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 }; 150 static ConfigTable PERF_TRACEPOINT_CONFIGS = { 151 152 }; 153 154 static const std::map<perf_type_id, std::string> PERF_TYPES = { 155 {PERF_TYPE_HARDWARE, "hardware"}, 156 {PERF_TYPE_SOFTWARE, "software"}, 157 {PERF_TYPE_TRACEPOINT, "tracepoint"}, 158 {PERF_TYPE_HW_CACHE, "hardware cache"}, 159 {PERF_TYPE_RAW, "raw"}, 160 }; 161 162 static std::map<perf_type_id, ConfigTable> TYPE_CONFIGS = { 163 {PERF_TYPE_HARDWARE, (PERF_HW_CONFIGS)}, {PERF_TYPE_SOFTWARE, (PERF_SW_CONFIGS)}, 164 {PERF_TYPE_HW_CACHE, (PERF_HW_CACHE_CONFIGS)}, {PERF_TYPE_RAW, (PERF_RAW_CONFIGS)}, 165 {PERF_TYPE_TRACEPOINT, (PERF_TRACEPOINT_CONFIGS)}, 166 }; 167 168 // default config 169 static const std::vector<__u64> DEFAULT_HW_CONFIGS = { 170 PERF_COUNT_HW_CPU_CYCLES, 171 #if defined(__aarch64__) 172 PERF_COUNT_HW_STALLED_CYCLES_FRONTEND, 173 PERF_COUNT_HW_STALLED_CYCLES_BACKEND, 174 #endif 175 PERF_COUNT_HW_INSTRUCTIONS, 176 PERF_COUNT_HW_BRANCH_INSTRUCTIONS, 177 PERF_COUNT_HW_BRANCH_MISSES, 178 }; 179 static const std::vector<__u64> DEFAULT_SW_CONFIGS = { 180 PERF_COUNT_SW_TASK_CLOCK, 181 PERF_COUNT_SW_CONTEXT_SWITCHES, 182 PERF_COUNT_SW_PAGE_FAULTS, 183 }; 184 static const std::map<perf_type_id, std::vector<__u64>> DEFAULT_TYPE_CONFIGS = { 185 {PERF_TYPE_HARDWARE, DEFAULT_HW_CONFIGS}, 186 {PERF_TYPE_SOFTWARE, DEFAULT_SW_CONFIGS}, 187 }; 188 189 struct read_format_event { 190 __u64 value; /* The value of the event */ 191 __u64 id; /* if PERF_FORMAT_ID */ 192 }; 193 194 struct read_format_group { 195 __u64 nr; /* The number of events */ 196 __u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */ 197 __u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */ 198 read_format_event events[1]; 199 }; 200 201 struct read_format_no_group { 202 __u64 value; /* The value of the event */ 203 __u64 time_enabled; /* if PERF_FORMAT_TOTAL_TIME_ENABLED */ 204 __u64 time_running; /* if PERF_FORMAT_TOTAL_TIME_RUNNING */ 205 __u64 id; /* if PERF_FORMAT_ID */ 206 }; 207 208 class PerfEvents { 209 public: 210 static constexpr uint64_t DEFAULT_SAMPLE_FREQUNCY = 4000; 211 static constexpr uint64_t DEFAULT_SAMPLE_PERIOD = 1; 212 static constexpr uint64_t DEFAULT_TIMEOUT = 10 * 1000; 213 static constexpr size_t MIN_BUFFER_SIZE = 64 * 1024 * 1024; 214 #ifdef LITTLE_MEMORY 215 static constexpr size_t MAX_BUFFER_SIZE = 128 * 1024 * 1024; 216 #else 217 static constexpr size_t MAX_BUFFER_SIZE = 256 * 1024 * 1024; 218 #endif 219 static constexpr size_t BUFFER_LOW_LEVEL = 10 * 1024 * 1024; 220 static constexpr size_t BUFFER_CRITICAL_LEVEL = 5 * 1024 * 1024; 221 222 PerfEvents(); 223 ~PerfEvents(); 224 225 bool AddEvents(const std::vector<std::string> &eventStrings, bool group = false); 226 bool PrepareTracking(void); 227 bool StartTracking(bool immediately = true); 228 bool StopTracking(void); 229 bool PauseTracking(void); 230 bool ResumeTracking(void); 231 /* call sequence 232 1. setXXX 233 2. AddEvents() 234 3. PrepareTracking 235 4. StartTracking (blocking...) 236 */ 237 bool EnableTracking(); 238 bool IsTrackRunning(); 239 240 void SetSystemTarget(bool); 241 void SetCpu(const std::vector<pid_t> cpus); // cpu id must be [0~N] 242 void SetPid(const std::vector<pid_t> pids); // tis is same as pid in kernel 243 void SetTimeOut(float timeOut); 244 void SetTimeReport(int); 245 void SetVerboseReport(bool); 246 bool AddOffCpuEvent(); 247 SetTrackedCommand(const std::vector<std::string> & trackedCommand)248 inline void SetTrackedCommand(const std::vector<std::string> &trackedCommand) 249 { 250 if (!trackedCommand.empty()) { 251 trackedCommand_ = TrackedCommand::CreateInstance(trackedCommand); 252 } 253 } 254 255 void SetSampleFrequency(unsigned int frequency); 256 void SetSamplePeriod(unsigned int period); 257 258 enum SampleStackType { 259 NONE, 260 FP, 261 DWARF, 262 }; 263 void SetSampleStackType(SampleStackType type); 264 void SetDwarfSampleStackSize(uint32_t stackSize); 265 void SetMmapPages(size_t mmapPages); 266 std::vector<AttrWithId> GetAttrWithId() const; 267 SetInherit(bool inherit)268 void SetInherit(bool inherit) 269 { 270 inherit_ = inherit; 271 }; SetClockId(int clockId)272 void SetClockId(int clockId) 273 { 274 clockId_ = clockId; 275 }; 276 bool SetBranchSampleType(uint64_t value); 277 bool AddDefaultEvent(perf_type_id type); 278 279 std::map<__u64, std::string> GetSupportEvents(perf_type_id type); 280 281 struct CountEvent { 282 bool userOnly = false; 283 bool kernelOnly = false; 284 __u64 eventCount = 0; 285 __u64 time_enabled = 0; 286 __u64 time_running = 0; 287 __u64 id = 0; 288 double used_cpus = 0; 289 }; 290 using StatCallBack = 291 std::function<void(const std::map<std::string, std::unique_ptr<PerfEvents::CountEvent>> &)>; 292 using RecordCallBack = std::function<bool(std::unique_ptr<PerfEventRecord>)>; 293 294 void SetStatCallBack(StatCallBack reportCallBack); 295 void SetRecordCallBack(RecordCallBack recordCallBack); GetLostSamples(size_t & lostSamples,size_t & lostNonSamples)296 void GetLostSamples(size_t &lostSamples, size_t &lostNonSamples) 297 { 298 lostSamples = lostSamples_; 299 lostNonSamples = lostNonSamples_; 300 } 301 302 // review: remove this function. GetStaticConfigName(perf_type_id type_id,__u64 config_id)303 static const std::string GetStaticConfigName(perf_type_id type_id, __u64 config_id) 304 { 305 auto typeConfigs = TYPE_CONFIGS.find(type_id); 306 if (typeConfigs != TYPE_CONFIGS.end()) { 307 auto configs = typeConfigs->second; 308 auto config = configs.find(config_id); 309 if (config != configs.end()) { 310 return config->second; 311 } else { 312 HLOGW("config not found for %u:%lld in %zu:%zu", type_id, config_id, 313 TYPE_CONFIGS.size(), configs.size()); 314 // dump all config size 315 for (auto types : TYPE_CONFIGS) { 316 HLOGV("type id %d %zu", types.first, types.second.size()); 317 } 318 } 319 } else { 320 HLOGW("type not found for %d in %zu", type_id, TYPE_CONFIGS.size()); 321 } 322 return "<not found>"; 323 }; 324 GetTraceConfigName(__u64 config_id)325 const std::string GetTraceConfigName(__u64 config_id) 326 { 327 auto config = traceConfigTable.find(config_id); 328 if (config != traceConfigTable.end()) { 329 return config->second; 330 } else { 331 HLOGW("config not found for %lld in traceConfigTable.", config_id); 332 } 333 return "<not found>"; 334 }; 335 336 static const std::string GetTypeName(perf_type_id type_id); 337 bool ParseEventName(const std::string &nameStr, std::string &name, bool &excludeUser, 338 bool &excludeKernel, bool &isTracePoint); 339 340 // mmap one fd for each cpu 341 struct MmapFd { 342 int fd; 343 perf_event_mmap_page *mmapPage = nullptr; 344 uint8_t *buf = nullptr; 345 size_t bufSize = 0; 346 // for read and sort 347 size_t dataSize = 0; 348 perf_event_header header; 349 uint64_t timestamp = 0; 350 const perf_event_attr *attr = nullptr; 351 size_t posCallChain = 0; 352 }; 353 354 private: 355 size_t recordEventCount_ = 0; // only for debug time 356 #ifdef HIPERF_DEBUG_TIME 357 std::chrono::microseconds recordCallBackTime_ = std::chrono::microseconds::zero(); 358 std::chrono::microseconds recordWaitDataTime_ = std::chrono::microseconds::zero(); 359 std::chrono::microseconds recordSleepTime_ = std::chrono::microseconds::zero(); 360 std::chrono::microseconds recordKernelReadTime_ = std::chrono::microseconds::zero(); 361 #endif 362 size_t lostSamples_ = 0; 363 size_t lostNonSamples_ = 0; 364 365 std::unique_ptr<RingBuffer> recordBuf_ {nullptr}; 366 bool recordBufReady_ = false; 367 std::mutex mtxRrecordBuf_; 368 std::condition_variable cvRecordBuf_; 369 std::thread readRecordBufThread_; 370 std::atomic_bool readRecordThreadRunning_ = false; 371 bool startedTracking_ = false; 372 bool isLowPriorityThread_ = false; 373 void RecordLoop(); 374 void StatLoop(); 375 bool IsRecordInMmap(); 376 void ReadRecordsFromMmaps(); 377 bool GetRecordFromMmap(MmapFd &mmap); 378 void GetRecordFieldFromMmap(MmapFd &mmap, void *dest, size_t pos, size_t size); 379 void MoveRecordToBuf(MmapFd &mmap); 380 size_t GetCallChainPosInSampleRecord(const perf_event_attr &attr); 381 size_t GetStackSizePosInSampleRecord(MmapFd &mmap); 382 bool CutStackAndMove(MmapFd &mmap); 383 void ReadRecordFromBuf(); 384 size_t CalcBufferSize(); 385 bool PrepareRecordThread(); 386 void WaitRecordThread(); 387 bool HaveTargetsExit(const std::chrono::steady_clock::time_point &startTime); 388 void ExitReadRecordBufThread(); 389 390 enum EventSpaceType { 391 UNKNOW = 0, 392 USER = 1, 393 KERNEL = 2, 394 USER_KERNEL = 3, 395 }; 396 uint8_t eventSpaceType_ = EventSpaceType::UNKNOW; 397 398 bool inherit_ = false; 399 std::vector<pid_t> pids_; 400 std::vector<pid_t> cpus_; 401 std::vector<OHOS::UniqueFd> groups_; 402 std::chrono::milliseconds timeOut_; // milliseconds 403 std::chrono::milliseconds timeReport_; // means same as timeOut 404 bool verboseReport_ = false; 405 bool prepared_ = false; 406 ConfigTable traceConfigTable; 407 408 unsigned int samplePeriod_ = 0; 409 unsigned int sampleFreq_ = 0; 410 411 struct FdItem { 412 OHOS::UniqueFd fd; 413 int cpu; 414 pid_t pid; 415 __u64 eventCount; 416 mutable uint64_t perf_id_ = 0; GetPrefIdFdItem417 uint64_t GetPrefId() const 418 { 419 if (perf_id_ == 0) { 420 read_format_no_group readNoGroupValue; 421 if (read(fd, &readNoGroupValue, sizeof(readNoGroupValue)) > 0) { 422 perf_id_ = readNoGroupValue.id; 423 } else { 424 HLOGW("read failed with fd %d", fd.Get()); 425 } 426 } 427 return perf_id_; 428 } 429 }; 430 struct EventItem { 431 std::string typeName; 432 std::string configName; 433 perf_event_attr attr = {}; 434 std::vector<FdItem> fdItems; 435 }; 436 struct EventGroupItem { 437 std::vector<EventItem> eventItems; 438 }; 439 std::vector<EventGroupItem> eventGroupItem_; 440 441 std::map<int, MmapFd> cpuMmap_; 442 std::vector<MmapFd *> MmapRecordHeap_; 443 444 #if !is_mingw 445 std::vector<struct pollfd> pollFds_; 446 #endif 447 const int pollTimeOut_ = 500; // ms 448 size_t pageSize_ = 4096; 449 bool systemTarget_ = false; 450 bool excludeHiperf_ = false; 451 pid_t selfPid_ = -1; 452 unsigned int mmapPages_ = 0; 453 int clockId_ = -1; 454 uint64_t branchSampleType_ = 0; 455 456 SampleStackType sampleStackType_ = SampleStackType::NONE; 457 uint32_t dwarfSampleStackSize_ = MAX_SAMPLE_STACK_SIZE; 458 459 // read records from the ring buffer singleton 460 void ReadRecordFromBuffer(); 461 void ReadRecordFromBufferThread(); 462 463 std::unique_ptr<TrackedCommand> trackedCommand_ = {}; 464 465 StatCallBack reportCallBack_; 466 RecordCallBack recordCallBack_; 467 468 void LoadTracepointEventTypesFromSystem(); 469 bool PerfEventsEnable(bool); 470 bool AddEvent(perf_type_id type, __u64 config, bool excludeUser = false, 471 bool excludeKernel = false, bool followGroup = false); 472 bool AddEvent(const std::string &eventString, bool followGroup = false); 473 bool IsEventSupport(perf_type_id type, __u64 config); 474 bool IsEventAttrSupport(perf_event_attr &attr); 475 476 std::chrono::time_point<std::chrono::steady_clock> trackingStartTime_; 477 std::chrono::time_point<std::chrono::steady_clock> trackingEndTime_; 478 std::chrono::time_point<std::chrono::steady_clock> readingStartTime_; 479 480 std::map<std::string, std::unique_ptr<CountEvent>> countEvents_; 481 482 void PutAllCpus(); 483 bool PrepareFdEvents(); 484 bool CreateFdEvents(); 485 bool StatReport(const __u64 &durationInSec); 486 bool CreateMmap(const FdItem &item, const perf_event_attr &attr); 487 GetDefaultAttr()488 const perf_event_attr *GetDefaultAttr() 489 { 490 HLOG_ASSERT(eventGroupItem_.size() > 0); 491 HLOG_ASSERT(eventGroupItem_[0].eventItems.size() > 0); 492 return &(eventGroupItem_.at(0).eventItems.at(0).attr); 493 }; 494 495 OHOS::UniqueFd Open(perf_event_attr &attr, pid_t pid = 0, int cpu = -1, int group_fd = -1, 496 unsigned long flags = 0); 497 std::unique_ptr<perf_event_attr> CreateDefaultAttr(perf_type_id type, __u64 config); 498 }; 499 } // namespace HiPerf 500 } // namespace Developtools 501 } // namespace OHOS 502 #endif // HIPERF_PERF_EVENTS_H 503