src/jsr166e/Striped64.java

/*
 * Written by Doug Lea with assistance from members of JCP JSR-166
 * Expert Group and released to the public domain, as explained at
 * http://creativecommons.org/publicdomain/zero/1.0/
 */

package jsr166e;
import java.util.Random;

/**
 * A package-local class holding common representation and mechanics
 * for classes supporting dynamic striping on 64bit values. The class
 * extends Number so that concrete subclasses must publicly do so.
 */
abstract class Striped64 extends Number {
    /*
     * This class maintains a lazily-initialized table of atomically
     * updated variables, plus an extra "base" field. The table size
     * is a power of two. Indexing uses masked per-thread hash codes.
     * Nearly all declarations in this class are package-private,
     * accessed directly by subclasses.
     *
     * Table entries are of class Cell; a variant of AtomicLong padded
     * to reduce cache contention on most processors. Padding is
     * overkill for most Atomics because they are usually irregularly
     * scattered in memory and thus don't interfere much with each
     * other. But Atomic objects residing in arrays will tend to be
     * placed adjacent to each other, and so will most often share
     * cache lines (with a huge negative performance impact) without
     * this precaution.
     *
     * In part because Cells are relatively large, we avoid creating
     * them until they are needed.  When there is no contention, all
     * updates are made to the base field.  Upon first contention (a
     * failed CAS on base update), the table is initialized to size 2.
     * The table size is doubled upon further contention until
     * reaching the nearest power of two greater than or equal to the
     * number of CPUS. Table slots remain empty (null) until they are
     * needed.
     *
     * A single spinlock ("busy") is used for initializing and
     * resizing the table, as well as populating slots with new Cells.
     * There is no need for a blocking lock; when the lock is not
     * available, threads try other slots (or the base).  During these
     * retries, there is increased contention and reduced locality,
     * which is still better than alternatives.
     *
     * Per-thread hash codes are initialized to random values.
     * Contention and/or table collisions are indicated by failed
     * CASes when performing an update operation (see method
     * retryUpdate). Upon a collision, if the table size is less than
     * the capacity, it is doubled in size unless some other thread
     * holds the lock. If a hashed slot is empty, and lock is
     * available, a new Cell is created. Otherwise, if the slot
     * exists, a CAS is tried.  Retries proceed by "double hashing",
     * using a secondary hash (Marsaglia XorShift) to try to find a
     * free slot.
     *
     * The table size is capped because, when there are more threads
     * than CPUs, supposing that each thread were bound to a CPU,
     * there would exist a perfect hash function mapping threads to
     * slots that eliminates collisions. When we reach capacity, we
     * search for this mapping by randomly varying the hash codes of
     * colliding threads.  Because search is random, and collisions
     * only become known via CAS failures, convergence can be slow,
     * and because threads are typically not bound to CPUS forever,
     * may not occur at all. However, despite these limitations,
     * observed contention rates are typically low in these cases.
     *
     * It is possible for a Cell to become unused when threads that
     * once hashed to it terminate, as well as in the case where
     * doubling the table causes no thread to hash to it under
     * expanded mask.  We do not try to detect or remove such cells,
     * under the assumption that for long-running instances, observed
     * contention levels will recur, so the cells will eventually be
     * needed again; and for short-lived ones, it does not matter.
     */

    /**
     * Padded variant of AtomicLong supporting only raw accesses plus CAS.
     * The value field is placed between pads, hoping that the JVM doesn't
     * reorder them.
     *
     * JVM intrinsics note: It would be possible to use a release-only
     * form of CAS here, if it were provided.
     */
    static final class Cell {
        volatile long p0, p1, p2, p3, p4, p5, p6;
        volatile long value;
        volatile long q0, q1, q2, q3, q4, q5, q6;
        Cell(long x) { value = x; }

        final boolean cas(long cmp, long val) {
            return UNSAFE.compareAndSwapLong(this, valueOffset, cmp, val);
        }

        // Unsafe mechanics
        private static final sun.misc.Unsafe UNSAFE;
        private static final long valueOffset;
        static {
            try {
                UNSAFE = getUnsafe();
                Class<?> ak = Cell.class;
                valueOffset = UNSAFE.objectFieldOffset
                    (ak.getDeclaredField("value"));
            } catch (Exception e) {
                throw new Error(e);
            }
        }

    }

    /**
     * ThreadLocal holding a single-slot int array holding hash code.
     * Unlike the JDK8 version of this class, we use a suboptimal
     * int[] representation to avoid introducing a new type that can
     * impede class-unloading when ThreadLocals are not removed.
     */
    static final ThreadLocal<int[]> threadHashCode = new ThreadLocal<int[]>();

    /**
     * Generator of new random hash codes
     */
    static final Random rng = new Random();

    /** Number of CPUS, to place bound on table size */
    static final int NCPU = Runtime.getRuntime().availableProcessors();

    /**
     * Table of cells. When non-null, size is a power of 2.
     */
    transient volatile Cell[] cells;

    /**
     * Base value, used mainly when there is no contention, but also as
     * a fallback during table initialization races. Updated via CAS.
     */
    transient volatile long base;

    /**
     * Spinlock (locked via CAS) used when resizing and/or creating Cells.
     */
    transient volatile int busy;

    /**
     * Package-private default constructor
     */
    Striped64() {
    }

    /**
     * CASes the base field.
     */
    final boolean casBase(long cmp, long val) {
        return UNSAFE.compareAndSwapLong(this, baseOffset, cmp, val);
    }

    /**
     * CASes the busy field from 0 to 1 to acquire lock.
     */
    final boolean casBusy() {
        return UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1);
    }

    /**
     * Computes the function of current and new value. Subclasses
     * should open-code this update function for most uses, but the
     * virtualized form is needed within retryUpdate.
     *
     * @param currentValue the current value (of either base or a cell)
     * @param newValue the argument from a user update call
     * @return result of the update function
     */
    abstract long fn(long currentValue, long newValue);

    /**
     * Handles cases of updates involving initialization, resizing,
     * creating new Cells, and/or contention. See above for
     * explanation. This method suffers the usual non-modularity
     * problems of optimistic retry code, relying on rechecked sets of
     * reads.
     *
     * @param x the value
     * @param hc the hash code holder
     * @param wasUncontended false if CAS failed before call
     */
    final void retryUpdate(long x, int[] hc, boolean wasUncontended) {
        int h;
        if (hc == null) {
            threadHashCode.set(hc = new int[1]); // Initialize randomly
            int r = rng.nextInt(); // Avoid zero to allow xorShift rehash
            h = hc[0] = (r == 0) ? 1 : r;
        }
        else
            h = hc[0];
        boolean collide = false;                // True if last slot nonempty
        for (;;) {
            Cell[] as; Cell a; int n; long v;
            if ((as = cells) != null && (n = as.length) > 0) {
                if ((a = as[(n - 1) & h]) == null) {
                    if (busy == 0) {            // Try to attach new Cell
                        Cell r = new Cell(x);   // Optimistically create
                        if (busy == 0 && casBusy()) {
                            boolean created = false;
                            try {               // Recheck under lock
                                Cell[] rs; int m, j;
                                if ((rs = cells) != null &&
                                    (m = rs.length) > 0 &&
                                    rs[j = (m - 1) & h] == null) {
                                    rs[j] = r;
                                    created = true;
                                }
                            } finally {
                                busy = 0;
                            }
                            if (created)
                                break;
                            continue;           // Slot is now non-empty
                        }
                    }
                    collide = false;
                }
                else if (!wasUncontended)       // CAS already known to fail
                    wasUncontended = true;      // Continue after rehash
                else if (a.cas(v = a.value, fn(v, x)))
                    break;
                else if (n >= NCPU || cells != as)
                    collide = false;            // At max size or stale
                else if (!collide)
                    collide = true;
                else if (busy == 0 && casBusy()) {
                    try {
                        if (cells == as) {      // Expand table unless stale
                            Cell[] rs = new Cell[n << 1];
                            for (int i = 0; i < n; ++i)
                                rs[i] = as[i];
                            cells = rs;
                        }
                    } finally {
                        busy = 0;
                    }
                    collide = false;
                    continue;                   // Retry with expanded table
                }
                h ^= h << 13;                   // Rehash
                h ^= h >>> 17;
                h ^= h << 5;
                hc[0] = h;                      // Record index for next time
            }
            else if (busy == 0 && cells == as && casBusy()) {
                boolean init = false;
                try {                           // Initialize table
                    if (cells == as) {
                        Cell[] rs = new Cell[2];
                        rs[h & 1] = new Cell(x);
                        cells = rs;
                        init = true;
                    }
                } finally {
                    busy = 0;
                }
                if (init)
                    break;
            }
            else if (casBase(v = base, fn(v, x)))
                break;                          // Fall back on using base
        }
    }


    /**
     * Sets base and all cells to the given value.
     */
    final void internalReset(long initialValue) {
        Cell[] as = cells;
        base = initialValue;
        if (as != null) {
            int n = as.length;
            for (int i = 0; i < n; ++i) {
                Cell a = as[i];
                if (a != null)
                    a.value = initialValue;
            }
        }
    }

    // Unsafe mechanics
    private static final sun.misc.Unsafe UNSAFE;
    private static final long baseOffset;
    private static final long busyOffset;
    static {
        try {
            UNSAFE = getUnsafe();
            Class<?> sk = Striped64.class;
            baseOffset = UNSAFE.objectFieldOffset
                (sk.getDeclaredField("base"));
            busyOffset = UNSAFE.objectFieldOffset
                (sk.getDeclaredField("busy"));
        } catch (Exception e) {
            throw new Error(e);
        }
    }

    /**
     * Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
     * Replace with a simple call to Unsafe.getUnsafe when integrating
     * into a jdk.
     *
     * @return a sun.misc.Unsafe
     */
    private static sun.misc.Unsafe getUnsafe() {
        try {
            return sun.misc.Unsafe.getUnsafe();
        } catch (SecurityException tryReflectionInstead) {}
        try {
            return java.security.AccessController.doPrivileged
            (new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() {
                public sun.misc.Unsafe run() throws Exception {
                    Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class;
                    for (java.lang.reflect.Field f : k.getDeclaredFields()) {
                        f.setAccessible(true);
                        Object x = f.get(null);
                        if (k.isInstance(x))
                            return k.cast(x);
                    }
                    throw new NoSuchFieldError("the Unsafe");
                }});
        } catch (java.security.PrivilegedActionException e) {
            throw new RuntimeException("Could not initialize intrinsics",
                                       e.getCause());
        }
    }
}
Revision:	1.9
Committed:	Sun May 4 22:38:19 2014 UTC (10 years ago) by dl
Branch:	MAIN
Changes since 1.8:	+17 -25 lines
Log Message:	Use int[] instead of HashCode holder class to enable class-unloading
#	Content
1	/*
2	* Written by Doug Lea with assistance from members of JCP JSR-166
3	* Expert Group and released to the public domain, as explained at
4	* http://creativecommons.org/publicdomain/zero/1.0/
5	*/
6
7	package jsr166e;
8	import java.util.Random;
9
10	/**
11	* A package-local class holding common representation and mechanics
12	* for classes supporting dynamic striping on 64bit values. The class
13	* extends Number so that concrete subclasses must publicly do so.
14	*/
15	abstract class Striped64 extends Number {
16	/*
17	* This class maintains a lazily-initialized table of atomically
18	* updated variables, plus an extra "base" field. The table size
19	* is a power of two. Indexing uses masked per-thread hash codes.
20	* Nearly all declarations in this class are package-private,
21	* accessed directly by subclasses.
22	*
23	* Table entries are of class Cell; a variant of AtomicLong padded
24	* to reduce cache contention on most processors. Padding is
25	* overkill for most Atomics because they are usually irregularly
26	* scattered in memory and thus don't interfere much with each
27	* other. But Atomic objects residing in arrays will tend to be
28	* placed adjacent to each other, and so will most often share
29	* cache lines (with a huge negative performance impact) without
30	* this precaution.
31	*
32	* In part because Cells are relatively large, we avoid creating
33	* them until they are needed. When there is no contention, all
34	* updates are made to the base field. Upon first contention (a
35	* failed CAS on base update), the table is initialized to size 2.
36	* The table size is doubled upon further contention until
37	* reaching the nearest power of two greater than or equal to the
38	* number of CPUS. Table slots remain empty (null) until they are
39	* needed.
40	*
41	* A single spinlock ("busy") is used for initializing and
42	* resizing the table, as well as populating slots with new Cells.
43	* There is no need for a blocking lock; when the lock is not
44	* available, threads try other slots (or the base). During these
45	* retries, there is increased contention and reduced locality,
46	* which is still better than alternatives.
47	*
48	* Per-thread hash codes are initialized to random values.
49	* Contention and/or table collisions are indicated by failed
50	* CASes when performing an update operation (see method
51	* retryUpdate). Upon a collision, if the table size is less than
52	* the capacity, it is doubled in size unless some other thread
53	* holds the lock. If a hashed slot is empty, and lock is
54	* available, a new Cell is created. Otherwise, if the slot
55	* exists, a CAS is tried. Retries proceed by "double hashing",
56	* using a secondary hash (Marsaglia XorShift) to try to find a
57	* free slot.
58	*
59	* The table size is capped because, when there are more threads
60	* than CPUs, supposing that each thread were bound to a CPU,
61	* there would exist a perfect hash function mapping threads to
62	* slots that eliminates collisions. When we reach capacity, we
63	* search for this mapping by randomly varying the hash codes of
64	* colliding threads. Because search is random, and collisions
65	* only become known via CAS failures, convergence can be slow,
66	* and because threads are typically not bound to CPUS forever,
67	* may not occur at all. However, despite these limitations,
68	* observed contention rates are typically low in these cases.
69	*
70	* It is possible for a Cell to become unused when threads that
71	* once hashed to it terminate, as well as in the case where
72	* doubling the table causes no thread to hash to it under
73	* expanded mask. We do not try to detect or remove such cells,
74	* under the assumption that for long-running instances, observed
75	* contention levels will recur, so the cells will eventually be
76	* needed again; and for short-lived ones, it does not matter.
77	*/
78
79	/**
80	* Padded variant of AtomicLong supporting only raw accesses plus CAS.
81	* The value field is placed between pads, hoping that the JVM doesn't
82	* reorder them.
83	*
84	* JVM intrinsics note: It would be possible to use a release-only
85	* form of CAS here, if it were provided.
86	*/
87	static final class Cell {
88	volatile long p0, p1, p2, p3, p4, p5, p6;
89	volatile long value;
90	volatile long q0, q1, q2, q3, q4, q5, q6;
91	Cell(long x) { value = x; }
92
93	final boolean cas(long cmp, long val) {
94	return UNSAFE.compareAndSwapLong(this, valueOffset, cmp, val);
95	}
96
97	// Unsafe mechanics
98	private static final sun.misc.Unsafe UNSAFE;
99	private static final long valueOffset;
100	static {
101	try {
102	UNSAFE = getUnsafe();
103	Class<?> ak = Cell.class;
104	valueOffset = UNSAFE.objectFieldOffset
105	(ak.getDeclaredField("value"));
106	} catch (Exception e) {
107	throw new Error(e);
108	}
109	}
110
111	}
112
113	/**
114	* ThreadLocal holding a single-slot int array holding hash code.
115	* Unlike the JDK8 version of this class, we use a suboptimal
116	* int[] representation to avoid introducing a new type that can
117	* impede class-unloading when ThreadLocals are not removed.
118	*/
119	static final ThreadLocal<int[]> threadHashCode = new ThreadLocal<int[]>();
120
121	/**
122	* Generator of new random hash codes
123	*/
124	static final Random rng = new Random();
125
126	/** Number of CPUS, to place bound on table size */
127	static final int NCPU = Runtime.getRuntime().availableProcessors();
128
129	/**
130	* Table of cells. When non-null, size is a power of 2.
131	*/
132	transient volatile Cell[] cells;
133
134	/**
135	* Base value, used mainly when there is no contention, but also as
136	* a fallback during table initialization races. Updated via CAS.
137	*/
138	transient volatile long base;
139
140	/**
141	* Spinlock (locked via CAS) used when resizing and/or creating Cells.
142	*/
143	transient volatile int busy;
144
145	/**
146	* Package-private default constructor
147	*/
148	Striped64() {
149	}
150
151	/**
152	* CASes the base field.
153	*/
154	final boolean casBase(long cmp, long val) {
155	return UNSAFE.compareAndSwapLong(this, baseOffset, cmp, val);
156	}
157
158	/**
159	* CASes the busy field from 0 to 1 to acquire lock.
160	*/
161	final boolean casBusy() {
162	return UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1);
163	}
164
165	/**
166	* Computes the function of current and new value. Subclasses
167	* should open-code this update function for most uses, but the
168	* virtualized form is needed within retryUpdate.
169	*
170	* @param currentValue the current value (of either base or a cell)
171	* @param newValue the argument from a user update call
172	* @return result of the update function
173	*/
174	abstract long fn(long currentValue, long newValue);
175
176	/**
177	* Handles cases of updates involving initialization, resizing,
178	* creating new Cells, and/or contention. See above for
179	* explanation. This method suffers the usual non-modularity
180	* problems of optimistic retry code, relying on rechecked sets of
181	* reads.
182	*
183	* @param x the value
184	* @param hc the hash code holder
185	* @param wasUncontended false if CAS failed before call
186	*/
187	final void retryUpdate(long x, int[] hc, boolean wasUncontended) {
188	int h;
189	if (hc == null) {
190	threadHashCode.set(hc = new int[1]); // Initialize randomly
191	int r = rng.nextInt(); // Avoid zero to allow xorShift rehash
192	h = hc[0] = (r == 0) ? 1 : r;
193	}
194	else
195	h = hc[0];
196	boolean collide = false; // True if last slot nonempty
197	for (;;) {
198	Cell[] as; Cell a; int n; long v;
199	if ((as = cells) != null && (n = as.length) > 0) {
200	if ((a = as[(n - 1) & h]) == null) {
201	if (busy == 0) { // Try to attach new Cell
202	Cell r = new Cell(x); // Optimistically create
203	if (busy == 0 && casBusy()) {
204	boolean created = false;
205	try { // Recheck under lock
206	Cell[] rs; int m, j;
207	if ((rs = cells) != null &&
208	(m = rs.length) > 0 &&
209	rs[j = (m - 1) & h] == null) {
210	rs[j] = r;
211	created = true;
212	}
213	} finally {
214	busy = 0;
215	}
216	if (created)
217	break;
218	continue; // Slot is now non-empty
219	}
220	}
221	collide = false;
222	}
223	else if (!wasUncontended) // CAS already known to fail
224	wasUncontended = true; // Continue after rehash
225	else if (a.cas(v = a.value, fn(v, x)))
226	break;
227	else if (n >= NCPU \|\| cells != as)
228	collide = false; // At max size or stale
229	else if (!collide)
230	collide = true;
231	else if (busy == 0 && casBusy()) {
232	try {
233	if (cells == as) { // Expand table unless stale
234	Cell[] rs = new Cell[n << 1];
235	for (int i = 0; i < n; ++i)
236	rs[i] = as[i];
237	cells = rs;
238	}
239	} finally {
240	busy = 0;
241	}
242	collide = false;
243	continue; // Retry with expanded table
244	}
245	h ^= h << 13; // Rehash
246	h ^= h >>> 17;
247	h ^= h << 5;
248	hc[0] = h; // Record index for next time
249	}
250	else if (busy == 0 && cells == as && casBusy()) {
251	boolean init = false;
252	try { // Initialize table
253	if (cells == as) {
254	Cell[] rs = new Cell[2];
255	rs[h & 1] = new Cell(x);
256	cells = rs;
257	init = true;
258	}
259	} finally {
260	busy = 0;
261	}
262	if (init)
263	break;
264	}
265	else if (casBase(v = base, fn(v, x)))
266	break; // Fall back on using base
267	}
268	}
269
270
271	/**
272	* Sets base and all cells to the given value.
273	*/
274	final void internalReset(long initialValue) {
275	Cell[] as = cells;
276	base = initialValue;
277	if (as != null) {
278	int n = as.length;
279	for (int i = 0; i < n; ++i) {
280	Cell a = as[i];
281	if (a != null)
282	a.value = initialValue;
283	}
284	}
285	}
286
287	// Unsafe mechanics
288	private static final sun.misc.Unsafe UNSAFE;
289	private static final long baseOffset;
290	private static final long busyOffset;
291	static {
292	try {
293	UNSAFE = getUnsafe();
294	Class<?> sk = Striped64.class;
295	baseOffset = UNSAFE.objectFieldOffset
296	(sk.getDeclaredField("base"));
297	busyOffset = UNSAFE.objectFieldOffset
298	(sk.getDeclaredField("busy"));
299	} catch (Exception e) {
300	throw new Error(e);
301	}
302	}
303
304	/**
305	* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package.
306	* Replace with a simple call to Unsafe.getUnsafe when integrating
307	* into a jdk.
308	*
309	* @return a sun.misc.Unsafe
310	*/
311	private static sun.misc.Unsafe getUnsafe() {
312	try {
313	return sun.misc.Unsafe.getUnsafe();
314	} catch (SecurityException tryReflectionInstead) {}
315	try {
316	return java.security.AccessController.doPrivileged
317	(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() {
318	public sun.misc.Unsafe run() throws Exception {
319	Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class;
320	for (java.lang.reflect.Field f : k.getDeclaredFields()) {
321	f.setAccessible(true);
322	Object x = f.get(null);
323	if (k.isInstance(x))
324	return k.cast(x);
325	}
326	throw new NoSuchFieldError("the Unsafe");
327	}});
328	} catch (java.security.PrivilegedActionException e) {
329	throw new RuntimeException("Could not initialize intrinsics",
330	e.getCause());
331	}
332	}
333	}