> |
if (cj != null && cj.status >= 0) |
| 591 |
|
cj.cancelIgnoringExceptions(); |
| 857 |
– |
} |
| 592 |
|
ForkJoinTask<?> cs = currentSteal; |
| 593 |
< |
if (cs != null) { |
| 860 |
< |
currentSteal = null; |
| 593 |
> |
if (cs != null && cs.status >= 0) |
| 594 |
|
cs.cancelIgnoringExceptions(); |
| 595 |
< |
} |
| 863 |
< |
while (base != sp) { |
| 595 |
> |
while (queueBase != queueTop) { |
| 596 |
|
ForkJoinTask<?> t = deqTask(); |
| 597 |
|
if (t != null) |
| 598 |
|
t.cancelIgnoringExceptions(); |
| 606 |
|
*/ |
| 607 |
|
final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
| 608 |
|
int n = 0; |
| 609 |
< |
while (base != sp) { |
| 609 |
> |
while (queueBase != queueTop) { |
| 610 |
|
ForkJoinTask<?> t = deqTask(); |
| 611 |
|
if (t != null) { |
| 612 |
|
c.add(t); |
| 619 |
|
// Support methods for ForkJoinTask |
| 620 |
|
|
| 621 |
|
/** |
| 622 |
+ |
* Returns an estimate of the number of tasks in the queue. |
| 623 |
+ |
*/ |
| 624 |
+ |
final int getQueueSize() { |
| 625 |
+ |
return queueTop - queueBase; |
| 626 |
+ |
} |
| 627 |
+ |
|
| 628 |
+ |
/** |
| 629 |
|
* Gets and removes a local task. |
| 630 |
|
* |
| 631 |
|
* @return a task, if available |
| 632 |
|
*/ |
| 633 |
|
final ForkJoinTask<?> pollLocalTask() { |
| 634 |
< |
while (sp != base) { |
| 896 |
< |
if (active || (active = pool.tryIncrementActiveCount())) |
| 897 |
< |
return locallyFifo? locallyDeqTask() : popTask(); |
| 898 |
< |
} |
| 899 |
< |
return null; |
| 634 |
> |
return locallyFifo ? locallyDeqTask() : popTask(); |
| 635 |
|
} |
| 636 |
|
|
| 637 |
|
/** |
| 640 |
|
* @return a task, if available |
| 641 |
|
*/ |
| 642 |
|
final ForkJoinTask<?> pollTask() { |
| 643 |
< |
ForkJoinTask<?> t; |
| 644 |
< |
return (t = pollLocalTask()) != null ? t : scan(); |
| 643 |
> |
ForkJoinWorkerThread[] ws; |
| 644 |
> |
ForkJoinTask<?> t = pollLocalTask(); |
| 645 |
> |
if (t != null || (ws = pool.workers) == null) |
| 646 |
> |
return t; |
| 647 |
> |
int n = ws.length; // cheap version of FJP.scan |
| 648 |
> |
int steps = n << 1; |
| 649 |
> |
int r = nextSeed(); |
| 650 |
> |
int i = 0; |
| 651 |
> |
while (i < steps) { |
| 652 |
> |
ForkJoinWorkerThread w = ws[(i++ + r) & (n - 1)]; |
| 653 |
> |
if (w != null && w.queueBase != w.queueTop && w.queue != null) { |
| 654 |
> |
if ((t = w.deqTask()) != null) |
| 655 |
> |
return t; |
| 656 |
> |
i = 0; |
| 657 |
> |
} |
| 658 |
> |
} |
| 659 |
> |
return null; |
| 660 |
|
} |
| 661 |
|
|
| 662 |
|
/** |
| 663 |
< |
* Possibly runs some tasks and/or blocks, until task is done. |
| 664 |
< |
* The main body is basically a big spinloop, alternating between |
| 665 |
< |
* calls to helpJoinTask and pool.tryAwaitJoin with increased |
| 666 |
< |
* patience parameters until either the task is done without |
| 667 |
< |
* waiting, or we have, if necessary, created or resumed a |
| 668 |
< |
* replacement for this thread while it blocks. |
| 663 |
> |
* The maximum stolen->joining link depth allowed in helpJoinTask, |
| 664 |
> |
* as well as the maximum number of retries (allowing on average |
| 665 |
> |
* one staleness retry per level) per attempt to instead try |
| 666 |
> |
* compensation. Depths for legitimate chains are unbounded, but |
| 667 |
> |
* we use a fixed constant to avoid (otherwise unchecked) cycles |
| 668 |
> |
* and bound staleness of traversal parameters at the expense of |
| 669 |
> |
* sometimes blocking when we could be helping. |
| 670 |
> |
*/ |
| 671 |
> |
private static final int MAX_HELP = 16; |
| 672 |
> |
|
| 673 |
> |
/** |
| 674 |
> |
* Possibly runs some tasks and/or blocks, until joinMe is done. |
| 675 |
|
* |
| 676 |
|
* @param joinMe the task to join |
| 677 |
< |
* @return task status on exit |
| 677 |
> |
* @return completion status on exit |
| 678 |
|
*/ |
| 679 |
|
final int joinTask(ForkJoinTask<?> joinMe) { |
| 924 |
– |
int stat; |
| 680 |
|
ForkJoinTask<?> prevJoin = currentJoin; |
| 681 |
|
currentJoin = joinMe; |
| 682 |
< |
if ((stat = joinMe.status) >= 0 && |
| 683 |
< |
(sp == base || (stat = localHelpJoinTask(joinMe)) >= 0)) { |
| 684 |
< |
ForkJoinPool p = pool; |
| 685 |
< |
int helpRetries = 2; // initial patience values |
| 686 |
< |
int awaitRetries = -1; // -1 is sentinel for replace-check only |
| 687 |
< |
do { |
| 688 |
< |
helpJoinTask(joinMe, helpRetries); |
| 689 |
< |
if ((stat = joinMe.status) < 0) |
| 690 |
< |
break; |
| 691 |
< |
boolean busy = p.tryAwaitJoin(joinMe, awaitRetries); |
| 692 |
< |
if ((stat = joinMe.status) < 0) |
| 693 |
< |
break; |
| 694 |
< |
if (awaitRetries == -1) |
| 695 |
< |
awaitRetries = 0; |
| 696 |
< |
else if (busy) |
| 697 |
< |
++awaitRetries; |
| 698 |
< |
if (helpRetries < p.parallelism) |
| 699 |
< |
helpRetries <<= 1; |
| 700 |
< |
Thread.yield(); // tame unbounded loop |
| 701 |
< |
} while (joinMe.status >= 0); |
| 682 |
> |
for (int s, retries = MAX_HELP;;) { |
| 683 |
> |
if ((s = joinMe.status) < 0) { |
| 684 |
> |
currentJoin = prevJoin; |
| 685 |
> |
return s; |
| 686 |
> |
} |
| 687 |
> |
if (retries > 0) { |
| 688 |
> |
if (queueTop != queueBase) { |
| 689 |
> |
if (!localHelpJoinTask(joinMe)) |
| 690 |
> |
retries = 0; // cannot help |
| 691 |
> |
} |
| 692 |
> |
else if (retries == MAX_HELP >>> 1) { |
| 693 |
> |
--retries; // check uncommon case |
| 694 |
> |
if (tryDeqAndExec(joinMe) >= 0) |
| 695 |
> |
Thread.yield(); // for politeness |
| 696 |
> |
} |
| 697 |
> |
else |
| 698 |
> |
retries = helpJoinTask(joinMe)? MAX_HELP : retries - 1; |
| 699 |
> |
} |
| 700 |
> |
else { |
| 701 |
> |
retries = MAX_HELP; // restart if not done |
| 702 |
> |
pool.tryAwaitJoin(joinMe); |
| 703 |
> |
} |
| 704 |
|
} |
| 948 |
– |
currentJoin = prevJoin; |
| 949 |
– |
return stat; |
| 705 |
|
} |
| 706 |
|
|
| 707 |
|
/** |
| 708 |
< |
* Run tasks in local queue until given task is done. |
| 708 |
> |
* If present, pops and executes the given task, or any other |
| 709 |
> |
* cancelled task |
| 710 |
|
* |
| 711 |
< |
* @param joinMe the task to join |
| 956 |
< |
* @return task status on exit |
| 711 |
> |
* @return false if any other non-cancelled task exists in local queue |
| 712 |
|
*/ |
| 713 |
< |
private int localHelpJoinTask(ForkJoinTask<?> joinMe) { |
| 714 |
< |
int stat, s; |
| 715 |
< |
ForkJoinTask<?>[] q; |
| 716 |
< |
while ((stat = joinMe.status) >= 0 && |
| 717 |
< |
base != (s = sp) && (q = queue) != null) { |
| 718 |
< |
ForkJoinTask<?> t; |
| 719 |
< |
int i = (q.length - 1) & --s; |
| 720 |
< |
long u = (i << qShift) + qBase; // raw offset |
| 721 |
< |
if ((t = q[i]) != null && |
| 722 |
< |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
| 723 |
< |
/* |
| 969 |
< |
* This recheck (and similarly in helpJoinTask) |
| 970 |
< |
* handles cases where joinMe is independently |
| 971 |
< |
* cancelled or forced even though there is other work |
| 972 |
< |
* available. Back out of the pop by putting t back |
| 973 |
< |
* into slot before we commit by writing sp. |
| 974 |
< |
*/ |
| 975 |
< |
if ((stat = joinMe.status) < 0) { |
| 976 |
< |
UNSAFE.putObjectVolatile(q, u, t); |
| 977 |
< |
break; |
| 978 |
< |
} |
| 979 |
< |
sp = s; |
| 980 |
< |
t.tryExec(); |
| 713 |
> |
private boolean localHelpJoinTask(ForkJoinTask<?> joinMe) { |
| 714 |
> |
int s, i; ForkJoinTask<?>[] q; ForkJoinTask<?> t; |
| 715 |
> |
if ((s = queueTop) != queueBase && (q = queue) != null && |
| 716 |
> |
(i = (q.length - 1) & --s) >= 0 && |
| 717 |
> |
(t = q[i]) != null) { |
| 718 |
> |
if (t != joinMe && t.status >= 0) |
| 719 |
> |
return false; |
| 720 |
> |
if (UNSAFE.compareAndSwapObject |
| 721 |
> |
(q, (i << ASHIFT) + ABASE, t, null)) { |
| 722 |
> |
queueTop = s; // or putOrderedInt |
| 723 |
> |
t.doExec(); |
| 724 |
|
} |
| 725 |
|
} |
| 726 |
< |
return stat; |
| 726 |
> |
return true; |
| 727 |
|
} |
| 728 |
|
|
| 729 |
|
/** |
| 730 |
< |
* Tries to locate and help perform tasks for a stealer of the |
| 731 |
< |
* given task, or in turn one of its stealers. Traces |
| 730 |
> |
* Tries to locate and execute tasks for a stealer of the given |
| 731 |
> |
* task, or in turn one of its stealers, Traces |
| 732 |
|
* currentSteal->currentJoin links looking for a thread working on |
| 733 |
|
* a descendant of the given task and with a non-empty queue to |
| 734 |
< |
* steal back and execute tasks from. Restarts search upon |
| 734 |
> |
* steal back and execute tasks from. The implementation is very |
| 735 |
> |
* branchy to cope with potential inconsistencies or loops |
| 736 |
|
* encountering chains that are stale, unknown, or of length |
| 737 |
< |
* greater than MAX_HELP_DEPTH links, to avoid unbounded cycles. |
| 738 |
< |
* |
| 995 |
< |
* The implementation is very branchy to cope with the restart |
| 996 |
< |
* cases. Returns void, not task status (which must be reread by |
| 997 |
< |
* caller anyway) to slightly simplify control paths. |
| 737 |
> |
* greater than MAX_HELP links. All of these cases are dealt with |
| 738 |
> |
* by just retrying by caller. |
| 739 |
|
* |
| 740 |
|
* @param joinMe the task to join |
| 741 |
+ |
* @param canSteal true if local queue is empty |
| 742 |
+ |
* @return true if ran a task |
| 743 |
|
*/ |
| 744 |
< |
final void helpJoinTask(ForkJoinTask<?> joinMe, int retries) { |
| 744 |
> |
private boolean helpJoinTask(ForkJoinTask<?> joinMe) { |
| 745 |
> |
boolean helped = false; |
| 746 |
> |
int m = pool.scanGuard & SMASK; |
| 747 |
|
ForkJoinWorkerThread[] ws = pool.workers; |
| 748 |
< |
int n; |
| 749 |
< |
if (ws == null || (n = ws.length) <= 1) |
| 750 |
< |
return; // need at least 2 workers |
| 751 |
< |
|
| 1007 |
< |
restart:while (joinMe.status >= 0 && --retries >= 0) { |
| 1008 |
< |
ForkJoinTask<?> task = joinMe; // base of chain |
| 1009 |
< |
ForkJoinWorkerThread thread = this; // thread with stolen task |
| 1010 |
< |
for (int depth = 0; depth < MAX_HELP_DEPTH; ++depth) { |
| 748 |
> |
if (ws != null && ws.length > m && joinMe.status >= 0) { |
| 749 |
> |
int levels = MAX_HELP; // remaining chain length |
| 750 |
> |
ForkJoinTask<?> task = joinMe; // base of chain |
| 751 |
> |
outer:for (ForkJoinWorkerThread thread = this;;) { |
| 752 |
|
// Try to find v, the stealer of task, by first using hint |
| 753 |
< |
ForkJoinWorkerThread v = ws[thread.stealHint & (n - 1)]; |
| 753 |
> |
ForkJoinWorkerThread v = ws[thread.stealHint & m]; |
| 754 |
|
if (v == null || v.currentSteal != task) { |
| 755 |
< |
for (int j = 0; ; ++j) { // search array |
| 1015 |
< |
if (task.status < 0 || j == n) |
| 1016 |
< |
continue restart; // stale or no stealer |
| 755 |
> |
for (int j = 0; ;) { // search array |
| 756 |
|
if ((v = ws[j]) != null && v.currentSteal == task) { |
| 757 |
< |
thread.stealHint = j; // save for next time |
| 758 |
< |
break; |
| 757 |
> |
thread.stealHint = j; |
| 758 |
> |
break; // save hint for next time |
| 759 |
|
} |
| 760 |
+ |
if (++j > m) |
| 761 |
+ |
break outer; // can't find stealer |
| 762 |
|
} |
| 763 |
|
} |
| 764 |
|
// Try to help v, using specialized form of deqTask |
| 765 |
< |
int b; |
| 766 |
< |
ForkJoinTask<?>[] q; |
| 767 |
< |
while ((b = v.base) != v.sp && (q = v.queue) != null) { |
| 768 |
< |
int i = (q.length - 1) & b; |
| 769 |
< |
long u = (i << qShift) + qBase; |
| 765 |
> |
for (;;) { |
| 766 |
> |
ForkJoinTask<?>[] q; int b, i; |
| 767 |
> |
if (joinMe.status < 0) |
| 768 |
> |
break outer; |
| 769 |
> |
if ((b = v.queueBase) == v.queueTop || |
| 770 |
> |
(q = v.queue) == null || |
| 771 |
> |
(i = (q.length-1) & b) < 0) |
| 772 |
> |
break; // empty |
| 773 |
> |
long u = (i << ASHIFT) + ABASE; |
| 774 |
|
ForkJoinTask<?> t = q[i]; |
| 775 |
< |
if (task.status < 0) // stale |
| 776 |
< |
continue restart; |
| 777 |
< |
if (v.base == b) { // recheck after reading t |
| 778 |
< |
if (t == null) // producer stalled |
| 779 |
< |
continue restart; // retry via restart |
| 780 |
< |
if (UNSAFE.compareAndSwapObject(q, u, t, null)) { |
| 781 |
< |
if (joinMe.status < 0) { |
| 782 |
< |
UNSAFE.putObjectVolatile(q, u, t); |
| 783 |
< |
return; // back out on cancel |
| 784 |
< |
} |
| 785 |
< |
ForkJoinTask<?> prevSteal = currentSteal; |
| 1041 |
< |
currentSteal = t; |
| 1042 |
< |
v.stealHint = poolIndex; |
| 1043 |
< |
v.base = b + 1; |
| 1044 |
< |
t.tryExec(); |
| 1045 |
< |
currentSteal = prevSteal; |
| 1046 |
< |
} |
| 775 |
> |
if (task.status < 0) |
| 776 |
> |
break outer; // stale |
| 777 |
> |
if (t != null && v.queueBase == b && |
| 778 |
> |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
| 779 |
> |
v.queueBase = b + 1; |
| 780 |
> |
v.stealHint = poolIndex; |
| 781 |
> |
ForkJoinTask<?> ps = currentSteal; |
| 782 |
> |
currentSteal = t; |
| 783 |
> |
t.doExec(); |
| 784 |
> |
currentSteal = ps; |
| 785 |
> |
helped = true; |
| 786 |
|
} |
| 1048 |
– |
if (joinMe.status < 0) |
| 1049 |
– |
return; |
| 787 |
|
} |
| 788 |
|
// Try to descend to find v's stealer |
| 789 |
|
ForkJoinTask<?> next = v.currentJoin; |
| 790 |
< |
if (next == null || task.status < 0) |
| 791 |
< |
continue restart; // no descendent or stale |
| 792 |
< |
if (joinMe.status < 0) |
| 793 |
< |
return; |
| 794 |
< |
task = next; |
| 795 |
< |
thread = v; |
| 790 |
> |
if (--levels > 0 && task.status >= 0 && |
| 791 |
> |
next != null && next != task) { |
| 792 |
> |
task = next; |
| 793 |
> |
thread = v; |
| 794 |
> |
} |
| 795 |
> |
else |
| 796 |
> |
break; // max levels, stale, dead-end, or cyclic |
| 797 |
|
} |
| 798 |
|
} |
| 799 |
+ |
return helped; |
| 800 |
|
} |
| 801 |
|
|
| 802 |
|
/** |
| 803 |
< |
* Returns an estimate of the number of tasks, offset by a |
| 804 |
< |
* function of number of idle workers. |
| 803 |
> |
* Performs an uncommon case for joinTask: If task t is at base of |
| 804 |
> |
* some workers queue, steals and executes it. |
| 805 |
> |
* |
| 806 |
> |
* @param t the task |
| 807 |
> |
* @return t's status |
| 808 |
> |
*/ |
| 809 |
> |
private int tryDeqAndExec(ForkJoinTask<?> t) { |
| 810 |
> |
int m = pool.scanGuard & SMASK; |
| 811 |
> |
ForkJoinWorkerThread[] ws = pool.workers; |
| 812 |
> |
if (ws != null && ws.length > m && t.status >= 0) { |
| 813 |
> |
for (int j = 0; j <= m; ++j) { |
| 814 |
> |
ForkJoinTask<?>[] q; int b, i; |
| 815 |
> |
ForkJoinWorkerThread v = ws[j]; |
| 816 |
> |
if (v != null && |
| 817 |
> |
(b = v.queueBase) != v.queueTop && |
| 818 |
> |
(q = v.queue) != null && |
| 819 |
> |
(i = (q.length - 1) & b) >= 0 && |
| 820 |
> |
q[i] == t) { |
| 821 |
> |
long u = (i << ASHIFT) + ABASE; |
| 822 |
> |
if (v.queueBase == b && |
| 823 |
> |
UNSAFE.compareAndSwapObject(q, u, t, null)) { |
| 824 |
> |
v.queueBase = b + 1; |
| 825 |
> |
v.stealHint = poolIndex; |
| 826 |
> |
ForkJoinTask<?> ps = currentSteal; |
| 827 |
> |
currentSteal = t; |
| 828 |
> |
t.doExec(); |
| 829 |
> |
currentSteal = ps; |
| 830 |
> |
} |
| 831 |
> |
break; |
| 832 |
> |
} |
| 833 |
> |
} |
| 834 |
> |
} |
| 835 |
> |
return t.status; |
| 836 |
> |
} |
| 837 |
> |
|
| 838 |
> |
/** |
| 839 |
> |
* Implements ForkJoinTask.getSurplusQueuedTaskCount(). Returns |
| 840 |
> |
* an estimate of the number of tasks, offset by a function of |
| 841 |
> |
* number of idle workers. |
| 842 |
|
* |
| 843 |
|
* This method provides a cheap heuristic guide for task |
| 844 |
|
* partitioning when programmers, frameworks, tools, or languages |
| 874 |
|
* When all threads are active, it is on average OK to estimate |
| 875 |
|
* surplus strictly locally. In steady-state, if one thread is |
| 876 |
|
* maintaining say 2 surplus tasks, then so are others. So we can |
| 877 |
< |
* just use estimated queue length (although note that (sp - base) |
| 878 |
< |
* can be an overestimate because of stealers lagging increments |
| 879 |
< |
* of base). However, this strategy alone leads to serious |
| 880 |
< |
* mis-estimates in some non-steady-state conditions (ramp-up, |
| 881 |
< |
* ramp-down, other stalls). We can detect many of these by |
| 882 |
< |
* further considering the number of "idle" threads, that are |
| 877 |
> |
* just use estimated queue length (although note that (queueTop - |
| 878 |
> |
* queueBase) can be an overestimate because of stealers lagging |
| 879 |
> |
* increments of queueBase). However, this strategy alone leads |
| 880 |
> |
* to serious mis-estimates in some non-steady-state conditions |
| 881 |
> |
* (ramp-up, ramp-down, other stalls). We can detect many of these |
| 882 |
> |
* by further considering the number of "idle" threads, that are |
| 883 |
|
* known to have zero queued tasks, so compensate by a factor of |
| 884 |
|
* (#idle/#active) threads. |
| 885 |
|
*/ |
| 886 |
|
final int getEstimatedSurplusTaskCount() { |
| 887 |
< |
return sp - base - pool.idlePerActive(); |
| 887 |
> |
return queueTop - queueBase - pool.idlePerActive(); |
| 888 |
|
} |
| 889 |
|
|
| 890 |
|
/** |
| 891 |
< |
* Runs tasks until {@code pool.isQuiescent()}. |
| 891 |
> |
* Runs tasks until {@code pool.isQuiescent()}. We piggyback on |
| 892 |
> |
* pool's active count ctl maintenance, but rather than blocking |
| 893 |
> |
* when tasks cannot be found, we rescan until all others cannot |
| 894 |
> |
* find tasks either. The bracketing by pool quiescerCounts |
| 895 |
> |
* updates suppresses pool auto-shutdown mechanics that could |
| 896 |
> |
* otherwise prematurely terminate the pool because all threads |
| 897 |
> |
* appear to be inactive. |
| 898 |
|
*/ |
| 899 |
|
final void helpQuiescePool() { |
| 900 |
+ |
boolean active = true; |
| 901 |
+ |
ForkJoinTask<?> ps = currentSteal; // to restore below |
| 902 |
+ |
ForkJoinPool p = pool; |
| 903 |
+ |
p.addQuiescerCount(1); |
| 904 |
|
for (;;) { |
| 905 |
< |
ForkJoinTask<?> t = pollLocalTask(); |
| 906 |
< |
if (t != null || (t = scan()) != null) { |
| 907 |
< |
t.tryExec(); |
| 908 |
< |
currentSteal = null; |
| 905 |
> |
ForkJoinWorkerThread[] ws = p.workers; |
| 906 |
> |
ForkJoinWorkerThread v = null; |
| 907 |
> |
int n; |
| 908 |
> |
if (queueTop != queueBase) |
| 909 |
> |
v = this; |
| 910 |
> |
else if (ws != null && (n = ws.length) > 1) { |
| 911 |
> |
ForkJoinWorkerThread w; |
| 912 |
> |
int r = nextSeed(); // cheap version of FJP.scan |
| 913 |
> |
int steps = n << 1; |
| 914 |
> |
for (int i = 0; i < steps; ++i) { |
| 915 |
> |
if ((w = ws[(i + r) & (n - 1)]) != null && |
| 916 |
> |
w.queueBase != w.queueTop) { |
| 917 |
> |
v = w; |
| 918 |
> |
break; |
| 919 |
> |
} |
| 920 |
> |
} |
| 921 |
> |
} |
| 922 |
> |
if (v != null) { |
| 923 |
> |
ForkJoinTask<?> t; |
| 924 |
> |
if (!active) { |
| 925 |
> |
active = true; |
| 926 |
> |
p.addActiveCount(1); |
| 927 |
> |
} |
| 928 |
> |
if ((t = (v != this) ? v.deqTask() : |
| 929 |
> |
locallyFifo? locallyDeqTask() : popTask()) != null) { |
| 930 |
> |
currentSteal = t; |
| 931 |
> |
t.doExec(); |
| 932 |
> |
currentSteal = ps; |
| 933 |
> |
} |
| 934 |
|
} |
| 935 |
|
else { |
| 1125 |
– |
ForkJoinPool p = pool; |
| 936 |
|
if (active) { |
| 937 |
< |
active = false; // inactivate |
| 938 |
< |
do {} while (!p.tryDecrementActiveCount()); |
| 937 |
> |
active = false; |
| 938 |
> |
p.addActiveCount(-1); |
| 939 |
|
} |
| 940 |
|
if (p.isQuiescent()) { |
| 941 |
< |
active = true; // re-activate |
| 942 |
< |
do {} while (!p.tryIncrementActiveCount()); |
| 943 |
< |
return; |
| 941 |
> |
p.addActiveCount(1); |
| 942 |
> |
p.addQuiescerCount(-1); |
| 943 |
> |
break; |
| 944 |
|
} |
| 945 |
|
} |
| 946 |
|
} |
| 947 |
|
} |
| 948 |
|
|
| 949 |
|
// Unsafe mechanics |
| 950 |
< |
|
| 951 |
< |
private static final sun.misc.Unsafe UNSAFE = getUnsafe(); |
| 952 |
< |
private static final long runStateOffset = |
| 1143 |
< |
objectFieldOffset("runState", ForkJoinWorkerThread.class); |
| 1144 |
< |
private static final long qBase = |
| 1145 |
< |
UNSAFE.arrayBaseOffset(ForkJoinTask[].class); |
| 1146 |
< |
private static final int qShift; |
| 950 |
> |
private static final sun.misc.Unsafe UNSAFE; |
| 951 |
> |
private static final long ABASE; |
| 952 |
> |
private static final int ASHIFT; |
| 953 |
|
|
| 954 |
|
static { |
| 955 |
< |
int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class); |
| 1150 |
< |
if ((s & (s-1)) != 0) |
| 1151 |
< |
throw new Error("data type scale not a power of two"); |
| 1152 |
< |
qShift = 31 - Integer.numberOfLeadingZeros(s); |
| 1153 |
< |
} |
| 1154 |
< |
|
| 1155 |
< |
private static long objectFieldOffset(String field, Class<?> klazz) { |
| 955 |
> |
int s; |
| 956 |
|
try { |
| 957 |
< |
return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field)); |
| 958 |
< |
} catch (NoSuchFieldException e) { |
| 959 |
< |
// Convert Exception to corresponding Error |
| 960 |
< |
NoSuchFieldError error = new NoSuchFieldError(field); |
| 961 |
< |
error.initCause(e); |
| 962 |
< |
throw error; |
| 957 |
> |
UNSAFE = getUnsafe(); |
| 958 |
> |
Class a = ForkJoinTask[].class; |
| 959 |
> |
ABASE = UNSAFE.arrayBaseOffset(a); |
| 960 |
> |
s = UNSAFE.arrayIndexScale(a); |
| 961 |
> |
} catch (Exception e) { |
| 962 |
> |
throw new Error(e); |
| 963 |
|
} |
| 964 |
+ |
if ((s & (s-1)) != 0) |
| 965 |
+ |
throw new Error("data type scale not a power of two"); |
| 966 |
+ |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
| 967 |
|
} |
| 968 |
|
|
| 969 |
|
/** |
