ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/jsr166/jsr166/src/jsr166e/Striped64.java
Revision: 1.3
Committed: Fri Aug 5 17:08:04 2011 UTC (12 years, 9 months ago) by jsr166
Branch: MAIN
Changes since 1.2: +2 -2 lines
Log Message:
typos

File Contents

# 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 import java.util.concurrent.atomic.AtomicInteger;
10 import java.util.concurrent.atomic.AtomicLong;
11
12 /**
13 * A package-local class holding common representation and mechanics
14 * for classes supporting dynamic striping on 64bit values. The class
15 * extends Number so that concrete subclasses can publicly do so.
16 */
17 abstract class Striped64 extends Number {
18 /*
19 * This class maintains a lazily-initialized table of atomically
20 * updated variables, plus an extra "base" field. The table size
21 * is a power of two. Indexing uses masked per-thread hash codes.
22 * Nearly all declarations in this class are package-private,
23 * accessed directly by subclasses.
24 *
25 * Table entries are of class Cell; a variant of AtomicLong padded
26 * to reduce cache contention on most processors. Padding is
27 * overkill for most Atomics because they are usually irregularly
28 * scattered in memory and thus don't interfere much with each
29 * other. But Atomic objects residing in arrays will tend to be
30 * placed adjacent to each other, and so will most often share
31 * cache lines (with a huge negative performance impact) without
32 * this precaution.
33 *
34 * In part because Cells are relatively large, we avoid creating
35 * them until they are needed. When there is no contention, all
36 * updates are made to the base field. Upon first contention (a
37 * failed CAS on base update), the table is initialized to size 2.
38 * The table size is doubled upon further contention until
39 * reaching the nearest power of two greater than or equal to the
40 * number of CPUS. Table slots remain empty (null) until they are
41 * needed.
42 *
43 * A single spinlock ("busy") is used for initializing and
44 * resizing the table, as well as populating slots with new Cells.
45 * There is no need for a blocking lock: When the lock is not
46 * available, threads try other slots (or the base). During these
47 * retries, there is increased contention and reduced locality,
48 * which is still better than alternatives.
49 *
50 * Per-thread hash codes are initialized to random values.
51 * Contention and/or table collisions are indicated by failed
52 * CASes when performing an update operation (see method
53 * retryUpdate). Upon a collision, if the table size is less than
54 * the capacity, it is doubled in size unless some other thread
55 * holds the lock. If a hashed slot is empty, and lock is
56 * available, a new Cell is created. Otherwise, if the slot
57 * exists, a CAS is tried. Retries proceed by "double hashing",
58 * using a secondary hash (Marsaglia XorShift) to try to find a
59 * free slot.
60 *
61 * The table size is capped because, when there are more threads
62 * than CPUs, supposing that each thread were bound to a CPU,
63 * there would exist a perfect hash function mapping threads to
64 * slots that eliminates collisions. When we reach capacity, we
65 * search for this mapping by randomly varying the hash codes of
66 * colliding threads. Because search is random, and collisions
67 * only become known via CAS failures, convergence can be slow,
68 * and because threads are typically not bound to CPUS forever,
69 * may not occur at all. However, despite these limitations,
70 * observed contention rates are typically low in these cases.
71 *
72 * It is possible for a Cell to become unused when threads that
73 * once hashed to it terminate, as well as in the case where
74 * doubling the table causes no thread to hash to it under
75 * expanded mask. We do not try to detect or remove such cells,
76 * under the assumption that for long-running instances, observed
77 * contention levels will recur, so the cells will eventually be
78 * needed again; and for short-lived ones, it does not matter.
79 */
80
81 /**
82 * Padded variant of AtomicLong supporting only raw accesses plus
83 * CAS The value field is placed between pads, hoping that the JVM
84 * doesn't reorder them.
85 *
86 * JVM intrinsics note: It would be possible to use a release-only
87 * form of CAS here, if it were provided.
88 */
89 static final class Cell {
90 volatile long p0, p1, p2, p3, p4, p5, p6;
91 volatile long value;
92 volatile long q0, q1, q2, q3, q4, q5, q6;
93 Cell(long x) { value = x; }
94
95 final boolean cas(long cmp, long val) {
96 return UNSAFE.compareAndSwapLong(this, valueOffset, cmp, val);
97 }
98
99 // Unsafe mechanics
100 private static final sun.misc.Unsafe UNSAFE;
101 private static final long valueOffset;
102 static {
103 try {
104 UNSAFE = getUnsafe();
105 Class<?> ak = Cell.class;
106 valueOffset = UNSAFE.objectFieldOffset
107 (ak.getDeclaredField("value"));
108 } catch (Exception e) {
109 throw new Error(e);
110 }
111 }
112
113 }
114
115 /**
116 * Holder for the thread-local hash code. The code is initially
117 * random, but may be set to a different value upon collisions.
118 */
119 static final class HashCode {
120 static final Random rng = new Random();
121 int code;
122 HashCode() {
123 int h = rng.nextInt(); // Avoid zero to allow xorShift rehash
124 code = (h == 0) ? 1 : h;
125 }
126 }
127
128 /**
129 * The corresponding ThreadLocal class
130 */
131 static final class ThreadHashCode extends ThreadLocal<HashCode> {
132 public HashCode initialValue() { return new HashCode(); }
133 }
134
135 /**
136 * Static per-thread hash codes. Shared across all instances to
137 * reduce ThreadLocal pollution and because adjustments due to
138 * collisions in one table are likely to be appropriate for
139 * others.
140 */
141 static final ThreadHashCode threadHashCode = new ThreadHashCode();
142
143 /** Number of CPUS, to place bound on table size */
144 static final int NCPU = Runtime.getRuntime().availableProcessors();
145
146 /**
147 * Table of cells. When non-null, size is a power of 2.
148 */
149 transient volatile Cell[] cells;
150
151 /**
152 * Base value, used mainly when there is no contention, but also as
153 * a fallback during table initialization races. Updated via CAS.
154 */
155 transient volatile long base;
156
157 /**
158 * Spinlock (locked via CAS) used when resizing and/or creating Cells.
159 */
160 transient volatile int busy;
161
162 /**
163 * Package-private default constructor
164 */
165 Striped64() {
166 }
167
168 /**
169 * CAS the base field
170 */
171 final boolean casBase(long cmp, long val) {
172 return UNSAFE.compareAndSwapLong(this, baseOffset, cmp, val);
173 }
174
175 /**
176 * CAS the busy field from 0 to 1 to acquire lock.
177 */
178 final boolean casBusy() {
179 return UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1);
180 }
181
182 /**
183 * Computes the function of current and new value. Subclasses
184 * should open-code this update function for most uses, but the
185 * virtualized form is needed within retryUpdate.
186 *
187 * @param currentValue the current value (of either base or a cell)
188 * @param newValue the argument from a user update call
189 * @return result of the update function
190 */
191 abstract long fn(long currentValue, long newValue);
192
193 /**
194 * Handles cases of updates involving initialization, resizing,
195 * creating new Cells, and/or contention. See above for
196 * explanation. This method suffers the usual non-modularity
197 * problems of optimistic retry code, relying on rechecked sets of
198 * reads.
199 *
200 * @param x the value
201 * @param hc the hash code holder
202 * @param wasUncontended false if CAS failed before call
203 */
204 final void retryUpdate(long x, HashCode hc, boolean wasUncontended) {
205 int h = hc.code;
206 boolean collide = false; // True if last slot nonempty
207 for (;;) {
208 Cell[] as; Cell a; int n; long v;
209 if ((as = cells) != null && (n = as.length) > 0) {
210 if ((a = as[(n - 1) & h]) == null) {
211 if (busy == 0) { // Try to attach new Cell
212 Cell r = new Cell(x); // Optimistically create
213 if (busy == 0 && casBusy()) {
214 boolean created = false;
215 try { // Recheck under lock
216 Cell[] rs; int m, j;
217 if ((rs = cells) != null &&
218 (m = rs.length) > 0 &&
219 rs[j = (m - 1) & h] == null) {
220 rs[j] = r;
221 created = true;
222 }
223 } finally {
224 busy = 0;
225 }
226 if (created)
227 break;
228 continue; // Slot is now non-empty
229 }
230 }
231 collide = false;
232 }
233 else if (!wasUncontended) // CAS already known to fail
234 wasUncontended = true; // Continue after rehash
235 else if (a.cas(v = a.value, fn(v, x)))
236 break;
237 else if (n >= NCPU || cells != as)
238 collide = false; // At max size or stale
239 else if (!collide)
240 collide = true;
241 else if (busy == 0 && casBusy()) {
242 try {
243 if (cells == as) { // Expand table unless stale
244 Cell[] rs = new Cell[n << 1];
245 for (int i = 0; i < n; ++i)
246 rs[i] = as[i];
247 cells = rs;
248 }
249 } finally {
250 busy = 0;
251 }
252 collide = false;
253 continue; // Retry with expanded table
254 }
255 h ^= h << 13; // Rehash
256 h ^= h >>> 17;
257 h ^= h << 5;
258 }
259 else if (busy == 0 && cells == as && casBusy()) {
260 boolean init = false;
261 try { // Initialize table
262 if (cells == as) {
263 Cell[] rs = new Cell[2];
264 rs[h & 1] = new Cell(x);
265 cells = rs;
266 init = true;
267 }
268 } finally {
269 busy = 0;
270 }
271 if (init)
272 break;
273 }
274 else if (casBase(v = base, fn(v, x)))
275 break; // Fall back on using base
276 }
277 hc.code = h; // Record index for next time
278 }
279
280
281 /**
282 * Set base and all cells to the given value
283 */
284 final void internalReset(long initialValue) {
285 Cell[] as = cells;
286 base = initialValue;
287 if (as != null) {
288 int n = as.length;
289 for (int i = 0; i < n; ++i) {
290 Cell a = as[i];
291 if (a != null)
292 a.value = initialValue;
293 }
294 }
295 }
296
297 // Unsafe mechanics
298 private static final sun.misc.Unsafe UNSAFE;
299 private static final long baseOffset;
300 private static final long busyOffset;
301 static {
302 try {
303 UNSAFE = getUnsafe();
304 Class<?> sk = Striped64.class;
305 baseOffset = UNSAFE.objectFieldOffset
306 (sk.getDeclaredField("base"));
307 busyOffset = UNSAFE.objectFieldOffset
308 (sk.getDeclaredField("busy"));
309 } catch (Exception e) {
310 throw new Error(e);
311 }
312 }
313
314 /**
315 * Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package.
316 * Replace with a simple call to Unsafe.getUnsafe when integrating
317 * into a jdk.
318 *
319 * @return a sun.misc.Unsafe
320 */
321 private static sun.misc.Unsafe getUnsafe() {
322 try {
323 return sun.misc.Unsafe.getUnsafe();
324 } catch (SecurityException se) {
325 try {
326 return java.security.AccessController.doPrivileged
327 (new java.security
328 .PrivilegedExceptionAction<sun.misc.Unsafe>() {
329 public sun.misc.Unsafe run() throws Exception {
330 java.lang.reflect.Field f = sun.misc
331 .Unsafe.class.getDeclaredField("theUnsafe");
332 f.setAccessible(true);
333 return (sun.misc.Unsafe) f.get(null);
334 }});
335 } catch (java.security.PrivilegedActionException e) {
336 throw new RuntimeException("Could not initialize intrinsics",
337 e.getCause());
338 }
339 }
340 }
341
342 }