--- jsr166/src/jsr166y/ForkJoinPool.java 2010/04/18 13:59:57 1.55 +++ jsr166/src/jsr166y/ForkJoinPool.java 2010/09/07 23:49:30 1.79 @@ -6,22 +6,28 @@ package jsr166y; -import java.util.concurrent.*; - import java.util.ArrayList; import java.util.Arrays; import java.util.Collection; import java.util.Collections; import java.util.List; +import java.util.concurrent.AbstractExecutorService; +import java.util.concurrent.Callable; +import java.util.concurrent.CountDownLatch; +import java.util.concurrent.ExecutorService; +import java.util.concurrent.Future; +import java.util.concurrent.RejectedExecutionException; +import java.util.concurrent.RunnableFuture; +import java.util.concurrent.TimeUnit; +import java.util.concurrent.TimeoutException; +import java.util.concurrent.atomic.AtomicInteger; import java.util.concurrent.locks.LockSupport; import java.util.concurrent.locks.ReentrantLock; -import java.util.concurrent.atomic.AtomicInteger; -import java.util.concurrent.CountDownLatch; /** * An {@link ExecutorService} for running {@link ForkJoinTask}s. * A {@code ForkJoinPool} provides the entry point for submissions - * from non-{@code ForkJoinTask}s, as well as management and + * from non-{@code ForkJoinTask} clients, as well as management and * monitoring operations. * *
A {@code ForkJoinPool} differs from other kinds of {@link @@ -30,33 +36,19 @@ import java.util.concurrent.CountDownLat * 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). A {@code ForkJoinPool} may also be used for mixed - * execution of some plain {@code Runnable}- or {@code Callable}- - * based activities along with {@code ForkJoinTask}s. When setting - * {@linkplain #setAsyncMode async mode}, a {@code ForkJoinPool} may - * also be appropriate for use with fine-grained tasks of any form - * that are never joined. Otherwise, other {@code ExecutorService} - * implementations are typically more appropriate choices. + * 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. * *
A {@code ForkJoinPool} is constructed with a given target * parallelism level; by default, equal to the number of available - * processors. Unless configured otherwise via {@link - * #setMaintainsParallelism}, the pool attempts to maintain this - * number of 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 performed in the face of blocked IO or other unmanaged - * synchronization. The nested {@link ManagedBlocker} interface - * enables extension of the kinds of synchronization accommodated. - * The target parallelism level may also be changed dynamically - * ({@link #setParallelism}). The total number of threads may be - * limited using method {@link #setMaximumPoolSize}, in which case it - * may become possible for the activities of a pool to stall due to - * the lack of available threads to process new tasks. When the pool - * is executing tasks, these and other configuration setting methods - * may only gradually affect actual pool sizes. It is normally best - * practice to invoke these methods only when the pool is known to be - * quiescent. + * 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 @@ -65,6 +57,40 @@ import java.util.concurrent.CountDownLat * {@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 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. + * + *
+ * | 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 @@ -89,7 +115,8 @@ import java.util.concurrent.CountDownLat * {@code IllegalArgumentException}. * *
This implementation rejects submitted tasks (that is, by throwing
- * {@link RejectedExecutionException}) only when the pool is shut down.
+ * {@link RejectedExecutionException}) only when the pool is shut down
+ * or internal resources have been exhausted.
*
* @since 1.7
* @author Doug Lea
@@ -116,21 +143,61 @@ public class ForkJoinPool extends Abstra
* of tasks profit from cache affinities, but others are harmed by
* cache pollution effects.)
*
+ * Beyond work-stealing support and essential bookkeeping, 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. Given that the creation costs of most threads on most
+ * systems mainly surrounds setting up runtime stacks, thread
+ * creation and switching is usually not much more expensive than
+ * stack creation and switching, and is more flexible). 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.helpJoinTask tracks joining->stealing
+ * links to try to find such a task.
+ *
+ * Compensating: Unless there are already enough live threads,
+ * method helpMaintainParallelism() may create or
+ * re-activate a spare thread to compensate for blocked
+ * joiners until they unblock.
+ *
+ * It is impossible to keep exactly the target (parallelism)
+ * number of threads running at any given time. Determining
+ * 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. Compensation
+ * occurs in slow-motion. It is triggered only upon timeouts of
+ * Object.wait used for joins. This reduces poor decisions that
+ * would otherwise be made when threads are waiting for others
+ * that are stalled because of unrelated activities such as
+ * garbage collection.
+ *
+ * The ManagedBlocker extension API can't use helping so relies
+ * only on compensation in method awaitBlocker.
+ *
* The main throughput advantages of work-stealing stem from
* decentralized control -- workers mostly steal tasks from each
* other. We do not want to negate this by creating bottlenecks
- * implementing the management responsibilities of this class. So
- * we use a collection of techniques that avoid, reduce, or cope
- * well with contention. These entail several instances of
- * bit-packing into CASable fields to maintain only the minimally
- * required atomicity. To enable such packing, we restrict maximum
- * parallelism to (1<<15)-1 (enabling twice this to fit into a 16
- * bit field), which is far in excess of normal operating range.
- * Even though updates to some of these bookkeeping fields do
- * sometimes contend with each other, they don't normally
- * cache-contend with updates to others enough to warrant memory
- * padding or isolation. So they are all held as fields of
- * ForkJoinPool objects. The main capabilities are as follows:
+ * implementing other management responsibilities. So we use a
+ * collection of techniques that avoid, reduce, or cope well with
+ * contention. These entail several instances of bit-packing into
+ * CASable fields to maintain only the minimally required
+ * atomicity. To enable such packing, we restrict maximum
+ * parallelism to (1<<15)-1 (enabling twice this (to accommodate
+ * unbalanced increments and decrements) to fit into a 16 bit
+ * field, which is far in excess of normal operating range. Even
+ * though updates to some of these bookkeeping fields do sometimes
+ * contend with each other, they don't normally cache-contend with
+ * updates to others enough to warrant memory padding or
+ * isolation. So they are all held as fields of ForkJoinPool
+ * objects. The main capabilities are as follows:
*
* 1. Creating and removing workers. Workers are recorded in the
* "workers" array. This is an array as opposed to some other data
@@ -140,39 +207,38 @@ public class ForkJoinPool extends Abstra
* (workerLock) but the array is otherwise concurrently readable,
* and accessed directly by workers. To simplify index-based
* operations, the array size is always a power of two, and all
- * readers must tolerate null slots. Currently, all but the first
- * 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.
+ * readers must tolerate null slots. Currently, 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 unusual
+ * code constructions here). In essence, the workers array serves
+ * as a WeakReference mechanism. Thus for example the event queue
+ * stores worker indices, not worker references. Access to the
+ * workers in associated methods (for example releaseEventWaiters)
+ * 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 shutdown, in which case
+ * it is OK to give up. On termination, we just clobber these
+ * data structures without trying to use them.
*
* 2. Bookkeeping for dynamically adding and removing workers. We
- * maintain a given level of parallelism (or, if
- * maintainsParallelism is false, at least avoid starvation). When
- * some workers are known to be blocked (on joins or via
+ * aim to approximately maintain the given level of parallelism.
+ * When some workers are known to be blocked (on joins or via
* ManagedBlocker), we may create or resume others to take their
* place until they unblock (see below). Implementing this
* requires counts of the number of "running" threads (i.e., those
- * that are neither blocked nor artifically suspended) as well as
+ * that are neither blocked nor artificially suspended) as well as
* the total number. These two values are packed into one field,
* "workerCounts" because we need accurate snapshots when deciding
- * to create, resume or suspend. To support these decisions,
- * updates must be prospective (not retrospective). For example,
- * the running count is decremented before blocking by a thread
- * about to block, but incremented by the thread about to unblock
- * it. (In a few cases, these prospective updates may need to be
- * rolled back, for example when deciding to create a new worker
- * but the thread factory fails or returns null. In these cases,
- * we are no worse off wrt other decisions than we would be
- * otherwise.) Updates to the workerCounts field sometimes
- * transiently encounter a fair amount of contention when join
- * dependencies are such that many threads block or unblock at
- * about the same time. We alleviate this by sometimes bundling
- * updates (for example blocking one thread on join and resuming a
- * spare cancel each other out), and in most other cases
- * performing an alternative action (like releasing waiters and
- * finding spares; see below) as a more productive form of
- * backoff.
+ * to create, resume or suspend. Note however that the
+ * correspondence of these counts to reality is not guaranteed. In
+ * particular updates for unblocked threads may lag until they
+ * actually wake up.
*
* 3. Maintaining global run state. The run state of the pool
* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
@@ -201,7 +267,7 @@ public class ForkJoinPool extends Abstra
* workers that previously could not find a task to now find one:
* Submission of a new task to the pool, or another worker pushing
* a task onto a previously empty queue. (We also use this
- * mechanism for termination and reconfiguration actions that
+ * mechanism for configuration and termination actions that
* require wakeups of idle workers). Each worker maintains its
* last known event count, and blocks when a scan for work did not
* find a task AND its lastEventCount matches the current
@@ -212,94 +278,66 @@ public class ForkJoinPool extends Abstra
* a record (field nextEventWaiter) for the next waiting worker.
* In addition to allowing simpler decisions about need for
* wakeup, the event count bits in eventWaiters serve the role of
- * tags to avoid ABA errors in Treiber stacks. To reduce delays
- * in task diffusion, workers not otherwise occupied may invoke
- * method releaseWaiters, that removes and signals (unparks)
- * workers not waiting on current count. To minimize task
- * production stalls associate with signalling, any worker pushing
- * a task on an empty queue invokes the weaker method signalWork,
- * that only releases idle workers until it detects interference
- * by other threads trying to release, and lets them take
- * over. The net effect is a tree-like diffusion of signals, where
- * released threads and possibly others) help with unparks. To
- * further reduce contention effects a bit, failed CASes to
- * increment field eventCount are tolerated without retries.
+ * tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
+ * released threads also try to release at most two others. The
+ * net effect is a tree-like diffusion of signals, where released
+ * threads (and possibly others) help with unparks. To further
+ * reduce contention effects a bit, failed CASes to increment
+ * field eventCount are tolerated without retries in signalWork.
* Conceptually they are merged into the same event, which is OK
* when their only purpose is to enable workers to scan for work.
*
- * 5. Managing suspension of extra workers. When a worker is about
- * to block waiting for a join (or via ManagedBlockers), we may
- * create a new thread to maintain parallelism level, or at least
- * avoid starvation (see below). Usually, extra threads are needed
- * for only very short periods, yet join dependencies are such
- * that we sometimes need them in bursts. Rather than create new
- * threads each time this happens, we suspend no-longer-needed
- * extra ones as "spares". For most purposes, we don't distinguish
- * "extra" spare threads from normal "core" threads: On each call
- * to preStep (the only point at which we can do this) a worker
- * checks to see if there are now too many running workers, and if
- * so, suspends itself. Methods preJoin and doBlock look for
- * suspended threads to resume before considering creating a new
- * replacement. We don't need a special data structure to maintain
- * spares; simply scanning the workers array looking for
- * worker.isSuspended() is fine because the calling thread is
- * otherwise not doing anything useful anyway; we are at least as
- * happy if after locating a spare, the caller doesn't actually
- * block because the join is ready before we try to adjust and
- * compensate. Note that this is intrinsically racy. One thread
- * may become a spare at about the same time as another is
- * needlessly being created. We counteract this and related slop
- * in part by requiring resumed spares to immediately recheck (in
- * preStep) to see whether they they should re-suspend. The only
- * effective difference between "extra" and "core" threads is that
- * we allow the "extra" ones to time out and die if they are not
- * resumed within a keep-alive interval of a few seconds. This is
- * implemented mainly within ForkJoinWorkerThread, but requires
- * some coordination (isTrimmed() -- meaning killed while
- * suspended) to correctly maintain pool counts.
- *
- * 6. Deciding when to create new workers. The main dynamic
- * control in this class is deciding when to create extra threads,
- * in methods preJoin and doBlock. We always need to create one
- * when the number of running threads becomes zero. But because
- * blocked joins are typically dependent, we don't necessarily
- * need or want one-to-one replacement. Using a one-to-one
- * compensation rule often leads to enough useless overhead
- * creating, suspending, resuming, and/or killing threads to
- * signficantly degrade throughput. We use a rule reflecting the
- * idea that, the more spare threads you already have, the more
- * evidence you need to create another one; where "evidence" is
- * expressed as the current deficit -- target minus running
- * threads. To reduce flickering and drift around target values,
- * the relation is quadratic: adding a spare if (dc*dc)>=(sc*pc)
- * (where dc is deficit, sc is number of spare threads and pc is
- * target parallelism.) This effectively reduces churn at the
- * price of systematically undershooting target parallelism when
- * many threads are blocked. However, biasing toward undeshooting
- * partially compensates for the above mechanics to suspend extra
- * threads, that normally lead to overshoot because we can only
- * suspend workers in-between top-level actions. It also better
- * copes with the fact that some of the methods in this class tend
- * to never become compiled (but are interpreted), so some
- * components of the entire set of controls might execute many
- * times faster than others. And similarly for cases where the
- * apparent lack of work is just due to GC stalls and other
- * transient system activity.
- *
- * 7. Maintaining other configuration parameters and monitoring
- * statistics. Updates to fields controlling parallelism level,
- * max size, etc can only meaningfully take effect for individual
- * threads upon their next top-level actions; i.e., between
- * stealing/running tasks/submission, which are separated by calls
- * to preStep. Memory ordering for these (assumed infrequent)
- * reconfiguration calls is ensured by using reads and writes to
- * volatile field workerCounts (that must be read in preStep anyway)
- * as "fences" -- user-level reads are preceded by reads of
- * workCounts, and writes are followed by no-op CAS to
- * workerCounts. The values reported by other management and
- * monitoring methods are either computed on demand, or are kept
- * in fields that are only updated when threads are otherwise
- * idle.
+ * 5. Managing suspension of extra workers. When a worker notices
+ * (usually upon timeout of a wait()) that there are too few
+ * running threads, we may create a new thread to maintain
+ * parallelism level, or at least avoid starvation. Usually, extra
+ * threads are needed for only very short periods, yet join
+ * dependencies are such that we sometimes need them in
+ * bursts. Rather than create new threads each time this happens,
+ * we suspend no-longer-needed extra ones as "spares". For most
+ * purposes, we don't distinguish "extra" spare threads from
+ * normal "core" threads: On each call to preStep (the only point
+ * at which we can do this) a worker checks to see if there are
+ * now too many running workers, and if so, suspends itself.
+ * Method helpMaintainParallelism looks for suspended threads to
+ * resume before considering creating a new replacement. The
+ * spares themselves are encoded on another variant of a Treiber
+ * Stack, headed at field "spareWaiters". Note that the use of
+ * spares is intrinsically racy. One thread may become a spare at
+ * about the same time as another is needlessly being created. We
+ * counteract this and related slop in part by requiring resumed
+ * spares to immediately recheck (in preStep) to see whether they
+ * should re-suspend.
+ *
+ * 6. Killing off unneeded workers. A timeout mechanism is used to
+ * shed unused workers: The oldest (first) event queue waiter uses
+ * a timed rather than hard wait. When this wait times out without
+ * a normal wakeup, it tries to shutdown any one (for convenience
+ * the newest) other spare or event waiter via
+ * tryShutdownUnusedWorker. This eventually reduces the number of
+ * worker threads to a minimum of one after a long enough period
+ * without use.
+ *
+ * 7. Deciding when to create new workers. The main dynamic
+ * control in this class is deciding when to create extra threads
+ * in method helpMaintainParallelism. We would like to keep
+ * exactly #parallelism threads running, which is an impossible
+ * task. We always need to create one when the number of running
+ * threads would become zero and all workers are busy. Beyond
+ * this, we must rely on heuristics that work well in the
+ * presence of transient phenomena such as GC stalls, dynamic
+ * compilation, and wake-up lags. These transients are extremely
+ * common -- we are normally trying to fully saturate the CPUs on
+ * a machine, so almost any activity other than running tasks
+ * impedes accuracy. Our main defense is to allow parallelism to
+ * lapse for a while during joins, and use a timeout to see if,
+ * after the resulting settling, there is still a need for
+ * additional workers. This also better copes with the fact that
+ * some of the methods in this class tend to never become compiled
+ * (but are interpreted), so some components of the entire set of
+ * controls might execute 100 times faster than others. And
+ * similarly for cases where the apparent lack of work is just due
+ * to GC stalls and other transient system activity.
*
* Beware that there is a lot of representation-level coupling
* among classes ForkJoinPool, ForkJoinWorkerThread, and
@@ -312,11 +350,13 @@ public class ForkJoinPool extends Abstra
*
* Style notes: There are lots of inline assignments (of form
* "while ((local = field) != 0)") which are usually the simplest
- * way to ensure read orderings. Also 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 a few
- * other coding oddities that help some methods perform reasonably
- * even when interpreted (not compiled).
+ * way to ensure the required read orderings (which are sometimes
+ * critical). Also 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), at the expense of some messy constructions that
+ * reduce byte code counts.
*
* The order of declarations in this file is: (1) statics (2)
* fields (along with constants used when unpacking some of them)
@@ -345,7 +385,7 @@ public class ForkJoinPool extends Abstra
* Default ForkJoinWorkerThreadFactory implementation; creates a
* new ForkJoinWorkerThread.
*/
- static class DefaultForkJoinWorkerThreadFactory
+ static class DefaultForkJoinWorkerThreadFactory
implements ForkJoinWorkerThreadFactory {
public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
return new ForkJoinWorkerThread(pool);
@@ -384,10 +424,33 @@ public class ForkJoinPool extends Abstra
new AtomicInteger();
/**
- * Absolute bound for parallelism level. Twice this number must
- * fit into a 16bit field to enable word-packing for some counts.
+ * The time to block in a join (see awaitJoin) before checking if
+ * a new worker should be (re)started to maintain parallelism
+ * level. The value should be short enough to maintain global
+ * responsiveness and progress but long enough to avoid
+ * counterproductive firings during GC stalls or unrelated system
+ * activity, and to not bog down systems with continual re-firings
+ * on GCs or legitimately long waits.
*/
- private static final int MAX_THREADS = 0x7fff;
+ private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
+
+ /**
+ * The wakeup interval (in nanoseconds) for the oldest worker
+ * waiting for an event to invoke tryShutdownUnusedWorker 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_NANOS =
+ 30L * 1000L * 1000L * 1000L; // 2 per minute
+
+ /**
+ * Absolute bound for parallelism level. Twice this number plus
+ * one (i.e., 0xfff) must fit into a 16bit field to enable
+ * word-packing for some counts and indices.
+ */
+ private static final int MAX_WORKERS = 0x7fff;
/**
* Array holding all worker threads in the pool. Array size must
@@ -413,7 +476,7 @@ public class ForkJoinPool extends Abstra
/**
* Latch released upon termination.
*/
- private final CountDownLatch terminationLatch;
+ private final Phaser termination;
/**
* Creation factory for worker threads.
@@ -427,37 +490,52 @@ public class ForkJoinPool extends Abstra
private volatile long stealCount;
/**
- * Encoded record of top of treiber stack of threads waiting for
+ * Encoded record of top of Treiber stack of threads waiting for
* events. The top 32 bits contain the count being waited for. The
- * bottom word contains one plus the pool index of waiting worker
- * thread.
+ * bottom 16 bits contains one plus the pool index of waiting
+ * worker thread. (Bits 16-31 are unused.)
*/
private volatile long eventWaiters;
private static final int EVENT_COUNT_SHIFT = 32;
- private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
+ private static final long WAITER_ID_MASK = (1L << 16) - 1L;
/**
* A counter for events that may wake up worker threads:
* - Submission of a new task to the pool
* - A worker pushing a task on an empty queue
- * - termination and reconfiguration
+ * - termination
*/
private volatile int eventCount;
/**
+ * Encoded record of top of Treiber stack of spare threads waiting
+ * for resumption. The top 16 bits contain an arbitrary count to
+ * avoid ABA effects. The bottom 16bits contains one plus the pool
+ * index of waiting worker thread.
+ */
+ private volatile int spareWaiters;
+
+ private static final int SPARE_COUNT_SHIFT = 16;
+ private static final int SPARE_ID_MASK = (1 << 16) - 1;
+
+ /**
* Lifecycle control. The low word contains the number of workers
* that are (probably) executing tasks. This value is atomically
* incremented before a worker gets a task to run, and decremented
- * when worker has no tasks and cannot find any. Bits 16-18
+ * when a worker has no tasks and cannot find any. Bits 16-18
* contain runLevel value. When all are zero, the pool is
* running. Level transitions are monotonic (running -> shutdown
* -> terminating -> terminated) so each transition adds a bit.
* These are bundled together to ensure consistent read for
* termination checks (i.e., that runLevel is at least SHUTDOWN
* and active threads is zero).
+ *
+ * Notes: Most direct CASes are dependent on these bitfield
+ * positions. Also, this field is non-private to enable direct
+ * performance-sensitive CASes in ForkJoinWorkerThread.
*/
- private volatile int runState;
+ volatile int runState;
// Note: The order among run level values matters.
private static final int RUNLEVEL_SHIFT = 16;
@@ -465,7 +543,6 @@ public class ForkJoinPool extends Abstra
private static final int TERMINATING = 1 << (RUNLEVEL_SHIFT + 1);
private static final int TERMINATED = 1 << (RUNLEVEL_SHIFT + 2);
private static final int ACTIVE_COUNT_MASK = (1 << RUNLEVEL_SHIFT) - 1;
- private static final int ONE_ACTIVE = 1; // active update delta
/**
* Holds number of total (i.e., created and not yet terminated)
@@ -474,8 +551,7 @@ public class ForkJoinPool extends Abstra
* making decisions about creating and suspending spare
* threads. Updated only by CAS. Note that adding a new worker
* requires incrementing both counts, since workers start off in
- * running state. This field is also used for memory-fencing
- * configuration parameters.
+ * running state.
*/
private volatile int workerCounts;
@@ -484,86 +560,73 @@ public class ForkJoinPool extends Abstra
private static final int ONE_RUNNING = 1;
private static final int ONE_TOTAL = 1 << TOTAL_COUNT_SHIFT;
- /*
- * Fields parallelism. maxPoolSize, locallyFifo,
- * maintainsParallelism, and ueh are non-volatile, but external
- * reads/writes use workerCount fences to ensure visability.
- */
-
/**
* The target parallelism level.
+ * Accessed directly by ForkJoinWorkerThreads.
*/
- private int parallelism;
-
- /**
- * The maximum allowed pool size.
- */
- private int maxPoolSize;
+ final int parallelism;
/**
* True if use local fifo, not default lifo, for local polling
- * Replicated by ForkJoinWorkerThreads
+ * Read by, and replicated by ForkJoinWorkerThreads
*/
- private boolean locallyFifo;
+ final boolean locallyFifo;
/**
- * Controls whether to add spares to maintain parallelism
+ * The uncaught exception handler used when any worker abruptly
+ * terminates.
*/
- private boolean maintainsParallelism;
-
- /**
- * The uncaught exception handler used when any worker
- * abruptly terminates
- */
- private Thread.UncaughtExceptionHandler ueh;
+ private final Thread.UncaughtExceptionHandler ueh;
/**
* Pool number, just for assigning useful names to worker threads
*/
private final int poolNumber;
- // utilities for updating fields
+ // Utilities for CASing fields. Note that most of these
+ // are usually manually inlined by callers
/**
- * Adds delta to running count. Used mainly by ForkJoinTask.
- *
- * @param delta the number to add
+ * Increments running count part of workerCounts
*/
- final void updateRunningCount(int delta) {
- int wc;
+ final void incrementRunningCount() {
+ int c;
do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc = workerCounts,
- wc + delta));
- }
-
- /**
- * Write fence for user modifications of pool parameters
- * (parallelism. etc). Note that it doesn't matter if CAS fails.
- */
- private void workerCountWriteFence() {
- int wc;
- UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc = workerCounts, wc);
+ c = workerCounts,
+ c + ONE_RUNNING));
}
/**
- * Read fence for external reads of pool parameters
- * (parallelism. maxPoolSize, etc).
+ * Tries to decrement running count unless already zero
*/
- private void workerCountReadFence() {
- int ignore = workerCounts;
+ final boolean tryDecrementRunningCount() {
+ int wc = workerCounts;
+ if ((wc & RUNNING_COUNT_MASK) == 0)
+ return false;
+ return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
+ wc, wc - ONE_RUNNING);
}
/**
- * Tries incrementing active count; fails on contention.
- * Called by workers before executing tasks.
+ * Forces decrement of encoded workerCounts, awaiting nonzero if
+ * (rarely) necessary when other count updates lag.
*
- * @return true on success
+ * @param dr -- either zero or ONE_RUNNING
+ * @param dt -- either zero or ONE_TOTAL
*/
- final boolean tryIncrementActiveCount() {
- int c;
- return UNSAFE.compareAndSwapInt(this, runStateOffset,
- c = runState, c + ONE_ACTIVE);
+ private void decrementWorkerCounts(int dr, int dt) {
+ for (;;) {
+ int wc = workerCounts;
+ if ((wc & RUNNING_COUNT_MASK) - dr < 0 ||
+ (wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
+ if ((runState & TERMINATED) != 0)
+ return; // lagging termination on a backout
+ Thread.yield();
+ }
+ if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
+ wc, wc - (dr + dt)))
+ return;
+ }
}
/**
@@ -573,7 +636,7 @@ public class ForkJoinPool extends Abstra
final boolean tryDecrementActiveCount() {
int c;
return UNSAFE.compareAndSwapInt(this, runStateOffset,
- c = runState, c - ONE_ACTIVE);
+ c = runState, c - 1);
}
/**
@@ -602,12 +665,12 @@ public class ForkJoinPool extends Abstra
lock.lock();
try {
ForkJoinWorkerThread[] ws = workers;
- int len = ws.length;
- if (k < 0 || k >= len || ws[k] != null) {
- for (k = 0; k < len && ws[k] != null; ++k)
+ int n = ws.length;
+ if (k < 0 || k >= n || ws[k] != null) {
+ for (k = 0; k < n && ws[k] != null; ++k)
;
- if (k == len)
- ws = Arrays.copyOf(ws, len << 1);
+ if (k == n)
+ ws = Arrays.copyOf(ws, n << 1);
}
ws[k] = w;
workers = ws; // volatile array write ensures slot visibility
@@ -618,11 +681,11 @@ public class ForkJoinPool extends Abstra
}
/**
- * Nulls out record of worker in workers array
+ * Nulls out record of worker in workers array.
*/
private void forgetWorker(ForkJoinWorkerThread w) {
int idx = w.poolIndex;
- // Locking helps method recordWorker avoid unecessary expansion
+ // Locking helps method recordWorker avoid unnecessary expansion
final ReentrantLock lock = this.workerLock;
lock.lock();
try {
@@ -634,405 +697,378 @@ public class ForkJoinPool extends Abstra
}
}
- // adding and removing workers
-
/**
- * Tries to create and add new worker. Assumes that worker counts
- * are already updated to accommodate the worker, so adjusts on
- * failure.
+ * Final callback from terminating worker. Removes record of
+ * worker from array, and adjusts counts. If pool is shutting
+ * down, tries to complete termination.
*
- * @return new worker or null if creation failed
+ * @param w the worker
*/
- private ForkJoinWorkerThread addWorker() {
- ForkJoinWorkerThread w = null;
- try {
- w = factory.newThread(this);
- } finally { // Adjust on either null or exceptional factory return
- if (w == null) {
- onWorkerCreationFailure();
- return null;
+ final void workerTerminated(ForkJoinWorkerThread w) {
+ forgetWorker(w);
+ decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
+ while (w.stealCount != 0) // collect final count
+ tryAccumulateStealCount(w);
+ tryTerminate(false);
+ }
+
+ // Waiting for and signalling events
+
+ /**
+ * Releases workers blocked on a count not equal to current count.
+ * Normally called after precheck that eventWaiters isn't zero to
+ * avoid wasted array checks. Gives up upon a change in count or
+ * upon releasing two workers, letting others take over.
+ */
+ private void releaseEventWaiters() {
+ ForkJoinWorkerThread[] ws = workers;
+ int n = ws.length;
+ long h = eventWaiters;
+ int ec = eventCount;
+ boolean releasedOne = false;
+ ForkJoinWorkerThread w; int id;
+ while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
+ (int)(h >>> EVENT_COUNT_SHIFT) != ec &&
+ id < n && (w = ws[id]) != null) {
+ if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
+ h, w.nextWaiter)) {
+ LockSupport.unpark(w);
+ if (releasedOne) // exit on second release
+ break;
+ releasedOne = true;
}
+ if (eventCount != ec)
+ break;
+ h = eventWaiters;
}
- w.start(recordWorker(w), locallyFifo, ueh);
- return w;
}
/**
- * Adjusts counts upon failure to create worker
+ * Tries to advance eventCount and releases waiters. Called only
+ * from workers.
*/
- private void onWorkerCreationFailure() {
- int c;
- do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- c = workerCounts,
- c - (ONE_RUNNING|ONE_TOTAL)));
- tryTerminate(false); // in case of failure during shutdown
+ final void signalWork() {
+ int c; // try to increment event count -- CAS failure OK
+ UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
+ if (eventWaiters != 0L)
+ releaseEventWaiters();
}
/**
- * Create enough total workers to establish target parallelism,
- * giving up if terminating or addWorker fails
+ * Adds the given worker to event queue and blocks until
+ * terminating or event count advances from the given value
+ *
+ * @param w the calling worker thread
+ * @param ec the count
*/
- private void ensureEnoughTotalWorkers() {
- int wc;
- while (runState < TERMINATING &&
- ((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism) {
- if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
- addWorker() == null))
+ private void eventSync(ForkJoinWorkerThread w, int ec) {
+ long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
+ long h;
+ while ((runState < SHUTDOWN || !tryTerminate(false)) &&
+ (((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
+ (int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
+ eventCount == ec) {
+ if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
+ w.nextWaiter = h, nh)) {
+ awaitEvent(w, ec);
break;
+ }
}
}
/**
- * Final callback from terminating worker. Removes record of
- * worker from array, and adjusts counts. If pool is shutting
- * down, tries to complete terminatation, else possibly replaces
- * the worker.
+ * Blocks the given worker (that has already been entered as an
+ * event waiter) until terminating or event count advances from
+ * the given value. The oldest (first) waiter uses a timed wait to
+ * occasionally one-by-one shrink the number of workers (to a
+ * minimum of one) if the pool has not been used for extended
+ * periods.
*
- * @param w the worker
+ * @param w the calling worker thread
+ * @param ec the count
*/
- final void workerTerminated(ForkJoinWorkerThread w) {
- if (w.active) { // force inactive
- w.active = false;
- do {} while (!tryDecrementActiveCount());
+ private void awaitEvent(ForkJoinWorkerThread w, int ec) {
+ while (eventCount == ec) {
+ if (tryAccumulateStealCount(w)) { // transfer while idle
+ boolean untimed = (w.nextWaiter != 0L ||
+ (workerCounts & RUNNING_COUNT_MASK) <= 1);
+ long startTime = untimed? 0 : System.nanoTime();
+ Thread.interrupted(); // clear/ignore interrupt
+ if (eventCount != ec || w.runState != 0 ||
+ runState >= TERMINATING) // recheck after clear
+ break;
+ if (untimed)
+ LockSupport.park(w);
+ else {
+ LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
+ if (eventCount != ec || w.runState != 0 ||
+ runState >= TERMINATING)
+ break;
+ if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
+ tryShutdownUnusedWorker(ec);
+ }
+ }
}
- forgetWorker(w);
-
- // decrement total count, and if was running, running count
- int unit = w.isTrimmed()? ONE_TOTAL : (ONE_RUNNING|ONE_TOTAL);
- int wc;
- do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc = workerCounts, wc - unit));
-
- accumulateStealCount(w); // collect final count
- if (!tryTerminate(false))
- ensureEnoughTotalWorkers();
}
- // Waiting for and signalling events
+ // Maintaining parallelism
/**
- * Ensures eventCount on exit is different (mod 2^32) than on
- * entry. CAS failures are OK -- any change in count suffices.
+ * Pushes worker onto the spare stack.
*/
- private void advanceEventCount() {
- int c;
- UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
+ final void pushSpare(ForkJoinWorkerThread w) {
+ int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
+ do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
+ w.nextSpare = spareWaiters,ns));
}
/**
- * Releases workers blocked on a count not equal to current count.
- */
- final void releaseWaiters() {
- long top;
- int id;
- while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
- (int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
- ForkJoinWorkerThread[] ws = workers;
- ForkJoinWorkerThread w;
- if (ws.length >= id && (w = ws[id - 1]) != null &&
- UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
- top, w.nextWaiter))
+ * Tries (once) to resume a spare if the number of running
+ * threads is less than target.
+ */
+ private void tryResumeSpare() {
+ int sw, id;
+ ForkJoinWorkerThread[] ws = workers;
+ int n = ws.length;
+ ForkJoinWorkerThread w;
+ if ((sw = spareWaiters) != 0 &&
+ (id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
+ id < n && (w = ws[id]) != null &&
+ (workerCounts & RUNNING_COUNT_MASK) < parallelism &&
+ spareWaiters == sw &&
+ UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
+ sw, w.nextSpare)) {
+ int c; // increment running count before resume
+ do {} while (!UNSAFE.compareAndSwapInt
+ (this, workerCountsOffset,
+ c = workerCounts, c + ONE_RUNNING));
+ if (w.tryUnsuspend())
LockSupport.unpark(w);
+ else // back out if w was shutdown
+ decrementWorkerCounts(ONE_RUNNING, 0);
}
}
/**
- * Advances eventCount and releases waiters until interference by
- * other releasing threads is detected.
+ * Tries to increase the number of running workers if below target
+ * parallelism: If a spare exists tries to resume it via
+ * tryResumeSpare. Otherwise, if not enough total workers or all
+ * existing workers are busy, adds a new worker. In all cases also
+ * helps wake up releasable workers waiting for work.
*/
- final void signalWork() {
- int ec;
- UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
- outer:for (;;) {
- long top = eventWaiters;
- ec = eventCount;
- for (;;) {
- ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
- int id = (int)(top & WAITER_INDEX_MASK);
- if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
- return;
- if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
- !UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
- top, top = w.nextWaiter))
- continue outer; // possibly stale; reread
- LockSupport.unpark(w);
- if (top != eventWaiters) // let someone else take over
- return;
+ private void helpMaintainParallelism() {
+ int pc = parallelism;
+ int wc, rs, tc;
+ while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
+ (rs = runState) < TERMINATING) {
+ if (spareWaiters != 0)
+ tryResumeSpare();
+ else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
+ (tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
+ break; // enough total
+ else if (runState == rs && workerCounts == wc &&
+ UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
+ wc + (ONE_RUNNING|ONE_TOTAL))) {
+ ForkJoinWorkerThread w = null;
+ try {
+ w = factory.newThread(this);
+ } finally { // adjust on null or exceptional factory return
+ if (w == null) {
+ decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
+ tryTerminate(false); // handle failure during shutdown
+ }
+ }
+ if (w == null)
+ break;
+ w.start(recordWorker(w), ueh);
+ if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
+ int c; // advance event count
+ UNSAFE.compareAndSwapInt(this, eventCountOffset,
+ c = eventCount, c+1);
+ break; // add at most one unless total below target
+ }
}
}
+ if (eventWaiters != 0L)
+ releaseEventWaiters();
}
/**
- * If worker is inactive, blocks until terminating or event count
- * advances from last value held by worker; in any case helps
- * release others.
+ * Callback from the oldest waiter in awaitEvent waking up after a
+ * period of non-use. If all workers are idle, tries (once) to
+ * shutdown an event waiter or a spare, if one exists. Note that
+ * we don't need CAS or locks here because the method is called
+ * only from one thread occasionally waking (and even misfires are
+ * OK). Note that until the shutdown worker fully terminates,
+ * workerCounts will overestimate total count, which is tolerable.
*
- * @param w the calling worker thread
+ * @param ec the event count waited on by caller (to abort
+ * attempt if count has since changed).
*/
- private void eventSync(ForkJoinWorkerThread w) {
- if (!w.active) {
- int prev = w.lastEventCount;
- long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
- ((long)(w.poolIndex + 1)));
- long top;
- while ((runState < SHUTDOWN || !tryTerminate(false)) &&
- (((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
- (int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
- eventCount == prev) {
- if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
- w.nextWaiter = top, nextTop)) {
- accumulateStealCount(w); // transfer steals while idle
- Thread.interrupted(); // clear/ignore interrupt
- while (eventCount == prev)
- w.doPark();
- break;
- }
+ private void tryShutdownUnusedWorker(int ec) {
+ if (runState == 0 && eventCount == ec) { // only trigger if all idle
+ ForkJoinWorkerThread[] ws = workers;
+ int n = ws.length;
+ ForkJoinWorkerThread w = null;
+ boolean shutdown = false;
+ int sw;
+ long h;
+ if ((sw = spareWaiters) != 0) { // prefer killing spares
+ int id = (sw & SPARE_ID_MASK) - 1;
+ if (id >= 0 && id < n && (w = ws[id]) != null &&
+ UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
+ sw, w.nextSpare))
+ shutdown = true;
+ }
+ else if ((h = eventWaiters) != 0L) {
+ long nh;
+ int id = ((int)(h & WAITER_ID_MASK)) - 1;
+ if (id >= 0 && id < n && (w = ws[id]) != null &&
+ (nh = w.nextWaiter) != 0L && // keep at least one worker
+ UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
+ shutdown = true;
+ }
+ if (w != null && shutdown) {
+ w.shutdown();
+ LockSupport.unpark(w);
}
- w.lastEventCount = eventCount;
}
- releaseWaiters();
+ releaseEventWaiters(); // in case of interference
}
/**
* Callback from workers invoked upon each top-level action (i.e.,
- * stealing a task or taking a submission and running
- * it). Performs one or both of the following:
+ * stealing a task or taking a submission and running it).
+ * Performs one or more of the following:
*
- * * If the worker cannot find work, updates its active status to
- * inactive and updates activeCount unless there is contention, in
- * which case it may try again (either in this or a subsequent
- * call). Additionally, awaits the next task event and/or helps
- * wake up other releasable waiters.
- *
- * * If there are too many running threads, suspends this worker
- * (first forcing inactivation if necessary). If it is not
- * resumed before a keepAlive elapses, the worker may be "trimmed"
- * -- killed while suspended within suspendAsSpare. Otherwise,
- * upon resume it rechecks to make sure that it is still needed.
+ * 1. If the worker is active and either did not run a task
+ * or there are too many workers, try to set its active status
+ * to inactive and update activeCount. On contention, we may
+ * try again in this or a subsequent call.
+ *
+ * 2. If not enough total workers, help create some.
+ *
+ * 3. If there are too many running workers, suspend this worker
+ * (first forcing inactive if necessary). If it is not needed,
+ * it may be shutdown while suspended (via
+ * tryShutdownUnusedWorker). Otherwise, upon resume it
+ * rechecks running thread count and need for event sync.
+ *
+ * 4. If worker did not run a task, await the next task event via
+ * eventSync if necessary (first forcing inactivation), upon
+ * which the worker may be shutdown via
+ * tryShutdownUnusedWorker. Otherwise, help release any
+ * existing event waiters that are now releasable,
*
* @param w the worker
- * @param worked false if the worker scanned for work but didn't
- * find any (in which case it may block waiting for work).
+ * @param ran true if worker ran a task since last call to this method
*/
- final void preStep(ForkJoinWorkerThread w, boolean worked) {
+ final void preStep(ForkJoinWorkerThread w, boolean ran) {
+ int wec = w.lastEventCount;
boolean active = w.active;
- boolean inactivate = !worked & active;
- for (;;) {
- if (inactivate) {
- int c = runState;
- if (UNSAFE.compareAndSwapInt(this, runStateOffset,
- c, c - ONE_ACTIVE))
- inactivate = active = w.active = false;
- }
+ boolean inactivate = false;
+ int pc = parallelism;
+ int rs;
+ while (w.runState == 0 && (rs = runState) < TERMINATING) {
+ if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
+ UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
+ inactivate = active = w.active = false;
int wc = workerCounts;
- if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
- if (!worked)
- eventSync(w);
- return;
- }
- if (!(inactivate |= active) && // must inactivate to suspend
- UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc, wc - ONE_RUNNING) &&
- !w.suspendAsSpare()) // false if trimmed
- return;
- }
- }
-
- /**
- * Adjusts counts and creates or resumes compensating threads for
- * a worker about to block on task joinMe, returning early if
- * joinMe becomes ready. First tries resuming an existing spare
- * (which usually also avoids any count adjustment), but must then
- * decrement running count to determine whether a new thread is
- * needed. See above for fuller explanation.
- */
- final void preJoin(ForkJoinTask> joinMe) {
- boolean dec = false; // true when running count decremented
- for (;;) {
- releaseWaiters(); // help other threads progress
-
- if (joinMe.status < 0) // surround spare search with done checks
- return;
- ForkJoinWorkerThread spare = null;
- for (ForkJoinWorkerThread w : workers) {
- if (w != null && w.isSuspended()) {
- spare = w;
+ if ((wc & RUNNING_COUNT_MASK) > pc) {
+ if (!(inactivate |= active) && // must inactivate to suspend
+ workerCounts == wc && // try to suspend as spare
+ UNSAFE.compareAndSwapInt(this, workerCountsOffset,
+ wc, wc - ONE_RUNNING))
+ w.suspendAsSpare();
+ }
+ else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
+ helpMaintainParallelism(); // not enough workers
+ else if (!ran) {
+ long h = eventWaiters;
+ int ec = eventCount;
+ if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
+ releaseEventWaiters(); // release others before waiting
+ else if (ec != wec) {
+ w.lastEventCount = ec; // no need to wait
break;
}
+ else if (!(inactivate |= active))
+ eventSync(w, wec); // must inactivate before sync
}
- if (joinMe.status < 0)
- return;
-
- if (spare != null && spare.tryUnsuspend()) {
- if (dec || joinMe.requestSignal() < 0) {
- int c;
- do {} while (!UNSAFE.compareAndSwapInt(this,
- workerCountsOffset,
- c = workerCounts,
- c + ONE_RUNNING));
- } // else no net count change
- LockSupport.unpark(spare);
- return;
- }
-
- int wc = workerCounts; // decrement running count
- if (!dec && (wc & RUNNING_COUNT_MASK) != 0 &&
- (dec = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc, wc -= ONE_RUNNING)) &&
- joinMe.requestSignal() < 0) { // cannot block
- int c; // back out
- do {} while (!UNSAFE.compareAndSwapInt(this,
- workerCountsOffset,
- c = workerCounts,
- c + ONE_RUNNING));
- return;
- }
-
- if (dec) {
- int tc = wc >>> TOTAL_COUNT_SHIFT;
- int pc = parallelism;
- int dc = pc - (wc & RUNNING_COUNT_MASK); // deficit count
- if ((dc < pc && (dc <= 0 || (dc * dc < (tc - pc) * pc) ||
- !maintainsParallelism)) ||
- tc >= maxPoolSize) // cannot add
- return;
- if (spare == null &&
- UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
- wc + (ONE_RUNNING|ONE_TOTAL))) {
- addWorker();
- return;
- }
- }
+ else
+ break;
}
}
/**
- * Same idea as preJoin but with too many differing details to
- * integrate: There are no task-based signal counts, and only one
- * way to do the actual blocking. So for simplicity it is directly
- * incorporated into this method.
+ * Helps and/or blocks awaiting join of the given task.
+ * See above for explanation.
+ *
+ * @param joinMe the task to join
+ * @param worker the current worker thread
*/
- final void doBlock(ManagedBlocker blocker, boolean maintainPar)
- throws InterruptedException {
- maintainPar &= maintainsParallelism; // override
- boolean dec = false;
- boolean done = false;
- for (;;) {
- releaseWaiters();
- if (done = blocker.isReleasable())
- break;
- ForkJoinWorkerThread spare = null;
- for (ForkJoinWorkerThread w : workers) {
- if (w != null && w.isSuspended()) {
- spare = w;
- break;
- }
- }
- if (done = blocker.isReleasable())
- break;
- if (spare != null && spare.tryUnsuspend()) {
- if (dec) {
- int c;
- do {} while (!UNSAFE.compareAndSwapInt(this,
- workerCountsOffset,
- c = workerCounts,
- c + ONE_RUNNING));
- }
- LockSupport.unpark(spare);
+ final void awaitJoin(ForkJoinTask> joinMe, ForkJoinWorkerThread worker) {
+ int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
+ while (joinMe.status >= 0) {
+ int wc;
+ worker.helpJoinTask(joinMe);
+ if (joinMe.status < 0)
break;
- }
- int wc = workerCounts;
- if (!dec && (wc & RUNNING_COUNT_MASK) != 0)
- dec = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc, wc -= ONE_RUNNING);
- if (dec) {
- int tc = wc >>> TOTAL_COUNT_SHIFT;
- int pc = parallelism;
- int dc = pc - (wc & RUNNING_COUNT_MASK);
- if ((dc < pc && (dc <= 0 || (dc * dc < (tc - pc) * pc) ||
- !maintainPar)) ||
- tc >= maxPoolSize)
- break;
- if (spare == null &&
- UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
- wc + (ONE_RUNNING|ONE_TOTAL))){
- addWorker();
- break;
- }
- }
- }
-
- try {
- if (!done)
- do {} while (!blocker.isReleasable() && !blocker.block());
- } finally {
- if (dec) {
- int c;
- do {} while (!UNSAFE.compareAndSwapInt(this,
- workerCountsOffset,
- c = workerCounts,
- c + ONE_RUNNING));
+ else if (retries > 0)
+ --retries;
+ else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
+ UNSAFE.compareAndSwapInt(this, workerCountsOffset,
+ wc, wc - ONE_RUNNING)) {
+ int stat, c; long h;
+ while ((stat = joinMe.status) >= 0 &&
+ (h = eventWaiters) != 0L && // help release others
+ (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
+ releaseEventWaiters();
+ if (stat >= 0 &&
+ ((workerCounts & RUNNING_COUNT_MASK) == 0 ||
+ (stat =
+ joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
+ helpMaintainParallelism(); // timeout or no running workers
+ do {} while (!UNSAFE.compareAndSwapInt
+ (this, workerCountsOffset,
+ c = workerCounts, c + ONE_RUNNING));
+ if (stat < 0)
+ break; // else restart
}
}
}
/**
- * Unless there are not enough other running threads, adjusts
- * counts for a a worker in performing helpJoin that cannot find
- * any work, so that this worker can now block.
- *
- * @return true if worker may block
+ * Same idea as awaitJoin, but no helping, retries, or timeouts.
*/
- final boolean preBlockHelpingJoin(ForkJoinTask> joinMe) {
- while (joinMe.status >= 0) {
- releaseWaiters(); // help other threads progress
-
- // if a spare exists, resume it to maintain parallelism level
- if ((workerCounts & RUNNING_COUNT_MASK) <= parallelism) {
- ForkJoinWorkerThread spare = null;
- for (ForkJoinWorkerThread w : workers) {
- if (w != null && w.isSuspended()) {
- spare = w;
- break;
- }
- }
- if (joinMe.status < 0)
- break;
- if (spare != null) {
- if (spare.tryUnsuspend()) {
- boolean canBlock = true;
- if (joinMe.requestSignal() < 0) {
- canBlock = false; // already done
- int c;
- do {} while (!UNSAFE.compareAndSwapInt
- (this, workerCountsOffset,
- c = workerCounts, c + ONE_RUNNING));
- }
- LockSupport.unpark(spare);
- return canBlock;
+ final void awaitBlocker(ManagedBlocker blocker)
+ throws InterruptedException {
+ while (!blocker.isReleasable()) {
+ int wc = workerCounts;
+ if ((wc & RUNNING_COUNT_MASK) != 0 &&
+ UNSAFE.compareAndSwapInt(this, workerCountsOffset,
+ wc, wc - ONE_RUNNING)) {
+ try {
+ while (!blocker.isReleasable()) {
+ long h = eventWaiters;
+ if (h != 0L &&
+ (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
+ releaseEventWaiters();
+ else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
+ runState < TERMINATING)
+ helpMaintainParallelism();
+ else if (blocker.block())
+ break;
}
- continue; // recheck -- another spare may exist
+ } finally {
+ int c;
+ do {} while (!UNSAFE.compareAndSwapInt
+ (this, workerCountsOffset,
+ c = workerCounts, c + ONE_RUNNING));
}
- }
-
- int wc = workerCounts; // reread to shorten CAS window
- int rc = wc & RUNNING_COUNT_MASK;
- if (rc <= 2) // keep this and at most one other thread alive
- break;
-
- if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc, wc - ONE_RUNNING)) {
- if (joinMe.requestSignal() >= 0)
- return true;
- int c; // back out
- do {} while (!UNSAFE.compareAndSwapInt
- (this, workerCountsOffset,
- c = workerCounts, c + ONE_RUNNING));
break;
}
}
- return false;
}
/**
@@ -1056,16 +1092,50 @@ public class ForkJoinPool extends Abstra
// Finish now if all threads terminated; else in some subsequent call
if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
advanceRunLevel(TERMINATED);
- terminationLatch.countDown();
+ termination.arrive();
}
return true;
}
/**
* Actions on transition to TERMINATING
+ *
+ * Runs up to four passes through workers: (0) shutting down each
+ * (without waking up if parked) to quickly spread notifications
+ * without unnecessary bouncing around event queues etc (1) wake
+ * up and help cancel tasks (2) interrupt (3) mop up races with
+ * interrupted workers
*/
private void startTerminating() {
- // Clear out and cancel submissions, ignoring exceptions
+ cancelSubmissions();
+ for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
+ int c; // advance event count
+ UNSAFE.compareAndSwapInt(this, eventCountOffset,
+ c = eventCount, c+1);
+ eventWaiters = 0L; // clobber lists
+ spareWaiters = 0;
+ for (ForkJoinWorkerThread w : workers) {
+ if (w != null) {
+ w.shutdown();
+ if (passes > 0 && !w.isTerminated()) {
+ w.cancelTasks();
+ LockSupport.unpark(w);
+ if (passes > 1) {
+ try {
+ w.interrupt();
+ } catch (SecurityException ignore) {
+ }
+ }
+ }
+ }
+ }
+ }
+ }
+
+ /**
+ * Clears out and cancels submissions, ignoring exceptions.
+ */
+ private void cancelSubmissions() {
ForkJoinTask> task;
while ((task = submissionQueue.poll()) != null) {
try {
@@ -1073,51 +1143,33 @@ public class ForkJoinPool extends Abstra
} catch (Throwable ignore) {
}
}
- // Propagate run level
- for (ForkJoinWorkerThread w : workers) {
- if (w != null)
- w.shutdown(); // also resumes suspended workers
- }
- // Ensure no straggling local tasks
- for (ForkJoinWorkerThread w : workers) {
- if (w != null)
- w.cancelTasks();
- }
- // Wake up idle workers
- advanceEventCount();
- releaseWaiters();
- // Unstick pending joins
- for (ForkJoinWorkerThread w : workers) {
- if (w != null && !w.isTerminated()) {
- try {
- w.interrupt();
- } catch (SecurityException ignore) {
- }
- }
- }
}
// misc support for ForkJoinWorkerThread
/**
- * Returns pool number
+ * Returns pool number.
*/
final int getPoolNumber() {
return poolNumber;
}
/**
- * Accumulates steal count from a worker, clearing
- * the worker's value
+ * Tries to accumulate steal count from a worker, clearing
+ * the worker's value if successful.
+ *
+ * @return true if worker steal count now zero
*/
- final void accumulateStealCount(ForkJoinWorkerThread w) {
+ final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
int sc = w.stealCount;
- if (sc != 0) {
- long c;
- w.stealCount = 0;
- do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
- c = stealCount, c + sc));
+ long c = stealCount;
+ // CAS even if zero, for fence effects
+ if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
+ if (sc != 0)
+ w.stealCount = 0;
+ return true;
}
+ return sc == 0;
}
/**
@@ -1125,10 +1177,13 @@ public class ForkJoinPool extends Abstra
* active thread.
*/
final int idlePerActive() {
- int ac = runState; // no mask -- artifically boosts during shutdown
- int pc = parallelism; // use targeted parallelism, not rc
+ int pc = parallelism; // use parallelism, not rc
+ int ac = runState; // no mask -- artificially boosts during shutdown
// Use exact results for small values, saturate past 4
- return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
+ return ((pc <= ac) ? 0 :
+ (pc >>> 1 <= ac) ? 1 :
+ (pc >>> 2 <= ac) ? 3 :
+ pc >>> 3);
}
// Public and protected methods
@@ -1137,8 +1192,9 @@ public class ForkJoinPool extends Abstra
/**
* Creates a {@code ForkJoinPool} with parallelism equal to {@link
- * java.lang.Runtime#availableProcessors}, and using the {@linkplain
- * #defaultForkJoinWorkerThreadFactory default thread factory}.
+ * java.lang.Runtime#availableProcessors}, using the {@linkplain
+ * #defaultForkJoinWorkerThreadFactory default thread factory},
+ * no UncaughtExceptionHandler, and non-async LIFO processing mode.
*
* @throws SecurityException if a security manager exists and
* the caller is not permitted to modify threads
@@ -1147,13 +1203,14 @@ public class ForkJoinPool extends Abstra
*/
public ForkJoinPool() {
this(Runtime.getRuntime().availableProcessors(),
- defaultForkJoinWorkerThreadFactory);
+ defaultForkJoinWorkerThreadFactory, null, false);
}
/**
* Creates a {@code ForkJoinPool} with the indicated parallelism
- * level and using the {@linkplain
- * #defaultForkJoinWorkerThreadFactory default thread factory}.
+ * level, the {@linkplain
+ * #defaultForkJoinWorkerThreadFactory default thread factory},
+ * no UncaughtExceptionHandler, and non-async LIFO processing mode.
*
* @param parallelism the parallelism level
* @throws IllegalArgumentException if parallelism less than or
@@ -1164,31 +1221,25 @@ public class ForkJoinPool extends Abstra
* java.lang.RuntimePermission}{@code ("modifyThread")}
*/
public ForkJoinPool(int parallelism) {
- this(parallelism, defaultForkJoinWorkerThreadFactory);
- }
-
- /**
- * Creates a {@code ForkJoinPool} with parallelism equal to {@link
- * java.lang.Runtime#availableProcessors}, and using the given
- * thread factory.
- *
- * @param factory the factory for creating new threads
- * @throws NullPointerException if the factory is null
- * @throws SecurityException if a security manager exists and
- * the caller is not permitted to modify threads
- * because it does not hold {@link
- * java.lang.RuntimePermission}{@code ("modifyThread")}
- */
- public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
- this(Runtime.getRuntime().availableProcessors(), factory);
+ this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
}
/**
- * Creates a {@code ForkJoinPool} with the given parallelism and
- * thread factory.
+ * Creates a {@code ForkJoinPool} with the given parameters.
*
- * @param parallelism the parallelism level
- * @param factory the factory for creating new threads
+ * @param parallelism the parallelism level. For default value,
+ * use {@link java.lang.Runtime#availableProcessors}.
+ * @param factory the factory for creating new threads. For default value,
+ * use {@link #defaultForkJoinWorkerThreadFactory}.
+ * @param handler the handler for internal worker threads that
+ * terminate due to unrecoverable errors encountered while executing
+ * tasks. For default value, use {@code null}.
+ * @param asyncMode if true,
+ * establishes local first-in-first-out scheduling mode for forked
+ * tasks that are never joined. This mode may be more appropriate
+ * than default locally stack-based mode in applications in which
+ * worker threads only process event-style asynchronous tasks.
+ * For default value, use {@code false}.
* @throws IllegalArgumentException if parallelism less than or
* equal to zero, or greater than implementation limit
* @throws NullPointerException if the factory is null
@@ -1197,25 +1248,25 @@ public class ForkJoinPool extends Abstra
* because it does not hold {@link
* java.lang.RuntimePermission}{@code ("modifyThread")}
*/
- public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
+ public ForkJoinPool(int parallelism,
+ ForkJoinWorkerThreadFactory factory,
+ Thread.UncaughtExceptionHandler handler,
+ boolean asyncMode) {
checkPermission();
if (factory == null)
throw new NullPointerException();
- if (parallelism <= 0 || parallelism > MAX_THREADS)
+ if (parallelism <= 0 || parallelism > MAX_WORKERS)
throw new IllegalArgumentException();
- this.poolNumber = poolNumberGenerator.incrementAndGet();
- int arraySize = initialArraySizeFor(parallelism);
this.parallelism = parallelism;
this.factory = factory;
- this.maxPoolSize = MAX_THREADS;
- this.maintainsParallelism = true;
+ this.ueh = handler;
+ this.locallyFifo = asyncMode;
+ int arraySize = initialArraySizeFor(parallelism);
this.workers = new ForkJoinWorkerThread[arraySize];
this.submissionQueue = new LinkedTransferQueue 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).
+ * 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). 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:
@@ -1850,6 +1757,26 @@ public class ForkJoinPool extends Abstra
* return hasLock || (hasLock = lock.tryLock());
* }
* }}
+ *
+ * Here is a class that possibly blocks waiting for an
+ * item on a given queue:
+ * If {@code maintainParallelism} is {@code true} and the pool
- * supports it ({@link #getMaintainsParallelism}), this method
- * attempts to maintain the pool's nominal parallelism. Otherwise
- * it activates a thread only if necessary to avoid complete
- * starvation. This option may be preferable when blockages use
- * timeouts, or are almost always brief.
+ * ensure sufficient parallelism while the current thread is blocked.
*
* If the caller is not a {@link ForkJoinTask}, this method is
* behaviorally equivalent to
@@ -1894,29 +1814,18 @@ public class ForkJoinPool extends Abstra
* first be expanded to ensure parallelism, and later adjusted.
*
* @param blocker the blocker
- * @param maintainParallelism if {@code true} and supported by
- * this pool, attempt to maintain the pool's nominal parallelism;
- * otherwise activate a thread only if necessary to avoid
- * complete starvation.
* @throws InterruptedException if blocker.block did so
*/
- public static void managedBlock(ManagedBlocker blocker,
- boolean maintainParallelism)
+ public static void managedBlock(ManagedBlocker blocker)
throws InterruptedException {
Thread t = Thread.currentThread();
- if (t instanceof ForkJoinWorkerThread)
- ((ForkJoinWorkerThread) t).pool.
- doBlock(blocker, maintainParallelism);
- else
- awaitBlocker(blocker);
- }
-
- /**
- * Performs Non-FJ blocking
- */
- private static void awaitBlocker(ManagedBlocker blocker)
- throws InterruptedException {
- do {} while (!blocker.isReleasable() && !blocker.block());
+ if (t instanceof ForkJoinWorkerThread) {
+ ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
+ w.pool.awaitBlocker(blocker);
+ }
+ else {
+ do {} while (!blocker.isReleasable() && !blocker.block());
+ }
}
// AbstractExecutorService overrides. These rely on undocumented
@@ -1941,10 +1850,11 @@ public class ForkJoinPool extends Abstra
private static final long eventCountOffset =
objectFieldOffset("eventCount", ForkJoinPool.class);
private static final long eventWaitersOffset =
- objectFieldOffset("eventWaiters",ForkJoinPool.class);
+ objectFieldOffset("eventWaiters", ForkJoinPool.class);
private static final long stealCountOffset =
- objectFieldOffset("stealCount",ForkJoinPool.class);
-
+ objectFieldOffset("stealCount", ForkJoinPool.class);
+ private static final long spareWaitersOffset =
+ objectFieldOffset("spareWaiters", ForkJoinPool.class);
private static long objectFieldOffset(String field, Class> klazz) {
try {
{@code
+ * class QueueTaker
*/
public static interface ManagedBlocker {
/**
@@ -1873,14 +1800,7 @@ public class ForkJoinPool extends Abstra
* 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 parallelism while the current thread is blocked.
- *
- *