| 1 | /* |
| 2 | * ompt-tsan.cpp -- Archer runtime library, TSan annotations for Archer |
| 3 | */ |
| 4 | |
| 5 | //===----------------------------------------------------------------------===// |
| 6 | // |
| 7 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 8 | // See https://llvm.org/LICENSE.txt for details. |
| 9 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef __STDC_FORMAT_MACROS |
| 14 | #define __STDC_FORMAT_MACROS |
| 15 | #endif |
| 16 | |
| 17 | #include <algorithm> |
| 18 | #include <atomic> |
| 19 | #include <cassert> |
| 20 | #include <cstdlib> |
| 21 | #include <cstring> |
| 22 | #include <dlfcn.h> |
| 23 | #include <inttypes.h> |
| 24 | #include <iostream> |
| 25 | #include <list> |
| 26 | #include <mutex> |
| 27 | #include <sstream> |
| 28 | #include <string> |
| 29 | #include <sys/resource.h> |
| 30 | #include <unistd.h> |
| 31 | #include <unordered_map> |
| 32 | #include <vector> |
| 33 | |
| 34 | #include "omp-tools.h" |
| 35 | |
| 36 | // Define attribute that indicates that the fall through from the previous |
| 37 | // case label is intentional and should not be diagnosed by a compiler |
| 38 | // Code from libcxx/include/__config |
| 39 | // Use a function like macro to imply that it must be followed by a semicolon |
| 40 | #if __cplusplus > 201402L && __has_cpp_attribute(fallthrough) |
| 41 | #define KMP_FALLTHROUGH() [[fallthrough]] |
| 42 | // icc cannot properly tell this attribute is absent so force off |
| 43 | #elif defined(__INTEL_COMPILER) |
| 44 | #define KMP_FALLTHROUGH() ((void)0) |
| 45 | #elif __has_cpp_attribute(clang::fallthrough) |
| 46 | #define KMP_FALLTHROUGH() [[clang::fallthrough]] |
| 47 | #elif __has_attribute(fallthrough) || __GNUC__ >= 7 |
| 48 | #define KMP_FALLTHROUGH() __attribute__((__fallthrough__)) |
| 49 | #else |
| 50 | #define KMP_FALLTHROUGH() ((void)0) |
| 51 | #endif |
| 52 | |
| 53 | static int hasReductionCallback; |
| 54 | |
| 55 | namespace { |
| 56 | class ArcherFlags { |
| 57 | public: |
| 58 | #if (LLVM_VERSION) >= 40 |
| 59 | int flush_shadow{0}; |
| 60 | #endif |
| 61 | int {0}; |
| 62 | int verbose{0}; |
| 63 | int enabled{1}; |
| 64 | int report_data_leak{0}; |
| 65 | int ignore_serial{0}; |
| 66 | std::atomic<int> all_memory{0}; |
| 67 | |
| 68 | ArcherFlags(const char *env) { |
| 69 | if (env) { |
| 70 | std::vector<std::string> tokens; |
| 71 | std::string token; |
| 72 | std::string str(env); |
| 73 | std::istringstream iss(str); |
| 74 | int tmp_int; |
| 75 | while (std::getline(in&: iss, str&: token, delim: ' ')) |
| 76 | tokens.push_back(x: token); |
| 77 | |
| 78 | for (std::vector<std::string>::iterator it = tokens.begin(); |
| 79 | it != tokens.end(); ++it) { |
| 80 | #if (LLVM_VERSION) >= 40 |
| 81 | if (sscanf(it->c_str(), "flush_shadow=%d" , &flush_shadow)) |
| 82 | continue; |
| 83 | #endif |
| 84 | if (sscanf(s: it->c_str(), format: "print_max_rss=%d" , &print_max_rss)) |
| 85 | continue; |
| 86 | if (sscanf(s: it->c_str(), format: "verbose=%d" , &verbose)) |
| 87 | continue; |
| 88 | if (sscanf(s: it->c_str(), format: "report_data_leak=%d" , &report_data_leak)) |
| 89 | continue; |
| 90 | if (sscanf(s: it->c_str(), format: "enable=%d" , &enabled)) |
| 91 | continue; |
| 92 | if (sscanf(s: it->c_str(), format: "ignore_serial=%d" , &ignore_serial)) |
| 93 | continue; |
| 94 | if (sscanf(s: it->c_str(), format: "all_memory=%d" , &tmp_int)) { |
| 95 | all_memory = tmp_int; |
| 96 | continue; |
| 97 | } |
| 98 | std::cerr << "Illegal values for ARCHER_OPTIONS variable: " << token |
| 99 | << std::endl; |
| 100 | } |
| 101 | } |
| 102 | } |
| 103 | }; |
| 104 | |
| 105 | class TsanFlags { |
| 106 | public: |
| 107 | int ignore_noninstrumented_modules; |
| 108 | |
| 109 | TsanFlags(const char *env) : ignore_noninstrumented_modules(0) { |
| 110 | if (env) { |
| 111 | std::vector<std::string> tokens; |
| 112 | std::string str(env); |
| 113 | auto end = str.end(); |
| 114 | auto it = str.begin(); |
| 115 | auto is_sep = [](char c) { |
| 116 | return c == ' ' || c == ',' || c == ':' || c == '\n' || c == '\t' || |
| 117 | c == '\r'; |
| 118 | }; |
| 119 | while (it != end) { |
| 120 | auto next_it = std::find_if(first: it, last: end, pred: is_sep); |
| 121 | tokens.emplace_back(args&: it, args&: next_it); |
| 122 | it = next_it; |
| 123 | if (it != end) { |
| 124 | ++it; |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | for (const auto &token : tokens) { |
| 129 | // we are interested in ignore_noninstrumented_modules to print a |
| 130 | // warning |
| 131 | if (sscanf(s: token.c_str(), format: "ignore_noninstrumented_modules=%d" , |
| 132 | &ignore_noninstrumented_modules)) |
| 133 | continue; |
| 134 | } |
| 135 | } |
| 136 | } |
| 137 | }; |
| 138 | } // namespace |
| 139 | |
| 140 | #if (LLVM_VERSION) >= 40 |
| 141 | extern "C" { |
| 142 | int __attribute__((weak)) __archer_get_omp_status(); |
| 143 | void __attribute__((weak)) __tsan_flush_memory() {} |
| 144 | } |
| 145 | #endif |
| 146 | static ArcherFlags *archer_flags; |
| 147 | |
| 148 | #ifndef TsanHappensBefore |
| 149 | |
| 150 | template <typename... Args> static void __ompt_tsan_func(Args...) {} |
| 151 | |
| 152 | #define DECLARE_TSAN_FUNCTION(name, ...) \ |
| 153 | static void (*name)(__VA_ARGS__) = __ompt_tsan_func<__VA_ARGS__>; |
| 154 | |
| 155 | // Thread Sanitizer is a tool that finds races in code. |
| 156 | // See http://code.google.com/p/data-race-test/wiki/DynamicAnnotations . |
| 157 | // tsan detects these exact functions by name. |
| 158 | extern "C" { |
| 159 | DECLARE_TSAN_FUNCTION(AnnotateHappensAfter, const char *, int, |
| 160 | const volatile void *) |
| 161 | DECLARE_TSAN_FUNCTION(AnnotateHappensBefore, const char *, int, |
| 162 | const volatile void *) |
| 163 | DECLARE_TSAN_FUNCTION(AnnotateIgnoreWritesBegin, const char *, int) |
| 164 | DECLARE_TSAN_FUNCTION(AnnotateIgnoreWritesEnd, const char *, int) |
| 165 | DECLARE_TSAN_FUNCTION(AnnotateNewMemory, const char *, int, |
| 166 | const volatile void *, size_t) |
| 167 | DECLARE_TSAN_FUNCTION(__tsan_func_entry, const void *) |
| 168 | DECLARE_TSAN_FUNCTION(__tsan_func_exit) |
| 169 | } |
| 170 | |
| 171 | // This marker is used to define a happens-before arc. The race detector will |
| 172 | // infer an arc from the begin to the end when they share the same pointer |
| 173 | // argument. |
| 174 | #define TsanHappensBefore(cv) AnnotateHappensBefore(__FILE__, __LINE__, cv) |
| 175 | |
| 176 | // This marker defines the destination of a happens-before arc. |
| 177 | #define TsanHappensAfter(cv) AnnotateHappensAfter(__FILE__, __LINE__, cv) |
| 178 | |
| 179 | // Ignore any races on writes between here and the next TsanIgnoreWritesEnd. |
| 180 | #define TsanIgnoreWritesBegin() AnnotateIgnoreWritesBegin(__FILE__, __LINE__) |
| 181 | |
| 182 | // Resume checking for racy writes. |
| 183 | #define TsanIgnoreWritesEnd() AnnotateIgnoreWritesEnd(__FILE__, __LINE__) |
| 184 | |
| 185 | // We don't really delete the clock for now |
| 186 | #define TsanDeleteClock(cv) |
| 187 | |
| 188 | // newMemory |
| 189 | #define TsanNewMemory(addr, size) \ |
| 190 | AnnotateNewMemory(__FILE__, __LINE__, addr, size) |
| 191 | #define TsanFreeMemory(addr, size) \ |
| 192 | AnnotateNewMemory(__FILE__, __LINE__, addr, size) |
| 193 | #endif |
| 194 | |
| 195 | // Function entry/exit |
| 196 | #define TsanFuncEntry(pc) __tsan_func_entry(pc) |
| 197 | #define TsanFuncExit() __tsan_func_exit() |
| 198 | |
| 199 | /// Required OMPT inquiry functions. |
| 200 | static ompt_get_parallel_info_t ompt_get_parallel_info; |
| 201 | static ompt_get_thread_data_t ompt_get_thread_data; |
| 202 | |
| 203 | typedef char ompt_tsan_clockid; |
| 204 | |
| 205 | static uint64_t my_next_id() { |
| 206 | static uint64_t ID = 0; |
| 207 | uint64_t ret = __sync_fetch_and_add(&ID, 1); |
| 208 | return ret; |
| 209 | } |
| 210 | |
| 211 | static int pagesize{0}; |
| 212 | |
| 213 | // Data structure to provide a threadsafe pool of reusable objects. |
| 214 | // DataPool<Type of objects> |
| 215 | namespace { |
| 216 | template <typename T> struct DataPool final { |
| 217 | static __thread DataPool<T> *ThreadDataPool; |
| 218 | std::mutex DPMutex{}; |
| 219 | |
| 220 | // store unused objects |
| 221 | std::vector<T *> DataPointer{}; |
| 222 | std::vector<T *> RemoteDataPointer{}; |
| 223 | |
| 224 | // store all allocated memory to finally release |
| 225 | std::list<void *> memory; |
| 226 | |
| 227 | // count remotely returned data (RemoteDataPointer.size()) |
| 228 | std::atomic<int> remote{0}; |
| 229 | |
| 230 | // totally allocated data objects in pool |
| 231 | int total{0}; |
| 232 | #ifdef DEBUG_DATA |
| 233 | int remoteReturn{0}; |
| 234 | int localReturn{0}; |
| 235 | |
| 236 | int getRemote() { return remoteReturn + remote; } |
| 237 | int getLocal() { return localReturn; } |
| 238 | #endif |
| 239 | int getTotal() { return total; } |
| 240 | int getMissing() { |
| 241 | return total - DataPointer.size() - RemoteDataPointer.size(); |
| 242 | } |
| 243 | |
| 244 | // fill the pool by allocating a page of memory |
| 245 | void newDatas() { |
| 246 | if (remote > 0) { |
| 247 | const std::lock_guard<std::mutex> lock(DPMutex); |
| 248 | // DataPointer is empty, so just swap the vectors |
| 249 | DataPointer.swap(RemoteDataPointer); |
| 250 | remote = 0; |
| 251 | return; |
| 252 | } |
| 253 | // calculate size of an object including padding to cacheline size |
| 254 | size_t elemSize = sizeof(T); |
| 255 | size_t paddedSize = (((elemSize - 1) / 64) + 1) * 64; |
| 256 | // number of padded elements to allocate |
| 257 | int ndatas = pagesize / paddedSize; |
| 258 | char *datas = (char *)malloc(size: ndatas * paddedSize); |
| 259 | memory.push_back(x: datas); |
| 260 | for (int i = 0; i < ndatas; i++) { |
| 261 | DataPointer.push_back(new (datas + i * paddedSize) T(this)); |
| 262 | } |
| 263 | total += ndatas; |
| 264 | } |
| 265 | |
| 266 | // get data from the pool |
| 267 | T *getData() { |
| 268 | T *ret; |
| 269 | if (DataPointer.empty()) |
| 270 | newDatas(); |
| 271 | ret = DataPointer.back(); |
| 272 | DataPointer.pop_back(); |
| 273 | return ret; |
| 274 | } |
| 275 | |
| 276 | // accesses to the thread-local datapool don't need locks |
| 277 | void returnOwnData(T *data) { |
| 278 | DataPointer.emplace_back(data); |
| 279 | #ifdef DEBUG_DATA |
| 280 | localReturn++; |
| 281 | #endif |
| 282 | } |
| 283 | |
| 284 | // returning to a remote datapool using lock |
| 285 | void returnData(T *data) { |
| 286 | const std::lock_guard<std::mutex> lock(DPMutex); |
| 287 | RemoteDataPointer.emplace_back(data); |
| 288 | remote++; |
| 289 | #ifdef DEBUG_DATA |
| 290 | remoteReturn++; |
| 291 | #endif |
| 292 | } |
| 293 | |
| 294 | ~DataPool() { |
| 295 | // we assume all memory is returned when the thread finished / destructor is |
| 296 | // called |
| 297 | if (archer_flags->report_data_leak && getMissing() != 0) { |
| 298 | printf("ERROR: While freeing DataPool (%s) we are missing %i data " |
| 299 | "objects.\n" , |
| 300 | __PRETTY_FUNCTION__, getMissing()); |
| 301 | exit(status: -3); |
| 302 | } |
| 303 | for (auto i : DataPointer) |
| 304 | if (i) |
| 305 | i->~T(); |
| 306 | for (auto i : RemoteDataPointer) |
| 307 | if (i) |
| 308 | i->~T(); |
| 309 | for (auto i : memory) |
| 310 | if (i) |
| 311 | free(ptr: i); |
| 312 | } |
| 313 | }; |
| 314 | |
| 315 | template <typename T> struct DataPoolEntry { |
| 316 | DataPool<T> *owner; |
| 317 | |
| 318 | static T *New() { return DataPool<T>::ThreadDataPool->getData(); } |
| 319 | |
| 320 | void Delete() { |
| 321 | static_cast<T *>(this)->Reset(); |
| 322 | if (owner == DataPool<T>::ThreadDataPool) |
| 323 | owner->returnOwnData(static_cast<T *>(this)); |
| 324 | else |
| 325 | owner->returnData(static_cast<T *>(this)); |
| 326 | } |
| 327 | |
| 328 | DataPoolEntry(DataPool<T> *dp) : owner(dp) {} |
| 329 | }; |
| 330 | |
| 331 | struct DependencyData; |
| 332 | typedef DataPool<DependencyData> DependencyDataPool; |
| 333 | template <> |
| 334 | __thread DependencyDataPool *DependencyDataPool::ThreadDataPool = nullptr; |
| 335 | |
| 336 | /// Data structure to store additional information for task dependency. |
| 337 | struct DependencyData final : DataPoolEntry<DependencyData> { |
| 338 | ompt_tsan_clockid in; |
| 339 | ompt_tsan_clockid out; |
| 340 | ompt_tsan_clockid inoutset; |
| 341 | void *GetInPtr() { return ∈ } |
| 342 | void *GetOutPtr() { return &out; } |
| 343 | void *GetInoutsetPtr() { return &inoutset; } |
| 344 | |
| 345 | void Reset() {} |
| 346 | |
| 347 | static DependencyData *New() { return DataPoolEntry<DependencyData>::New(); } |
| 348 | |
| 349 | DependencyData(DataPool<DependencyData> *dp) |
| 350 | : DataPoolEntry<DependencyData>(dp) {} |
| 351 | }; |
| 352 | |
| 353 | struct TaskDependency { |
| 354 | void *inPtr; |
| 355 | void *outPtr; |
| 356 | void *inoutsetPtr; |
| 357 | ompt_dependence_type_t type; |
| 358 | TaskDependency(DependencyData *depData, ompt_dependence_type_t type) |
| 359 | : inPtr(depData->GetInPtr()), outPtr(depData->GetOutPtr()), |
| 360 | inoutsetPtr(depData->GetInoutsetPtr()), type(type) {} |
| 361 | void AnnotateBegin() { |
| 362 | if (type == ompt_dependence_type_out || |
| 363 | type == ompt_dependence_type_inout || |
| 364 | type == ompt_dependence_type_mutexinoutset) { |
| 365 | TsanHappensAfter(inPtr); |
| 366 | TsanHappensAfter(outPtr); |
| 367 | TsanHappensAfter(inoutsetPtr); |
| 368 | } else if (type == ompt_dependence_type_in) { |
| 369 | TsanHappensAfter(outPtr); |
| 370 | TsanHappensAfter(inoutsetPtr); |
| 371 | } else if (type == ompt_dependence_type_inoutset) { |
| 372 | TsanHappensAfter(inPtr); |
| 373 | TsanHappensAfter(outPtr); |
| 374 | } |
| 375 | } |
| 376 | void AnnotateEnd() { |
| 377 | if (type == ompt_dependence_type_out || |
| 378 | type == ompt_dependence_type_inout || |
| 379 | type == ompt_dependence_type_mutexinoutset) { |
| 380 | TsanHappensBefore(outPtr); |
| 381 | } else if (type == ompt_dependence_type_in) { |
| 382 | TsanHappensBefore(inPtr); |
| 383 | } else if (type == ompt_dependence_type_inoutset) { |
| 384 | TsanHappensBefore(inoutsetPtr); |
| 385 | } |
| 386 | } |
| 387 | }; |
| 388 | |
| 389 | struct ParallelData; |
| 390 | typedef DataPool<ParallelData> ParallelDataPool; |
| 391 | template <> |
| 392 | __thread ParallelDataPool *ParallelDataPool::ThreadDataPool = nullptr; |
| 393 | |
| 394 | /// Data structure to store additional information for parallel regions. |
| 395 | struct ParallelData final : DataPoolEntry<ParallelData> { |
| 396 | |
| 397 | // Parallel fork is just another barrier, use Barrier[1] |
| 398 | |
| 399 | /// Two addresses for relationships with barriers. |
| 400 | ompt_tsan_clockid Barrier[2]; |
| 401 | |
| 402 | const void *codePtr; |
| 403 | |
| 404 | void *GetParallelPtr() { return &(Barrier[1]); } |
| 405 | |
| 406 | void *GetBarrierPtr(unsigned Index) { return &(Barrier[Index]); } |
| 407 | |
| 408 | ParallelData *Init(const void *codeptr) { |
| 409 | codePtr = codeptr; |
| 410 | return this; |
| 411 | } |
| 412 | |
| 413 | void Reset() {} |
| 414 | |
| 415 | static ParallelData *New(const void *codeptr) { |
| 416 | return DataPoolEntry<ParallelData>::New()->Init(codeptr); |
| 417 | } |
| 418 | |
| 419 | ParallelData(DataPool<ParallelData> *dp) : DataPoolEntry<ParallelData>(dp) {} |
| 420 | }; |
| 421 | |
| 422 | static inline ParallelData *ToParallelData(ompt_data_t *parallel_data) { |
| 423 | return reinterpret_cast<ParallelData *>(parallel_data->ptr); |
| 424 | } |
| 425 | |
| 426 | struct Taskgroup; |
| 427 | typedef DataPool<Taskgroup> TaskgroupPool; |
| 428 | template <> __thread TaskgroupPool *TaskgroupPool::ThreadDataPool = nullptr; |
| 429 | |
| 430 | /// Data structure to support stacking of taskgroups and allow synchronization. |
| 431 | struct Taskgroup final : DataPoolEntry<Taskgroup> { |
| 432 | /// Its address is used for relationships of the taskgroup's task set. |
| 433 | ompt_tsan_clockid Ptr; |
| 434 | |
| 435 | /// Reference to the parent taskgroup. |
| 436 | Taskgroup *Parent; |
| 437 | |
| 438 | void *GetPtr() { return &Ptr; } |
| 439 | |
| 440 | Taskgroup *Init(Taskgroup *parent) { |
| 441 | Parent = parent; |
| 442 | return this; |
| 443 | } |
| 444 | |
| 445 | void Reset() {} |
| 446 | |
| 447 | static Taskgroup *New(Taskgroup *Parent) { |
| 448 | return DataPoolEntry<Taskgroup>::New()->Init(parent: Parent); |
| 449 | } |
| 450 | |
| 451 | Taskgroup(DataPool<Taskgroup> *dp) : DataPoolEntry<Taskgroup>(dp) {} |
| 452 | }; |
| 453 | |
| 454 | enum ArcherTaskFlag { ArcherTaskFulfilled = 0x00010000 }; |
| 455 | |
| 456 | struct TaskData; |
| 457 | typedef DataPool<TaskData> TaskDataPool; |
| 458 | template <> __thread TaskDataPool *TaskDataPool::ThreadDataPool = nullptr; |
| 459 | |
| 460 | /// Data structure to store additional information for tasks. |
| 461 | struct TaskData final : DataPoolEntry<TaskData> { |
| 462 | /// Its address is used for relationships of this task. |
| 463 | ompt_tsan_clockid Task{0}; |
| 464 | |
| 465 | /// Child tasks use its address to declare a relationship to a taskwait in |
| 466 | /// this task. |
| 467 | ompt_tsan_clockid Taskwait{0}; |
| 468 | |
| 469 | /// Child tasks use its address to model omp_all_memory dependencies |
| 470 | ompt_tsan_clockid AllMemory[2]{0}; |
| 471 | |
| 472 | /// Index of which barrier to use next. |
| 473 | char BarrierIndex{0}; |
| 474 | |
| 475 | /// Whether this task is currently executing a barrier. |
| 476 | bool InBarrier{false}; |
| 477 | |
| 478 | /// Whether this task is an included task. |
| 479 | int TaskType{0}; |
| 480 | |
| 481 | /// count execution phase |
| 482 | int execution{0}; |
| 483 | |
| 484 | /// Count how often this structure has been put into child tasks + 1. |
| 485 | std::atomic_int RefCount{1}; |
| 486 | |
| 487 | /// Reference to the parent that created this task. |
| 488 | TaskData *Parent{nullptr}; |
| 489 | |
| 490 | /// Reference to the team of this task. |
| 491 | ParallelData *Team{nullptr}; |
| 492 | |
| 493 | /// Reference to the current taskgroup that this task either belongs to or |
| 494 | /// that it just created. |
| 495 | Taskgroup *TaskGroup{nullptr}; |
| 496 | |
| 497 | /// Dependency information for this task. |
| 498 | TaskDependency *Dependencies{nullptr}; |
| 499 | |
| 500 | /// Number of dependency entries. |
| 501 | unsigned DependencyCount{0}; |
| 502 | |
| 503 | // The dependency-map stores DependencyData objects representing |
| 504 | // the dependency variables used on the sibling tasks created from |
| 505 | // this task |
| 506 | // We expect a rare need for the dependency-map, so alloc on demand |
| 507 | std::unordered_map<void *, DependencyData *> *DependencyMap{nullptr}; |
| 508 | |
| 509 | #ifdef DEBUG |
| 510 | int freed{0}; |
| 511 | #endif |
| 512 | |
| 513 | bool isIncluded() { return TaskType & ompt_task_undeferred; } |
| 514 | bool isUntied() { return TaskType & ompt_task_untied; } |
| 515 | bool isFinal() { return TaskType & ompt_task_final; } |
| 516 | bool isMergable() { return TaskType & ompt_task_mergeable; } |
| 517 | bool isMerged() { return TaskType & ompt_task_merged; } |
| 518 | |
| 519 | bool isExplicit() { return TaskType & ompt_task_explicit; } |
| 520 | bool isImplicit() { return TaskType & ompt_task_implicit; } |
| 521 | bool isInitial() { return TaskType & ompt_task_initial; } |
| 522 | bool isTarget() { return TaskType & ompt_task_target; } |
| 523 | |
| 524 | bool isFulfilled() { return TaskType & ArcherTaskFulfilled; } |
| 525 | void setFulfilled() { TaskType |= ArcherTaskFulfilled; } |
| 526 | |
| 527 | void setAllMemoryDep() { AllMemory[0] = 1; } |
| 528 | bool hasAllMemoryDep() { return AllMemory[0]; } |
| 529 | |
| 530 | void *GetTaskPtr() { return &Task; } |
| 531 | |
| 532 | void *GetTaskwaitPtr() { return &Taskwait; } |
| 533 | |
| 534 | void *GetLastAllMemoryPtr() { return AllMemory; } |
| 535 | void *GetNextAllMemoryPtr() { return AllMemory + 1; } |
| 536 | |
| 537 | TaskData *Init(TaskData *parent, int taskType) { |
| 538 | TaskType = taskType; |
| 539 | Parent = parent; |
| 540 | Team = Parent->Team; |
| 541 | BarrierIndex = Parent->BarrierIndex; |
| 542 | if (Parent != nullptr) { |
| 543 | Parent->RefCount++; |
| 544 | // Copy over pointer to taskgroup. This task may set up its own stack |
| 545 | // but for now belongs to its parent's taskgroup. |
| 546 | TaskGroup = Parent->TaskGroup; |
| 547 | } |
| 548 | return this; |
| 549 | } |
| 550 | |
| 551 | TaskData *Init(ParallelData *team, int taskType) { |
| 552 | TaskType = taskType; |
| 553 | execution = 1; |
| 554 | Team = team; |
| 555 | return this; |
| 556 | } |
| 557 | |
| 558 | void Reset() { |
| 559 | InBarrier = false; |
| 560 | TaskType = 0; |
| 561 | execution = 0; |
| 562 | BarrierIndex = 0; |
| 563 | RefCount = 1; |
| 564 | Parent = nullptr; |
| 565 | Team = nullptr; |
| 566 | TaskGroup = nullptr; |
| 567 | if (DependencyMap) { |
| 568 | for (auto i : *DependencyMap) |
| 569 | i.second->Delete(); |
| 570 | delete DependencyMap; |
| 571 | } |
| 572 | DependencyMap = nullptr; |
| 573 | if (Dependencies) |
| 574 | free(ptr: Dependencies); |
| 575 | Dependencies = nullptr; |
| 576 | DependencyCount = 0; |
| 577 | #ifdef DEBUG |
| 578 | freed = 0; |
| 579 | #endif |
| 580 | } |
| 581 | |
| 582 | static TaskData *New(TaskData *parent, int taskType) { |
| 583 | return DataPoolEntry<TaskData>::New()->Init(parent, taskType); |
| 584 | } |
| 585 | |
| 586 | static TaskData *New(ParallelData *team, int taskType) { |
| 587 | return DataPoolEntry<TaskData>::New()->Init(team, taskType); |
| 588 | } |
| 589 | |
| 590 | TaskData(DataPool<TaskData> *dp) : DataPoolEntry<TaskData>(dp) {} |
| 591 | }; |
| 592 | } // namespace |
| 593 | |
| 594 | static inline TaskData *ToTaskData(ompt_data_t *task_data) { |
| 595 | if (task_data) |
| 596 | return reinterpret_cast<TaskData *>(task_data->ptr); |
| 597 | return nullptr; |
| 598 | } |
| 599 | |
| 600 | /// Store a mutex for each wait_id to resolve race condition with callbacks. |
| 601 | static std::unordered_map<ompt_wait_id_t, std::mutex> Locks; |
| 602 | static std::mutex LocksMutex; |
| 603 | |
| 604 | static void ompt_tsan_thread_begin(ompt_thread_t thread_type, |
| 605 | ompt_data_t *thread_data) { |
| 606 | ParallelDataPool::ThreadDataPool = new ParallelDataPool; |
| 607 | TsanNewMemory(ParallelDataPool::ThreadDataPool, |
| 608 | sizeof(ParallelDataPool::ThreadDataPool)); |
| 609 | TaskgroupPool::ThreadDataPool = new TaskgroupPool; |
| 610 | TsanNewMemory(TaskgroupPool::ThreadDataPool, |
| 611 | sizeof(TaskgroupPool::ThreadDataPool)); |
| 612 | TaskDataPool::ThreadDataPool = new TaskDataPool; |
| 613 | TsanNewMemory(TaskDataPool::ThreadDataPool, |
| 614 | sizeof(TaskDataPool::ThreadDataPool)); |
| 615 | DependencyDataPool::ThreadDataPool = new DependencyDataPool; |
| 616 | TsanNewMemory(DependencyDataPool::ThreadDataPool, |
| 617 | sizeof(DependencyDataPool::ThreadDataPool)); |
| 618 | thread_data->value = my_next_id(); |
| 619 | } |
| 620 | |
| 621 | static void ompt_tsan_thread_end(ompt_data_t *thread_data) { |
| 622 | TsanIgnoreWritesBegin(); |
| 623 | delete ParallelDataPool::ThreadDataPool; |
| 624 | delete TaskgroupPool::ThreadDataPool; |
| 625 | delete TaskDataPool::ThreadDataPool; |
| 626 | delete DependencyDataPool::ThreadDataPool; |
| 627 | TsanIgnoreWritesEnd(); |
| 628 | } |
| 629 | |
| 630 | /// OMPT event callbacks for handling parallel regions. |
| 631 | |
| 632 | static void ompt_tsan_parallel_begin(ompt_data_t *parent_task_data, |
| 633 | const ompt_frame_t *parent_task_frame, |
| 634 | ompt_data_t *parallel_data, |
| 635 | uint32_t requested_team_size, int flag, |
| 636 | const void *codeptr_ra) { |
| 637 | ParallelData *Data = ParallelData::New(codeptr: codeptr_ra); |
| 638 | parallel_data->ptr = Data; |
| 639 | |
| 640 | TsanHappensBefore(Data->GetParallelPtr()); |
| 641 | if (archer_flags->ignore_serial && ToTaskData(task_data: parent_task_data)->isInitial()) |
| 642 | TsanIgnoreWritesEnd(); |
| 643 | } |
| 644 | |
| 645 | static void ompt_tsan_parallel_end(ompt_data_t *parallel_data, |
| 646 | ompt_data_t *task_data, int flag, |
| 647 | const void *codeptr_ra) { |
| 648 | if (archer_flags->ignore_serial && ToTaskData(task_data)->isInitial()) |
| 649 | TsanIgnoreWritesBegin(); |
| 650 | ParallelData *Data = ToParallelData(parallel_data); |
| 651 | TsanHappensAfter(Data->GetBarrierPtr(0)); |
| 652 | TsanHappensAfter(Data->GetBarrierPtr(1)); |
| 653 | |
| 654 | Data->Delete(); |
| 655 | |
| 656 | #if (LLVM_VERSION >= 40) |
| 657 | if (&__archer_get_omp_status) { |
| 658 | if (__archer_get_omp_status() == 0 && archer_flags->flush_shadow) |
| 659 | __tsan_flush_memory(); |
| 660 | } |
| 661 | #endif |
| 662 | } |
| 663 | |
| 664 | static void ompt_tsan_implicit_task(ompt_scope_endpoint_t endpoint, |
| 665 | ompt_data_t *parallel_data, |
| 666 | ompt_data_t *task_data, |
| 667 | unsigned int team_size, |
| 668 | unsigned int thread_num, int type) { |
| 669 | switch (endpoint) { |
| 670 | case ompt_scope_begin: |
| 671 | if (type & ompt_task_initial) { |
| 672 | parallel_data->ptr = ParallelData::New(codeptr: nullptr); |
| 673 | } |
| 674 | task_data->ptr = TaskData::New(team: ToParallelData(parallel_data), taskType: type); |
| 675 | TsanHappensAfter(ToParallelData(parallel_data)->GetParallelPtr()); |
| 676 | TsanFuncEntry(ToParallelData(parallel_data)->codePtr); |
| 677 | break; |
| 678 | case ompt_scope_end: { |
| 679 | TaskData *Data = ToTaskData(task_data); |
| 680 | #ifdef DEBUG |
| 681 | assert(Data->freed == 0 && "Implicit task end should only be called once!" ); |
| 682 | Data->freed = 1; |
| 683 | #endif |
| 684 | assert(Data->RefCount == 1 && |
| 685 | "All tasks should have finished at the implicit barrier!" ); |
| 686 | if (type & ompt_task_initial) { |
| 687 | Data->Team->Delete(); |
| 688 | } |
| 689 | Data->Delete(); |
| 690 | TsanFuncExit(); |
| 691 | break; |
| 692 | } |
| 693 | case ompt_scope_beginend: |
| 694 | // Should not occur according to OpenMP 5.1 |
| 695 | // Tested in OMPT tests |
| 696 | break; |
| 697 | } |
| 698 | } |
| 699 | |
| 700 | static void ompt_tsan_sync_region(ompt_sync_region_t kind, |
| 701 | ompt_scope_endpoint_t endpoint, |
| 702 | ompt_data_t *parallel_data, |
| 703 | ompt_data_t *task_data, |
| 704 | const void *codeptr_ra) { |
| 705 | TaskData *Data = ToTaskData(task_data); |
| 706 | switch (endpoint) { |
| 707 | case ompt_scope_begin: |
| 708 | case ompt_scope_beginend: |
| 709 | TsanFuncEntry(codeptr_ra); |
| 710 | switch (kind) { |
| 711 | case ompt_sync_region_barrier_implementation: |
| 712 | case ompt_sync_region_barrier_implicit: |
| 713 | case ompt_sync_region_barrier_explicit: |
| 714 | case ompt_sync_region_barrier_implicit_parallel: |
| 715 | case ompt_sync_region_barrier_implicit_workshare: |
| 716 | case ompt_sync_region_barrier_teams: |
| 717 | case ompt_sync_region_barrier: { |
| 718 | char BarrierIndex = Data->BarrierIndex; |
| 719 | TsanHappensBefore(Data->Team->GetBarrierPtr(BarrierIndex)); |
| 720 | |
| 721 | if (hasReductionCallback < ompt_set_always) { |
| 722 | // We ignore writes inside the barrier. These would either occur during |
| 723 | // 1. reductions performed by the runtime which are guaranteed to be |
| 724 | // race-free. |
| 725 | // 2. execution of another task. |
| 726 | // For the latter case we will re-enable tracking in task_switch. |
| 727 | Data->InBarrier = true; |
| 728 | TsanIgnoreWritesBegin(); |
| 729 | } |
| 730 | |
| 731 | break; |
| 732 | } |
| 733 | |
| 734 | case ompt_sync_region_taskwait: |
| 735 | break; |
| 736 | |
| 737 | case ompt_sync_region_taskgroup: |
| 738 | Data->TaskGroup = Taskgroup::New(Parent: Data->TaskGroup); |
| 739 | break; |
| 740 | |
| 741 | case ompt_sync_region_reduction: |
| 742 | // should never be reached |
| 743 | break; |
| 744 | } |
| 745 | if (endpoint == ompt_scope_begin) |
| 746 | break; |
| 747 | KMP_FALLTHROUGH(); |
| 748 | case ompt_scope_end: |
| 749 | TsanFuncExit(); |
| 750 | switch (kind) { |
| 751 | case ompt_sync_region_barrier_implementation: |
| 752 | case ompt_sync_region_barrier_implicit: |
| 753 | case ompt_sync_region_barrier_explicit: |
| 754 | case ompt_sync_region_barrier_implicit_parallel: |
| 755 | case ompt_sync_region_barrier_implicit_workshare: |
| 756 | case ompt_sync_region_barrier_teams: |
| 757 | case ompt_sync_region_barrier: { |
| 758 | if (hasReductionCallback < ompt_set_always) { |
| 759 | // We want to track writes after the barrier again. |
| 760 | Data->InBarrier = false; |
| 761 | TsanIgnoreWritesEnd(); |
| 762 | } |
| 763 | |
| 764 | char BarrierIndex = Data->BarrierIndex; |
| 765 | // Barrier will end after it has been entered by all threads. |
| 766 | if (parallel_data) |
| 767 | TsanHappensAfter(Data->Team->GetBarrierPtr(BarrierIndex)); |
| 768 | |
| 769 | // It is not guaranteed that all threads have exited this barrier before |
| 770 | // we enter the next one. So we will use a different address. |
| 771 | // We are however guaranteed that this current barrier is finished |
| 772 | // by the time we exit the next one. So we can then reuse the first |
| 773 | // address. |
| 774 | Data->BarrierIndex = (BarrierIndex + 1) % 2; |
| 775 | break; |
| 776 | } |
| 777 | |
| 778 | case ompt_sync_region_taskwait: { |
| 779 | if (Data->execution > 1) |
| 780 | TsanHappensAfter(Data->GetTaskwaitPtr()); |
| 781 | break; |
| 782 | } |
| 783 | |
| 784 | case ompt_sync_region_taskgroup: { |
| 785 | assert(Data->TaskGroup != nullptr && |
| 786 | "Should have at least one taskgroup!" ); |
| 787 | |
| 788 | TsanHappensAfter(Data->TaskGroup->GetPtr()); |
| 789 | |
| 790 | // Delete this allocated taskgroup, all descendent task are finished by |
| 791 | // now. |
| 792 | Taskgroup *Parent = Data->TaskGroup->Parent; |
| 793 | Data->TaskGroup->Delete(); |
| 794 | Data->TaskGroup = Parent; |
| 795 | break; |
| 796 | } |
| 797 | |
| 798 | case ompt_sync_region_reduction: |
| 799 | // Should not occur according to OpenMP 5.1 |
| 800 | // Tested in OMPT tests |
| 801 | break; |
| 802 | } |
| 803 | break; |
| 804 | } |
| 805 | } |
| 806 | |
| 807 | static void ompt_tsan_reduction(ompt_sync_region_t kind, |
| 808 | ompt_scope_endpoint_t endpoint, |
| 809 | ompt_data_t *parallel_data, |
| 810 | ompt_data_t *task_data, |
| 811 | const void *codeptr_ra) { |
| 812 | switch (endpoint) { |
| 813 | case ompt_scope_begin: |
| 814 | switch (kind) { |
| 815 | case ompt_sync_region_reduction: |
| 816 | TsanIgnoreWritesBegin(); |
| 817 | break; |
| 818 | default: |
| 819 | break; |
| 820 | } |
| 821 | break; |
| 822 | case ompt_scope_end: |
| 823 | switch (kind) { |
| 824 | case ompt_sync_region_reduction: |
| 825 | TsanIgnoreWritesEnd(); |
| 826 | break; |
| 827 | default: |
| 828 | break; |
| 829 | } |
| 830 | break; |
| 831 | case ompt_scope_beginend: |
| 832 | // Should not occur according to OpenMP 5.1 |
| 833 | // Tested in OMPT tests |
| 834 | // Would have no implications for DR detection |
| 835 | break; |
| 836 | } |
| 837 | } |
| 838 | |
| 839 | /// OMPT event callbacks for handling tasks. |
| 840 | |
| 841 | static void ompt_tsan_task_create( |
| 842 | ompt_data_t *parent_task_data, /* id of parent task */ |
| 843 | const ompt_frame_t *parent_frame, /* frame data for parent task */ |
| 844 | ompt_data_t *new_task_data, /* id of created task */ |
| 845 | int type, int has_dependences, |
| 846 | const void *codeptr_ra) /* pointer to outlined function */ |
| 847 | { |
| 848 | TaskData *Data; |
| 849 | assert(new_task_data->ptr == NULL && |
| 850 | "Task data should be initialized to NULL" ); |
| 851 | if (type & ompt_task_initial) { |
| 852 | ompt_data_t *parallel_data; |
| 853 | int team_size = 1; |
| 854 | ompt_get_parallel_info(0, ¶llel_data, &team_size); |
| 855 | ParallelData *PData = ParallelData::New(codeptr: nullptr); |
| 856 | parallel_data->ptr = PData; |
| 857 | |
| 858 | Data = TaskData::New(team: PData, taskType: type); |
| 859 | new_task_data->ptr = Data; |
| 860 | } else if (type & ompt_task_undeferred) { |
| 861 | Data = TaskData::New(parent: ToTaskData(task_data: parent_task_data), taskType: type); |
| 862 | new_task_data->ptr = Data; |
| 863 | } else if (type & ompt_task_explicit || type & ompt_task_target) { |
| 864 | Data = TaskData::New(parent: ToTaskData(task_data: parent_task_data), taskType: type); |
| 865 | new_task_data->ptr = Data; |
| 866 | |
| 867 | // Use the newly created address. We cannot use a single address from the |
| 868 | // parent because that would declare wrong relationships with other |
| 869 | // sibling tasks that may be created before this task is started! |
| 870 | TsanHappensBefore(Data->GetTaskPtr()); |
| 871 | ToTaskData(task_data: parent_task_data)->execution++; |
| 872 | } |
| 873 | } |
| 874 | |
| 875 | static void freeTask(TaskData *task) { |
| 876 | while (task != nullptr && --task->RefCount == 0) { |
| 877 | TaskData *Parent = task->Parent; |
| 878 | task->Delete(); |
| 879 | task = Parent; |
| 880 | } |
| 881 | } |
| 882 | |
| 883 | // LastAllMemoryPtr marks the beginning of an all_memory epoch |
| 884 | // NextAllMemoryPtr marks the end of an all_memory epoch |
| 885 | // All tasks with depend begin execution after LastAllMemoryPtr |
| 886 | // and end before NextAllMemoryPtr |
| 887 | static void releaseDependencies(TaskData *task) { |
| 888 | if (archer_flags->all_memory) { |
| 889 | if (task->hasAllMemoryDep()) { |
| 890 | TsanHappensBefore(task->Parent->GetLastAllMemoryPtr()); |
| 891 | TsanHappensBefore(task->Parent->GetNextAllMemoryPtr()); |
| 892 | } else if (task->DependencyCount) |
| 893 | TsanHappensBefore(task->Parent->GetNextAllMemoryPtr()); |
| 894 | } |
| 895 | for (unsigned i = 0; i < task->DependencyCount; i++) { |
| 896 | task->Dependencies[i].AnnotateEnd(); |
| 897 | } |
| 898 | } |
| 899 | |
| 900 | static void acquireDependencies(TaskData *task) { |
| 901 | if (archer_flags->all_memory) { |
| 902 | if (task->hasAllMemoryDep()) |
| 903 | TsanHappensAfter(task->Parent->GetNextAllMemoryPtr()); |
| 904 | else if (task->DependencyCount) |
| 905 | TsanHappensAfter(task->Parent->GetLastAllMemoryPtr()); |
| 906 | } |
| 907 | for (unsigned i = 0; i < task->DependencyCount; i++) { |
| 908 | task->Dependencies[i].AnnotateBegin(); |
| 909 | } |
| 910 | } |
| 911 | |
| 912 | static void completeTask(TaskData *FromTask) { |
| 913 | if (!FromTask) |
| 914 | return; |
| 915 | // Task-end happens after a possible omp_fulfill_event call |
| 916 | if (FromTask->isFulfilled()) |
| 917 | TsanHappensAfter(FromTask->GetTaskPtr()); |
| 918 | // Included tasks are executed sequentially, no need to track |
| 919 | // synchronization |
| 920 | if (!FromTask->isIncluded()) { |
| 921 | // Task will finish before a barrier in the surrounding parallel region |
| 922 | // ... |
| 923 | ParallelData *PData = FromTask->Team; |
| 924 | TsanHappensBefore(PData->GetBarrierPtr(FromTask->BarrierIndex)); |
| 925 | |
| 926 | // ... and before an eventual taskwait by the parent thread. |
| 927 | TsanHappensBefore(FromTask->Parent->GetTaskwaitPtr()); |
| 928 | |
| 929 | if (FromTask->TaskGroup != nullptr) { |
| 930 | // This task is part of a taskgroup, so it will finish before the |
| 931 | // corresponding taskgroup_end. |
| 932 | TsanHappensBefore(FromTask->TaskGroup->GetPtr()); |
| 933 | } |
| 934 | } |
| 935 | // release dependencies |
| 936 | releaseDependencies(task: FromTask); |
| 937 | } |
| 938 | |
| 939 | static void suspendTask(TaskData *FromTask) { |
| 940 | if (!FromTask) |
| 941 | return; |
| 942 | // Task may be resumed at a later point in time. |
| 943 | TsanHappensBefore(FromTask->GetTaskPtr()); |
| 944 | } |
| 945 | |
| 946 | static void switchTasks(TaskData *FromTask, TaskData *ToTask) { |
| 947 | // Legacy handling for missing reduction callback |
| 948 | if (hasReductionCallback < ompt_set_always) { |
| 949 | if (FromTask && FromTask->InBarrier) { |
| 950 | // We want to ignore writes in the runtime code during barriers, |
| 951 | // but not when executing tasks with user code! |
| 952 | TsanIgnoreWritesEnd(); |
| 953 | } |
| 954 | if (ToTask && ToTask->InBarrier) { |
| 955 | // We want to ignore writes in the runtime code during barriers, |
| 956 | // but not when executing tasks with user code! |
| 957 | TsanIgnoreWritesBegin(); |
| 958 | } |
| 959 | } |
| 960 | //// Not yet used |
| 961 | // if (FromTask) |
| 962 | // FromTask->deactivate(); |
| 963 | // if (ToTask) |
| 964 | // ToTask->activate(); |
| 965 | } |
| 966 | |
| 967 | static void endTask(TaskData *FromTask) { |
| 968 | if (!FromTask) |
| 969 | return; |
| 970 | } |
| 971 | |
| 972 | static void startTask(TaskData *ToTask) { |
| 973 | if (!ToTask) |
| 974 | return; |
| 975 | // Handle dependencies on first execution of the task |
| 976 | if (ToTask->execution == 0) { |
| 977 | ToTask->execution++; |
| 978 | acquireDependencies(task: ToTask); |
| 979 | } |
| 980 | // 1. Task will begin execution after it has been created. |
| 981 | // 2. Task will resume after it has been switched away. |
| 982 | TsanHappensAfter(ToTask->GetTaskPtr()); |
| 983 | } |
| 984 | |
| 985 | static void ompt_tsan_task_schedule(ompt_data_t *first_task_data, |
| 986 | ompt_task_status_t prior_task_status, |
| 987 | ompt_data_t *second_task_data) { |
| 988 | |
| 989 | // |
| 990 | // The necessary action depends on prior_task_status: |
| 991 | // |
| 992 | // ompt_task_early_fulfill = 5, |
| 993 | // -> ignored |
| 994 | // |
| 995 | // ompt_task_late_fulfill = 6, |
| 996 | // -> first completed, first freed, second ignored |
| 997 | // |
| 998 | // ompt_task_complete = 1, |
| 999 | // ompt_task_cancel = 3, |
| 1000 | // -> first completed, first freed, second starts |
| 1001 | // |
| 1002 | // ompt_taskwait_complete = 8, |
| 1003 | // -> first starts, first completes, first freed, second ignored |
| 1004 | // |
| 1005 | // ompt_task_detach = 4, |
| 1006 | // ompt_task_yield = 2, |
| 1007 | // ompt_task_switch = 7 |
| 1008 | // -> first suspended, second starts |
| 1009 | // |
| 1010 | |
| 1011 | TaskData *FromTask = ToTaskData(task_data: first_task_data); |
| 1012 | TaskData *ToTask = ToTaskData(task_data: second_task_data); |
| 1013 | |
| 1014 | switch (prior_task_status) { |
| 1015 | case ompt_task_early_fulfill: |
| 1016 | TsanHappensBefore(FromTask->GetTaskPtr()); |
| 1017 | FromTask->setFulfilled(); |
| 1018 | return; |
| 1019 | case ompt_task_late_fulfill: |
| 1020 | TsanHappensAfter(FromTask->GetTaskPtr()); |
| 1021 | completeTask(FromTask); |
| 1022 | freeTask(task: FromTask); |
| 1023 | return; |
| 1024 | case ompt_taskwait_complete: |
| 1025 | acquireDependencies(task: FromTask); |
| 1026 | freeTask(task: FromTask); |
| 1027 | return; |
| 1028 | case ompt_task_complete: |
| 1029 | completeTask(FromTask); |
| 1030 | endTask(FromTask); |
| 1031 | switchTasks(FromTask, ToTask); |
| 1032 | freeTask(task: FromTask); |
| 1033 | return; |
| 1034 | case ompt_task_cancel: |
| 1035 | completeTask(FromTask); |
| 1036 | endTask(FromTask); |
| 1037 | switchTasks(FromTask, ToTask); |
| 1038 | freeTask(task: FromTask); |
| 1039 | startTask(ToTask); |
| 1040 | return; |
| 1041 | case ompt_task_detach: |
| 1042 | endTask(FromTask); |
| 1043 | suspendTask(FromTask); |
| 1044 | switchTasks(FromTask, ToTask); |
| 1045 | startTask(ToTask); |
| 1046 | return; |
| 1047 | case ompt_task_yield: |
| 1048 | suspendTask(FromTask); |
| 1049 | switchTasks(FromTask, ToTask); |
| 1050 | startTask(ToTask); |
| 1051 | return; |
| 1052 | case ompt_task_switch: |
| 1053 | suspendTask(FromTask); |
| 1054 | switchTasks(FromTask, ToTask); |
| 1055 | startTask(ToTask); |
| 1056 | return; |
| 1057 | } |
| 1058 | } |
| 1059 | |
| 1060 | static void ompt_tsan_dependences(ompt_data_t *task_data, |
| 1061 | const ompt_dependence_t *deps, int ndeps) { |
| 1062 | if (ndeps > 0) { |
| 1063 | // Copy the data to use it in task_switch and task_end. |
| 1064 | TaskData *Data = ToTaskData(task_data); |
| 1065 | if (!Data->Parent) { |
| 1066 | // Return since doacross dependences are not supported yet. |
| 1067 | return; |
| 1068 | } |
| 1069 | if (!Data->Parent->DependencyMap) |
| 1070 | Data->Parent->DependencyMap = |
| 1071 | new std::unordered_map<void *, DependencyData *>(); |
| 1072 | Data->Dependencies = |
| 1073 | (TaskDependency *)malloc(size: sizeof(TaskDependency) * ndeps); |
| 1074 | Data->DependencyCount = ndeps; |
| 1075 | for (int i = 0, d = 0; i < ndeps; i++, d++) { |
| 1076 | if (deps[i].dependence_type == ompt_dependence_type_out_all_memory || |
| 1077 | deps[i].dependence_type == ompt_dependence_type_inout_all_memory) { |
| 1078 | Data->setAllMemoryDep(); |
| 1079 | Data->DependencyCount--; |
| 1080 | if (!archer_flags->all_memory) { |
| 1081 | printf(format: "The application uses omp_all_memory, but Archer was\n" |
| 1082 | "started to not consider omp_all_memory. This can lead\n" |
| 1083 | "to false data race alerts.\n" |
| 1084 | "Include all_memory=1 in ARCHER_OPTIONS to consider\n" |
| 1085 | "omp_all_memory from the beginning.\n" ); |
| 1086 | archer_flags->all_memory = 1; |
| 1087 | } |
| 1088 | d--; |
| 1089 | continue; |
| 1090 | } |
| 1091 | auto ret = Data->Parent->DependencyMap->insert( |
| 1092 | x: std::make_pair(x: deps[i].variable.ptr, y: nullptr)); |
| 1093 | if (ret.second) { |
| 1094 | ret.first->second = DependencyData::New(); |
| 1095 | } |
| 1096 | new ((void *)(Data->Dependencies + d)) |
| 1097 | TaskDependency(ret.first->second, deps[i].dependence_type); |
| 1098 | } |
| 1099 | |
| 1100 | // This callback is executed before this task is first started. |
| 1101 | TsanHappensBefore(Data->GetTaskPtr()); |
| 1102 | } |
| 1103 | } |
| 1104 | |
| 1105 | /// OMPT event callbacks for handling locking. |
| 1106 | static void ompt_tsan_mutex_acquired(ompt_mutex_t kind, ompt_wait_id_t wait_id, |
| 1107 | const void *codeptr_ra) { |
| 1108 | |
| 1109 | // Acquire our own lock to make sure that |
| 1110 | // 1. the previous release has finished. |
| 1111 | // 2. the next acquire doesn't start before we have finished our release. |
| 1112 | LocksMutex.lock(); |
| 1113 | std::mutex &Lock = Locks[wait_id]; |
| 1114 | LocksMutex.unlock(); |
| 1115 | |
| 1116 | Lock.lock(); |
| 1117 | TsanHappensAfter(&Lock); |
| 1118 | } |
| 1119 | |
| 1120 | static void ompt_tsan_mutex_released(ompt_mutex_t kind, ompt_wait_id_t wait_id, |
| 1121 | const void *codeptr_ra) { |
| 1122 | LocksMutex.lock(); |
| 1123 | std::mutex &Lock = Locks[wait_id]; |
| 1124 | LocksMutex.unlock(); |
| 1125 | TsanHappensBefore(&Lock); |
| 1126 | |
| 1127 | Lock.unlock(); |
| 1128 | } |
| 1129 | |
| 1130 | // callback , signature , variable to store result , required support level |
| 1131 | #define SET_OPTIONAL_CALLBACK_T(event, type, result, level) \ |
| 1132 | do { \ |
| 1133 | ompt_callback_##type##_t tsan_##event = &ompt_tsan_##event; \ |
| 1134 | result = ompt_set_callback(ompt_callback_##event, \ |
| 1135 | (ompt_callback_t)tsan_##event); \ |
| 1136 | if (result < level) \ |
| 1137 | printf("Registered callback '" #event "' is not supported at " #level \ |
| 1138 | " (%i)\n", \ |
| 1139 | result); \ |
| 1140 | } while (0) |
| 1141 | |
| 1142 | #define SET_CALLBACK_T(event, type) \ |
| 1143 | do { \ |
| 1144 | int res; \ |
| 1145 | SET_OPTIONAL_CALLBACK_T(event, type, res, ompt_set_always); \ |
| 1146 | } while (0) |
| 1147 | |
| 1148 | #define SET_CALLBACK(event) SET_CALLBACK_T(event, event) |
| 1149 | |
| 1150 | #define findTsanFunction(f, fSig) \ |
| 1151 | do { \ |
| 1152 | void *fp = dlsym(RTLD_DEFAULT, #f); \ |
| 1153 | if (fp) \ |
| 1154 | f = fSig fp; \ |
| 1155 | else \ |
| 1156 | printf("Unable to find TSan function " #f ".\n"); \ |
| 1157 | } while (0) |
| 1158 | |
| 1159 | #define findTsanFunctionSilent(f, fSig) f = fSig dlsym(RTLD_DEFAULT, #f) |
| 1160 | |
| 1161 | static int ompt_tsan_initialize(ompt_function_lookup_t lookup, int device_num, |
| 1162 | ompt_data_t *tool_data) { |
| 1163 | const char *options = getenv(name: "TSAN_OPTIONS" ); |
| 1164 | TsanFlags tsan_flags(options); |
| 1165 | |
| 1166 | ompt_set_callback_t ompt_set_callback = |
| 1167 | (ompt_set_callback_t)lookup("ompt_set_callback" ); |
| 1168 | if (ompt_set_callback == NULL) { |
| 1169 | std::cerr << "Could not set callback, exiting..." << std::endl; |
| 1170 | std::exit(status: 1); |
| 1171 | } |
| 1172 | ompt_get_parallel_info = |
| 1173 | (ompt_get_parallel_info_t)lookup("ompt_get_parallel_info" ); |
| 1174 | ompt_get_thread_data = (ompt_get_thread_data_t)lookup("ompt_get_thread_data" ); |
| 1175 | |
| 1176 | if (ompt_get_parallel_info == NULL) { |
| 1177 | fprintf(stderr, format: "Could not get inquiry function 'ompt_get_parallel_info', " |
| 1178 | "exiting...\n" ); |
| 1179 | exit(status: 1); |
| 1180 | } |
| 1181 | |
| 1182 | findTsanFunction(AnnotateHappensAfter, |
| 1183 | (void (*)(const char *, int, const volatile void *))); |
| 1184 | findTsanFunction(AnnotateHappensBefore, |
| 1185 | (void (*)(const char *, int, const volatile void *))); |
| 1186 | findTsanFunction(AnnotateIgnoreWritesBegin, (void (*)(const char *, int))); |
| 1187 | findTsanFunction(AnnotateIgnoreWritesEnd, (void (*)(const char *, int))); |
| 1188 | findTsanFunction( |
| 1189 | AnnotateNewMemory, |
| 1190 | (void (*)(const char *, int, const volatile void *, size_t))); |
| 1191 | findTsanFunction(__tsan_func_entry, (void (*)(const void *))); |
| 1192 | findTsanFunction(__tsan_func_exit, (void (*)(void))); |
| 1193 | |
| 1194 | SET_CALLBACK(thread_begin); |
| 1195 | SET_CALLBACK(thread_end); |
| 1196 | SET_CALLBACK(parallel_begin); |
| 1197 | SET_CALLBACK(implicit_task); |
| 1198 | SET_CALLBACK(sync_region); |
| 1199 | SET_CALLBACK(parallel_end); |
| 1200 | |
| 1201 | SET_CALLBACK(task_create); |
| 1202 | SET_CALLBACK(task_schedule); |
| 1203 | SET_CALLBACK(dependences); |
| 1204 | |
| 1205 | SET_CALLBACK_T(mutex_acquired, mutex); |
| 1206 | SET_CALLBACK_T(mutex_released, mutex); |
| 1207 | SET_OPTIONAL_CALLBACK_T(reduction, sync_region, hasReductionCallback, |
| 1208 | ompt_set_never); |
| 1209 | |
| 1210 | if (!tsan_flags.ignore_noninstrumented_modules) |
| 1211 | fprintf(stderr, |
| 1212 | format: "Warning: please export " |
| 1213 | "TSAN_OPTIONS='ignore_noninstrumented_modules=1' " |
| 1214 | "to avoid false positive reports from the OpenMP runtime!\n" ); |
| 1215 | if (archer_flags->ignore_serial) |
| 1216 | TsanIgnoreWritesBegin(); |
| 1217 | |
| 1218 | return 1; // success |
| 1219 | } |
| 1220 | |
| 1221 | static void ompt_tsan_finalize(ompt_data_t *tool_data) { |
| 1222 | if (archer_flags->ignore_serial) |
| 1223 | TsanIgnoreWritesEnd(); |
| 1224 | if (archer_flags->print_max_rss) { |
| 1225 | struct rusage end; |
| 1226 | getrusage(RUSAGE_SELF, usage: &end); |
| 1227 | printf(format: "MAX RSS[KiB] during execution: %ld\n" , end.ru_maxrss); |
| 1228 | } |
| 1229 | |
| 1230 | if (archer_flags) |
| 1231 | delete archer_flags; |
| 1232 | } |
| 1233 | |
| 1234 | extern "C" ompt_start_tool_result_t * |
| 1235 | ompt_start_tool(unsigned int omp_version, const char *runtime_version) { |
| 1236 | const char *options = getenv(name: "ARCHER_OPTIONS" ); |
| 1237 | archer_flags = new ArcherFlags(options); |
| 1238 | if (!archer_flags->enabled) { |
| 1239 | if (archer_flags->verbose) |
| 1240 | std::cout << "Archer disabled, stopping operation" << std::endl; |
| 1241 | delete archer_flags; |
| 1242 | return NULL; |
| 1243 | } |
| 1244 | |
| 1245 | pagesize = getpagesize(); |
| 1246 | |
| 1247 | static ompt_start_tool_result_t ompt_start_tool_result = { |
| 1248 | .initialize: &ompt_tsan_initialize, .finalize: &ompt_tsan_finalize, .tool_data: {.value: 0}}; |
| 1249 | |
| 1250 | // The OMPT start-up code uses dlopen with RTLD_LAZY. Therefore, we cannot |
| 1251 | // rely on dlopen to fail if TSan is missing, but would get a runtime error |
| 1252 | // for the first TSan call. We use __tsan_init to detect whether |
| 1253 | // an implementation of the Annotation interface is available in the |
| 1254 | // execution or disable the tool (by returning NULL). |
| 1255 | |
| 1256 | void (*__tsan_init)(void) = nullptr; |
| 1257 | |
| 1258 | findTsanFunctionSilent(__tsan_init, (void (*)(void))); |
| 1259 | if (!__tsan_init) // if we are not running on TSAN, give a different |
| 1260 | // tool the chance to be loaded |
| 1261 | { |
| 1262 | if (archer_flags->verbose) |
| 1263 | std::cout << "Archer detected OpenMP application without TSan; " |
| 1264 | "stopping operation" |
| 1265 | << std::endl; |
| 1266 | delete archer_flags; |
| 1267 | return NULL; |
| 1268 | } |
| 1269 | |
| 1270 | if (archer_flags->verbose) |
| 1271 | std::cout << "Archer detected OpenMP application with TSan, supplying " |
| 1272 | "OpenMP synchronization semantics" |
| 1273 | << std::endl; |
| 1274 | return &ompt_start_tool_result; |
| 1275 | } |
| 1276 | |