4632N/A * Copyright (c) 2010, 2011, Oracle and/or its affiliates. All rights reserved. 4632N/A * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4632N/A * This code is free software; you can redistribute it and/or modify it 4632N/A * under the terms of the GNU General Public License version 2 only, as 4632N/A * published by the Free Software Foundation. Oracle designates this 4632N/A * particular file as subject to the "Classpath" exception as provided 4632N/A * by Oracle in the LICENSE file that accompanied this code. 4632N/A * This code is distributed in the hope that it will be useful, but WITHOUT 4632N/A * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 4632N/A * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 4632N/A * version 2 for more details (a copy is included in the LICENSE file that 4632N/A * You should have received a copy of the GNU General Public License version 4632N/A * 2 along with this work; if not, write to the Free Software Foundation, 4632N/A * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 4632N/A * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 4632N/A * or visit www.oracle.com if you need additional information or have any 4632N/A * Lazily associate a computed value with (potentially) every type. 4632N/A * For example, if a dynamic language needs to construct a message dispatch 4639N/A * table for each class encountered at a message send call site, 4632N/A * it can use a {@code ClassValue} to cache information needed to 4691N/A * perform the message send quickly, for each class encountered. 4691N/A * @author John Rose, JSR 292 EG 4632N/A * Sole constructor. (For invocation by subclass constructors, typically 4639N/A * Computes the given class's derived value for this {@code ClassValue}. 4639N/A * This method will be invoked within the first thread that accesses 4668N/A * the value with the {@link #get get} method. 4668N/A * Normally, this method is invoked at most once per class, 4639N/A * but it may be invoked again if there has been a call to 4691N/A * If this method throws an exception, the corresponding call to {@code get} 4632N/A * will terminate abnormally with that exception, and no class value will be recorded. 4632N/A * @param type the type whose class value must be computed 4632N/A * @return the newly computed value associated with this {@code ClassValue}, for the given class or interface 4632N/A * Returns the value for the given class. 4632N/A * If no value has yet been computed, it is obtained by 4632N/A * an invocation of the {@link #computeValue computeValue} method. 4632N/A * The actual installation of the value on the class 4639N/A * At that point, if several racing threads have 4639N/A * computed values, one is chosen, and returned to 4695N/A * The {@code type} parameter is typically a class, but it may be any type, 4632N/A * such as an interface, a primitive type (like {@code int.class}), or {@code void.class}. 4632N/A * In the absence of {@code remove} calls, a class value has a simple 4639N/A * state diagram: uninitialized and initialized. 4639N/A * When {@code remove} calls are made, 4632N/A * the rules for value observation are more complex. 4632N/A * See the documentation for {@link #remove remove} for more information. 4632N/A * @param type the type whose class value must be computed or retrieved 4632N/A * @return the current value associated with this {@code ClassValue}, for the given class or interface 4632N/A * @throws NullPointerException if the argument is null 4695N/A // non-racing this.hashCodeForCache : final int 4632N/A // racing e : current value <=> stale value from current cache or from stale cache 4639N/A // invariant: e is null or an Entry with readable Entry.version and Entry.value 4632N/A // invariant: No false positive matches. False negatives are OK if rare. 4668N/A // The key fact that makes this work: if this.version == e.version, 4632N/A // then this thread has a right to observe (final) e.value. 4632N/A // The fast path can fail for any of these reasons: 4632N/A // 1. no entry has been computed yet 4632N/A // 2. hash code collision (before or after reduction mod cache.length) 4639N/A // 3. an entry has been removed (either on this type or another) 4639N/A // 4. the GC has somehow managed to delete e.version and clear the reference 4632N/A * Removes the associated value for the given class. 4632N/A * If this value is subsequently {@linkplain #get read} for the same class, 4639N/A * its value will be reinitialized by invoking its {@link #computeValue computeValue} method. 4639N/A * This may result in an additional invocation of the 4632N/A * {@code computeValue} method for the given class. 4632N/A * In order to explain the interaction between {@code get} and {@code remove} calls, 4632N/A * we must model the state transitions of a class value to take into account 4632N/A * the alternation between uninitialized and initialized states. 4639N/A * To do this, number these states sequentially from zero, and note that 4639N/A * uninitialized (or removed) states are numbered with even numbers, 4632N/A * while initialized (or re-initialized) states have odd numbers. 4632N/A * When a thread {@code T} removes a class value in state {@code 2N}, 4632N/A * nothing happens, since the class value is already uninitialized. 4632N/A * Otherwise, the state is advanced atomically to {@code 2N+1}. 4632N/A * When a thread {@code T} queries a class value in state {@code 2N}, 4639N/A * the thread first attempts to initialize the class value to state {@code 2N+1} 4639N/A * by invoking {@code computeValue} and installing the resulting value. 4632N/A * When {@code T} attempts to install the newly computed value, 4632N/A * if the state is still at {@code 2N}, the class value will be initialized 4639N/A * with the computed value, advancing it to state {@code 2N+1}. 4632N/A * Otherwise, whether the new state is even or odd, 4632N/A * {@code T} will discard the newly computed value 4632N/A * and retry the {@code get} operation. 4632N/A * Discarding and retrying is an important proviso, 4639N/A * since otherwise {@code T} could potentially install 4639N/A * a disastrously stale value. For example: 4639N/A * <li>{@code T} calls {@code CV.get(C)} and sees state {@code 2N} 4632N/A * <li>{@code T} quickly computes a time-dependent value {@code V0} and gets ready to install it 4632N/A * <li>{@code T} is hit by an unlucky paging or scheduling event, and goes to sleep for a long time 4632N/A * <li>...meanwhile, {@code T2} also calls {@code CV.get(C)} and sees state {@code 2N} 4632N/A * <li>{@code T2} quickly computes a similar time-dependent value {@code V1} and installs it on {@code CV.get(C)} 4695N/A * <li>{@code T2} (or a third thread) then calls {@code CV.remove(C)}, undoing {@code T2}'s work 4695N/A * <li> the previous actions of {@code T2} are repeated several times 4695N/A * <li> also, the relevant computed values change over time: {@code V1}, {@code V2}, ... 4695N/A * <li>...meanwhile, {@code T} wakes up and attempts to install {@code V0}; <em>this must fail</em> 4695N/A * We can assume in the above scenario that {@code CV.computeValue} uses locks to properly 4639N/A * observe the time-dependent states as it computes {@code V1}, etc. 4639N/A * This does not remove the threat of a stale value, since there is a window of time 4632N/A * between the return of {@code computeValue} in {@code T} and the installation 4668N/A * of the the new value. No user synchronization is possible during this time. 4632N/A * @param type the type whose class value must be removed 4632N/A * @throws NullPointerException if the argument is null 4632N/A // Possible functionality for JSR 292 MR 1 4639N/A /** Return the cache, if it exists, else a dummy empty cache. */ 4632N/A // racing type.classValueMap{.cacheArray} : null => new Entry[X] <=> new Entry[Y] 4632N/A // invariant: returned value is safe to dereference and check for an Entry 4668N/A /** Initial, one-element, empty cache used by all Class instances. Must never be filled. */ 4668N/A * Slow tail of ClassValue.get to retry at nearby locations in the cache, 4668N/A * or take a slow lock and check the hash table. 4668N/A * Called only if the first probe was empty or a collision. 4668N/A * This is a separate method, so compilers can process it independently. 4639N/A // Hack to suppress warnings on the (T) cast, which is a no-op. 4639N/A /** Called when the fast path of get fails, and cache reprobe also fails. 4639N/A // The fail-safe recovery is to fall back to the underlying classValueMap. 4632N/A // Try to make a real entry for the promised version. 4632N/A // Whether computeValue throws or returns normally, 4632N/A // be sure to remove the empty entry. 4632N/A // else try again, in case a racing thread called remove (so e == null) 4632N/A /** Check that e is non-null, matches this ClassValue, and is live. */ 4639N/A // racing e.version : null (blank) => unique Version token => null (GC-ed version) 4639N/A // non-racing this.version : v1 => v2 => ... (updates are read faithfully from volatile) 4639N/A // invariant: No false positives on version match. Null is OK for false negative. 4639N/A // invariant: If version matches, then e.value is readable (final set in Entry.<init>) 4888N/A /** Internal hash code for accessing Class.classValueMap.cacheArray. */ 4888N/A /** Value stream for hashCodeForCache. See similar structure in ThreadLocal. */ 4888N/A /** Good for power-of-two tables. See similar structure in ThreadLocal. */ 4639N/A /** Mask a hash code to be positive but not too large, to prevent wraparound. */ 4639N/A * Private key for retrieval of this object from ClassValueMap. 4639N/A * This ClassValue's identity, expressed as an opaque object. 4639N/A * The main object {@code ClassValue.this} is incorrect since 4639N/A * subclasses may override {@code ClassValue.equals}, which 4639N/A * could confuse keys in the ClassValueMap. 4660N/A * Current version for retrieving this class value from the cache. 4639N/A * Any number of computeValue calls can be cached in association with one version. 4660N/A * But the version changes when a remove (on any type) is executed. 4639N/A * A version change invalidates all cache entries for the affected ClassValue, 4639N/A * by marking them as stale. Stale cache entries do not force another call 4639N/A * to computeValue, but they do require a synchronized visit to a backing map. 4639N/A * All user-visible state changes on the ClassValue take place under 4639N/A * a lock inside the synchronized methods of ClassValueMap. 4639N/A * Readers (of ClassValue.get) are notified of such state changes 4660N/A * when this.version is bumped to a new token. 4660N/A * This variable must be volatile so that an unsynchronized reader 4660N/A * will receive the notification without delay. 4660N/A * If version were not volatile, one thread T1 could persistently hold onto 4639N/A * a stale value this.value == V1, while while another thread T2 advances 4639N/A * (under a lock) to this.value == V2. This will typically be harmless, 4639N/A * but if T1 and T2 interact causally via some other channel, such that 4639N/A * T1's further actions are constrained (in the JMM) to happen after 4639N/A * the V2 event, then T1's observation of V1 will be an error. 4639N/A * The practical effect of making this.version be volatile is that it cannot 4639N/A * be hoisted out of a loop (by an optimizing JIT) or otherwise cached. 4639N/A * Some machines may also require a barrier instruction to execute /** One binding of a value to a class via a ClassValue. * <li> promise if value == Entry.this * <li> else dead if version == null * <li> else stale if version != classValue.version * Promises are never put into the cache; they only live in the * backing map while a computeValue call is in flight. * Once an entry goes stale, it can be reset at any time final Object value;
// usually of type T, but sometimes (Entry)this this.
value =
value;
// for a regular entry, value is of type T /** For creating a promise. */ this.
value =
this;
// for a promise, value is not of type T, but Entry! /** Fetch the value. This entry must not be a promise. */ if (v ==
null)
return false;
// value = null -- caller must drop /** Return the backing map associated with this type. */ // racing type.classValueMap : null (blank) => unique ClassValueMap // if a null is observed, a map is created (lazily, synchronously, uniquely) // all further access to that map is synchronized // happens about once per type // Note that explicitVersion might be different from this.version. // As soon as the Entry is put into the cache, the value will be // reachable via a data race (as defined by the Java Memory Model). // This race is benign, assuming the value object itself can be // read safely by multiple threads. This is up to the user. // The entry and version fields themselves can be safely read via // a race because they are either final or have controlled states. // If the pointer from the entry to the version is still null, // or if the version goes immediately dead and is nulled out, // the reader will take the slow path and retry under a lock. // The following class could also be top level and non-public: /** A backing map for all ClassValues, relative a single given type. * Gives a fully serialized "true state" for each pair (ClassValue cv, Class type). * Also manages an unserialized fast-path cache. /** Number of entries initially allocated to each type when first used with any ClassValue. * It would be pointless to make this much smaller than the Class and ClassValueMap objects themselves. /** Build a backing map for ClassValues, relative the given type. * Also, create an empty cache array and install it on the class. /** Initiate a query. Store a promise (placeholder) if there is no value yet. */ // The presence of a promise means that a value is pending for v. // Eventually, finishEntry will overwrite the promise. // Note that the promise is never entered into the cache! // Somebody else has asked the same question. // there is already a completed entry here; report it // There is a stale but valid entry here; make it fresh again. // Once an entry is in the hash table, we don't care what its version is. // Add to the cache, to enable the fast path, next time. /** Finish a query. Overwrite a matching placeholder. Drop stale incoming values. */ // We can get here during exception processing, unwinding from computeValue. // If e0 matches the intended entry, there has not been a remove call // between the previous startEntry and now. So now overwrite e0. // Add to the cache, to enable the fast path, next time. // Some sort of mismatch; caller must try again. // make all cache elements for this guy go stale: /** Change the value for an entry. */ // no value change => no version change needed // Add to the cache, to enable the fast path, next time. // Statics do not need synchronization. /** Load the cache entry at the given (hashed) location. */ // non-racing cache.length : constant // racing cache[i & (mask)] : null <=> Entry // invariant: returned value is null or well-constructed (ready to match) /** Look in the cache, at the home location for the given ClassValue. */ /** Given that first probe was a collision, retry at nearby locations. */ // Probe the cache carefully, in a range of slots. return null;
// if nobody is at home, no need to search nearby // assume !classValue.match(e2), but do not assert, because of races break;
// only search within non-null runs // relocate colliding entry e2 (from cache[home]) to first empty slot ?
e2 // put e2 here if it fits // Remember first empty slot, if any: /** How far out of place is e? */ if (
cv ==
null)
return 0;
// entry is not live! /// Below this line all functions are private, and assume synchronized access. /** Make sure the cache load stays below its limit, if possible. */ /** Remove stale entries in the given range. * Should be executed under a Map lock. continue;
// skip null and live entries // avoid breaking up a non-null run /** Clearing a cache slot risks disconnecting following entries * from the head of a non-null run, which would allow them * to be found via reprobes. Find an entry after cache[begin] * to plug into the hole, or return null if none is needed. if (
e2 ==
null)
break;
// End of non-null run. if (!
e2.
isLive())
continue;
// Doomed anyway. if (
dis2 ==
0)
continue;
// e2 already optimally placed // e2 can replace entry at cache[home1] // Put e2 exactly where he belongs. // And keep going, so we can favor larger dislocations. // Be conservative, to avoid breaking up a non-null run. /** Remove stale entries in the range near classValue. */ /** Remove all stale entries, everywhere. */ /** Add the given entry to the cache, in its home location, unless it is out of date. */ /** Add the given entry to the cache, in its home location. */ if (
e2 ==
null)
return;
// done // try to move e2 somewhere else in his probe range // Note: At this point, e2 is just dropped from the cache. /** Store the given entry. Update cacheLoad, and return any live victim. * 'Gently' means return self rather than dislocating a live victim. // do not overwrite a live entry /** Note an entry that is about to be overwritten. * If it is not live, quietly replace it by null. * If it is an actual null, increment cacheLoad, * because the caller is going to store something /** Percent loading of cache before resize. */ /** Maximum number of probes to attempt. */ // N.B. Set PROBE_LIMIT=0 to disable all fast paths.