trace.cc revision 41b175aba41c9365a1c53b8a1afbd17129c87c14
1/* 2 * Copyright (C) 2011 The Android Open Source Project 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 "trace.h" 18 19#include <sys/uio.h> 20#include <unistd.h> 21 22#define ATRACE_TAG ATRACE_TAG_DALVIK 23#include "cutils/trace.h" 24 25#include "base/casts.h" 26#include "base/stl_util.h" 27#include "base/time_utils.h" 28#include "base/unix_file/fd_file.h" 29#include "class_linker.h" 30#include "common_throws.h" 31#include "debugger.h" 32#include "dex_file-inl.h" 33#include "instrumentation.h" 34#include "mirror/art_method-inl.h" 35#include "mirror/class-inl.h" 36#include "mirror/dex_cache-inl.h" 37#include "mirror/object_array-inl.h" 38#include "mirror/object-inl.h" 39#include "os.h" 40#include "scoped_thread_state_change.h" 41#include "ScopedLocalRef.h" 42#include "thread.h" 43#include "thread_list.h" 44#include "utils.h" 45#include "entrypoints/quick/quick_entrypoints.h" 46 47namespace art { 48 49// File format: 50// header 51// record 0 52// record 1 53// ... 54// 55// Header format: 56// u4 magic ('SLOW') 57// u2 version 58// u2 offset to data 59// u8 start date/time in usec 60// u2 record size in bytes (version >= 2 only) 61// ... padding to 32 bytes 62// 63// Record format v1: 64// u1 thread ID 65// u4 method ID | method action 66// u4 time delta since start, in usec 67// 68// Record format v2: 69// u2 thread ID 70// u4 method ID | method action 71// u4 time delta since start, in usec 72// 73// Record format v3: 74// u2 thread ID 75// u4 method ID | method action 76// u4 time delta since start, in usec 77// u4 wall time since start, in usec (when clock == "dual" only) 78// 79// 32 bits of microseconds is 70 minutes. 80// 81// All values are stored in little-endian order. 82 83enum TraceAction { 84 kTraceMethodEnter = 0x00, // method entry 85 kTraceMethodExit = 0x01, // method exit 86 kTraceUnroll = 0x02, // method exited by exception unrolling 87 // 0x03 currently unused 88 kTraceMethodActionMask = 0x03, // two bits 89}; 90 91static constexpr uint8_t kOpNewMethod = 1U; 92static constexpr uint8_t kOpNewThread = 2U; 93 94class BuildStackTraceVisitor : public StackVisitor { 95 public: 96 explicit BuildStackTraceVisitor(Thread* thread) 97 : StackVisitor(thread, nullptr, StackVisitor::StackWalkKind::kIncludeInlinedFrames), 98 method_trace_(Trace::AllocStackTrace()) {} 99 100 bool VisitFrame() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 101 mirror::ArtMethod* m = GetMethod(); 102 // Ignore runtime frames (in particular callee save). 103 if (!m->IsRuntimeMethod()) { 104 method_trace_->push_back(m); 105 } 106 return true; 107 } 108 109 // Returns a stack trace where the topmost frame corresponds with the first element of the vector. 110 std::vector<mirror::ArtMethod*>* GetStackTrace() const { 111 return method_trace_; 112 } 113 114 private: 115 std::vector<mirror::ArtMethod*>* const method_trace_; 116}; 117 118static const char kTraceTokenChar = '*'; 119static const uint16_t kTraceHeaderLength = 32; 120static const uint32_t kTraceMagicValue = 0x574f4c53; 121static const uint16_t kTraceVersionSingleClock = 2; 122static const uint16_t kTraceVersionDualClock = 3; 123static const uint16_t kTraceRecordSizeSingleClock = 10; // using v2 124static const uint16_t kTraceRecordSizeDualClock = 14; // using v3 with two timestamps 125 126TraceClockSource Trace::default_clock_source_ = kDefaultTraceClockSource; 127 128Trace* volatile Trace::the_trace_ = nullptr; 129pthread_t Trace::sampling_pthread_ = 0U; 130std::unique_ptr<std::vector<mirror::ArtMethod*>> Trace::temp_stack_trace_; 131 132// The key identifying the tracer to update instrumentation. 133static constexpr const char* kTracerInstrumentationKey = "Tracer"; 134 135static mirror::ArtMethod* DecodeTraceMethodId(uint32_t tmid) { 136 return reinterpret_cast<mirror::ArtMethod*>(tmid & ~kTraceMethodActionMask); 137} 138 139static TraceAction DecodeTraceAction(uint32_t tmid) { 140 return static_cast<TraceAction>(tmid & kTraceMethodActionMask); 141} 142 143static uint32_t EncodeTraceMethodAndAction(mirror::ArtMethod* method, 144 TraceAction action) { 145 uint32_t tmid = PointerToLowMemUInt32(method) | action; 146 DCHECK_EQ(method, DecodeTraceMethodId(tmid)); 147 return tmid; 148} 149 150std::vector<mirror::ArtMethod*>* Trace::AllocStackTrace() { 151 if (temp_stack_trace_.get() != nullptr) { 152 return temp_stack_trace_.release(); 153 } else { 154 return new std::vector<mirror::ArtMethod*>(); 155 } 156} 157 158void Trace::FreeStackTrace(std::vector<mirror::ArtMethod*>* stack_trace) { 159 stack_trace->clear(); 160 temp_stack_trace_.reset(stack_trace); 161} 162 163void Trace::SetDefaultClockSource(TraceClockSource clock_source) { 164#if defined(__linux__) 165 default_clock_source_ = clock_source; 166#else 167 if (clock_source != TraceClockSource::kWall) { 168 LOG(WARNING) << "Ignoring tracing request to use CPU time."; 169 } 170#endif 171} 172 173static uint16_t GetTraceVersion(TraceClockSource clock_source) { 174 return (clock_source == TraceClockSource::kDual) ? kTraceVersionDualClock 175 : kTraceVersionSingleClock; 176} 177 178static uint16_t GetRecordSize(TraceClockSource clock_source) { 179 return (clock_source == TraceClockSource::kDual) ? kTraceRecordSizeDualClock 180 : kTraceRecordSizeSingleClock; 181} 182 183bool Trace::UseThreadCpuClock() { 184 return (clock_source_ == TraceClockSource::kThreadCpu) || 185 (clock_source_ == TraceClockSource::kDual); 186} 187 188bool Trace::UseWallClock() { 189 return (clock_source_ == TraceClockSource::kWall) || 190 (clock_source_ == TraceClockSource::kDual); 191} 192 193void Trace::MeasureClockOverhead() { 194 if (UseThreadCpuClock()) { 195 Thread::Current()->GetCpuMicroTime(); 196 } 197 if (UseWallClock()) { 198 MicroTime(); 199 } 200} 201 202// Compute an average time taken to measure clocks. 203uint32_t Trace::GetClockOverheadNanoSeconds() { 204 Thread* self = Thread::Current(); 205 uint64_t start = self->GetCpuMicroTime(); 206 207 for (int i = 4000; i > 0; i--) { 208 MeasureClockOverhead(); 209 MeasureClockOverhead(); 210 MeasureClockOverhead(); 211 MeasureClockOverhead(); 212 MeasureClockOverhead(); 213 MeasureClockOverhead(); 214 MeasureClockOverhead(); 215 MeasureClockOverhead(); 216 } 217 218 uint64_t elapsed_us = self->GetCpuMicroTime() - start; 219 return static_cast<uint32_t>(elapsed_us / 32); 220} 221 222// TODO: put this somewhere with the big-endian equivalent used by JDWP. 223static void Append2LE(uint8_t* buf, uint16_t val) { 224 *buf++ = static_cast<uint8_t>(val); 225 *buf++ = static_cast<uint8_t>(val >> 8); 226} 227 228// TODO: put this somewhere with the big-endian equivalent used by JDWP. 229static void Append4LE(uint8_t* buf, uint32_t val) { 230 *buf++ = static_cast<uint8_t>(val); 231 *buf++ = static_cast<uint8_t>(val >> 8); 232 *buf++ = static_cast<uint8_t>(val >> 16); 233 *buf++ = static_cast<uint8_t>(val >> 24); 234} 235 236// TODO: put this somewhere with the big-endian equivalent used by JDWP. 237static void Append8LE(uint8_t* buf, uint64_t val) { 238 *buf++ = static_cast<uint8_t>(val); 239 *buf++ = static_cast<uint8_t>(val >> 8); 240 *buf++ = static_cast<uint8_t>(val >> 16); 241 *buf++ = static_cast<uint8_t>(val >> 24); 242 *buf++ = static_cast<uint8_t>(val >> 32); 243 *buf++ = static_cast<uint8_t>(val >> 40); 244 *buf++ = static_cast<uint8_t>(val >> 48); 245 *buf++ = static_cast<uint8_t>(val >> 56); 246} 247 248static void GetSample(Thread* thread, void* arg) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 249 BuildStackTraceVisitor build_trace_visitor(thread); 250 build_trace_visitor.WalkStack(); 251 std::vector<mirror::ArtMethod*>* stack_trace = build_trace_visitor.GetStackTrace(); 252 Trace* the_trace = reinterpret_cast<Trace*>(arg); 253 the_trace->CompareAndUpdateStackTrace(thread, stack_trace); 254} 255 256static void ClearThreadStackTraceAndClockBase(Thread* thread, void* arg ATTRIBUTE_UNUSED) { 257 thread->SetTraceClockBase(0); 258 std::vector<mirror::ArtMethod*>* stack_trace = thread->GetStackTraceSample(); 259 thread->SetStackTraceSample(nullptr); 260 delete stack_trace; 261} 262 263void Trace::CompareAndUpdateStackTrace(Thread* thread, 264 std::vector<mirror::ArtMethod*>* stack_trace) { 265 CHECK_EQ(pthread_self(), sampling_pthread_); 266 std::vector<mirror::ArtMethod*>* old_stack_trace = thread->GetStackTraceSample(); 267 // Update the thread's stack trace sample. 268 thread->SetStackTraceSample(stack_trace); 269 // Read timer clocks to use for all events in this trace. 270 uint32_t thread_clock_diff = 0; 271 uint32_t wall_clock_diff = 0; 272 ReadClocks(thread, &thread_clock_diff, &wall_clock_diff); 273 if (old_stack_trace == nullptr) { 274 // If there's no previous stack trace sample for this thread, log an entry event for all 275 // methods in the trace. 276 for (std::vector<mirror::ArtMethod*>::reverse_iterator rit = stack_trace->rbegin(); 277 rit != stack_trace->rend(); ++rit) { 278 LogMethodTraceEvent(thread, *rit, instrumentation::Instrumentation::kMethodEntered, 279 thread_clock_diff, wall_clock_diff); 280 } 281 } else { 282 // If there's a previous stack trace for this thread, diff the traces and emit entry and exit 283 // events accordingly. 284 std::vector<mirror::ArtMethod*>::reverse_iterator old_rit = old_stack_trace->rbegin(); 285 std::vector<mirror::ArtMethod*>::reverse_iterator rit = stack_trace->rbegin(); 286 // Iterate bottom-up over both traces until there's a difference between them. 287 while (old_rit != old_stack_trace->rend() && rit != stack_trace->rend() && *old_rit == *rit) { 288 old_rit++; 289 rit++; 290 } 291 // Iterate top-down over the old trace until the point where they differ, emitting exit events. 292 for (std::vector<mirror::ArtMethod*>::iterator old_it = old_stack_trace->begin(); 293 old_it != old_rit.base(); ++old_it) { 294 LogMethodTraceEvent(thread, *old_it, instrumentation::Instrumentation::kMethodExited, 295 thread_clock_diff, wall_clock_diff); 296 } 297 // Iterate bottom-up over the new trace from the point where they differ, emitting entry events. 298 for (; rit != stack_trace->rend(); ++rit) { 299 LogMethodTraceEvent(thread, *rit, instrumentation::Instrumentation::kMethodEntered, 300 thread_clock_diff, wall_clock_diff); 301 } 302 FreeStackTrace(old_stack_trace); 303 } 304} 305 306void* Trace::RunSamplingThread(void* arg) { 307 Runtime* runtime = Runtime::Current(); 308 intptr_t interval_us = reinterpret_cast<intptr_t>(arg); 309 CHECK_GE(interval_us, 0); 310 CHECK(runtime->AttachCurrentThread("Sampling Profiler", true, runtime->GetSystemThreadGroup(), 311 !runtime->IsAotCompiler())); 312 313 while (true) { 314 usleep(interval_us); 315 ATRACE_BEGIN("Profile sampling"); 316 Thread* self = Thread::Current(); 317 Trace* the_trace; 318 { 319 MutexLock mu(self, *Locks::trace_lock_); 320 the_trace = the_trace_; 321 if (the_trace == nullptr) { 322 break; 323 } 324 } 325 326 runtime->GetThreadList()->SuspendAll(__FUNCTION__); 327 { 328 MutexLock mu(self, *Locks::thread_list_lock_); 329 runtime->GetThreadList()->ForEach(GetSample, the_trace); 330 } 331 runtime->GetThreadList()->ResumeAll(); 332 ATRACE_END(); 333 } 334 335 runtime->DetachCurrentThread(); 336 return nullptr; 337} 338 339void Trace::Start(const char* trace_filename, int trace_fd, size_t buffer_size, int flags, 340 TraceOutputMode output_mode, TraceMode trace_mode, int interval_us) { 341 Thread* self = Thread::Current(); 342 { 343 MutexLock mu(self, *Locks::trace_lock_); 344 if (the_trace_ != nullptr) { 345 LOG(ERROR) << "Trace already in progress, ignoring this request"; 346 return; 347 } 348 } 349 350 // Check interval if sampling is enabled 351 if (trace_mode == TraceMode::kSampling && interval_us <= 0) { 352 LOG(ERROR) << "Invalid sampling interval: " << interval_us; 353 ScopedObjectAccess soa(self); 354 ThrowRuntimeException("Invalid sampling interval: %d", interval_us); 355 return; 356 } 357 358 // Open trace file if not going directly to ddms. 359 std::unique_ptr<File> trace_file; 360 if (output_mode != TraceOutputMode::kDDMS) { 361 if (trace_fd < 0) { 362 trace_file.reset(OS::CreateEmptyFile(trace_filename)); 363 } else { 364 trace_file.reset(new File(trace_fd, "tracefile")); 365 trace_file->DisableAutoClose(); 366 } 367 if (trace_file.get() == nullptr) { 368 PLOG(ERROR) << "Unable to open trace file '" << trace_filename << "'"; 369 ScopedObjectAccess soa(self); 370 ThrowRuntimeException("Unable to open trace file '%s'", trace_filename); 371 return; 372 } 373 } 374 375 Runtime* runtime = Runtime::Current(); 376 377 // Enable count of allocs if specified in the flags. 378 bool enable_stats = false; 379 380 runtime->GetThreadList()->SuspendAll(__FUNCTION__); 381 382 // Create Trace object. 383 { 384 MutexLock mu(self, *Locks::trace_lock_); 385 if (the_trace_ != nullptr) { 386 LOG(ERROR) << "Trace already in progress, ignoring this request"; 387 } else { 388 enable_stats = (flags && kTraceCountAllocs) != 0; 389 the_trace_ = new Trace(trace_file.release(), trace_filename, buffer_size, flags, output_mode, 390 trace_mode); 391 if (trace_mode == TraceMode::kSampling) { 392 CHECK_PTHREAD_CALL(pthread_create, (&sampling_pthread_, nullptr, &RunSamplingThread, 393 reinterpret_cast<void*>(interval_us)), 394 "Sampling profiler thread"); 395 the_trace_->interval_us_ = interval_us; 396 } else { 397 runtime->GetInstrumentation()->AddListener(the_trace_, 398 instrumentation::Instrumentation::kMethodEntered | 399 instrumentation::Instrumentation::kMethodExited | 400 instrumentation::Instrumentation::kMethodUnwind); 401 // TODO: In full-PIC mode, we don't need to fully deopt. 402 runtime->GetInstrumentation()->EnableMethodTracing(kTracerInstrumentationKey); 403 } 404 } 405 } 406 407 runtime->GetThreadList()->ResumeAll(); 408 409 // Can't call this when holding the mutator lock. 410 if (enable_stats) { 411 runtime->SetStatsEnabled(true); 412 } 413} 414 415void Trace::StopTracing(bool finish_tracing, bool flush_file) { 416 bool stop_alloc_counting = false; 417 Runtime* const runtime = Runtime::Current(); 418 Trace* the_trace = nullptr; 419 pthread_t sampling_pthread = 0U; 420 { 421 MutexLock mu(Thread::Current(), *Locks::trace_lock_); 422 if (the_trace_ == nullptr) { 423 LOG(ERROR) << "Trace stop requested, but no trace currently running"; 424 } else { 425 the_trace = the_trace_; 426 the_trace_ = nullptr; 427 sampling_pthread = sampling_pthread_; 428 } 429 } 430 // Make sure that we join before we delete the trace since we don't want to have 431 // the sampling thread access a stale pointer. This finishes since the sampling thread exits when 432 // the_trace_ is null. 433 if (sampling_pthread != 0U) { 434 CHECK_PTHREAD_CALL(pthread_join, (sampling_pthread, nullptr), "sampling thread shutdown"); 435 sampling_pthread_ = 0U; 436 } 437 runtime->GetThreadList()->SuspendAll(__FUNCTION__); 438 439 if (the_trace != nullptr) { 440 stop_alloc_counting = (the_trace->flags_ & Trace::kTraceCountAllocs) != 0; 441 if (finish_tracing) { 442 the_trace->FinishTracing(); 443 } 444 445 if (the_trace->trace_mode_ == TraceMode::kSampling) { 446 MutexLock mu(Thread::Current(), *Locks::thread_list_lock_); 447 runtime->GetThreadList()->ForEach(ClearThreadStackTraceAndClockBase, nullptr); 448 } else { 449 runtime->GetInstrumentation()->DisableMethodTracing(kTracerInstrumentationKey); 450 runtime->GetInstrumentation()->RemoveListener( 451 the_trace, instrumentation::Instrumentation::kMethodEntered | 452 instrumentation::Instrumentation::kMethodExited | 453 instrumentation::Instrumentation::kMethodUnwind); 454 } 455 if (the_trace->trace_file_.get() != nullptr) { 456 // Do not try to erase, so flush and close explicitly. 457 if (flush_file) { 458 if (the_trace->trace_file_->Flush() != 0) { 459 PLOG(ERROR) << "Could not flush trace file."; 460 } 461 } else { 462 the_trace->trace_file_->MarkUnchecked(); // Do not trigger guard. 463 } 464 if (the_trace->trace_file_->Close() != 0) { 465 PLOG(ERROR) << "Could not close trace file."; 466 } 467 } 468 delete the_trace; 469 } 470 runtime->GetThreadList()->ResumeAll(); 471 if (stop_alloc_counting) { 472 // Can be racy since SetStatsEnabled is not guarded by any locks. 473 runtime->SetStatsEnabled(false); 474 } 475} 476 477void Trace::Abort() { 478 // Do not write anything anymore. 479 StopTracing(false, false); 480} 481 482void Trace::Stop() { 483 // Finish writing. 484 StopTracing(true, true); 485} 486 487void Trace::Shutdown() { 488 if (GetMethodTracingMode() != kTracingInactive) { 489 Stop(); 490 } 491} 492 493void Trace::Pause() { 494 bool stop_alloc_counting = false; 495 Runtime* runtime = Runtime::Current(); 496 Trace* the_trace = nullptr; 497 498 pthread_t sampling_pthread = 0U; 499 { 500 MutexLock mu(Thread::Current(), *Locks::trace_lock_); 501 if (the_trace_ == nullptr) { 502 LOG(ERROR) << "Trace pause requested, but no trace currently running"; 503 return; 504 } else { 505 the_trace = the_trace_; 506 sampling_pthread = sampling_pthread_; 507 } 508 } 509 510 if (sampling_pthread != 0U) { 511 { 512 MutexLock mu(Thread::Current(), *Locks::trace_lock_); 513 the_trace_ = nullptr; 514 } 515 CHECK_PTHREAD_CALL(pthread_join, (sampling_pthread, nullptr), "sampling thread shutdown"); 516 sampling_pthread_ = 0U; 517 { 518 MutexLock mu(Thread::Current(), *Locks::trace_lock_); 519 the_trace_ = the_trace; 520 } 521 } 522 523 if (the_trace != nullptr) { 524 runtime->GetThreadList()->SuspendAll(__FUNCTION__); 525 stop_alloc_counting = (the_trace->flags_ & Trace::kTraceCountAllocs) != 0; 526 527 if (the_trace->trace_mode_ == TraceMode::kSampling) { 528 MutexLock mu(Thread::Current(), *Locks::thread_list_lock_); 529 runtime->GetThreadList()->ForEach(ClearThreadStackTraceAndClockBase, nullptr); 530 } else { 531 runtime->GetInstrumentation()->DisableMethodTracing(kTracerInstrumentationKey); 532 runtime->GetInstrumentation()->RemoveListener(the_trace, 533 instrumentation::Instrumentation::kMethodEntered | 534 instrumentation::Instrumentation::kMethodExited | 535 instrumentation::Instrumentation::kMethodUnwind); 536 } 537 runtime->GetThreadList()->ResumeAll(); 538 } 539 540 if (stop_alloc_counting) { 541 // Can be racy since SetStatsEnabled is not guarded by any locks. 542 Runtime::Current()->SetStatsEnabled(false); 543 } 544} 545 546void Trace::Resume() { 547 Thread* self = Thread::Current(); 548 Trace* the_trace; 549 { 550 MutexLock mu(self, *Locks::trace_lock_); 551 if (the_trace_ == nullptr) { 552 LOG(ERROR) << "No trace to resume (or sampling mode), ignoring this request"; 553 return; 554 } 555 the_trace = the_trace_; 556 } 557 558 Runtime* runtime = Runtime::Current(); 559 560 // Enable count of allocs if specified in the flags. 561 bool enable_stats = (the_trace->flags_ && kTraceCountAllocs) != 0; 562 563 runtime->GetThreadList()->SuspendAll(__FUNCTION__); 564 565 // Reenable. 566 if (the_trace->trace_mode_ == TraceMode::kSampling) { 567 CHECK_PTHREAD_CALL(pthread_create, (&sampling_pthread_, nullptr, &RunSamplingThread, 568 reinterpret_cast<void*>(the_trace->interval_us_)), "Sampling profiler thread"); 569 } else { 570 runtime->GetInstrumentation()->AddListener(the_trace, 571 instrumentation::Instrumentation::kMethodEntered | 572 instrumentation::Instrumentation::kMethodExited | 573 instrumentation::Instrumentation::kMethodUnwind); 574 // TODO: In full-PIC mode, we don't need to fully deopt. 575 runtime->GetInstrumentation()->EnableMethodTracing(kTracerInstrumentationKey); 576 } 577 578 runtime->GetThreadList()->ResumeAll(); 579 580 // Can't call this when holding the mutator lock. 581 if (enable_stats) { 582 runtime->SetStatsEnabled(true); 583 } 584} 585 586TracingMode Trace::GetMethodTracingMode() { 587 MutexLock mu(Thread::Current(), *Locks::trace_lock_); 588 if (the_trace_ == nullptr) { 589 return kTracingInactive; 590 } else { 591 switch (the_trace_->trace_mode_) { 592 case TraceMode::kSampling: 593 return kSampleProfilingActive; 594 case TraceMode::kMethodTracing: 595 return kMethodTracingActive; 596 } 597 LOG(FATAL) << "Unreachable"; 598 UNREACHABLE(); 599 } 600} 601 602static constexpr size_t kMinBufSize = 18U; // Trace header is up to 18B. 603 604Trace::Trace(File* trace_file, const char* trace_name, size_t buffer_size, int flags, 605 TraceOutputMode output_mode, TraceMode trace_mode) 606 : trace_file_(trace_file), 607 buf_(new uint8_t[std::max(kMinBufSize, buffer_size)]()), 608 flags_(flags), trace_output_mode_(output_mode), trace_mode_(trace_mode), 609 clock_source_(default_clock_source_), 610 buffer_size_(std::max(kMinBufSize, buffer_size)), 611 start_time_(MicroTime()), clock_overhead_ns_(GetClockOverheadNanoSeconds()), cur_offset_(0), 612 overflow_(false), interval_us_(0), streaming_lock_(nullptr) { 613 uint16_t trace_version = GetTraceVersion(clock_source_); 614 if (output_mode == TraceOutputMode::kStreaming) { 615 trace_version |= 0xF0U; 616 } 617 // Set up the beginning of the trace. 618 memset(buf_.get(), 0, kTraceHeaderLength); 619 Append4LE(buf_.get(), kTraceMagicValue); 620 Append2LE(buf_.get() + 4, trace_version); 621 Append2LE(buf_.get() + 6, kTraceHeaderLength); 622 Append8LE(buf_.get() + 8, start_time_); 623 if (trace_version >= kTraceVersionDualClock) { 624 uint16_t record_size = GetRecordSize(clock_source_); 625 Append2LE(buf_.get() + 16, record_size); 626 } 627 static_assert(18 <= kMinBufSize, "Minimum buffer size not large enough for trace header"); 628 629 // Update current offset. 630 cur_offset_.StoreRelaxed(kTraceHeaderLength); 631 632 if (output_mode == TraceOutputMode::kStreaming) { 633 streaming_file_name_ = trace_name; 634 streaming_lock_ = new Mutex("tracing lock"); 635 seen_threads_.reset(new ThreadIDBitSet()); 636 } 637} 638 639Trace::~Trace() { 640 delete streaming_lock_; 641} 642 643static void DumpBuf(uint8_t* buf, size_t buf_size, TraceClockSource clock_source) 644 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 645 uint8_t* ptr = buf + kTraceHeaderLength; 646 uint8_t* end = buf + buf_size; 647 648 while (ptr < end) { 649 uint32_t tmid = ptr[2] | (ptr[3] << 8) | (ptr[4] << 16) | (ptr[5] << 24); 650 mirror::ArtMethod* method = DecodeTraceMethodId(tmid); 651 TraceAction action = DecodeTraceAction(tmid); 652 LOG(INFO) << PrettyMethod(method) << " " << static_cast<int>(action); 653 ptr += GetRecordSize(clock_source); 654 } 655} 656 657static void GetVisitedMethodsFromBitSets( 658 const std::map<mirror::DexCache*, DexIndexBitSet*>& seen_methods, 659 std::set<mirror::ArtMethod*>* visited_methods) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 660 for (auto& e : seen_methods) { 661 DexIndexBitSet* bit_set = e.second; 662 for (uint32_t i = 0; i < bit_set->size(); ++i) { 663 if ((*bit_set)[i]) { 664 visited_methods->insert(e.first->GetResolvedMethod(i)); 665 } 666 } 667 } 668} 669 670void Trace::FinishTracing() { 671 size_t final_offset = 0; 672 673 std::set<mirror::ArtMethod*> visited_methods; 674 if (trace_output_mode_ == TraceOutputMode::kStreaming) { 675 // Write the secondary file with all the method names. 676 GetVisitedMethodsFromBitSets(seen_methods_, &visited_methods); 677 678 // Clean up. 679 for (auto& e : seen_methods_) { 680 delete e.second; 681 } 682 } else { 683 final_offset = cur_offset_.LoadRelaxed(); 684 GetVisitedMethods(final_offset, &visited_methods); 685 } 686 687 // Compute elapsed time. 688 uint64_t elapsed = MicroTime() - start_time_; 689 690 std::ostringstream os; 691 692 os << StringPrintf("%cversion\n", kTraceTokenChar); 693 os << StringPrintf("%d\n", GetTraceVersion(clock_source_)); 694 os << StringPrintf("data-file-overflow=%s\n", overflow_ ? "true" : "false"); 695 if (UseThreadCpuClock()) { 696 if (UseWallClock()) { 697 os << StringPrintf("clock=dual\n"); 698 } else { 699 os << StringPrintf("clock=thread-cpu\n"); 700 } 701 } else { 702 os << StringPrintf("clock=wall\n"); 703 } 704 os << StringPrintf("elapsed-time-usec=%" PRIu64 "\n", elapsed); 705 if (trace_output_mode_ != TraceOutputMode::kStreaming) { 706 size_t num_records = (final_offset - kTraceHeaderLength) / GetRecordSize(clock_source_); 707 os << StringPrintf("num-method-calls=%zd\n", num_records); 708 } 709 os << StringPrintf("clock-call-overhead-nsec=%d\n", clock_overhead_ns_); 710 os << StringPrintf("vm=art\n"); 711 os << StringPrintf("pid=%d\n", getpid()); 712 if ((flags_ & kTraceCountAllocs) != 0) { 713 os << StringPrintf("alloc-count=%d\n", Runtime::Current()->GetStat(KIND_ALLOCATED_OBJECTS)); 714 os << StringPrintf("alloc-size=%d\n", Runtime::Current()->GetStat(KIND_ALLOCATED_BYTES)); 715 os << StringPrintf("gc-count=%d\n", Runtime::Current()->GetStat(KIND_GC_INVOCATIONS)); 716 } 717 os << StringPrintf("%cthreads\n", kTraceTokenChar); 718 DumpThreadList(os); 719 os << StringPrintf("%cmethods\n", kTraceTokenChar); 720 DumpMethodList(os, visited_methods); 721 os << StringPrintf("%cend\n", kTraceTokenChar); 722 std::string header(os.str()); 723 724 if (trace_output_mode_ == TraceOutputMode::kStreaming) { 725 File file; 726 if (!file.Open(streaming_file_name_ + ".sec", O_CREAT | O_WRONLY)) { 727 LOG(WARNING) << "Could not open secondary trace file!"; 728 return; 729 } 730 if (!file.WriteFully(header.c_str(), header.length())) { 731 file.Erase(); 732 std::string detail(StringPrintf("Trace data write failed: %s", strerror(errno))); 733 PLOG(ERROR) << detail; 734 ThrowRuntimeException("%s", detail.c_str()); 735 } 736 if (file.FlushCloseOrErase() != 0) { 737 PLOG(ERROR) << "Could not write secondary file"; 738 } 739 } else { 740 if (trace_file_.get() == nullptr) { 741 iovec iov[2]; 742 iov[0].iov_base = reinterpret_cast<void*>(const_cast<char*>(header.c_str())); 743 iov[0].iov_len = header.length(); 744 iov[1].iov_base = buf_.get(); 745 iov[1].iov_len = final_offset; 746 Dbg::DdmSendChunkV(CHUNK_TYPE("MPSE"), iov, 2); 747 const bool kDumpTraceInfo = false; 748 if (kDumpTraceInfo) { 749 LOG(INFO) << "Trace sent:\n" << header; 750 DumpBuf(buf_.get(), final_offset, clock_source_); 751 } 752 } else { 753 if (!trace_file_->WriteFully(header.c_str(), header.length()) || 754 !trace_file_->WriteFully(buf_.get(), final_offset)) { 755 std::string detail(StringPrintf("Trace data write failed: %s", strerror(errno))); 756 PLOG(ERROR) << detail; 757 ThrowRuntimeException("%s", detail.c_str()); 758 } 759 } 760 } 761} 762 763void Trace::DexPcMoved(Thread* thread, mirror::Object* this_object, 764 mirror::ArtMethod* method, uint32_t new_dex_pc) { 765 UNUSED(thread, this_object, method, new_dex_pc); 766 // We're not recorded to listen to this kind of event, so complain. 767 LOG(ERROR) << "Unexpected dex PC event in tracing " << PrettyMethod(method) << " " << new_dex_pc; 768} 769 770void Trace::FieldRead(Thread* thread, mirror::Object* this_object, 771 mirror::ArtMethod* method, uint32_t dex_pc, ArtField* field) 772 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 773 UNUSED(thread, this_object, method, dex_pc, field); 774 // We're not recorded to listen to this kind of event, so complain. 775 LOG(ERROR) << "Unexpected field read event in tracing " << PrettyMethod(method) << " " << dex_pc; 776} 777 778void Trace::FieldWritten(Thread* thread, mirror::Object* this_object, 779 mirror::ArtMethod* method, uint32_t dex_pc, ArtField* field, 780 const JValue& field_value) 781 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 782 UNUSED(thread, this_object, method, dex_pc, field, field_value); 783 // We're not recorded to listen to this kind of event, so complain. 784 LOG(ERROR) << "Unexpected field write event in tracing " << PrettyMethod(method) << " " << dex_pc; 785} 786 787void Trace::MethodEntered(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED, 788 mirror::ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED) { 789 uint32_t thread_clock_diff = 0; 790 uint32_t wall_clock_diff = 0; 791 ReadClocks(thread, &thread_clock_diff, &wall_clock_diff); 792 LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodEntered, 793 thread_clock_diff, wall_clock_diff); 794} 795 796void Trace::MethodExited(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED, 797 mirror::ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED, 798 const JValue& return_value ATTRIBUTE_UNUSED) { 799 uint32_t thread_clock_diff = 0; 800 uint32_t wall_clock_diff = 0; 801 ReadClocks(thread, &thread_clock_diff, &wall_clock_diff); 802 LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodExited, 803 thread_clock_diff, wall_clock_diff); 804} 805 806void Trace::MethodUnwind(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED, 807 mirror::ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED) { 808 uint32_t thread_clock_diff = 0; 809 uint32_t wall_clock_diff = 0; 810 ReadClocks(thread, &thread_clock_diff, &wall_clock_diff); 811 LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodUnwind, 812 thread_clock_diff, wall_clock_diff); 813} 814 815void Trace::ExceptionCaught(Thread* thread, mirror::Throwable* exception_object) 816 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 817 UNUSED(thread, exception_object); 818 LOG(ERROR) << "Unexpected exception caught event in tracing"; 819} 820 821void Trace::BackwardBranch(Thread* /*thread*/, mirror::ArtMethod* method, 822 int32_t /*dex_pc_offset*/) 823 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 824 LOG(ERROR) << "Unexpected backward branch event in tracing" << PrettyMethod(method); 825} 826 827void Trace::ReadClocks(Thread* thread, uint32_t* thread_clock_diff, uint32_t* wall_clock_diff) { 828 if (UseThreadCpuClock()) { 829 uint64_t clock_base = thread->GetTraceClockBase(); 830 if (UNLIKELY(clock_base == 0)) { 831 // First event, record the base time in the map. 832 uint64_t time = thread->GetCpuMicroTime(); 833 thread->SetTraceClockBase(time); 834 } else { 835 *thread_clock_diff = thread->GetCpuMicroTime() - clock_base; 836 } 837 } 838 if (UseWallClock()) { 839 *wall_clock_diff = MicroTime() - start_time_; 840 } 841} 842 843bool Trace::RegisterMethod(mirror::ArtMethod* method) { 844 mirror::DexCache* dex_cache = method->GetDexCache(); 845 if (dex_cache->GetResolvedMethod(method->GetDexMethodIndex()) != method) { 846 DCHECK(dex_cache->GetResolvedMethod(method->GetDexMethodIndex()) == nullptr); 847 dex_cache->SetResolvedMethod(method->GetDexMethodIndex(), method); 848 } 849 if (seen_methods_.find(dex_cache) == seen_methods_.end()) { 850 seen_methods_.insert(std::make_pair(dex_cache, new DexIndexBitSet())); 851 } 852 DexIndexBitSet* bit_set = seen_methods_.find(dex_cache)->second; 853 if (!(*bit_set)[method->GetDexMethodIndex()]) { 854 bit_set->set(method->GetDexMethodIndex()); 855 return true; 856 } 857 return false; 858} 859 860bool Trace::RegisterThread(Thread* thread) { 861 pid_t tid = thread->GetTid(); 862 CHECK_LT(0U, static_cast<uint32_t>(tid)); 863 CHECK_LT(static_cast<uint32_t>(tid), 65536U); 864 865 if (!(*seen_threads_)[tid]) { 866 seen_threads_->set(tid); 867 return true; 868 } 869 return false; 870} 871 872static std::string GetMethodLine(mirror::ArtMethod* method) 873 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 874 return StringPrintf("%p\t%s\t%s\t%s\t%s\n", method, 875 PrettyDescriptor(method->GetDeclaringClassDescriptor()).c_str(), method->GetName(), 876 method->GetSignature().ToString().c_str(), method->GetDeclaringClassSourceFile()); 877} 878 879void Trace::WriteToBuf(const uint8_t* src, size_t src_size) { 880 int32_t old_offset = cur_offset_.LoadRelaxed(); 881 int32_t new_offset = old_offset + static_cast<int32_t>(src_size); 882 if (dchecked_integral_cast<size_t>(new_offset) > buffer_size_) { 883 // Flush buffer. 884 if (!trace_file_->WriteFully(buf_.get(), old_offset)) { 885 PLOG(WARNING) << "Failed streaming a tracing event."; 886 } 887 888 // Check whether the data is too large for the buffer, then write immediately. 889 if (src_size >= buffer_size_) { 890 if (!trace_file_->WriteFully(src, src_size)) { 891 PLOG(WARNING) << "Failed streaming a tracing event."; 892 } 893 cur_offset_.StoreRelease(0); // Buffer is empty now. 894 return; 895 } 896 897 old_offset = 0; 898 new_offset = static_cast<int32_t>(src_size); 899 } 900 cur_offset_.StoreRelease(new_offset); 901 // Fill in data. 902 memcpy(buf_.get() + old_offset, src, src_size); 903} 904 905void Trace::LogMethodTraceEvent(Thread* thread, mirror::ArtMethod* method, 906 instrumentation::Instrumentation::InstrumentationEvent event, 907 uint32_t thread_clock_diff, uint32_t wall_clock_diff) { 908 // Advance cur_offset_ atomically. 909 int32_t new_offset; 910 int32_t old_offset = 0; 911 912 // We do a busy loop here trying to acquire the next offset. 913 if (trace_output_mode_ != TraceOutputMode::kStreaming) { 914 do { 915 old_offset = cur_offset_.LoadRelaxed(); 916 new_offset = old_offset + GetRecordSize(clock_source_); 917 if (static_cast<size_t>(new_offset) > buffer_size_) { 918 overflow_ = true; 919 return; 920 } 921 } while (!cur_offset_.CompareExchangeWeakSequentiallyConsistent(old_offset, new_offset)); 922 } 923 924 TraceAction action = kTraceMethodEnter; 925 switch (event) { 926 case instrumentation::Instrumentation::kMethodEntered: 927 action = kTraceMethodEnter; 928 break; 929 case instrumentation::Instrumentation::kMethodExited: 930 action = kTraceMethodExit; 931 break; 932 case instrumentation::Instrumentation::kMethodUnwind: 933 action = kTraceUnroll; 934 break; 935 default: 936 UNIMPLEMENTED(FATAL) << "Unexpected event: " << event; 937 } 938 939 uint32_t method_value = EncodeTraceMethodAndAction(method, action); 940 941 // Write data 942 uint8_t* ptr; 943 static constexpr size_t kPacketSize = 14U; // The maximum size of data in a packet. 944 uint8_t stack_buf[kPacketSize]; // Space to store a packet when in streaming mode. 945 if (trace_output_mode_ == TraceOutputMode::kStreaming) { 946 ptr = stack_buf; 947 } else { 948 ptr = buf_.get() + old_offset; 949 } 950 951 Append2LE(ptr, thread->GetTid()); 952 Append4LE(ptr + 2, method_value); 953 ptr += 6; 954 955 if (UseThreadCpuClock()) { 956 Append4LE(ptr, thread_clock_diff); 957 ptr += 4; 958 } 959 if (UseWallClock()) { 960 Append4LE(ptr, wall_clock_diff); 961 } 962 static_assert(kPacketSize == 2 + 4 + 4 + 4, "Packet size incorrect."); 963 964 if (trace_output_mode_ == TraceOutputMode::kStreaming) { 965 MutexLock mu(Thread::Current(), *streaming_lock_); // To serialize writing. 966 if (RegisterMethod(method)) { 967 // Write a special block with the name. 968 std::string method_line(GetMethodLine(method)); 969 uint8_t buf2[5]; 970 Append2LE(buf2, 0); 971 buf2[2] = kOpNewMethod; 972 Append2LE(buf2 + 3, static_cast<uint16_t>(method_line.length())); 973 WriteToBuf(buf2, sizeof(buf2)); 974 WriteToBuf(reinterpret_cast<const uint8_t*>(method_line.c_str()), method_line.length()); 975 } 976 if (RegisterThread(thread)) { 977 // It might be better to postpone this. Threads might not have received names... 978 std::string thread_name; 979 thread->GetThreadName(thread_name); 980 uint8_t buf2[7]; 981 Append2LE(buf2, 0); 982 buf2[2] = kOpNewThread; 983 Append2LE(buf2 + 3, static_cast<uint16_t>(thread->GetTid())); 984 Append2LE(buf2 + 5, static_cast<uint16_t>(thread_name.length())); 985 WriteToBuf(buf2, sizeof(buf2)); 986 WriteToBuf(reinterpret_cast<const uint8_t*>(thread_name.c_str()), thread_name.length()); 987 } 988 WriteToBuf(stack_buf, sizeof(stack_buf)); 989 } 990} 991 992void Trace::GetVisitedMethods(size_t buf_size, 993 std::set<mirror::ArtMethod*>* visited_methods) { 994 uint8_t* ptr = buf_.get() + kTraceHeaderLength; 995 uint8_t* end = buf_.get() + buf_size; 996 997 while (ptr < end) { 998 uint32_t tmid = ptr[2] | (ptr[3] << 8) | (ptr[4] << 16) | (ptr[5] << 24); 999 mirror::ArtMethod* method = DecodeTraceMethodId(tmid); 1000 visited_methods->insert(method); 1001 ptr += GetRecordSize(clock_source_); 1002 } 1003} 1004 1005void Trace::DumpMethodList(std::ostream& os, const std::set<mirror::ArtMethod*>& visited_methods) { 1006 for (const auto& method : visited_methods) { 1007 os << GetMethodLine(method); 1008 } 1009} 1010 1011static void DumpThread(Thread* t, void* arg) { 1012 std::ostream& os = *reinterpret_cast<std::ostream*>(arg); 1013 std::string name; 1014 t->GetThreadName(name); 1015 os << t->GetTid() << "\t" << name << "\n"; 1016} 1017 1018void Trace::DumpThreadList(std::ostream& os) { 1019 Thread* self = Thread::Current(); 1020 for (auto it : exited_threads_) { 1021 os << it.first << "\t" << it.second << "\n"; 1022 } 1023 Locks::thread_list_lock_->AssertNotHeld(self); 1024 MutexLock mu(self, *Locks::thread_list_lock_); 1025 Runtime::Current()->GetThreadList()->ForEach(DumpThread, &os); 1026} 1027 1028void Trace::StoreExitingThreadInfo(Thread* thread) { 1029 MutexLock mu(thread, *Locks::trace_lock_); 1030 if (the_trace_ != nullptr) { 1031 std::string name; 1032 thread->GetThreadName(name); 1033 // The same thread/tid may be used multiple times. As SafeMap::Put does not allow to override 1034 // a previous mapping, use SafeMap::Overwrite. 1035 the_trace_->exited_threads_.Overwrite(thread->GetTid(), name); 1036 } 1037} 1038 1039Trace::TraceOutputMode Trace::GetOutputMode() { 1040 MutexLock mu(Thread::Current(), *Locks::trace_lock_); 1041 CHECK(the_trace_ != nullptr) << "Trace output mode requested, but no trace currently running"; 1042 return the_trace_->trace_output_mode_; 1043} 1044 1045Trace::TraceMode Trace::GetMode() { 1046 MutexLock mu(Thread::Current(), *Locks::trace_lock_); 1047 CHECK(the_trace_ != nullptr) << "Trace mode requested, but no trace currently running"; 1048 return the_trace_->trace_mode_; 1049} 1050 1051size_t Trace::GetBufferSize() { 1052 MutexLock mu(Thread::Current(), *Locks::trace_lock_); 1053 CHECK(the_trace_ != nullptr) << "Trace mode requested, but no trace currently running"; 1054 return the_trace_->buffer_size_; 1055} 1056 1057} // namespace art 1058