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 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 * processors. The pool attempts to maintain enough active (or * available) threads by dynamically adding, suspending, or resuming * internal worker threads, even if some tasks are stalled waiting to * join others. However, no such adjustments are guaranteed in the * face of blocked IO or other unmanaged synchronization. The nested * {@link ManagedBlocker} interface enables extension of the kinds of * synchronization accommodated. * *
In addition to execution and lifecycle control methods, this * class provides status check methods (for example * {@link #getStealCount}) that are intended to aid in developing, * tuning, and monitoring fork/join applications. Also, method * {@link #toString} returns indications of pool state in a * convenient form for informal monitoring. * *
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 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. * *
| * | Call from non-fork/join clients | *Call from within fork/join computations | *
| Arrange async execution | *{@link #execute(ForkJoinTask)} | *{@link ForkJoinTask#fork} | *
| Await and obtain result | *{@link #invoke(ForkJoinTask)} | *{@link ForkJoinTask#invoke} | *
| Arrange exec and obtain Future | *{@link #submit(ForkJoinTask)} | *{@link ForkJoinTask#fork} (ForkJoinTasks are Futures) | *
Sample Usage. Normally a single {@code ForkJoinPool} is * used for all parallel task execution in a program or subsystem. * Otherwise, use would not usually outweigh the construction and * bookkeeping overhead of creating a large set of threads. For * example, a common pool could be used for the {@code SortTasks} * illustrated in {@link RecursiveAction}. Because {@code * ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon * daemon} mode, there is typically no need to explicitly {@link * #shutdown} such a pool upon program exit. * *
{@code
* static final ForkJoinPool mainPool = new ForkJoinPool();
* ...
* public void sort(long[] array) {
* mainPool.invoke(new SortTask(array, 0, array.length));
* }}
*
* Implementation notes: This implementation restricts the * maximum number of running threads to 32767. Attempts to create * pools with greater than the maximum number result in * {@code IllegalArgumentException}. * *
This implementation rejects submitted tasks (that is, by throwing
* {@link RejectedExecutionException}) only when the pool is shut down
* or internal resources have been exhausted.
*
* @since 1.7
* @author Doug Lea
*/
public class ForkJoinPool extends AbstractExecutorService {
/*
* Implementation Overview
*
* 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 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. 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.
*
* 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. 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 (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 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;
* 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
* time out and terminate if the pool has remained quiescent for
* SHRINK_RATE nanosecs. This will slowly propagate, eventually
* terminating all workers after long periods of non-use.
*
* 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.
*
* Compensating: Unless there are already enough live threads,
* 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, 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. 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
* inline assignments (of form "while ((local = field) != 0)")
* which are usually the simplest way to ensure the required read
* orderings (which are sometimes critical). This leads to a
* "C"-like style of listing declarations of these locals at the
* heads of methods or blocks. There are several occurrences of
* the unusual "do {} while (!cas...)" which is the simplest way
* to force an update of a CAS'ed variable. There are also other
* coding oddities that help some methods perform reasonably even
* when interpreted (not compiled).
*
* 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) 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.
*/
/**
* Factory for creating new {@link ForkJoinWorkerThread}s.
* A {@code ForkJoinWorkerThreadFactory} must be defined and used
* for {@code ForkJoinWorkerThread} subclasses that extend base
* functionality or initialize threads with different contexts.
*/
public static interface ForkJoinWorkerThreadFactory {
/**
* Returns a new worker thread operating in the given pool.
*
* @param pool the pool this thread works in
* @throws NullPointerException if the pool is null
*/
public ForkJoinWorkerThread newThread(ForkJoinPool pool);
}
/**
* Default ForkJoinWorkerThreadFactory implementation; creates a
* new ForkJoinWorkerThread.
*/
static class DefaultForkJoinWorkerThreadFactory
implements ForkJoinWorkerThreadFactory {
public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
return new ForkJoinWorkerThread(pool);
}
}
/**
* Creates a new ForkJoinWorkerThread. This factory is used unless
* overridden in ForkJoinPool constructors.
*/
public static final ForkJoinWorkerThreadFactory
defaultForkJoinWorkerThreadFactory;
/**
* Permission required for callers of methods that may start or
* kill threads.
*/
private static final RuntimePermission modifyThreadPermission;
/**
* If there is a security manager, makes sure caller has
* permission to modify threads.
*/
private static void checkPermission() {
SecurityManager security = System.getSecurityManager();
if (security != null)
security.checkPermission(modifyThreadPermission);
}
/**
* Generator for assigning sequence numbers as pool names.
*/
private static final AtomicInteger poolNumberGenerator;
/**
* 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 >>> 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 =
* (int)ctl. The ec field is never accessed alone, but always
* together with id and st. The offsets of counts by the target
* parallelism and the positionings of fields makes it possible to
* perform the most common checks via sign tests of fields: When
* ac is negative, there are not enough active workers, when tc is
* negative, there are not enough total workers, when id is
* negative, there is at least one waiting worker, and when e is
* 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.
*/
// bit positions/shifts for fields
private static final int AC_SHIFT = 48;
private static final int TC_SHIFT = 32;
private static final int ST_SHIFT = 31;
private static final int EC_SHIFT = 16;
// bounds
private static final int MAX_ID = 0x7fff; // max poolIndex
private static final int SMASK = 0xffff; // mask short bits
private static final int SHORT_SIGN = 1 << 15;
private static final int INT_SIGN = 1 << 31;
// masks
private static final long STOP_BIT = 0x0001L << ST_SHIFT;
private static final long AC_MASK = ((long)SMASK) << AC_SHIFT;
private static final long TC_MASK = ((long)SMASK) << TC_SHIFT;
// units for incrementing and decrementing
private static final long TC_UNIT = 1L << TC_SHIFT;
private static final long AC_UNIT = 1L << AC_SHIFT;
// masks and units for dealing with u = (int)(ctl >>> 32)
private static final int UAC_SHIFT = AC_SHIFT - 32;
private static final int UTC_SHIFT = TC_SHIFT - 32;
private static final int UAC_MASK = SMASK << UAC_SHIFT;
private static final int UTC_MASK = SMASK << UTC_SHIFT;
private static final int UAC_UNIT = 1 << UAC_SHIFT;
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 E_SEQ = 1 << EC_SHIFT;
// 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;
/**
* 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.
*/
private static final long SHRINK_RATE =
4L * 1000L * 1000L * 1000L; // 4 seconds
/**
* The timeout value for attempted shrinkage, includes
* some slop to cope with system timer imprecision.
*/
private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
/**
* 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.
*/
private static final int MAX_HELP_DEPTH = 16;
/*
* 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.
*/
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);
}
}
/**
* 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.
*/
static final class EmptyTask extends ForkJoinTask A {@code ManagedBlocker} provides two methods. Method
* {@code isReleasable} must return {@code true} if blocking is
* not necessary. Method {@code block} blocks the current thread
* if necessary (perhaps internally invoking {@code isReleasable}
* before actually blocking). These actions are performed by any
* thread invoking {@link ForkJoinPool#managedBlock}. The
* unusual methods in this API accommodate synchronizers that may,
* but don't usually, block for long periods. Similarly, they
* allow more efficient internal handling of cases in which
* additional workers may be, but usually are not, needed to
* ensure sufficient parallelism. Toward this end,
* implementations of method {@code isReleasable} must be amenable
* to repeated invocation.
*
* For example, here is a ManagedBlocker based on a
* ReentrantLock:
* Here is a class that possibly blocks waiting for an
* item on a given queue:
* If the caller is not a {@link ForkJoinTask}, this method is
* behaviorally equivalent to
a * {@code
* class ManagedLocker implements ManagedBlocker {
* final ReentrantLock lock;
* boolean hasLock = false;
* ManagedLocker(ReentrantLock lock) { this.lock = lock; }
* public boolean block() {
* if (!hasLock)
* lock.lock();
* return true;
* }
* public boolean isReleasable() {
* return hasLock || (hasLock = lock.tryLock());
* }
* }}
*
* {@code
* class QueueTaker
*/
public static interface ManagedBlocker {
/**
* Possibly blocks the current thread, for example waiting for
* a lock or condition.
*
* @return {@code true} if no additional blocking is necessary
* (i.e., if isReleasable would return true)
* @throws InterruptedException if interrupted while waiting
* (the method is not required to do so, but is allowed to)
*/
boolean block() throws InterruptedException;
/**
* Returns {@code true} if blocking is unnecessary.
*/
boolean isReleasable();
}
/**
* Blocks in accord with the given blocker. If the current thread
* is a {@link ForkJoinWorkerThread}, this method possibly
* arranges for a spare thread to be activated if necessary to
* ensure sufficient parallelism while the current thread is blocked.
*
* {@code
* while (!blocker.isReleasable())
* if (blocker.block())
* return;
* }
*
* If the caller is a {@code ForkJoinTask}, then the pool may
* first be expanded to ensure parallelism, and later adjusted.
*
* @param blocker the blocker
* @throws InterruptedException if blocker.block did so
*/
public static void managedBlock(ManagedBlocker blocker)
throws InterruptedException {
Thread t = Thread.currentThread();
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;
}
}
}
// AbstractExecutorService overrides. These rely on undocumented
// fact that ForkJoinTask.adapt returns ForkJoinTasks that also
// implement RunnableFuture.
protected