--- jsr166/src/jsr166e/ForkJoinPool.java 2012/10/30 14:23:03 1.9 +++ jsr166/src/jsr166e/ForkJoinPool.java 2012/12/14 16:33:42 1.29 @@ -11,19 +11,13 @@ import java.util.Arrays; import java.util.Collection; import java.util.Collections; import java.util.List; -import java.util.Random; import java.util.concurrent.AbstractExecutorService; import java.util.concurrent.Callable; import java.util.concurrent.ExecutorService; import java.util.concurrent.Future; import java.util.concurrent.RejectedExecutionException; import java.util.concurrent.RunnableFuture; -import java.util.concurrent.ThreadLocalRandom; import java.util.concurrent.TimeUnit; -import java.util.concurrent.atomic.AtomicInteger; -import java.util.concurrent.atomic.AtomicLong; -import java.util.concurrent.locks.AbstractQueuedSynchronizer; -import java.util.concurrent.locks.Condition; /** * An {@link ExecutorService} for running {@link ForkJoinTask}s. @@ -43,16 +37,12 @@ import java.util.concurrent.locks.Condit * ForkJoinPool}s may also be appropriate for use with event-style * tasks that are never joined. * - *

A static {@link #commonPool} is available and appropriate for + *

A static {@link #commonPool()} is available and appropriate for * most applications. The common pool is used by any ForkJoinTask that * is not explicitly submitted to a specified pool. Using the common * pool normally reduces resource usage (its threads are slowly * reclaimed during periods of non-use, and reinstated upon subsequent - * use). The common pool is by default constructed with default - * parameters, but these may be controlled by setting any or all of - * the three properties {@code - * java.util.concurrent.ForkJoinPool.common.{parallelism, - * threadFactory, exceptionHandler}}. + * use). * *

For applications that require separate or custom pools, a {@code * ForkJoinPool} may be constructed with a given target parallelism @@ -61,7 +51,7 @@ import java.util.concurrent.locks.Condit * 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 + * blocked I/O or other unmanaged synchronization. The nested {@link * ManagedBlocker} interface enables extension of the kinds of * synchronization accommodated. * @@ -72,7 +62,7 @@ import java.util.concurrent.locks.Condit * {@link #toString} returns indications of pool state in a * convenient form for informal monitoring. * - *

As is the case with other ExecutorServices, there are three + *

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 @@ -107,6 +97,17 @@ import java.util.concurrent.locks.Condit * * * + *

The common pool is by default constructed with default + * parameters, but these may be controlled by setting three {@link + * System#getProperty system properties} with prefix {@code + * java.util.concurrent.ForkJoinPool.common}: {@code parallelism} -- + * an integer greater than zero, {@code threadFactory} -- the class + * name of a {@link ForkJoinWorkerThreadFactory}, and {@code + * exceptionHandler} -- the class name of a {@link + * java.lang.Thread.UncaughtExceptionHandler + * Thread.UncaughtExceptionHandler}. Upon any error in establishing + * these settings, default parameters are used. + * *

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 @@ -193,21 +194,24 @@ public class ForkJoinPool extends Abstra * 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 submission queues + * processing). Instead, we randomly associate submission queues * with submitting threads, using a form of hashing. The * ThreadLocal Submitter class contains a value initially used as * a hash code for choosing existing queues, but may be randomly * repositioned upon contention with other submitters. 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. 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 move on to try or create - * other queues -- they block only when creating and registering - * new queues. + * essence, submitters act like workers except that they are + * restricted to executing local tasks that they submitted (or in + * the case of CountedCompleters, others with the same root task). + * However, because most shared/external queue operations are more + * expensive than internal, and because, at steady state, external + * submitters will compete for CPU with workers, ForkJoinTask.join + * and related methods disable them from repeatedly helping to + * process tasks if all workers are active. 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 qlock), because submitters encountering a busy queue move + * on to try or create other queues -- they block only when + * creating and registering new queues. * * Management * ========== @@ -229,11 +233,13 @@ public class ForkJoinPool extends Abstra * 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. + * Field "plock" is a form of sequence lock with a saturating + * shutdown bit (similarly for per-queue "qlocks"), mainly + * protecting updates to the workQueues array, as well as to + * enable shutdown. When used as a lock, it is normally only very + * briefly held, so is nearly always available after at most a + * brief spin, but we use a monitor-based backup strategy to + * block when needed. * * Recording WorkQueues. WorkQueues are recorded in the * "workQueues" array that is created upon first use and expanded @@ -242,9 +248,11 @@ public class ForkJoinPool extends Abstra * 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. + * tolerate null slots. Worker queues are at odd indices. Shared + * (submission) queues are at even indices, up to a maximum of 64 + * slots, to limit growth even if array needs to expand to add + * more workers. Grouping them together in this way simplifies and + * speeds up task scanning. * * All worker thread creation is on-demand, triggered by task * submissions, replacement of terminated workers, and/or @@ -305,14 +313,21 @@ public class ForkJoinPool extends Abstra * * 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. + * execute. However, many other threads may notice the same task + * and each signal to wake up a thread that might take it. So in + * general, pools will be over-signalled. When a submission is + * added or another worker adds a task to a queue that has fewer + * than two tasks, they signal waiting workers (or trigger + * creation of new ones if fewer than the given parallelism level + * -- signalWork), and may leave a hint to the unparked worker to + * help signal others upon wakeup). These primary signals are + * buttressed by others (see method helpSignal) whenever other + * threads scan for work or do not have a task to process. On + * most platforms, signalling (unpark) overhead time is noticeably + * long, and the time between signalling a thread and it actually + * making progress can be very noticeably long, so it is worth + * offloading these delays from critical paths as much as + * possible. * * Trimming workers. To release resources after periods of lack of * use, a worker starting to wait when the pool is quiescent will @@ -323,8 +338,8 @@ public class ForkJoinPool extends Abstra * periods of non-use. * * Shutdown and Termination. A call to shutdownNow atomically sets - * a runState bit and then (non-atomically) sets each worker's - * runState status, cancels all unprocessed tasks, and wakes up + * a plock bit and then (non-atomically) sets each worker's + * qlock 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 @@ -352,13 +367,13 @@ public class ForkJoinPool extends Abstra * 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). + * A third form (implemented in tryRemoveAndExec) 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. @@ -380,12 +395,12 @@ public class ForkJoinPool extends Abstra * steals, rather than use per-task bookkeeping. This sometimes * requires a linear scan of workQueues 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 + * stale/wrong) of where to locate them. It 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 @@ -393,6 +408,12 @@ public class ForkJoinPool extends Abstra * to find work (see MAX_HELP) and fall back to suspending the * worker and if necessary replacing it with another. * + * Helping actions for CountedCompleters are much simpler: Method + * helpComplete can take and execute any task with the same root + * as the task being waited on. However, this still entails some + * traversal of completer chains, so is less efficient than using + * CountedCompleters without explicit joins. + * * 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 @@ -414,30 +435,41 @@ public class ForkJoinPool extends Abstra * intractable) game with an opponent that may choose the worst * (for us) active thread to stall at any time. We take several * precautions to bound losses (and thus bound gains), mainly in - * methods tryCompensate and awaitJoin: (1) We only try - * compensation after attempting enough helping steps (measured - * via counting and timing) that we have already consumed the - * estimated cost of creating and activating a new thread. (2) We - * allow up to 50% of threads to be blocked before initially - * adding any others, and unless completely saturated, check that - * some work is available for a new worker before adding. Also, we - * create up to only 50% more threads until entering a mode that - * only adds a thread if all others are possibly blocked. All - * together, this means that we might be half as fast to react, - * and create half as many threads as possible in the ideal case, - * but present vastly fewer anomalies in all other cases compared - * to both more aggressive and more conservative alternatives. - * - * 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. Several methods intrinsically - * sprawl because they must accumulate sets of consistent reads of - * volatiles held in local variables. Methods signalWork() and - * scan() are the main bottlenecks, so are especially heavily + * methods tryCompensate and awaitJoin. + * + * Common Pool + * =========== + * + * The static commonPool always exists after static + * initialization. Since it (or any other created pool) need + * never be used, we minimize initial construction overhead and + * footprint to the setup of about a dozen fields, with no nested + * allocation. Most bootstrapping occurs within method + * fullExternalPush during the first submission to the pool. + * + * When external threads submit to the common pool, they can + * perform some subtask processing (see externalHelpJoin and + * related methods). We do not need to record whether these + * submissions are to the common pool -- if not, externalHelpJoin + * returns quickly (at the most helping to signal some common pool + * workers). These submitters would otherwise be blocked waiting + * for completion, so the extra effort (with liberally sprinkled + * task status checks) in inapplicable cases amounts to an odd + * form of limited spin-wait before blocking in ForkJoinTask.join. + * + * 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. Several methods intrinsically sprawl because + * they must accumulate sets of consistent reads of volatiles held + * in local variables. 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 @@ -445,7 +477,8 @@ public class ForkJoinPool extends Abstra * 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 + * CAS'ed variable. There are also other coding oddities (including + * several unnecessary-looking hoisted null checks) that help * some methods perform reasonably even when interpreted (not * compiled). * @@ -494,20 +527,41 @@ public class ForkJoinPool extends Abstra * Default ForkJoinWorkerThreadFactory implementation; creates a * new ForkJoinWorkerThread. */ - static class DefaultForkJoinWorkerThreadFactory + static final class DefaultForkJoinWorkerThreadFactory implements ForkJoinWorkerThreadFactory { - public ForkJoinWorkerThread newThread(ForkJoinPool pool) { + public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { return new ForkJoinWorkerThread(pool); } } /** + * Per-thread records for threads that submit to pools. Currently + * holds only pseudo-random seed / index that is used to choose + * submission queues in method externalPush. In the future, this may + * also incorporate a means to implement different task rejection + * and resubmission policies. + * + * Seeds for submitters and workers/workQueues work in basically + * the same way but are initialized and updated using slightly + * different mechanics. Both are initialized using the same + * approach as in class ThreadLocal, where successive values are + * unlikely to collide with previous values. Seeds are then + * randomly modified upon collisions using xorshifts, which + * requires a non-zero seed. + */ + static final class Submitter { + int seed; + Submitter(int s) { seed = 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 { + private static final long serialVersionUID = -7721805057305804111L; EmptyTask() { status = ForkJoinTask.NORMAL; } // force done public final Void getRawResult() { return null; } public final void setRawResult(Void x) {} @@ -528,27 +582,31 @@ public class ForkJoinPool extends Abstra * * 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 commonly -(base - top) to force volatile read of base - * before top) still estimates size. + * for push, or under lock for external/shared push, 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 commonly -(base - top) to force volatile read of + * base before top) still estimates size. The lock ("qlock") is + * forced to -1 on termination, causing all further lock attempts + * to fail. (Note: we don't need CAS for termination state because + * upon pool shutdown, all shared-queues will stop being used + * anyway.) Nearly all lock bodies are set up so that exceptions + * within lock bodies are "impossible" (modulo JVM errors that + * would cause failure anyway.) * * 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. + * its publication in the queue. 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. + * 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 @@ -562,10 +620,10 @@ public class ForkJoinPool extends Abstra * 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 + * properties of JVM field layout rules.) We also take care in * allocating, sizing and resizing the array. Non-shared queue - * arrays are initialized (via method growArray) by workers before - * use. Others are allocated on first use. + * arrays are initialized by workers before use. Others are + * allocated on first use. */ static final class WorkQueue { /** @@ -588,16 +646,17 @@ public class ForkJoinPool extends Abstra */ static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M - volatile long totalSteals; // cumulative number of steals + // Heuristic padding to ameliorate unfortunate memory placements + volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; + 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 hint; // steal or signal hint (index) 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 + final int mode; // 0: lifo, > 0: fifo, < 0: shared + int nsteals; // number of steals + volatile int qlock; // 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) @@ -606,14 +665,16 @@ public class ForkJoinPool extends Abstra volatile Thread parker; // == owner during call to park; else null volatile 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; - Object p08, p09, p0a, p0b, p0c, p0d, p0e; - WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) { - this.mode = mode; + volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; + volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d; + + WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode, + int seed) { this.pool = pool; this.owner = owner; + this.mode = mode; + this.seed = seed; // Place indices in the center of array (that is not yet allocated) base = top = INITIAL_QUEUE_CAPACITY >>> 1; } @@ -626,7 +687,7 @@ public class ForkJoinPool extends Abstra return (n >= 0) ? 0 : -n; // ignore transient negative } - /** + /** * Provides a more accurate estimate of whether this queue has * any tasks than does queueSize, by checking whether a * near-empty queue has at least one unclaimed task. @@ -638,12 +699,13 @@ public class ForkJoinPool extends Abstra (n == -1 && ((a = array) == null || (m = a.length - 1) < 0 || - U.getObjectVolatile - (a, ((m & (s - 1)) << ASHIFT) + ABASE) == null))); + U.getObject + (a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null))); } /** - * Pushes a task. Call only by owner in unshared queues. + * Pushes a task. Call only by owner in unshared queues. (The + * shared-queue version is embedded in method externalPush.) * * @param task the task. Caller must ensure non-null. * @throw RejectedExecutionException if array cannot be resized @@ -652,42 +714,43 @@ public class ForkJoinPool extends Abstra ForkJoinTask[] a; ForkJoinPool p; int s = top, m, n; if ((a = array) != null) { // ignore if queue removed - U.putOrderedObject - (a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task); + int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE; + U.putOrderedObject(a, j, task); if ((n = (top = s + 1) - base) <= 2) { if ((p = pool) != null) - p.signalWork(); + p.signalWork(this); } else if (n >= m) - growArray(true); + growArray(); } } - /** - * 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 + /** + * 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. */ - final boolean trySharedPush(ForkJoinTask task) { - boolean submitted = false; - if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { - ForkJoinTask[] a = array; - int s = top; - try { - if ((a != null && a.length > s + 1 - base) || - (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 - } + final ForkJoinTask[] growArray() { + ForkJoinTask[] oldA = array; + int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; + if (size > MAXIMUM_QUEUE_CAPACITY) + throw new RejectedExecutionException("Queue capacity exceeded"); + 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 submitted; + return a; } /** @@ -710,90 +773,6 @@ public class ForkJoinPool extends Abstra return null; } - final ForkJoinTask sharedPop() { - ForkJoinTask task = null; - if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { - try { - ForkJoinTask[] a; int m; - if ((a = array) != null && (m = a.length - 1) >= 0) { - for (int s; (s = top - 1) - base >= 0;) { - long j = ((m & s) << ASHIFT) + ABASE; - ForkJoinTask t = - (ForkJoinTask)U.getObject(a, j); - if (t == null) - break; - if (U.compareAndSwapObject(a, j, t, null)) { - top = s; - task = t; - break; - } - } - } - } finally { - runState = 0; - } - } - return task; - } - - /** - * Version of pop that takes top element only if it - * its root is the given CountedCompleter. - */ - final ForkJoinTask popCC(CountedCompleter root) { - ForkJoinTask[] a; int m; - if (root != null && (a = array) != null && (m = a.length - 1) >= 0) { - for (int s; (s = top - 1) - base >= 0;) { - long j = ((m & s) << ASHIFT) + ABASE; - ForkJoinTask t = - (ForkJoinTask)U.getObject(a, j); - if (t == null || !(t instanceof CountedCompleter) || - ((CountedCompleter)t).getRoot() != root) - break; - if (U.compareAndSwapObject(a, j, t, null)) { - top = s; - return t; - } - if (root.status < 0) - break; - } - } - return null; - } - - /** - * Shared version of popCC - */ - final ForkJoinTask sharedPopCC(CountedCompleter root) { - ForkJoinTask task = null; - if (root != null && - runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { - try { - ForkJoinTask[] a; int m; - if ((a = array) != null && (m = a.length - 1) >= 0) { - for (int s; (s = top - 1) - base >= 0;) { - long j = ((m & s) << ASHIFT) + ABASE; - ForkJoinTask t = - (ForkJoinTask)U.getObject(a, j); - if (t == null || !(t instanceof CountedCompleter) || - ((CountedCompleter)t).getRoot() != root) - break; - if (U.compareAndSwapObject(a, j, t, null)) { - top = s; - task = t; - break; - } - if (root.status < 0) - break; - } - } - } finally { - runState = 0; - } - } - return task; - } - /** * Takes a task in FIFO order if b is base of queue and a task * can be claimed without contention. Specialized versions @@ -831,7 +810,7 @@ public class ForkJoinPool extends Abstra else if (base == b) { if (b + 1 == top) break; - Thread.yield(); // wait for lagging update + Thread.yield(); // wait for lagging update (very rare) } } return null; @@ -858,6 +837,7 @@ public class ForkJoinPool extends Abstra /** * Pops the given task only if it is at the current top. + * (A shared version is available only via FJP.tryExternalUnpush) */ final boolean tryUnpush(ForkJoinTask t) { ForkJoinTask[] a; int s; @@ -871,79 +851,6 @@ public class ForkJoinPool extends Abstra } /** - * Version of tryUnpush for shared queues; called by non-FJ - * submitters after prechecking that task probably exists. - */ - final boolean trySharedUnpush(ForkJoinTask t) { - boolean success = false; - if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { - try { - ForkJoinTask[] a; int s; - if ((a = array) != null && (s = top) != base && - U.compareAndSwapObject - (a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) { - top = s; - success = true; - } - } finally { - runState = 0; // unlock - } - } - return success; - } - - /** - * Polls the given task only if it is at the current base. - */ - final boolean pollFor(ForkJoinTask task) { - ForkJoinTask[] a; int b; - if ((b = base) - top < 0 && (a = array) != null) { - int j = (((a.length - 1) & b) << ASHIFT) + ABASE; - if (U.getObjectVolatile(a, j) == task && base == b && - U.compareAndSwapObject(a, j, task, null)) { - base = b + 1; - return true; - } - } - return false; - } - - /** - * 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() { @@ -967,7 +874,7 @@ public class ForkJoinPool extends Abstra return seed = r ^= r << 5; } - // Execution methods + // Specialized execution methods /** * Pops and runs tasks until empty. @@ -996,16 +903,14 @@ public class ForkJoinPool extends Abstra } /** - * If present, removes from queue and executes the given task, or - * any other cancelled task. Returns (true) immediately on any CAS + * If present, removes from queue and executes the given task, + * or any other cancelled task. Returns (true) on any CAS * or consistency check failure so caller can retry. * - * @return 0 if no progress can be made, else positive - * (this unusual convention simplifies use with tryHelpStealer.) + * @return false if no progress can be made, else true; */ - final int tryRemoveAndExec(ForkJoinTask task) { - int stat = 1; - boolean removed = false, empty = true; + final boolean tryRemoveAndExec(ForkJoinTask task) { + boolean stat = true, removed = false, empty = true; ForkJoinTask[] a; int m, s, b, n; if ((a = array) != null && (m = a.length - 1) >= 0 && (n = (s = top) - (b = base)) > 0) { @@ -1035,7 +940,7 @@ public class ForkJoinPool extends Abstra } if (--n == 0) { if (!empty && base == b) - stat = 0; + stat = false; break; } } @@ -1046,21 +951,50 @@ public class ForkJoinPool extends Abstra } /** + * Polls for and executes the given task or any other task in + * its CountedCompleter computation + */ + final boolean pollAndExecCC(ForkJoinTask root) { + ForkJoinTask[] a; int b; Object o; + outer: while ((b = base) - top < 0 && (a = array) != null) { + long j = (((a.length - 1) & b) << ASHIFT) + ABASE; + if ((o = U.getObject(a, j)) == null || + !(o instanceof CountedCompleter)) + break; + for (CountedCompleter t = (CountedCompleter)o, r = t;;) { + if (r == root) { + if (base == b && + U.compareAndSwapObject(a, j, t, null)) { + base = b + 1; + t.doExec(); + return true; + } + else + break; // restart + } + if ((r = r.completer) == null) + break outer; // not part of root computation + } + } + return false; + } + + /** * Executes a top-level task and any local tasks remaining * after execution. */ final void runTask(ForkJoinTask t) { if (t != null) { - currentSteal = t; - t.doExec(); - if (top != base) { // process remaining local tasks + (currentSteal = t).doExec(); + currentSteal = null; + if (base - top < 0) { // process remaining local tasks if (mode == 0) popAndExecAll(); else pollAndExecAll(); } ++nsteals; - currentSteal = null; + hint = -1; } } @@ -1070,8 +1004,7 @@ public class ForkJoinPool extends Abstra final void runSubtask(ForkJoinTask t) { if (t != null) { ForkJoinTask ps = currentSteal; - currentSteal = t; - t.doExec(); + (currentSteal = t).doExec(); currentSteal = ps; } } @@ -1106,7 +1039,7 @@ public class ForkJoinPool extends Abstra // Unsafe mechanics private static final sun.misc.Unsafe U; - private static final long RUNSTATE; + private static final long QLOCK; private static final int ABASE; private static final int ASHIFT; static { @@ -1115,8 +1048,8 @@ public class ForkJoinPool extends Abstra U = getUnsafe(); Class k = WorkQueue.class; Class ak = ForkJoinTask[].class; - RUNSTATE = U.objectFieldOffset - (k.getDeclaredField("runState")); + QLOCK = U.objectFieldOffset + (k.getDeclaredField("qlock")); ABASE = U.arrayBaseOffset(ak); s = U.arrayIndexScale(ak); } catch (Exception e) { @@ -1128,35 +1061,6 @@ public class ForkJoinPool extends Abstra } } - /** - * Per-thread records for threads that submit to pools. Currently - * holds only pseudo-random seed / index that is used to choose - * submission queues in method doSubmit. In the future, this may - * also incorporate a means to implement different task rejection - * and resubmission policies. - * - * Seeds for submitters and workers/workQueues work in basically - * the same way but are initialized and updated using slightly - * different mechanics. Both are initialized using the same - * approach as in class ThreadLocal, where successive values are - * unlikely to collide with previous values. This is done during - * registration for workers, but requires a separate AtomicInteger - * for submitters. Seeds are then randomly modified upon - * collisions using xorshifts, which requires a non-zero seed. - */ - static final class Submitter { - int seed; - Submitter() { - int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT); - seed = (s == 0) ? 1 : s; // ensure non-zero - } - } - - /** ThreadLocal class for Submitters */ - static final class ThreadSubmitter extends ThreadLocal { - public Submitter initialValue() { return new Submitter(); } - } - // static fields (initialized in static initializer below) /** @@ -1166,15 +1070,26 @@ public class ForkJoinPool extends Abstra public static final ForkJoinWorkerThreadFactory defaultForkJoinWorkerThreadFactory; + /** + * Per-thread submission bookkeeping. Shared across all pools + * to reduce ThreadLocal pollution and because random motion + * to avoid contention in one pool is likely to hold for others. + * Lazily initialized on first submission (but null-checked + * in other contexts to avoid unnecessary initialization). + */ + static final ThreadLocal submitters; - /** Property prefix for constructing common pool */ - private static final String propPrefix = - "java.util.concurrent.ForkJoinPool.common."; + /** + * Permission required for callers of methods that may start or + * kill threads. + */ + private static final RuntimePermission modifyThreadPermission; /** * Common (static) pool. Non-null for public use unless a static - * construction exception, but internal usages must null-check on - * use. + * construction exception, but internal usages null-check on use + * to paranoically avoid potential initialization circularities + * as well as to simplify generated code. */ static final ForkJoinPool commonPool; @@ -1184,49 +1099,43 @@ public class ForkJoinPool extends Abstra static final int commonPoolParallelism; /** - * Generator for assigning sequence numbers as pool names. + * Sequence number for creating workerNamePrefix. */ - private static final AtomicInteger poolNumberGenerator; + private static int poolNumberSequence; /** - * Generator for initial hashes/seeds for submitters. Accessed by - * Submitter class constructor. + * Return the next sequence number. We don't expect this to + * ever contend so use simple builtin sync. */ - static final AtomicInteger nextSubmitterSeed; + private static final synchronized int nextPoolId() { + return ++poolNumberSequence; + } - /** - * Permission required for callers of methods that may start or - * kill threads. - */ - private static final RuntimePermission modifyThreadPermission; + // static constants /** - * Per-thread submission bookkeeping. Shared across all pools - * to reduce ThreadLocal pollution and because random motion - * to avoid contention in one pool is likely to hold for others. + * Initial timeout value (in nanoseconds) for the thread + * triggering quiescence to park waiting for new work. On timeout, + * the thread will instead try to shrink the number of + * workers. The value should be large enough to avoid overly + * aggressive shrinkage during most transient stalls (long GCs + * etc). */ - private static final ThreadSubmitter submitters; - - // static constants + private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec /** - * Initial timeout value (in nanoseconds) for the thread triggering - * quiescence to park waiting for new work. On timeout, the thread - * will instead try to shrink the number of workers. + * Timeout value when there are more threads than parallelism level */ - private static final long IDLE_TIMEOUT = 1000L * 1000L * 1000L; // 1sec + private static final long FAST_IDLE_TIMEOUT = 200L * 1000L * 1000L; /** - * Timeout value when there are more threads than parallelism level + * Tolerance for idle timeouts, to cope with timer undershoots */ - private static final long FAST_IDLE_TIMEOUT = 100L * 1000L * 1000L; + private static final long TIMEOUT_SLOP = 2000000L; // 20ms /** * The maximum stolen->joining link depth allowed in method - * tryHelpStealer. Must be a power of two. This value also - * controls the maximum number of times to try to help join a task - * without any apparent progress or change in pool state before - * giving up and blocking (see awaitJoin). Depths for legitimate + * tryHelpStealer. Must be a power of two. 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 @@ -1235,16 +1144,6 @@ public class ForkJoinPool extends Abstra private static final int MAX_HELP = 64; /** - * Secondary time-based bound (in nanosecs) for helping attempts - * before trying compensated blocking in awaitJoin. Used in - * conjunction with MAX_HELP to reduce variance due to different - * polling rates associated with different helping options. The - * value should roughly approximate the time required to create - * and/or activate a worker thread. - */ - private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec - - /** * Increment for seed generators. See class ThreadLocal for * explanation. */ @@ -1278,14 +1177,14 @@ public class ForkJoinPool extends Abstra * scan for them to avoid queuing races. Note however that * eventCount updates lag releases so usage requires care. * - * Field runState is an int packed with: + * Field plock is an int packed with: * SHUTDOWN: true if shutdown is enabled (1 bit) - * SEQ: a sequence number updated upon (de)registering workers (30 bits) - * INIT: set true after workQueues array construction (1 bit) + * SEQ: a sequence lock, with PL_LOCK bit set if locked (30 bits) + * SIGNAL: set when threads may be waiting on the lock (1 bit) * * The sequence number enables simple consistency checks: * Staleness of read-only operations on the workQueues array can - * be checked by comparing runState before vs after the reads. + * be checked by comparing plock before vs after the reads. */ // bit positions/shifts for fields @@ -1297,7 +1196,8 @@ public class ForkJoinPool extends Abstra // bounds private static final int SMASK = 0xffff; // short bits private static final int MAX_CAP = 0x7fff; // max #workers - 1 - private static final int SQMASK = 0xfffe; // even short bits + private static final int EVENMASK = 0xfffe; // even short bits + private static final int SQMASK = 0x007e; // max 64 (even) slots private static final int SHORT_SIGN = 1 << 15; private static final int INT_SIGN = 1 << 31; @@ -1322,131 +1222,199 @@ public class ForkJoinPool extends Abstra private static final int E_MASK = 0x7fffffff; // no STOP_BIT private static final int E_SEQ = 1 << EC_SHIFT; - // runState bits + // plock bits private static final int SHUTDOWN = 1 << 31; + private static final int PL_LOCK = 2; + private static final int PL_SIGNAL = 1; + private static final int PL_SPINS = 1 << 8; // access mode for WorkQueue static final int LIFO_QUEUE = 0; static final int FIFO_QUEUE = 1; static final int SHARED_QUEUE = -1; + // bounds for #steps in scan loop -- must be power 2 minus 1 + private static final int MIN_SCAN = 0x1ff; // cover estimation slop + private static final int MAX_SCAN = 0x1ffff; // 4 * max workers + // Instance fields /* - * Field layout order in this class tends to matter more than one - * would like. Runtime layout order is only loosely related to + * Field layout of 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. */ + // Heuristic padding to ameliorate unfortunate memory placements + volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; + volatile long stealCount; // collects worker counts volatile long ctl; // main pool control - final int parallelism; // parallelism level - final int localMode; // per-worker scheduling mode - volatile int nextWorkerNumber; // to create worker name string - final int submitMask; // submit queue index bound - int nextSeed; // for initializing worker seeds - volatile int mainLock; // spinlock for array updates - volatile int runState; // shutdown status and seq + volatile int plock; // shutdown status and seqLock + volatile int indexSeed; // worker/submitter index seed + final int config; // mode and parallelism level WorkQueue[] workQueues; // main registry - final ForkJoinWorkerThreadFactory factory; // factory for new workers + final ForkJoinWorkerThreadFactory factory; final Thread.UncaughtExceptionHandler ueh; // per-worker UEH final String workerNamePrefix; // to create worker name string - /* - * Mechanics for main lock protecting worker array updates. Uses - * the same strategy as ConcurrentHashMap bins -- a spinLock for - * normal cases, but falling back to builtin lock when (rarely) - * needed. See internal ConcurrentHashMap documentation for - * explanation. - */ - - static final int LOCK_WAITING = 2; // bit to indicate need for signal - static final int MAX_LOCK_SPINS = 1 << 8; + volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; + volatile Object pad18, pad19, pad1a, pad1b; - private void tryAwaitMainLock() { - int spins = MAX_LOCK_SPINS, r = 0, h; - while (((h = mainLock) & 1) != 0) { - if (r == 0) - r = ThreadLocalRandom.current().nextInt(); // randomize spins + /* + * Acquires the plock lock to protect worker array and related + * updates. This method is called only if an initial CAS on plock + * fails. This acts as a spinLock for normal cases, but falls back + * to builtin monitor to block when (rarely) needed. This would be + * a terrible idea for a highly contended lock, but works fine as + * a more conservative alternative to a pure spinlock. See + * internal ConcurrentHashMap documentation for further + * explanation of nearly the same construction. + */ + private int acquirePlock() { + int spins = PL_SPINS, r = 0, ps, nps; + for (;;) { + if (((ps = plock) & PL_LOCK) == 0 && + U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK)) + return nps; + else if (r == 0) { // randomize spins if possible + Thread t = Thread.currentThread(); WorkQueue w; Submitter z; + if ((t instanceof ForkJoinWorkerThread) && + (w = ((ForkJoinWorkerThread)t).workQueue) != null) + r = w.seed; + else if ((z = submitters.get()) != null) + r = z.seed; + else + r = 1; + } else if (spins >= 0) { r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift if (r >= 0) --spins; } - else if (U.compareAndSwapInt(this, MAINLOCK, h, h | LOCK_WAITING)) { + else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) { synchronized (this) { - if ((mainLock & LOCK_WAITING) != 0) { + if ((plock & PL_SIGNAL) != 0) { try { wait(); } catch (InterruptedException ie) { - Thread.currentThread().interrupt(); + try { + Thread.currentThread().interrupt(); + } catch (SecurityException ignore) { + } } } else - notifyAll(); // possibly won race vs signaller + notifyAll(); } - break; } } } - // Creating, registering, and deregistering workers - /** - * Tries to create and start a worker + * Unlocks and signals any thread waiting for plock. Called only + * when CAS of seq value for unlock fails. */ - private void addWorker() { - Throwable ex = null; - ForkJoinWorkerThread wt = null; - try { - if ((wt = factory.newThread(this)) != null) { - wt.start(); - return; - } - } catch (Throwable e) { - ex = e; - } - deregisterWorker(wt, ex); // adjust counts etc on failure + private void releasePlock(int ps) { + plock = ps; + synchronized (this) { notifyAll(); } } /** - * Callback from ForkJoinWorkerThread constructor to assign a - * public name. This must be separate from registerWorker because - * it is called during the "super" constructor call in - * ForkJoinWorkerThread. + * Performs secondary initialization, called when plock is zero. + * Creates workQueue array and sets plock to a valid value. The + * lock body must be exception-free (so no try/finally) so we + * optimistically allocate new array outside the lock and throw + * away if (very rarely) not needed. (A similar tactic is used in + * fullExternalPush.) Because the plock seq value can eventually + * wrap around zero, this method harmlessly fails to reinitialize + * if workQueues exists, while still advancing plock. + * + * Additionally tries to create the first worker. */ - final String nextWorkerName() { - int n; - do {} while(!U.compareAndSwapInt(this, NEXTWORKERNUMBER, - n = nextWorkerNumber, ++n)); - return workerNamePrefix.concat(Integer.toString(n)); + private void initWorkers() { + WorkQueue[] ws, nws; int ps; + int p = config & SMASK; // find power of two table size + int n = (p > 1) ? p - 1 : 1; // ensure at least 2 slots + n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; + n = (n + 1) << 1; + if ((ws = workQueues) == null || ws.length == 0) + nws = new WorkQueue[n]; + else + nws = null; + if (((ps = plock) & PL_LOCK) != 0 || + !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) + ps = acquirePlock(); + if (((ws = workQueues) == null || ws.length == 0) && nws != null) + workQueues = nws; + int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); + if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) + releasePlock(nps); + tryAddWorker(); } /** - * Callback from ForkJoinWorkerThread constructor to establish its - * poolIndex and record its WorkQueue. To avoid scanning bias due - * to packing entries in front of the workQueues array, we treat - * the array as a simple power-of-two hash table using per-thread - * seed as hash, expanding as needed. - * - * @param w the worker's queue + * Tries to create and start one worker if fewer than target + * parallelism level exist. Adjusts counts etc on failure. */ - final void registerWorker(WorkQueue w) { - while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) - tryAwaitMainLock(); + private void tryAddWorker() { + long c; int u; + while ((u = (int)((c = ctl) >>> 32)) < 0 && + (u & SHORT_SIGN) != 0 && (int)c == 0) { + long nc = (long)(((u + UTC_UNIT) & UTC_MASK) | + ((u + UAC_UNIT) & UAC_MASK)) << 32; + if (U.compareAndSwapLong(this, CTL, c, nc)) { + ForkJoinWorkerThreadFactory fac; + Throwable ex = null; + ForkJoinWorkerThread wt = null; + try { + if ((fac = factory) != null && + (wt = fac.newThread(this)) != null) { + wt.start(); + break; + } + } catch (Throwable e) { + ex = e; + } + deregisterWorker(wt, ex); + break; + } + } + } + + // Registering and deregistering workers + + /** + * Callback from ForkJoinWorkerThread to establish and record its + * WorkQueue. To avoid scanning bias due to packing entries in + * front of the workQueues array, we treat the array as a simple + * power-of-two hash table using per-thread seed as hash, + * expanding as needed. + * + * @param wt the worker thread + * @return the worker's queue + */ + final WorkQueue registerWorker(ForkJoinWorkerThread wt) { + Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps; + wt.setDaemon(true); + if ((handler = ueh) != null) + wt.setUncaughtExceptionHandler(handler); + do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed, + s += SEED_INCREMENT) || + s == 0); // skip 0 + WorkQueue w = new WorkQueue(this, wt, config >>> 16, s); + if (((ps = plock) & PL_LOCK) != 0 || + !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) + ps = acquirePlock(); + int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); try { - WorkQueue[] ws; - if ((ws = workQueues) == null) - ws = workQueues = new WorkQueue[submitMask + 1]; - if (w != null) { - int rs, n = ws.length, m = n - 1; - int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence - w.seed = (s == 0) ? 1 : s; // ensure non-zero seed - int r = (s << 1) | 1; // use odd-numbered indices - if (ws[r &= m] != null) { // collision - int probes = 0; // step by approx half size - int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2; + if ((ws = workQueues) != null) { // skip if shutting down + int n = ws.length, m = n - 1; + int r = (s << 1) | 1; // use odd-numbered indices + if (ws[r &= m] != null) { // collision + int probes = 0; // step by approx half size + int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; while (ws[r = (r + step) & m] != null) { if (++probes >= n) { workQueues = ws = Arrays.copyOf(ws, n <<= 1); @@ -1455,47 +1423,46 @@ public class ForkJoinPool extends Abstra } } } - w.eventCount = w.poolIndex = r; // establish before recording - ws[r] = w; // also update seq - runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN); + w.eventCount = w.poolIndex = r; // volatile write orders + ws[r] = w; } } finally { - if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { - mainLock = 0; - synchronized (this) { notifyAll(); }; - } + if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) + releasePlock(nps); } - + wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex))); + return w; } /** * Final callback from terminating worker, as well as upon failure - * to construct or start a worker in addWorker. Removes record of - * worker from array, and adjusts counts. If pool is shutting - * down, tries to complete termination. + * to construct or start a worker. Removes record of worker from + * array, and adjusts counts. If pool is shutting down, tries to + * complete termination. * - * @param wt the worker thread or null if addWorker failed + * @param wt the worker thread or null if construction failed * @param ex the exception causing failure, or null if none */ final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { WorkQueue w = null; if (wt != null && (w = wt.workQueue) != null) { - w.runState = -1; // ensure runState is set - long steals = w.totalSteals + w.nsteals, sc; - do {} while(!U.compareAndSwapLong(this, STEALCOUNT, - sc = stealCount, sc + steals)); - int idx = w.poolIndex; - while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) - tryAwaitMainLock(); + int ps; + w.qlock = -1; // ensure set + long ns = w.nsteals, sc; // collect steal count + do {} while (!U.compareAndSwapLong(this, STEALCOUNT, + sc = stealCount, sc + ns)); + if (((ps = plock) & PL_LOCK) != 0 || + !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) + ps = acquirePlock(); + int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); try { + int idx = w.poolIndex; WorkQueue[] ws = workQueues; if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) ws[idx] = null; } finally { - if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { - mainLock = 0; - synchronized (this) { notifyAll(); }; - } + if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) + releasePlock(nps); } } @@ -1505,16 +1472,37 @@ public class ForkJoinPool extends Abstra ((c - TC_UNIT) & TC_MASK) | (c & ~(AC_MASK|TC_MASK))))); - if (!tryTerminate(false, false) && w != null) { + if (!tryTerminate(false, false) && w != null && w.array != null) { w.cancelAll(); // cancel remaining tasks - if (w.array != null) // suppress signal if never ran - signalWork(); // wake up or create replacement - if (ex == null) // help clean refs on way out - ForkJoinTask.helpExpungeStaleExceptions(); + int e, u, i, n; WorkQueue[] ws; WorkQueue v; Thread p; + while ((u = (int)((c = ctl) >>> 32)) < 0) { + if ((e = (int)c) > 0) { // activate or create replacement + if ((ws = workQueues) != null && + ws.length > (i = e & SMASK) && + (v = ws[i]) != null && v.eventCount == (e | INT_SIGN)) { + long nc = (((long)(v.nextWait & E_MASK)) | + ((long)(u + UAC_UNIT) << 32)); + if (U.compareAndSwapLong(this, CTL, c, nc)) { + v.eventCount = (e + E_SEQ) & E_MASK; + if ((p = v.parker) != null) + U.unpark(p); + break; + } + } + else + break; + } + else { + if ((short)u < 0) + tryAddWorker(); + break; + } + } } - - if (ex != null) // rethrow - U.throwException(ex); + if (ex == null) // help clean refs on way out + ForkJoinTask.helpExpungeStaleExceptions(); + else // rethrow + ForkJoinTask.rethrow(ex); } // Submissions @@ -1522,137 +1510,102 @@ public class ForkJoinPool extends Abstra /** * Unless shutting down, adds the given task to a submission queue * at submitter's current queue index (modulo submission - * range). If no queue exists at the index, one is created. If - * the queue is busy, another index is randomly chosen. The - * submitMask bounds the effective number of queues to the - * (nearest power of two for) parallelism level. + * range). Only the most common path is directly handled in this + * method. All others are relayed to fullExternalPush. * * @param task the task. Caller must ensure non-null. */ - private void doSubmit(ForkJoinTask task) { - Submitter s = submitters.get(); - for (int r = s.seed, m = submitMask;;) { - WorkQueue[] ws; WorkQueue q; - int k = r & m & SQMASK; // use only even indices - if (runState < 0) - throw new RejectedExecutionException(); // shutting down - else if ((ws = workQueues) == null || ws.length <= k) { - while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) - tryAwaitMainLock(); - try { - if (workQueues == null) - workQueues = new WorkQueue[submitMask + 1]; - } finally { - if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { - mainLock = 0; - synchronized (this) { notifyAll(); }; - } - } - } - else if ((q = ws[k]) == null) { // create new queue - WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE); - while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) - tryAwaitMainLock(); - try { - int rs = runState; // to update seq - if (ws == workQueues && ws[k] == null) { - ws[k] = nq; - runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN)); - } - } finally { - if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { - mainLock = 0; - synchronized (this) { notifyAll(); }; - } - } - } - else if (q.trySharedPush(task)) { - signalWork(); + final void externalPush(ForkJoinTask task) { + WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask[] a; + if ((z = submitters.get()) != null && plock > 0 && + (ws = workQueues) != null && (m = (ws.length - 1)) >= 0 && + (q = ws[m & z.seed & SQMASK]) != null && + U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock + int b = q.base, s = q.top, n, an; + if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) { + int j = (((an - 1) & s) << ASHIFT) + ABASE; + U.putOrderedObject(a, j, task); + q.top = s + 1; // push on to deque + q.qlock = 0; + if (n <= 2) + signalWork(q); return; } - else if (m > 1) { // move to a different index + q.qlock = 0; + } + fullExternalPush(task); + } + + /** + * Full version of externalPush. This method is called, among + * other times, upon the first submission of the first task to the + * pool, so must perform secondary initialization (via + * initWorkers). It also detects first submission by an external + * thread by looking up its ThreadLocal, and creates a new shared + * queue if the one at index if empty or contended. The plock lock + * body must be exception-free (so no try/finally) so we + * optimistically allocate new queues outside the lock and throw + * them away if (very rarely) not needed. + */ + private void fullExternalPush(ForkJoinTask task) { + int r = 0; // random index seed + for (Submitter z = submitters.get();;) { + WorkQueue[] ws; WorkQueue q; int ps, m, k; + if (z == null) { + if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed, + r += SEED_INCREMENT) && r != 0) + submitters.set(z = new Submitter(r)); + } + else if (r == 0) { // move to a different index + r = z.seed; r ^= r << 13; // same xorshift as WorkQueues r ^= r >>> 17; - s.seed = r ^= r << 5; + z.seed = r ^ (r << 5); + } + else if ((ps = plock) < 0) + throw new RejectedExecutionException(); + else if (ps == 0 || (ws = workQueues) == null || + (m = ws.length - 1) < 0) + initWorkers(); + else if ((q = ws[k = r & m & SQMASK]) != null) { + if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) { + ForkJoinTask[] a = q.array; + int s = q.top; + boolean submitted = false; + try { // locked version of push + if ((a != null && a.length > s + 1 - q.base) || + (a = q.growArray()) != null) { // must presize + int j = (((a.length - 1) & s) << ASHIFT) + ABASE; + U.putOrderedObject(a, j, task); + q.top = s + 1; + submitted = true; + } + } finally { + q.qlock = 0; // unlock + } + if (submitted) { + signalWork(q); + return; + } + } + r = 0; // move on failure + } + else if (((ps = plock) & PL_LOCK) == 0) { // create new queue + q = new WorkQueue(this, null, SHARED_QUEUE, r); + if (((ps = plock) & PL_LOCK) != 0 || + !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) + ps = acquirePlock(); + if ((ws = workQueues) != null && k < ws.length && ws[k] == null) + ws[k] = q; + int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); + if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) + releasePlock(nps); } else - Thread.yield(); // yield if no alternatives + r = 0; // try elsewhere while lock held } } - /** - * Submits the given (non-null) task to the common pool, if possible. - */ - static void submitToCommonPool(ForkJoinTask task) { - ForkJoinPool p; - if ((p = commonPool) == null) - throw new RejectedExecutionException("Common Pool Unavailable"); - p.doSubmit(task); - } - - /** - * Returns true if caller is (or may be) submitter to the common - * pool, and not all workers are active, and there appear to be - * tasks in the associated submission queue. - */ - static boolean canHelpCommonPool() { - ForkJoinPool p; WorkQueue[] ws; WorkQueue q; - int k = submitters.get().seed & SQMASK; - return ((p = commonPool) != null && - (int)(p.ctl >> AC_SHIFT) < 0 && - (ws = p.workQueues) != null && - ws.length > (k &= p.submitMask) && - (q = ws[k]) != null && - q.top - q.base > 0); - } - - /** - * Returns true if the given task was submitted to common pool - * and has not yet commenced execution, and is available for - * removal according to execution policies; if so removing the - * submission from the pool. - * - * @param task the task - * @return true if successful - */ - static boolean tryUnsubmitFromCommonPool(ForkJoinTask task) { - // Peek, looking for task and eligibility before - // using trySharedUnpush to actually take it under lock - ForkJoinPool p; WorkQueue[] ws; WorkQueue q; - ForkJoinTask[] a; int s; - int k = submitters.get().seed & SQMASK; - return ((p = commonPool) != null && - (int)(p.ctl >> AC_SHIFT) < 0 && - (ws = p.workQueues) != null && - ws.length > (k &= p.submitMask) && - (q = ws[k]) != null && - (a = q.array) != null && - (s = q.top - 1) - q.base >= 0 && - s >= 0 && s < a.length && - a[s] == task && - q.trySharedUnpush(task)); - } - - /** - * Tries to pop a task from common pool with given root - */ - static ForkJoinTask popCCFromCommonPool(CountedCompleter root) { - ForkJoinPool p; WorkQueue[] ws; WorkQueue q; - ForkJoinTask t; - int k = submitters.get().seed & SQMASK; - if (root != null && - (p = commonPool) != null && - (int)(p.ctl >> AC_SHIFT) < 0 && - (ws = p.workQueues) != null && - ws.length > (k &= p.submitMask) && - (q = ws[k]) != null && q.top - q.base > 0 && - root.status < 0 && - (t = q.sharedPopCC(root)) != null) - return t; - return null; - } - - // Maintaining ctl counts /** @@ -1664,37 +1617,37 @@ public class ForkJoinPool extends Abstra } /** - * Tries to create one or activate one or more workers if too few are active. + * Tries to create or activate a worker if too few are active. + * + * @param q the (non-null) queue holding tasks to be signalled */ - final void signalWork() { - long c; int u; - while ((u = (int)((c = ctl) >>> 32)) < 0) { // too few active - WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p; - if ((e = (int)c) > 0) { // at least one waiting - if (ws != null && (i = e & SMASK) < ws.length && + final void signalWork(WorkQueue q) { + int hint = q.poolIndex; + long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p; + while ((u = (int)((c = ctl) >>> 32)) < 0) { + if ((e = (int)c) > 0) { + if ((ws = workQueues) != null && ws.length > (i = e & SMASK) && (w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { long nc = (((long)(w.nextWait & E_MASK)) | ((long)(u + UAC_UNIT) << 32)); if (U.compareAndSwapLong(this, CTL, c, nc)) { + w.hint = hint; w.eventCount = (e + E_SEQ) & E_MASK; if ((p = w.parker) != null) - U.unpark(p); // activate and release + U.unpark(p); break; } + if (q.top - q.base <= 0) + break; } else break; } - else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total - long nc = (long)(((u + UTC_UNIT) & UTC_MASK) | - ((u + UAC_UNIT) & UAC_MASK)) << 32; - if (U.compareAndSwapLong(this, CTL, c, nc)) { - addWorker(); - break; - } - } - else + else { + if ((short)u < 0) + tryAddWorker(); break; + } } } @@ -1704,8 +1657,8 @@ public class ForkJoinPool extends Abstra * Top-level runloop for workers, called by ForkJoinWorkerThread.run. */ final void runWorker(WorkQueue w) { - w.growArray(false); // initialize queue array in this thread - do { w.runTask(scan(w)); } while (w.runState >= 0); + w.growArray(); // allocate queue + do { w.runTask(scan(w)); } while (w.qlock >= 0); } /** @@ -1716,115 +1669,93 @@ public class ForkJoinPool extends Abstra * contention, or state changes that indicate possible success on * re-invocation. * - * The scan searches for tasks across a random permutation of - * queues (starting at a random index and stepping by a random - * relative prime, checking each at least once). The scan - * terminates upon either finding a non-empty queue, or completing - * the sweep. If the worker is not inactivated, it takes and - * returns a task from this queue. On failure to find a task, we - * take one of the following actions, after which the caller will - * retry calling this method unless terminated. + * The scan searches for tasks across queues (starting at a random + * index, and relying on registerWorker to irregularly scatter + * them within array to avoid bias), checking each at least twice. + * The scan terminates upon either finding a non-empty queue, or + * completing the sweep. If the worker is not inactivated, it + * takes and returns a task from this queue. Otherwise, if not + * activated, it signals workers (that may include itself) and + * returns so caller can retry. Also returns for true if the + * worker array may have changed during an empty scan. On failure + * to find a task, we take one of the following actions, after + * which the caller will retry calling this method unless + * terminated. * * * If pool is terminating, terminate the worker. * - * * If not a complete sweep, try to release a waiting worker. If - * the scan terminated because the worker is inactivated, then the - * released worker will often be the calling worker, and it can - * succeed obtaining a task on the next call. Or maybe it is - * another worker, but with same net effect. Releasing in other - * cases as well ensures that we have enough workers running. - * * * If not already enqueued, try to inactivate and enqueue the * worker on wait queue. Or, if inactivating has caused the pool * to be quiescent, relay to idleAwaitWork to check for * termination and possibly shrink pool. * - * * If already inactive, and the caller has run a task since the - * last empty scan, return (to allow rescan) unless others are - * also inactivated. Field WorkQueue.rescans counts down on each - * scan to ensure eventual inactivation and blocking. - * - * * If already enqueued and none of the above apply, park - * awaiting signal, + * * If already enqueued and none of the above apply, possibly + * (with 1/2 probability) park awaiting signal, else lingering to + * help scan and signal. * * @param w the worker (via its WorkQueue) * @return a task or null if none found */ private final ForkJoinTask scan(WorkQueue w) { - WorkQueue[] ws; // first update random seed - int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; - int rs = runState, m; // volatile read order matters - if ((ws = workQueues) != null && (m = ws.length - 1) > 0) { - int ec = w.eventCount; // ec is negative if inactive - int step = (r >>> 16) | 1; // relative prime - for (int j = (m + 1) << 2; ; r += step) { - WorkQueue q; ForkJoinTask t; ForkJoinTask[] a; int b; - if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 && - (a = q.array) != null) { // probably nonempty + WorkQueue[] ws; int m; + int ps = plock; // read plock before ws + if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) { + int ec = w.eventCount; // ec is negative if inactive + int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; + int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN; + do { + WorkQueue q; ForkJoinTask[] a; int b; + if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 && + (a = q.array) != null) { // probably nonempty int i = (((a.length - 1) & b) << ASHIFT) + ABASE; - t = (ForkJoinTask)U.getObjectVolatile(a, i); + ForkJoinTask t = (ForkJoinTask) + U.getObjectVolatile(a, i); if (q.base == b && ec >= 0 && t != null && U.compareAndSwapObject(a, i, t, null)) { - if (q.top - (q.base = b + 1) > 0) - signalWork(); // help pushes signal - return t; - } - else if (ec < 0 || j <= m) { - rs = 0; // mark scan as imcomplete - break; // caller can retry after release - } - } - if (--j < 0) - break; - } - - long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns; - if (e < 0) // decode ctl on empty scan - w.runState = -1; // pool is terminating - else if (rs == 0 || rs != runState) { // incomplete scan - WorkQueue v; Thread p; // try to release a waiter - if (e > 0 && a < 0 && w.eventCount == ec && - (v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) { - long nc = ((long)(v.nextWait & E_MASK) | - ((c + AC_UNIT) & (AC_MASK|TC_MASK))); - if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) { - v.eventCount = (e + E_SEQ) & E_MASK; - if ((p = v.parker) != null) - U.unpark(p); - } - } - } - else if (ec >= 0) { // try to enqueue/inactivate - long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); - w.nextWait = e; - w.eventCount = ec | INT_SIGN; // mark as inactive - if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) - w.eventCount = ec; // unmark on CAS failure - else { - if ((ns = w.nsteals) != 0) { - w.nsteals = 0; // set rescans if ran task - w.rescans = (a > 0) ? 0 : a + parallelism; - w.totalSteals += ns; - } - if (a == 1 - parallelism) // quiescent + if ((q.base = b + 1) - q.top < 0) + signalWork(q); + return t; // taken + } + else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) { + w.hint = (r + j) & m; // help signal below + break; // cannot take + } + } + } while (--j >= 0); + + long c, sc; int e, ns, h; + if ((h = w.hint) < 0) { + if ((ns = w.nsteals) != 0) { + if (U.compareAndSwapLong(this, STEALCOUNT, + sc = stealCount, sc + ns)) + w.nsteals = 0; // collect steals + } + else if (plock != ps) // consistency check + ; // skip + else if ((e = (int)(c = ctl)) < 0) + w.qlock = -1; // pool is terminating + else if (ec >= 0) { // try to enqueue/inactivate + long nc = ((long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK))); + w.nextWait = e; // link and mark inactive + w.eventCount = ec | INT_SIGN; + if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) + w.eventCount = ec; // unmark on CAS failure + else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK)) idleAwaitWork(w, nc, c); } - } - else if (w.eventCount < 0) { // already queued - int ac = a + parallelism; - if ((nr = w.rescans) > 0) // continue rescanning - w.rescans = (ac < nr) ? ac : nr - 1; - else if (((w.seed >>> 16) & ac) == 0) { // randomize park - Thread.interrupted(); // clear status + else if (w.eventCount < 0) { // block Thread wt = Thread.currentThread(); + Thread.interrupted(); // clear status U.putObject(wt, PARKBLOCKER, this); - w.parker = wt; // emulate LockSupport.park - if (w.eventCount < 0) // recheck + w.parker = wt; // emulate LockSupport.park + if (w.eventCount < 0) // recheck U.park(false, 0L); w.parker = null; U.putObject(wt, PARKBLOCKER, null); } } + if (h >= 0 || w.hint >= 0) // signal others before retry + helpSignalHint(w); } return null; } @@ -1842,11 +1773,11 @@ public class ForkJoinPool extends Abstra * @param prevCtl the ctl value to restore if thread is terminated */ private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) { - if (w.eventCount < 0 && !tryTerminate(false, false) && - (int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) { + if (w != null && w.eventCount < 0 && + !tryTerminate(false, false) && (int)prevCtl != 0) { int dc = -(short)(currentCtl >>> TC_SHIFT); long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT; - long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop + long deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP; Thread wt = Thread.currentThread(); while (ctl == currentCtl) { Thread.interrupted(); // timed variant of version in scan() @@ -1861,7 +1792,7 @@ public class ForkJoinPool extends Abstra if (deadline - System.nanoTime() <= 0L && U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) { w.eventCount = (w.eventCount + E_SEQ) | E_MASK; - w.runState = -1; // shrink + w.qlock = -1; // shrink break; } } @@ -1869,6 +1800,82 @@ public class ForkJoinPool extends Abstra } /** + * Scans through queues looking for work while joining a task; if + * any present, signals. May return early if more signalling is + * detectably unneeded. + * + * @param task return early if done + * @param origin an index to start scan + */ + private void helpSignal(ForkJoinTask task, int origin) { + WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s; + if (task != null && task.status >= 0 && + (u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 && + (ws = workQueues) != null && (m = ws.length - 1) >= 0) { + outer: for (int k = origin, j = m; j >= 0; --j) { + WorkQueue q = ws[k++ & m]; + for (int n = m;;) { // limit to at most m signals + if (task.status < 0) + break outer; + if (q == null || + ((s = -q.base + q.top) <= n && (n = s) <= 0)) + break; + if ((u = (int)((c = ctl) >>> 32)) >= 0 || + (e = (int)c) <= 0 || m < (i = e & SMASK) || + (w = ws[i]) == null) + break outer; + long nc = (((long)(w.nextWait & E_MASK)) | + ((long)(u + UAC_UNIT) << 32)); + if (w.eventCount == (e | INT_SIGN) && + U.compareAndSwapLong(this, CTL, c, nc)) { + w.eventCount = (e + E_SEQ) & E_MASK; + if ((p = w.parker) != null) + U.unpark(p); + if (--n <= 0) + break; + } + } + } + } + } + + /** + * Signals other workers if tasks are present in hinted queue. + * + * @param caller the worker with the hint + */ + private void helpSignalHint(WorkQueue caller) { + WorkQueue[] ws; WorkQueue q, w; Thread p; long c; int h, m, u, e, i, s; + if (caller != null && (h = caller.hint) >= 0) { + caller.hint = -1; + if ((u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 && + (ws = workQueues) != null && (m = ws.length - 1) >= 0 && + (q = ws[h & m]) != null) { + for (int n = 2;;) { // limit to at most 2 signals + int idleCount = (caller.eventCount < 0) ? 0 : -1; + if (((s = idleCount - q.base + q.top) <= n && + (n = s) <= 0) || + (u = (int)((c = ctl) >>> 32)) >= 0 || + (e = (int)c) <= 0 || m < (i = e & SMASK) || + (w = ws[i]) == null) + break; + long nc = (((long)(w.nextWait & E_MASK)) | + ((long)(u + UAC_UNIT) << 32)); + if (w.eventCount == (e | INT_SIGN) && + U.compareAndSwapLong(this, CTL, c, nc)) { + w.hint = h; + w.eventCount = (e + E_SEQ) & E_MASK; + if ((p = w.parker) != null) + U.unpark(p); + if (--n <= 0) + break; + } + } + } + } + } + + /** * Tries to locate and execute tasks for a stealer of the given * task, or in turn one of its stealers, Traces currentSteal -> * currentJoin links looking for a thread working on a descendant @@ -1899,7 +1906,7 @@ public class ForkJoinPool extends Abstra } if ((ws = workQueues) == null || (m = ws.length - 1) <= 0) break restart; // shutting down - if ((v = ws[h = (j.stealHint | 1) & m]) == null || + if ((v = ws[h = (j.hint | 1) & m]) == null || v.currentSteal != subtask) { for (int origin = h;;) { // find stealer if (((h = (h + 2) & m) & 15) == 1 && @@ -1907,7 +1914,7 @@ public class ForkJoinPool extends Abstra continue restart; // occasional staleness check if ((v = ws[h]) != null && v.currentSteal == subtask) { - j.stealHint = h; // save hint + j.hint = h; // save hint break; } if (h == origin) @@ -1955,88 +1962,78 @@ public class ForkJoinPool extends Abstra } /** - * If task is at base of some steal queue, steals and executes it. + * Analog of tryHelpStealer for CountedCompleters. Tries to steal + * and run tasks within the target's computation. + * + * @param task the task to join + * @param mode if shared, exit upon completing any task + * if all workers are active * - * @param joiner the joining worker - * @param task the task */ - private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask task) { - WorkQueue[] ws; - if ((ws = workQueues) != null) { - for (int j = 1; j < ws.length && task.status >= 0; j += 2) { - WorkQueue q = ws[j]; - if (q != null && q.pollFor(task)) { - joiner.runSubtask(task); - break; + private int helpComplete(ForkJoinTask task, int mode) { + WorkQueue[] ws; WorkQueue q; int m, n, s, u; + if (task != null && (ws = workQueues) != null && + (m = ws.length - 1) >= 0) { + for (int j = 1, origin = j;;) { + if ((s = task.status) < 0) + return s; + if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) { + origin = j; + if (mode == SHARED_QUEUE && + ((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)) + break; } + else if ((j = (j + 2) & m) == origin) + break; } } + return 0; } /** * Tries to decrement active count (sometimes implicitly) and * possibly release or create a compensating worker in preparation * for blocking. Fails on contention or termination. Otherwise, - * adds a new thread if no idle workers are available and either - * pool would become completely starved or: (at least half - * starved, and fewer than 50% spares exist, and there is at least - * one task apparently available). Even though the availability - * check requires a full scan, it is worthwhile in reducing false - * alarms. - * - * @param task if non-null, a task being waited for - * @param blocker if non-null, a blocker being waited for - * @return true if the caller can block, else should recheck and retry + * adds a new thread if no idle workers are available and pool + * may become starved. */ - final boolean tryCompensate(ForkJoinTask task, ManagedBlocker blocker) { - int pc = parallelism, e; - long c = ctl; - WorkQueue[] ws = workQueues; - if ((e = (int)c) >= 0 && ws != null) { - int u, a, ac, hc; - int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc; - boolean replace = false; - if ((a = u >> UAC_SHIFT) <= 0) { - if ((ac = a + pc) <= 1) - replace = true; - else if ((e > 0 || (task != null && - ac <= (hc = pc >>> 1) && tc < pc + hc))) { - WorkQueue w; - for (int j = 0; j < ws.length; ++j) { - if ((w = ws[j]) != null && !w.isEmpty()) { - replace = true; - break; // in compensation range and tasks available - } - } + final boolean tryCompensate() { + int pc = config & SMASK, e, i, tc; long c; + WorkQueue[] ws; WorkQueue w; Thread p; + if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) { + if (e != 0 && (i = e & SMASK) < ws.length && + (w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { + long nc = ((long)(w.nextWait & E_MASK) | + (c & (AC_MASK|TC_MASK))); + if (U.compareAndSwapLong(this, CTL, c, nc)) { + w.eventCount = (e + E_SEQ) & E_MASK; + if ((p = w.parker) != null) + U.unpark(p); + return true; // replace with idle worker } } - if ((task == null || task.status >= 0) && // recheck need to block - (blocker == null || !blocker.isReleasable()) && ctl == c) { - if (!replace) { // no compensation - long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); - if (U.compareAndSwapLong(this, CTL, c, nc)) - return true; - } - else if (e != 0) { // release an idle worker - WorkQueue w; Thread p; int i; - if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) { - long nc = ((long)(w.nextWait & E_MASK) | - (c & (AC_MASK|TC_MASK))); - if (w.eventCount == (e | INT_SIGN) && - U.compareAndSwapLong(this, CTL, c, nc)) { - w.eventCount = (e + E_SEQ) & E_MASK; - if ((p = w.parker) != null) - U.unpark(p); + else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 && + (int)(c >> AC_SHIFT) + pc > 1) { + long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); + if (U.compareAndSwapLong(this, CTL, c, nc)) + return true; // no compensation + } + else if (tc + pc < MAX_CAP) { + long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); + if (U.compareAndSwapLong(this, CTL, c, nc)) { + ForkJoinWorkerThreadFactory fac; + Throwable ex = null; + ForkJoinWorkerThread wt = null; + try { + if ((fac = factory) != null && + (wt = fac.newThread(this)) != null) { + wt.start(); return true; } + } catch (Throwable rex) { + ex = rex; } - } - else if (tc < MAX_CAP) { // create replacement - long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); - if (U.compareAndSwapLong(this, CTL, c, nc)) { - addWorker(); - return true; - } + deregisterWorker(wt, ex); // clean up and return false } } } @@ -2051,25 +2048,25 @@ public class ForkJoinPool extends Abstra * @return task status on exit */ final int awaitJoin(WorkQueue joiner, ForkJoinTask task) { - int s; - if ((s = task.status) >= 0) { + int s = 0; + if (joiner != null && task != null && (s = task.status) >= 0) { ForkJoinTask prevJoin = joiner.currentJoin; joiner.currentJoin = task; - long startTime = 0L; - for (int k = 0;;) { - if ((s = (joiner.isEmpty() ? // try to help - tryHelpStealer(joiner, task) : - joiner.tryRemoveAndExec(task))) == 0 && + do {} while ((s = task.status) >= 0 && !joiner.isEmpty() && + joiner.tryRemoveAndExec(task)); // process local tasks + if (s >= 0 && (s = task.status) >= 0) { + helpSignal(task, joiner.poolIndex); + if ((s = task.status) >= 0 && + (task instanceof CountedCompleter)) + s = helpComplete(task, LIFO_QUEUE); + } + while (s >= 0 && (s = task.status) >= 0) { + if ((!joiner.isEmpty() || // try helping + (s = tryHelpStealer(joiner, task)) == 0) && (s = task.status) >= 0) { - if (k == 0) { - startTime = System.nanoTime(); - tryPollForAndExec(joiner, task); // check uncommon case - } - else if ((k & (MAX_HELP - 1)) == 0 && - System.nanoTime() - startTime >= - COMPENSATION_DELAY && - tryCompensate(task, null)) { - if (task.trySetSignal()) { + helpSignal(task, joiner.poolIndex); + if ((s = task.status) >= 0 && tryCompensate()) { + if (task.trySetSignal() && (s = task.status) >= 0) { synchronized (task) { if (task.status >= 0) { try { // see ForkJoinTask @@ -2086,13 +2083,8 @@ public class ForkJoinPool extends Abstra (this, CTL, c = ctl, c + AC_UNIT)); } } - if (s < 0 || (s = task.status) < 0) { - joiner.currentJoin = prevJoin; - break; - } - else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1) - Thread.yield(); // for politeness } + joiner.currentJoin = prevJoin; } return s; } @@ -2104,16 +2096,26 @@ public class ForkJoinPool extends Abstra * * @param joiner the joining worker * @param task the task - * @return task status on exit */ - final int helpJoinOnce(WorkQueue joiner, ForkJoinTask task) { + final void helpJoinOnce(WorkQueue joiner, ForkJoinTask task) { int s; - while ((s = task.status) >= 0 && - (joiner.isEmpty() ? - tryHelpStealer(joiner, task) : - joiner.tryRemoveAndExec(task)) != 0) - ; - return s; + if (joiner != null && task != null && (s = task.status) >= 0) { + ForkJoinTask prevJoin = joiner.currentJoin; + joiner.currentJoin = task; + do {} while ((s = task.status) >= 0 && !joiner.isEmpty() && + joiner.tryRemoveAndExec(task)); + if (s >= 0 && (s = task.status) >= 0) { + helpSignal(task, joiner.poolIndex); + if ((s = task.status) >= 0 && + (task instanceof CountedCompleter)) + s = helpComplete(task, LIFO_QUEUE); + } + if (s >= 0 && joiner.isEmpty()) { + do {} while (task.status >= 0 && + tryHelpStealer(joiner, task) > 0); + } + joiner.currentJoin = prevJoin; + } } /** @@ -2121,21 +2123,22 @@ public class ForkJoinPool extends Abstra * during a random, then cyclic scan, else null. This method must * be retried by caller if, by the time it tries to use the queue, * it is empty. + * @param r a (random) seed for scanning */ - private WorkQueue findNonEmptyStealQueue(WorkQueue w) { - // Similar to loop in scan(), but ignoring submissions - int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; - int step = (r >>> 16) | 1; + private WorkQueue findNonEmptyStealQueue(int r) { for (WorkQueue[] ws;;) { - int rs = runState, m; + int ps = plock, m, n; if ((ws = workQueues) == null || (m = ws.length - 1) < 1) return null; - for (int j = (m + 1) << 2; ; r += step) { - WorkQueue q = ws[((r << 1) | 1) & m]; - if (q != null && !q.isEmpty()) + for (int j = (m + 1) << 2; ;) { + WorkQueue q = ws[(((r + j) << 1) | 1) & m]; + if (q != null && (n = q.base - q.top) < 0) { + if (n < -1) + signalWork(q); return q; + } else if (--j < 0) { - if (runState == rs) + if (plock == ps) return null; break; } @@ -2154,7 +2157,8 @@ public class ForkJoinPool extends Abstra ForkJoinTask localTask; // exhaust local queue while ((localTask = w.nextLocalTask()) != null) localTask.doExec(); - WorkQueue q = findNonEmptyStealQueue(w); + // Similar to loop in scan(), but ignoring submissions + WorkQueue q = findNonEmptyStealQueue(w.nextSeed()); if (q != null) { ForkJoinTask t; int b; if (!active) { // re-establish active count @@ -2175,7 +2179,7 @@ public class ForkJoinPool extends Abstra } else c = ctl; // re-increment on exit - if ((int)(c >> AC_SHIFT) + parallelism == 0) { + if ((int)(c >> AC_SHIFT) + (config & SMASK) == 0) { do {} while (!U.compareAndSwapLong (this, CTL, c = ctl, c + AC_UNIT)); break; @@ -2185,31 +2189,6 @@ public class ForkJoinPool extends Abstra } /** - * Restricted version of helpQuiescePool for non-FJ callers - */ - static void externalHelpQuiescePool() { - ForkJoinPool p; WorkQueue[] ws; WorkQueue q, sq; - ForkJoinTask[] a; int b; - ForkJoinTask t = null; - int k = submitters.get().seed & SQMASK; - if ((p = commonPool) != null && - (int)(p.ctl >> AC_SHIFT) < 0 && - (ws = p.workQueues) != null && - ws.length > (k &= p.submitMask) && - (q = ws[k]) != null) { - while (q.top - q.base > 0) { - if ((t = q.sharedPop()) != null) - break; - } - if (t == null && (sq = p.findNonEmptyStealQueue(q)) != null && - (b = sq.base) - sq.top < 0) - t = sq.pollAt(b); - if (t != null) - t.doExec(); - } - } - - /** * Gets and removes a local or stolen task for the given worker. * * @return a task, if available @@ -2219,7 +2198,7 @@ public class ForkJoinPool extends Abstra WorkQueue q; int b; if ((t = w.nextLocalTask()) != null) return t; - if ((q = findNonEmptyStealQueue(w)) == null) + if ((q = findNonEmptyStealQueue(w.nextSeed())) == null) return null; if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) return t; @@ -2227,33 +2206,64 @@ public class ForkJoinPool extends Abstra } /** - * Returns the approximate (non-atomic) number of idle threads per - * active thread to offset steal queue size for method - * ForkJoinTask.getSurplusQueuedTaskCount(). - */ - final int idlePerActive() { - // Approximate at powers of two for small values, saturate past 4 - int p = parallelism; - int a = p + (int)(ctl >> AC_SHIFT); - return (a > (p >>>= 1) ? 0 : - a > (p >>>= 1) ? 1 : - a > (p >>>= 1) ? 2 : - a > (p >>>= 1) ? 4 : - 8); - } - - /** - * Returns approximate submission queue length for the given caller - */ - static int getEstimatedSubmitterQueueLength() { - ForkJoinPool p; WorkQueue[] ws; WorkQueue q; - int k = submitters.get().seed & SQMASK; - return ((p = commonPool) != null && - p.runState >= 0 && - (ws = p.workQueues) != null && - ws.length > (k &= p.submitMask) && - (q = ws[k]) != null) ? - q.queueSize() : 0; + * Returns a cheap heuristic guide for task partitioning when + * programmers, frameworks, tools, or languages have little or no + * idea about task granularity. In essence by offering this + * method, we ask users only about tradeoffs in overhead vs + * expected throughput and its variance, rather than how finely to + * partition tasks. + * + * In a steady state strict (tree-structured) computation, each + * thread makes available for stealing enough tasks for other + * threads to remain active. Inductively, if all threads play by + * the same rules, each thread should make available only a + * constant number of tasks. + * + * The minimum useful constant is just 1. But using a value of 1 + * would require immediate replenishment upon each steal to + * maintain enough tasks, which is infeasible. Further, + * partitionings/granularities of offered tasks should minimize + * steal rates, which in general means that threads nearer the top + * of computation tree should generate more than those nearer the + * bottom. In perfect steady state, each thread is at + * approximately the same level of computation tree. However, + * producing extra tasks amortizes the uncertainty of progress and + * diffusion assumptions. + * + * So, users will want to use values larger, but not much larger + * than 1 to both smooth over transient shortages and hedge + * against uneven progress; as traded off against the cost of + * extra task overhead. We leave the user to pick a threshold + * value to compare with the results of this call to guide + * decisions, but recommend values such as 3. + * + * When all threads are active, it is on average OK to estimate + * surplus strictly locally. In steady-state, if one thread is + * maintaining say 2 surplus tasks, then so are others. So we can + * just use estimated queue length. However, this strategy alone + * leads to serious mis-estimates in some non-steady-state + * conditions (ramp-up, ramp-down, other stalls). We can detect + * many of these by further considering the number of "idle" + * threads, that are known to have zero queued tasks, so + * compensate by a factor of (#idle/#active) threads. + * + * Note: The approximation of #busy workers as #active workers is + * not very good under current signalling scheme, and should be + * improved. + */ + static int getSurplusQueuedTaskCount() { + Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; + if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) { + int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK; + int n = (q = wt.workQueue).top - q.base; + int a = (int)(pool.ctl >> AC_SHIFT) + p; + return n - (a > (p >>>= 1) ? 0 : + a > (p >>>= 1) ? 1 : + a > (p >>>= 1) ? 2 : + a > (p >>>= 1) ? 4 : + 8); + } + return 0; } // Termination @@ -2273,31 +2283,30 @@ public class ForkJoinPool extends Abstra * @return true if now terminating or terminated */ private boolean tryTerminate(boolean now, boolean enable) { + if (this == commonPool) // cannot shut down + return false; for (long c;;) { if (((c = ctl) & STOP_BIT) != 0) { // already terminating - if ((short)(c >>> TC_SHIFT) == -parallelism) { - synchronized(this) { + if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) { + synchronized (this) { notifyAll(); // signal when 0 workers } } return true; } - if (runState >= 0) { // not yet enabled + if (plock >= 0) { // not yet enabled + int ps; if (!enable) return false; - while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) - tryAwaitMainLock(); - try { - runState |= SHUTDOWN; - } finally { - if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { - mainLock = 0; - synchronized (this) { notifyAll(); }; - } - } + if (((ps = plock) & PL_LOCK) != 0 || + !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) + ps = acquirePlock(); + int nps = SHUTDOWN; + if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) + releasePlock(nps); } if (!now) { // check if idle & no tasks - if ((int)(c >> AC_SHIFT) != -parallelism || + if ((int)(c >> AC_SHIFT) != -(config & SMASK) || hasQueuedSubmissions()) return false; // Check for unqueued inactive workers. One pass suffices. @@ -2317,7 +2326,7 @@ public class ForkJoinPool extends Abstra int n = ws.length; for (int i = 0; i < n; ++i) { if ((w = ws[i]) != null) { - w.runState = -1; + w.qlock = -1; if (pass > 0) { w.cancelAll(); if (pass > 1) @@ -2336,7 +2345,7 @@ public class ForkJoinPool extends Abstra if (w.eventCount == (e | INT_SIGN) && U.compareAndSwapLong(this, CTL, cc, nc)) { w.eventCount = (e + E_SEQ) & E_MASK; - w.runState = -1; + w.qlock = -1; if ((p = w.parker) != null) U.unpark(p); } @@ -2347,6 +2356,147 @@ public class ForkJoinPool extends Abstra } } + // external operations on common pool + + /** + * Returns common pool queue for a thread that has submitted at + * least one task. + */ + static WorkQueue commonSubmitterQueue() { + ForkJoinPool p; WorkQueue[] ws; int m; Submitter z; + return ((z = submitters.get()) != null && + (p = commonPool) != null && + (ws = p.workQueues) != null && + (m = ws.length - 1) >= 0) ? + ws[m & z.seed & SQMASK] : null; + } + + /** + * Tries to pop the given task from submitter's queue in common pool. + */ + static boolean tryExternalUnpush(ForkJoinTask t) { + ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z; + ForkJoinTask[] a; int m, s; + if (t != null && + (z = submitters.get()) != null && + (p = commonPool) != null && + (ws = p.workQueues) != null && + (m = ws.length - 1) >= 0 && + (q = ws[m & z.seed & SQMASK]) != null && + (s = q.top) != q.base && + (a = q.array) != null) { + long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE; + if (U.getObject(a, j) == t && + U.compareAndSwapInt(q, QLOCK, 0, 1)) { + if (q.array == a && q.top == s && // recheck + U.compareAndSwapObject(a, j, t, null)) { + q.top = s - 1; + q.qlock = 0; + return true; + } + q.qlock = 0; + } + } + return false; + } + + /** + * Tries to pop and run local tasks within the same computation + * as the given root. On failure, tries to help complete from + * other queues via helpComplete. + */ + private void externalHelpComplete(WorkQueue q, ForkJoinTask root) { + ForkJoinTask[] a; int m; + if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 && + root != null && root.status >= 0) { + for (;;) { + int s, u; Object o; CountedCompleter task = null; + if ((s = q.top) - q.base > 0) { + long j = ((m & (s - 1)) << ASHIFT) + ABASE; + if ((o = U.getObject(a, j)) != null && + (o instanceof CountedCompleter)) { + CountedCompleter t = (CountedCompleter)o, r = t; + do { + if (r == root) { + if (U.compareAndSwapInt(q, QLOCK, 0, 1)) { + if (q.array == a && q.top == s && + U.compareAndSwapObject(a, j, t, null)) { + q.top = s - 1; + task = t; + } + q.qlock = 0; + } + break; + } + } while ((r = r.completer) != null); + } + } + if (task != null) + task.doExec(); + if (root.status < 0 || + (u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0) + break; + if (task == null) { + helpSignal(root, q.poolIndex); + if (root.status >= 0) + helpComplete(root, SHARED_QUEUE); + break; + } + } + } + } + + /** + * Tries to help execute or signal availability of the given task + * from submitter's queue in common pool. + */ + static void externalHelpJoin(ForkJoinTask t) { + // Some hard-to-avoid overlap with tryExternalUnpush + ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z; + ForkJoinTask[] a; int m, s, n; + if (t != null && + (z = submitters.get()) != null && + (p = commonPool) != null && + (ws = p.workQueues) != null && + (m = ws.length - 1) >= 0 && + (q = ws[m & z.seed & SQMASK]) != null && + (a = q.array) != null) { + int am = a.length - 1; + if ((s = q.top) != q.base) { + long j = ((am & (s - 1)) << ASHIFT) + ABASE; + if (U.getObject(a, j) == t && + U.compareAndSwapInt(q, QLOCK, 0, 1)) { + if (q.array == a && q.top == s && + U.compareAndSwapObject(a, j, t, null)) { + q.top = s - 1; + q.qlock = 0; + t.doExec(); + } + else + q.qlock = 0; + } + } + if (t.status >= 0) { + if (t instanceof CountedCompleter) + p.externalHelpComplete(q, t); + else + p.helpSignal(t, q.poolIndex); + } + } + } + + /** + * Restricted version of helpQuiescePool for external callers + */ + static void externalHelpQuiescePool() { + ForkJoinPool p; ForkJoinTask t; WorkQueue q; int b; + if ((p = commonPool) != null && + (q = p.findNonEmptyStealQueue(1)) != null && + (b = q.base) - q.top < 0 && + (t = q.pollAt(b)) != null) + t.doExec(); + } + // Exported methods // Constructors @@ -2418,40 +2568,30 @@ public class ForkJoinPool extends Abstra throw new NullPointerException(); if (parallelism <= 0 || parallelism > MAX_CAP) throw new IllegalArgumentException(); - this.parallelism = parallelism; this.factory = factory; this.ueh = handler; - this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE; + this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0); long np = (long)(-parallelism); // offset ctl counts this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); - // Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2. - int n = parallelism - 1; - n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; - this.submitMask = ((n + 1) << 1) - 1; - int pn = poolNumberGenerator.incrementAndGet(); + int pn = nextPoolId(); StringBuilder sb = new StringBuilder("ForkJoinPool-"); sb.append(Integer.toString(pn)); sb.append("-worker-"); this.workerNamePrefix = sb.toString(); - this.runState = 1; // set init flag } /** * Constructor for common pool, suitable only for static initialization. * Basically the same as above, but uses smallest possible initial footprint. */ - ForkJoinPool(int parallelism, int submitMask, + ForkJoinPool(int parallelism, long ctl, ForkJoinWorkerThreadFactory factory, Thread.UncaughtExceptionHandler handler) { + this.config = parallelism; + this.ctl = ctl; this.factory = factory; this.ueh = handler; - this.submitMask = submitMask; - this.parallelism = parallelism; - long np = (long)(-parallelism); - this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); - this.localMode = LIFO_QUEUE; this.workerNamePrefix = "ForkJoinPool.commonPool-worker-"; - this.runState = 1; } /** @@ -2460,10 +2600,8 @@ public class ForkJoinPool extends Abstra * @return the common pool instance */ public static ForkJoinPool commonPool() { - ForkJoinPool p; - if ((p = commonPool) == null) - throw new Error("Common Pool Unavailable"); - return p; + // assert commonPool != null : "static init error"; + return commonPool; } // Execution methods @@ -2487,7 +2625,7 @@ public class ForkJoinPool extends Abstra public T invoke(ForkJoinTask task) { if (task == null) throw new NullPointerException(); - doSubmit(task); + externalPush(task); return task.join(); } @@ -2502,7 +2640,7 @@ public class ForkJoinPool extends Abstra public void execute(ForkJoinTask task) { if (task == null) throw new NullPointerException(); - doSubmit(task); + externalPush(task); } // AbstractExecutorService methods @@ -2520,7 +2658,7 @@ public class ForkJoinPool extends Abstra job = (ForkJoinTask) task; else job = new ForkJoinTask.AdaptedRunnableAction(task); - doSubmit(job); + externalPush(job); } /** @@ -2535,7 +2673,7 @@ public class ForkJoinPool extends Abstra public ForkJoinTask submit(ForkJoinTask task) { if (task == null) throw new NullPointerException(); - doSubmit(task); + externalPush(task); return task; } @@ -2546,7 +2684,7 @@ public class ForkJoinPool extends Abstra */ public ForkJoinTask submit(Callable task) { ForkJoinTask job = new ForkJoinTask.AdaptedCallable(task); - doSubmit(job); + externalPush(job); return job; } @@ -2557,7 +2695,7 @@ public class ForkJoinPool extends Abstra */ public ForkJoinTask submit(Runnable task, T result) { ForkJoinTask job = new ForkJoinTask.AdaptedRunnable(task, result); - doSubmit(job); + externalPush(job); return job; } @@ -2574,7 +2712,7 @@ public class ForkJoinPool extends Abstra job = (ForkJoinTask) task; else job = new ForkJoinTask.AdaptedRunnableAction(task); - doSubmit(job); + externalPush(job); return job; } @@ -2596,7 +2734,7 @@ public class ForkJoinPool extends Abstra try { for (Callable t : tasks) { ForkJoinTask f = new ForkJoinTask.AdaptedCallable(t); - doSubmit(f); + externalPush(f); fs.add(f); } for (ForkJoinTask f : fs) @@ -2635,7 +2773,7 @@ public class ForkJoinPool extends Abstra * @return the targeted parallelism level of this pool */ public int getParallelism() { - return parallelism; + return config & SMASK; } /** @@ -2656,7 +2794,7 @@ public class ForkJoinPool extends Abstra * @return the number of worker threads */ public int getPoolSize() { - return parallelism + (short)(ctl >>> TC_SHIFT); + return (config & SMASK) + (short)(ctl >>> TC_SHIFT); } /** @@ -2666,7 +2804,7 @@ public class ForkJoinPool extends Abstra * @return {@code true} if this pool uses async mode */ public boolean getAsyncMode() { - return localMode != 0; + return (config >>> 16) == FIFO_QUEUE; } /** @@ -2697,7 +2835,7 @@ public class ForkJoinPool extends Abstra * @return the number of active threads */ public int getActiveThreadCount() { - int r = parallelism + (int)(ctl >> AC_SHIFT); + int r = (config & SMASK) + (int)(ctl >> AC_SHIFT); return (r <= 0) ? 0 : r; // suppress momentarily negative values } @@ -2713,7 +2851,7 @@ public class ForkJoinPool extends Abstra * @return {@code true} if all threads are currently idle */ public boolean isQuiescent() { - return (int)(ctl >> AC_SHIFT) + parallelism == 0; + return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0; } /** @@ -2733,7 +2871,7 @@ public class ForkJoinPool extends Abstra if ((ws = workQueues) != null) { for (int i = 1; i < ws.length; i += 2) { if ((w = ws[i]) != null) - count += w.totalSteals; + count += w.nsteals; } } return count; @@ -2869,14 +3007,14 @@ public class ForkJoinPool extends Abstra qs += size; else { qt += size; - st += w.totalSteals; + st += w.nsteals; if (w.isApparentlyUnblocked()) ++rc; } } } } - int pc = parallelism; + int pc = (config & SMASK); int tc = pc + (short)(c >>> TC_SHIFT); int ac = pc + (int)(c >> AC_SHIFT); if (ac < 0) // ignore transient negative @@ -2885,7 +3023,7 @@ public class ForkJoinPool extends Abstra if ((c & STOP_BIT) != 0) level = (tc == 0) ? "Terminated" : "Terminating"; else - level = runState < 0 ? "Shutting down" : "Running"; + level = plock < 0 ? "Shutting down" : "Running"; return super.toString() + "[" + level + ", parallelism = " + pc + @@ -2914,8 +3052,7 @@ public class ForkJoinPool extends Abstra */ public void shutdown() { checkPermission(); - if (this != commonPool) - tryTerminate(false, true); + tryTerminate(false, true); } /** @@ -2938,8 +3075,7 @@ public class ForkJoinPool extends Abstra */ public List shutdownNow() { checkPermission(); - if (this != commonPool) - tryTerminate(true, true); + tryTerminate(true, true); return Collections.emptyList(); } @@ -2951,7 +3087,7 @@ public class ForkJoinPool extends Abstra public boolean isTerminated() { long c = ctl; return ((c & STOP_BIT) != 0L && - (short)(c >>> TC_SHIFT) == -parallelism); + (short)(c >>> TC_SHIFT) == -(config & SMASK)); } /** @@ -2959,7 +3095,7 @@ public class ForkJoinPool extends Abstra * commenced but not yet completed. This method may be useful for * debugging. A return of {@code true} reported a sufficient * period after shutdown may indicate that submitted tasks have - * ignored or suppressed interruption, or are waiting for IO, + * ignored or suppressed interruption, or are waiting for I/O, * causing this executor not to properly terminate. (See the * advisory notes for class {@link ForkJoinTask} stating that * tasks should not normally entail blocking operations. But if @@ -2970,7 +3106,7 @@ public class ForkJoinPool extends Abstra public boolean isTerminating() { long c = ctl; return ((c & STOP_BIT) != 0L && - (short)(c >>> TC_SHIFT) != -parallelism); + (short)(c >>> TC_SHIFT) != -(config & SMASK)); } /** @@ -2979,13 +3115,15 @@ public class ForkJoinPool extends Abstra * @return {@code true} if this pool has been shut down */ public boolean isShutdown() { - return runState < 0; + return plock < 0; } /** - * Blocks until all tasks have completed execution after a shutdown - * request, or the timeout occurs, or the current thread is - * interrupted, whichever happens first. + * Blocks until all tasks have completed execution after a + * shutdown request, or the timeout occurs, or the current thread + * is interrupted, whichever happens first. Note that the {@link + * #commonPool()} never terminates until program shutdown so + * this method will always time out. * * @param timeout the maximum time to wait * @param unit the time unit of the timeout argument @@ -3000,7 +3138,7 @@ public class ForkJoinPool extends Abstra return true; long startTime = System.nanoTime(); boolean terminated = false; - synchronized(this) { + synchronized (this) { for (long waitTime = nanos, millis = 0L;;) { if (terminated = isTerminated() || waitTime <= 0L || @@ -3109,19 +3247,37 @@ public class ForkJoinPool extends Abstra 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(null, blocker)) { - try { - do {} while (!blocker.isReleasable() && !blocker.block()); - } finally { - if (p != null) + if (t instanceof ForkJoinWorkerThread) { + ForkJoinPool p = ((ForkJoinWorkerThread)t).pool; + while (!blocker.isReleasable()) { // variant of helpSignal + WorkQueue[] ws; WorkQueue q; int m, u; + if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) { + for (int i = 0; i <= m; ++i) { + if (blocker.isReleasable()) + return; + if ((q = ws[i]) != null && q.base - q.top < 0) { + p.signalWork(q); + if ((u = (int)(p.ctl >>> 32)) >= 0 || + (u >> UAC_SHIFT) >= 0) + break; + } + } + } + if (p.tryCompensate()) { + try { + do {} while (!blocker.isReleasable() && + !blocker.block()); + } finally { p.incrementActiveCount(); + } + break; } - break; } } + else { + do {} while (!blocker.isReleasable() && + !blocker.block()); + } } // AbstractExecutorService overrides. These rely on undocumented @@ -3142,33 +3298,31 @@ public class ForkJoinPool extends Abstra private static final long PARKBLOCKER; private static final int ABASE; private static final int ASHIFT; - private static final long NEXTWORKERNUMBER; private static final long STEALCOUNT; - private static final long MAINLOCK; + private static final long PLOCK; + private static final long INDEXSEED; + private static final long QLOCK; static { - poolNumberGenerator = new AtomicInteger(); - nextSubmitterSeed = new AtomicInteger(0x55555555); - modifyThreadPermission = new RuntimePermission("modifyThread"); - defaultForkJoinWorkerThreadFactory = - new DefaultForkJoinWorkerThreadFactory(); - submitters = new ThreadSubmitter(); - int s; + int s; // initialize field offsets for CAS etc try { U = getUnsafe(); Class k = ForkJoinPool.class; - Class ak = ForkJoinTask[].class; CTL = U.objectFieldOffset (k.getDeclaredField("ctl")); - NEXTWORKERNUMBER = U.objectFieldOffset - (k.getDeclaredField("nextWorkerNumber")); STEALCOUNT = U.objectFieldOffset (k.getDeclaredField("stealCount")); - MAINLOCK = U.objectFieldOffset - (k.getDeclaredField("mainLock")); + PLOCK = U.objectFieldOffset + (k.getDeclaredField("plock")); + INDEXSEED = U.objectFieldOffset + (k.getDeclaredField("indexSeed")); Class tk = Thread.class; PARKBLOCKER = U.objectFieldOffset (tk.getDeclaredField("parkBlocker")); + Class wk = WorkQueue.class; + QLOCK = U.objectFieldOffset + (wk.getDeclaredField("qlock")); + Class ak = ForkJoinTask[].class; ABASE = U.arrayBaseOffset(ak); s = U.arrayIndexScale(ak); ASHIFT = 31 - Integer.numberOfLeadingZeros(s); @@ -3177,31 +3331,47 @@ public class ForkJoinPool extends Abstra } if ((s & (s-1)) != 0) throw new Error("data type scale not a power of two"); - try { // Establish common pool - String pp = System.getProperty(propPrefix + "parallelism"); - String fp = System.getProperty(propPrefix + "threadFactory"); - String up = System.getProperty(propPrefix + "exceptionHandler"); - ForkJoinWorkerThreadFactory fac = (fp == null) ? - defaultForkJoinWorkerThreadFactory : - ((ForkJoinWorkerThreadFactory)ClassLoader. - getSystemClassLoader().loadClass(fp).newInstance()); - Thread.UncaughtExceptionHandler ueh = (up == null)? null : - ((Thread.UncaughtExceptionHandler)ClassLoader. - getSystemClassLoader().loadClass(up).newInstance()); - int par; - if ((pp == null || (par = Integer.parseInt(pp)) <= 0)) - par = Runtime.getRuntime().availableProcessors(); - if (par > MAX_CAP) - par = MAX_CAP; - commonPoolParallelism = par; - int n = par - 1; // precompute submit mask - n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; - n |= n >>> 8; n |= n >>> 16; - int mask = ((n + 1) << 1) - 1; - commonPool = new ForkJoinPool(par, mask, fac, ueh); - } catch (Exception e) { - throw new Error(e); - } + + submitters = new ThreadLocal(); + ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory = + new DefaultForkJoinWorkerThreadFactory(); + modifyThreadPermission = new RuntimePermission("modifyThread"); + + /* + * Establish common pool parameters. For extra caution, + * computations to set up common pool state are here; the + * constructor just assigns these values to fields. + */ + + int par = 0; + Thread.UncaughtExceptionHandler handler = null; + try { // TBD: limit or report ignored exceptions? + String pp = System.getProperty + ("java.util.concurrent.ForkJoinPool.common.parallelism"); + String hp = System.getProperty + ("java.util.concurrent.ForkJoinPool.common.exceptionHandler"); + String fp = System.getProperty + ("java.util.concurrent.ForkJoinPool.common.threadFactory"); + if (fp != null) + fac = ((ForkJoinWorkerThreadFactory)ClassLoader. + getSystemClassLoader().loadClass(fp).newInstance()); + if (hp != null) + handler = ((Thread.UncaughtExceptionHandler)ClassLoader. + getSystemClassLoader().loadClass(hp).newInstance()); + if (pp != null) + par = Integer.parseInt(pp); + } catch (Exception ignore) { + } + + if (par <= 0) + par = Runtime.getRuntime().availableProcessors(); + if (par > MAX_CAP) + par = MAX_CAP; + commonPoolParallelism = par; + long np = (long)(-par); // precompute initial ctl value + long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); + + commonPool = new ForkJoinPool(par, ct, fac, handler); } /**