1 /* 2 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 3 * 4 * This code is free software; you can redistribute it and/or modify it 5 * under the terms of the GNU General Public License version 2 only, as 6 * published by the Free Software Foundation. Oracle designates this 7 * particular file as subject to the "Classpath" exception as provided 8 * by Oracle in the LICENSE file that accompanied this code. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 */ 24 25 /* 26 * This file is available under and governed by the GNU General Public 27 * License version 2 only, as published by the Free Software Foundation. 28 * However, the following notice accompanied the original version of this 29 * file: 30 * 31 * Written by Doug Lea with assistance from members of JCP JSR-166 32 * Expert Group and released to the public domain, as explained at 33 * http://creativecommons.org/publicdomain/zero/1.0/ 34 */ 35 36 package java.util.concurrent.atomic; 37 38 import java.lang.invoke.MethodHandles; 39 import java.lang.invoke.VarHandle; 40 import java.util.Arrays; 41 import java.util.concurrent.ThreadLocalRandom; 42 import java.util.function.DoubleBinaryOperator; 43 import java.util.function.LongBinaryOperator; 44 45 /** 46 * A package-local class holding common representation and mechanics 47 * for classes supporting dynamic striping on 64bit values. The class 48 * extends Number so that concrete subclasses must publicly do so. 49 */ 50 @SuppressWarnings("serial") 51 abstract class Striped64 extends Number { 52 /* 53 * This class maintains a lazily-initialized table of atomically 54 * updated variables, plus an extra "base" field. The table size 55 * is a power of two. Indexing uses masked per-thread hash codes. 56 * Nearly all declarations in this class are package-private, 57 * accessed directly by subclasses. 58 * 59 * Table entries are of class Cell; a variant of AtomicLong padded 60 * (via @Contended) to reduce cache contention. Padding is 61 * overkill for most Atomics because they are usually irregularly 62 * scattered in memory and thus don't interfere much with each 63 * other. But Atomic objects residing in arrays will tend to be 64 * placed adjacent to each other, and so will most often share 65 * cache lines (with a huge negative performance impact) without 66 * this precaution. 67 * 68 * In part because Cells are relatively large, we avoid creating 69 * them until they are needed. When there is no contention, all 70 * updates are made to the base field. Upon first contention (a 71 * failed CAS on base update), the table is initialized to size 2. 72 * The table size is doubled upon further contention until 73 * reaching the nearest power of two greater than or equal to the 74 * number of CPUS. Table slots remain empty (null) until they are 75 * needed. 76 * 77 * A single spinlock ("cellsBusy") is used for initializing and 78 * resizing the table, as well as populating slots with new Cells. 79 * There is no need for a blocking lock; when the lock is not 80 * available, threads try other slots (or the base). During these 81 * retries, there is increased contention and reduced locality, 82 * which is still better than alternatives. 83 * 84 * The Thread probe fields maintained via ThreadLocalRandom serve 85 * as per-thread hash codes. We let them remain uninitialized as 86 * zero (if they come in this way) until they contend at slot 87 * 0. They are then initialized to values that typically do not 88 * often conflict with others. Contention and/or table collisions 89 * are indicated by failed CASes when performing an update 90 * operation. Upon a collision, if the table size is less than 91 * the capacity, it is doubled in size unless some other thread 92 * holds the lock. If a hashed slot is empty, and lock is 93 * available, a new Cell is created. Otherwise, if the slot 94 * exists, a CAS is tried. Retries proceed by "double hashing", 95 * using a secondary hash (Marsaglia XorShift) to try to find a 96 * free slot. 97 * 98 * The table size is capped because, when there are more threads 99 * than CPUs, supposing that each thread were bound to a CPU, 100 * there would exist a perfect hash function mapping threads to 101 * slots that eliminates collisions. When we reach capacity, we 102 * search for this mapping by randomly varying the hash codes of 103 * colliding threads. Because search is random, and collisions 104 * only become known via CAS failures, convergence can be slow, 105 * and because threads are typically not bound to CPUS forever, 106 * may not occur at all. However, despite these limitations, 107 * observed contention rates are typically low in these cases. 108 * 109 * It is possible for a Cell to become unused when threads that 110 * once hashed to it terminate, as well as in the case where 111 * doubling the table causes no thread to hash to it under 112 * expanded mask. We do not try to detect or remove such cells, 113 * under the assumption that for long-running instances, observed 114 * contention levels will recur, so the cells will eventually be 115 * needed again; and for short-lived ones, it does not matter. 116 */ 117 118 /** 119 * Padded variant of AtomicLong supporting only raw accesses plus CAS. 120 * 121 * JVM intrinsics note: It would be possible to use a release-only 122 * form of CAS here, if it were provided. 123 */ 124 @jdk.internal.vm.annotation.Contended static final class Cell { 125 volatile long value; 126 Cell(long x) { value = x; } 127 final boolean cas(long cmp, long val) { 128 return VALUE.compareAndSet(this, cmp, val); 129 } 130 final void reset() { 131 VALUE.setVolatile(this, 0L); 132 } 133 final void reset(long identity) { 134 VALUE.setVolatile(this, identity); 135 } 136 final long getAndSet(long val) { 137 return (long)VALUE.getAndSet(this, val); 138 } 139 140 // VarHandle mechanics 141 private static final VarHandle VALUE; 142 static { 143 try { 144 MethodHandles.Lookup l = MethodHandles.lookup(); 145 VALUE = l.findVarHandle(Cell.class, "value", long.class); 146 } catch (ReflectiveOperationException e) { 147 throw new ExceptionInInitializerError(e); 148 } 149 } 150 } 151 152 /** Number of CPUS, to place bound on table size */ 153 static final int NCPU = Runtime.getRuntime().availableProcessors(); 154 155 /** 156 * Table of cells. When non-null, size is a power of 2. 157 */ 158 transient volatile Cell[] cells; 159 160 /** 161 * Base value, used mainly when there is no contention, but also as 162 * a fallback during table initialization races. Updated via CAS. 163 */ 164 transient volatile long base; 165 166 /** 167 * Spinlock (locked via CAS) used when resizing and/or creating Cells. 168 */ 169 transient volatile int cellsBusy; 170 171 /** 172 * Package-private default constructor. 173 */ 174 Striped64() { 175 } 176 177 /** 178 * CASes the base field. 179 */ 180 final boolean casBase(long cmp, long val) { 181 return BASE.compareAndSet(this, cmp, val); 182 } 183 184 final long getAndSetBase(long val) { 185 return (long)BASE.getAndSet(this, val); 186 } 187 188 /** 189 * CASes the cellsBusy field from 0 to 1 to acquire lock. 190 */ 191 final boolean casCellsBusy() { 192 return CELLSBUSY.compareAndSet(this, 0, 1); 193 } 194 195 /** 196 * Returns the probe value for the current thread. 197 * Duplicated from ThreadLocalRandom because of packaging restrictions. 198 */ 199 static final int getProbe() { 200 return (int) THREAD_PROBE.get(Thread.currentThread()); 201 } 202 203 /** 204 * Pseudo-randomly advances and records the given probe value for the 205 * given thread. 206 * Duplicated from ThreadLocalRandom because of packaging restrictions. 207 */ 208 static final int advanceProbe(int probe) { 209 probe ^= probe << 13; // xorshift 210 probe ^= probe >>> 17; 211 probe ^= probe << 5; 212 THREAD_PROBE.set(Thread.currentThread(), probe); 213 return probe; 214 } 215 216 /** 217 * Handles cases of updates involving initialization, resizing, 218 * creating new Cells, and/or contention. See above for 219 * explanation. This method suffers the usual non-modularity 220 * problems of optimistic retry code, relying on rechecked sets of 221 * reads. 222 * 223 * @param x the value 224 * @param fn the update function, or null for add (this convention 225 * avoids the need for an extra field or function in LongAdder). 226 * @param wasUncontended false if CAS failed before call 227 */ 228 final void longAccumulate(long x, LongBinaryOperator fn, 229 boolean wasUncontended) { 230 int h; 231 if ((h = getProbe()) == 0) { 232 ThreadLocalRandom.current(); // force initialization 233 h = getProbe(); 234 wasUncontended = true; 235 } 236 boolean collide = false; // True if last slot nonempty 237 done: for (;;) { 238 Cell[] cs; Cell c; int n; long v; 239 if ((cs = cells) != null && (n = cs.length) > 0) { 240 if ((c = cs[(n - 1) & h]) == null) { 241 if (cellsBusy == 0) { // Try to attach new Cell 242 Cell r = new Cell(x); // Optimistically create 243 if (cellsBusy == 0 && casCellsBusy()) { 244 try { // Recheck under lock 245 Cell[] rs; int m, j; 246 if ((rs = cells) != null && 247 (m = rs.length) > 0 && 248 rs[j = (m - 1) & h] == null) { 249 rs[j] = r; 250 break done; 251 } 252 } finally { 253 cellsBusy = 0; 254 } 255 continue; // Slot is now non-empty 256 } 257 } 258 collide = false; 259 } 260 else if (!wasUncontended) // CAS already known to fail 261 wasUncontended = true; // Continue after rehash 262 else if (c.cas(v = c.value, 263 (fn == null) ? v + x : fn.applyAsLong(v, x))) 264 break; 265 else if (n >= NCPU || cells != cs) 266 collide = false; // At max size or stale 267 else if (!collide) 268 collide = true; 269 else if (cellsBusy == 0 && casCellsBusy()) { 270 try { 271 if (cells == cs) // Expand table unless stale 272 cells = Arrays.copyOf(cs, n << 1); 273 } finally { 274 cellsBusy = 0; 275 } 276 collide = false; 277 continue; // Retry with expanded table 278 } 279 h = advanceProbe(h); 280 } 281 else if (cellsBusy == 0 && cells == cs && casCellsBusy()) { 282 try { // Initialize table 283 if (cells == cs) { 284 Cell[] rs = new Cell[2]; 285 rs[h & 1] = new Cell(x); 286 cells = rs; 287 break done; 288 } 289 } finally { 290 cellsBusy = 0; 291 } 292 } 293 // Fall back on using base 294 else if (casBase(v = base, 295 (fn == null) ? v + x : fn.applyAsLong(v, x))) 296 break done; 297 } 298 } 299 300 private static long apply(DoubleBinaryOperator fn, long v, double x) { 301 double d = Double.longBitsToDouble(v); 302 d = (fn == null) ? d + x : fn.applyAsDouble(d, x); 303 return Double.doubleToRawLongBits(d); 304 } 305 306 /** 307 * Same as longAccumulate, but injecting long/double conversions 308 * in too many places to sensibly merge with long version, given 309 * the low-overhead requirements of this class. So must instead be 310 * maintained by copy/paste/adapt. 311 */ 312 final void doubleAccumulate(double x, DoubleBinaryOperator fn, 313 boolean wasUncontended) { 314 int h; 315 if ((h = getProbe()) == 0) { 316 ThreadLocalRandom.current(); // force initialization 317 h = getProbe(); 318 wasUncontended = true; 319 } 320 boolean collide = false; // True if last slot nonempty 321 done: for (;;) { 322 Cell[] cs; Cell c; int n; long v; 323 if ((cs = cells) != null && (n = cs.length) > 0) { 324 if ((c = cs[(n - 1) & h]) == null) { 325 if (cellsBusy == 0) { // Try to attach new Cell 326 Cell r = new Cell(Double.doubleToRawLongBits(x)); 327 if (cellsBusy == 0 && casCellsBusy()) { 328 try { // Recheck under lock 329 Cell[] rs; int m, j; 330 if ((rs = cells) != null && 331 (m = rs.length) > 0 && 332 rs[j = (m - 1) & h] == null) { 333 rs[j] = r; 334 break done; 335 } 336 } finally { 337 cellsBusy = 0; 338 } 339 continue; // Slot is now non-empty 340 } 341 } 342 collide = false; 343 } 344 else if (!wasUncontended) // CAS already known to fail 345 wasUncontended = true; // Continue after rehash 346 else if (c.cas(v = c.value, apply(fn, v, x))) 347 break; 348 else if (n >= NCPU || cells != cs) 349 collide = false; // At max size or stale 350 else if (!collide) 351 collide = true; 352 else if (cellsBusy == 0 && casCellsBusy()) { 353 try { 354 if (cells == cs) // Expand table unless stale 355 cells = Arrays.copyOf(cs, n << 1); 356 } finally { 357 cellsBusy = 0; 358 } 359 collide = false; 360 continue; // Retry with expanded table 361 } 362 h = advanceProbe(h); 363 } 364 else if (cellsBusy == 0 && cells == cs && casCellsBusy()) { 365 try { // Initialize table 366 if (cells == cs) { 367 Cell[] rs = new Cell[2]; 368 rs[h & 1] = new Cell(Double.doubleToRawLongBits(x)); 369 cells = rs; 370 break done; 371 } 372 } finally { 373 cellsBusy = 0; 374 } 375 } 376 // Fall back on using base 377 else if (casBase(v = base, apply(fn, v, x))) 378 break done; 379 } 380 } 381 382 // VarHandle mechanics 383 private static final VarHandle BASE; 384 private static final VarHandle CELLSBUSY; 385 private static final VarHandle THREAD_PROBE; 386 static { 387 try { 388 MethodHandles.Lookup l = MethodHandles.lookup(); 389 BASE = l.findVarHandle(Striped64.class, 390 "base", long.class); 391 CELLSBUSY = l.findVarHandle(Striped64.class, 392 "cellsBusy", int.class); 393 l = java.security.AccessController.doPrivileged( 394 new java.security.PrivilegedAction<>() { 395 public MethodHandles.Lookup run() { 396 try { 397 return MethodHandles.privateLookupIn(Thread.class, MethodHandles.lookup()); 398 } catch (ReflectiveOperationException e) { 399 throw new ExceptionInInitializerError(e); 400 } 401 }}); 402 THREAD_PROBE = l.findVarHandle(Thread.class, 403 "threadLocalRandomProbe", int.class); 404 } catch (ReflectiveOperationException e) { 405 throw new ExceptionInInitializerError(e); 406 } 407 } 408 409 }