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 |
import java.io.IOException; |
12 |
import java.io.Serializable; |
13 |
import java.io.ObjectInputStream; |
14 |
import java.io.ObjectOutputStream; |
15 |
|
16 |
/** |
17 |
* One or more variables that together maintain an initially zero sum. |
18 |
* When updates (method {@link #add}) are contended across threads, |
19 |
* the set of variables may grow dynamically to reduce contention. |
20 |
* |
21 |
* <p> This class is usually preferable to {@link AtomicLong} when |
22 |
* multiple threads update a common sum that is used for purposes such |
23 |
* as collecting statistics, not for fine-grained synchronization |
24 |
* control. Under low update contention, the two classes have similar |
25 |
* characteristics. But under high contention, expected throughput of |
26 |
* this class is significantly higher, at the expense of higher space |
27 |
* consumption. |
28 |
* |
29 |
* <p> Method {@link #sum} returns the current combined total across |
30 |
* the variables maintaining the sum. This value is <em>NOT</em> an |
31 |
* atomic snapshot: Invocation of <code>sum</code> in the absence of |
32 |
* concurrent updates returns an accurate result, but concurrent |
33 |
* updates that occur while the sum is being calculated might not be |
34 |
* incorporated. The sum may also be <code>reset</code> to zero, as |
35 |
* an alternative to creating a new adder. However, method {@link |
36 |
* #reset} is intrinsically racy, so should only be used when it is |
37 |
* known that no threads are concurrently updating the sum. |
38 |
* |
39 |
* <p><em>jsr166e note: This class is targeted to be placed in |
40 |
* java.util.concurrent.atomic<em> |
41 |
* |
42 |
* @author Doug Lea |
43 |
*/ |
44 |
public class LongAdder implements Serializable { |
45 |
private static final long serialVersionUID = 7249069246863182397L; |
46 |
|
47 |
/* |
48 |
* A LongAdder maintains a lazily-initialized table of atomically |
49 |
* updated variables, plus an extra "base" field. The table size |
50 |
* is a power of two. Indexing uses masked per-thread hash codes |
51 |
* |
52 |
* Table entries are of class Cell; a variant of AtomicLong padded |
53 |
* to reduce cache contention on most processors. Padding is |
54 |
* overkill for most Atomics because they are usually irregularly |
55 |
* scattered in memory and thus don't interfere much with each |
56 |
* other. But Atomic objects residing in arrays will tend to be |
57 |
* placed adjacent to each other, and so will most often share |
58 |
* cache lines (with a huge negative performance impact) without |
59 |
* this precaution. |
60 |
* |
61 |
* In part because Cells are relatively large, we avoid creating |
62 |
* them until they are needed. When there is no contention, all |
63 |
* updates are made to the base field. Upon first contention (a |
64 |
* failed CAS on base update), the table is initialized to size 2. |
65 |
* The table size is doubled upon further contention until |
66 |
* reaching the nearest power of two greater than or equal to the |
67 |
* number of CPUS. |
68 |
* |
69 |
* Per-thread hash codes are initialized to random values. |
70 |
* Contention and/or table collisions are indicated by failed |
71 |
* CASes when performing an add operation (see method |
72 |
* retryAdd). Upon a collision, if the table size is less than the |
73 |
* capacity, it is doubled in size unless some other thread holds |
74 |
* the lock. If a hashed slot is empty, and lock is available, a |
75 |
* new Cell is created. Otherwise, if the slot exists, a CAS is |
76 |
* tried. Retries proceed by "double hashing", using a secondary |
77 |
* hash (Marsaglia XorShift) to try to find a free slot. |
78 |
* |
79 |
* The table size is capped because, when there are more threads |
80 |
* than CPUs, supposing that each thread were bound to a CPU, |
81 |
* there would exist a perfect hash function mapping threads to |
82 |
* slots that eliminates collisions. When we reach capacity, we |
83 |
* search for this mapping by randomly varying the hash codes of |
84 |
* colliding threads. Because search is random, and collisions |
85 |
* only become known via CAS failures, convergence can be slow, |
86 |
* and because threads are typically not bound to CPUS forever, |
87 |
* may not occur at all. However, despite these limitations, |
88 |
* observed contention rates are typically low in these cases. |
89 |
* |
90 |
* A single spinlock is used for initializing and resizing the |
91 |
* table, as well as populating slots with new Cells. There is no |
92 |
* need for a blocking lock: Upon lock contention, threads try |
93 |
* other slots (or the base) rather than blocking. During these |
94 |
* retries, there is increased contention and reduced locality, |
95 |
* which is still better than alternatives. |
96 |
* |
97 |
* It is possible for a Cell to become unused when threads that |
98 |
* once hashed to it terminate, as well as in the case where |
99 |
* doubling the table causes no thread to hash to it under |
100 |
* expanded mask. We do not try to detect or remove such cells, |
101 |
* under the assumption that for long-running adders, observed |
102 |
* contention levels will recur, so the cells will eventually be |
103 |
* needed again; and for short-lived ones, it does not matter. |
104 |
* |
105 |
* JVM intrinsics note: It would be possible to use a release-only |
106 |
* form of CAS here, if it were provided. |
107 |
*/ |
108 |
|
109 |
/** |
110 |
* Padded variant of AtomicLong. The value field is placed |
111 |
* between pads, hoping that the JVM doesn't reorder them. |
112 |
* Updates are via inlined CAS in methods add and retryAdd. |
113 |
*/ |
114 |
static final class Cell { |
115 |
volatile long p0, p1, p2, p3, p4, p5, p6; |
116 |
volatile long value; |
117 |
volatile long q0, q1, q2, q3, q4, q5, q6; |
118 |
Cell(long x) { value = x; } |
119 |
} |
120 |
|
121 |
/** |
122 |
* Holder for the thread-local hash code. The code is initially |
123 |
* random, but may be set to a different value upon collisions. |
124 |
*/ |
125 |
static final class HashCode { |
126 |
static final Random rng = new Random(); |
127 |
int code; |
128 |
HashCode() { |
129 |
int h = rng.nextInt(); // Avoid zero to allow xorShift rehash |
130 |
code = (h == 0) ? 1 : h; |
131 |
} |
132 |
} |
133 |
|
134 |
/** |
135 |
* The corresponding ThreadLocal class |
136 |
*/ |
137 |
static final class ThreadHashCode extends ThreadLocal<HashCode> { |
138 |
public HashCode initialValue() { return new HashCode(); } |
139 |
} |
140 |
|
141 |
/** |
142 |
* Static per-thread hash codes. Shared across all LongAdders |
143 |
* to reduce ThreadLocal pollution and because adjustments due to |
144 |
* collisions in one table are likely to be appropriate for |
145 |
* others. |
146 |
*/ |
147 |
static final ThreadHashCode threadHashCode = new ThreadHashCode(); |
148 |
|
149 |
/** Nomber of CPUS, to place bound on table size */ |
150 |
private static final int NCPU = Runtime.getRuntime().availableProcessors(); |
151 |
|
152 |
/** |
153 |
* Table of cells. When non-null, size is a power of 2. |
154 |
*/ |
155 |
private transient volatile Cell[] cells; |
156 |
|
157 |
/** |
158 |
* Base sum, used mainly when there is no contention, but also as |
159 |
* a fallback during table initializion races. Updated via CAS. |
160 |
*/ |
161 |
private transient volatile long base; |
162 |
|
163 |
/** |
164 |
* Spinlock (locked via CAS) used when resizing and/or creating Cells. |
165 |
*/ |
166 |
private transient volatile int busy; |
167 |
|
168 |
/** |
169 |
* Creates a new adder with initial sum of zero. |
170 |
*/ |
171 |
public LongAdder() { |
172 |
} |
173 |
|
174 |
/** |
175 |
* Adds the given value. |
176 |
* |
177 |
* @param x the value to add |
178 |
*/ |
179 |
public void add(long x) { |
180 |
Cell[] as; long v; HashCode hc; Cell a; int n; |
181 |
if ((as = cells) != null || |
182 |
!UNSAFE.compareAndSwapLong(this, baseOffset, v = base, v + x)) { |
183 |
boolean uncontended = true; |
184 |
int h = (hc = threadHashCode.get()).code; |
185 |
if (as == null || (n = as.length) < 1 || |
186 |
(a = as[(n - 1) & h]) == null || |
187 |
!(uncontended = UNSAFE.compareAndSwapLong(a, valueOffset, |
188 |
v = a.value, v + x))) |
189 |
retryAdd(x, hc, uncontended); |
190 |
} |
191 |
} |
192 |
|
193 |
/** |
194 |
* Handle cases of add 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 to add |
201 |
* @param hc the hash code holder |
202 |
* @param wasUncontended false if CAS failed before call |
203 |
*/ |
204 |
private void retryAdd(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 && |
214 |
UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) { |
215 |
boolean created = false; |
216 |
try { // Recheck under lock |
217 |
Cell[] rs; int m, j; |
218 |
if ((rs = cells) != null && |
219 |
(m = rs.length) > 0 && |
220 |
rs[j = (m - 1) & h] == null) { |
221 |
rs[j] = r; |
222 |
created = true; |
223 |
} |
224 |
} finally { |
225 |
busy = 0; |
226 |
} |
227 |
if (created) |
228 |
break; |
229 |
continue; // Slot is now non-empty |
230 |
} |
231 |
} |
232 |
collide = false; |
233 |
} |
234 |
else if (!wasUncontended) // CAS already known to fail |
235 |
wasUncontended = true; // Continue after rehash |
236 |
else if (UNSAFE.compareAndSwapLong(a, valueOffset, |
237 |
v = a.value, v + x)) |
238 |
break; |
239 |
else if (n >= NCPU || cells != as) |
240 |
collide = false; // At max size or stale |
241 |
else if (!collide) |
242 |
collide = true; |
243 |
else if (busy == 0 && // Try to expand table |
244 |
UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) { |
245 |
try { |
246 |
if (cells == as) { |
247 |
Cell[] rs = new Cell[n << 1]; |
248 |
for (int i = 0; i < n; ++i) |
249 |
rs[i] = as[i]; |
250 |
cells = rs; |
251 |
} |
252 |
} finally { |
253 |
busy = 0; |
254 |
} |
255 |
collide = false; |
256 |
continue; // Retry with expanded table |
257 |
} |
258 |
h ^= h << 13; // Rehash |
259 |
h ^= h >>> 17; |
260 |
h ^= h << 5; |
261 |
} |
262 |
else if (busy == 0 && cells == as && |
263 |
UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) { |
264 |
boolean init = false; |
265 |
try { // Initialize table |
266 |
if (cells == as) { |
267 |
Cell[] rs = new Cell[2]; |
268 |
rs[h & 1] = new Cell(x); |
269 |
cells = rs; |
270 |
init = true; |
271 |
} |
272 |
} finally { |
273 |
busy = 0; |
274 |
} |
275 |
if (init) |
276 |
break; |
277 |
} |
278 |
else if (UNSAFE.compareAndSwapLong(this, baseOffset, |
279 |
v = base, v + x)) |
280 |
break; // Fall back on using base |
281 |
} |
282 |
hc.code = h; // Record index for next time |
283 |
} |
284 |
|
285 |
/** |
286 |
* Equivalent to {@code add(1)}. |
287 |
*/ |
288 |
public void increment() { |
289 |
add(1L); |
290 |
} |
291 |
|
292 |
/** |
293 |
* Equivalent to {@code add(-1)}. |
294 |
*/ |
295 |
public void decrement() { |
296 |
add(-1L); |
297 |
} |
298 |
|
299 |
/** |
300 |
* Returns the current sum. The result is only guaranteed to be |
301 |
* accurate in the absence of concurrent updates. Otherwise, it |
302 |
* may fail to reflect one or more updates occuring while |
303 |
* calculating the result. |
304 |
* |
305 |
* @return the sum |
306 |
*/ |
307 |
public long sum() { |
308 |
Cell[] as = cells; |
309 |
long sum = base; |
310 |
if (as != null) { |
311 |
int n = as.length; |
312 |
for (int i = 0; i < n; ++i) { |
313 |
Cell a = as[i]; |
314 |
if (a != null) |
315 |
sum += a.value; |
316 |
} |
317 |
} |
318 |
return sum; |
319 |
} |
320 |
|
321 |
/** |
322 |
* Resets variables maintaining the sum to zero. This is |
323 |
* effective in setting the sum to zero only if there are no |
324 |
* concurrent updates. |
325 |
*/ |
326 |
public void reset() { |
327 |
Cell[] as = cells; |
328 |
base = 0L; |
329 |
if (as != null) { |
330 |
int n = as.length; |
331 |
for (int i = 0; i < n; ++i) { |
332 |
Cell a = as[i]; |
333 |
if (a != null) |
334 |
a.value = 0L; |
335 |
} |
336 |
} |
337 |
} |
338 |
|
339 |
/** |
340 |
* Equivalent in effect to {@link #sum} followed by {@link |
341 |
* #reset}. This method may apply for example during quiescent |
342 |
* points between multithreaded computations. If there are |
343 |
* updates concurrent with this method, the returned value is |
344 |
* <em>not</em> guaranteed to be the final sum occurring before |
345 |
* the reset. |
346 |
* |
347 |
* @return the sum |
348 |
*/ |
349 |
public long sumThenReset() { |
350 |
Cell[] as = cells; |
351 |
long sum = base; |
352 |
base = 0L; |
353 |
if (as != null) { |
354 |
int n = as.length; |
355 |
for (int i = 0; i < n; ++i) { |
356 |
Cell a = as[i]; |
357 |
if (a != null) { |
358 |
sum += a.value; |
359 |
a.value = 0L; |
360 |
} |
361 |
} |
362 |
} |
363 |
return sum; |
364 |
} |
365 |
|
366 |
private void writeObject(java.io.ObjectOutputStream s) |
367 |
throws java.io.IOException { |
368 |
s.defaultWriteObject(); |
369 |
s.writeLong(sum()); |
370 |
} |
371 |
|
372 |
private void readObject(ObjectInputStream s) |
373 |
throws IOException, ClassNotFoundException { |
374 |
s.defaultReadObject(); |
375 |
busy = 0; |
376 |
cells = null; |
377 |
base = s.readLong(); |
378 |
} |
379 |
|
380 |
// Unsafe mechanics |
381 |
private static final sun.misc.Unsafe UNSAFE; |
382 |
private static final long baseOffset; |
383 |
private static final long busyOffset; |
384 |
private static final long valueOffset; |
385 |
static { |
386 |
try { |
387 |
UNSAFE = getUnsafe(); |
388 |
Class<?> sk = LongAdder.class; |
389 |
baseOffset = UNSAFE.objectFieldOffset |
390 |
(sk.getDeclaredField("base")); |
391 |
busyOffset = UNSAFE.objectFieldOffset |
392 |
(sk.getDeclaredField("busy")); |
393 |
Class<?> ak = Cell.class; |
394 |
valueOffset = UNSAFE.objectFieldOffset |
395 |
(ak.getDeclaredField("value")); |
396 |
} catch (Exception e) { |
397 |
throw new Error(e); |
398 |
} |
399 |
} |
400 |
|
401 |
/** |
402 |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
403 |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
404 |
* into a jdk. |
405 |
* |
406 |
* @return a sun.misc.Unsafe |
407 |
*/ |
408 |
private static sun.misc.Unsafe getUnsafe() { |
409 |
try { |
410 |
return sun.misc.Unsafe.getUnsafe(); |
411 |
} catch (SecurityException se) { |
412 |
try { |
413 |
return java.security.AccessController.doPrivileged |
414 |
(new java.security |
415 |
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
416 |
public sun.misc.Unsafe run() throws Exception { |
417 |
java.lang.reflect.Field f = sun.misc |
418 |
.Unsafe.class.getDeclaredField("theUnsafe"); |
419 |
f.setAccessible(true); |
420 |
return (sun.misc.Unsafe) f.get(null); |
421 |
}}); |
422 |
} catch (java.security.PrivilegedActionException e) { |
423 |
throw new RuntimeException("Could not initialize intrinsics", |
424 |
e.getCause()); |
425 |
} |
426 |
} |
427 |
} |
428 |
|
429 |
} |