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 |
|
9 |
import java.util.Random; |
10 |
|
11 |
/** |
12 |
* A package-local class holding common representation and mechanics |
13 |
* for classes supporting dynamic striping on 64bit values. The class |
14 |
* extends Number so that concrete subclasses must publicly do so. |
15 |
*/ |
16 |
abstract class Striped64 extends Number { |
17 |
/* |
18 |
* This class maintains a lazily-initialized table of atomically |
19 |
* updated variables, plus an extra "base" field. The table size |
20 |
* is a power of two. Indexing uses masked per-thread hash codes. |
21 |
* Nearly all declarations in this class are package-private, |
22 |
* accessed directly by subclasses. |
23 |
* |
24 |
* Table entries are of class Cell; a variant of AtomicLong padded |
25 |
* to reduce cache contention on most processors. Padding is |
26 |
* overkill for most Atomics because they are usually irregularly |
27 |
* scattered in memory and thus don't interfere much with each |
28 |
* other. But Atomic objects residing in arrays will tend to be |
29 |
* placed adjacent to each other, and so will most often share |
30 |
* cache lines (with a huge negative performance impact) without |
31 |
* this precaution. |
32 |
* |
33 |
* In part because Cells are relatively large, we avoid creating |
34 |
* them until they are needed. When there is no contention, all |
35 |
* updates are made to the base field. Upon first contention (a |
36 |
* failed CAS on base update), the table is initialized to size 2. |
37 |
* The table size is doubled upon further contention until |
38 |
* reaching the nearest power of two greater than or equal to the |
39 |
* number of CPUS. Table slots remain empty (null) until they are |
40 |
* needed. |
41 |
* |
42 |
* A single spinlock ("busy") is used for initializing and |
43 |
* resizing the table, as well as populating slots with new Cells. |
44 |
* There is no need for a blocking lock; when the lock is not |
45 |
* available, threads try other slots (or the base). During these |
46 |
* retries, there is increased contention and reduced locality, |
47 |
* which is still better than alternatives. |
48 |
* |
49 |
* Per-thread hash codes are initialized to random values. |
50 |
* Contention and/or table collisions are indicated by failed |
51 |
* CASes when performing an update operation (see method |
52 |
* retryUpdate). Upon a collision, if the table size is less than |
53 |
* the capacity, it is doubled in size unless some other thread |
54 |
* holds the lock. If a hashed slot is empty, and lock is |
55 |
* available, a new Cell is created. Otherwise, if the slot |
56 |
* exists, a CAS is tried. Retries proceed by "double hashing", |
57 |
* using a secondary hash (Marsaglia XorShift) to try to find a |
58 |
* free slot. |
59 |
* |
60 |
* The table size is capped because, when there are more threads |
61 |
* than CPUs, supposing that each thread were bound to a CPU, |
62 |
* there would exist a perfect hash function mapping threads to |
63 |
* slots that eliminates collisions. When we reach capacity, we |
64 |
* search for this mapping by randomly varying the hash codes of |
65 |
* colliding threads. Because search is random, and collisions |
66 |
* only become known via CAS failures, convergence can be slow, |
67 |
* and because threads are typically not bound to CPUS forever, |
68 |
* may not occur at all. However, despite these limitations, |
69 |
* observed contention rates are typically low in these cases. |
70 |
* |
71 |
* It is possible for a Cell to become unused when threads that |
72 |
* once hashed to it terminate, as well as in the case where |
73 |
* doubling the table causes no thread to hash to it under |
74 |
* expanded mask. We do not try to detect or remove such cells, |
75 |
* under the assumption that for long-running instances, observed |
76 |
* contention levels will recur, so the cells will eventually be |
77 |
* needed again; and for short-lived ones, it does not matter. |
78 |
*/ |
79 |
|
80 |
/** |
81 |
* Padded variant of AtomicLong supporting only raw accesses plus CAS. |
82 |
* The value field is placed between pads, hoping that the JVM doesn't |
83 |
* reorder them. |
84 |
* |
85 |
* JVM intrinsics note: It would be possible to use a release-only |
86 |
* form of CAS here, if it were provided. |
87 |
*/ |
88 |
static final class Cell { |
89 |
volatile long p0, p1, p2, p3, p4, p5, p6; |
90 |
volatile long value; |
91 |
volatile long q0, q1, q2, q3, q4, q5, q6; |
92 |
Cell(long x) { value = x; } |
93 |
|
94 |
final boolean cas(long cmp, long val) { |
95 |
return UNSAFE.compareAndSwapLong(this, valueOffset, cmp, val); |
96 |
} |
97 |
|
98 |
// Unsafe mechanics |
99 |
private static final sun.misc.Unsafe UNSAFE; |
100 |
private static final long valueOffset; |
101 |
static { |
102 |
try { |
103 |
UNSAFE = getUnsafe(); |
104 |
Class<?> ak = Cell.class; |
105 |
valueOffset = UNSAFE.objectFieldOffset |
106 |
(ak.getDeclaredField("value")); |
107 |
} catch (Exception e) { |
108 |
throw new Error(e); |
109 |
} |
110 |
} |
111 |
|
112 |
} |
113 |
|
114 |
/** |
115 |
* ThreadLocal holding a single-slot int array holding hash code. |
116 |
* Unlike the JDK8 version of this class, we use a suboptimal |
117 |
* int[] representation to avoid introducing a new type that can |
118 |
* impede class-unloading when ThreadLocals are not removed. |
119 |
*/ |
120 |
static final ThreadLocal<int[]> threadHashCode = new ThreadLocal<int[]>(); |
121 |
|
122 |
/** |
123 |
* Generator of new random hash codes |
124 |
*/ |
125 |
static final Random rng = new Random(); |
126 |
|
127 |
/** Number of CPUS, to place bound on table size */ |
128 |
static final int NCPU = Runtime.getRuntime().availableProcessors(); |
129 |
|
130 |
/** |
131 |
* Table of cells. When non-null, size is a power of 2. |
132 |
*/ |
133 |
transient volatile Cell[] cells; |
134 |
|
135 |
/** |
136 |
* Base value, used mainly when there is no contention, but also as |
137 |
* a fallback during table initialization races. Updated via CAS. |
138 |
*/ |
139 |
transient volatile long base; |
140 |
|
141 |
/** |
142 |
* Spinlock (locked via CAS) used when resizing and/or creating Cells. |
143 |
*/ |
144 |
transient volatile int busy; |
145 |
|
146 |
/** |
147 |
* Package-private default constructor |
148 |
*/ |
149 |
Striped64() { |
150 |
} |
151 |
|
152 |
/** |
153 |
* CASes the base field. |
154 |
*/ |
155 |
final boolean casBase(long cmp, long val) { |
156 |
return UNSAFE.compareAndSwapLong(this, baseOffset, cmp, val); |
157 |
} |
158 |
|
159 |
/** |
160 |
* CASes the busy field from 0 to 1 to acquire lock. |
161 |
*/ |
162 |
final boolean casBusy() { |
163 |
return UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1); |
164 |
} |
165 |
|
166 |
/** |
167 |
* Computes the function of current and new value. Subclasses |
168 |
* should open-code this update function for most uses, but the |
169 |
* virtualized form is needed within retryUpdate. |
170 |
* |
171 |
* @param currentValue the current value (of either base or a cell) |
172 |
* @param newValue the argument from a user update call |
173 |
* @return result of the update function |
174 |
*/ |
175 |
abstract long fn(long currentValue, long newValue); |
176 |
|
177 |
/** |
178 |
* Handles cases of updates involving initialization, resizing, |
179 |
* creating new Cells, and/or contention. See above for |
180 |
* explanation. This method suffers the usual non-modularity |
181 |
* problems of optimistic retry code, relying on rechecked sets of |
182 |
* reads. |
183 |
* |
184 |
* @param x the value |
185 |
* @param hc the hash code holder |
186 |
* @param wasUncontended false if CAS failed before call |
187 |
*/ |
188 |
final void retryUpdate(long x, int[] hc, boolean wasUncontended) { |
189 |
int h; |
190 |
if (hc == null) { |
191 |
threadHashCode.set(hc = new int[1]); // Initialize randomly |
192 |
int r = rng.nextInt(); // Avoid zero to allow xorShift rehash |
193 |
h = hc[0] = (r == 0) ? 1 : r; |
194 |
} |
195 |
else |
196 |
h = hc[0]; |
197 |
boolean collide = false; // True if last slot nonempty |
198 |
for (;;) { |
199 |
Cell[] as; Cell a; int n; long v; |
200 |
if ((as = cells) != null && (n = as.length) > 0) { |
201 |
if ((a = as[(n - 1) & h]) == null) { |
202 |
if (busy == 0) { // Try to attach new Cell |
203 |
Cell r = new Cell(x); // Optimistically create |
204 |
if (busy == 0 && casBusy()) { |
205 |
boolean created = false; |
206 |
try { // Recheck under lock |
207 |
Cell[] rs; int m, j; |
208 |
if ((rs = cells) != null && |
209 |
(m = rs.length) > 0 && |
210 |
rs[j = (m - 1) & h] == null) { |
211 |
rs[j] = r; |
212 |
created = true; |
213 |
} |
214 |
} finally { |
215 |
busy = 0; |
216 |
} |
217 |
if (created) |
218 |
break; |
219 |
continue; // Slot is now non-empty |
220 |
} |
221 |
} |
222 |
collide = false; |
223 |
} |
224 |
else if (!wasUncontended) // CAS already known to fail |
225 |
wasUncontended = true; // Continue after rehash |
226 |
else if (a.cas(v = a.value, fn(v, x))) |
227 |
break; |
228 |
else if (n >= NCPU || cells != as) |
229 |
collide = false; // At max size or stale |
230 |
else if (!collide) |
231 |
collide = true; |
232 |
else if (busy == 0 && casBusy()) { |
233 |
try { |
234 |
if (cells == as) { // Expand table unless stale |
235 |
Cell[] rs = new Cell[n << 1]; |
236 |
for (int i = 0; i < n; ++i) |
237 |
rs[i] = as[i]; |
238 |
cells = rs; |
239 |
} |
240 |
} finally { |
241 |
busy = 0; |
242 |
} |
243 |
collide = false; |
244 |
continue; // Retry with expanded table |
245 |
} |
246 |
h ^= h << 13; // Rehash |
247 |
h ^= h >>> 17; |
248 |
h ^= h << 5; |
249 |
hc[0] = h; // Record index for next time |
250 |
} |
251 |
else if (busy == 0 && cells == as && casBusy()) { |
252 |
boolean init = false; |
253 |
try { // Initialize table |
254 |
if (cells == as) { |
255 |
Cell[] rs = new Cell[2]; |
256 |
rs[h & 1] = new Cell(x); |
257 |
cells = rs; |
258 |
init = true; |
259 |
} |
260 |
} finally { |
261 |
busy = 0; |
262 |
} |
263 |
if (init) |
264 |
break; |
265 |
} |
266 |
else if (casBase(v = base, fn(v, x))) |
267 |
break; // Fall back on using base |
268 |
} |
269 |
} |
270 |
|
271 |
|
272 |
/** |
273 |
* Sets base and all cells to the given value. |
274 |
*/ |
275 |
final void internalReset(long initialValue) { |
276 |
Cell[] as = cells; |
277 |
base = initialValue; |
278 |
if (as != null) { |
279 |
int n = as.length; |
280 |
for (int i = 0; i < n; ++i) { |
281 |
Cell a = as[i]; |
282 |
if (a != null) |
283 |
a.value = initialValue; |
284 |
} |
285 |
} |
286 |
} |
287 |
|
288 |
// Unsafe mechanics |
289 |
private static final sun.misc.Unsafe UNSAFE; |
290 |
private static final long baseOffset; |
291 |
private static final long busyOffset; |
292 |
static { |
293 |
try { |
294 |
UNSAFE = getUnsafe(); |
295 |
Class<?> sk = Striped64.class; |
296 |
baseOffset = UNSAFE.objectFieldOffset |
297 |
(sk.getDeclaredField("base")); |
298 |
busyOffset = UNSAFE.objectFieldOffset |
299 |
(sk.getDeclaredField("busy")); |
300 |
} catch (Exception e) { |
301 |
throw new Error(e); |
302 |
} |
303 |
} |
304 |
|
305 |
/** |
306 |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
307 |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
308 |
* into a jdk. |
309 |
* |
310 |
* @return a sun.misc.Unsafe |
311 |
*/ |
312 |
private static sun.misc.Unsafe getUnsafe() { |
313 |
try { |
314 |
return sun.misc.Unsafe.getUnsafe(); |
315 |
} catch (SecurityException tryReflectionInstead) {} |
316 |
try { |
317 |
return java.security.AccessController.doPrivileged |
318 |
(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
319 |
public sun.misc.Unsafe run() throws Exception { |
320 |
Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class; |
321 |
for (java.lang.reflect.Field f : k.getDeclaredFields()) { |
322 |
f.setAccessible(true); |
323 |
Object x = f.get(null); |
324 |
if (k.isInstance(x)) |
325 |
return k.cast(x); |
326 |
} |
327 |
throw new NoSuchFieldError("the Unsafe"); |
328 |
}}); |
329 |
} catch (java.security.PrivilegedActionException e) { |
330 |
throw new RuntimeException("Could not initialize intrinsics", |
331 |
e.getCause()); |
332 |
} |
333 |
} |
334 |
} |