--- jsr166/src/jsr166y/ForkJoinPool.java 2011/12/23 00:58:29 1.110 +++ jsr166/src/jsr166y/ForkJoinPool.java 2012/01/26 00:08:13 1.111 @@ -5,7 +5,6 @@ */ package jsr166y; - import java.util.ArrayList; import java.util.Arrays; import java.util.Collection; @@ -20,7 +19,7 @@ import java.util.concurrent.RejectedExec import java.util.concurrent.RunnableFuture; import java.util.concurrent.TimeUnit; import java.util.concurrent.atomic.AtomicInteger; -import java.util.concurrent.locks.LockSupport; +import java.util.concurrent.atomic.AtomicLong; import java.util.concurrent.locks.ReentrantLock; import java.util.concurrent.locks.Condition; @@ -33,12 +32,14 @@ import java.util.concurrent.locks.Condit *

A {@code ForkJoinPool} differs from other kinds of {@link * ExecutorService} mainly by virtue of employing * work-stealing: all threads in the pool attempt to find and - * execute subtasks created by other active tasks (eventually blocking - * waiting for work if none exist). This enables efficient processing - * when most tasks spawn other subtasks (as do most {@code - * ForkJoinTask}s). When setting asyncMode to true in - * constructors, {@code ForkJoinPool}s may also be appropriate for use - * with event-style tasks that are never joined. + * execute tasks submitted to the pool and/or created by other active + * tasks (eventually blocking waiting for work if none exist). This + * enables efficient processing when most tasks spawn other subtasks + * (as do most {@code ForkJoinTask}s), as well as when many small + * tasks are submitted to the pool from external clients. Especially + * when setting asyncMode to true in constructors, {@code + * ForkJoinPool}s may also be appropriate for use with event-style + * tasks that are never joined. * *

A {@code ForkJoinPool} is constructed with a given target * parallelism level; by default, equal to the number of available @@ -59,14 +60,16 @@ import java.util.concurrent.locks.Condit * *

As is the case with other ExecutorServices, there are three * main task execution methods summarized in the following - * table. These are designed to be used by clients not already engaged - * in fork/join computations in the current pool. The main forms of - * these methods accept instances of {@code ForkJoinTask}, but - * overloaded forms also allow mixed execution of plain {@code - * Runnable}- or {@code Callable}- based activities as well. However, - * tasks that are already executing in a pool should normally - * NOT use these pool execution methods, but instead use the - * within-computation forms listed in the table. + * table. These are designed to be used primarily by clients not + * already engaged in fork/join computations in the current pool. The + * main forms of these methods accept instances of {@code + * ForkJoinTask}, but overloaded forms also allow mixed execution of + * plain {@code Runnable}- or {@code Callable}- based activities as + * well. However, tasks that are already executing in a pool should + * normally instead use the within-computation forms listed in the + * table unless using async event-style tasks that are not usually + * joined, in which case there is little difference among choice of + * methods. * * * @@ -125,108 +128,195 @@ public class ForkJoinPool extends Abstra /* * Implementation Overview * - * This class provides the central bookkeeping and control for a - * set of worker threads: Submissions from non-FJ threads enter - * into a submission queue. Workers take these tasks and typically - * split them into subtasks that may be stolen by other workers. - * Preference rules give first priority to processing tasks from - * their own queues (LIFO or FIFO, depending on mode), then to - * randomized FIFO steals of tasks in other worker queues, and - * lastly to new submissions. + * This class and its nested classes provide the main + * functionality and control for a set of worker threads: + * Submissions from non-FJ threads enter into submission + * queues. Workers take these tasks and typically split them into + * subtasks that may be stolen by other workers. Preference rules + * give first priority to processing tasks from their own queues + * (LIFO or FIFO, depending on mode), then to randomized FIFO + * steals of tasks in other queues. + * + * WorkQueues. + * ========== + * + * Most operations occur within work-stealing queues (in nested + * class WorkQueue). These are special forms of Deques that + * support only three of the four possible end-operations -- push, + * pop, and poll (aka steal), under the further constraints that + * push and pop are called only from the owning thread (or, as + * extended here, under a lock), while poll may be called from + * other threads. (If you are unfamiliar with them, you probably + * want to read Herlihy and Shavit's book "The Art of + * Multiprocessor programming", chapter 16 describing these in + * more detail before proceeding.) The main work-stealing queue + * design is roughly similar to those in the papers "Dynamic + * Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 + * (http://research.sun.com/scalable/pubs/index.html) and + * "Idempotent work stealing" by Michael, Saraswat, and Vechev, + * PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). + * The main differences ultimately stem from gc requirements that + * we null out taken slots as soon as we can, to maintain as small + * a footprint as possible even in programs generating huge + * numbers of tasks. To accomplish this, we shift the CAS + * arbitrating pop vs poll (steal) from being on the indices + * ("base" and "top") to the slots themselves. So, both a + * successful pop and poll mainly entail a CAS of a slot from + * non-null to null. Because we rely on CASes of references, we + * do not need tag bits on base or top. They are simple ints as + * used in any circular array-based queue (see for example + * ArrayDeque). Updates to the indices must still be ordered in a + * way that guarantees that top == base means the queue is empty, + * but otherwise may err on the side of possibly making the queue + * appear nonempty when a push, pop, or poll have not fully + * committed. Note that this means that the poll operation, + * considered individually, is not wait-free. One thief cannot + * successfully continue until another in-progress one (or, if + * previously empty, a push) completes. However, in the + * aggregate, we ensure at least probabilistic non-blockingness. + * If an attempted steal fails, a thief always chooses a different + * random victim target to try next. So, in order for one thief to + * progress, it suffices for any in-progress poll or new push on + * any empty queue to complete. + * + * This approach also enables support a user mode in which local + * task processing is in FIFO, not LIFO order, simply by using + * poll rather than pop. This can be useful in message-passing + * frameworks in which tasks are never joined. However neither + * mode considers affinities, loads, cache localities, etc, so + * rarely provide the best possible performance on a given + * machine, but portably provide good throughput by averaging over + * these factors. (Further, even if we did try to use such + * information, we do not usually have a basis for exploiting + * it. For example, some sets of tasks profit from cache + * affinities, but others are harmed by cache pollution effects.) + * + * WorkQueues are also used in a similar way for tasks submitted + * to the pool. We cannot mix these tasks in the same queues used + * for work-stealing (this would contaminate lifo/fifo + * processing). Instead, we loosely associate (via hashing) + * submission queues with submitting threads, and randomly scan + * these queues as well when looking for work. In essence, + * submitters act like workers except that they never take tasks, + * and they are multiplexed on to a finite number of shared work + * queues. However, classes are set up so that future extensions + * could allow submitters to optionally help perform tasks as + * well. Pool submissions from internal workers are also allowed, + * but use randomized rather than thread-hashed queue indices to + * avoid imbalance. Insertion of tasks in shared mode requires a + * lock (mainly to protect in the case of resizing) but we use + * only a simple spinlock (using bits in field runState), because + * submitters encountering a busy queue try others so never block. + * + * Management. + * ========== * * The main throughput advantages of work-stealing stem from * decentralized control -- workers mostly take tasks from * themselves or each other. We cannot negate this in the * implementation of other management responsibilities. The main * tactic for avoiding bottlenecks is packing nearly all - * essentially atomic control state into a single 64bit volatile - * variable ("ctl"). This variable is read on the order of 10-100 - * times as often as it is modified (always via CAS). (There is - * some additional control state, for example variable "shutdown" - * for which we can cope with uncoordinated updates.) This - * streamlines synchronization and control at the expense of messy - * constructions needed to repack status bits upon updates. - * Updates tend not to contend with each other except during - * bursts while submitted tasks begin or end. In some cases when - * they do contend, threads can instead do something else - * (usually, scan for tasks) until contention subsides. - * - * To enable packing, we restrict maximum parallelism to (1<<15)-1 - * (which is far in excess of normal operating range) to allow - * ids, counts, and their negations (used for thresholding) to fit - * into 16bit fields. - * - * Recording Workers. Workers are recorded in the "workers" array - * that is created upon pool construction and expanded if (rarely) - * necessary. This is an array as opposed to some other data - * structure to support index-based random steals by workers. - * Updates to the array recording new workers and unrecording - * terminated ones are protected from each other by a seqLock - * (scanGuard) but the array is otherwise concurrently readable, - * and accessed directly by workers. To simplify index-based + * essentially atomic control state into two volatile variables + * that are by far most often read (not written) as status and + * consistency checks + * + * Field "ctl" contains 64 bits holding all the information needed + * to atomically decide to add, inactivate, enqueue (on an event + * queue), dequeue, and/or re-activate workers. To enable this + * packing, we restrict maximum parallelism to (1<<15)-1 (which is + * far in excess of normal operating range) to allow ids, counts, + * and their negations (used for thresholding) to fit into 16bit + * fields. + * + * Field "runState" contains 32 bits needed to register and + * deregister WorkQueues, as well as to enable shutdown. It is + * only modified under a lock (normally briefly held, but + * occasionally protecting allocations and resizings) but even + * when locked remains available to check consistency. + * + * Recording WorkQueues. WorkQueues are recorded in the + * "workQueues" array that is created upon pool construction and + * expanded if necessary. Updates to the array while recording + * new workers and unrecording terminated ones are protected from + * each other by a lock but the array is otherwise concurrently + * readable, and accessed directly. To simplify index-based * operations, the array size is always a power of two, and all - * readers must tolerate null slots. To avoid flailing during - * start-up, the array is presized to hold twice #parallelism - * workers (which is unlikely to need further resizing during - * execution). But to avoid dealing with so many null slots, - * variable scanGuard includes a mask for the nearest power of two - * that contains all current workers. All worker thread creation - * is on-demand, triggered by task submissions, replacement of - * terminated workers, and/or compensation for blocked - * workers. However, all other support code is set up to work with - * other policies. To ensure that we do not hold on to worker - * references that would prevent GC, ALL accesses to workers are - * via indices into the workers array (which is one source of some - * of the messy code constructions here). In essence, the workers - * array serves as a weak reference mechanism. Thus for example - * the wait queue field of ctl stores worker indices, not worker - * references. Access to the workers in associated methods (for - * example signalWork) must both index-check and null-check the - * IDs. All such accesses ignore bad IDs by returning out early - * from what they are doing, since this can only be associated - * with termination, in which case it is OK to give up. - * - * All uses of the workers array, as well as queue arrays, check - * that the array is non-null (even if previously non-null). This - * allows nulling during termination, which is currently not - * necessary, but remains an option for resource-revocation-based - * shutdown schemes. + * readers must tolerate null slots. Shared (submission) queues + * are at even indices, worker queues at odd indices. Grouping + * them together in this way simplifies and speeds up task + * scanning. To avoid flailing during start-up, the array is + * presized to hold twice #parallelism workers (which is unlikely + * to need further resizing during execution). But to avoid + * dealing with so many null slots, variable runState includes a + * mask for the nearest power of two that contains all current + * workers. All worker thread creation is on-demand, triggered by + * task submissions, replacement of terminated workers, and/or + * compensation for blocked workers. However, all other support + * code is set up to work with other policies. To ensure that we + * do not hold on to worker references that would prevent GC, ALL + * accesses to workQueues are via indices into the workQueues + * array (which is one source of some of the messy code + * constructions here). In essence, the workQueues array serves as + * a weak reference mechanism. Thus for example the wait queue + * field of ctl stores indices, not references. Access to the + * workQueues in associated methods (for example signalWork) must + * both index-check and null-check the IDs. All such accesses + * ignore bad IDs by returning out early from what they are doing, + * since this can only be associated with termination, in which + * case it is OK to give up. + * + * All uses of the workQueues array check that it is non-null + * (even if previously non-null). This allows nulling during + * termination, which is currently not necessary, but remains an + * option for resource-revocation-based shutdown schemes. It also + * helps reduce JIT issuance of uncommon-trap code, which tends to + * unnecessarily complicate control flow in some methods. * - * Wait Queuing. Unlike HPC work-stealing frameworks, we cannot + * Event Queuing. Unlike HPC work-stealing frameworks, we cannot * let workers spin indefinitely scanning for tasks when none can * be found immediately, and we cannot start/resume workers unless * there appear to be tasks available. On the other hand, we must * quickly prod them into action when new tasks are submitted or - * generated. We park/unpark workers after placing in an event - * wait queue when they cannot find work. This "queue" is actually - * a simple Treiber stack, headed by the "id" field of ctl, plus a - * 15bit counter value to both wake up waiters (by advancing their - * count) and avoid ABA effects. Successors are held in worker - * field "nextWait". Queuing deals with several intrinsic races, - * mainly that a task-producing thread can miss seeing (and + * generated. In many usages, ramp-up time to activate workers is + * the main limiting factor in overall performance (this is + * compounded at program start-up by JIT compilation and + * allocation). So we try to streamline this as much as possible. + * We park/unpark workers after placing in an event wait queue + * when they cannot find work. This "queue" is actually a simple + * Treiber stack, headed by the "id" field of ctl, plus a 15bit + * counter value (that reflects the number of times a worker has + * been inactivated) to avoid ABA effects (we need only as many + * version numbers as worker threads). Successors are held in + * field WorkQueue.nextWait. Queuing deals with several intrinsic + * races, mainly that a task-producing thread can miss seeing (and * signalling) another thread that gave up looking for work but * has not yet entered the wait queue. We solve this by requiring - * a full sweep of all workers both before (in scan()) and after - * (in tryAwaitWork()) a newly waiting worker is added to the wait - * queue. During a rescan, the worker might release some other - * queued worker rather than itself, which has the same net - * effect. Because enqueued workers may actually be rescanning - * rather than waiting, we set and clear the "parked" field of - * ForkJoinWorkerThread to reduce unnecessary calls to unpark. - * (Use of the parked field requires a secondary recheck to avoid - * missed signals.) + * a full sweep of all workers (via repeated calls to method + * scan()) both before and after a newly waiting worker is added + * to the wait queue. During a rescan, the worker might release + * some other queued worker rather than itself, which has the same + * net effect. Because enqueued workers may actually be rescanning + * rather than waiting, we set and clear the "parker" field of + * Workqueues to reduce unnecessary calls to unpark. (this + * requires a secondary recheck to avoid missed signals.) Note + * the unusual conventions about Thread.interrupts surrounding + * parking and other blocking: Because interrupts are used solely + * to alert threads to check termination, which is checked anyway + * upon blocking, we clear status (using Thread.interrupted) + * before any call to park, so that park does not immediately + * return due to status being set via some other unrelated call to + * interrupt in user code. * * Signalling. We create or wake up workers only when there * appears to be at least one task they might be able to find and * execute. When a submission is added or another worker adds a - * task to a queue that previously had two or fewer tasks, they + * task to a queue that previously had fewer than two tasks, they * signal waiting workers (or trigger creation of new ones if * fewer than the given parallelism level -- see signalWork). - * These primary signals are buttressed by signals during rescans - * as well as those performed when a worker steals a task and - * notices that there are more tasks too; together these cover the - * signals needed in cases when more than two tasks are pushed - * but untaken. + * These primary signals are buttressed by signals during rescans; + * together these cover the signals needed in cases when more + * tasks are pushed but untaken, and improve performance compared + * to having one thread wake up all workers. * * Trimming workers. To release resources after periods of lack of * use, a worker starting to wait when the pool is quiescent will @@ -234,82 +324,102 @@ public class ForkJoinPool extends Abstra * SHRINK_RATE nanosecs. This will slowly propagate, eventually * terminating all workers after long periods of non-use. * - * Submissions. External submissions are maintained in an - * array-based queue that is structured identically to - * ForkJoinWorkerThread queues except for the use of - * submissionLock in method addSubmission. Unlike the case for - * worker queues, multiple external threads can add new - * submissions, so adding requires a lock. - * - * Compensation. Beyond work-stealing support and lifecycle - * control, the main responsibility of this framework is to take - * actions when one worker is waiting to join a task stolen (or - * always held by) another. Because we are multiplexing many - * tasks on to a pool of workers, we can't just let them block (as - * in Thread.join). We also cannot just reassign the joiner's - * run-time stack with another and replace it later, which would - * be a form of "continuation", that even if possible is not - * necessarily a good idea since we sometimes need both an - * unblocked task and its continuation to progress. Instead we - * combine two tactics: + * Shutdown and Termination. A call to shutdownNow atomically sets + * a runState bit and then (non-atomically) sets each workers + * runState status, cancels all unprocessed tasks, and wakes up + * all waiting workers. Detecting whether termination should + * commence after a non-abrupt shutdown() call requires more work + * and bookkeeping. We need consensus about quiescence (i.e., that + * there is no more work). The active count provides a primary + * indication but non-abrupt shutdown still requires a rechecking + * scan for any workers that are inactive but not queued. + * + * Joining Tasks. + * ============== + * + * Any of several actions may be taken when one worker is waiting + * to join a task stolen (or always held by) another. Because we + * are multiplexing many tasks on to a pool of workers, we can't + * just let them block (as in Thread.join). We also cannot just + * reassign the joiner's run-time stack with another and replace + * it later, which would be a form of "continuation", that even if + * possible is not necessarily a good idea since we sometimes need + * both an unblocked task and its continuation to + * progress. Instead we combine two tactics: * * Helping: Arranging for the joiner to execute some task that it - * would be running if the steal had not occurred. Method - * ForkJoinWorkerThread.joinTask tracks joining->stealing - * links to try to find such a task. + * would be running if the steal had not occurred. * * Compensating: Unless there are already enough live threads, - * method tryPreBlock() may create or re-activate a spare - * thread to compensate for blocked joiners until they - * unblock. + * method tryCompensate() may create or re-activate a spare + * thread to compensate for blocked joiners until they unblock. + * + * A third form (implemented in tryRemoveAndExec and + * tryPollForAndExec) amounts to helping a hypothetical + * compensator: If we can readily tell that a possible action of a + * compensator is to steal and execute the task being joined, the + * joining thread can do so directly, without the need for a + * compensation thread (although at the expense of larger run-time + * stacks, but the tradeoff is typically worthwhile). * * The ManagedBlocker extension API can't use helping so relies * only on compensation in method awaitBlocker. * + * The algorithm in tryHelpStealer entails a form of "linear" + * helping: Each worker records (in field currentSteal) the most + * recent task it stole from some other worker. Plus, it records + * (in field currentJoin) the task it is currently actively + * joining. Method tryHelpStealer uses these markers to try to + * find a worker to help (i.e., steal back a task from and execute + * it) that could hasten completion of the actively joined task. + * In essence, the joiner executes a task that would be on its own + * local deque had the to-be-joined task not been stolen. This may + * be seen as a conservative variant of the approach in Wagner & + * Calder "Leapfrogging: a portable technique for implementing + * efficient futures" SIGPLAN Notices, 1993 + * (http://portal.acm.org/citation.cfm?id=155354). It differs in + * that: (1) We only maintain dependency links across workers upon + * steals, rather than use per-task bookkeeping. This sometimes + * requires a linear scan of workers array to locate stealers, but + * often doesn't because stealers leave hints (that may become + * stale/wrong) of where to locate them. A stealHint is only a + * hint because a worker might have had multiple steals and the + * hint records only one of them (usually the most current). + * Hinting isolates cost to when it is needed, rather than adding + * to per-task overhead. (2) It is "shallow", ignoring nesting + * and potentially cyclic mutual steals. (3) It is intentionally + * racy: field currentJoin is updated only while actively joining, + * which means that we miss links in the chain during long-lived + * tasks, GC stalls etc (which is OK since blocking in such cases + * is usually a good idea). (4) We bound the number of attempts + * to find work (see MAX_HELP_DEPTH) and fall back to suspending + * the worker and if necessary replacing it with another. + * * It is impossible to keep exactly the target parallelism number * of threads running at any given time. Determining the * existence of conservatively safe helping targets, the * availability of already-created spares, and the apparent need - * to create new spares are all racy and require heuristic - * guidance, so we rely on multiple retries of each. Currently, - * in keeping with on-demand signalling policy, we compensate only - * if blocking would leave less than one active (non-waiting, - * non-blocked) worker. Additionally, to avoid some false alarms - * due to GC, lagging counters, system activity, etc, compensated - * blocking for joins is only attempted after rechecks stabilize - * (retries are interspersed with Thread.yield, for good - * citizenship). The variable blockedCount, incremented before - * blocking and decremented after, is sometimes needed to - * distinguish cases of waiting for work vs blocking on joins or - * other managed sync. Both cases are equivalent for most pool - * control, so we can update non-atomically. (Additionally, - * contention on blockedCount alleviates some contention on ctl). - * - * Shutdown and Termination. A call to shutdownNow atomically sets - * the ctl stop bit and then (non-atomically) sets each workers - * "terminate" status, cancels all unprocessed tasks, and wakes up - * all waiting workers. Detecting whether termination should - * commence after a non-abrupt shutdown() call requires more work - * and bookkeeping. We need consensus about quiescence (i.e., that - * there is no more work) which is reflected in active counts so - * long as there are no current blockers, as well as possible - * re-evaluations during independent changes in blocking or - * quiescing workers. + * to create new spares are all racy, so we rely on multiple + * retries of each. Currently, in keeping with on-demand + * signalling policy, we compensate only if blocking would leave + * less than one active (non-waiting, non-blocked) worker. + * Additionally, to avoid some false alarms due to GC, lagging + * counters, system activity, etc, compensated blocking for joins + * is only attempted after rechecks stabilize in + * ForkJoinTask.awaitJoin. (Retries are interspersed with + * Thread.yield, for good citizenship.) * * Style notes: There is a lot of representation-level coupling * among classes ForkJoinPool, ForkJoinWorkerThread, and - * ForkJoinTask. Most fields of ForkJoinWorkerThread maintain - * data structures managed by ForkJoinPool, so are directly - * accessed. Conversely we allow access to "workers" array by - * workers, and direct access to ForkJoinTask.status by both - * ForkJoinPool and ForkJoinWorkerThread. There is little point - * trying to reduce this, since any associated future changes in - * representations will need to be accompanied by algorithmic - * changes anyway. All together, these low-level implementation - * choices produce as much as a factor of 4 performance - * improvement compared to naive implementations, and enable the - * processing of billions of tasks per second, at the expense of - * some ugliness. + * ForkJoinTask. The fields of WorkQueue maintain data structures + * managed by ForkJoinPool, so are directly accessed. There is + * little point trying to reduce this, since any associated future + * changes in representations will need to be accompanied by + * algorithmic changes anyway. All together, these low-level + * implementation choices produce as much as a factor of 4 + * performance improvement compared to naive implementations, and + * enable the processing of billions of tasks per second, at the + * expense of some ugliness. * * Methods signalWork() and scan() are the main bottlenecks so are * especially heavily micro-optimized/mangled. There are lots of @@ -326,11 +436,11 @@ public class ForkJoinPool extends Abstra * The order of declarations in this file is: (1) declarations of * statics (2) fields (along with constants used when unpacking * some of them), listed in an order that tends to reduce - * contention among them a bit under most JVMs. (3) internal - * control methods (4) callbacks and other support for - * ForkJoinTask and ForkJoinWorkerThread classes, (5) exported - * methods (plus a few little helpers). (6) static block - * initializing all statics in a minimally dependent order. + * contention among them a bit under most JVMs; (3) nested + * classes; (4) internal control methods; (5) callbacks and other + * support for ForkJoinTask methods; (6) exported methods (plus a + * few little helpers); (7) static block initializing all statics + * in a minimally dependent order. */ /** @@ -389,86 +499,14 @@ public class ForkJoinPool extends Abstra private static final AtomicInteger poolNumberGenerator; /** - * Generator for initial random seeds for worker victim - * selection. This is used only to create initial seeds. Random - * steals use a cheaper xorshift generator per steal attempt. We - * don't expect much contention on seedGenerator, so just use a - * plain Random. - */ - static final Random workerSeedGenerator; - - /** - * Array holding all worker threads in the pool. Initialized upon - * construction. Array size must be a power of two. Updates and - * replacements are protected by scanGuard, but the array is - * always kept in a consistent enough state to be randomly - * accessed without locking by workers performing work-stealing, - * as well as other traversal-based methods in this class, so long - * as reads memory-acquire by first reading ctl. All readers must - * tolerate that some array slots may be null. - */ - ForkJoinWorkerThread[] workers; - - /** - * Initial size for submission queue array. Must be a power of - * two. In many applications, these always stay small so we use a - * small initial cap. - */ - private static final int INITIAL_QUEUE_CAPACITY = 8; - - /** - * Maximum size for submission queue array. Must be a power of two - * less than or equal to 1 << (31 - width of array entry) to - * ensure lack of index wraparound, but is capped at a lower - * value to help users trap runaway computations. - */ - private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M - - /** - * Array serving as submission queue. Initialized upon construction. - */ - private ForkJoinTask[] submissionQueue; - - /** - * Lock protecting submissions array for addSubmission - */ - private final ReentrantLock submissionLock; - - /** - * Condition for awaitTermination, using submissionLock for - * convenience. - */ - private final Condition termination; - - /** - * Creation factory for worker threads. - */ - private final ForkJoinWorkerThreadFactory factory; - - /** - * The uncaught exception handler used when any worker abruptly - * terminates. - */ - final Thread.UncaughtExceptionHandler ueh; - - /** - * Prefix for assigning names to worker threads - */ - private final String workerNamePrefix; - - /** - * Sum of per-thread steal counts, updated only when threads are - * idle or terminating. - */ - private volatile long stealCount; - - /** - * Main pool control -- a long packed with: + * Bits and masks for control variables + * + * Field ctl is a long packed with: * AC: Number of active running workers minus target parallelism (16 bits) * TC: Number of total workers minus target parallelism (16 bits) * ST: true if pool is terminating (1 bit) * EC: the wait count of top waiting thread (15 bits) - * ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits) + * ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits) * * When convenient, we can extract the upper 32 bits of counts and * the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e = @@ -482,8 +520,25 @@ public class ForkJoinPool extends Abstra * negative, the pool is terminating. To deal with these possibly * negative fields, we use casts in and out of "short" and/or * signed shifts to maintain signedness. + * + * When a thread is queued (inactivated), its eventCount field is + * negative, which is the only way to tell if a worker is + * prevented from executing tasks, even though it must continue to + * scan for them to avoid queuing races. + * + * Field runState is an int packed with: + * SHUTDOWN: true if shutdown is enabled (1 bit) + * SEQ: a sequence number updated upon (de)registering workers (15 bits) + * MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits) + * + * The combination of mask and sequence number enables simple + * consistency checks: Staleness of read-only operations on the + * workers and queues arrays can be checked by comparing runState + * before vs after the reads. The low 16 bits (i.e, anding with + * SMASK) hold (the smallest power of two covering all worker + * indices, minus one. The mask for queues (vs workers) is twice + * this value plus 1. */ - volatile long ctl; // bit positions/shifts for fields private static final int AC_SHIFT = 48; @@ -515,113 +570,666 @@ public class ForkJoinPool extends Abstra private static final int UTC_UNIT = 1 << UTC_SHIFT; // masks and units for dealing with e = (int)ctl - private static final int E_MASK = 0x7fffffff; // no STOP_BIT - private static final int EC_UNIT = 1 << EC_SHIFT; + private static final int E_MASK = 0x7fffffff; // no STOP_BIT + private static final int E_SEQ = 1 << EC_SHIFT; - /** - * The target parallelism level. - */ - final int parallelism; + // runState bits + private static final int SHUTDOWN = 1 << 31; + private static final int RS_SEQ = 1 << 16; + private static final int RS_SEQ_MASK = 0x7fff0000; + + // access mode for WorkQueue + static final int LIFO_QUEUE = 0; + static final int FIFO_QUEUE = 1; + static final int SHARED_QUEUE = -1; /** - * Index (mod submission queue length) of next element to take - * from submission queue. Usage is identical to that for - * per-worker queues -- see ForkJoinWorkerThread internal - * documentation. + * The wakeup interval (in nanoseconds) for a worker waiting for a + * task when the pool is quiescent to instead try to shrink the + * number of workers. The exact value does not matter too + * much. It must be short enough to release resources during + * sustained periods of idleness, but not so short that threads + * are continually re-created. */ - volatile int queueBase; + private static final long SHRINK_RATE = + 4L * 1000L * 1000L * 1000L; // 4 seconds /** - * Index (mod submission queue length) of next element to add - * in submission queue. Usage is identical to that for - * per-worker queues -- see ForkJoinWorkerThread internal - * documentation. + * The timeout value for attempted shrinkage, includes + * some slop to cope with system timer imprecision. */ - int queueTop; + private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10); /** - * True when shutdown() has been called. + * The maximum stolen->joining link depth allowed in tryHelpStealer. + * Depths for legitimate chains are unbounded, but we use a fixed + * constant to avoid (otherwise unchecked) cycles and to bound + * staleness of traversal parameters at the expense of sometimes + * blocking when we could be helping. */ - volatile boolean shutdown; + private static final int MAX_HELP_DEPTH = 16; - /** - * True if use local fifo, not default lifo, for local polling. - * Read by, and replicated by ForkJoinWorkerThreads. + /* + * Field layout order in this class tends to matter more than one + * would like. Runtime layout order is only loosely related to + * declaration order and may differ across JVMs, but the following + * empirically works OK on current JVMs. + */ + + volatile long ctl; // main pool control + final int parallelism; // parallelism level + final int localMode; // per-worker scheduling mode + int nextPoolIndex; // hint used in registerWorker + volatile int runState; // shutdown status, seq, and mask + WorkQueue[] workQueues; // main registry + final ReentrantLock lock; // for registration + final Condition termination; // for awaitTermination + final ForkJoinWorkerThreadFactory factory; // factory for new workers + final Thread.UncaughtExceptionHandler ueh; // per-worker UEH + final AtomicLong stealCount; // collect counts when terminated + final AtomicInteger nextWorkerNumber; // to create worker name string + final String workerNamePrefix; // Prefix for assigning worker names + + /** + * Queues supporting work-stealing as well as external task + * submission. See above for main rationale and algorithms. + * Implementation relies heavily on "Unsafe" intrinsics + * and selective use of "volatile": + * + * Field "base" is the index (mod array.length) of the least valid + * queue slot, which is always the next position to steal (poll) + * from if nonempty. Reads and writes require volatile orderings + * but not CAS, because updates are only performed after slot + * CASes. + * + * Field "top" is the index (mod array.length) of the next queue + * slot to push to or pop from. It is written only by owner thread + * for push, or under lock for trySharedPush, and accessed by + * other threads only after reading (volatile) base. Both top and + * base are allowed to wrap around on overflow, but (top - base) + * (or more comonly -(base - top) to force volatile read of base + * before top) still estimates size. + * + * The array slots are read and written using the emulation of + * volatiles/atomics provided by Unsafe. Insertions must in + * general use putOrderedObject as a form of releasing store to + * ensure that all writes to the task object are ordered before + * its publication in the queue. (Although we can avoid one case + * of this when locked in trySharedPush.) All removals entail a + * CAS to null. The array is always a power of two. To ensure + * safety of Unsafe array operations, all accesses perform + * explicit null checks and implicit bounds checks via + * power-of-two masking. + * + * In addition to basic queuing support, this class contains + * fields described elsewhere to control execution. It turns out + * to work better memory-layout-wise to include them in this + * class rather than a separate class. + * + * Performance on most platforms is very sensitive to placement of + * instances of both WorkQueues and their arrays -- we absolutely + * do not want multiple WorkQueue instances or multiple queue + * arrays sharing cache lines. (It would be best for queue objects + * and their arrays to share, but there is nothing available to + * help arrange that). Unfortunately, because they are recorded + * in a common array, WorkQueue instances are often moved to be + * adjacent by garbage collectors. To reduce impact, we use field + * padding that works OK on common platforms; this effectively + * trades off slightly slower average field access for the sake of + * avoiding really bad worst-case access. (Until better JVM + * support is in place, this padding is dependent on transient + * properties of JVM field layout rules.) We also take care in + * allocating and sizing and resizing the array. Non-shared queue + * arrays are initialized (via method growArray) by workers before + * use. Others are allocated on first use. */ - final boolean locallyFifo; + static final class WorkQueue { + /** + * Capacity of work-stealing queue array upon initialization. + * Must be a power of two; at least 4, but set larger to + * reduce cacheline sharing among queues. + */ + static final int INITIAL_QUEUE_CAPACITY = 1 << 8; + + /** + * Maximum size for queue arrays. Must be a power of two less + * than or equal to 1 << (31 - width of array entry) to ensure + * lack of wraparound of index calculations, but defined to a + * value a bit less than this to help users trap runaway + * programs before saturating systems. + */ + static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M + + volatile long totalSteals; // cumulative number of steals + int seed; // for random scanning; initialize nonzero + volatile int eventCount; // encoded inactivation count; < 0 if inactive + int nextWait; // encoded record of next event waiter + int rescans; // remaining scans until block + int nsteals; // top-level task executions since last idle + final int mode; // lifo, fifo, or shared + int poolIndex; // index of this queue in pool (or 0) + int stealHint; // index of most recent known stealer + volatile int runState; // 1: locked, -1: terminate; else 0 + volatile int base; // index of next slot for poll + int top; // index of next slot for push + ForkJoinTask[] array; // the elements (initially unallocated) + final ForkJoinWorkerThread owner; // owning thread or null if shared + volatile Thread parker; // == owner during call to park; else null + ForkJoinTask currentJoin; // task being joined in awaitJoin + ForkJoinTask currentSteal; // current non-local task being executed + // Heuristic padding to ameliorate unfortunate memory placements + Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a; + + WorkQueue(ForkJoinWorkerThread owner, int mode) { + this.owner = owner; + this.mode = mode; + // Place indices in the center of array (that is not yet allocated) + base = top = INITIAL_QUEUE_CAPACITY >>> 1; + } + + /** + * Returns number of tasks in the queue + */ + final int queueSize() { + int n = base - top; // non-owner callers must read base first + return (n >= 0) ? 0 : -n; + } + + /** + * Pushes a task. Call only by owner in unshared queues. + * + * @param task the task. Caller must ensure non-null. + * @param p, if non-null, pool to signal if necessary + * @throw RejectedExecutionException if array cannot + * be resized + */ + final void push(ForkJoinTask task, ForkJoinPool p) { + boolean signal = false; + ForkJoinTask[] a; + int s = top, m, n; + if ((a = array) != null) { // ignore if queue removed + U.putOrderedObject + (a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task); + if ((n = (top = s + 1) - base) <= 2) { + if (p != null) + p.signalWork(); + } + else if (n >= m) + growArray(true); + } + } + + /** + * Pushes a task if lock is free and array is either big + * enough or can be resized to be big enough. + * + * @param task the task. Caller must ensure non-null. + * @return true if submitted + */ + final boolean trySharedPush(ForkJoinTask task) { + boolean submitted = false; + if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { + ForkJoinTask[] a = array; + int s = top, n = s - base; + try { + if ((a != null && n < a.length - 1) || + (a = growArray(false)) != null) { // must presize + int j = (((a.length - 1) & s) << ASHIFT) + ABASE; + U.putObject(a, (long)j, task); // don't need "ordered" + top = s + 1; + submitted = true; + } + } finally { + runState = 0; // unlock + } + } + return submitted; + } + + /** + * Takes next task, if one exists, in FIFO order. + */ + final ForkJoinTask poll() { + ForkJoinTask[] a; int b, i; + while ((b = base) - top < 0 && (a = array) != null && + (i = (a.length - 1) & b) >= 0) { + int j = (i << ASHIFT) + ABASE; + ForkJoinTask t = (ForkJoinTask)U.getObjectVolatile(a, j); + if (t != null && base == b && + U.compareAndSwapObject(a, j, t, null)) { + base = b + 1; + return t; + } + } + return null; + } + + /** + * Takes next task, if one exists, in LIFO order. + * Call only by owner in unshared queues. + */ + final ForkJoinTask pop() { + ForkJoinTask t; int m; + ForkJoinTask[] a = array; + if (a != null && (m = a.length - 1) >= 0) { + for (int s; (s = top - 1) - base >= 0;) { + int j = ((m & s) << ASHIFT) + ABASE; + if ((t = (ForkJoinTask)U.getObjectVolatile(a, j)) == null) + break; + if (U.compareAndSwapObject(a, j, t, null)) { + top = s; + return t; + } + } + } + return null; + } + + /** + * Takes next task, if one exists, in order specified by mode. + */ + final ForkJoinTask nextLocalTask() { + return mode == 0 ? pop() : poll(); + } + + /** + * Returns next task, if one exists, in order specified by mode. + */ + final ForkJoinTask peek() { + ForkJoinTask[] a = array; int m; + if (a == null || (m = a.length - 1) < 0) + return null; + int i = mode == 0 ? top - 1 : base; + int j = ((i & m) << ASHIFT) + ABASE; + return (ForkJoinTask)U.getObjectVolatile(a, j); + } + + /** + * Returns task at index b if b is current base of queue. + */ + final ForkJoinTask pollAt(int b) { + ForkJoinTask[] a; int i; + ForkJoinTask task = null; + if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) { + int j = (i << ASHIFT) + ABASE; + ForkJoinTask t = (ForkJoinTask)U.getObjectVolatile(a, j); + if (t != null && base == b && + U.compareAndSwapObject(a, j, t, null)) { + base = b + 1; + task = t; + } + } + return task; + } + + /** + * Pops the given task only if it is at the current top. + */ + final boolean tryUnpush(ForkJoinTask t) { + ForkJoinTask[] a; int s; + if ((a = array) != null && (s = top) != base && + U.compareAndSwapObject + (a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) { + top = s; + return true; + } + return false; + } + + /** + * Polls the given task only if it is at the current base. + */ + final boolean pollFor(ForkJoinTask task) { + ForkJoinTask[] a; int b, i; + if ((b = base) - top < 0 && (a = array) != null && + (i = (a.length - 1) & b) >= 0) { + int j = (i << ASHIFT) + ABASE; + if (U.getObjectVolatile(a, j) == task && base == b && + U.compareAndSwapObject(a, j, task, null)) { + base = b + 1; + return true; + } + } + return false; + } + + /** + * If present, removes from queue and executes the given task, or + * any other cancelled task. Returns (true) immediately on any CAS + * or consistency check failure so caller can retry. + * + * @return false if no progress can be made + */ + final boolean tryRemoveAndExec(ForkJoinTask task) { + boolean removed = false, empty = true, progress = true; + ForkJoinTask[] a; int m, s, b, n; + if ((a = array) != null && (m = a.length - 1) >= 0 && + (n = (s = top) - (b = base)) > 0) { + for (ForkJoinTask t;;) { // traverse from s to b + int j = ((--s & m) << ASHIFT) + ABASE; + t = (ForkJoinTask)U.getObjectVolatile(a, j); + if (t == null) // inconsistent length + break; + else if (t == task) { + if (s + 1 == top) { // pop + if (!U.compareAndSwapObject(a, j, task, null)) + break; + top = s; + removed = true; + } + else if (base == b) // replace with proxy + removed = U.compareAndSwapObject(a, j, task, + new EmptyTask()); + break; + } + else if (t.status >= 0) + empty = false; + else if (s + 1 == top) { // pop and throw away + if (U.compareAndSwapObject(a, j, t, null)) + top = s; + break; + } + if (--n == 0) { + if (!empty && base == b) + progress = false; + break; + } + } + } + if (removed) + task.doExec(); + return progress; + } + + /** + * Initializes or doubles the capacity of array. Call either + * by owner or with lock held -- it is OK for base, but not + * top, to move while resizings are in progress. + * + * @param rejectOnFailure if true, throw exception if capacity + * exceeded (relayed ultimately to user); else return null. + */ + final ForkJoinTask[] growArray(boolean rejectOnFailure) { + ForkJoinTask[] oldA = array; + int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; + if (size <= MAXIMUM_QUEUE_CAPACITY) { + int oldMask, t, b; + ForkJoinTask[] a = array = new ForkJoinTask[size]; + if (oldA != null && (oldMask = oldA.length - 1) >= 0 && + (t = top) - (b = base) > 0) { + int mask = size - 1; + do { + ForkJoinTask x; + int oldj = ((b & oldMask) << ASHIFT) + ABASE; + int j = ((b & mask) << ASHIFT) + ABASE; + x = (ForkJoinTask)U.getObjectVolatile(oldA, oldj); + if (x != null && + U.compareAndSwapObject(oldA, oldj, x, null)) + U.putObjectVolatile(a, j, x); + } while (++b != t); + } + return a; + } + else if (!rejectOnFailure) + return null; + else + throw new RejectedExecutionException("Queue capacity exceeded"); + } + + /** + * Removes and cancels all known tasks, ignoring any exceptions + */ + final void cancelAll() { + ForkJoinTask.cancelIgnoringExceptions(currentJoin); + ForkJoinTask.cancelIgnoringExceptions(currentSteal); + for (ForkJoinTask t; (t = poll()) != null; ) + ForkJoinTask.cancelIgnoringExceptions(t); + } + + // Execution methods + + /** + * Removes and runs tasks until empty, using local mode + * ordering. + */ + final void runLocalTasks() { + if (base - top < 0) { + for (ForkJoinTask t; (t = nextLocalTask()) != null; ) + t.doExec(); + } + } + + /** + * Executes a top-level task and any local tasks remaining + * after execution. + * + * @return true unless terminating + */ + final boolean runTask(ForkJoinTask t) { + boolean alive = true; + if (t != null) { + currentSteal = t; + t.doExec(); + runLocalTasks(); + ++nsteals; + currentSteal = null; + } + else if (runState < 0) // terminating + alive = false; + return alive; + } + + /** + * Executes a non-top-level (stolen) task + */ + final void runSubtask(ForkJoinTask t) { + if (t != null) { + ForkJoinTask ps = currentSteal; + currentSteal = t; + t.doExec(); + currentSteal = ps; + } + } + + /** + * Computes next value for random probes. Scans don't require + * a very high quality generator, but also not a crummy one. + * Marsaglia xor-shift is cheap and works well enough. Note: + * This is manually inlined in several usages in ForkJoinPool + * to avoid writes inside busy scan loops. + */ + final int nextSeed() { + int r = seed; + r ^= r << 13; + r ^= r >>> 17; + r ^= r << 5; + return seed = r; + } + + // Unsafe mechanics + private static final sun.misc.Unsafe U; + private static final long RUNSTATE; + private static final int ABASE; + private static final int ASHIFT; + static { + int s; + try { + U = getUnsafe(); + Class k = WorkQueue.class; + Class ak = ForkJoinTask[].class; + RUNSTATE = U.objectFieldOffset + (k.getDeclaredField("runState")); + ABASE = U.arrayBaseOffset(ak); + s = U.arrayIndexScale(ak); + } catch (Exception e) { + throw new Error(e); + } + if ((s & (s-1)) != 0) + throw new Error("data type scale not a power of two"); + ASHIFT = 31 - Integer.numberOfLeadingZeros(s); + } + } /** - * The number of threads in ForkJoinWorkerThreads.helpQuiescePool. - * When non-zero, suppresses automatic shutdown when active - * counts become zero. + * Class for artificial tasks that are used to replace the target + * of local joins if they are removed from an interior queue slot + * in WorkQueue.tryRemoveAndExec. We don't need the proxy to + * actually do anything beyond having a unique identity. */ - volatile int quiescerCount; + static final class EmptyTask extends ForkJoinTask { + EmptyTask() { status = ForkJoinTask.NORMAL; } // force done + public Void getRawResult() { return null; } + public void setRawResult(Void x) {} + public boolean exec() { return true; } + } /** - * The number of threads blocked in join. + * Computes a hash code for the given thread. This method is + * expected to provide higher-quality hash codes than those using + * method hashCode(). */ - volatile int blockedCount; + static final int hashThread(Thread t) { + long id = (t == null)? 0L : t.getId(); // Use MurmurHash of thread id + int h = (int)id ^ (int)(id >>> 32); + h ^= h >>> 16; + h *= 0x85ebca6b; + h ^= h >>> 13; + h *= 0xc2b2ae35; + return h ^ (h >>> 16); + } /** - * Counter for worker Thread names (unrelated to their poolIndex) + * Top-level runloop for workers */ - private volatile int nextWorkerNumber; + final void runWorker(ForkJoinWorkerThread wt) { + WorkQueue w = wt.workQueue; + w.growArray(false); // Initialize queue array and seed in this thread + w.seed = hashThread(Thread.currentThread()) | (1 << 31); // force < 0 + + do {} while (w.runTask(scan(w))); + } + + // Creating, registering and deregistering workers /** - * The index for the next created worker. Accessed under scanGuard. + * Tries to create and start a worker */ - private int nextWorkerIndex; + private void addWorker() { + Throwable ex = null; + ForkJoinWorkerThread w = null; + try { + if ((w = factory.newThread(this)) != null) { + w.start(); + return; + } + } catch (Throwable e) { + ex = e; + } + deregisterWorker(w, ex); + } /** - * SeqLock and index masking for updates to workers array. Locked - * when SG_UNIT is set. Unlocking clears bit by adding - * SG_UNIT. Staleness of read-only operations can be checked by - * comparing scanGuard to value before the reads. The low 16 bits - * (i.e, anding with SMASK) hold (the smallest power of two - * covering all worker indices, minus one, and is used to avoid - * dealing with large numbers of null slots when the workers array - * is overallocated. + * Callback from ForkJoinWorkerThread constructor to assign a + * public name. This must be separate from registerWorker because + * it is called during the "super" constructor call in + * ForkJoinWorkerThread. */ - volatile int scanGuard; - - private static final int SG_UNIT = 1 << 16; + final String nextWorkerName() { + return workerNamePrefix.concat + (Integer.toString(nextWorkerNumber.addAndGet(1))); + } /** - * The wakeup interval (in nanoseconds) for a worker waiting for a - * task when the pool is quiescent to instead try to shrink the - * number of workers. The exact value does not matter too - * much. It must be short enough to release resources during - * sustained periods of idleness, but not so short that threads - * are continually re-created. + * Callback from ForkJoinWorkerThread constructor to establish and + * record its WorkQueue + * + * @param wt the worker thread */ - private static final long SHRINK_RATE = - 4L * 1000L * 1000L * 1000L; // 4 seconds + final void registerWorker(ForkJoinWorkerThread wt) { + WorkQueue w = wt.workQueue; + ReentrantLock lock = this.lock; + lock.lock(); + try { + int k = nextPoolIndex; + WorkQueue[] ws = workQueues; + if (ws != null) { // ignore on shutdown + int n = ws.length; + if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) { + for (k = 1; k < n && ws[k] != null; k += 2) + ; // workers are at odd indices + if (k >= n) // resize + workQueues = ws = Arrays.copyOf(ws, n << 1); + } + w.poolIndex = k; + w.eventCount = ~(k >>> 1) & SMASK; // Set up wait count + ws[k] = w; // record worker + nextPoolIndex = k + 2; + int rs = runState; + int m = rs & SMASK; // recalculate runState mask + if (k > m) + m = (m << 1) + 1; + runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m; + } + } finally { + lock.unlock(); + } + } /** - * Top-level loop for worker threads: On each step: if the - * previous step swept through all queues and found no tasks, or - * there are excess threads, then possibly blocks. Otherwise, - * scans for and, if found, executes a task. Returns when pool - * and/or worker terminate. + * Final callback from terminating worker, as well as failure to + * construct or start a worker in addWorker. Removes record of + * worker from array, and adjusts counts. If pool is shutting + * down, tries to complete termination. * - * @param w the worker + * @param wt the worker thread or null if addWorker failed + * @param ex the exception causing failure, or null if none */ - final void work(ForkJoinWorkerThread w) { - boolean swept = false; // true on empty scans - long c; - while (!w.terminate && (int)(c = ctl) >= 0) { - int a; // active count - if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0) - swept = scan(w, a); - else if (tryAwaitWork(w, c)) - swept = false; + final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { + WorkQueue w = null; + if (wt != null && (w = wt.workQueue) != null) { + w.runState = -1; // ensure runState is set + stealCount.getAndAdd(w.totalSteals + w.nsteals); + int idx = w.poolIndex; + ReentrantLock lock = this.lock; + lock.lock(); + try { // remove record from array + WorkQueue[] ws = workQueues; + if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) + ws[nextPoolIndex = idx] = null; + } finally { + lock.unlock(); + } + } + + long c; // adjust ctl counts + do {} while (!U.compareAndSwapLong + (this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) | + ((c - TC_UNIT) & TC_MASK) | + (c & ~(AC_MASK|TC_MASK))))); + + if (!tryTerminate(false) && w != null) { + w.cancelAll(); // cancel remaining tasks + if (w.array != null) // suppress signal if never ran + signalWork(); // wake up or create replacement } + + if (ex != null) // rethrow + U.throwException(ex); } - // Signalling + + // Maintaining ctl counts /** - * Wakes up or creates a worker. + * Increments active count; mainly called upon return from blocking + */ + final void incrementActiveCount() { + long c; + do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT)); + } + + /** + * Activates or creates a worker */ final void signalWork() { /* @@ -637,731 +1245,611 @@ public class ForkJoinPool extends Abstra */ long c; int e, u; while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) & - (INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) { - if (e > 0) { // release a waiting worker - int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws; - if ((ws = workers) == null || - (i = ~e & SMASK) >= ws.length || - (w = ws[i]) == null) + (INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) { + WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p; + if (e == 0) { // add a new worker + if (U.compareAndSwapLong + (this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) | + ((u + UAC_UNIT) & UAC_MASK)) << 32)) { + addWorker(); break; - long nc = (((long)(w.nextWait & E_MASK)) | - ((long)(u + UAC_UNIT) << 32)); - if (w.eventCount == e && - UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { - w.eventCount = (e + EC_UNIT) & E_MASK; - if (w.parked) - UNSAFE.unpark(w); + } + } + else if (e > 0 && ws != null && + (i = ((~e << 1) | 1) & SMASK) < ws.length && + (w = ws[i]) != null && + w.eventCount == (e | INT_SIGN)) { + if (U.compareAndSwapLong + (this, CTL, c, (((long)(w.nextWait & E_MASK)) | + ((long)(u + UAC_UNIT) << 32)))) { + w.eventCount = (e + E_SEQ) & E_MASK; + if ((p = w.parker) != null) + U.unpark(p); // release a waiting worker break; } } - else if (UNSAFE.compareAndSwapLong - (this, ctlOffset, c, - (long)(((u + UTC_UNIT) & UTC_MASK) | - ((u + UAC_UNIT) & UAC_MASK)) << 32)) { - addWorker(); + else break; - } } } /** - * Variant of signalWork to help release waiters on rescans. - * Tries once to release a waiter if active count < 0. + * Tries to decrement active count (sometimes implicitly) and + * possibly release or create a compensating worker in preparation + * for blocking. Fails on contention or termination. * - * @return false if failed due to contention, else true + * @return true if the caller can block, else should recheck and retry */ - private boolean tryReleaseWaiter() { - long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws; - if ((e = (int)(c = ctl)) > 0 && - (int)(c >> AC_SHIFT) < 0 && - (ws = workers) != null && - (i = ~e & SMASK) < ws.length && - (w = ws[i]) != null) { - long nc = ((long)(w.nextWait & E_MASK) | - ((c + AC_UNIT) & (AC_MASK|TC_MASK))); - if (w.eventCount != e || - !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) - return false; - w.eventCount = (e + EC_UNIT) & E_MASK; - if (w.parked) - UNSAFE.unpark(w); - } - return true; - } - - // Scanning for tasks + final boolean tryCompensate() { + WorkQueue[] ws; WorkQueue w; Thread p; + int pc = parallelism, e, u, ac, tc, i; + long c = ctl; - /** - * Scans for and, if found, executes one task. Scans start at a - * random index of workers array, and randomly select the first - * (2*#workers)-1 probes, and then, if all empty, resort to 2 - * circular sweeps, which is necessary to check quiescence. and - * taking a submission only if no stealable tasks were found. The - * steal code inside the loop is a specialized form of - * ForkJoinWorkerThread.deqTask, followed bookkeeping to support - * helpJoinTask and signal propagation. The code for submission - * queues is almost identical. On each steal, the worker completes - * not only the task, but also all local tasks that this task may - * have generated. On detecting staleness or contention when - * trying to take a task, this method returns without finishing - * sweep, which allows global state rechecks before retry. - * - * @param w the worker - * @param a the number of active workers - * @return true if swept all queues without finding a task - */ - private boolean scan(ForkJoinWorkerThread w, int a) { - int g = scanGuard; // mask 0 avoids useless scans if only one active - int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK; - ForkJoinWorkerThread[] ws = workers; - if (ws == null || ws.length <= m) // staleness check - return false; - for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) { - ForkJoinTask t; ForkJoinTask[] q; int b, i; - ForkJoinWorkerThread v = ws[k & m]; - if (v != null && (b = v.queueBase) != v.queueTop && - (q = v.queue) != null && (i = (q.length - 1) & b) >= 0) { - long u = (i << ASHIFT) + ABASE; - if ((t = q[i]) != null && v.queueBase == b && - UNSAFE.compareAndSwapObject(q, u, t, null)) { - int d = (v.queueBase = b + 1) - v.queueTop; - v.stealHint = w.poolIndex; - if (d != 0) - signalWork(); // propagate if nonempty - w.execTask(t); + if ((e = (int)c) >= 0) { + if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 && + e != 0 && (ws = workQueues) != null && + (i = ((~e << 1) | 1) & SMASK) < ws.length && + (w = ws[i]) != null) { + if (w.eventCount == (e | INT_SIGN) && + U.compareAndSwapLong + (this, CTL, c, ((long)(w.nextWait & E_MASK) | + (c & (AC_MASK|TC_MASK))))) { + w.eventCount = (e + E_SEQ) & E_MASK; + if ((p = w.parker) != null) + U.unpark(p); + return true; // release an idle worker } - r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5); - return false; // store next seed } - else if (j < 0) { // xorshift - r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5; + else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) { + long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); + if (U.compareAndSwapLong(this, CTL, c, nc)) + return true; // no compensation needed + } + else if (tc + pc < MAX_ID) { + long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); + if (U.compareAndSwapLong(this, CTL, c, nc)) { + addWorker(); + return true; // create replacement + } } - else - ++k; - } - if (scanGuard != g) // staleness check - return false; - else { // try to take submission - ForkJoinTask t; ForkJoinTask[] q; int b, i; - if ((b = queueBase) != queueTop && - (q = submissionQueue) != null && - (i = (q.length - 1) & b) >= 0) { - long u = (i << ASHIFT) + ABASE; - if ((t = q[i]) != null && queueBase == b && - UNSAFE.compareAndSwapObject(q, u, t, null)) { - queueBase = b + 1; - w.execTask(t); - } - return false; - } - return true; // all queues empty } + return false; } + // Submissions + /** - * Tries to enqueue worker w in wait queue and await change in - * worker's eventCount. If the pool is quiescent and there is - * more than one worker, possibly terminates worker upon exit. - * Otherwise, before blocking, rescans queues to avoid missed - * signals. Upon finding work, releases at least one worker - * (which may be the current worker). Rescans restart upon - * detected staleness or failure to release due to - * contention. Note the unusual conventions about Thread.interrupt - * here and elsewhere: Because interrupts are used solely to alert - * threads to check termination, which is checked here anyway, we - * clear status (using Thread.interrupted) before any call to - * park, so that park does not immediately return due to status - * being set via some other unrelated call to interrupt in user - * code. - * - * @param w the calling worker - * @param c the ctl value on entry - * @return true if waited or another thread was released upon enq + * Unless shutting down, adds the given task to some submission + * queue; using a randomly chosen queue index if the caller is a + * ForkJoinWorkerThread, else one based on caller thread's hash + * code. If no queue exists at the index, one is created. If the + * queue is busy, another is chosen by sweeping through the queues + * array. */ - private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) { - int v = w.eventCount; - w.nextWait = (int)c; // w's successor record - long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); - if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { - long d = ctl; // return true if lost to a deq, to force scan - return (int)d != (int)c && (d & AC_MASK) >= (c & AC_MASK); - } - for (int sc = w.stealCount; sc != 0;) { // accumulate stealCount - long s = stealCount; - if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc)) - sc = w.stealCount = 0; - else if (w.eventCount != v) - return true; // update next time - } - if ((!shutdown || !tryTerminate(false)) && - (int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 && - blockedCount == 0 && quiescerCount == 0) - idleAwaitWork(w, nc, c, v); // quiescent - for (boolean rescanned = false;;) { - if (w.eventCount != v) - return true; - if (!rescanned) { - int g = scanGuard, m = g & SMASK; - ForkJoinWorkerThread[] ws = workers; - if (ws != null && m < ws.length) { - rescanned = true; - for (int i = 0; i <= m; ++i) { - ForkJoinWorkerThread u = ws[i]; - if (u != null) { - if (u.queueBase != u.queueTop && - !tryReleaseWaiter()) - rescanned = false; // contended - if (w.eventCount != v) - return true; - } + private void doSubmit(ForkJoinTask task) { + if (task == null) + throw new NullPointerException(); + Thread t = Thread.currentThread(); + int r = ((t instanceof ForkJoinWorkerThread) ? + ((ForkJoinWorkerThread)t).workQueue.nextSeed() : hashThread(t)); + for (;;) { + int rs = runState, m = rs & SMASK; + int j = r &= (m & ~1); // even numbered queues + WorkQueue[] ws = workQueues; + if (rs < 0 || ws == null) + throw new RejectedExecutionException(); // shutting down + if (ws.length > m) { // consistency check + for (WorkQueue q;;) { // circular sweep + if (((q = ws[j]) != null || + (q = tryAddSharedQueue(j)) != null) && + q.trySharedPush(task)) { + signalWork(); + return; + } + if ((j = (j + 2) & m) == r) { + Thread.yield(); // all queues busy + break; } } - if (scanGuard != g || // stale - (queueBase != queueTop && !tryReleaseWaiter())) - rescanned = false; - if (!rescanned) - Thread.yield(); // reduce contention - else - Thread.interrupted(); // clear before park - } - else { - w.parked = true; // must recheck - if (w.eventCount != v) { - w.parked = false; - return true; - } - LockSupport.park(this); - rescanned = w.parked = false; } } } /** - * If inactivating worker w has caused pool to become - * quiescent, check for pool termination, and wait for event - * for up to SHRINK_RATE nanosecs (rescans are unnecessary in - * this case because quiescence reflects consensus about lack - * of work). On timeout, if ctl has not changed, terminate the - * worker. Upon its termination (see deregisterWorker), it may - * wake up another worker to possibly repeat this process. + * Tries to add and register a new queue at the given index. * - * @param w the calling worker - * @param currentCtl the ctl value after enqueuing w - * @param prevCtl the ctl value if w terminated - * @param v the eventCount w awaits change - */ - private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl, - long prevCtl, int v) { - if (w.eventCount == v) { - if (shutdown) - tryTerminate(false); - ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs - while (ctl == currentCtl) { - long startTime = System.nanoTime(); - w.parked = true; - if (w.eventCount == v) // must recheck - LockSupport.parkNanos(this, SHRINK_RATE); - w.parked = false; - if (w.eventCount != v) - break; - else if (System.nanoTime() - startTime < - SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop - Thread.interrupted(); // spurious wakeup - else if (UNSAFE.compareAndSwapLong(this, ctlOffset, - currentCtl, prevCtl)) { - w.terminate = true; // restore previous - w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK; - break; + * @param idx the workQueues array index to register the queue + * @return the queue, or null if could not add because could + * not acquire lock or idx is unusable + */ + private WorkQueue tryAddSharedQueue(int idx) { + WorkQueue q = null; + ReentrantLock lock = this.lock; + if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) { + // create queue outside of lock but only if apparently free + WorkQueue nq = new WorkQueue(null, SHARED_QUEUE); + if (lock.tryLock()) { + try { + WorkQueue[] ws = workQueues; + if (ws != null && idx < ws.length) { + if ((q = ws[idx]) == null) { + int rs; // update runState seq + ws[idx] = q = nq; + runState = (((rs = runState) & SHUTDOWN) | + ((rs + RS_SEQ) & ~SHUTDOWN)); + } + } + } finally { + lock.unlock(); } } } + return q; } - // Submissions + // Scanning for tasks /** - * Enqueues the given task in the submissionQueue. Same idea as - * ForkJoinWorkerThread.pushTask except for use of submissionLock. - * - * @param t the task - */ - private void addSubmission(ForkJoinTask t) { - final ReentrantLock lock = this.submissionLock; - lock.lock(); - try { - ForkJoinTask[] q; int s, m; - if ((q = submissionQueue) != null) { // ignore if queue removed - long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE; - UNSAFE.putOrderedObject(q, u, t); - queueTop = s + 1; - if (s - queueBase == m) - growSubmissionQueue(); + * Scans for and, if found, returns one task, else possibly + * inactivates the worker. This method operates on single reads of + * volatile state and is designed to be re-invoked continuously in + * part because it returns upon detecting inconsistencies, + * contention, or state changes that indicate possible success on + * re-invocation. + * + * The scan searches for tasks across queues, randomly selecting + * the first #queues probes, favoring steals 2:1 over submissions + * (by exploiting even/odd indexing), and then performing a + * circular sweep of all queues. The scan terminates upon either + * finding a non-empty queue, or completing a full sweep. If the + * worker is not inactivated, it takes and returns a task from + * this queue. On failure to find a task, we take one of the + * following actions, after which the caller will retry calling + * this method unless terminated. + * + * * If not a complete sweep, try to release a waiting worker. If + * the scan terminated because the worker is inactivated, then the + * released worker will often be the calling worker, and it can + * succeed obtaining a task on the next call. Or maybe it is + * another worker, but with same net effect. Releasing in other + * cases as well ensures that we have enough workers running. + * + * * If the caller has run a task since the the last empty scan, + * return (to allow rescan) if other workers are not also yet + * enqueued. Field WorkQueue.rescans counts down on each scan to + * ensure eventual inactivation, and occasional calls to + * Thread.yield to help avoid interference with more useful + * activities on the system. + * + * * If pool is terminating, terminate the worker + * + * * If not already enqueued, try to inactivate and enqueue the + * worker on wait queue. + * + * * If already enqueued and none of the above apply, either park + * awaiting signal, or if this is the most recent waiter and pool + * is quiescent, relay to idleAwaitWork to check for termination + * and possibly shrink pool. + * + * @param w the worker (via its WorkQueue) + * @return a task or null of none found + */ + private final ForkJoinTask scan(WorkQueue w) { + boolean swept = false; // true after full empty scan + WorkQueue[] ws; // volatile read order matters + int r = w.seed, ec = w.eventCount; // ec is negative if inactive + int rs = runState, m = rs & SMASK; + if ((ws = workQueues) != null && ws.length > m) { + ForkJoinTask task = null; + for (int k = 0, j = -2 - m; ; ++j) { + WorkQueue q; int b; + if (j < 0) { // random probes while j negative + r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1); + } // worker (not submit) for odd j + else // cyclic scan when j >= 0 + k += (m >>> 1) | 1; // step by half to reduce bias + + if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) { + if (ec >= 0) + task = q.pollAt(b); // steal + break; + } + else if (j > m) { + if (rs == runState) // staleness check + swept = true; + break; + } + } + w.seed = r; // save seed for next scan + if (task != null) + return task; + } + + // Decode ctl on empty scan + long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns; + if (!swept) { // try to release a waiter + WorkQueue v; Thread p; + if (e > 0 && a < 0 && ws != null && + (v = ws[((~e << 1) | 1) & m]) != null && + v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong + (this, CTL, c, ((long)(v.nextWait & E_MASK) | + ((c + AC_UNIT) & (AC_MASK|TC_MASK))))) { + v.eventCount = (e + E_SEQ) & E_MASK; + if ((p = v.parker) != null) + U.unpark(p); + } + } + else if ((nr = w.rescans) > 0) { // continue rescanning + int ac = a + parallelism; + if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 && + w.eventCount == ec) + Thread.yield(); // 1 bit randomness for yield call + } + else if (e < 0) // pool is terminating + w.runState = -1; + else if (ec >= 0) { // try to enqueue + long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); + w.nextWait = e; + w.eventCount = ec | INT_SIGN; // mark as inactive + if (!U.compareAndSwapLong(this, CTL, c, nc)) + w.eventCount = ec; // back out on CAS failure + else if ((ns = w.nsteals) != 0) { // set rescans if ran task + if (a <= 0) // ... unless too many active + w.rescans = a + parallelism; + w.nsteals = 0; + w.totalSteals += ns; + } + } + else{ // already queued + if (parallelism == -a) + idleAwaitWork(w); // quiescent + if (w.eventCount == ec) { + Thread.interrupted(); // clear status + ForkJoinWorkerThread wt = w.owner; + U.putObject(wt, PARKBLOCKER, this); + w.parker = wt; // emulate LockSupport.park + if (w.eventCount == ec) // recheck + U.park(false, 0L); // block + w.parker = null; + U.putObject(wt, PARKBLOCKER, null); } - } finally { - lock.unlock(); } - signalWork(); + return null; } - // (pollSubmission is defined below with exported methods) - /** - * Creates or doubles submissionQueue array. - * Basically identical to ForkJoinWorkerThread version. + * If inactivating worker w has caused pool to become quiescent, + * check for pool termination, and, so long as this is not the + * only worker, wait for event for up to SHRINK_RATE nanosecs On + * timeout, if ctl has not changed, terminate the worker, which + * will in turn wake up another worker to possibly repeat this + * process. + * + * @param w the calling worker */ - private void growSubmissionQueue() { - ForkJoinTask[] oldQ = submissionQueue; - int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY; - if (size > MAXIMUM_QUEUE_CAPACITY) - throw new RejectedExecutionException("Queue capacity exceeded"); - if (size < INITIAL_QUEUE_CAPACITY) - size = INITIAL_QUEUE_CAPACITY; - ForkJoinTask[] q = submissionQueue = new ForkJoinTask[size]; - int mask = size - 1; - int top = queueTop; - int oldMask; - if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) { - for (int b = queueBase; b != top; ++b) { - long u = ((b & oldMask) << ASHIFT) + ABASE; - Object x = UNSAFE.getObjectVolatile(oldQ, u); - if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null)) - UNSAFE.putObjectVolatile - (q, ((b & mask) << ASHIFT) + ABASE, x); + private void idleAwaitWork(WorkQueue w) { + long c; int nw, ec; + if (!tryTerminate(false) && + (int)((c = ctl) >> AC_SHIFT) + parallelism == 0 && + (ec = w.eventCount) == ((int)c | INT_SIGN) && + (nw = w.nextWait) != 0) { + long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout + ((c + AC_UNIT) & AC_MASK) | (c & TC_MASK)); + ForkJoinTask.helpExpungeStaleExceptions(); // help clean + ForkJoinWorkerThread wt = w.owner; + while (ctl == c) { + long startTime = System.nanoTime(); + Thread.interrupted(); // timed variant of version in scan() + U.putObject(wt, PARKBLOCKER, this); + w.parker = wt; + if (ctl == c) + U.park(false, SHRINK_RATE); + w.parker = null; + U.putObject(wt, PARKBLOCKER, null); + if (ctl != c) + break; + if (System.nanoTime() - startTime >= SHRINK_TIMEOUT && + U.compareAndSwapLong(this, CTL, c, nc)) { + w.runState = -1; // shrink + w.eventCount = (ec + E_SEQ) | E_MASK; + break; + } } } } - // Blocking support - /** - * Tries to increment blockedCount, decrement active count - * (sometimes implicitly) and possibly release or create a - * compensating worker in preparation for blocking. Fails - * on contention or termination. - * - * @return true if the caller can block, else should recheck and retry - */ - private boolean tryPreBlock() { - int b = blockedCount; - if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) { - int pc = parallelism; - do { - ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w; - int e, ac, tc, i; - long c = ctl; - int u = (int)(c >>> 32); - if ((e = (int)c) < 0) { - // skip -- terminating - } - else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 && - (ws = workers) != null && - (i = ~e & SMASK) < ws.length && - (w = ws[i]) != null) { - long nc = ((long)(w.nextWait & E_MASK) | - (c & (AC_MASK|TC_MASK))); - if (w.eventCount == e && - UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { - w.eventCount = (e + EC_UNIT) & E_MASK; - if (w.parked) - UNSAFE.unpark(w); - return true; // release an idle worker + * Tries to locate and execute tasks for a stealer of the given + * task, or in turn one of its stealers, Traces currentSteal -> + * currentJoin links looking for a thread working on a descendant + * of the given task and with a non-empty queue to steal back and + * execute tasks from. The first call to this method upon a + * waiting join will often entail scanning/search, (which is OK + * because the joiner has nothing better to do), but this method + * leaves hints in workers to speed up subsequent calls. The + * implementation is very branchy to cope with potential + * inconsistencies or loops encountering chains that are stale, + * unknown, or of length greater than MAX_HELP_DEPTH links. All + * of these cases are dealt with by just retrying by caller. + * + * @param joiner the joining worker + * @param task the task to join + * @return true if found or ran a task (and so is immediately retryable) + */ + final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask task) { + ForkJoinTask subtask; // current target + boolean progress = false; + int depth = 0; // current chain depth + int m = runState & SMASK; + WorkQueue[] ws = workQueues; + + if (ws != null && ws.length > m && (subtask = task).status >= 0) { + outer:for (WorkQueue j = joiner;;) { + // Try to find the stealer of subtask, by first using hint + WorkQueue stealer = null; + WorkQueue v = ws[j.stealHint & m]; + if (v != null && v.currentSteal == subtask) + stealer = v; + else { + for (int i = 1; i <= m; i += 2) { + if ((v = ws[i]) != null && v.currentSteal == subtask) { + stealer = v; + j.stealHint = i; // save hint + break; + } } + if (stealer == null) + break; } - else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) { - long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); - if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) - return true; // no compensation needed - } - else if (tc + pc < MAX_ID) { - long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); - if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) { - addWorker(); - return true; // create a replacement + + for (WorkQueue q = stealer;;) { // Try to help stealer + ForkJoinTask t; int b; + if (task.status < 0) + break outer; + if ((b = q.base) - q.top < 0) { + progress = true; + if (subtask.status < 0) + break outer; // stale + if ((t = q.pollAt(b)) != null) { + stealer.stealHint = joiner.poolIndex; + joiner.runSubtask(t); + } + } + else { // empty - try to descend to find stealer's stealer + ForkJoinTask next = stealer.currentJoin; + if (++depth == MAX_HELP_DEPTH || subtask.status < 0 || + next == null || next == subtask) + break outer; // max depth, stale, dead-end, cyclic + subtask = next; + j = stealer; + break; } } - // try to back out on any failure and let caller retry - } while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset, - b = blockedCount, b - 1)); + } } - return false; + return progress; } /** - * Decrements blockedCount and increments active count. - */ - private void postBlock() { - long c; - do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, // no mask - c = ctl, c + AC_UNIT)); - int b; - do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset, - b = blockedCount, b - 1)); - } - - /** - * Possibly blocks waiting for the given task to complete, or - * cancels the task if terminating. Fails to wait if contended. + * If task is at base of some steal queue, steals and executes it. * - * @param joinMe the task + * @param joiner the joining worker + * @param task the task */ - final void tryAwaitJoin(ForkJoinTask joinMe) { - Thread.interrupted(); // clear interrupts before checking termination - if (joinMe.status >= 0) { - if (tryPreBlock()) { - joinMe.tryAwaitDone(0L); - postBlock(); + final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask task) { + WorkQueue[] ws; + int m = runState & SMASK; + if ((ws = workQueues) != null && ws.length > m) { + for (int j = 1; j <= m && task.status >= 0; j += 2) { + WorkQueue q = ws[j]; + if (q != null && q.pollFor(task)) { + joiner.runSubtask(task); + break; + } } - else if ((ctl & STOP_BIT) != 0L) - joinMe.cancelIgnoringExceptions(); } } /** - * Possibly blocks the given worker waiting for joinMe to - * complete or timeout. - * - * @param joinMe the task - * @param millis the wait time for underlying Object.wait - */ - final void timedAwaitJoin(ForkJoinTask joinMe, long nanos) { - while (joinMe.status >= 0) { - Thread.interrupted(); - if ((ctl & STOP_BIT) != 0L) { - joinMe.cancelIgnoringExceptions(); - break; - } - if (tryPreBlock()) { - long last = System.nanoTime(); - while (joinMe.status >= 0) { - long millis = TimeUnit.NANOSECONDS.toMillis(nanos); - if (millis <= 0) - break; - joinMe.tryAwaitDone(millis); - if (joinMe.status < 0) - break; - if ((ctl & STOP_BIT) != 0L) { - joinMe.cancelIgnoringExceptions(); - break; + * Returns a non-empty steal queue, if is found during a random, + * then cyclic scan, else null. This method must be retried by + * caller if, by the time it tries to use the queue, it is empty. + */ + private WorkQueue findNonEmptyStealQueue(WorkQueue w) { + int r = w.seed; // Same idea as scan(), but ignoring submissions + for (WorkQueue[] ws;;) { + int m = runState & SMASK; + if ((ws = workQueues) == null) + return null; + if (ws.length > m) { + WorkQueue q; + for (int n = m << 2, k = r, j = -n;;) { + r ^= r << 13; r ^= r >>> 17; r ^= r << 5; + if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) { + w.seed = r; + return q; } - long now = System.nanoTime(); - nanos -= now - last; - last = now; + else if (j > n) + return null; + else + k = (j++ < 0) ? r : k + ((m >>> 1) | 1); + } - postBlock(); - break; } } } /** - * If necessary, compensates for blocker, and blocks. - */ - private void awaitBlocker(ManagedBlocker blocker) - throws InterruptedException { - while (!blocker.isReleasable()) { - if (tryPreBlock()) { - try { - do {} while (!blocker.isReleasable() && !blocker.block()); - } finally { - postBlock(); + * Runs tasks until {@code isQuiescent()}. We piggyback on + * active count ctl maintenance, but rather than blocking + * when tasks cannot be found, we rescan until all others cannot + * find tasks either. + */ + final void helpQuiescePool(WorkQueue w) { + for (boolean active = true;;) { + w.runLocalTasks(); // exhaust local queue + WorkQueue q = findNonEmptyStealQueue(w); + if (q != null) { + ForkJoinTask t; + if (!active) { // re-establish active count + long c; + active = true; + do {} while (!U.compareAndSwapLong + (this, CTL, c = ctl, c + AC_UNIT)); + } + if ((t = q.poll()) != null) + w.runSubtask(t); + } + else { + long c; + if (active) { // decrement active count without queuing + active = false; + do {} while (!U.compareAndSwapLong + (this, CTL, c = ctl, c -= AC_UNIT)); + } + else + c = ctl; // re-increment on exit + if ((int)(c >> AC_SHIFT) + parallelism == 0) { + do {} while (!U.compareAndSwapLong + (this, CTL, c = ctl, c + AC_UNIT)); + break; } - break; } } } - // Creating, registering and deregistering workers - /** - * Tries to create and start a worker; minimally rolls back counts - * on failure. + * Gets and removes a local or stolen task for the given worker + * + * @return a task, if available */ - private void addWorker() { - Throwable ex = null; - ForkJoinWorkerThread t = null; - try { - t = factory.newThread(this); - } catch (Throwable e) { - ex = e; - } - if (t == null) { // null or exceptional factory return - long c; // adjust counts - do {} while (!UNSAFE.compareAndSwapLong - (this, ctlOffset, c = ctl, - (((c - AC_UNIT) & AC_MASK) | - ((c - TC_UNIT) & TC_MASK) | - (c & ~(AC_MASK|TC_MASK))))); - // Propagate exception if originating from an external caller - if (!tryTerminate(false) && ex != null && - !(Thread.currentThread() instanceof ForkJoinWorkerThread)) - UNSAFE.throwException(ex); + final ForkJoinTask nextTaskFor(WorkQueue w) { + for (ForkJoinTask t;;) { + WorkQueue q; + if ((t = w.nextLocalTask()) != null) + return t; + if ((q = findNonEmptyStealQueue(w)) == null) + return null; + if ((t = q.poll()) != null) + return t; } - else - t.start(); } /** - * Callback from ForkJoinWorkerThread constructor to assign a - * public name + * Returns the approximate (non-atomic) number of idle threads per + * active thread to offset steal queue size for method + * ForkJoinTask.getSurplusQueuedTaskCount(). */ - final String nextWorkerName() { - for (int n;;) { - if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset, - n = nextWorkerNumber, ++n)) - return workerNamePrefix + n; - } + final int idlePerActive() { + // Approximate at powers of two for small values, saturate past 4 + int p = parallelism; + int a = p + (int)(ctl >> AC_SHIFT); + return (a > (p >>>= 1) ? 0 : + a > (p >>>= 1) ? 1 : + a > (p >>>= 1) ? 2 : + a > (p >>>= 1) ? 4 : + 8); } - /** - * Callback from ForkJoinWorkerThread constructor to - * determine its poolIndex and record in workers array. - * - * @param w the worker - * @return the worker's pool index - */ - final int registerWorker(ForkJoinWorkerThread w) { - /* - * In the typical case, a new worker acquires the lock, uses - * next available index and returns quickly. Since we should - * not block callers (ultimately from signalWork or - * tryPreBlock) waiting for the lock needed to do this, we - * instead help release other workers while waiting for the - * lock. - */ - for (int g;;) { - ForkJoinWorkerThread[] ws; - if (((g = scanGuard) & SG_UNIT) == 0 && - UNSAFE.compareAndSwapInt(this, scanGuardOffset, - g, g | SG_UNIT)) { - int k = nextWorkerIndex; - try { - if ((ws = workers) != null) { // ignore on shutdown - int n = ws.length; - if (k < 0 || k >= n || ws[k] != null) { - for (k = 0; k < n && ws[k] != null; ++k) - ; - if (k == n) - ws = workers = Arrays.copyOf(ws, n << 1); - } - ws[k] = w; - nextWorkerIndex = k + 1; - int m = g & SMASK; - g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1); - } - } finally { - scanGuard = g; - } - return k; - } - else if ((ws = workers) != null) { // help release others - for (ForkJoinWorkerThread u : ws) { - if (u != null && u.queueBase != u.queueTop) { - if (tryReleaseWaiter()) - break; - } - } - } - } - } + // Termination /** - * Final callback from terminating worker. Removes record of - * worker from array, and adjusts counts. If pool is shutting - * down, tries to complete termination. - * - * @param w the worker + * Sets SHUTDOWN bit of runState under lock */ - final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) { - int idx = w.poolIndex; - int sc = w.stealCount; - int steps = 0; - // Remove from array, adjust worker counts and collect steal count. - // We can intermix failed removes or adjusts with steal updates - do { - long s, c; - int g; - if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 && - UNSAFE.compareAndSwapInt(this, scanGuardOffset, - g, g |= SG_UNIT)) { - ForkJoinWorkerThread[] ws = workers; - if (ws != null && idx >= 0 && - idx < ws.length && ws[idx] == w) - ws[idx] = null; // verify - nextWorkerIndex = idx; - scanGuard = g + SG_UNIT; - steps = 1; - } - if (steps == 1 && - UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl, - (((c - AC_UNIT) & AC_MASK) | - ((c - TC_UNIT) & TC_MASK) | - (c & ~(AC_MASK|TC_MASK))))) - steps = 2; - if (sc != 0 && - UNSAFE.compareAndSwapLong(this, stealCountOffset, - s = stealCount, s + sc)) - sc = 0; - } while (steps != 2 || sc != 0); - if (!tryTerminate(false)) { - if (ex != null) // possibly replace if died abnormally - signalWork(); - else - tryReleaseWaiter(); + private void enableShutdown() { + ReentrantLock lock = this.lock; + if (runState >= 0) { + lock.lock(); // don't need try/finally + runState |= SHUTDOWN; + lock.unlock(); } } - // Shutdown and termination - /** - * Possibly initiates and/or completes termination. + * Possibly initiates and/or completes termination. Upon + * termination, cancels all queued tasks and then * * @param now if true, unconditionally terminate, else only - * if shutdown and empty queue and no active workers + * if no work and no active workers * @return true if now terminating or terminated */ private boolean tryTerminate(boolean now) { - long c; - while (((c = ctl) & STOP_BIT) == 0) { + for (long c;;) { + if (((c = ctl) & STOP_BIT) != 0) { // already terminating + if ((short)(c >>> TC_SHIFT) == -parallelism) { + ReentrantLock lock = this.lock; // signal when no workers + lock.lock(); // don't need try/finally + termination.signalAll(); // signal when 0 workers + lock.unlock(); + } + return true; + } if (!now) { - if ((int)(c >> AC_SHIFT) != -parallelism) + if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 || + hasQueuedSubmissions()) return false; - if (!shutdown || blockedCount != 0 || quiescerCount != 0 || - queueBase != queueTop) { - if (ctl == c) // staleness check - return false; - continue; + // Check for unqueued inactive workers. One pass suffices. + WorkQueue[] ws = workQueues; WorkQueue w; + if (ws != null) { + int n = ws.length; + for (int i = 1; i < n; i += 2) { + if ((w = ws[i]) != null && w.eventCount >= 0) + return false; + } } } - if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT)) + if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) startTerminating(); } - if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers - final ReentrantLock lock = this.submissionLock; - lock.lock(); - try { - termination.signalAll(); - } finally { - lock.unlock(); - } - } - return true; } /** - * Runs up to three passes through workers: (0) Setting - * termination status for each worker, followed by wakeups up to - * queued workers; (1) helping cancel tasks; (2) interrupting - * lagging threads (likely in external tasks, but possibly also - * blocked in joins). Each pass repeats previous steps because of - * potential lagging thread creation. + * Initiates termination: Runs three passes through workQueues: + * (0) Setting termination status, followed by wakeups of queued + * workers; (1) cancelling all tasks; (2) interrupting lagging + * threads (likely in external tasks, but possibly also blocked in + * joins). Each pass repeats previous steps because of potential + * lagging thread creation. */ private void startTerminating() { - cancelSubmissions(); for (int pass = 0; pass < 3; ++pass) { - ForkJoinWorkerThread[] ws = workers; + WorkQueue[] ws = workQueues; if (ws != null) { - for (ForkJoinWorkerThread w : ws) { - if (w != null) { - w.terminate = true; + WorkQueue w; Thread wt; + int n = ws.length; + for (int j = 0; j < n; ++j) { + if ((w = ws[j]) != null) { + w.runState = -1; if (pass > 0) { - w.cancelTasks(); - if (pass > 1 && !w.isInterrupted()) { + w.cancelAll(); + if (pass > 1 && (wt = w.owner) != null && + !wt.isInterrupted()) { try { - w.interrupt(); + wt.interrupt(); } catch (SecurityException ignore) { } } } } } - terminateWaiters(); - } - } - } - - /** - * Polls and cancels all submissions. Called only during termination. - */ - private void cancelSubmissions() { - while (queueBase != queueTop) { - ForkJoinTask task = pollSubmission(); - if (task != null) { - try { - task.cancel(false); - } catch (Throwable ignore) { - } - } - } - } - - /** - * Tries to set the termination status of waiting workers, and - * then wakes them up (after which they will terminate). - */ - private void terminateWaiters() { - ForkJoinWorkerThread[] ws = workers; - if (ws != null) { - ForkJoinWorkerThread w; long c; int i, e; - int n = ws.length; - while ((i = ~(e = (int)(c = ctl)) & SMASK) < n && - (w = ws[i]) != null && w.eventCount == (e & E_MASK)) { - if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, - (long)(w.nextWait & E_MASK) | - ((c + AC_UNIT) & AC_MASK) | - (c & (TC_MASK|STOP_BIT)))) { - w.terminate = true; - w.eventCount = e + EC_UNIT; - if (w.parked) - UNSAFE.unpark(w); + // Wake up workers parked on event queue + int i, e; long c; Thread p; + while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n && + (w = ws[i]) != null && + w.eventCount == (e | INT_SIGN)) { + long nc = ((long)(w.nextWait & E_MASK) | + ((c + AC_UNIT) & AC_MASK) | + (c & (TC_MASK|STOP_BIT))); + if (U.compareAndSwapLong(this, CTL, c, nc)) { + w.eventCount = (e + E_SEQ) & E_MASK; + if ((p = w.parker) != null) + U.unpark(p); + } } } } } - // misc ForkJoinWorkerThread support - - /** - * Increments or decrements quiescerCount. Needed only to prevent - * triggering shutdown if a worker is transiently inactive while - * checking quiescence. - * - * @param delta 1 for increment, -1 for decrement - */ - final void addQuiescerCount(int delta) { - int c; - do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset, - c = quiescerCount, c + delta)); - } - - /** - * Directly increments or decrements active count without queuing. - * This method is used to transiently assert inactivation while - * checking quiescence. - * - * @param delta 1 for increment, -1 for decrement - */ - final void addActiveCount(int delta) { - long d = (long)delta << AC_SHIFT; - long c; - do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, - c = ctl, c + d)); - } - - /** - * Returns the approximate (non-atomic) number of idle threads per - * active thread. - */ - final int idlePerActive() { - // Approximate at powers of two for small values, saturate past 4 - int p = parallelism; - int a = p + (int)(ctl >> AC_SHIFT); - return (a > (p >>>= 1) ? 0 : - a > (p >>>= 1) ? 1 : - a > (p >>>= 1) ? 2 : - a > (p >>>= 1) ? 4 : - 8); - } - // Exported methods // Constructors @@ -1436,24 +1924,30 @@ public class ForkJoinPool extends Abstra this.parallelism = parallelism; this.factory = factory; this.ueh = handler; - this.locallyFifo = asyncMode; + this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE; + this.nextPoolIndex = 1; long np = (long)(-parallelism); // offset ctl counts this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); - this.submissionQueue = new ForkJoinTask[INITIAL_QUEUE_CAPACITY]; - // initialize workers array with room for 2*parallelism if possible + // initialize workQueues array with room for 2*parallelism if possible int n = parallelism << 1; if (n >= MAX_ID) n = MAX_ID; else { // See Hackers Delight, sec 3.2, where n < (1 << 16) n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; } - workers = new ForkJoinWorkerThread[n + 1]; - this.submissionLock = new ReentrantLock(); - this.termination = submissionLock.newCondition(); + this.workQueues = new WorkQueue[(n + 1) << 1]; + ReentrantLock lck = this.lock = new ReentrantLock(); + this.termination = lck.newCondition(); + this.stealCount = new AtomicLong(); + this.nextWorkerNumber = new AtomicInteger(); StringBuilder sb = new StringBuilder("ForkJoinPool-"); sb.append(poolNumberGenerator.incrementAndGet()); sb.append("-worker-"); this.workerNamePrefix = sb.toString(); + // Create initial submission queue + WorkQueue sq = tryAddSharedQueue(0); + if (sq != null) + sq.growArray(false); } // Execution methods @@ -1475,34 +1969,8 @@ public class ForkJoinPool extends Abstra * scheduled for execution */ public T invoke(ForkJoinTask task) { - Thread t = Thread.currentThread(); - if (task == null) - throw new NullPointerException(); - if (shutdown) - throw new RejectedExecutionException(); - if ((t instanceof ForkJoinWorkerThread) && - ((ForkJoinWorkerThread)t).pool == this) - return task.invoke(); // bypass submit if in same pool - else { - addSubmission(task); - return task.join(); - } - } - - /** - * Unless terminating, forks task if within an ongoing FJ - * computation in the current pool, else submits as external task. - */ - private void forkOrSubmit(ForkJoinTask task) { - ForkJoinWorkerThread w; - Thread t = Thread.currentThread(); - if (shutdown) - throw new RejectedExecutionException(); - if ((t instanceof ForkJoinWorkerThread) && - (w = (ForkJoinWorkerThread)t).pool == this) - w.pushTask(task); - else - addSubmission(task); + doSubmit(task); + return task.join(); } /** @@ -1514,9 +1982,7 @@ public class ForkJoinPool extends Abstra * scheduled for execution */ public void execute(ForkJoinTask task) { - if (task == null) - throw new NullPointerException(); - forkOrSubmit(task); + doSubmit(task); } // AbstractExecutorService methods @@ -1534,7 +2000,7 @@ public class ForkJoinPool extends Abstra job = (ForkJoinTask) task; else job = ForkJoinTask.adapt(task, null); - forkOrSubmit(job); + doSubmit(job); } /** @@ -1547,9 +2013,7 @@ public class ForkJoinPool extends Abstra * scheduled for execution */ public ForkJoinTask submit(ForkJoinTask task) { - if (task == null) - throw new NullPointerException(); - forkOrSubmit(task); + doSubmit(task); return task; } @@ -1562,7 +2026,7 @@ public class ForkJoinPool extends Abstra if (task == null) throw new NullPointerException(); ForkJoinTask job = ForkJoinTask.adapt(task); - forkOrSubmit(job); + doSubmit(job); return job; } @@ -1575,7 +2039,7 @@ public class ForkJoinPool extends Abstra if (task == null) throw new NullPointerException(); ForkJoinTask job = ForkJoinTask.adapt(task, result); - forkOrSubmit(job); + doSubmit(job); return job; } @@ -1592,7 +2056,7 @@ public class ForkJoinPool extends Abstra job = (ForkJoinTask) task; else job = ForkJoinTask.adapt(task, null); - forkOrSubmit(job); + doSubmit(job); return job; } @@ -1669,7 +2133,7 @@ public class ForkJoinPool extends Abstra * @return {@code true} if this pool uses async mode */ public boolean getAsyncMode() { - return locallyFifo; + return localMode != 0; } /** @@ -1681,8 +2145,21 @@ public class ForkJoinPool extends Abstra * @return the number of worker threads */ public int getRunningThreadCount() { - int r = parallelism + (int)(ctl >> AC_SHIFT); - return (r <= 0) ? 0 : r; // suppress momentarily negative values + int rc = 0; + WorkQueue[] ws; WorkQueue w; + if ((ws = workQueues) != null) { + int n = ws.length; + for (int i = 1; i < n; i += 2) { + Thread.State s; ForkJoinWorkerThread wt; + if ((w = ws[i]) != null && (wt = w.owner) != null && + w.eventCount >= 0 && + (s = wt.getState()) != Thread.State.BLOCKED && + s != Thread.State.WAITING && + s != Thread.State.TIMED_WAITING) + ++rc; + } + } + return rc; } /** @@ -1693,7 +2170,7 @@ public class ForkJoinPool extends Abstra * @return the number of active threads */ public int getActiveThreadCount() { - int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount; + int r = parallelism + (int)(ctl >> AC_SHIFT); return (r <= 0) ? 0 : r; // suppress momentarily negative values } @@ -1709,7 +2186,7 @@ public class ForkJoinPool extends Abstra * @return {@code true} if all threads are currently idle */ public boolean isQuiescent() { - return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0; + return (int)(ctl >> AC_SHIFT) + parallelism == 0; } /** @@ -1724,7 +2201,16 @@ public class ForkJoinPool extends Abstra * @return the number of steals */ public long getStealCount() { - return stealCount; + long count = stealCount.get(); + WorkQueue[] ws; WorkQueue w; + if ((ws = workQueues) != null) { + int n = ws.length; + for (int i = 1; i < n; i += 2) { + if ((w = ws[i]) != null) + count += w.totalSteals; + } + } + return count; } /** @@ -1739,12 +2225,13 @@ public class ForkJoinPool extends Abstra */ public long getQueuedTaskCount() { long count = 0; - ForkJoinWorkerThread[] ws; - if ((short)(ctl >>> TC_SHIFT) > -parallelism && - (ws = workers) != null) { - for (ForkJoinWorkerThread w : ws) - if (w != null) - count -= w.queueBase - w.queueTop; // must read base first + WorkQueue[] ws; WorkQueue w; + if ((ws = workQueues) != null) { + int n = ws.length; + for (int i = 1; i < n; i += 2) { + if ((w = ws[i]) != null) + count += w.queueSize(); + } } return count; } @@ -1757,7 +2244,16 @@ public class ForkJoinPool extends Abstra * @return the number of queued submissions */ public int getQueuedSubmissionCount() { - return -queueBase + queueTop; + int count = 0; + WorkQueue[] ws; WorkQueue w; + if ((ws = workQueues) != null) { + int n = ws.length; + for (int i = 0; i < n; i += 2) { + if ((w = ws[i]) != null) + count += w.queueSize(); + } + } + return count; } /** @@ -1767,7 +2263,15 @@ public class ForkJoinPool extends Abstra * @return {@code true} if there are any queued submissions */ public boolean hasQueuedSubmissions() { - return queueBase != queueTop; + WorkQueue[] ws; WorkQueue w; + if ((ws = workQueues) != null) { + int n = ws.length; + for (int i = 0; i < n; i += 2) { + if ((w = ws[i]) != null && w.queueSize() != 0) + return true; + } + } + return false; } /** @@ -1778,16 +2282,12 @@ public class ForkJoinPool extends Abstra * @return the next submission, or {@code null} if none */ protected ForkJoinTask pollSubmission() { - ForkJoinTask t; ForkJoinTask[] q; int b, i; - while ((b = queueBase) != queueTop && - (q = submissionQueue) != null && - (i = (q.length - 1) & b) >= 0) { - long u = (i << ASHIFT) + ABASE; - if ((t = q[i]) != null && - queueBase == b && - UNSAFE.compareAndSwapObject(q, u, t, null)) { - queueBase = b + 1; - return t; + WorkQueue[] ws; WorkQueue w; ForkJoinTask t; + if ((ws = workQueues) != null) { + int n = ws.length; + for (int i = 0; i < n; i += 2) { + if ((w = ws[i]) != null && (t = w.poll()) != null) + return t; } } return null; @@ -1812,20 +2312,18 @@ public class ForkJoinPool extends Abstra */ protected int drainTasksTo(Collection> c) { int count = 0; - while (queueBase != queueTop) { - ForkJoinTask t = pollSubmission(); - if (t != null) { - c.add(t); - ++count; + WorkQueue[] ws; WorkQueue w; ForkJoinTask t; + if ((ws = workQueues) != null) { + int n = ws.length; + for (int i = 0; i < n; ++i) { + if ((w = ws[i]) != null) { + while ((t = w.poll()) != null) { + c.add(t); + ++count; + } + } } } - ForkJoinWorkerThread[] ws; - if ((short)(ctl >>> TC_SHIFT) > -parallelism && - (ws = workers) != null) { - for (ForkJoinWorkerThread w : ws) - if (w != null) - count += w.drainTasksTo(c); - } return count; } @@ -1840,18 +2338,18 @@ public class ForkJoinPool extends Abstra long st = getStealCount(); long qt = getQueuedTaskCount(); long qs = getQueuedSubmissionCount(); + int rc = getRunningThreadCount(); int pc = parallelism; long c = ctl; int tc = pc + (short)(c >>> TC_SHIFT); - int rc = pc + (int)(c >> AC_SHIFT); - if (rc < 0) // ignore transient negative - rc = 0; - int ac = rc + blockedCount; + int ac = pc + (int)(c >> AC_SHIFT); + if (ac < 0) // ignore transient negative + ac = 0; String level; if ((c & STOP_BIT) != 0) level = (tc == 0) ? "Terminated" : "Terminating"; else - level = shutdown ? "Shutting down" : "Running"; + level = runState < 0 ? "Shutting down" : "Running"; return super.toString() + "[" + level + ", parallelism = " + pc + @@ -1878,7 +2376,7 @@ public class ForkJoinPool extends Abstra */ public void shutdown() { checkPermission(); - shutdown = true; + enableShutdown(); tryTerminate(false); } @@ -1900,7 +2398,7 @@ public class ForkJoinPool extends Abstra */ public List shutdownNow() { checkPermission(); - shutdown = true; + enableShutdown(); tryTerminate(true); return Collections.emptyList(); } @@ -1936,19 +2434,12 @@ public class ForkJoinPool extends Abstra } /** - * Returns true if terminating or terminated. Used by ForkJoinWorkerThread. - */ - final boolean isAtLeastTerminating() { - return (ctl & STOP_BIT) != 0L; - } - - /** * Returns {@code true} if this pool has been shut down. * * @return {@code true} if this pool has been shut down */ public boolean isShutdown() { - return shutdown; + return runState < 0; } /** @@ -1965,7 +2456,7 @@ public class ForkJoinPool extends Abstra public boolean awaitTermination(long timeout, TimeUnit unit) throws InterruptedException { long nanos = unit.toNanos(timeout); - final ReentrantLock lock = this.submissionLock; + final ReentrantLock lock = this.lock; lock.lock(); try { for (;;) { @@ -2061,7 +2552,7 @@ public class ForkJoinPool extends Abstra * *

If the caller is not a {@link ForkJoinTask}, this method is * behaviorally equivalent to - * 

 {@code
+a     *  
 {@code
      * while (!blocker.isReleasable())
      *   if (blocker.block())
      *     return;
@@ -2076,12 +2567,18 @@ public class ForkJoinPool extends Abstra
     public static void managedBlock(ManagedBlocker blocker)
         throws InterruptedException {
         Thread t = Thread.currentThread();
-        if (t instanceof ForkJoinWorkerThread) {
-            ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
-            w.pool.awaitBlocker(blocker);
-        }
-        else {
-            do {} while (!blocker.isReleasable() && !blocker.block());
+        ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
+                          ((ForkJoinWorkerThread)t).pool : null);
+        while (!blocker.isReleasable()) {
+            if (p == null || p.tryCompensate()) {
+                try {
+                    do {} while (!blocker.isReleasable() && !blocker.block());
+                } finally {
+                    if (p != null)
+                        p.incrementActiveCount();
+                }
+                break;
+            }
         }
     }
 
@@ -2098,46 +2595,30 @@ public class ForkJoinPool extends Abstra
     }
 
     // Unsafe mechanics
-    private static final sun.misc.Unsafe UNSAFE;
-    private static final long ctlOffset;
-    private static final long stealCountOffset;
-    private static final long blockedCountOffset;
-    private static final long quiescerCountOffset;
-    private static final long scanGuardOffset;
-    private static final long nextWorkerNumberOffset;
-    private static final long ABASE;
-    private static final int ASHIFT;
+    private static final sun.misc.Unsafe U;
+    private static final long CTL;
+    private static final long RUNSTATE;
+    private static final long PARKBLOCKER;
 
     static {
         poolNumberGenerator = new AtomicInteger();
-        workerSeedGenerator = new Random();
         modifyThreadPermission = new RuntimePermission("modifyThread");
         defaultForkJoinWorkerThreadFactory =
             new DefaultForkJoinWorkerThreadFactory();
+        int s;
         try {
-            UNSAFE = getUnsafe();
+            U = getUnsafe();
             Class k = ForkJoinPool.class;
-            ctlOffset = UNSAFE.objectFieldOffset
+            Class tk = Thread.class;
+            CTL = U.objectFieldOffset
                 (k.getDeclaredField("ctl"));
-            stealCountOffset = UNSAFE.objectFieldOffset
-                (k.getDeclaredField("stealCount"));
-            blockedCountOffset = UNSAFE.objectFieldOffset
-                (k.getDeclaredField("blockedCount"));
-            quiescerCountOffset = UNSAFE.objectFieldOffset
-                (k.getDeclaredField("quiescerCount"));
-            scanGuardOffset = UNSAFE.objectFieldOffset
-                (k.getDeclaredField("scanGuard"));
-            nextWorkerNumberOffset = UNSAFE.objectFieldOffset
-                (k.getDeclaredField("nextWorkerNumber"));
+            RUNSTATE = U.objectFieldOffset
+                (k.getDeclaredField("runState"));
+            PARKBLOCKER = U.objectFieldOffset
+                (tk.getDeclaredField("parkBlocker"));
         } catch (Exception e) {
             throw new Error(e);
         }
-        Class a = ForkJoinTask[].class;
-        ABASE = UNSAFE.arrayBaseOffset(a);
-        int s = UNSAFE.arrayIndexScale(a);
-        if ((s & (s-1)) != 0)
-            throw new Error("data type scale not a power of two");
-        ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
     }
 
     /**