/*
* Written by Doug Lea, Bill Scherer, and Michael Scott with
* assistance from members of JCP JSR-166 Expert Group and released to
* the public domain, as explained at
* http://creativecommons.org/licenses/publicdomain
*/
package java.util.concurrent;
import java.util.concurrent.atomic.*;
import java.util.concurrent.locks.LockSupport;
/**
* A synchronization point at which threads can pair and swap elements
* within pairs. Each thread presents some object on entry to the
* {@link #exchange exchange} method, matches with a partner thread,
* and receives its partner's object on return. An Exchanger may be
* viewed as a bidirectional form of a {@link SynchronousQueue}.
* Exchangers may be useful in applications such as genetic algorithms
* and pipeline designs.
*
*
Sample Usage:
* Here are the highlights of a class that uses an {@code Exchanger}
* to swap buffers between threads so that the thread filling the
* buffer gets a freshly emptied one when it needs it, handing off the
* filled one to the thread emptying the buffer.
*
Memory consistency effects: For each pair of threads that
* successfully exchange objects via an {@code Exchanger}, actions
* prior to the {@code exchange()} in each thread
* happen-before
* those subsequent to a return from the corresponding {@code exchange()}
* in the other thread.
*
* @since 1.5
* @author Doug Lea and Bill Scherer and Michael Scott
* @param The type of objects that may be exchanged
*/
public class Exchanger {
/*
* Algorithm Description:
*
* The basic idea is to maintain a "slot", which is a reference to
* a Node containing both an Item to offer and a "hole" waiting to
* get filled in. If an incoming "occupying" thread sees that the
* slot is null, it CAS'es (compareAndSets) a Node there and waits
* for another to invoke exchange. That second "fulfilling" thread
* sees that the slot is non-null, and so CASes it back to null,
* also exchanging items by CASing the hole, plus waking up the
* occupying thread if it is blocked. In each case CAS'es may
* fail because a slot at first appears non-null but is null upon
* CAS, or vice-versa. So threads may need to retry these
* actions.
*
* This simple approach works great when there are only a few
* threads using an Exchanger, but performance rapidly
* deteriorates due to CAS contention on the single slot when
* there are lots of threads using an exchanger. So instead we use
* an "arena"; basically a kind of hash table with a dynamically
* varying number of slots, any one of which can be used by
* threads performing an exchange. Incoming threads pick slots
* based on a hash of their Thread ids. If an incoming thread
* fails to CAS in its chosen slot, it picks an alternative slot
* instead. And similarly from there. If a thread successfully
* CASes into a slot but no other thread arrives, it tries
* another, heading toward the zero slot, which always exists even
* if the table shrinks. The particular mechanics controlling this
* are as follows:
*
* Waiting: Slot zero is special in that it is the only slot that
* exists when there is no contention. A thread occupying slot
* zero will block if no thread fulfills it after a short spin.
* In other cases, occupying threads eventually give up and try
* another slot. Waiting threads spin for a while (a period that
* should be a little less than a typical context-switch time)
* before either blocking (if slot zero) or giving up (if other
* slots) and restarting. There is no reason for threads to block
* unless there are unlikely to be any other threads present.
* Occupants are mainly avoiding memory contention so sit there
* quietly polling for a shorter period than it would take to
* block and then unblock them. Non-slot-zero waits that elapse
* because of lack of other threads waste around one extra
* context-switch time per try, which is still on average much
* faster than alternative approaches.
*
* Sizing: Usually, using only a few slots suffices to reduce
* contention. Especially with small numbers of threads, using
* too many slots can lead to just as poor performance as using
* too few of them, and there's not much room for error. The
* variable "max" maintains the number of slots actually in
* use. It is increased when a thread sees too many CAS
* failures. (This is analogous to resizing a regular hash table
* based on a target load factor, except here, growth steps are
* just one-by-one rather than proportional.) Growth requires
* contention failures in each of three tried slots. Requiring
* multiple failures for expansion copes with the fact that some
* failed CASes are not due to contention but instead to simple
* races between two threads or thread pre-emptions occurring
* between reading and CASing. Also, very transient peak
* contention can be much higher than the average sustainable
* levels. The max limit is decreased on average 50% of the times
* that a non-slot-zero wait elapses without being fulfilled.
* Threads experiencing elapsed waits move closer to zero, so
* eventually find existing (or future) threads even if the table
* has been shrunk due to inactivity. The chosen mechanics and
* thresholds for growing and shrinking are intrinsically
* entangled with indexing and hashing inside the exchange code,
* and can't be nicely abstracted out.
*
* Hashing: Each thread picks its initial slot to use in accord
* with a simple hashcode. The sequence is the same on each
* encounter by any given thread, but effectively random across
* threads. Using arenas encounters the classic cost vs quality
* tradeoffs of all hash tables. Here, we use a one-step FNV-1a
* hash code based on the current thread's Thread.getId(), along
* with a cheap approximation to a mod operation to select an
* index. The downside of optimizing index selection in this way
* is that the code is hardwired to use a maximum table size of
* 32. But this value more than suffices for known platforms and
* applications.
*
* Probing: On sensed contention of a selected slot, we probe
* sequentially through the table, analogously to linear probing
* after collision in a hash table. (We move circularly, in
* reverse order, to mesh best with table growth and shrinkage
* rules.) Except that to minimize the effects of false-alarms
* and cache thrashing, we try the first selected slot twice
* before moving.
*
* Padding: Even with contention management, slots are heavily
* contended, so use cache-padding to avoid poor memory
* performance. Because of this, slots are lazily constructed
* only when used, to avoid wasting this space unnecessarily.
* While isolation of locations is not much of an issue at first
* in an application, as time goes on and garbage-collectors
* perform compaction, slots are very likely to be moved adjacent
* to each other, which can cause much thrashing of cache lines on
* MPs unless padding is employed.
*
* This is an improvement of the algorithm described in the paper
* "A Scalable Elimination-based Exchange Channel" by William
* Scherer, Doug Lea, and Michael Scott in Proceedings of SCOOL05
* workshop. Available at: http://hdl.handle.net/1802/2104
*/
/** The number of CPUs, for sizing and spin control */
private static final int NCPU = Runtime.getRuntime().availableProcessors();
/**
* The capacity of the arena. Set to a value that provides more
* than enough space to handle contention. On small machines
* most slots won't be used, but it is still not wasted because
* the extra space provides some machine-level address padding
* to minimize interference with heavily CAS'ed Slot locations.
* And on very large machines, performance eventually becomes
* bounded by memory bandwidth, not numbers of threads/CPUs.
* This constant cannot be changed without also modifying
* indexing and hashing algorithms.
*/
private static final int CAPACITY = 32;
/**
* The value of "max" that will hold all threads without
* contention. When this value is less than CAPACITY, some
* otherwise wasted expansion can be avoided.
*/
private static final int FULL =
Math.max(0, Math.min(CAPACITY, NCPU / 2) - 1);
/**
* The number of times to spin (doing nothing except polling a
* memory location) before blocking or giving up while waiting to
* be fulfilled. Should be zero on uniprocessors. On
* multiprocessors, this value should be large enough so that two
* threads exchanging items as fast as possible block only when
* one of them is stalled (due to GC or preemption), but not much
* longer, to avoid wasting CPU resources. Seen differently, this
* value is a little over half the number of cycles of an average
* context switch time on most systems. The value here is
* approximately the average of those across a range of tested
* systems.
*/
private static final int SPINS = (NCPU == 1) ? 0 : 2000;
/**
* The number of times to spin before blocking in timed waits.
* Timed waits spin more slowly because checking the time takes
* time. The best value relies mainly on the relative rate of
* System.nanoTime vs memory accesses. The value is empirically
* derived to work well across a variety of systems.
*/
private static final int TIMED_SPINS = SPINS / 20;
/**
* Sentinel item representing cancellation of a wait due to
* interruption, timeout, or elapsed spin-waits. This value is
* placed in holes on cancellation, and used as a return value
* from waiting methods to indicate failure to set or get hole.
*/
private static final Object CANCEL = new Object();
/**
* Value representing null arguments/returns from public
* methods. This disambiguates from internal requirement that
* holes start out as null to mean they are not yet set.
*/
private static final Object NULL_ITEM = new Object();
/**
* Nodes hold partially exchanged data. This class
* opportunistically subclasses AtomicReference to represent the
* hole. So get() returns hole, and compareAndSet CAS'es value
* into hole. This class cannot be parameterized as "V" because
* of the use of non-V CANCEL sentinels.
*/
private static final class Node extends AtomicReference