1 /* 2 * Copyright (c) 2021-2022 Huawei Device Co., Ltd. 3 * Licensed under the Apache License, Version 2.0 (the "License"); 4 * you may not use this file except in compliance with the License. 5 * You may obtain a copy of the License at 6 * 7 * http://www.apache.org/licenses/LICENSE-2.0 8 * 9 * Unless required by applicable law or agreed to in writing, software 10 * distributed under the License is distributed on an "AS IS" BASIS, 11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 * See the License for the specific language governing permissions and 13 * limitations under the License. 14 */ 15 #ifndef HIPERF_PERF_EVENTS_H 16 #define HIPERF_PERF_EVENTS_H 17 18 #include <atomic> 19 #include <cassert> 20 #include <chrono> 21 #include <cinttypes> 22 #include <condition_variable> 23 #include <deque> 24 #include <map> 25 #include <memory> 26 #include <string> 27 #include <thread> 28 #include <variant> 29 #include <vector> 30 #if !is_mingw 31 #include <poll.h> 32 #endif 33 34 #include <sys/types.h> 35 #include <unique_fd.h> 36 #include <linux/perf_event.h> 37 38 #include "debug_logger.h" 39 #include "perf_event_record.h" 40 #include "ring_buffer.h" 41 #include "tracked_command.h" 42 #include "utilities.h" 43 #include "virtual_runtime.h" 44 45 // this for some performance debug 46 #define HIDEBUG_SKIP_CALLBACK 0 47 48 namespace OHOS { 49 namespace Developtools { 50 namespace HiPerf { 51 using ConfigTable = std::map<__u64, const std::string>; 52 using SharedConfigTable = std::unique_ptr<ConfigTable>; 53 54 static const std::string PERF_EVENT_PARANOID = "/proc/sys/kernel/perf_event_paranoid"; 55 static const std::string PERF_DISABLE_PARAM = "security.perf_harden"; 56 57 // define convert from linux/perf_event.h 58 // description from https://man7.org/linux/man-pages/man2/perf_event_open.2.html 59 60 static const ConfigTable PERF_HW_CONFIGS = { 61 {PERF_COUNT_HW_CPU_CYCLES, "hw-cpu-cycles"}, 62 {PERF_COUNT_HW_INSTRUCTIONS, "hw-instructions"}, 63 {PERF_COUNT_HW_CACHE_REFERENCES, "hw-cache-references"}, 64 {PERF_COUNT_HW_CACHE_MISSES, "hw-cache-misses"}, 65 {PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "hw-branch-instructions"}, 66 {PERF_COUNT_HW_BRANCH_MISSES, "hw-branch-misses"}, 67 {PERF_COUNT_HW_BUS_CYCLES, "hw-bus-cycles"}, 68 {PERF_COUNT_HW_STALLED_CYCLES_FRONTEND, "hw-stalled-cycles-frontend"}, 69 {PERF_COUNT_HW_STALLED_CYCLES_BACKEND, "hw-stalled-cycles-backend"}, 70 {PERF_COUNT_HW_REF_CPU_CYCLES, "hw-ref-cpu-cycles"}, 71 }; 72 static const ConfigTable PERF_HW_CACHE_CONFIGS = { 73 {PERF_COUNT_HW_CACHE_L1D, "hw-cache-l1d"}, {PERF_COUNT_HW_CACHE_L1I, "hw-cache-l1i"}, 74 {PERF_COUNT_HW_CACHE_LL, "hw-cache-ll"}, {PERF_COUNT_HW_CACHE_DTLB, "hw-cache-dtlb"}, 75 {PERF_COUNT_HW_CACHE_ITLB, "hw-cache-itlb"}, {PERF_COUNT_HW_CACHE_BPU, "hw-cache-bpu"}, 76 {PERF_COUNT_HW_CACHE_NODE, "hw-cache-node"}, 77 }; 78 static const ConfigTable PERF_HW_CACHE_OP_CONFIGS = { 79 {PERF_COUNT_HW_CACHE_OP_READ, "hw-cache-op-read"}, 80 {PERF_COUNT_HW_CACHE_OP_WRITE, "hw-cache-op-write"}, 81 {PERF_COUNT_HW_CACHE_OP_PREFETCH, "hw-cache-op-prefetch"}, 82 }; 83 static const ConfigTable PERF_HW_CACHE_OP_RESULT_CONFIGS = { 84 {PERF_COUNT_HW_CACHE_RESULT_ACCESS, "hw-cache-result-access"}, 85 {PERF_COUNT_HW_CACHE_RESULT_MISS, "hw-cache-result-miss"}, 86 }; 87 static const ConfigTable PERF_SW_CONFIGS = { 88 {PERF_COUNT_SW_CPU_CLOCK, "sw-cpu-clock"}, 89 {PERF_COUNT_SW_TASK_CLOCK, "sw-task-clock"}, 90 {PERF_COUNT_SW_PAGE_FAULTS, "sw-page-faults"}, 91 {PERF_COUNT_SW_CONTEXT_SWITCHES, "sw-context-switches"}, 92 {PERF_COUNT_SW_CPU_MIGRATIONS, "sw-cpu-migrations"}, 93 {PERF_COUNT_SW_PAGE_FAULTS_MIN, "sw-page-faults-min"}, 94 {PERF_COUNT_SW_PAGE_FAULTS_MAJ, "sw-page-faults-maj"}, 95 {PERF_COUNT_SW_ALIGNMENT_FAULTS, "sw-alignment-faults"}, 96 {PERF_COUNT_SW_EMULATION_FAULTS, "sw-emulation-faults"}, 97 {PERF_COUNT_SW_DUMMY, "sw-dummy"}, 98 {PERF_COUNT_SW_BPF_OUTPUT, "sw-bpf-output"}, 99 }; 100 static const ConfigTable PERF_RAW_CONFIGS = { 101 {0x0, "raw-sw-incr"}, 102 {0x1, "raw-l1-icache-refill"}, 103 {0x2, "raw-l1-itlb-refill"}, 104 {0x3, "raw-l1-dcache-refill"}, 105 {0x4, "raw-l1-dcache"}, 106 {0x5, "raw-l1-dtlb-refill"}, 107 {0x6, "raw-load-retired"}, 108 {0x7, "raw-store-retired"}, 109 {0x8, "raw-instruction-retired"}, 110 {0x9, "raw-exception-taken"}, 111 {0xa, "raw-exception-return"}, 112 {0xb, "raw-cid-write-retired"}, 113 {0xc, "raw-pc-write-retired"}, 114 {0xd, "raw-br-immed-retired"}, 115 {0xe, "raw-br-return-retired"}, 116 {0xf, "raw-unaligned-ldst-retired"}, 117 {0x10, "raw-br-mis-pred"}, 118 {0x11, "raw-cpu-cycles"}, 119 {0x12, "raw-br-pred"}, 120 {0x13, "raw-mem-access"}, 121 {0x14, "raw-l1-icache"}, 122 {0x15, "raw-l1-dcache-wb"}, 123 {0x16, "raw-l2-dcache"}, 124 {0x17, "raw-l2-dcache-refill"}, 125 {0x18, "raw-l2-dcache-wb"}, 126 {0x19, "raw-bus-access"}, 127 {0x1a, "raw-memory-error"}, 128 {0x1b, "raw-inst-spec"}, 129 {0x1c, "raw-ttbr-write-retired"}, 130 {0x1d, "raw-bus-cycles"}, 131 {0x1f, "raw-l1-dcache-allocate"}, 132 {0x20, "raw-l2-dcache-allocate"}, 133 {0x21, "raw-br-retired"}, 134 {0x22, "raw-br-mis-pred-retired"}, 135 {0x23, "raw-stall-frontend"}, 136 {0x24, "raw-stall-backend"}, 137 {0x25, "raw-l1-dtlb"}, 138 {0x26, "raw-l1-itlb"}, 139 {0x27, "raw-l2-icache"}, 140 {0x28, "raw-l2-icache-refill"}, 141 {0x29, "raw-l3-dcache-allocate"}, 142 {0x2a, "raw-l3-dcache-refill"}, 143 {0x2b, "raw-l3-dcache"}, 144 {0x2c, "raw-l3-dcache-wb"}, 145 {0x2d, "raw-l2-dtlb-refill"}, 146 {0x2e, "raw-l2-itlb-refill"}, 147 {0x2f, "raw-l2-dtlb"}, 148 {0x30, "raw-l2-itlb"}, 149 // newadd 150 {0x31, "raw-remote-access"}, 151 {0x32, "raw-ll-cache"}, 152 {0x33, "raw-ll-cache-miss"}, 153 {0x34, "raw-dtlb-walk"}, 154 {0x35, "raw-itlb-walk"}, 155 {0x36, "raw-ll-cache-rd"}, 156 {0x37, "raw-ll-cache-miss-rd"}, 157 {0x38, "raw-remote-access-rd"}, 158 {0x39, "raw-l1d-cache-lmiss-rd"}, 159 {0x3A, "raw-op-retired"}, 160 {0x3B, "raw-op-spec"}, 161 {0x3C, "raw-stall"}, 162 {0x3D, "raw-stall-slot-backend"}, 163 {0x3E, "raw-stall-slot-frontend"}, 164 {0x3F, "raw-stall-slot"}, 165 {0x40, "raw-l1d-cache-rd"}, 166 {0x4000, "raw-sample-pop"}, 167 {0x4001, "raw-sample-feed"}, 168 {0x4002, "raw-sample-filtrate"}, 169 {0x4003, "raw-sample-collision"}, 170 {0x4004, "raw-cnt-cycles"}, 171 {0x4005, "raw-stall-backend-mem"}, 172 {0x4006, "raw-l1i-cache-lmiss"}, 173 {0x4009, "raw-l2d-cache-lmiss-rd"}, 174 {0x400A, "raw-l2i-cache-lmiss"}, 175 {0x400B, "raw-l3d-cache-lmiss-rd"}, 176 {0x8002, "raw-sve-inst-retired"}, 177 {0x8006, "raw-sve-inst-spec"}, 178 // Refer to "Table K3.1 ARM recommendations for IMPLEMENTATION DEFINED event numbers" in ARMv8 sp 179 {0x0040, "raw-l1d-cache-rd"}, 180 {0x0041, "raw-l1d-cache-wr"}, 181 {0x0042, "raw-l1d-cache-refill-rd"}, 182 {0x0043, "raw-l1d-cache-refill-wr"}, 183 {0x0044, "raw-l1d-cache-refill-inner"}, 184 {0x0045, "raw-l1d-cache-refill-outer"}, 185 {0x0046, "raw-l1d-cache-wb-victim"}, 186 {0x0047, "raw-l1d-cache-wb-clean"}, 187 {0x0048, "raw-l1d-cache-inval"}, 188 // 0x0049-0x004B - Reserved 189 {0x004C, "raw-l1d-tlb-refill-rd"}, 190 {0x004D, "raw-l1d-tlb-refill-wr"}, 191 {0x004E, "raw-l1d-tlb-rd"}, 192 {0x004F, "raw-l1d-tlb-wr"}, 193 {0x0050, "raw-l2d-cache-rd"}, 194 {0x0051, "raw-l2d-cache-wr"}, 195 {0x0052, "raw-l2d-cache-refill-rd"}, 196 {0x0053, "raw-l2d-cache-refill-wr"}, 197 // 0x0054-0x0055 - Reserved 198 {0x0056, "raw-l2d-cache-wb-victim"}, 199 {0x0057, "raw-l2d-cache-wb-clean"}, 200 {0x0058, "raw-l2d-cache-inval"}, 201 // 0x0059-0x005B - Reserved 202 {0x005C, "raw-l2d-tlb-refill-rd"}, 203 {0x005D, "raw-l2d-tlb-refill-wr"}, 204 {0x005E, "raw-l2d-tlb-rd"}, 205 {0x005F, "raw-l2d-tlb-wr"}, 206 {0x0060, "raw-bus-access-rd"}, 207 {0x0061, "raw-bus-access-wr"}, 208 {0x0062, "raw-bus-access-shared"}, 209 {0x0063, "raw-bus-access-not-shared"}, 210 {0x0064, "raw-bus-access-normal"}, 211 {0x0065, "raw-bus-access-periph"}, 212 {0x0066, "raw-mem-access-rd"}, 213 {0x0067, "raw-mem-access-wr"}, 214 {0x0068, "raw-unaligned-ld-spec"}, 215 {0x0069, "raw-unaligned-st-spec"}, 216 {0x006A, "raw-unaligned-ldst-spec"}, 217 // 0x006B - Reserved 218 {0x006C, "raw-ldrex-spec"}, 219 {0x006D, "raw-strex-pass-spec"}, 220 {0x006E, "raw-strex-fail-spec"}, 221 {0x006F, "raw-strex-spec"}, 222 {0x0070, "raw-ld-spec"}, 223 {0x0071, "raw-st-spec"}, 224 {0x0072, "raw-ldst-spec"}, 225 {0x0073, "raw-dp-spec"}, 226 {0x0074, "raw-ase-spec"}, 227 {0x0075, "raw-vfp-spec"}, 228 {0x0076, "raw-pc-write-spec"}, 229 {0x0077, "raw-crypto-spec"}, 230 {0x0078, "raw-br-immed-spec"}, 231 {0x0079, "raw-br-return-spec"}, 232 {0x007A, "raw-br-indirect-spec"}, 233 // 0x007B - Reserved 234 {0x007C, "raw-isb-spec"}, 235 {0x007D, "raw-dsb-spec"}, 236 {0x007E, "raw-dmb-spec"}, 237 // 0x007F-0x0080 - Reserved 238 {0x0081, "raw-exc-undef"}, 239 {0x0082, "raw-exc-svc"}, 240 {0x0083, "raw-exc-pabort"}, 241 {0x0084, "raw-exc-dabort"}, 242 // 0x0085 - Reserved 243 {0x0086, "raw-exc-irq"}, 244 {0x0087, "raw-exc-fiq"}, 245 {0x0088, "raw-exc-smc"}, 246 // 0x0089 - Reserved 247 {0x008A, "raw-exc-hvc"}, 248 {0x008B, "raw-exc-trap-pabort"}, 249 {0x008C, "raw-exc-trap-dabort"}, 250 {0x008D, "raw-exc-trap-other"}, 251 {0x008E, "raw-exc-trap-irq"}, 252 {0x008F, "raw-exc-trap-fiq"}, 253 {0x0090, "raw-rc-ld-spec"}, 254 {0x0091, "raw-rc-st-spec"}, 255 // 0x0092-0x009F - Reserved 256 {0x00A0, "raw-l3d-cache-rd"}, 257 {0x00A1, "raw-l3d-cache-wr"}, 258 {0x00A2, "raw-l3d-cache-refill-rd"}, 259 {0x00A3, "raw-l3d-cache-refill-wr"}, 260 // 0x00A4-0x00A5 - Reserved 261 {0x00A6, "raw-l3d-cache-wb-victim"}, 262 {0x00A7, "raw-l3d-cache-wb-clean"}, 263 {0x00A8, "raw-l3d-cache-inval"}, 264 }; 265 static ConfigTable PERF_TRACEPOINT_CONFIGS = { 266 267 }; 268 269 static const std::map<perf_type_id, std::string> PERF_TYPES = { 270 {PERF_TYPE_HARDWARE, "hardware"}, 271 {PERF_TYPE_SOFTWARE, "software"}, 272 {PERF_TYPE_TRACEPOINT, "tracepoint"}, 273 {PERF_TYPE_HW_CACHE, "hardware cache"}, 274 {PERF_TYPE_RAW, "raw"}, 275 }; 276 277 static std::map<perf_type_id, ConfigTable> TYPE_CONFIGS = { 278 {PERF_TYPE_HARDWARE, (PERF_HW_CONFIGS)}, {PERF_TYPE_SOFTWARE, (PERF_SW_CONFIGS)}, 279 {PERF_TYPE_HW_CACHE, (PERF_HW_CACHE_CONFIGS)}, {PERF_TYPE_RAW, (PERF_RAW_CONFIGS)}, 280 {PERF_TYPE_TRACEPOINT, (PERF_TRACEPOINT_CONFIGS)}, 281 }; 282 283 // default config 284 static const std::vector<__u64> DEFAULT_HW_CONFIGS = { 285 PERF_COUNT_HW_CPU_CYCLES, 286 #if defined(__aarch64__) 287 PERF_COUNT_HW_STALLED_CYCLES_FRONTEND, 288 PERF_COUNT_HW_STALLED_CYCLES_BACKEND, 289 #endif 290 PERF_COUNT_HW_INSTRUCTIONS, 291 PERF_COUNT_HW_BRANCH_INSTRUCTIONS, 292 PERF_COUNT_HW_BRANCH_MISSES, 293 }; 294 static const std::vector<__u64> DEFAULT_SW_CONFIGS = { 295 PERF_COUNT_SW_TASK_CLOCK, 296 PERF_COUNT_SW_CONTEXT_SWITCHES, 297 PERF_COUNT_SW_PAGE_FAULTS, 298 }; 299 static const std::map<perf_type_id, std::vector<__u64>> DEFAULT_TYPE_CONFIGS = { 300 {PERF_TYPE_HARDWARE, DEFAULT_HW_CONFIGS}, 301 {PERF_TYPE_SOFTWARE, DEFAULT_SW_CONFIGS}, 302 }; 303 304 struct read_format_event { 305 __u64 value; /* The value of the event */ 306 __u64 id; /* if PERF_FORMAT_ID */ 307 }; 308 309 struct read_format_group { 310 __u64 nr; /* The number of events */ 311 __u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */ 312 __u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */ 313 read_format_event events[1]; 314 }; 315 316 struct read_format_no_group { 317 __u64 value; /* The value of the event */ 318 __u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */ 319 __u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */ 320 __u64 id; /* if PERF_FORMAT_ID */ 321 }; 322 323 class PerfEvents { 324 public: 325 static constexpr uint64_t DEFAULT_SAMPLE_FREQUNCY = 4000; 326 static constexpr uint64_t DEFAULT_SAMPLE_PERIOD = 1; 327 static constexpr uint64_t DEFAULT_TIMEOUT = 10 * 1000; 328 static constexpr size_t MIN_BUFFER_SIZE = 64 * 1024 * 1024; 329 static constexpr size_t BUFFER_LOW_LEVEL = 10 * 1024 * 1024; 330 static constexpr size_t BUFFER_CRITICAL_LEVEL = 5 * 1024 * 1024; 331 static constexpr size_t MAX_BUFFER_SIZE_LITTLE = 128 * 1024 * 1024; 332 static constexpr size_t MAX_BUFFER_SIZE_LARGE = 256 * 1024 * 1024; 333 334 PerfEvents(); 335 ~PerfEvents(); 336 337 bool AddEvents(const std::vector<std::string> &eventStrings, bool group = false); 338 bool PrepareTracking(void); 339 bool StartTracking(bool immediately = true); 340 bool StopTracking(void); 341 bool PauseTracking(void); 342 bool ResumeTracking(void); 343 /* call sequence 344 1. setXXX 345 2. AddEvents() 346 3. PrepareTracking 347 4. StartTracking (blocking...) 348 */ 349 bool EnableTracking(); 350 bool IsTrackRunning(); 351 352 void SetSystemTarget(bool); 353 void SetCpu(const std::vector<pid_t> cpus); // cpu id must be [0~N] 354 void SetPid(const std::vector<pid_t> pids); // tis is same as pid in kernel 355 void SetTimeOut(float timeOut); 356 void SetTimeReport(int); 357 void SetVerboseReport(bool); 358 bool AddOffCpuEvent(); 359 SetTrackedCommand(const std::vector<std::string> & trackedCommand)360 inline void SetTrackedCommand(const std::vector<std::string> &trackedCommand) 361 { 362 if (!trackedCommand.empty()) { 363 trackedCommand_ = TrackedCommand::CreateInstance(trackedCommand); 364 } 365 } 366 367 void SetSampleFrequency(unsigned int frequency); 368 void SetSamplePeriod(unsigned int period); 369 370 enum SampleStackType { 371 NONE, 372 FP, 373 DWARF, 374 }; 375 void SetSampleStackType(SampleStackType type); 376 void SetDwarfSampleStackSize(uint32_t stackSize); 377 void SetMmapPages(size_t mmapPages); 378 std::vector<AttrWithId> GetAttrWithId() const; 379 SetInherit(bool inherit)380 void SetInherit(bool inherit) 381 { 382 inherit_ = inherit; 383 }; SetClockId(int clockId)384 void SetClockId(int clockId) 385 { 386 clockId_ = clockId; 387 }; 388 void SetPerCpu(bool perCpu); 389 void SetPerThread(bool perThread); 390 bool SetBranchSampleType(uint64_t value); 391 bool AddDefaultEvent(perf_type_id type); 392 393 std::map<__u64, std::string> GetSupportEvents(perf_type_id type); 394 395 struct Summary { 396 int cpu; 397 pid_t tid; 398 __u64 eventCount = 0; 399 __u64 time_enabled = 0; 400 __u64 time_running = 0; SummarySummary401 Summary(const int cpu, const pid_t tid, const __u64 eventCount, 402 const __u64 time_enabled, const __u64 time_running) 403 : cpu(cpu), tid(tid), eventCount(eventCount), time_enabled(time_enabled), time_running(time_running) 404 { 405 } 406 }; 407 408 struct ReportSum { 409 int cpu; 410 pid_t pid; 411 pid_t tid; 412 double scaleSum = 1.0; 413 double commentSum = 0; 414 __u64 eventCountSum = 0; 415 __u64 enabledSum = 0; 416 __u64 runningSum = 0; 417 std::string configName = ""; 418 std::string threadName = ""; 419 }; 420 struct CountEvent { 421 bool userOnly = false; 422 bool kernelOnly = false; 423 __u64 eventCount = 0; 424 __u64 time_enabled = 0; 425 __u64 time_running = 0; 426 __u64 id = 0; 427 double used_cpus = 0; 428 std::vector<Summary> summaries; 429 }; 430 using StatCallBack = 431 std::function<void(const std::map<std::string, std::unique_ptr<PerfEvents::CountEvent>> &)>; 432 using RecordCallBack = std::function<bool(std::unique_ptr<PerfEventRecord>)>; 433 434 void SetStatCallBack(StatCallBack reportCallBack); 435 void SetRecordCallBack(RecordCallBack recordCallBack); GetLostSamples(size_t & lostSamples,size_t & lostNonSamples)436 void GetLostSamples(size_t &lostSamples, size_t &lostNonSamples) 437 { 438 lostSamples = lostSamples_; 439 lostNonSamples = lostNonSamples_; 440 } 441 442 // review: remove this function. GetStaticConfigName(perf_type_id type_id,__u64 config_id)443 static const std::string GetStaticConfigName(perf_type_id type_id, __u64 config_id) 444 { 445 auto typeConfigs = TYPE_CONFIGS.find(type_id); 446 if (typeConfigs != TYPE_CONFIGS.end()) { 447 auto configs = typeConfigs->second; 448 auto config = configs.find(config_id); 449 if (config != configs.end()) { 450 return config->second; 451 } else { 452 HLOGW("config not found for %u:%lld in %zu:%zu", type_id, config_id, 453 TYPE_CONFIGS.size(), configs.size()); 454 // dump all config size 455 for (auto types : TYPE_CONFIGS) { 456 HLOGV("type id %d %zu", types.first, types.second.size()); 457 } 458 } 459 } else { 460 HLOGW("type not found for %d in %zu", type_id, TYPE_CONFIGS.size()); 461 } 462 return "<not found>"; 463 }; 464 GetTraceConfigName(__u64 config_id)465 const std::string GetTraceConfigName(__u64 config_id) 466 { 467 auto config = traceConfigTable.find(config_id); 468 if (config != traceConfigTable.end()) { 469 return config->second; 470 } else { 471 HLOGW("config not found for %lld in traceConfigTable.", config_id); 472 } 473 return "<not found>"; 474 }; 475 476 static const std::string GetTypeName(perf_type_id type_id); 477 bool ParseEventName(const std::string &nameStr, std::string &name, bool &excludeUser, 478 bool &excludeKernel, bool &isTracePoint); 479 480 // mmap one fd for each cpu 481 struct MmapFd { 482 int fd; 483 perf_event_mmap_page *mmapPage = nullptr; 484 uint8_t *buf = nullptr; 485 size_t bufSize = 0; 486 // for read and sort 487 size_t dataSize = 0; 488 perf_event_header header; 489 uint64_t timestamp = 0; 490 const perf_event_attr *attr = nullptr; 491 size_t posCallChain = 0; 492 }; 493 494 private: 495 size_t recordEventCount_ = 0; // only for debug time 496 #ifdef HIPERF_DEBUG_TIME 497 std::chrono::microseconds recordCallBackTime_ = std::chrono::microseconds::zero(); 498 std::chrono::microseconds recordWaitDataTime_ = std::chrono::microseconds::zero(); 499 std::chrono::microseconds recordSleepTime_ = std::chrono::microseconds::zero(); 500 std::chrono::microseconds recordKernelReadTime_ = std::chrono::microseconds::zero(); 501 #endif 502 size_t lostSamples_ = 0; 503 size_t lostNonSamples_ = 0; 504 505 std::unique_ptr<RingBuffer> recordBuf_ {nullptr}; 506 bool recordBufReady_ = false; 507 std::mutex mtxRrecordBuf_; 508 std::condition_variable cvRecordBuf_; 509 std::thread readRecordBufThread_; 510 std::atomic_bool readRecordThreadRunning_ = false; 511 bool startedTracking_ = false; 512 bool isLowPriorityThread_ = false; 513 void RecordLoop(); 514 void StatLoop(); 515 bool IsRecordInMmap(); 516 void ReadRecordsFromMmaps(); 517 bool GetRecordFromMmap(MmapFd &mmap); 518 void GetRecordFieldFromMmap(MmapFd &mmap, void *dest, size_t pos, size_t size); 519 void MoveRecordToBuf(MmapFd &mmap); 520 size_t GetCallChainPosInSampleRecord(const perf_event_attr &attr); 521 size_t GetStackSizePosInSampleRecord(MmapFd &mmap); 522 bool CutStackAndMove(MmapFd &mmap); 523 void ReadRecordFromBuf(); 524 size_t CalcBufferSize(); 525 bool PrepareRecordThread(); 526 void WaitRecordThread(); 527 bool HaveTargetsExit(const std::chrono::steady_clock::time_point &startTime); 528 void ExitReadRecordBufThread(); 529 530 enum EventSpaceType { 531 UNKNOW = 0, 532 USER = 1, 533 KERNEL = 2, 534 USER_KERNEL = 3, 535 }; 536 uint8_t eventSpaceType_ = EventSpaceType::UNKNOW; 537 538 bool inherit_ = false; 539 std::vector<pid_t> pids_; 540 std::vector<pid_t> cpus_; 541 std::vector<OHOS::UniqueFd> groups_; 542 std::chrono::milliseconds timeOut_; // milliseconds 543 std::chrono::milliseconds timeReport_; // means same as timeOut 544 bool perCpu_ = false; 545 bool perThread_ = false; 546 bool verboseReport_ = false; 547 bool prepared_ = false; 548 ConfigTable traceConfigTable; 549 550 unsigned int samplePeriod_ = 0; 551 unsigned int sampleFreq_ = 0; 552 553 struct FdItem { 554 OHOS::UniqueFd fd; 555 int cpu; 556 pid_t pid; 557 __u64 eventCount; 558 mutable uint64_t perf_id_ = 0; GetPrefIdFdItem559 uint64_t GetPrefId() const 560 { 561 if (perf_id_ == 0) { 562 read_format_no_group readNoGroupValue; 563 if (read(fd, &readNoGroupValue, sizeof(readNoGroupValue)) > 0) { 564 perf_id_ = readNoGroupValue.id; 565 } else { 566 HLOGW("read failed with fd %d", fd.Get()); 567 } 568 } 569 return perf_id_; 570 } 571 }; 572 struct EventItem { 573 std::string typeName; 574 std::string configName; 575 perf_event_attr attr = {}; 576 std::vector<FdItem> fdItems; 577 }; 578 struct EventGroupItem { 579 std::vector<EventItem> eventItems; 580 }; 581 std::vector<EventGroupItem> eventGroupItem_; 582 583 std::map<int, MmapFd> cpuMmap_; 584 std::vector<MmapFd *> MmapRecordHeap_; 585 586 #if !is_mingw 587 std::vector<struct pollfd> pollFds_; 588 #endif 589 const int pollTimeOut_ = 500; // ms 590 size_t pageSize_ = 4096; 591 bool systemTarget_ = false; 592 bool excludeHiperf_ = false; 593 pid_t selfPid_ = -1; 594 unsigned int mmapPages_ = 0; 595 int clockId_ = -1; 596 uint64_t branchSampleType_ = 0; 597 598 SampleStackType sampleStackType_ = SampleStackType::NONE; 599 uint32_t dwarfSampleStackSize_ = MAX_SAMPLE_STACK_SIZE; 600 601 // read records from the ring buffer singleton 602 void ReadRecordFromBuffer(); 603 void ReadRecordFromBufferThread(); 604 605 std::unique_ptr<TrackedCommand> trackedCommand_ = {}; 606 607 StatCallBack reportCallBack_; 608 RecordCallBack recordCallBack_; 609 610 void LoadTracepointEventTypesFromSystem(); 611 bool PerfEventsEnable(bool); 612 bool AddEvent(perf_type_id type, __u64 config, bool excludeUser = false, 613 bool excludeKernel = false, bool followGroup = false); 614 bool AddEvent(const std::string &eventString, bool followGroup = false); 615 bool IsEventSupport(perf_type_id type, __u64 config); 616 bool IsEventAttrSupport(perf_event_attr &attr); 617 618 std::chrono::time_point<std::chrono::steady_clock> trackingStartTime_; 619 std::chrono::time_point<std::chrono::steady_clock> trackingEndTime_; 620 std::chrono::time_point<std::chrono::steady_clock> readingStartTime_; 621 622 std::map<std::string, std::unique_ptr<CountEvent>> countEvents_; 623 624 void PutAllCpus(); 625 bool PrepareFdEvents(); 626 bool CreateFdEvents(); 627 bool StatReport(const __u64 &durationInSec); 628 bool CreateMmap(const FdItem &item, const perf_event_attr &attr); 629 GetDefaultAttr()630 const perf_event_attr *GetDefaultAttr() 631 { 632 HLOG_ASSERT(eventGroupItem_.size() > 0); 633 HLOG_ASSERT(eventGroupItem_[0].eventItems.size() > 0); 634 return &(eventGroupItem_.at(0).eventItems.at(0).attr); 635 }; 636 637 OHOS::UniqueFd Open(perf_event_attr &attr, pid_t pid = 0, int cpu = -1, int group_fd = -1, 638 unsigned long flags = 0); 639 std::unique_ptr<perf_event_attr> CreateDefaultAttr(perf_type_id type, __u64 config); 640 }; 641 } // namespace HiPerf 642 } // namespace Developtools 643 } // namespace OHOS 644 #endif // HIPERF_PERF_EVENTS_H 645