folly/test/SharedMutexTest.cpp

   1 /*
   2  * Copyright 2015 Facebook, Inc.
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at
   7  *
   8  *   http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 #include <folly/SharedMutex.h>
  18
  19 #include <stdlib.h>
  20 #include <thread>
  21 #include <vector>
  22 #include <boost/optional.hpp>
  23 #include <folly/Benchmark.h>
  24 #include <folly/MPMCQueue.h>
  25 #include <folly/Random.h>
  26 #include <folly/test/DeterministicSchedule.h>
  27 #include <gflags/gflags.h>
  28 #include <gtest/gtest.h>
  29
  30 #include <boost/thread/shared_mutex.hpp>
  31 #include <folly/RWSpinLock.h>
  32
  33 using namespace folly;
  34 using namespace folly::test;
  35 using namespace std;
  36 using namespace chrono;
  37
  38 typedef DeterministicSchedule DSched;
  39 typedef SharedMutexImpl<true, void, DeterministicAtomic, true>
  40     DSharedMutexReadPriority;
  41 typedef SharedMutexImpl<false, void, DeterministicAtomic, true>
  42     DSharedMutexWritePriority;
  43
  44 COMMON_CONCURRENCY_SHARED_MUTEX_DECLARE_STATIC_STORAGE(
  45     DSharedMutexReadPriority);
  46 COMMON_CONCURRENCY_SHARED_MUTEX_DECLARE_STATIC_STORAGE(
  47     DSharedMutexWritePriority);
  48
  49 template <typename Lock>
  50 void runBasicTest() {
  51   Lock lock;
  52   SharedMutexToken token1;
  53   SharedMutexToken token2;
  54   SharedMutexToken token3;
  55
  56   EXPECT_TRUE(lock.try_lock());
  57   EXPECT_FALSE(lock.try_lock());
  58   EXPECT_FALSE(lock.try_lock_shared(token1));
  59   lock.unlock();
  60
  61   EXPECT_TRUE(lock.try_lock_shared(token1));
  62   EXPECT_FALSE(lock.try_lock());
  63   EXPECT_TRUE(lock.try_lock_shared(token2));
  64   lock.lock_shared(token3);
  65   lock.unlock_shared(token3);
  66   lock.unlock_shared(token2);
  67   lock.unlock_shared(token1);
  68
  69   lock.lock();
  70   lock.unlock();
  71
  72   lock.lock_shared(token1);
  73   lock.lock_shared(token2);
  74   lock.unlock_shared(token1);
  75   lock.unlock_shared(token2);
  76
  77   lock.lock();
  78   lock.unlock_and_lock_shared(token1);
  79   lock.lock_shared(token2);
  80   lock.unlock_shared(token2);
  81   lock.unlock_shared(token1);
  82 }
  83
  84 TEST(SharedMutex, basic) {
  85   runBasicTest<SharedMutexReadPriority>();
  86   runBasicTest<SharedMutexWritePriority>();
  87 }
  88
  89 template <typename Lock>
  90 void runBasicHoldersTest() {
  91   Lock lock;
  92   SharedMutexToken token;
  93
  94   {
  95     typename Lock::WriteHolder holder(lock);
  96     EXPECT_FALSE(lock.try_lock());
  97     EXPECT_FALSE(lock.try_lock_shared(token));
  98
  99     typename Lock::WriteHolder holder2(std::move(holder));
 100     typename Lock::WriteHolder holder3;
 101     holder3 = std::move(holder2);
 102
 103     typename Lock::UpgradeHolder holder4(std::move(holder3));
 104     typename Lock::WriteHolder holder5(std::move(holder4));
 105
 106     typename Lock::ReadHolder holder6(std::move(holder5));
 107
 108     EXPECT_FALSE(lock.try_lock());
 109     EXPECT_TRUE(lock.try_lock_shared(token));
 110     lock.unlock_shared(token);
 111   }
 112
 113   {
 114     typename Lock::WriteHolder holder(lock);
 115     EXPECT_FALSE(lock.try_lock());
 116   }
 117
 118   {
 119     typename Lock::ReadHolder holder(lock);
 120     typename Lock::ReadHolder holder2(lock);
 121     typename Lock::UpgradeHolder holder3(lock);
 122   }
 123
 124   {
 125     typename Lock::UpgradeHolder holder(lock);
 126     typename Lock::ReadHolder holder2(lock);
 127     typename Lock::ReadHolder holder3(std::move(holder));
 128   }
 129 }
 130
 131 TEST(SharedMutex, basic_holders) {
 132   runBasicHoldersTest<SharedMutexReadPriority>();
 133   runBasicHoldersTest<SharedMutexWritePriority>();
 134 }
 135
 136 template <typename Lock>
 137 void runManyReadLocksTestWithTokens() {
 138   Lock lock;
 139
 140   vector<SharedMutexToken> tokens;
 141   for (int i = 0; i < 1000; ++i) {
 142     tokens.emplace_back();
 143     EXPECT_TRUE(lock.try_lock_shared(tokens.back()));
 144   }
 145   for (auto& token : tokens) {
 146     lock.unlock_shared(token);
 147   }
 148   EXPECT_TRUE(lock.try_lock());
 149   lock.unlock();
 150 }
 151
 152 TEST(SharedMutex, many_read_locks_with_tokens) {
 153   runManyReadLocksTestWithTokens<SharedMutexReadPriority>();
 154   runManyReadLocksTestWithTokens<SharedMutexWritePriority>();
 155 }
 156
 157 template <typename Lock>
 158 void runManyReadLocksTestWithoutTokens() {
 159   Lock lock;
 160
 161   for (int i = 0; i < 1000; ++i) {
 162     EXPECT_TRUE(lock.try_lock_shared());
 163   }
 164   for (int i = 0; i < 1000; ++i) {
 165     lock.unlock_shared();
 166   }
 167   EXPECT_TRUE(lock.try_lock());
 168   lock.unlock();
 169 }
 170
 171 TEST(SharedMutex, many_read_locks_without_tokens) {
 172   runManyReadLocksTestWithoutTokens<SharedMutexReadPriority>();
 173   runManyReadLocksTestWithoutTokens<SharedMutexWritePriority>();
 174 }
 175
 176 template <typename Lock>
 177 void runTimeoutInPastTest() {
 178   Lock lock;
 179
 180   EXPECT_TRUE(lock.try_lock_for(milliseconds(0)));
 181   lock.unlock();
 182   EXPECT_TRUE(lock.try_lock_for(milliseconds(-1)));
 183   lock.unlock();
 184   EXPECT_TRUE(lock.try_lock_shared_for(milliseconds(0)));
 185   lock.unlock_shared();
 186   EXPECT_TRUE(lock.try_lock_shared_for(milliseconds(-1)));
 187   lock.unlock_shared();
 188   EXPECT_TRUE(lock.try_lock_until(system_clock::now() - milliseconds(1)));
 189   lock.unlock();
 190   EXPECT_TRUE(
 191       lock.try_lock_shared_until(system_clock::now() - milliseconds(1)));
 192   lock.unlock_shared();
 193   EXPECT_TRUE(lock.try_lock_until(steady_clock::now() - milliseconds(1)));
 194   lock.unlock();
 195   EXPECT_TRUE(
 196       lock.try_lock_shared_until(steady_clock::now() - milliseconds(1)));
 197   lock.unlock_shared();
 198 }
 199
 200 TEST(SharedMutex, timeout_in_past) {
 201   runTimeoutInPastTest<SharedMutexReadPriority>();
 202   runTimeoutInPastTest<SharedMutexWritePriority>();
 203 }
 204
 205 template <class Func>
 206 bool funcHasDuration(milliseconds expectedDuration, Func func) {
 207   // elapsed time should eventually fall within expectedDuration +- 25%
 208   for (int tries = 0; tries < 100; ++tries) {
 209     auto start = steady_clock::now();
 210     func();
 211     auto elapsed = steady_clock::now() - start;
 212     if (elapsed > expectedDuration - expectedDuration / 4 &&
 213         elapsed < expectedDuration + expectedDuration / 4) {
 214       return true;
 215     }
 216   }
 217   return false;
 218 }
 219
 220 template <typename Lock>
 221 void runFailingTryTimeoutTest() {
 222   Lock lock;
 223   lock.lock();
 224   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 225     EXPECT_FALSE(lock.try_lock_for(milliseconds(10)));
 226   }));
 227   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 228     typename Lock::Token token;
 229     EXPECT_FALSE(lock.try_lock_shared_for(milliseconds(10), token));
 230   }));
 231   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 232     EXPECT_FALSE(lock.try_lock_upgrade_for(milliseconds(10)));
 233   }));
 234   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 235     EXPECT_FALSE(lock.try_lock_until(steady_clock::now() + milliseconds(10)));
 236   }));
 237   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 238     typename Lock::Token token;
 239     EXPECT_FALSE(lock.try_lock_shared_until(
 240         steady_clock::now() + milliseconds(10), token));
 241   }));
 242   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 243     EXPECT_FALSE(
 244         lock.try_lock_upgrade_until(steady_clock::now() + milliseconds(10)));
 245   }));
 246   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 247     EXPECT_FALSE(lock.try_lock_until(system_clock::now() + milliseconds(10)));
 248   }));
 249   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 250     typename Lock::Token token;
 251     EXPECT_FALSE(lock.try_lock_shared_until(
 252         system_clock::now() + milliseconds(10), token));
 253   }));
 254   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 255     EXPECT_FALSE(
 256         lock.try_lock_upgrade_until(system_clock::now() + milliseconds(10)));
 257   }));
 258   lock.unlock();
 259
 260   lock.lock_shared();
 261   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 262     EXPECT_FALSE(lock.try_lock_for(milliseconds(10)));
 263   }));
 264   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 265     EXPECT_FALSE(lock.try_lock_until(steady_clock::now() + milliseconds(10)));
 266   }));
 267   EXPECT_TRUE(funcHasDuration(milliseconds(10), [&] {
 268     EXPECT_FALSE(lock.try_lock_until(system_clock::now() + milliseconds(10)));
 269   }));
 270   lock.unlock_shared();
 271
 272   lock.lock();
 273   for (int p = 0; p < 8; ++p) {
 274     EXPECT_FALSE(lock.try_lock_for(nanoseconds(1 << p)));
 275   }
 276   lock.unlock();
 277
 278   for (int p = 0; p < 8; ++p) {
 279     typename Lock::ReadHolder holder1(lock);
 280     typename Lock::ReadHolder holder2(lock);
 281     typename Lock::ReadHolder holder3(lock);
 282     EXPECT_FALSE(lock.try_lock_for(nanoseconds(1 << p)));
 283   }
 284 }
 285
 286 TEST(SharedMutex, failing_try_timeout) {
 287   runFailingTryTimeoutTest<SharedMutexReadPriority>();
 288   runFailingTryTimeoutTest<SharedMutexWritePriority>();
 289 }
 290
 291 template <typename Lock>
 292 void runBasicUpgradeTest() {
 293   Lock lock;
 294   typename Lock::Token token1;
 295   typename Lock::Token token2;
 296
 297   lock.lock_upgrade();
 298   EXPECT_FALSE(lock.try_lock());
 299   EXPECT_TRUE(lock.try_lock_shared(token1));
 300   lock.unlock_shared(token1);
 301   lock.unlock_upgrade();
 302
 303   lock.lock_upgrade();
 304   lock.unlock_upgrade_and_lock();
 305   EXPECT_FALSE(lock.try_lock_shared(token1));
 306   lock.unlock();
 307
 308   lock.lock_upgrade();
 309   lock.unlock_upgrade_and_lock_shared(token1);
 310   lock.lock_upgrade();
 311   lock.unlock_upgrade_and_lock_shared(token2);
 312   lock.unlock_shared(token1);
 313   lock.unlock_shared(token2);
 314
 315   lock.lock();
 316   lock.unlock_and_lock_upgrade();
 317   EXPECT_TRUE(lock.try_lock_shared(token1));
 318   lock.unlock_upgrade();
 319   lock.unlock_shared(token1);
 320 }
 321
 322 TEST(SharedMutex, basic_upgrade_tests) {
 323   runBasicUpgradeTest<SharedMutexReadPriority>();
 324   runBasicUpgradeTest<SharedMutexWritePriority>();
 325 }
 326
 327 TEST(SharedMutex, read_has_prio) {
 328   SharedMutexReadPriority lock;
 329   SharedMutexToken token1;
 330   SharedMutexToken token2;
 331   lock.lock_shared(token1);
 332   bool exclusiveAcquired = false;
 333   auto writer = thread([&] {
 334     lock.lock();
 335     exclusiveAcquired = true;
 336     lock.unlock();
 337   });
 338
 339   // lock() can't complete until we unlock token1, but it should stake
 340   // its claim with regards to other exclusive or upgrade locks.  We can
 341   // use try_lock_upgrade to poll for that eventuality.
 342   while (lock.try_lock_upgrade()) {
 343     lock.unlock_upgrade();
 344     this_thread::yield();
 345   }
 346   EXPECT_FALSE(exclusiveAcquired);
 347
 348   // Even though lock() is stuck we should be able to get token2
 349   EXPECT_TRUE(lock.try_lock_shared(token2));
 350   lock.unlock_shared(token1);
 351   lock.unlock_shared(token2);
 352   writer.join();
 353   EXPECT_TRUE(exclusiveAcquired);
 354 }
 355
 356 TEST(SharedMutex, write_has_prio) {
 357   SharedMutexWritePriority lock;
 358   SharedMutexToken token1;
 359   SharedMutexToken token2;
 360   lock.lock_shared(token1);
 361   auto writer = thread([&] {
 362     lock.lock();
 363     lock.unlock();
 364   });
 365
 366   // eventually lock() should block readers
 367   while (lock.try_lock_shared(token2)) {
 368     lock.unlock_shared(token2);
 369     this_thread::yield();
 370   }
 371
 372   lock.unlock_shared(token1);
 373   writer.join();
 374 }
 375
 376 struct TokenLocker {
 377   SharedMutexToken token;
 378
 379   template <typename T>
 380   void lock(T* lock) {
 381     lock->lock();
 382   }
 383
 384   template <typename T>
 385   void unlock(T* lock) {
 386     lock->unlock();
 387   }
 388
 389   template <typename T>
 390   void lock_shared(T* lock) {
 391     lock->lock_shared(token);
 392   }
 393
 394   template <typename T>
 395   void unlock_shared(T* lock) {
 396     lock->unlock_shared(token);
 397   }
 398 };
 399
 400 struct Locker {
 401   template <typename T>
 402   void lock(T* lock) {
 403     lock->lock();
 404   }
 405
 406   template <typename T>
 407   void unlock(T* lock) {
 408     lock->unlock();
 409   }
 410
 411   template <typename T>
 412   void lock_shared(T* lock) {
 413     lock->lock_shared();
 414   }
 415
 416   template <typename T>
 417   void unlock_shared(T* lock) {
 418     lock->unlock_shared();
 419   }
 420 };
 421
 422 struct EnterLocker {
 423   template <typename T>
 424   void lock(T* lock) {
 425     lock->lock(0);
 426   }
 427
 428   template <typename T>
 429   void unlock(T* lock) {
 430     lock->unlock();
 431   }
 432
 433   template <typename T>
 434   void lock_shared(T* lock) {
 435     lock->enter(0);
 436   }
 437
 438   template <typename T>
 439   void unlock_shared(T* lock) {
 440     lock->leave();
 441   }
 442 };
 443
 444 struct PosixRWLock {
 445   pthread_rwlock_t lock_;
 446
 447   PosixRWLock() { pthread_rwlock_init(&lock_, nullptr); }
 448
 449   ~PosixRWLock() { pthread_rwlock_destroy(&lock_); }
 450
 451   void lock() { pthread_rwlock_wrlock(&lock_); }
 452
 453   void unlock() { pthread_rwlock_unlock(&lock_); }
 454
 455   void lock_shared() { pthread_rwlock_rdlock(&lock_); }
 456
 457   void unlock_shared() { pthread_rwlock_unlock(&lock_); }
 458 };
 459
 460 struct PosixMutex {
 461   pthread_mutex_t lock_;
 462
 463   PosixMutex() { pthread_mutex_init(&lock_, nullptr); }
 464
 465   ~PosixMutex() { pthread_mutex_destroy(&lock_); }
 466
 467   void lock() { pthread_mutex_lock(&lock_); }
 468
 469   void unlock() { pthread_mutex_unlock(&lock_); }
 470
 471   void lock_shared() { pthread_mutex_lock(&lock_); }
 472
 473   void unlock_shared() { pthread_mutex_unlock(&lock_); }
 474 };
 475
 476 template <template <typename> class Atom, typename Lock, typename Locker>
 477 static void runContendedReaders(size_t numOps,
 478                                 size_t numThreads,
 479                                 bool useSeparateLocks) {
 480   char padding1[64];
 481   (void)padding1;
 482   Lock globalLock;
 483   int valueProtectedByLock = 10;
 484   char padding2[64];
 485   (void)padding2;
 486   Atom<bool> go(false);
 487   Atom<bool>* goPtr = &go; // workaround for clang bug
 488   vector<thread> threads(numThreads);
 489
 490   BENCHMARK_SUSPEND {
 491     for (size_t t = 0; t < numThreads; ++t) {
 492       threads[t] = DSched::thread([&, t, numThreads] {
 493         Lock privateLock;
 494         Lock* lock = useSeparateLocks ? &privateLock : &globalLock;
 495         Locker locker;
 496         while (!goPtr->load()) {
 497           this_thread::yield();
 498         }
 499         for (size_t op = t; op < numOps; op += numThreads) {
 500           locker.lock_shared(lock);
 501           // note: folly::doNotOptimizeAway reads and writes to its arg,
 502           // so the following two lines are very different than a call
 503           // to folly::doNotOptimizeAway(valueProtectedByLock);
 504           auto copy = valueProtectedByLock;
 505           folly::doNotOptimizeAway(copy);
 506           locker.unlock_shared(lock);
 507         }
 508       });
 509     }
 510   }
 511
 512   go.store(true);
 513   for (auto& thr : threads) {
 514     DSched::join(thr);
 515   }
 516 }
 517
 518 static void folly_rwspin_reads(uint numOps,
 519                                size_t numThreads,
 520                                bool useSeparateLocks) {
 521   runContendedReaders<atomic, RWSpinLock, Locker>(
 522       numOps, numThreads, useSeparateLocks);
 523 }
 524
 525 static void shmtx_wr_pri_reads(uint numOps,
 526                                size_t numThreads,
 527                                bool useSeparateLocks) {
 528   runContendedReaders<atomic, SharedMutexWritePriority, TokenLocker>(
 529       numOps, numThreads, useSeparateLocks);
 530 }
 531
 532 static void shmtx_w_bare_reads(uint numOps,
 533                                size_t numThreads,
 534                                bool useSeparateLocks) {
 535   runContendedReaders<atomic, SharedMutexWritePriority, Locker>(
 536       numOps, numThreads, useSeparateLocks);
 537 }
 538
 539 static void shmtx_rd_pri_reads(uint numOps,
 540                                size_t numThreads,
 541                                bool useSeparateLocks) {
 542   runContendedReaders<atomic, SharedMutexReadPriority, TokenLocker>(
 543       numOps, numThreads, useSeparateLocks);
 544 }
 545
 546 static void shmtx_r_bare_reads(uint numOps,
 547                                size_t numThreads,
 548                                bool useSeparateLocks) {
 549   runContendedReaders<atomic, SharedMutexReadPriority, Locker>(
 550       numOps, numThreads, useSeparateLocks);
 551 }
 552
 553 static void folly_ticket_reads(uint numOps,
 554                                size_t numThreads,
 555                                bool useSeparateLocks) {
 556   runContendedReaders<atomic, RWTicketSpinLock64, Locker>(
 557       numOps, numThreads, useSeparateLocks);
 558 }
 559
 560 static void boost_shared_reads(uint numOps,
 561                                size_t numThreads,
 562                                bool useSeparateLocks) {
 563   runContendedReaders<atomic, boost::shared_mutex, Locker>(
 564       numOps, numThreads, useSeparateLocks);
 565 }
 566
 567 static void pthrd_rwlock_reads(uint numOps,
 568                                size_t numThreads,
 569                                bool useSeparateLocks) {
 570   runContendedReaders<atomic, PosixRWLock, Locker>(
 571       numOps, numThreads, useSeparateLocks);
 572 }
 573
 574 template <template <typename> class Atom, typename Lock, typename Locker>
 575 static void runMixed(size_t numOps,
 576                      size_t numThreads,
 577                      double writeFraction,
 578                      bool useSeparateLocks) {
 579   char padding1[64];
 580   (void)padding1;
 581   Lock globalLock;
 582   int valueProtectedByLock = 0;
 583   char padding2[64];
 584   (void)padding2;
 585   Atom<bool> go(false);
 586   Atom<bool>* goPtr = &go; // workaround for clang bug
 587   vector<thread> threads(numThreads);
 588
 589   BENCHMARK_SUSPEND {
 590     for (size_t t = 0; t < numThreads; ++t) {
 591       threads[t] = DSched::thread([&, t, numThreads] {
 592         struct drand48_data buffer;
 593         srand48_r(t, &buffer);
 594         long writeThreshold = writeFraction * 0x7fffffff;
 595         Lock privateLock;
 596         Lock* lock = useSeparateLocks ? &privateLock : &globalLock;
 597         Locker locker;
 598         while (!goPtr->load()) {
 599           this_thread::yield();
 600         }
 601         for (size_t op = t; op < numOps; op += numThreads) {
 602           long randVal;
 603           lrand48_r(&buffer, &randVal);
 604           bool writeOp = randVal < writeThreshold;
 605           if (writeOp) {
 606             locker.lock(lock);
 607             if (!useSeparateLocks) {
 608               ++valueProtectedByLock;
 609             }
 610             locker.unlock(lock);
 611           } else {
 612             locker.lock_shared(lock);
 613             auto v = valueProtectedByLock;
 614             folly::doNotOptimizeAway(v);
 615             locker.unlock_shared(lock);
 616           }
 617         }
 618       });
 619     }
 620   }
 621
 622   go.store(true);
 623   for (auto& thr : threads) {
 624     DSched::join(thr);
 625   }
 626 }
 627
 628 static void folly_rwspin(size_t numOps,
 629                          size_t numThreads,
 630                          double writeFraction,
 631                          bool useSeparateLocks) {
 632   runMixed<atomic, RWSpinLock, Locker>(
 633       numOps, numThreads, writeFraction, useSeparateLocks);
 634 }
 635
 636 static void shmtx_wr_pri(uint numOps,
 637                          size_t numThreads,
 638                          double writeFraction,
 639                          bool useSeparateLocks) {
 640   runMixed<atomic, SharedMutexWritePriority, TokenLocker>(
 641       numOps, numThreads, writeFraction, useSeparateLocks);
 642 }
 643
 644 static void shmtx_w_bare(uint numOps,
 645                          size_t numThreads,
 646                          double writeFraction,
 647                          bool useSeparateLocks) {
 648   runMixed<atomic, SharedMutexWritePriority, Locker>(
 649       numOps, numThreads, writeFraction, useSeparateLocks);
 650 }
 651
 652 static void shmtx_rd_pri(uint numOps,
 653                          size_t numThreads,
 654                          double writeFraction,
 655                          bool useSeparateLocks) {
 656   runMixed<atomic, SharedMutexReadPriority, TokenLocker>(
 657       numOps, numThreads, writeFraction, useSeparateLocks);
 658 }
 659
 660 static void shmtx_r_bare(uint numOps,
 661                          size_t numThreads,
 662                          double writeFraction,
 663                          bool useSeparateLocks) {
 664   runMixed<atomic, SharedMutexReadPriority, Locker>(
 665       numOps, numThreads, writeFraction, useSeparateLocks);
 666 }
 667
 668 static void folly_ticket(size_t numOps,
 669                          size_t numThreads,
 670                          double writeFraction,
 671                          bool useSeparateLocks) {
 672   runMixed<atomic, RWTicketSpinLock64, Locker>(
 673       numOps, numThreads, writeFraction, useSeparateLocks);
 674 }
 675
 676 static void boost_shared(size_t numOps,
 677                          size_t numThreads,
 678                          double writeFraction,
 679                          bool useSeparateLocks) {
 680   runMixed<atomic, boost::shared_mutex, Locker>(
 681       numOps, numThreads, writeFraction, useSeparateLocks);
 682 }
 683
 684 static void pthrd_rwlock(size_t numOps,
 685                          size_t numThreads,
 686                          double writeFraction,
 687                          bool useSeparateLocks) {
 688   runMixed<atomic, PosixRWLock, Locker>(
 689       numOps, numThreads, writeFraction, useSeparateLocks);
 690 }
 691
 692 static void pthrd_mutex_(size_t numOps,
 693                          size_t numThreads,
 694                          double writeFraction,
 695                          bool useSeparateLocks) {
 696   runMixed<atomic, PosixMutex, Locker>(
 697       numOps, numThreads, writeFraction, useSeparateLocks);
 698 }
 699
 700 template <typename Lock, template <typename> class Atom>
 701 static void runAllAndValidate(size_t numOps, size_t numThreads) {
 702   Lock globalLock;
 703   Atom<int> globalExclusiveCount(0);
 704   Atom<int> globalUpgradeCount(0);
 705   Atom<int> globalSharedCount(0);
 706
 707   Atom<bool> go(false);
 708
 709   // clang crashes on access to Atom<> captured by ref in closure
 710   Atom<int>* globalExclusiveCountPtr = &globalExclusiveCount;
 711   Atom<int>* globalUpgradeCountPtr = &globalUpgradeCount;
 712   Atom<int>* globalSharedCountPtr = &globalSharedCount;
 713   Atom<bool>* goPtr = &go;
 714
 715   vector<thread> threads(numThreads);
 716
 717   BENCHMARK_SUSPEND {
 718     for (size_t t = 0; t < numThreads; ++t) {
 719       threads[t] = DSched::thread([&, t, numThreads] {
 720         struct drand48_data buffer;
 721         srand48_r(t, &buffer);
 722
 723         bool exclusive = false;
 724         bool upgrade = false;
 725         bool shared = false;
 726         bool ourGlobalTokenUsed = false;
 727         SharedMutexToken ourGlobalToken;
 728
 729         Lock privateLock;
 730         vector<SharedMutexToken> privateTokens;
 731
 732         while (!goPtr->load()) {
 733           this_thread::yield();
 734         }
 735         for (size_t op = t; op < numOps; op += numThreads) {
 736           // randVal in [0,1000)
 737           long randVal;
 738           lrand48_r(&buffer, &randVal);
 739           randVal = (long)((randVal * (uint64_t)1000) / 0x7fffffff);
 740
 741           // make as many assertions as possible about the global state
 742           if (exclusive) {
 743             EXPECT_EQ(1, globalExclusiveCountPtr->load(memory_order_acquire));
 744             EXPECT_EQ(0, globalUpgradeCountPtr->load(memory_order_acquire));
 745             EXPECT_EQ(0, globalSharedCountPtr->load(memory_order_acquire));
 746           }
 747           if (upgrade) {
 748             EXPECT_EQ(0, globalExclusiveCountPtr->load(memory_order_acquire));
 749             EXPECT_EQ(1, globalUpgradeCountPtr->load(memory_order_acquire));
 750           }
 751           if (shared) {
 752             EXPECT_EQ(0, globalExclusiveCountPtr->load(memory_order_acquire));
 753             EXPECT_TRUE(globalSharedCountPtr->load(memory_order_acquire) > 0);
 754           } else {
 755             EXPECT_FALSE(ourGlobalTokenUsed);
 756           }
 757
 758           // independent 20% chance we do something to the private lock
 759           if (randVal < 200) {
 760             // it's okay to take multiple private shared locks because
 761             // we never take an exclusive lock, so reader versus writer
 762             // priority doesn't cause deadlocks
 763             if (randVal < 100 && privateTokens.size() > 0) {
 764               auto i = randVal % privateTokens.size();
 765               privateLock.unlock_shared(privateTokens[i]);
 766               privateTokens.erase(privateTokens.begin() + i);
 767             } else {
 768               SharedMutexToken token;
 769               privateLock.lock_shared(token);
 770               privateTokens.push_back(token);
 771             }
 772             continue;
 773           }
 774
 775           // if we've got a lock, the only thing we can do is release it
 776           // or transform it into a different kind of lock
 777           if (exclusive) {
 778             exclusive = false;
 779             --*globalExclusiveCountPtr;
 780             if (randVal < 500) {
 781               globalLock.unlock();
 782             } else if (randVal < 700) {
 783               globalLock.unlock_and_lock_shared();
 784               ++*globalSharedCountPtr;
 785               shared = true;
 786             } else if (randVal < 900) {
 787               globalLock.unlock_and_lock_shared(ourGlobalToken);
 788               ++*globalSharedCountPtr;
 789               shared = true;
 790               ourGlobalTokenUsed = true;
 791             } else {
 792               globalLock.unlock_and_lock_upgrade();
 793               ++*globalUpgradeCountPtr;
 794               upgrade = true;
 795             }
 796           } else if (upgrade) {
 797             upgrade = false;
 798             --*globalUpgradeCountPtr;
 799             if (randVal < 500) {
 800               globalLock.unlock_upgrade();
 801             } else if (randVal < 700) {
 802               globalLock.unlock_upgrade_and_lock_shared();
 803               ++*globalSharedCountPtr;
 804               shared = true;
 805             } else if (randVal < 900) {
 806               globalLock.unlock_upgrade_and_lock_shared(ourGlobalToken);
 807               ++*globalSharedCountPtr;
 808               shared = true;
 809               ourGlobalTokenUsed = true;
 810             } else {
 811               globalLock.unlock_upgrade_and_lock();
 812               ++*globalExclusiveCountPtr;
 813               exclusive = true;
 814             }
 815           } else if (shared) {
 816             shared = false;
 817             --*globalSharedCountPtr;
 818             if (ourGlobalTokenUsed) {
 819               globalLock.unlock_shared(ourGlobalToken);
 820               ourGlobalTokenUsed = false;
 821             } else {
 822               globalLock.unlock_shared();
 823             }
 824           } else if (randVal < 400) {
 825             // 40% chance of shared lock with token, 5 ways to get it
 826
 827             // delta t goes from -1 millis to 7 millis
 828             auto dt = microseconds(10 * (randVal - 100));
 829
 830             if (randVal < 400) {
 831               globalLock.lock_shared(ourGlobalToken);
 832               shared = true;
 833             } else if (randVal < 500) {
 834               shared = globalLock.try_lock_shared(ourGlobalToken);
 835             } else if (randVal < 600) {
 836               shared = globalLock.try_lock_shared_for(dt, ourGlobalToken);
 837             } else if (randVal < 800) {
 838               shared = globalLock.try_lock_shared_until(
 839                   system_clock::now() + dt, ourGlobalToken);
 840             }
 841             if (shared) {
 842               ourGlobalTokenUsed = true;
 843               ++*globalSharedCountPtr;
 844             }
 845           } else if (randVal < 800) {
 846             // 40% chance of shared lock without token
 847             auto dt = microseconds(10 * (randVal - 100));
 848             if (randVal < 400) {
 849               globalLock.lock_shared();
 850               shared = true;
 851             } else if (randVal < 500) {
 852               shared = globalLock.try_lock_shared();
 853             } else if (randVal < 600) {
 854               shared = globalLock.try_lock_shared_for(dt);
 855             } else if (randVal < 800) {
 856               shared = globalLock.try_lock_shared_until(
 857                   system_clock::now() + dt);
 858             }
 859             if (shared) {
 860               ++*globalSharedCountPtr;
 861             }
 862           } else if (randVal < 900) {
 863             // 10% change of upgrade lock
 864             globalLock.lock_upgrade();
 865             upgrade = true;
 866             ++*globalUpgradeCountPtr;
 867           } else {
 868             // 10% chance of exclusive lock, 5 ways to get it
 869
 870             // delta t goes from -1 millis to 9 millis
 871             auto dt = microseconds(100 * (randVal - 910));
 872
 873             if (randVal < 400) {
 874               globalLock.lock();
 875               exclusive = true;
 876             } else if (randVal < 500) {
 877               exclusive = globalLock.try_lock();
 878             } else if (randVal < 600) {
 879               exclusive = globalLock.try_lock_for(dt);
 880             } else if (randVal < 700) {
 881               exclusive = globalLock.try_lock_until(steady_clock::now() + dt);
 882             } else {
 883               exclusive = globalLock.try_lock_until(system_clock::now() + dt);
 884             }
 885             if (exclusive) {
 886               ++*globalExclusiveCountPtr;
 887             }
 888           }
 889         }
 890
 891         if (exclusive) {
 892           --*globalExclusiveCountPtr;
 893           globalLock.unlock();
 894         }
 895         if (upgrade) {
 896           --*globalUpgradeCountPtr;
 897           globalLock.unlock_upgrade();
 898         }
 899         if (shared) {
 900           --*globalSharedCountPtr;
 901           if (ourGlobalTokenUsed) {
 902             globalLock.unlock_shared(ourGlobalToken);
 903             ourGlobalTokenUsed = false;
 904           } else {
 905             globalLock.unlock_shared();
 906           }
 907         }
 908         for (auto& token : privateTokens) {
 909           privateLock.unlock_shared(token);
 910         }
 911       });
 912     }
 913   }
 914
 915   go.store(true);
 916   for (auto& thr : threads) {
 917     DSched::join(thr);
 918   }
 919 }
 920
 921 TEST(SharedMutex, deterministic_concurrent_readers_of_one_lock_read_prio) {
 922   for (int pass = 0; pass < 3; ++pass) {
 923     DSched sched(DSched::uniform(pass));
 924     runContendedReaders<DeterministicAtomic,
 925                         DSharedMutexReadPriority,
 926                         Locker>(1000, 3, false);
 927   }
 928 }
 929
 930 TEST(SharedMutex, deterministic_concurrent_readers_of_one_lock_write_prio) {
 931   for (int pass = 0; pass < 3; ++pass) {
 932     DSched sched(DSched::uniform(pass));
 933     runContendedReaders<DeterministicAtomic,
 934                         DSharedMutexWritePriority,
 935                         Locker>(1000, 3, false);
 936   }
 937 }
 938
 939 TEST(SharedMutex, concurrent_readers_of_one_lock_read_prio) {
 940   for (int pass = 0; pass < 10; ++pass) {
 941     runContendedReaders<atomic, SharedMutexReadPriority, Locker>(
 942         100000, 32, false);
 943   }
 944 }
 945
 946 TEST(SharedMutex, concurrent_readers_of_one_lock_write_prio) {
 947   for (int pass = 0; pass < 10; ++pass) {
 948     runContendedReaders<atomic, SharedMutexWritePriority, Locker>(
 949         100000, 32, false);
 950   }
 951 }
 952
 953 TEST(SharedMutex, deterministic_readers_of_concurrent_locks_read_prio) {
 954   for (int pass = 0; pass < 3; ++pass) {
 955     DSched sched(DSched::uniform(pass));
 956     runContendedReaders<DeterministicAtomic,
 957                         DSharedMutexReadPriority,
 958                         Locker>(1000, 3, true);
 959   }
 960 }
 961
 962 TEST(SharedMutex, deterministic_readers_of_concurrent_locks_write_prio) {
 963   for (int pass = 0; pass < 3; ++pass) {
 964     DSched sched(DSched::uniform(pass));
 965     runContendedReaders<DeterministicAtomic,
 966                         DSharedMutexWritePriority,
 967                         Locker>(1000, 3, true);
 968   }
 969 }
 970
 971 TEST(SharedMutex, readers_of_concurrent_locks_read_prio) {
 972   for (int pass = 0; pass < 10; ++pass) {
 973     runContendedReaders<atomic, SharedMutexReadPriority, TokenLocker>(
 974         100000, 32, true);
 975   }
 976 }
 977
 978 TEST(SharedMutex, readers_of_concurrent_locks_write_prio) {
 979   for (int pass = 0; pass < 10; ++pass) {
 980     runContendedReaders<atomic, SharedMutexWritePriority, TokenLocker>(
 981         100000, 32, true);
 982   }
 983 }
 984
 985 TEST(SharedMutex, deterministic_mixed_mostly_read_read_prio) {
 986   for (int pass = 0; pass < 3; ++pass) {
 987     DSched sched(DSched::uniform(pass));
 988     runMixed<DeterministicAtomic, DSharedMutexReadPriority, Locker>(
 989         1000, 3, 0.1, false);
 990   }
 991 }
 992
 993 TEST(SharedMutex, deterministic_mixed_mostly_read_write_prio) {
 994   for (int pass = 0; pass < 3; ++pass) {
 995     DSched sched(DSched::uniform(pass));
 996     runMixed<DeterministicAtomic, DSharedMutexWritePriority, Locker>(
 997         1000, 3, 0.1, false);
 998   }
 999 }
1000
1001 TEST(SharedMutex, mixed_mostly_read_read_prio) {
1002   for (int pass = 0; pass < 5; ++pass) {
1003     runMixed<atomic, SharedMutexReadPriority, TokenLocker>(
1004         50000, 32, 0.1, false);
1005   }
1006 }
1007
1008 TEST(SharedMutex, mixed_mostly_read_write_prio) {
1009   for (int pass = 0; pass < 5; ++pass) {
1010     runMixed<atomic, SharedMutexWritePriority, TokenLocker>(
1011         50000, 32, 0.1, false);
1012   }
1013 }
1014
1015 TEST(SharedMutex, deterministic_mixed_mostly_write_read_prio) {
1016   for (int pass = 0; pass < 1; ++pass) {
1017     DSched sched(DSched::uniform(pass));
1018     runMixed<DeterministicAtomic, DSharedMutexReadPriority, TokenLocker>(
1019         1000, 10, 0.9, false);
1020   }
1021 }
1022
1023 TEST(SharedMutex, deterministic_mixed_mostly_write_write_prio) {
1024   for (int pass = 0; pass < 1; ++pass) {
1025     DSched sched(DSched::uniform(pass));
1026     runMixed<DeterministicAtomic, DSharedMutexWritePriority, TokenLocker>(
1027         1000, 10, 0.9, false);
1028   }
1029 }
1030
1031 TEST(SharedMutex, deterministic_lost_wakeup_write_prio) {
1032   for (int pass = 0; pass < 10; ++pass) {
1033     DSched sched(DSched::uniformSubset(pass, 2, 200));
1034     runMixed<DeterministicAtomic, DSharedMutexWritePriority, TokenLocker>(
1035         1000, 3, 1.0, false);
1036   }
1037 }
1038
1039 TEST(SharedMutex, mixed_mostly_write_read_prio) {
1040   for (int pass = 0; pass < 5; ++pass) {
1041     runMixed<atomic, SharedMutexReadPriority, TokenLocker>(
1042         50000, 300, 0.9, false);
1043   }
1044 }
1045
1046 TEST(SharedMutex, mixed_mostly_write_write_prio) {
1047   for (int pass = 0; pass < 5; ++pass) {
1048     runMixed<atomic, SharedMutexWritePriority, TokenLocker>(
1049         50000, 300, 0.9, false);
1050   }
1051 }
1052
1053 TEST(SharedMutex, deterministic_all_ops_read_prio) {
1054   for (int pass = 0; pass < 5; ++pass) {
1055     DSched sched(DSched::uniform(pass));
1056     runAllAndValidate<DSharedMutexReadPriority, DeterministicAtomic>(1000, 8);
1057   }
1058 }
1059
1060 TEST(SharedMutex, deterministic_all_ops_write_prio) {
1061   for (int pass = 0; pass < 5; ++pass) {
1062     DSched sched(DSched::uniform(pass));
1063     runAllAndValidate<DSharedMutexWritePriority, DeterministicAtomic>(1000, 8);
1064   }
1065 }
1066
1067 TEST(SharedMutex, all_ops_read_prio) {
1068   for (int pass = 0; pass < 5; ++pass) {
1069     runAllAndValidate<SharedMutexReadPriority, atomic>(100000, 32);
1070   }
1071 }
1072
1073 TEST(SharedMutex, all_ops_write_prio) {
1074   for (int pass = 0; pass < 5; ++pass) {
1075     runAllAndValidate<SharedMutexWritePriority, atomic>(100000, 32);
1076   }
1077 }
1078
1079 FOLLY_ASSUME_FBVECTOR_COMPATIBLE(
1080     boost::optional<boost::optional<SharedMutexToken>>)
1081
1082 // Setup is a set of threads that either grab a shared lock, or exclusive
1083 // and then downgrade it, or upgrade then upgrade and downgrade, then
1084 // enqueue the shared lock to a second set of threads that just performs
1085 // unlocks.  Half of the shared locks use tokens, the others don't.
1086 template <typename Lock, template <typename> class Atom>
1087 static void runRemoteUnlock(size_t numOps,
1088                             double preWriteFraction,
1089                             double preUpgradeFraction,
1090                             size_t numSendingThreads,
1091                             size_t numReceivingThreads) {
1092   Lock globalLock;
1093   MPMCQueue<boost::optional<boost::optional<SharedMutexToken>>, Atom>
1094     queue(10);
1095   auto queuePtr = &queue; // workaround for clang crash
1096
1097   Atom<bool> go(false);
1098   auto goPtr = &go; // workaround for clang crash
1099   Atom<int> pendingSenders(numSendingThreads);
1100   auto pendingSendersPtr = &pendingSenders; // workaround for clang crash
1101   vector<thread> threads(numSendingThreads + numReceivingThreads);
1102
1103   BENCHMARK_SUSPEND {
1104     for (size_t t = 0; t < threads.size(); ++t) {
1105       threads[t] = DSched::thread([&, t, numSendingThreads] {
1106         if (t >= numSendingThreads) {
1107           // we're a receiver
1108           typename decltype(queue)::value_type elem;
1109           while (true) {
1110             queuePtr->blockingRead(elem);
1111             if (!elem) {
1112               // EOF, pass the EOF token
1113               queuePtr->blockingWrite(std::move(elem));
1114               break;
1115             }
1116             if (*elem) {
1117               globalLock.unlock_shared(**elem);
1118             } else {
1119               globalLock.unlock_shared();
1120             }
1121           }
1122           return;
1123         }
1124         // else we're a sender
1125
1126         struct drand48_data buffer;
1127         srand48_r(t, &buffer);
1128
1129         while (!goPtr->load()) {
1130           this_thread::yield();
1131         }
1132         for (size_t op = t; op < numOps; op += numSendingThreads) {
1133           long unscaledRandVal;
1134           lrand48_r(&buffer, &unscaledRandVal);
1135
1136           // randVal in [0,1]
1137           double randVal = ((double)unscaledRandVal) / 0x7fffffff;
1138
1139           // extract a bit and rescale
1140           bool useToken = randVal >= 0.5;
1141           randVal = (randVal - (useToken ? 0.5 : 0.0)) * 2;
1142
1143           boost::optional<SharedMutexToken> maybeToken;
1144
1145           if (useToken) {
1146             SharedMutexToken token;
1147             if (randVal < preWriteFraction) {
1148               globalLock.lock();
1149               globalLock.unlock_and_lock_shared(token);
1150             } else if (randVal < preWriteFraction + preUpgradeFraction / 2) {
1151               globalLock.lock_upgrade();
1152               globalLock.unlock_upgrade_and_lock_shared(token);
1153             } else if (randVal < preWriteFraction + preUpgradeFraction) {
1154               globalLock.lock_upgrade();
1155               globalLock.unlock_upgrade_and_lock();
1156               globalLock.unlock_and_lock_shared(token);
1157             } else {
1158               globalLock.lock_shared(token);
1159             }
1160             maybeToken = token;
1161           } else {
1162             if (randVal < preWriteFraction) {
1163               globalLock.lock();
1164               globalLock.unlock_and_lock_shared();
1165             } else if (randVal < preWriteFraction + preUpgradeFraction / 2) {
1166               globalLock.lock_upgrade();
1167               globalLock.unlock_upgrade_and_lock_shared();
1168             } else if (randVal < preWriteFraction + preUpgradeFraction) {
1169               globalLock.lock_upgrade();
1170               globalLock.unlock_upgrade_and_lock();
1171               globalLock.unlock_and_lock_shared();
1172             } else {
1173               globalLock.lock_shared();
1174             }
1175           }
1176
1177           // blockingWrite is emplace-like, so this automatically adds
1178           // another level of wrapping
1179           queuePtr->blockingWrite(maybeToken);
1180         }
1181         if (--*pendingSendersPtr == 0) {
1182           queuePtr->blockingWrite(boost::none);
1183         }
1184       });
1185     }
1186   }
1187
1188   go.store(true);
1189   for (auto& thr : threads) {
1190     DSched::join(thr);
1191   }
1192 }
1193
1194 TEST(SharedMutex, deterministic_remote_write_prio) {
1195   for (int pass = 0; pass < 1; ++pass) {
1196     DSched sched(DSched::uniform(pass));
1197     runRemoteUnlock<DSharedMutexWritePriority, DeterministicAtomic>(
1198         500, 0.1, 0.1, 5, 5);
1199   }
1200 }
1201
1202 TEST(SharedMutex, deterministic_remote_read_prio) {
1203   for (int pass = 0; pass < 1; ++pass) {
1204     DSched sched(DSched::uniform(pass));
1205     runRemoteUnlock<DSharedMutexReadPriority, DeterministicAtomic>(
1206         500, 0.1, 0.1, 5, 5);
1207   }
1208 }
1209
1210 TEST(SharedMutex, remote_write_prio) {
1211   for (int pass = 0; pass < 10; ++pass) {
1212     runRemoteUnlock<SharedMutexWritePriority, atomic>(100000, 0.1, 0.1, 5, 5);
1213   }
1214 }
1215
1216 TEST(SharedMutex, remote_read_prio) {
1217   for (int pass = 0; pass < 100; ++pass) {
1218     runRemoteUnlock<SharedMutexReadPriority, atomic>(100000, 0.1, 0.1, 5, 5);
1219   }
1220 }
1221
1222 static void burn(size_t n) {
1223   for (size_t i = 0; i < n; ++i) {
1224     folly::doNotOptimizeAway(i);
1225   }
1226 }
1227
1228 // Two threads and three locks, arranged so that they have to proceed
1229 // in turn with reader/writer conflict
1230 template <typename Lock, template <typename> class Atom = atomic>
1231 static void runPingPong(size_t numRounds, size_t burnCount) {
1232   char padding1[56];
1233   (void)padding1;
1234   pair<Lock, char[56]> locks[3];
1235   char padding2[56];
1236   (void)padding2;
1237
1238   Atom<int> avail(0);
1239   auto availPtr = &avail; // workaround for clang crash
1240   Atom<bool> go(false);
1241   auto goPtr = &go; // workaround for clang crash
1242   vector<thread> threads(2);
1243
1244   locks[0].first.lock();
1245   locks[1].first.lock();
1246   locks[2].first.lock_shared();
1247
1248   BENCHMARK_SUSPEND {
1249     threads[0] = DSched::thread([&] {
1250       ++*availPtr;
1251       while (!goPtr->load()) {
1252         this_thread::yield();
1253       }
1254       for (size_t i = 0; i < numRounds; ++i) {
1255         locks[i % 3].first.unlock();
1256         locks[(i + 2) % 3].first.lock();
1257         burn(burnCount);
1258       }
1259     });
1260     threads[1] = DSched::thread([&] {
1261       ++*availPtr;
1262       while (!goPtr->load()) {
1263         this_thread::yield();
1264       }
1265       for (size_t i = 0; i < numRounds; ++i) {
1266         locks[i % 3].first.lock_shared();
1267         burn(burnCount);
1268         locks[(i + 2) % 3].first.unlock_shared();
1269       }
1270     });
1271
1272     while (avail.load() < 2) {
1273       this_thread::yield();
1274     }
1275   }
1276
1277   go.store(true);
1278   for (auto& thr : threads) {
1279     DSched::join(thr);
1280   }
1281   locks[numRounds % 3].first.unlock();
1282   locks[(numRounds + 1) % 3].first.unlock();
1283   locks[(numRounds + 2) % 3].first.unlock_shared();
1284 }
1285
1286 static void folly_rwspin_ping_pong(size_t n, size_t scale, size_t burnCount) {
1287   runPingPong<RWSpinLock>(n / scale, burnCount);
1288 }
1289
1290 static void shmtx_w_bare_ping_pong(size_t n, size_t scale, size_t burnCount) {
1291   runPingPong<SharedMutexWritePriority>(n / scale, burnCount);
1292 }
1293
1294 static void shmtx_r_bare_ping_pong(size_t n, size_t scale, size_t burnCount) {
1295   runPingPong<SharedMutexReadPriority>(n / scale, burnCount);
1296 }
1297
1298 static void folly_ticket_ping_pong(size_t n, size_t scale, size_t burnCount) {
1299   runPingPong<RWTicketSpinLock64>(n / scale, burnCount);
1300 }
1301
1302 static void boost_shared_ping_pong(size_t n, size_t scale, size_t burnCount) {
1303   runPingPong<boost::shared_mutex>(n / scale, burnCount);
1304 }
1305
1306 static void pthrd_rwlock_ping_pong(size_t n, size_t scale, size_t burnCount) {
1307   runPingPong<PosixRWLock>(n / scale, burnCount);
1308 }
1309
1310 TEST(SharedMutex, deterministic_ping_pong_write_prio) {
1311   for (int pass = 0; pass < 1; ++pass) {
1312     DSched sched(DSched::uniform(pass));
1313     runPingPong<DSharedMutexWritePriority, DeterministicAtomic>(500, 0);
1314   }
1315 }
1316
1317 TEST(SharedMutex, deterministic_ping_pong_read_prio) {
1318   for (int pass = 0; pass < 1; ++pass) {
1319     DSched sched(DSched::uniform(pass));
1320     runPingPong<DSharedMutexReadPriority, DeterministicAtomic>(500, 0);
1321   }
1322 }
1323
1324 TEST(SharedMutex, ping_pong_write_prio) {
1325   for (int pass = 0; pass < 1; ++pass) {
1326     runPingPong<SharedMutexWritePriority, atomic>(50000, 0);
1327   }
1328 }
1329
1330 TEST(SharedMutex, ping_pong_read_prio) {
1331   for (int pass = 0; pass < 1; ++pass) {
1332     runPingPong<SharedMutexReadPriority, atomic>(50000, 0);
1333   }
1334 }
1335
1336 // This is here so you can tell how much of the runtime reported by the
1337 // more complex harnesses is due to the harness, although due to the
1338 // magic of compiler optimization it may also be slower
1339 BENCHMARK(single_thread_lock_shared_unlock_shared, iters) {
1340   SharedMutex lock;
1341   for (size_t n = 0; n < iters; ++n) {
1342     SharedMutex::Token token;
1343     lock.lock_shared(token);
1344     folly::doNotOptimizeAway(0);
1345     lock.unlock_shared(token);
1346   }
1347 }
1348
1349 BENCHMARK(single_thread_lock_unlock, iters) {
1350   SharedMutex lock;
1351   for (size_t n = 0; n < iters; ++n) {
1352     lock.lock();
1353     folly::doNotOptimizeAway(0);
1354     lock.unlock();
1355   }
1356 }
1357
1358 #define BENCH_BASE(args...) BENCHMARK_NAMED_PARAM(args)
1359 #define BENCH_REL(args...) BENCHMARK_RELATIVE_NAMED_PARAM(args)
1360
1361 // 100% reads.  Best-case scenario for deferred locks.  Lock is colocated
1362 // with read data, so inline lock takes cache miss every time but deferred
1363 // lock has only cache hits and local access.
1364 BENCHMARK_DRAW_LINE()
1365 BENCHMARK_DRAW_LINE()
1366 BENCH_BASE(folly_rwspin_reads, 1thread, 1, false)
1367 BENCH_REL (shmtx_wr_pri_reads, 1thread, 1, false)
1368 BENCH_REL (shmtx_w_bare_reads, 1thread, 1, false)
1369 BENCH_REL (shmtx_rd_pri_reads, 1thread, 1, false)
1370 BENCH_REL (shmtx_r_bare_reads, 1thread, 1, false)
1371 BENCH_REL (folly_ticket_reads, 1thread, 1, false)
1372 BENCH_REL (boost_shared_reads, 1thread, 1, false)
1373 BENCH_REL (pthrd_rwlock_reads, 1thread, 1, false)
1374 BENCHMARK_DRAW_LINE()
1375 BENCH_BASE(folly_rwspin_reads, 2thread, 2, false)
1376 BENCH_REL (shmtx_wr_pri_reads, 2thread, 2, false)
1377 BENCH_REL (shmtx_w_bare_reads, 2thread, 2, false)
1378 BENCH_REL (shmtx_rd_pri_reads, 2thread, 2, false)
1379 BENCH_REL (shmtx_r_bare_reads, 2thread, 2, false)
1380 BENCH_REL (folly_ticket_reads, 2thread, 2, false)
1381 BENCH_REL (boost_shared_reads, 2thread, 2, false)
1382 BENCH_REL (pthrd_rwlock_reads, 2thread, 2, false)
1383 BENCHMARK_DRAW_LINE()
1384 BENCH_BASE(folly_rwspin_reads, 4thread, 4, false)
1385 BENCH_REL (shmtx_wr_pri_reads, 4thread, 4, false)
1386 BENCH_REL (shmtx_w_bare_reads, 4thread, 4, false)
1387 BENCH_REL (shmtx_rd_pri_reads, 4thread, 4, false)
1388 BENCH_REL (shmtx_r_bare_reads, 4thread, 4, false)
1389 BENCH_REL (folly_ticket_reads, 4thread, 4, false)
1390 BENCH_REL (boost_shared_reads, 4thread, 4, false)
1391 BENCH_REL (pthrd_rwlock_reads, 4thread, 4, false)
1392 BENCHMARK_DRAW_LINE()
1393 BENCH_BASE(folly_rwspin_reads, 8thread, 8, false)
1394 BENCH_REL (shmtx_wr_pri_reads, 8thread, 8, false)
1395 BENCH_REL (shmtx_w_bare_reads, 8thread, 8, false)
1396 BENCH_REL (shmtx_rd_pri_reads, 8thread, 8, false)
1397 BENCH_REL (shmtx_r_bare_reads, 8thread, 8, false)
1398 BENCH_REL (folly_ticket_reads, 8thread, 8, false)
1399 BENCH_REL (boost_shared_reads, 8thread, 8, false)
1400 BENCH_REL (pthrd_rwlock_reads, 8thread, 8, false)
1401 BENCHMARK_DRAW_LINE()
1402 BENCH_BASE(folly_rwspin_reads, 16thread, 16, false)
1403 BENCH_REL (shmtx_wr_pri_reads, 16thread, 16, false)
1404 BENCH_REL (shmtx_w_bare_reads, 16thread, 16, false)
1405 BENCH_REL (shmtx_rd_pri_reads, 16thread, 16, false)
1406 BENCH_REL (shmtx_r_bare_reads, 16thread, 16, false)
1407 BENCH_REL (folly_ticket_reads, 16thread, 16, false)
1408 BENCH_REL (boost_shared_reads, 16thread, 16, false)
1409 BENCH_REL (pthrd_rwlock_reads, 16thread, 16, false)
1410 BENCHMARK_DRAW_LINE()
1411 BENCH_BASE(folly_rwspin_reads, 32thread, 32, false)
1412 BENCH_REL (shmtx_wr_pri_reads, 32thread, 32, false)
1413 BENCH_REL (shmtx_w_bare_reads, 32thread, 32, false)
1414 BENCH_REL (shmtx_rd_pri_reads, 32thread, 32, false)
1415 BENCH_REL (shmtx_r_bare_reads, 32thread, 32, false)
1416 BENCH_REL (folly_ticket_reads, 32thread, 32, false)
1417 BENCH_REL (boost_shared_reads, 32thread, 32, false)
1418 BENCH_REL (pthrd_rwlock_reads, 32thread, 32, false)
1419 BENCHMARK_DRAW_LINE()
1420 BENCH_BASE(folly_rwspin_reads, 64thread, 64, false)
1421 BENCH_REL (shmtx_wr_pri_reads, 64thread, 64, false)
1422 BENCH_REL (shmtx_w_bare_reads, 64thread, 64, false)
1423 BENCH_REL (shmtx_rd_pri_reads, 64thread, 64, false)
1424 BENCH_REL (shmtx_r_bare_reads, 64thread, 64, false)
1425 BENCH_REL (folly_ticket_reads, 64thread, 64, false)
1426 BENCH_REL (boost_shared_reads, 64thread, 64, false)
1427 BENCH_REL (pthrd_rwlock_reads, 64thread, 64, false)
1428
1429 // 1 lock used by everybody, 100% writes.  Threads only hurt, but it is
1430 // good to not fail catastrophically.  Compare to single_thread_lock_unlock
1431 // to see the overhead of the generic driver (and its pseudo-random number
1432 // generator).  pthrd_mutex_ is a pthread_mutex_t (default, not adaptive),
1433 // which is better than any of the reader-writer locks for this scenario.
1434 BENCHMARK_DRAW_LINE()
1435 BENCHMARK_DRAW_LINE()
1436 BENCH_BASE(folly_rwspin, 1thread_all_write, 1, 1.0, false)
1437 BENCH_REL (shmtx_wr_pri, 1thread_all_write, 1, 1.0, false)
1438 BENCH_REL (shmtx_rd_pri, 1thread_all_write, 1, 1.0, false)
1439 BENCH_REL (folly_ticket, 1thread_all_write, 1, 1.0, false)
1440 BENCH_REL (boost_shared, 1thread_all_write, 1, 1.0, false)
1441 BENCH_REL (pthrd_rwlock, 1thread_all_write, 1, 1.0, false)
1442 BENCH_REL (pthrd_mutex_, 1thread_all_write, 1, 1.0, false)
1443 BENCHMARK_DRAW_LINE()
1444 BENCH_BASE(folly_rwspin, 2thread_all_write, 2, 1.0, false)
1445 BENCH_REL (shmtx_wr_pri, 2thread_all_write, 2, 1.0, false)
1446 BENCH_REL (shmtx_rd_pri, 2thread_all_write, 2, 1.0, false)
1447 BENCH_REL (folly_ticket, 2thread_all_write, 2, 1.0, false)
1448 BENCH_REL (boost_shared, 2thread_all_write, 2, 1.0, false)
1449 BENCH_REL (pthrd_rwlock, 2thread_all_write, 2, 1.0, false)
1450 BENCH_REL (pthrd_mutex_, 2thread_all_write, 2, 1.0, false)
1451 BENCHMARK_DRAW_LINE()
1452 BENCH_BASE(folly_rwspin, 4thread_all_write, 4, 1.0, false)
1453 BENCH_REL (shmtx_wr_pri, 4thread_all_write, 4, 1.0, false)
1454 BENCH_REL (shmtx_rd_pri, 4thread_all_write, 4, 1.0, false)
1455 BENCH_REL (folly_ticket, 4thread_all_write, 4, 1.0, false)
1456 BENCH_REL (boost_shared, 4thread_all_write, 4, 1.0, false)
1457 BENCH_REL (pthrd_rwlock, 4thread_all_write, 4, 1.0, false)
1458 BENCH_REL (pthrd_mutex_, 4thread_all_write, 4, 1.0, false)
1459 BENCHMARK_DRAW_LINE()
1460 BENCH_BASE(folly_rwspin, 8thread_all_write, 8, 1.0, false)
1461 BENCH_REL (shmtx_wr_pri, 8thread_all_write, 8, 1.0, false)
1462 BENCH_REL (shmtx_rd_pri, 8thread_all_write, 8, 1.0, false)
1463 BENCH_REL (folly_ticket, 8thread_all_write, 8, 1.0, false)
1464 BENCH_REL (boost_shared, 8thread_all_write, 8, 1.0, false)
1465 BENCH_REL (pthrd_rwlock, 8thread_all_write, 8, 1.0, false)
1466 BENCH_REL (pthrd_mutex_, 8thread_all_write, 8, 1.0, false)
1467 BENCHMARK_DRAW_LINE()
1468 BENCH_BASE(folly_rwspin, 16thread_all_write, 16, 1.0, false)
1469 BENCH_REL (shmtx_wr_pri, 16thread_all_write, 16, 1.0, false)
1470 BENCH_REL (shmtx_rd_pri, 16thread_all_write, 16, 1.0, false)
1471 BENCH_REL (folly_ticket, 16thread_all_write, 16, 1.0, false)
1472 BENCH_REL (boost_shared, 16thread_all_write, 16, 1.0, false)
1473 BENCH_REL (pthrd_rwlock, 16thread_all_write, 16, 1.0, false)
1474 BENCH_REL (pthrd_mutex_, 16thread_all_write, 16, 1.0, false)
1475 BENCHMARK_DRAW_LINE()
1476 BENCH_BASE(folly_rwspin, 32thread_all_write, 32, 1.0, false)
1477 BENCH_REL (shmtx_wr_pri, 32thread_all_write, 32, 1.0, false)
1478 BENCH_REL (shmtx_rd_pri, 32thread_all_write, 32, 1.0, false)
1479 BENCH_REL (folly_ticket, 32thread_all_write, 32, 1.0, false)
1480 BENCH_REL (boost_shared, 32thread_all_write, 32, 1.0, false)
1481 BENCH_REL (pthrd_rwlock, 32thread_all_write, 32, 1.0, false)
1482 BENCH_REL (pthrd_mutex_, 32thread_all_write, 32, 1.0, false)
1483 BENCHMARK_DRAW_LINE()
1484 BENCH_BASE(folly_rwspin, 64thread_all_write, 64, 1.0, false)
1485 BENCH_REL (shmtx_wr_pri, 64thread_all_write, 64, 1.0, false)
1486 BENCH_REL (shmtx_rd_pri, 64thread_all_write, 64, 1.0, false)
1487 BENCH_REL (folly_ticket, 64thread_all_write, 64, 1.0, false)
1488 BENCH_REL (boost_shared, 64thread_all_write, 64, 1.0, false)
1489 BENCH_REL (pthrd_rwlock, 64thread_all_write, 64, 1.0, false)
1490 BENCH_REL (pthrd_mutex_, 64thread_all_write, 64, 1.0, false)
1491
1492 // 1 lock used by everybody, 10% writes.  Not much scaling to be had.  Perf
1493 // is best at 1 thread, once you've got multiple threads > 8 threads hurts.
1494 BENCHMARK_DRAW_LINE()
1495 BENCHMARK_DRAW_LINE()
1496 BENCH_BASE(folly_rwspin, 1thread_10pct_write, 1, 0.10, false)
1497 BENCH_REL (shmtx_wr_pri, 1thread_10pct_write, 1, 0.10, false)
1498 BENCH_REL (shmtx_rd_pri, 1thread_10pct_write, 1, 0.10, false)
1499 BENCH_REL (folly_ticket, 1thread_10pct_write, 1, 0.10, false)
1500 BENCH_REL (boost_shared, 1thread_10pct_write, 1, 0.10, false)
1501 BENCH_REL (pthrd_rwlock, 1thread_10pct_write, 1, 0.10, false)
1502 BENCHMARK_DRAW_LINE()
1503 BENCH_BASE(folly_rwspin, 2thread_10pct_write, 2, 0.10, false)
1504 BENCH_REL (shmtx_wr_pri, 2thread_10pct_write, 2, 0.10, false)
1505 BENCH_REL (shmtx_rd_pri, 2thread_10pct_write, 2, 0.10, false)
1506 BENCH_REL (folly_ticket, 2thread_10pct_write, 2, 0.10, false)
1507 BENCH_REL (boost_shared, 2thread_10pct_write, 2, 0.10, false)
1508 BENCH_REL (pthrd_rwlock, 2thread_10pct_write, 2, 0.10, false)
1509 BENCHMARK_DRAW_LINE()
1510 BENCH_BASE(folly_rwspin, 4thread_10pct_write, 4, 0.10, false)
1511 BENCH_REL (shmtx_wr_pri, 4thread_10pct_write, 4, 0.10, false)
1512 BENCH_REL (shmtx_rd_pri, 4thread_10pct_write, 4, 0.10, false)
1513 BENCH_REL (folly_ticket, 4thread_10pct_write, 4, 0.10, false)
1514 BENCH_REL (boost_shared, 4thread_10pct_write, 4, 0.10, false)
1515 BENCH_REL (pthrd_rwlock, 4thread_10pct_write, 4, 0.10, false)
1516 BENCHMARK_DRAW_LINE()
1517 BENCH_BASE(folly_rwspin, 8thread_10pct_write, 8, 0.10, false)
1518 BENCH_REL (shmtx_wr_pri, 8thread_10pct_write, 8, 0.10, false)
1519 BENCH_REL (shmtx_rd_pri, 8thread_10pct_write, 8, 0.10, false)
1520 BENCH_REL (folly_ticket, 8thread_10pct_write, 8, 0.10, false)
1521 BENCH_REL (boost_shared, 8thread_10pct_write, 8, 0.10, false)
1522 BENCH_REL (pthrd_rwlock, 8thread_10pct_write, 8, 0.10, false)
1523 BENCHMARK_DRAW_LINE()
1524 BENCH_BASE(folly_rwspin, 16thread_10pct_write, 16, 0.10, false)
1525 BENCH_REL (shmtx_wr_pri, 16thread_10pct_write, 16, 0.10, false)
1526 BENCH_REL (shmtx_rd_pri, 16thread_10pct_write, 16, 0.10, false)
1527 BENCH_REL (folly_ticket, 16thread_10pct_write, 16, 0.10, false)
1528 BENCH_REL (boost_shared, 16thread_10pct_write, 16, 0.10, false)
1529 BENCH_REL (pthrd_rwlock, 16thread_10pct_write, 16, 0.10, false)
1530 BENCHMARK_DRAW_LINE()
1531 BENCH_BASE(folly_rwspin, 32thread_10pct_write, 32, 0.10, false)
1532 BENCH_REL (shmtx_wr_pri, 32thread_10pct_write, 32, 0.10, false)
1533 BENCH_REL (shmtx_rd_pri, 32thread_10pct_write, 32, 0.10, false)
1534 BENCH_REL (folly_ticket, 32thread_10pct_write, 32, 0.10, false)
1535 BENCH_REL (boost_shared, 32thread_10pct_write, 32, 0.10, false)
1536 BENCH_REL (pthrd_rwlock, 32thread_10pct_write, 32, 0.10, false)
1537 BENCHMARK_DRAW_LINE()
1538 BENCH_BASE(folly_rwspin, 64thread_10pct_write, 64, 0.10, false)
1539 BENCH_REL (shmtx_wr_pri, 64thread_10pct_write, 64, 0.10, false)
1540 BENCH_REL (shmtx_rd_pri, 64thread_10pct_write, 64, 0.10, false)
1541 BENCH_REL (folly_ticket, 64thread_10pct_write, 64, 0.10, false)
1542 BENCH_REL (boost_shared, 64thread_10pct_write, 64, 0.10, false)
1543 BENCH_REL (pthrd_rwlock, 64thread_10pct_write, 64, 0.10, false)
1544
1545 // 1 lock used by everybody, 1% writes.  This is a more realistic example
1546 // than the concurrent_*_reads benchmark, but still shows SharedMutex locks
1547 // winning over all of the others
1548 BENCHMARK_DRAW_LINE()
1549 BENCHMARK_DRAW_LINE()
1550 BENCH_BASE(folly_rwspin, 1thread_1pct_write, 1, 0.01, false)
1551 BENCH_REL (shmtx_wr_pri, 1thread_1pct_write, 1, 0.01, false)
1552 BENCH_REL (shmtx_w_bare, 1thread_1pct_write, 1, 0.01, false)
1553 BENCH_REL (shmtx_rd_pri, 1thread_1pct_write, 1, 0.01, false)
1554 BENCH_REL (shmtx_r_bare, 1thread_1pct_write, 1, 0.01, false)
1555 BENCH_REL (folly_ticket, 1thread_1pct_write, 1, 0.01, false)
1556 BENCH_REL (boost_shared, 1thread_1pct_write, 1, 0.01, false)
1557 BENCH_REL (pthrd_rwlock, 1thread_1pct_write, 1, 0.01, false)
1558 BENCHMARK_DRAW_LINE()
1559 BENCH_BASE(folly_rwspin, 2thread_1pct_write, 2, 0.01, false)
1560 BENCH_REL (shmtx_wr_pri, 2thread_1pct_write, 2, 0.01, false)
1561 BENCH_REL (shmtx_w_bare, 2thread_1pct_write, 2, 0.01, false)
1562 BENCH_REL (shmtx_rd_pri, 2thread_1pct_write, 2, 0.01, false)
1563 BENCH_REL (shmtx_r_bare, 2thread_1pct_write, 2, 0.01, false)
1564 BENCH_REL (folly_ticket, 2thread_1pct_write, 2, 0.01, false)
1565 BENCH_REL (boost_shared, 2thread_1pct_write, 2, 0.01, false)
1566 BENCH_REL (pthrd_rwlock, 2thread_1pct_write, 2, 0.01, false)
1567 BENCHMARK_DRAW_LINE()
1568 BENCH_BASE(folly_rwspin, 4thread_1pct_write, 4, 0.01, false)
1569 BENCH_REL (shmtx_wr_pri, 4thread_1pct_write, 4, 0.01, false)
1570 BENCH_REL (shmtx_w_bare, 4thread_1pct_write, 4, 0.01, false)
1571 BENCH_REL (shmtx_rd_pri, 4thread_1pct_write, 4, 0.01, false)
1572 BENCH_REL (shmtx_r_bare, 4thread_1pct_write, 4, 0.01, false)
1573 BENCH_REL (folly_ticket, 4thread_1pct_write, 4, 0.01, false)
1574 BENCH_REL (boost_shared, 4thread_1pct_write, 4, 0.01, false)
1575 BENCH_REL (pthrd_rwlock, 4thread_1pct_write, 4, 0.01, false)
1576 BENCHMARK_DRAW_LINE()
1577 BENCH_BASE(folly_rwspin, 8thread_1pct_write, 8, 0.01, false)
1578 BENCH_REL (shmtx_wr_pri, 8thread_1pct_write, 8, 0.01, false)
1579 BENCH_REL (shmtx_w_bare, 8thread_1pct_write, 8, 0.01, false)
1580 BENCH_REL (shmtx_rd_pri, 8thread_1pct_write, 8, 0.01, false)
1581 BENCH_REL (shmtx_r_bare, 8thread_1pct_write, 8, 0.01, false)
1582 BENCH_REL (folly_ticket, 8thread_1pct_write, 8, 0.01, false)
1583 BENCH_REL (boost_shared, 8thread_1pct_write, 8, 0.01, false)
1584 BENCH_REL (pthrd_rwlock, 8thread_1pct_write, 8, 0.01, false)
1585 BENCHMARK_DRAW_LINE()
1586 BENCH_BASE(folly_rwspin, 16thread_1pct_write, 16, 0.01, false)
1587 BENCH_REL (shmtx_wr_pri, 16thread_1pct_write, 16, 0.01, false)
1588 BENCH_REL (shmtx_w_bare, 16thread_1pct_write, 16, 0.01, false)
1589 BENCH_REL (shmtx_rd_pri, 16thread_1pct_write, 16, 0.01, false)
1590 BENCH_REL (shmtx_r_bare, 16thread_1pct_write, 16, 0.01, false)
1591 BENCH_REL (folly_ticket, 16thread_1pct_write, 16, 0.01, false)
1592 BENCH_REL (boost_shared, 16thread_1pct_write, 16, 0.01, false)
1593 BENCH_REL (pthrd_rwlock, 16thread_1pct_write, 16, 0.01, false)
1594 BENCHMARK_DRAW_LINE()
1595 BENCH_BASE(folly_rwspin, 32thread_1pct_write, 32, 0.01, false)
1596 BENCH_REL (shmtx_wr_pri, 32thread_1pct_write, 32, 0.01, false)
1597 BENCH_REL (shmtx_w_bare, 32thread_1pct_write, 32, 0.01, false)
1598 BENCH_REL (shmtx_rd_pri, 32thread_1pct_write, 32, 0.01, false)
1599 BENCH_REL (shmtx_r_bare, 32thread_1pct_write, 32, 0.01, false)
1600 BENCH_REL (folly_ticket, 32thread_1pct_write, 32, 0.01, false)
1601 BENCH_REL (boost_shared, 32thread_1pct_write, 32, 0.01, false)
1602 BENCH_REL (pthrd_rwlock, 32thread_1pct_write, 32, 0.01, false)
1603 BENCHMARK_DRAW_LINE()
1604 BENCH_BASE(folly_rwspin, 64thread_1pct_write, 64, 0.01, false)
1605 BENCH_REL (shmtx_wr_pri, 64thread_1pct_write, 64, 0.01, false)
1606 BENCH_REL (shmtx_w_bare, 64thread_1pct_write, 64, 0.01, false)
1607 BENCH_REL (shmtx_rd_pri, 64thread_1pct_write, 64, 0.01, false)
1608 BENCH_REL (shmtx_r_bare, 64thread_1pct_write, 64, 0.01, false)
1609 BENCH_REL (folly_ticket, 64thread_1pct_write, 64, 0.01, false)
1610 BENCH_REL (boost_shared, 64thread_1pct_write, 64, 0.01, false)
1611 BENCH_REL (pthrd_rwlock, 64thread_1pct_write, 64, 0.01, false)
1612
1613 // Worst case scenario for deferred locks. No actual sharing, likely that
1614 // read operations will have to first set the kDeferredReadersPossibleBit,
1615 // and likely that writers will have to scan deferredReaders[].
1616 BENCHMARK_DRAW_LINE()
1617 BENCH_BASE(folly_rwspin, 2thr_2lock_50pct_write, 2, 0.50, true)
1618 BENCH_REL (shmtx_wr_pri, 2thr_2lock_50pct_write, 2, 0.50, true)
1619 BENCH_REL (shmtx_rd_pri, 2thr_2lock_50pct_write, 2, 0.50, true)
1620 BENCH_BASE(folly_rwspin, 4thr_4lock_50pct_write, 4, 0.50, true)
1621 BENCH_REL (shmtx_wr_pri, 4thr_4lock_50pct_write, 4, 0.50, true)
1622 BENCH_REL (shmtx_rd_pri, 4thr_4lock_50pct_write, 4, 0.50, true)
1623 BENCH_BASE(folly_rwspin, 8thr_8lock_50pct_write, 8, 0.50, true)
1624 BENCH_REL (shmtx_wr_pri, 8thr_8lock_50pct_write, 8, 0.50, true)
1625 BENCH_REL (shmtx_rd_pri, 8thr_8lock_50pct_write, 8, 0.50, true)
1626 BENCH_BASE(folly_rwspin, 16thr_16lock_50pct_write, 16, 0.50, true)
1627 BENCH_REL (shmtx_wr_pri, 16thr_16lock_50pct_write, 16, 0.50, true)
1628 BENCH_REL (shmtx_rd_pri, 16thr_16lock_50pct_write, 16, 0.50, true)
1629 BENCH_BASE(folly_rwspin, 32thr_32lock_50pct_write, 32, 0.50, true)
1630 BENCH_REL (shmtx_wr_pri, 32thr_32lock_50pct_write, 32, 0.50, true)
1631 BENCH_REL (shmtx_rd_pri, 32thr_32lock_50pct_write, 32, 0.50, true)
1632 BENCH_BASE(folly_rwspin, 64thr_64lock_50pct_write, 64, 0.50, true)
1633 BENCH_REL (shmtx_wr_pri, 64thr_64lock_50pct_write, 64, 0.50, true)
1634 BENCH_REL (shmtx_rd_pri, 64thr_64lock_50pct_write, 64, 0.50, true)
1635 BENCHMARK_DRAW_LINE()
1636 BENCH_BASE(folly_rwspin, 2thr_2lock_10pct_write, 2, 0.10, true)
1637 BENCH_REL (shmtx_wr_pri, 2thr_2lock_10pct_write, 2, 0.10, true)
1638 BENCH_REL (shmtx_rd_pri, 2thr_2lock_10pct_write, 2, 0.10, true)
1639 BENCH_BASE(folly_rwspin, 4thr_4lock_10pct_write, 4, 0.10, true)
1640 BENCH_REL (shmtx_wr_pri, 4thr_4lock_10pct_write, 4, 0.10, true)
1641 BENCH_REL (shmtx_rd_pri, 4thr_4lock_10pct_write, 4, 0.10, true)
1642 BENCH_BASE(folly_rwspin, 8thr_8lock_10pct_write, 8, 0.10, true)
1643 BENCH_REL (shmtx_wr_pri, 8thr_8lock_10pct_write, 8, 0.10, true)
1644 BENCH_REL (shmtx_rd_pri, 8thr_8lock_10pct_write, 8, 0.10, true)
1645 BENCH_BASE(folly_rwspin, 16thr_16lock_10pct_write, 16, 0.10, true)
1646 BENCH_REL (shmtx_wr_pri, 16thr_16lock_10pct_write, 16, 0.10, true)
1647 BENCH_REL (shmtx_rd_pri, 16thr_16lock_10pct_write, 16, 0.10, true)
1648 BENCH_BASE(folly_rwspin, 32thr_32lock_10pct_write, 32, 0.10, true)
1649 BENCH_REL (shmtx_wr_pri, 32thr_32lock_10pct_write, 32, 0.10, true)
1650 BENCH_REL (shmtx_rd_pri, 32thr_32lock_10pct_write, 32, 0.10, true)
1651 BENCH_BASE(folly_rwspin, 64thr_64lock_10pct_write, 64, 0.10, true)
1652 BENCH_REL (shmtx_wr_pri, 64thr_64lock_10pct_write, 64, 0.10, true)
1653 BENCH_REL (shmtx_rd_pri, 64thr_64lock_10pct_write, 64, 0.10, true)
1654 BENCHMARK_DRAW_LINE()
1655 BENCH_BASE(folly_rwspin, 2thr_2lock_1pct_write, 2, 0.01, true)
1656 BENCH_REL (shmtx_wr_pri, 2thr_2lock_1pct_write, 2, 0.01, true)
1657 BENCH_REL (shmtx_rd_pri, 2thr_2lock_1pct_write, 2, 0.01, true)
1658 BENCH_BASE(folly_rwspin, 4thr_4lock_1pct_write, 4, 0.01, true)
1659 BENCH_REL (shmtx_wr_pri, 4thr_4lock_1pct_write, 4, 0.01, true)
1660 BENCH_REL (shmtx_rd_pri, 4thr_4lock_1pct_write, 4, 0.01, true)
1661 BENCH_BASE(folly_rwspin, 8thr_8lock_1pct_write, 8, 0.01, true)
1662 BENCH_REL (shmtx_wr_pri, 8thr_8lock_1pct_write, 8, 0.01, true)
1663 BENCH_REL (shmtx_rd_pri, 8thr_8lock_1pct_write, 8, 0.01, true)
1664 BENCH_BASE(folly_rwspin, 16thr_16lock_1pct_write, 16, 0.01, true)
1665 BENCH_REL (shmtx_wr_pri, 16thr_16lock_1pct_write, 16, 0.01, true)
1666 BENCH_REL (shmtx_rd_pri, 16thr_16lock_1pct_write, 16, 0.01, true)
1667 BENCH_BASE(folly_rwspin, 32thr_32lock_1pct_write, 32, 0.01, true)
1668 BENCH_REL (shmtx_wr_pri, 32thr_32lock_1pct_write, 32, 0.01, true)
1669 BENCH_REL (shmtx_rd_pri, 32thr_32lock_1pct_write, 32, 0.01, true)
1670 BENCH_BASE(folly_rwspin, 64thr_64lock_1pct_write, 64, 0.01, true)
1671 BENCH_REL (shmtx_wr_pri, 64thr_64lock_1pct_write, 64, 0.01, true)
1672 BENCH_REL (shmtx_rd_pri, 64thr_64lock_1pct_write, 64, 0.01, true)
1673
1674 // Ping-pong tests have a scaled number of iterations, because their burn
1675 // loop would make them too slow otherwise.  Ping-pong with burn count of
1676 // 100k or 300k shows the advantage of soft-spin, reducing the cost of
1677 // each wakeup by about 20 usec.  (Take benchmark reported difference,
1678 // ~400 nanos, multiply by the scale of 100, then divide by 2 because
1679 // each round has two wakeups.)
1680 BENCHMARK_DRAW_LINE()
1681 BENCHMARK_DRAW_LINE()
1682 BENCH_BASE(folly_rwspin_ping_pong, burn0, 1, 0)
1683 BENCH_REL (shmtx_w_bare_ping_pong, burn0, 1, 0)
1684 BENCH_REL (shmtx_r_bare_ping_pong, burn0, 1, 0)
1685 BENCH_REL (folly_ticket_ping_pong, burn0, 1, 0)
1686 BENCH_REL (boost_shared_ping_pong, burn0, 1, 0)
1687 BENCH_REL (pthrd_rwlock_ping_pong, burn0, 1, 0)
1688 BENCHMARK_DRAW_LINE()
1689 BENCH_BASE(folly_rwspin_ping_pong, burn100k, 100, 100000)
1690 BENCH_REL (shmtx_w_bare_ping_pong, burn100k, 100, 100000)
1691 BENCH_REL (shmtx_r_bare_ping_pong, burn100k, 100, 100000)
1692 BENCH_REL (folly_ticket_ping_pong, burn100k, 100, 100000)
1693 BENCH_REL (boost_shared_ping_pong, burn100k, 100, 100000)
1694 BENCH_REL (pthrd_rwlock_ping_pong, burn100k, 100, 100000)
1695 BENCHMARK_DRAW_LINE()
1696 BENCH_BASE(folly_rwspin_ping_pong, burn300k, 100, 300000)
1697 BENCH_REL (shmtx_w_bare_ping_pong, burn300k, 100, 300000)
1698 BENCH_REL (shmtx_r_bare_ping_pong, burn300k, 100, 300000)
1699 BENCH_REL (folly_ticket_ping_pong, burn300k, 100, 300000)
1700 BENCH_REL (boost_shared_ping_pong, burn300k, 100, 300000)
1701 BENCH_REL (pthrd_rwlock_ping_pong, burn300k, 100, 300000)
1702 BENCHMARK_DRAW_LINE()
1703 BENCH_BASE(folly_rwspin_ping_pong, burn1M, 1000, 1000000)
1704 BENCH_REL (shmtx_w_bare_ping_pong, burn1M, 1000, 1000000)
1705 BENCH_REL (shmtx_r_bare_ping_pong, burn1M, 1000, 1000000)
1706 BENCH_REL (folly_ticket_ping_pong, burn1M, 1000, 1000000)
1707 BENCH_REL (boost_shared_ping_pong, burn1M, 1000, 1000000)
1708 BENCH_REL (pthrd_rwlock_ping_pong, burn1M, 1000, 1000000)
1709
1710 // Reproduce with 10 minutes and
1711 //   sudo nice -n -20 \
1712 //     shared_mutex_test --benchmark --bm_min_iters=1000000
1713 //
1714 // Comparison use folly::RWSpinLock as the baseline, with the
1715 // following row being the default SharedMutex (using *Holder or
1716 // Token-ful methods).
1717 // ============================================================================
1718 // folly/experimental/test/SharedMutexTest.cpp     relative  time/iter  iters/s
1719 // ============================================================================
1720 // single_thread_lock_shared_unlock_shared                     22.78ns   43.89M
1721 // single_thread_lock_unlock                                   26.01ns   38.45M
1722 // ----------------------------------------------------------------------------
1723 // ----------------------------------------------------------------------------
1724 // folly_rwspin_reads(1thread)                                 15.09ns   66.25M
1725 // shmtx_wr_pri_reads(1thread)                       69.89%    21.60ns   46.30M
1726 // shmtx_w_bare_reads(1thread)                       58.25%    25.91ns   38.59M
1727 // shmtx_rd_pri_reads(1thread)                       72.50%    20.82ns   48.03M
1728 // shmtx_r_bare_reads(1thread)                       58.27%    25.91ns   38.60M
1729 // folly_ticket_reads(1thread)                       54.80%    27.55ns   36.30M
1730 // boost_shared_reads(1thread)                       10.88%   138.80ns    7.20M
1731 // pthrd_rwlock_reads(1thread)                       40.68%    37.11ns   26.95M
1732 // ----------------------------------------------------------------------------
1733 // folly_rwspin_reads(2thread)                                 92.63ns   10.80M
1734 // shmtx_wr_pri_reads(2thread)                      462.86%    20.01ns   49.97M
1735 // shmtx_w_bare_reads(2thread)                      430.53%    21.51ns   46.48M
1736 // shmtx_rd_pri_reads(2thread)                      487.13%    19.01ns   52.59M
1737 // shmtx_r_bare_reads(2thread)                      433.35%    21.37ns   46.79M
1738 // folly_ticket_reads(2thread)                       69.82%   132.67ns    7.54M
1739 // boost_shared_reads(2thread)                       36.66%   252.63ns    3.96M
1740 // pthrd_rwlock_reads(2thread)                      127.76%    72.50ns   13.79M
1741 // ----------------------------------------------------------------------------
1742 // folly_rwspin_reads(4thread)                                 97.45ns   10.26M
1743 // shmtx_wr_pri_reads(4thread)                      978.22%     9.96ns  100.38M
1744 // shmtx_w_bare_reads(4thread)                      908.35%    10.73ns   93.21M
1745 // shmtx_rd_pri_reads(4thread)                     1032.29%     9.44ns  105.93M
1746 // shmtx_r_bare_reads(4thread)                      912.38%    10.68ns   93.63M
1747 // folly_ticket_reads(4thread)                       46.08%   211.46ns    4.73M
1748 // boost_shared_reads(4thread)                       25.00%   389.74ns    2.57M
1749 // pthrd_rwlock_reads(4thread)                       47.53%   205.01ns    4.88M
1750 // ----------------------------------------------------------------------------
1751 // folly_rwspin_reads(8thread)                                147.24ns    6.79M
1752 // shmtx_wr_pri_reads(8thread)                     2915.66%     5.05ns  198.02M
1753 // shmtx_w_bare_reads(8thread)                     2699.32%     5.45ns  183.32M
1754 // shmtx_rd_pri_reads(8thread)                     3092.58%     4.76ns  210.03M
1755 // shmtx_r_bare_reads(8thread)                     2744.63%     5.36ns  186.40M
1756 // folly_ticket_reads(8thread)                       54.84%   268.47ns    3.72M
1757 // boost_shared_reads(8thread)                       42.40%   347.30ns    2.88M
1758 // pthrd_rwlock_reads(8thread)                       78.90%   186.63ns    5.36M
1759 // ----------------------------------------------------------------------------
1760 // folly_rwspin_reads(16thread)                               166.25ns    6.02M
1761 // shmtx_wr_pri_reads(16thread)                    6133.03%     2.71ns  368.91M
1762 // shmtx_w_bare_reads(16thread)                    5936.05%     2.80ns  357.06M
1763 // shmtx_rd_pri_reads(16thread)                    6786.57%     2.45ns  408.22M
1764 // shmtx_r_bare_reads(16thread)                    5995.54%     2.77ns  360.64M
1765 // folly_ticket_reads(16thread)                      56.35%   295.01ns    3.39M
1766 // boost_shared_reads(16thread)                      51.62%   322.08ns    3.10M
1767 // pthrd_rwlock_reads(16thread)                      92.47%   179.79ns    5.56M
1768 // ----------------------------------------------------------------------------
1769 // folly_rwspin_reads(32thread)                               107.72ns    9.28M
1770 // shmtx_wr_pri_reads(32thread)                    6772.80%     1.59ns  628.77M
1771 // shmtx_w_bare_reads(32thread)                    6236.13%     1.73ns  578.94M
1772 // shmtx_rd_pri_reads(32thread)                    8143.32%     1.32ns  756.00M
1773 // shmtx_r_bare_reads(32thread)                    6485.18%     1.66ns  602.06M
1774 // folly_ticket_reads(32thread)                      35.12%   306.73ns    3.26M
1775 // boost_shared_reads(32thread)                      28.19%   382.17ns    2.62M
1776 // pthrd_rwlock_reads(32thread)                      65.29%   164.99ns    6.06M
1777 // ----------------------------------------------------------------------------
1778 // folly_rwspin_reads(64thread)                               119.46ns    8.37M
1779 // shmtx_wr_pri_reads(64thread)                    6744.92%     1.77ns  564.60M
1780 // shmtx_w_bare_reads(64thread)                    6268.50%     1.91ns  524.72M
1781 // shmtx_rd_pri_reads(64thread)                    7508.56%     1.59ns  628.52M
1782 // shmtx_r_bare_reads(64thread)                    6299.53%     1.90ns  527.32M
1783 // folly_ticket_reads(64thread)                      37.42%   319.26ns    3.13M
1784 // boost_shared_reads(64thread)                      32.58%   366.70ns    2.73M
1785 // pthrd_rwlock_reads(64thread)                      73.64%   162.24ns    6.16M
1786 // ----------------------------------------------------------------------------
1787 // ----------------------------------------------------------------------------
1788 // folly_rwspin(1thread_all_write)                             25.51ns   39.19M
1789 // shmtx_wr_pri(1thread_all_write)                   97.38%    26.20ns   38.17M
1790 // shmtx_rd_pri(1thread_all_write)                   97.55%    26.16ns   38.23M
1791 // folly_ticket(1thread_all_write)                   90.98%    28.04ns   35.66M
1792 // boost_shared(1thread_all_write)                   16.80%   151.89ns    6.58M
1793 // pthrd_rwlock(1thread_all_write)                   63.86%    39.96ns   25.03M
1794 // pthrd_mutex_(1thread_all_write)                   82.05%    31.09ns   32.16M
1795 // ----------------------------------------------------------------------------
1796 // folly_rwspin(2thread_all_write)                            100.70ns    9.93M
1797 // shmtx_wr_pri(2thread_all_write)                   40.83%   246.61ns    4.05M
1798 // shmtx_rd_pri(2thread_all_write)                   40.53%   248.44ns    4.03M
1799 // folly_ticket(2thread_all_write)                   58.49%   172.17ns    5.81M
1800 // boost_shared(2thread_all_write)                   24.26%   415.00ns    2.41M
1801 // pthrd_rwlock(2thread_all_write)                   41.35%   243.49ns    4.11M
1802 // pthrd_mutex_(2thread_all_write)                  146.91%    68.55ns   14.59M
1803 // ----------------------------------------------------------------------------
1804 // folly_rwspin(4thread_all_write)                            199.52ns    5.01M
1805 // shmtx_wr_pri(4thread_all_write)                   51.71%   385.86ns    2.59M
1806 // shmtx_rd_pri(4thread_all_write)                   49.43%   403.62ns    2.48M
1807 // folly_ticket(4thread_all_write)                  117.88%   169.26ns    5.91M
1808 // boost_shared(4thread_all_write)                    9.81%     2.03us  491.48K
1809 // pthrd_rwlock(4thread_all_write)                   28.23%   706.69ns    1.42M
1810 // pthrd_mutex_(4thread_all_write)                  111.54%   178.88ns    5.59M
1811 // ----------------------------------------------------------------------------
1812 // folly_rwspin(8thread_all_write)                            304.61ns    3.28M
1813 // shmtx_wr_pri(8thread_all_write)                   69.77%   436.59ns    2.29M
1814 // shmtx_rd_pri(8thread_all_write)                   66.58%   457.51ns    2.19M
1815 // folly_ticket(8thread_all_write)                  141.00%   216.03ns    4.63M
1816 // boost_shared(8thread_all_write)                    6.11%     4.99us  200.59K
1817 // pthrd_rwlock(8thread_all_write)                   38.03%   800.88ns    1.25M
1818 // pthrd_mutex_(8thread_all_write)                  177.66%   171.45ns    5.83M
1819 // ----------------------------------------------------------------------------
1820 // folly_rwspin(16thread_all_write)                           576.97ns    1.73M
1821 // shmtx_wr_pri(16thread_all_write)                 105.72%   545.77ns    1.83M
1822 // shmtx_rd_pri(16thread_all_write)                 105.13%   548.83ns    1.82M
1823 // folly_ticket(16thread_all_write)                 161.70%   356.82ns    2.80M
1824 // boost_shared(16thread_all_write)                   7.73%     7.46us  134.03K
1825 // pthrd_rwlock(16thread_all_write)                  96.88%   595.54ns    1.68M
1826 // pthrd_mutex_(16thread_all_write)                 330.44%   174.61ns    5.73M
1827 // ----------------------------------------------------------------------------
1828 // folly_rwspin(32thread_all_write)                             1.41us  707.76K
1829 // shmtx_wr_pri(32thread_all_write)                 240.46%   587.58ns    1.70M
1830 // shmtx_rd_pri(32thread_all_write)                 393.71%   358.87ns    2.79M
1831 // folly_ticket(32thread_all_write)                 325.07%   434.65ns    2.30M
1832 // boost_shared(32thread_all_write)                  18.57%     7.61us  131.43K
1833 // pthrd_rwlock(32thread_all_write)                 266.78%   529.62ns    1.89M
1834 // pthrd_mutex_(32thread_all_write)                 877.89%   160.94ns    6.21M
1835 // ----------------------------------------------------------------------------
1836 // folly_rwspin(64thread_all_write)                             1.76us  566.94K
1837 // shmtx_wr_pri(64thread_all_write)                 255.67%   689.91ns    1.45M
1838 // shmtx_rd_pri(64thread_all_write)                 468.82%   376.23ns    2.66M
1839 // folly_ticket(64thread_all_write)                 294.72%   598.49ns    1.67M
1840 // boost_shared(64thread_all_write)                  23.39%     7.54us  132.58K
1841 // pthrd_rwlock(64thread_all_write)                 321.39%   548.83ns    1.82M
1842 // pthrd_mutex_(64thread_all_write)                1165.04%   151.40ns    6.61M
1843 // ----------------------------------------------------------------------------
1844 // ----------------------------------------------------------------------------
1845 // folly_rwspin(1thread_10pct_write)                           19.51ns   51.26M
1846 // shmtx_wr_pri(1thread_10pct_write)                 83.25%    23.43ns   42.67M
1847 // shmtx_rd_pri(1thread_10pct_write)                 83.31%    23.42ns   42.71M
1848 // folly_ticket(1thread_10pct_write)                 70.88%    27.52ns   36.34M
1849 // boost_shared(1thread_10pct_write)                 13.09%   148.99ns    6.71M
1850 // pthrd_rwlock(1thread_10pct_write)                 47.41%    41.15ns   24.30M
1851 // ----------------------------------------------------------------------------
1852 // folly_rwspin(2thread_10pct_write)                          159.42ns    6.27M
1853 // shmtx_wr_pri(2thread_10pct_write)                188.44%    84.60ns   11.82M
1854 // shmtx_rd_pri(2thread_10pct_write)                188.29%    84.67ns   11.81M
1855 // folly_ticket(2thread_10pct_write)                140.28%   113.64ns    8.80M
1856 // boost_shared(2thread_10pct_write)                 42.09%   378.81ns    2.64M
1857 // pthrd_rwlock(2thread_10pct_write)                103.86%   153.49ns    6.51M
1858 // ----------------------------------------------------------------------------
1859 // folly_rwspin(4thread_10pct_write)                          193.35ns    5.17M
1860 // shmtx_wr_pri(4thread_10pct_write)                184.30%   104.91ns    9.53M
1861 // shmtx_rd_pri(4thread_10pct_write)                163.76%   118.07ns    8.47M
1862 // folly_ticket(4thread_10pct_write)                124.07%   155.84ns    6.42M
1863 // boost_shared(4thread_10pct_write)                 16.32%     1.18us  843.92K
1864 // pthrd_rwlock(4thread_10pct_write)                 48.59%   397.94ns    2.51M
1865 // ----------------------------------------------------------------------------
1866 // folly_rwspin(8thread_10pct_write)                          373.17ns    2.68M
1867 // shmtx_wr_pri(8thread_10pct_write)                252.02%   148.08ns    6.75M
1868 // shmtx_rd_pri(8thread_10pct_write)                203.59%   183.30ns    5.46M
1869 // folly_ticket(8thread_10pct_write)                184.37%   202.40ns    4.94M
1870 // boost_shared(8thread_10pct_write)                 15.85%     2.35us  424.72K
1871 // pthrd_rwlock(8thread_10pct_write)                 83.03%   449.45ns    2.22M
1872 // ----------------------------------------------------------------------------
1873 // folly_rwspin(16thread_10pct_write)                         742.87ns    1.35M
1874 // shmtx_wr_pri(16thread_10pct_write)               344.27%   215.78ns    4.63M
1875 // shmtx_rd_pri(16thread_10pct_write)               287.04%   258.80ns    3.86M
1876 // folly_ticket(16thread_10pct_write)               277.25%   267.94ns    3.73M
1877 // boost_shared(16thread_10pct_write)                15.33%     4.85us  206.30K
1878 // pthrd_rwlock(16thread_10pct_write)               158.34%   469.16ns    2.13M
1879 // ----------------------------------------------------------------------------
1880 // folly_rwspin(32thread_10pct_write)                         799.97ns    1.25M
1881 // shmtx_wr_pri(32thread_10pct_write)               351.40%   227.65ns    4.39M
1882 // shmtx_rd_pri(32thread_10pct_write)               341.71%   234.11ns    4.27M
1883 // folly_ticket(32thread_10pct_write)               245.91%   325.31ns    3.07M
1884 // boost_shared(32thread_10pct_write)                 7.72%    10.36us   96.56K
1885 // pthrd_rwlock(32thread_10pct_write)               165.87%   482.30ns    2.07M
1886 // ----------------------------------------------------------------------------
1887 // folly_rwspin(64thread_10pct_write)                           1.12us  892.01K
1888 // shmtx_wr_pri(64thread_10pct_write)               429.84%   260.81ns    3.83M
1889 // shmtx_rd_pri(64thread_10pct_write)               456.93%   245.35ns    4.08M
1890 // folly_ticket(64thread_10pct_write)               219.21%   511.42ns    1.96M
1891 // boost_shared(64thread_10pct_write)                 5.43%    20.65us   48.44K
1892 // pthrd_rwlock(64thread_10pct_write)               233.93%   479.23ns    2.09M
1893 // ----------------------------------------------------------------------------
1894 // ----------------------------------------------------------------------------
1895 // folly_rwspin(1thread_1pct_write)                            18.88ns   52.98M
1896 // shmtx_wr_pri(1thread_1pct_write)                  81.53%    23.15ns   43.19M
1897 // shmtx_w_bare(1thread_1pct_write)                  67.90%    27.80ns   35.97M
1898 // shmtx_rd_pri(1thread_1pct_write)                  81.50%    23.16ns   43.18M
1899 // shmtx_r_bare(1thread_1pct_write)                  67.74%    27.86ns   35.89M
1900 // folly_ticket(1thread_1pct_write)                  68.68%    27.48ns   36.39M
1901 // boost_shared(1thread_1pct_write)                  12.80%   147.51ns    6.78M
1902 // pthrd_rwlock(1thread_1pct_write)                  45.81%    41.20ns   24.27M
1903 // ----------------------------------------------------------------------------
1904 // folly_rwspin(2thread_1pct_write)                           125.85ns    7.95M
1905 // shmtx_wr_pri(2thread_1pct_write)                 359.04%    35.05ns   28.53M
1906 // shmtx_w_bare(2thread_1pct_write)                 475.60%    26.46ns   37.79M
1907 // shmtx_rd_pri(2thread_1pct_write)                 332.75%    37.82ns   26.44M
1908 // shmtx_r_bare(2thread_1pct_write)                 115.64%   108.83ns    9.19M
1909 // folly_ticket(2thread_1pct_write)                 140.24%    89.74ns   11.14M
1910 // boost_shared(2thread_1pct_write)                  40.62%   309.82ns    3.23M
1911 // pthrd_rwlock(2thread_1pct_write)                 134.67%    93.45ns   10.70M
1912 // ----------------------------------------------------------------------------
1913 // folly_rwspin(4thread_1pct_write)                           126.70ns    7.89M
1914 // shmtx_wr_pri(4thread_1pct_write)                 422.20%    30.01ns   33.32M
1915 // shmtx_w_bare(4thread_1pct_write)                 403.52%    31.40ns   31.85M
1916 // shmtx_rd_pri(4thread_1pct_write)                 282.50%    44.85ns   22.30M
1917 // shmtx_r_bare(4thread_1pct_write)                  66.30%   191.10ns    5.23M
1918 // folly_ticket(4thread_1pct_write)                  91.93%   137.83ns    7.26M
1919 // boost_shared(4thread_1pct_write)                  22.74%   557.10ns    1.80M
1920 // pthrd_rwlock(4thread_1pct_write)                  55.66%   227.62ns    4.39M
1921 // ----------------------------------------------------------------------------
1922 // folly_rwspin(8thread_1pct_write)                           169.42ns    5.90M
1923 // shmtx_wr_pri(8thread_1pct_write)                 567.81%    29.84ns   33.51M
1924 // shmtx_w_bare(8thread_1pct_write)                 519.18%    32.63ns   30.64M
1925 // shmtx_rd_pri(8thread_1pct_write)                 172.36%    98.30ns   10.17M
1926 // shmtx_r_bare(8thread_1pct_write)                  75.56%   224.21ns    4.46M
1927 // folly_ticket(8thread_1pct_write)                 104.03%   162.85ns    6.14M
1928 // boost_shared(8thread_1pct_write)                  22.01%   769.73ns    1.30M
1929 // pthrd_rwlock(8thread_1pct_write)                  71.79%   235.99ns    4.24M
1930 // ----------------------------------------------------------------------------
1931 // folly_rwspin(16thread_1pct_write)                          385.88ns    2.59M
1932 // shmtx_wr_pri(16thread_1pct_write)               1039.03%    37.14ns   26.93M
1933 // shmtx_w_bare(16thread_1pct_write)                997.26%    38.69ns   25.84M
1934 // shmtx_rd_pri(16thread_1pct_write)                263.60%   146.39ns    6.83M
1935 // shmtx_r_bare(16thread_1pct_write)                173.16%   222.85ns    4.49M
1936 // folly_ticket(16thread_1pct_write)                179.37%   215.13ns    4.65M
1937 // boost_shared(16thread_1pct_write)                 26.95%     1.43us  698.42K
1938 // pthrd_rwlock(16thread_1pct_write)                166.70%   231.48ns    4.32M
1939 // ----------------------------------------------------------------------------
1940 // folly_rwspin(32thread_1pct_write)                          382.49ns    2.61M
1941 // shmtx_wr_pri(32thread_1pct_write)               1046.64%    36.54ns   27.36M
1942 // shmtx_w_bare(32thread_1pct_write)                922.87%    41.45ns   24.13M
1943 // shmtx_rd_pri(32thread_1pct_write)                251.93%   151.82ns    6.59M
1944 // shmtx_r_bare(32thread_1pct_write)                176.44%   216.78ns    4.61M
1945 // folly_ticket(32thread_1pct_write)                131.07%   291.82ns    3.43M
1946 // boost_shared(32thread_1pct_write)                 12.77%     2.99us  333.95K
1947 // pthrd_rwlock(32thread_1pct_write)                173.43%   220.55ns    4.53M
1948 // ----------------------------------------------------------------------------
1949 // folly_rwspin(64thread_1pct_write)                          510.54ns    1.96M
1950 // shmtx_wr_pri(64thread_1pct_write)               1378.27%    37.04ns   27.00M
1951 // shmtx_w_bare(64thread_1pct_write)               1178.24%    43.33ns   23.08M
1952 // shmtx_rd_pri(64thread_1pct_write)                325.29%   156.95ns    6.37M
1953 // shmtx_r_bare(64thread_1pct_write)                247.82%   206.02ns    4.85M
1954 // folly_ticket(64thread_1pct_write)                117.87%   433.13ns    2.31M
1955 // boost_shared(64thread_1pct_write)                  9.45%     5.40us  185.09K
1956 // pthrd_rwlock(64thread_1pct_write)                236.72%   215.68ns    4.64M
1957 // ----------------------------------------------------------------------------
1958 // folly_rwspin(2thr_2lock_50pct_write)                        10.85ns   92.15M
1959 // shmtx_wr_pri(2thr_2lock_50pct_write)              81.73%    13.28ns   75.32M
1960 // shmtx_rd_pri(2thr_2lock_50pct_write)              81.82%    13.26ns   75.40M
1961 // folly_rwspin(4thr_4lock_50pct_write)                         5.29ns  188.90M
1962 // shmtx_wr_pri(4thr_4lock_50pct_write)              80.89%     6.54ns  152.80M
1963 // shmtx_rd_pri(4thr_4lock_50pct_write)              81.07%     6.53ns  153.14M
1964 // folly_rwspin(8thr_8lock_50pct_write)                         2.63ns  380.57M
1965 // shmtx_wr_pri(8thr_8lock_50pct_write)              80.56%     3.26ns  306.57M
1966 // shmtx_rd_pri(8thr_8lock_50pct_write)              80.29%     3.27ns  305.54M
1967 // folly_rwspin(16thr_16lock_50pct_write)                       1.31ns  764.70M
1968 // shmtx_wr_pri(16thr_16lock_50pct_write)            79.32%     1.65ns  606.54M
1969 // shmtx_rd_pri(16thr_16lock_50pct_write)            79.62%     1.64ns  608.84M
1970 // folly_rwspin(32thr_32lock_50pct_write)                       1.20ns  836.75M
1971 // shmtx_wr_pri(32thr_32lock_50pct_write)            91.67%     1.30ns  767.07M
1972 // shmtx_rd_pri(32thr_32lock_50pct_write)            92.00%     1.30ns  769.82M
1973 // folly_rwspin(64thr_64lock_50pct_write)                       1.39ns  717.80M
1974 // shmtx_wr_pri(64thr_64lock_50pct_write)            93.21%     1.49ns  669.08M
1975 // shmtx_rd_pri(64thr_64lock_50pct_write)            92.49%     1.51ns  663.89M
1976 // ----------------------------------------------------------------------------
1977 // folly_rwspin(2thr_2lock_10pct_write)                        10.24ns   97.70M
1978 // shmtx_wr_pri(2thr_2lock_10pct_write)              76.46%    13.39ns   74.70M
1979 // shmtx_rd_pri(2thr_2lock_10pct_write)              76.35%    13.41ns   74.60M
1980 // folly_rwspin(4thr_4lock_10pct_write)                         5.02ns  199.03M
1981 // shmtx_wr_pri(4thr_4lock_10pct_write)              75.83%     6.63ns  150.91M
1982 // shmtx_rd_pri(4thr_4lock_10pct_write)              76.10%     6.60ns  151.46M
1983 // folly_rwspin(8thr_8lock_10pct_write)                         2.47ns  405.50M
1984 // shmtx_wr_pri(8thr_8lock_10pct_write)              74.54%     3.31ns  302.27M
1985 // shmtx_rd_pri(8thr_8lock_10pct_write)              74.85%     3.29ns  303.52M
1986 // folly_rwspin(16thr_16lock_10pct_write)                       1.22ns  818.68M
1987 // shmtx_wr_pri(16thr_16lock_10pct_write)            73.35%     1.67ns  600.47M
1988 // shmtx_rd_pri(16thr_16lock_10pct_write)            73.38%     1.66ns  600.73M
1989 // folly_rwspin(32thr_32lock_10pct_write)                       1.21ns  827.95M
1990 // shmtx_wr_pri(32thr_32lock_10pct_write)            96.13%     1.26ns  795.89M
1991 // shmtx_rd_pri(32thr_32lock_10pct_write)            96.01%     1.26ns  794.95M
1992 // folly_rwspin(64thr_64lock_10pct_write)                       1.40ns  716.17M
1993 // shmtx_wr_pri(64thr_64lock_10pct_write)            96.91%     1.44ns  694.03M
1994 // shmtx_rd_pri(64thr_64lock_10pct_write)            96.85%     1.44ns  693.64M
1995 // ----------------------------------------------------------------------------
1996 // folly_rwspin(2thr_2lock_1pct_write)                         10.11ns   98.91M
1997 // shmtx_wr_pri(2thr_2lock_1pct_write)               75.07%    13.47ns   74.25M
1998 // shmtx_rd_pri(2thr_2lock_1pct_write)               74.98%    13.48ns   74.16M
1999 // folly_rwspin(4thr_4lock_1pct_write)                          4.96ns  201.77M
2000 // shmtx_wr_pri(4thr_4lock_1pct_write)               74.59%     6.64ns  150.49M
2001 // shmtx_rd_pri(4thr_4lock_1pct_write)               74.60%     6.64ns  150.51M
2002 // folly_rwspin(8thr_8lock_1pct_write)                          2.44ns  410.42M
2003 // shmtx_wr_pri(8thr_8lock_1pct_write)               73.68%     3.31ns  302.41M
2004 // shmtx_rd_pri(8thr_8lock_1pct_write)               73.38%     3.32ns  301.16M
2005 // folly_rwspin(16thr_16lock_1pct_write)                        1.21ns  827.53M
2006 // shmtx_wr_pri(16thr_16lock_1pct_write)             72.11%     1.68ns  596.74M
2007 // shmtx_rd_pri(16thr_16lock_1pct_write)             72.23%     1.67ns  597.73M
2008 // folly_rwspin(32thr_32lock_1pct_write)                        1.22ns  819.53M
2009 // shmtx_wr_pri(32thr_32lock_1pct_write)             98.17%     1.24ns  804.50M
2010 // shmtx_rd_pri(32thr_32lock_1pct_write)             98.21%     1.24ns  804.86M
2011 // folly_rwspin(64thr_64lock_1pct_write)                        1.41ns  710.26M
2012 // shmtx_wr_pri(64thr_64lock_1pct_write)             97.81%     1.44ns  694.71M
2013 // shmtx_rd_pri(64thr_64lock_1pct_write)             99.44%     1.42ns  706.28M
2014 // ----------------------------------------------------------------------------
2015 // ----------------------------------------------------------------------------
2016 // folly_rwspin_ping_pong(burn0)                              641.24ns    1.56M
2017 // shmtx_w_bare_ping_pong(burn0)                     91.07%   704.12ns    1.42M
2018 // shmtx_r_bare_ping_pong(burn0)                     78.70%   814.84ns    1.23M
2019 // folly_ticket_ping_pong(burn0)                     85.67%   748.53ns    1.34M
2020 // boost_shared_ping_pong(burn0)                      5.58%    11.50us   86.96K
2021 // pthrd_rwlock_ping_pong(burn0)                      8.81%     7.28us  137.40K
2022 // ----------------------------------------------------------------------------
2023 // folly_rwspin_ping_pong(burn100k)                           678.97ns    1.47M
2024 // shmtx_w_bare_ping_pong(burn100k)                  99.73%   680.78ns    1.47M
2025 // shmtx_r_bare_ping_pong(burn100k)                  98.67%   688.13ns    1.45M
2026 // folly_ticket_ping_pong(burn100k)                  99.31%   683.68ns    1.46M
2027 // boost_shared_ping_pong(burn100k)                  58.23%     1.17us  857.64K
2028 // pthrd_rwlock_ping_pong(burn100k)                  57.43%     1.18us  845.86K
2029 // ----------------------------------------------------------------------------
2030 // folly_rwspin_ping_pong(burn300k)                             2.03us  492.99K
2031 // shmtx_w_bare_ping_pong(burn300k)                  99.98%     2.03us  492.88K
2032 // shmtx_r_bare_ping_pong(burn300k)                  99.94%     2.03us  492.68K
2033 // folly_ticket_ping_pong(burn300k)                  99.88%     2.03us  492.40K
2034 // boost_shared_ping_pong(burn300k)                  81.43%     2.49us  401.47K
2035 // pthrd_rwlock_ping_pong(burn300k)                  83.22%     2.44us  410.29K
2036 // ----------------------------------------------------------------------------
2037 // folly_rwspin_ping_pong(burn1M)                             677.07ns    1.48M
2038 // shmtx_w_bare_ping_pong(burn1M)                   100.50%   673.74ns    1.48M
2039 // shmtx_r_bare_ping_pong(burn1M)                   100.14%   676.12ns    1.48M
2040 // folly_ticket_ping_pong(burn1M)                   100.44%   674.14ns    1.48M
2041 // boost_shared_ping_pong(burn1M)                    93.04%   727.72ns    1.37M
2042 // pthrd_rwlock_ping_pong(burn1M)                    94.52%   716.30ns    1.40M
2043 // ============================================================================
2044
2045 int main(int argc, char** argv) {
2046   (void)folly_rwspin_reads;
2047   (void)shmtx_wr_pri_reads;
2048   (void)shmtx_w_bare_reads;
2049   (void)shmtx_rd_pri_reads;
2050   (void)shmtx_r_bare_reads;
2051   (void)folly_ticket_reads;
2052   (void)boost_shared_reads;
2053   (void)pthrd_rwlock_reads;
2054   (void)folly_rwspin;
2055   (void)shmtx_wr_pri;
2056   (void)shmtx_w_bare;
2057   (void)shmtx_rd_pri;
2058   (void)shmtx_r_bare;
2059   (void)folly_ticket;
2060   (void)boost_shared;
2061   (void)pthrd_rwlock;
2062   (void)pthrd_mutex_;
2063   (void)folly_rwspin_ping_pong;
2064   (void)shmtx_w_bare_ping_pong;
2065   (void)shmtx_r_bare_ping_pong;
2066   (void)folly_ticket_ping_pong;
2067   (void)boost_shared_ping_pong;
2068   (void)pthrd_rwlock_ping_pong;
2069
2070   testing::InitGoogleTest(&argc, argv);
2071   gflags::ParseCommandLineFlags(&argc, &argv, true);
2072   int rv = RUN_ALL_TESTS();
2073   folly::runBenchmarksOnFlag();
2074   return rv;
2075 }