--- jsr166/src/jsr166y/ForkJoinPool.java 2010/08/13 16:21:23 1.63
+++ jsr166/src/jsr166y/ForkJoinPool.java 2010/09/07 06:32:45 1.74
@@ -7,7 +7,6 @@
package jsr166y;
import java.util.concurrent.*;
-
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collection;
@@ -69,7 +68,7 @@ import java.util.concurrent.CountDownLat
* | Call from within fork/join computations |
*
*
- * | Arange async execution |
+ * Arrange async execution |
* {@link #execute(ForkJoinTask)} |
* {@link ForkJoinTask#fork} |
*
@@ -140,7 +139,7 @@ public class ForkJoinPool extends Abstra
* Beyond work-stealing support and essential bookkeeping, the
* main responsibility of this framework is to take actions when
* one worker is waiting to join a task stolen (or always held by)
- * another. Becauae we are multiplexing many tasks on to a pool
+ * another. Because we are multiplexing many tasks on to a pool
* of workers, we can't just let them block (as in Thread.join).
* We also cannot just reassign the joiner's run-time stack with
* another and replace it later, which would be a form of
@@ -157,19 +156,21 @@ public class ForkJoinPool extends Abstra
* links to try to find such a task.
*
* Compensating: Unless there are already enough live threads,
- * method helpMaintainParallelism() may create or or
+ * method helpMaintainParallelism() may create or
* re-activate a spare thread to compensate for blocked
* joiners until they unblock.
*
- * Because the determining existence of conservatively safe
- * helping targets, the availability of already-created spares,
- * and the apparent need to create new spares are all racy and
- * require heuristic guidance, we rely on multiple retries of
- * each. Further, because it is impossible to keep exactly the
- * target (parallelism) number of threads running at any given
- * time, we allow compensation during joins to fail, and enlist
- * all other threads to help out whenever they are not otherwise
- * occupied (i.e., mainly in method preStep).
+ * It is impossible to keep exactly the target (parallelism)
+ * number of threads running at any given time. Determining
+ * existence of conservatively safe helping targets, the
+ * availability of already-created spares, and the apparent need
+ * to create new spares are all racy and require heuristic
+ * guidance, so we rely on multiple retries of each. Compensation
+ * occurs in slow-motion. It is triggered only upon timeouts of
+ * Object.wait used for joins. This reduces poor decisions that
+ * would otherwise be made when threads are waiting for others
+ * that are stalled because of unrelated activities such as
+ * garbage collection.
*
* The ManagedBlocker extension API can't use helping so relies
* only on compensation in method awaitBlocker.
@@ -224,11 +225,11 @@ public class ForkJoinPool extends Abstra
* ManagedBlocker), we may create or resume others to take their
* place until they unblock (see below). Implementing this
* requires counts of the number of "running" threads (i.e., those
- * that are neither blocked nor artifically suspended) as well as
+ * that are neither blocked nor artificially suspended) as well as
* the total number. These two values are packed into one field,
* "workerCounts" because we need accurate snapshots when deciding
* to create, resume or suspend. Note however that the
- * correspondance of these counts to reality is not guaranteed. In
+ * correspondence of these counts to reality is not guaranteed. In
* particular updates for unblocked threads may lag until they
* actually wake up.
*
@@ -259,80 +260,77 @@ public class ForkJoinPool extends Abstra
* workers that previously could not find a task to now find one:
* Submission of a new task to the pool, or another worker pushing
* a task onto a previously empty queue. (We also use this
- * mechanism for termination actions that require wakeups of idle
- * workers). Each worker maintains its last known event count,
- * and blocks when a scan for work did not find a task AND its
- * lastEventCount matches the current eventCount. Waiting idle
- * workers are recorded in a variant of Treiber stack headed by
- * field eventWaiters which, when nonzero, encodes the thread
- * index and count awaited for by the worker thread most recently
- * calling eventSync. This thread in turn has a record (field
- * nextEventWaiter) for the next waiting worker. In addition to
- * allowing simpler decisions about need for wakeup, the event
- * count bits in eventWaiters serve the role of tags to avoid ABA
- * errors in Treiber stacks. To reduce delays in task diffusion,
- * workers not otherwise occupied may invoke method
- * releaseEventWaiters, that removes and signals (unparks) workers
- * not waiting on current count. To reduce stalls, To minimize
- * task production stalls associate with signalling, any worker
- * pushing a task on an empty queue invokes the weaker method
- * signalWork, that only releases idle workers until it detects
- * interference by other threads trying to release, and lets them
- * take over. The net effect is a tree-like diffusion of signals,
- * where released threads (and possibly others) help with unparks.
- * To further reduce contention effects a bit, failed CASes to
- * increment field eventCount are tolerated without retries.
+ * mechanism for configuration and termination actions that
+ * require wakeups of idle workers). Each worker maintains its
+ * last known event count, and blocks when a scan for work did not
+ * find a task AND its lastEventCount matches the current
+ * eventCount. Waiting idle workers are recorded in a variant of
+ * Treiber stack headed by field eventWaiters which, when nonzero,
+ * encodes the thread index and count awaited for by the worker
+ * thread most recently calling eventSync. This thread in turn has
+ * a record (field nextEventWaiter) for the next waiting worker.
+ * In addition to allowing simpler decisions about need for
+ * wakeup, the event count bits in eventWaiters serve the role of
+ * tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
+ * released threads also try to release at most two others. The
+ * net effect is a tree-like diffusion of signals, where released
+ * threads (and possibly others) help with unparks. To further
+ * reduce contention effects a bit, failed CASes to increment
+ * field eventCount are tolerated without retries in signalWork.
* Conceptually they are merged into the same event, which is OK
* when their only purpose is to enable workers to scan for work.
*
- * 5. Managing suspension of extra workers. When a worker is about
- * to block waiting for a join (or via ManagedBlockers), we may
- * create a new thread to maintain parallelism level, or at least
- * avoid starvation. Usually, extra threads are needed for only
- * very short periods, yet join dependencies are such that we
- * sometimes need them in bursts. Rather than create new threads
- * each time this happens, we suspend no-longer-needed extra ones
- * as "spares". For most purposes, we don't distinguish "extra"
- * spare threads from normal "core" threads: On each call to
- * preStep (the only point at which we can do this) a worker
- * checks to see if there are now too many running workers, and if
- * so, suspends itself. Method helpMaintainParallelism looks for
- * suspended threads to resume before considering creating a new
- * replacement. The spares themselves are encoded on another
- * variant of a Treiber Stack, headed at field "spareWaiters".
- * Note that the use of spares is intrinsically racy. One thread
- * may become a spare at about the same time as another is
- * needlessly being created. We counteract this and related slop
- * in part by requiring resumed spares to immediately recheck (in
- * preStep) to see whether they they should re-suspend. To avoid
- * long-term build-up of spares, the oldest spare (see
- * ForkJoinWorkerThread.suspendAsSpare) occasionally wakes up if
- * not signalled and calls tryTrimSpare, which uses two different
- * thresholds: Always killing if the number of spares is greater
- * that 25% of total, and killing others only at a slower rate
- * (UNUSED_SPARE_TRIM_RATE_NANOS).
+ * 5. Managing suspension of extra workers. When a worker notices
+ * (usually upon timeout of a wait()) that there are too few
+ * running threads, we may create a new thread to maintain
+ * parallelism level, or at least avoid starvation. Usually, extra
+ * threads are needed for only very short periods, yet join
+ * dependencies are such that we sometimes need them in
+ * bursts. Rather than create new threads each time this happens,
+ * we suspend no-longer-needed extra ones as "spares". For most
+ * purposes, we don't distinguish "extra" spare threads from
+ * normal "core" threads: On each call to preStep (the only point
+ * at which we can do this) a worker checks to see if there are
+ * now too many running workers, and if so, suspends itself.
+ * Method helpMaintainParallelism looks for suspended threads to
+ * resume before considering creating a new replacement. The
+ * spares themselves are encoded on another variant of a Treiber
+ * Stack, headed at field "spareWaiters". Note that the use of
+ * spares is intrinsically racy. One thread may become a spare at
+ * about the same time as another is needlessly being created. We
+ * counteract this and related slop in part by requiring resumed
+ * spares to immediately recheck (in preStep) to see whether they
+ * should re-suspend.
+ *
+ * 6. Killing off unneeded workers. A timeout mechanism is used to
+ * shed unused workers: The oldest (first) event queue waiter uses
+ * a timed rather than hard wait. When this wait times out without
+ * a normal wakeup, it tries to shutdown any one (for convenience
+ * the newest) other spare or event waiter via
+ * tryShutdownUnusedWorker. This eventually reduces the number of
+ * worker threads to a minimum of one after a long enough period
+ * without use.
*
- * 6. Deciding when to create new workers. The main dynamic
+ * 7. Deciding when to create new workers. The main dynamic
* control in this class is deciding when to create extra threads
* in method helpMaintainParallelism. We would like to keep
- * exactly #parallelism threads running, which is an impossble
+ * exactly #parallelism threads running, which is an impossible
* task. We always need to create one when the number of running
* threads would become zero and all workers are busy. Beyond
- * this, we must rely on heuristics that work well in the the
- * presence of transients phenomena such as GC stalls, dynamic
+ * this, we must rely on heuristics that work well in the
+ * presence of transient phenomena such as GC stalls, dynamic
* compilation, and wake-up lags. These transients are extremely
* common -- we are normally trying to fully saturate the CPUs on
* a machine, so almost any activity other than running tasks
- * impedes accuracy. Our main defense is to allow some slack in
- * creation thresholds, using rules that reflect the fact that the
- * more threads we have running, the more likely that we are
- * underestimating the number running threads. The rules also
- * better cope with the fact that some of the methods in this
- * class tend to never become compiled (but are interpreted), so
- * some components of the entire set of controls might execute 100
- * times faster than others. And similarly for cases where the
- * apparent lack of work is just due to GC stalls and other
- * transient system activity.
+ * impedes accuracy. Our main defense is to allow parallelism to
+ * lapse for a while during joins, and use a timeout to see if,
+ * after the resulting settling, there is still a need for
+ * additional workers. This also better copes with the fact that
+ * some of the methods in this class tend to never become compiled
+ * (but are interpreted), so some components of the entire set of
+ * controls might execute 100 times faster than others. And
+ * similarly for cases where the apparent lack of work is just due
+ * to GC stalls and other transient system activity.
*
* Beware that there is a lot of representation-level coupling
* among classes ForkJoinPool, ForkJoinWorkerThread, and
@@ -347,7 +345,7 @@ public class ForkJoinPool extends Abstra
* "while ((local = field) != 0)") which are usually the simplest
* way to ensure the required read orderings (which are sometimes
* critical). Also several occurrences of the unusual "do {}
- * while(!cas...)" which is the simplest way to force an update of
+ * while (!cas...)" which is the simplest way to force an update of
* a CAS'ed variable. There are also other coding oddities that
* help some methods perform reasonably even when interpreted (not
* compiled), at the expense of some messy constructions that
@@ -419,6 +417,27 @@ public class ForkJoinPool extends Abstra
new AtomicInteger();
/**
+ * The time to block in a join (see awaitJoin) before checking if
+ * a new worker should be (re)started to maintain parallelism
+ * level. The value should be short enough to maintain global
+ * responsiveness and progress but long enough to avoid
+ * counterproductive firings during GC stalls or unrelated system
+ * activity, and to not bog down systems with continual re-firings
+ * on GCs or legitimately long waits.
+ */
+ private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
+
+ /**
+ * The wakeup interval (in nanoseconds) for the oldest worker
+ * waiting for an event invokes tryShutdownUnusedWorker to shrink
+ * the number of workers. The exact value does not matter too
+ * much, but should be long enough to slowly release resources
+ * during long periods without use without disrupting normal use.
+ */
+ private static final long SHRINK_RATE_NANOS =
+ 30L * 1000L * 1000L * 1000L; // 2 per minute
+
+ /**
* Absolute bound for parallelism level. Twice this number plus
* one (i.e., 0xfff) must fit into a 16bit field to enable
* word-packing for some counts and indices.
@@ -463,18 +482,7 @@ public class ForkJoinPool extends Abstra
private volatile long stealCount;
/**
- * The last nanoTime that a spare thread was trimmed
- */
- private volatile long trimTime;
-
- /**
- * The rate at which to trim unused spares
- */
- static final long UNUSED_SPARE_TRIM_RATE_NANOS =
- 1000L * 1000L * 1000L; // 1 sec
-
- /**
- * Encoded record of top of treiber stack of threads waiting for
+ * Encoded record of top of Treiber stack of threads waiting for
* events. The top 32 bits contain the count being waited for. The
* bottom 16 bits contains one plus the pool index of waiting
* worker thread. (Bits 16-31 are unused.)
@@ -493,7 +501,7 @@ public class ForkJoinPool extends Abstra
private volatile int eventCount;
/**
- * Encoded record of top of treiber stack of spare threads waiting
+ * Encoded record of top of Treiber stack of spare threads waiting
* for resumption. The top 16 bits contain an arbitrary count to
* avoid ABA effects. The bottom 16bits contains one plus the pool
* index of waiting worker thread.
@@ -514,8 +522,12 @@ public class ForkJoinPool extends Abstra
* These are bundled together to ensure consistent read for
* termination checks (i.e., that runLevel is at least SHUTDOWN
* and active threads is zero).
+ *
+ * Notes: Most direct CASes are dependent on these bitfield
+ * positions. Also, this field is non-private to enable direct
+ * performance-sensitive CASes in ForkJoinWorkerThread.
*/
- private volatile int runState;
+ volatile int runState;
// Note: The order among run level values matters.
private static final int RUNLEVEL_SHIFT = 16;
@@ -523,7 +535,6 @@ public class ForkJoinPool extends Abstra
private static final int TERMINATING = 1 << (RUNLEVEL_SHIFT + 1);
private static final int TERMINATED = 1 << (RUNLEVEL_SHIFT + 2);
private static final int ACTIVE_COUNT_MASK = (1 << RUNLEVEL_SHIFT) - 1;
- private static final int ONE_ACTIVE = 1; // active update delta
/**
* Holds number of total (i.e., created and not yet terminated)
@@ -564,9 +575,8 @@ public class ForkJoinPool extends Abstra
*/
private final int poolNumber;
-
- // Utilities for CASing fields. Note that several of these
- // are manually inlined by callers
+ // Utilities for CASing fields. Note that most of these
+ // are usually manually inlined by callers
/**
* Increments running count part of workerCounts
@@ -599,11 +609,12 @@ public class ForkJoinPool extends Abstra
private void decrementWorkerCounts(int dr, int dt) {
for (;;) {
int wc = workerCounts;
- if (wc == 0 && (runState & TERMINATED) != 0)
- return; // lagging termination on a backout
if ((wc & RUNNING_COUNT_MASK) - dr < 0 ||
- (wc >>> TOTAL_COUNT_SHIFT) - dt < 0)
+ (wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
+ if ((runState & TERMINATED) != 0)
+ return; // lagging termination on a backout
Thread.yield();
+ }
if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
wc, wc - (dr + dt)))
return;
@@ -611,34 +622,13 @@ public class ForkJoinPool extends Abstra
}
/**
- * Increments event count
- */
- private void advanceEventCount() {
- int c;
- do {} while(!UNSAFE.compareAndSwapInt(this, eventCountOffset,
- c = eventCount, c+1));
- }
-
- /**
- * Tries incrementing active count; fails on contention.
- * Called by workers before executing tasks.
- *
- * @return true on success
- */
- final boolean tryIncrementActiveCount() {
- int c;
- return UNSAFE.compareAndSwapInt(this, runStateOffset,
- c = runState, c + ONE_ACTIVE);
- }
-
- /**
* Tries decrementing active count; fails on contention.
* Called when workers cannot find tasks to run.
*/
final boolean tryDecrementActiveCount() {
int c;
return UNSAFE.compareAndSwapInt(this, runStateOffset,
- c = runState, c - ONE_ACTIVE);
+ c = runState, c - 1);
}
/**
@@ -683,11 +673,11 @@ public class ForkJoinPool extends Abstra
}
/**
- * Nulls out record of worker in workers array
+ * Nulls out record of worker in workers array.
*/
private void forgetWorker(ForkJoinWorkerThread w) {
int idx = w.poolIndex;
- // Locking helps method recordWorker avoid unecessary expansion
+ // Locking helps method recordWorker avoid unnecessary expansion
final ReentrantLock lock = this.workerLock;
lock.lock();
try {
@@ -699,31 +689,10 @@ public class ForkJoinPool extends Abstra
}
}
- // adding and removing workers
-
- /**
- * Tries to create and add new worker. Assumes that worker counts
- * are already updated to accommodate the worker, so adjusts on
- * failure.
- */
- private void addWorker() {
- ForkJoinWorkerThread w = null;
- try {
- w = factory.newThread(this);
- } finally { // Adjust on either null or exceptional factory return
- if (w == null) {
- decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
- tryTerminate(false); // in case of failure during shutdown
- }
- }
- if (w != null)
- w.start(recordWorker(w), ueh);
- }
-
/**
* Final callback from terminating worker. Removes record of
* worker from array, and adjusts counts. If pool is shutting
- * down, tries to complete terminatation.
+ * down, tries to complete termination.
*
* @param w the worker
*/
@@ -740,25 +709,29 @@ public class ForkJoinPool extends Abstra
/**
* Releases workers blocked on a count not equal to current count.
* Normally called after precheck that eventWaiters isn't zero to
- * avoid wasted array checks.
- *
- * @param signalling true if caller is a signalling worker so can
- * exit upon (conservatively) detected contention by other threads
- * who will continue to release
+ * avoid wasted array checks. Gives up upon a change in count or
+ * upon releasing two workers, letting others take over.
*/
- private void releaseEventWaiters(boolean signalling) {
+ private void releaseEventWaiters() {
ForkJoinWorkerThread[] ws = workers;
int n = ws.length;
- long h; // head of stack
- ForkJoinWorkerThread w; int id, ec;
- while ((id = ((int)((h = eventWaiters) & WAITER_ID_MASK)) - 1) >= 0 &&
- (int)(h >>> EVENT_COUNT_SHIFT) != (ec = eventCount) &&
+ long h = eventWaiters;
+ int ec = eventCount;
+ boolean releasedOne = false;
+ ForkJoinWorkerThread w; int id;
+ while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
+ (int)(h >>> EVENT_COUNT_SHIFT) != ec &&
id < n && (w = ws[id]) != null) {
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
- h, h = w.nextWaiter))
+ h, w.nextWaiter)) {
LockSupport.unpark(w);
- if (signalling && (eventCount != ec || eventWaiters != h))
+ if (releasedOne) // exit on second release
+ break;
+ releasedOne = true;
+ }
+ if (eventCount != ec)
break;
+ h = eventWaiters;
}
}
@@ -770,268 +743,315 @@ public class ForkJoinPool extends Abstra
int c; // try to increment event count -- CAS failure OK
UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
if (eventWaiters != 0L)
- releaseEventWaiters(true);
+ releaseEventWaiters();
}
/**
- * Blocks worker until terminating or event count
- * advances from last value held by worker
+ * Adds the given worker to event queue and blocks until
+ * terminating or event count advances from the given value
*
* @param w the calling worker thread
+ * @param ec the count
*/
- private void eventSync(ForkJoinWorkerThread w) {
- int wec = w.lastEventCount;
- long nh = (((long)wec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
+ private void eventSync(ForkJoinWorkerThread w, int ec) {
+ long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
long h;
while ((runState < SHUTDOWN || !tryTerminate(false)) &&
- ((h = eventWaiters) == 0L ||
- (int)(h >>> EVENT_COUNT_SHIFT) == wec) &&
- eventCount == wec) {
+ (((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
+ (int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
+ eventCount == ec) {
if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
w.nextWaiter = h, nh)) {
- while (runState < TERMINATING && eventCount == wec) {
- if (!tryAccumulateStealCount(w)) // transfer while idle
- continue;
- Thread.interrupted(); // clear/ignore interrupt
- if (eventCount != wec)
- break;
+ awaitEvent(w, ec);
+ break;
+ }
+ }
+ }
+
+ /**
+ * Blocks the given worker (that has already been entered as an
+ * event waiter) until terminating or event count advances from
+ * the given value. The oldest (first) waiter uses a timed wait to
+ * occasionally one-by-one shrink the number of workers (to a
+ * minimum of one) if the pool has not been used for extended
+ * periods.
+ *
+ * @param w the calling worker thread
+ * @param ec the count
+ */
+ private void awaitEvent(ForkJoinWorkerThread w, int ec) {
+ while (eventCount == ec) {
+ if (tryAccumulateStealCount(w)) { // transfer while idle
+ boolean untimed = (w.nextWaiter != 0L ||
+ (workerCounts & RUNNING_COUNT_MASK) <= 1);
+ long startTime = untimed? 0 : System.nanoTime();
+ Thread.interrupted(); // clear/ignore interrupt
+ if (eventCount != ec || w.runState != 0 ||
+ runState >= TERMINATING) // recheck after clear
+ break;
+ if (untimed)
LockSupport.park(w);
+ else {
+ LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
+ if (eventCount != ec || w.runState != 0 ||
+ runState >= TERMINATING)
+ break;
+ if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
+ tryShutdownUnusedWorker(ec);
}
- break;
}
}
- w.lastEventCount = eventCount;
}
- // Maintaining spares
+ // Maintaining parallelism
/**
- * Pushes worker onto the spare stack
+ * Pushes worker onto the spare stack.
*/
final void pushSpare(ForkJoinWorkerThread w) {
- int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex+1);
+ int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
w.nextSpare = spareWaiters,ns));
}
/**
- * Tries (once) to resume a spare if running count is less than
- * target parallelism. Fails on contention or stale workers.
+ * Tries (once) to resume a spare if the number of running
+ * threads is less than target.
*/
private void tryResumeSpare() {
int sw, id;
+ ForkJoinWorkerThread[] ws = workers;
+ int n = ws.length;
ForkJoinWorkerThread w;
- ForkJoinWorkerThread[] ws;
- if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
- id < (ws = workers).length && (w = ws[id]) != null &&
+ if ((sw = spareWaiters) != 0 &&
+ (id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
+ id < n && (w = ws[id]) != null &&
(workerCounts & RUNNING_COUNT_MASK) < parallelism &&
- eventWaiters == 0L &&
spareWaiters == sw &&
UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
- sw, w.nextSpare) &&
- w.tryUnsuspend()) {
- int c; // try increment; if contended, finish after unpark
- boolean inc = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- c = workerCounts,
- c + ONE_RUNNING);
- LockSupport.unpark(w);
- if (!inc) {
- do {} while(!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- c = workerCounts,
- c + ONE_RUNNING));
- }
+ sw, w.nextSpare)) {
+ int c; // increment running count before resume
+ do {} while (!UNSAFE.compareAndSwapInt
+ (this, workerCountsOffset,
+ c = workerCounts, c + ONE_RUNNING));
+ if (w.tryUnsuspend())
+ LockSupport.unpark(w);
+ else // back out if w was shutdown
+ decrementWorkerCounts(ONE_RUNNING, 0);
}
}
/**
- * Callback from oldest spare occasionally waking up. Tries
- * (once) to shutdown a spare if more than 25% spare overage, or
- * if UNUSED_SPARE_TRIM_RATE_NANOS have elapsed and there are at
- * least #parallelism running threads. Note that we don't need CAS
- * or locks here because the method is called only from the oldest
- * suspended spare occasionally waking (and even misfires are OK).
- *
- * @param now the wake up nanoTime of caller
- */
- final void tryTrimSpare(long now) {
- long lastTrim = trimTime;
- trimTime = now;
- helpMaintainParallelism(); // first, help wake up any needed spares
- int sw, id;
- ForkJoinWorkerThread w;
- ForkJoinWorkerThread[] ws;
+ * Tries to increase the number of running workers if below target
+ * parallelism: If a spare exists tries to resume it via
+ * tryResumeSpare. Otherwise, if not enough total workers or all
+ * existing workers are busy, adds a new worker. In all cases also
+ * helps wake up releasable workers waiting for work.
+ */
+ private void helpMaintainParallelism() {
int pc = parallelism;
- int wc = workerCounts;
- if ((wc & RUNNING_COUNT_MASK) >= pc &&
- (((wc >>> TOTAL_COUNT_SHIFT) - pc) > (pc >>> 2) + 1 ||// approx 25%
- now - lastTrim >= UNUSED_SPARE_TRIM_RATE_NANOS) &&
- (id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
- id < (ws = workers).length && (w = ws[id]) != null &&
- UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
- sw, w.nextSpare))
- w.shutdown(false);
+ int wc, rs, tc;
+ while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
+ (rs = runState) < TERMINATING) {
+ if (spareWaiters != 0)
+ tryResumeSpare();
+ else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
+ (tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
+ break; // enough total
+ else if (runState == rs && workerCounts == wc &&
+ UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
+ wc + (ONE_RUNNING|ONE_TOTAL))) {
+ ForkJoinWorkerThread w = null;
+ try {
+ w = factory.newThread(this);
+ } finally { // adjust on null or exceptional factory return
+ if (w == null) {
+ decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
+ tryTerminate(false); // handle failure during shutdown
+ }
+ }
+ if (w == null)
+ break;
+ w.start(recordWorker(w), ueh);
+ if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
+ int c; // advance event count
+ UNSAFE.compareAndSwapInt(this, eventCountOffset,
+ c = eventCount, c+1);
+ break; // add at most one unless total below target
+ }
+ }
+ }
+ if (eventWaiters != 0L)
+ releaseEventWaiters();
}
/**
- * Does at most one of:
- *
- * 1. Help wake up existing workers waiting for work via
- * releaseEventWaiters. (If any exist, then it probably doesn't
- * matter right now if under target parallelism level.)
+ * Callback from the oldest waiter in awaitEvent waking up after a
+ * period of non-use. If all workers are idle, tries (once) to
+ * shutdown an event waiter or a spare, if one exists. Note that
+ * we don't need CAS or locks here because the method is called
+ * only from one thread occasionally waking (and even misfires are
+ * OK). Note that until the shutdown worker fully terminates,
+ * workerCounts will overestimate total count, which is tolerable.
*
- * 2. If below parallelism level and a spare exists, try (once)
- * to resume it via tryResumeSpare.
- *
- * 3. If neither of the above, tries (once) to add a new
- * worker if either there are not enough total, or if all
- * existing workers are busy, there are either no running
- * workers or the deficit is at least twice the surplus.
+ * @param ec the event count waited on by caller (to abort
+ * attempt if count has since changed).
*/
- private void helpMaintainParallelism() {
- // uglified to work better when not compiled
- int pc, wc, rc, tc, rs; long h;
- if ((h = eventWaiters) != 0L) {
- if ((int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
- releaseEventWaiters(false); // avoid useless call
- }
- else if ((pc = parallelism) >
- (rc = ((wc = workerCounts) & RUNNING_COUNT_MASK))) {
- if (spareWaiters != 0)
- tryResumeSpare();
- else if ((rs = runState) < TERMINATING &&
- ((tc = wc >>> TOTAL_COUNT_SHIFT) < pc ||
- (tc == (rs & ACTIVE_COUNT_MASK) && // all busy
- (rc == 0 || // must add
- rc < pc - ((tc - pc) << 1)) && // within slack
- tc < MAX_WORKERS && runState == rs)) && // recheck busy
- workerCounts == wc &&
- UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
- wc + (ONE_RUNNING|ONE_TOTAL)))
- addWorker();
+ private void tryShutdownUnusedWorker(int ec) {
+ if (runState == 0 && eventCount == ec) { // only trigger if all idle
+ ForkJoinWorkerThread[] ws = workers;
+ int n = ws.length;
+ ForkJoinWorkerThread w = null;
+ boolean shutdown = false;
+ int sw;
+ long h;
+ if ((sw = spareWaiters) != 0) { // prefer killing spares
+ int id = (sw & SPARE_ID_MASK) - 1;
+ if (id >= 0 && id < n && (w = ws[id]) != null &&
+ UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
+ sw, w.nextSpare))
+ shutdown = true;
+ }
+ else if ((h = eventWaiters) != 0L) {
+ long nh;
+ int id = ((int)(h & WAITER_ID_MASK)) - 1;
+ if (id >= 0 && id < n && (w = ws[id]) != null &&
+ (nh = w.nextWaiter) != 0L && // keep at least one worker
+ UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
+ shutdown = true;
+ }
+ if (w != null && shutdown) {
+ w.shutdown();
+ LockSupport.unpark(w);
+ }
}
+ releaseEventWaiters(); // in case of interference
}
/**
* Callback from workers invoked upon each top-level action (i.e.,
- * stealing a task or taking a submission and running
- * it). Performs one or more of the following:
+ * stealing a task or taking a submission and running it).
+ * Performs one or more of the following:
*
- * 1. If the worker cannot find work (misses > 0), updates its
- * active status to inactive and updates activeCount unless
- * this is the first miss and there is contention, in which
- * case it may try again (either in this or a subsequent
- * call).
- *
- * 2. If there are at least 2 misses, awaits the next task event
- * via eventSync
- *
- * 3. If there are too many running threads, suspends this worker
- * (first forcing inactivation if necessary). If it is not
- * needed, it may be killed while suspended via
- * tryTrimSpare. Otherwise, upon resume it rechecks to make
- * sure that it is still needed.
- *
- * 4. Helps release and/or reactivate other workers via
- * helpMaintainParallelism
+ * 1. If the worker is active and either did not run a task
+ * or there are too many workers, try to set its active status
+ * to inactive and update activeCount. On contention, we may
+ * try again in this or a subsequent call.
+ *
+ * 2. If not enough total workers, help create some.
+ *
+ * 3. If there are too many running workers, suspend this worker
+ * (first forcing inactive if necessary). If it is not needed,
+ * it may be shutdown while suspended (via
+ * tryShutdownUnusedWorker). Otherwise, upon resume it
+ * rechecks running thread count and need for event sync.
+ *
+ * 4. If worker did not run a task, await the next task event via
+ * eventSync if necessary (first forcing inactivation), upon
+ * which the worker may be shutdown via
+ * tryShutdownUnusedWorker. Otherwise, help release any
+ * existing event waiters that are now releasable,
*
* @param w the worker
- * @param misses the number of scans by caller failing to find work
- * (saturating at 2 just to avoid wraparound)
+ * @param ran true if worker ran a task since last call to this method
*/
- final void preStep(ForkJoinWorkerThread w, int misses) {
+ final void preStep(ForkJoinWorkerThread w, boolean ran) {
+ int wec = w.lastEventCount;
boolean active = w.active;
+ boolean inactivate = false;
int pc = parallelism;
- for (;;) {
+ int rs;
+ while (w.runState == 0 && (rs = runState) < TERMINATING) {
+ if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
+ UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
+ inactivate = active = w.active = false;
int wc = workerCounts;
- int rc = wc & RUNNING_COUNT_MASK;
- if (active && (misses > 0 || rc > pc)) {
- int rs; // try inactivate
- if (UNSAFE.compareAndSwapInt(this, runStateOffset,
- rs = runState, rs - ONE_ACTIVE))
- active = w.active = false;
- else if (misses > 1 || rc > pc ||
- (rs & ACTIVE_COUNT_MASK) >= pc)
- continue; // force inactivate
- }
- if (misses > 1) {
- misses = 0; // don't re-sync
- eventSync(w); // continue loop to recheck rc
- }
- else if (rc > pc) {
- if (workerCounts == wc && // try to suspend as spare
+ if ((wc & RUNNING_COUNT_MASK) > pc) {
+ if (!(inactivate |= active) && // must inactivate to suspend
+ workerCounts == wc && // try to suspend as spare
UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc, wc - ONE_RUNNING) &&
- !w.suspendAsSpare()) // false if killed
+ wc, wc - ONE_RUNNING))
+ w.suspendAsSpare();
+ }
+ else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
+ helpMaintainParallelism(); // not enough workers
+ else if (!ran) {
+ long h = eventWaiters;
+ int ec = eventCount;
+ if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
+ releaseEventWaiters(); // release others before waiting
+ else if (ec != wec) {
+ w.lastEventCount = ec; // no need to wait
break;
+ }
+ else if (!(inactivate |= active))
+ eventSync(w, wec); // must inactivate before sync
}
- else {
- if (rc < pc || eventWaiters != 0L)
- helpMaintainParallelism();
+ else
break;
- }
}
}
/**
* Helps and/or blocks awaiting join of the given task.
- * Alternates between helpJoinTask() and helpMaintainParallelism()
- * as many times as there is a deficit in running count (or longer
- * if running count would become zero), then blocks if task still
- * not done.
+ * See above for explanation.
*
* @param joinMe the task to join
+ * @param worker the current worker thread
*/
final void awaitJoin(ForkJoinTask joinMe, ForkJoinWorkerThread worker) {
- int threshold = parallelism; // descend blocking thresholds
+ int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
while (joinMe.status >= 0) {
- boolean block; int wc;
+ int wc;
worker.helpJoinTask(joinMe);
if (joinMe.status < 0)
break;
- if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
- if (threshold > 0)
- --threshold;
- else
- advanceEventCount(); // force release
- block = false;
- }
- else
- block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc, wc - ONE_RUNNING);
- helpMaintainParallelism();
- if (block) {
- int c;
- joinMe.internalAwaitDone();
+ else if (retries > 0)
+ --retries;
+ else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
+ UNSAFE.compareAndSwapInt(this, workerCountsOffset,
+ wc, wc - ONE_RUNNING)) {
+ int stat, c; long h;
+ while ((stat = joinMe.status) >= 0 &&
+ (h = eventWaiters) != 0L && // help release others
+ (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
+ releaseEventWaiters();
+ if (stat >= 0 &&
+ ((workerCounts & RUNNING_COUNT_MASK) == 0 ||
+ (stat =
+ joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
+ helpMaintainParallelism(); // timeout or no running workers
do {} while (!UNSAFE.compareAndSwapInt
(this, workerCountsOffset,
c = workerCounts, c + ONE_RUNNING));
- break;
+ if (stat < 0)
+ break; // else restart
}
}
}
/**
- * Same idea as awaitJoin, but no helping
+ * Same idea as awaitJoin, but no helping, retries, or timeouts.
*/
final void awaitBlocker(ManagedBlocker blocker)
throws InterruptedException {
- int threshold = parallelism;
while (!blocker.isReleasable()) {
- boolean block; int wc;
- if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
- if (threshold > 0)
- --threshold;
- else
- advanceEventCount();
- block = false;
- }
- else
- block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
- wc, wc - ONE_RUNNING);
- helpMaintainParallelism();
- if (block) {
+ int wc = workerCounts;
+ if ((wc & RUNNING_COUNT_MASK) != 0 &&
+ UNSAFE.compareAndSwapInt(this, workerCountsOffset,
+ wc, wc - ONE_RUNNING)) {
try {
- do {} while (!blocker.isReleasable() && !blocker.block());
+ while (!blocker.isReleasable()) {
+ long h = eventWaiters;
+ if (h != 0L &&
+ (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
+ releaseEventWaiters();
+ else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
+ runState < TERMINATING)
+ helpMaintainParallelism();
+ else if (blocker.block())
+ break;
+ }
} finally {
int c;
do {} while (!UNSAFE.compareAndSwapInt
@@ -1073,23 +1093,22 @@ public class ForkJoinPool extends Abstra
* Actions on transition to TERMINATING
*
* Runs up to four passes through workers: (0) shutting down each
- * quietly (without waking up if parked) to quickly spread
- * notifications without unnecessary bouncing around event queues
- * etc (1) wake up and help cancel tasks (2) interrupt (3) mop up
- * races with interrupted workers
+ * (without waking up if parked) to quickly spread notifications
+ * without unnecessary bouncing around event queues etc (1) wake
+ * up and help cancel tasks (2) interrupt (3) mop up races with
+ * interrupted workers
*/
private void startTerminating() {
cancelSubmissions();
for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
- advanceEventCount();
+ int c; // advance event count
+ UNSAFE.compareAndSwapInt(this, eventCountOffset,
+ c = eventCount, c+1);
eventWaiters = 0L; // clobber lists
spareWaiters = 0;
- ForkJoinWorkerThread[] ws = workers;
- int n = ws.length;
- for (int i = 0; i < n; ++i) {
- ForkJoinWorkerThread w = ws[i];
+ for (ForkJoinWorkerThread w : workers) {
if (w != null) {
- w.shutdown(true);
+ w.shutdown();
if (passes > 0 && !w.isTerminated()) {
w.cancelTasks();
LockSupport.unpark(w);
@@ -1106,7 +1125,7 @@ public class ForkJoinPool extends Abstra
}
/**
- * Clear out and cancel submissions, ignoring exceptions
+ * Clears out and cancels submissions, ignoring exceptions.
*/
private void cancelSubmissions() {
ForkJoinTask task;
@@ -1121,15 +1140,15 @@ public class ForkJoinPool extends Abstra
// misc support for ForkJoinWorkerThread
/**
- * Returns pool number
+ * Returns pool number.
*/
final int getPoolNumber() {
return poolNumber;
}
/**
- * Tries to accumulates steal count from a worker, clearing
- * the worker's value.
+ * Tries to accumulate steal count from a worker, clearing
+ * the worker's value if successful.
*
* @return true if worker steal count now zero
*/
@@ -1151,9 +1170,12 @@ public class ForkJoinPool extends Abstra
*/
final int idlePerActive() {
int pc = parallelism; // use parallelism, not rc
- int ac = runState; // no mask -- artifically boosts during shutdown
+ int ac = runState; // no mask -- artificially boosts during shutdown
// Use exact results for small values, saturate past 4
- return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
+ return ((pc <= ac) ? 0 :
+ (pc >>> 1 <= ac) ? 1 :
+ (pc >>> 2 <= ac) ? 3 :
+ pc >>> 3);
}
// Public and protected methods
@@ -1203,13 +1225,13 @@ public class ForkJoinPool extends Abstra
* use {@link #defaultForkJoinWorkerThreadFactory}.
* @param handler the handler for internal worker threads that
* terminate due to unrecoverable errors encountered while executing
- * tasks. For default value, use null.
+ * tasks. For default value, use {@code null}.
* @param asyncMode if true,
* establishes local first-in-first-out scheduling mode for forked
* tasks that are never joined. This mode may be more appropriate
* than default locally stack-based mode in applications in which
* worker threads only process event-style asynchronous tasks.
- * For default value, use false.
+ * For default value, use {@code false}.
* @throws IllegalArgumentException if parallelism less than or
* equal to zero, or greater than implementation limit
* @throws NullPointerException if the factory is null
@@ -1237,7 +1259,6 @@ public class ForkJoinPool extends Abstra
this.workerLock = new ReentrantLock();
this.termination = new Phaser(1);
this.poolNumber = poolNumberGenerator.incrementAndGet();
- this.trimTime = System.nanoTime();
}
/**
@@ -1245,8 +1266,9 @@ public class ForkJoinPool extends Abstra
* @param pc the initial parallelism level
*/
private static int initialArraySizeFor(int pc) {
- // See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
+ // If possible, initially allocate enough space for one spare
int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
+ // See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
size |= size >>> 1;
size |= size >>> 2;
size |= size >>> 4;
@@ -1265,8 +1287,9 @@ public class ForkJoinPool extends Abstra
if (runState >= SHUTDOWN)
throw new RejectedExecutionException();
submissionQueue.offer(task);
- advanceEventCount();
- helpMaintainParallelism(); // start or wake up workers
+ int c; // try to increment event count -- CAS failure OK
+ UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
+ helpMaintainParallelism(); // create, start, or resume some workers
}
/**
@@ -1285,9 +1308,6 @@ public class ForkJoinPool extends Abstra
/**
* Arranges for (asynchronous) execution of the given task.
- * If the caller is already engaged in a fork/join computation in
- * the current pool, this method is equivalent in effect to
- * {@link ForkJoinTask#fork}.
*
* @param task the task
* @throws NullPointerException if the task is null
@@ -1316,9 +1336,6 @@ public class ForkJoinPool extends Abstra
/**
* Submits a ForkJoinTask for execution.
- * If the caller is already engaged in a fork/join computation in
- * the current pool, this method is equivalent in effect to
- * {@link ForkJoinTask#fork}.
*
* @param task the task to submit
* @return the task
@@ -1509,13 +1526,9 @@ public class ForkJoinPool extends Abstra
*/
public long getQueuedTaskCount() {
long count = 0;
- ForkJoinWorkerThread[] ws = workers;
- int n = ws.length;
- for (int i = 0; i < n; ++i) {
- ForkJoinWorkerThread w = ws[i];
+ for (ForkJoinWorkerThread w : workers)
if (w != null)
count += w.getQueueSize();
- }
return count;
}
@@ -1570,13 +1583,9 @@ public class ForkJoinPool extends Abstra
*/
protected int drainTasksTo(Collection> c) {
int count = submissionQueue.drainTo(c);
- ForkJoinWorkerThread[] ws = workers;
- int n = ws.length;
- for (int i = 0; i < n; ++i) {
- ForkJoinWorkerThread w = ws[i];
+ for (ForkJoinWorkerThread w : workers)
if (w != null)
count += w.drainTasksTo(c);
- }
return count;
}
@@ -1703,7 +1712,7 @@ public class ForkJoinPool extends Abstra
throws InterruptedException {
try {
return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
- } catch(TimeoutException ex) {
+ } catch (TimeoutException ex) {
return false;
}
}
@@ -1750,11 +1759,11 @@ public class ForkJoinPool extends Abstra
* QueueTaker(BlockingQueue q) { this.queue = q; }
* public boolean block() throws InterruptedException {
* if (item == null)
- * item = queue.take
+ * item = queue.take();
* return true;
* }
* public boolean isReleasable() {
- * return item != null || (item = queue.poll) != null;
+ * return item != null || (item = queue.poll()) != null;
* }
* public E getItem() { // call after pool.managedBlock completes
* return item;