trace.cc revision 3d21bdf8894e780d349c481e5c9e29fe1556051c 
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 "art_method-inl.h"
26#include "base/casts.h"
27#include "base/stl_util.h"
28#include "base/time_utils.h"
29#include "base/unix_file/fd_file.h"
30#include "class_linker.h"
31#include "common_throws.h"
32#include "debugger.h"
33#include "dex_file-inl.h"
34#include "instrumentation.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    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<ArtMethod*>* GetStackTrace() const {
111    return method_trace_;
112  }
113
114 private:
115  std::vector<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 = 14;  // using v2
124static const uint16_t kTraceRecordSizeDualClock   = 18;  // 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<ArtMethod*>> Trace::temp_stack_trace_;
131
132// The key identifying the tracer to update instrumentation.
133static constexpr const char* kTracerInstrumentationKey = "Tracer";
134
135static ArtMethod* DecodeTraceMethodId(uint64_t tmid) {
136  return reinterpret_cast<ArtMethod*>(tmid & ~kTraceMethodActionMask);
137}
138
139static TraceAction DecodeTraceAction(uint32_t tmid) {
140  return static_cast<TraceAction>(tmid & kTraceMethodActionMask);
141}
142
143static uint64_t EncodeTraceMethodAndAction(ArtMethod* method, TraceAction action) {
144  auto tmid = static_cast<uint64_t>(reinterpret_cast<uintptr_t>(method)) | action;
145  DCHECK_EQ(method, DecodeTraceMethodId(tmid));
146  return tmid;
147}
148
149std::vector<ArtMethod*>* Trace::AllocStackTrace() {
150  if (temp_stack_trace_.get() != nullptr) {
151    return temp_stack_trace_.release();
152  } else {
153    return new std::vector<ArtMethod*>();
154  }
155}
156
157void Trace::FreeStackTrace(std::vector<ArtMethod*>* stack_trace) {
158  stack_trace->clear();
159  temp_stack_trace_.reset(stack_trace);
160}
161
162void Trace::SetDefaultClockSource(TraceClockSource clock_source) {
163#if defined(__linux__)
164  default_clock_source_ = clock_source;
165#else
166  if (clock_source != TraceClockSource::kWall) {
167    LOG(WARNING) << "Ignoring tracing request to use CPU time.";
168  }
169#endif
170}
171
172static uint16_t GetTraceVersion(TraceClockSource clock_source) {
173  return (clock_source == TraceClockSource::kDual) ? kTraceVersionDualClock
174                                                    : kTraceVersionSingleClock;
175}
176
177static uint16_t GetRecordSize(TraceClockSource clock_source) {
178  return (clock_source == TraceClockSource::kDual) ? kTraceRecordSizeDualClock
179                                                    : kTraceRecordSizeSingleClock;
180}
181
182bool Trace::UseThreadCpuClock() {
183  return (clock_source_ == TraceClockSource::kThreadCpu) ||
184      (clock_source_ == TraceClockSource::kDual);
185}
186
187bool Trace::UseWallClock() {
188  return (clock_source_ == TraceClockSource::kWall) ||
189      (clock_source_ == TraceClockSource::kDual);
190}
191
192void Trace::MeasureClockOverhead() {
193  if (UseThreadCpuClock()) {
194    Thread::Current()->GetCpuMicroTime();
195  }
196  if (UseWallClock()) {
197    MicroTime();
198  }
199}
200
201// Compute an average time taken to measure clocks.
202uint32_t Trace::GetClockOverheadNanoSeconds() {
203  Thread* self = Thread::Current();
204  uint64_t start = self->GetCpuMicroTime();
205
206  for (int i = 4000; i > 0; i--) {
207    MeasureClockOverhead();
208    MeasureClockOverhead();
209    MeasureClockOverhead();
210    MeasureClockOverhead();
211    MeasureClockOverhead();
212    MeasureClockOverhead();
213    MeasureClockOverhead();
214    MeasureClockOverhead();
215  }
216
217  uint64_t elapsed_us = self->GetCpuMicroTime() - start;
218  return static_cast<uint32_t>(elapsed_us / 32);
219}
220
221// TODO: put this somewhere with the big-endian equivalent used by JDWP.
222static void Append2LE(uint8_t* buf, uint16_t val) {
223  *buf++ = static_cast<uint8_t>(val);
224  *buf++ = static_cast<uint8_t>(val >> 8);
225}
226
227// TODO: put this somewhere with the big-endian equivalent used by JDWP.
228static void Append4LE(uint8_t* buf, uint32_t val) {
229  *buf++ = static_cast<uint8_t>(val);
230  *buf++ = static_cast<uint8_t>(val >> 8);
231  *buf++ = static_cast<uint8_t>(val >> 16);
232  *buf++ = static_cast<uint8_t>(val >> 24);
233}
234
235// TODO: put this somewhere with the big-endian equivalent used by JDWP.
236static void Append8LE(uint8_t* buf, uint64_t val) {
237  *buf++ = static_cast<uint8_t>(val);
238  *buf++ = static_cast<uint8_t>(val >> 8);
239  *buf++ = static_cast<uint8_t>(val >> 16);
240  *buf++ = static_cast<uint8_t>(val >> 24);
241  *buf++ = static_cast<uint8_t>(val >> 32);
242  *buf++ = static_cast<uint8_t>(val >> 40);
243  *buf++ = static_cast<uint8_t>(val >> 48);
244  *buf++ = static_cast<uint8_t>(val >> 56);
245}
246
247static void GetSample(Thread* thread, void* arg) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
248  BuildStackTraceVisitor build_trace_visitor(thread);
249  build_trace_visitor.WalkStack();
250  std::vector<ArtMethod*>* stack_trace = build_trace_visitor.GetStackTrace();
251  Trace* the_trace = reinterpret_cast<Trace*>(arg);
252  the_trace->CompareAndUpdateStackTrace(thread, stack_trace);
253}
254
255static void ClearThreadStackTraceAndClockBase(Thread* thread, void* arg ATTRIBUTE_UNUSED) {
256  thread->SetTraceClockBase(0);
257  std::vector<ArtMethod*>* stack_trace = thread->GetStackTraceSample();
258  thread->SetStackTraceSample(nullptr);
259  delete stack_trace;
260}
261
262void Trace::CompareAndUpdateStackTrace(Thread* thread,
263                                       std::vector<ArtMethod*>* stack_trace) {
264  CHECK_EQ(pthread_self(), sampling_pthread_);
265  std::vector<ArtMethod*>* old_stack_trace = thread->GetStackTraceSample();
266  // Update the thread's stack trace sample.
267  thread->SetStackTraceSample(stack_trace);
268  // Read timer clocks to use for all events in this trace.
269  uint32_t thread_clock_diff = 0;
270  uint32_t wall_clock_diff = 0;
271  ReadClocks(thread, &thread_clock_diff, &wall_clock_diff);
272  if (old_stack_trace == nullptr) {
273    // If there's no previous stack trace sample for this thread, log an entry event for all
274    // methods in the trace.
275    for (std::vector<ArtMethod*>::reverse_iterator rit = stack_trace->rbegin();
276         rit != stack_trace->rend(); ++rit) {
277      LogMethodTraceEvent(thread, *rit, instrumentation::Instrumentation::kMethodEntered,
278                          thread_clock_diff, wall_clock_diff);
279    }
280  } else {
281    // If there's a previous stack trace for this thread, diff the traces and emit entry and exit
282    // events accordingly.
283    std::vector<ArtMethod*>::reverse_iterator old_rit = old_stack_trace->rbegin();
284    std::vector<ArtMethod*>::reverse_iterator rit = stack_trace->rbegin();
285    // Iterate bottom-up over both traces until there's a difference between them.
286    while (old_rit != old_stack_trace->rend() && rit != stack_trace->rend() && *old_rit == *rit) {
287      old_rit++;
288      rit++;
289    }
290    // Iterate top-down over the old trace until the point where they differ, emitting exit events.
291    for (std::vector<ArtMethod*>::iterator old_it = old_stack_trace->begin();
292         old_it != old_rit.base(); ++old_it) {
293      LogMethodTraceEvent(thread, *old_it, instrumentation::Instrumentation::kMethodExited,
294                          thread_clock_diff, wall_clock_diff);
295    }
296    // Iterate bottom-up over the new trace from the point where they differ, emitting entry events.
297    for (; rit != stack_trace->rend(); ++rit) {
298      LogMethodTraceEvent(thread, *rit, instrumentation::Instrumentation::kMethodEntered,
299                          thread_clock_diff, wall_clock_diff);
300    }
301    FreeStackTrace(old_stack_trace);
302  }
303}
304
305void* Trace::RunSamplingThread(void* arg) {
306  Runtime* runtime = Runtime::Current();
307  intptr_t interval_us = reinterpret_cast<intptr_t>(arg);
308  CHECK_GE(interval_us, 0);
309  CHECK(runtime->AttachCurrentThread("Sampling Profiler", true, runtime->GetSystemThreadGroup(),
310                                     !runtime->IsAotCompiler()));
311
312  while (true) {
313    usleep(interval_us);
314    ATRACE_BEGIN("Profile sampling");
315    Thread* self = Thread::Current();
316    Trace* the_trace;
317    {
318      MutexLock mu(self, *Locks::trace_lock_);
319      the_trace = the_trace_;
320      if (the_trace == nullptr) {
321        break;
322      }
323    }
324
325    runtime->GetThreadList()->SuspendAll(__FUNCTION__);
326    {
327      MutexLock mu(self, *Locks::thread_list_lock_);
328      runtime->GetThreadList()->ForEach(GetSample, the_trace);
329    }
330    runtime->GetThreadList()->ResumeAll();
331    ATRACE_END();
332  }
333
334  runtime->DetachCurrentThread();
335  return nullptr;
336}
337
338void Trace::Start(const char* trace_filename, int trace_fd, size_t buffer_size, int flags,
339                  TraceOutputMode output_mode, TraceMode trace_mode, int interval_us) {
340  Thread* self = Thread::Current();
341  {
342    MutexLock mu(self, *Locks::trace_lock_);
343    if (the_trace_ != nullptr) {
344      LOG(ERROR) << "Trace already in progress, ignoring this request";
345      return;
346    }
347  }
348
349  // Check interval if sampling is enabled
350  if (trace_mode == TraceMode::kSampling && interval_us <= 0) {
351    LOG(ERROR) << "Invalid sampling interval: " << interval_us;
352    ScopedObjectAccess soa(self);
353    ThrowRuntimeException("Invalid sampling interval: %d", interval_us);
354    return;
355  }
356
357  // Open trace file if not going directly to ddms.
358  std::unique_ptr<File> trace_file;
359  if (output_mode != TraceOutputMode::kDDMS) {
360    if (trace_fd < 0) {
361      trace_file.reset(OS::CreateEmptyFile(trace_filename));
362    } else {
363      trace_file.reset(new File(trace_fd, "tracefile"));
364      trace_file->DisableAutoClose();
365    }
366    if (trace_file.get() == nullptr) {
367      PLOG(ERROR) << "Unable to open trace file '" << trace_filename << "'";
368      ScopedObjectAccess soa(self);
369      ThrowRuntimeException("Unable to open trace file '%s'", trace_filename);
370      return;
371    }
372  }
373
374  Runtime* runtime = Runtime::Current();
375
376  // Enable count of allocs if specified in the flags.
377  bool enable_stats = false;
378
379  runtime->GetThreadList()->SuspendAll(__FUNCTION__);
380
381  // Create Trace object.
382  {
383    MutexLock mu(self, *Locks::trace_lock_);
384    if (the_trace_ != nullptr) {
385      LOG(ERROR) << "Trace already in progress, ignoring this request";
386    } else {
387      enable_stats = (flags && kTraceCountAllocs) != 0;
388      the_trace_ = new Trace(trace_file.release(), trace_filename, buffer_size, flags, output_mode,
389                             trace_mode);
390      if (trace_mode == TraceMode::kSampling) {
391        CHECK_PTHREAD_CALL(pthread_create, (&sampling_pthread_, nullptr, &RunSamplingThread,
392                                            reinterpret_cast<void*>(interval_us)),
393                                            "Sampling profiler thread");
394        the_trace_->interval_us_ = interval_us;
395      } else {
396        runtime->GetInstrumentation()->AddListener(the_trace_,
397                                                   instrumentation::Instrumentation::kMethodEntered |
398                                                   instrumentation::Instrumentation::kMethodExited |
399                                                   instrumentation::Instrumentation::kMethodUnwind);
400        // TODO: In full-PIC mode, we don't need to fully deopt.
401        runtime->GetInstrumentation()->EnableMethodTracing(kTracerInstrumentationKey);
402      }
403    }
404  }
405
406  runtime->GetThreadList()->ResumeAll();
407
408  // Can't call this when holding the mutator lock.
409  if (enable_stats) {
410    runtime->SetStatsEnabled(true);
411  }
412}
413
414void Trace::StopTracing(bool finish_tracing, bool flush_file) {
415  bool stop_alloc_counting = false;
416  Runtime* const runtime = Runtime::Current();
417  Trace* the_trace = nullptr;
418  pthread_t sampling_pthread = 0U;
419  {
420    MutexLock mu(Thread::Current(), *Locks::trace_lock_);
421    if (the_trace_ == nullptr) {
422      LOG(ERROR) << "Trace stop requested, but no trace currently running";
423    } else {
424      the_trace = the_trace_;
425      the_trace_ = nullptr;
426      sampling_pthread = sampling_pthread_;
427    }
428  }
429  // Make sure that we join before we delete the trace since we don't want to have
430  // the sampling thread access a stale pointer. This finishes since the sampling thread exits when
431  // the_trace_ is null.
432  if (sampling_pthread != 0U) {
433    CHECK_PTHREAD_CALL(pthread_join, (sampling_pthread, nullptr), "sampling thread shutdown");
434    sampling_pthread_ = 0U;
435  }
436  runtime->GetThreadList()->SuspendAll(__FUNCTION__);
437
438  if (the_trace != nullptr) {
439    stop_alloc_counting = (the_trace->flags_ & Trace::kTraceCountAllocs) != 0;
440    if (finish_tracing) {
441      the_trace->FinishTracing();
442    }
443
444    if (the_trace->trace_mode_ == TraceMode::kSampling) {
445      MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
446      runtime->GetThreadList()->ForEach(ClearThreadStackTraceAndClockBase, nullptr);
447    } else {
448      runtime->GetInstrumentation()->DisableMethodTracing(kTracerInstrumentationKey);
449      runtime->GetInstrumentation()->RemoveListener(
450          the_trace, instrumentation::Instrumentation::kMethodEntered |
451          instrumentation::Instrumentation::kMethodExited |
452          instrumentation::Instrumentation::kMethodUnwind);
453    }
454    if (the_trace->trace_file_.get() != nullptr) {
455      // Do not try to erase, so flush and close explicitly.
456      if (flush_file) {
457        if (the_trace->trace_file_->Flush() != 0) {
458          PLOG(ERROR) << "Could not flush trace file.";
459        }
460      } else {
461        the_trace->trace_file_->MarkUnchecked();  // Do not trigger guard.
462      }
463      if (the_trace->trace_file_->Close() != 0) {
464        PLOG(ERROR) << "Could not close trace file.";
465      }
466    }
467    delete the_trace;
468  }
469  runtime->GetThreadList()->ResumeAll();
470  if (stop_alloc_counting) {
471    // Can be racy since SetStatsEnabled is not guarded by any locks.
472    runtime->SetStatsEnabled(false);
473  }
474}
475
476void Trace::Abort() {
477  // Do not write anything anymore.
478  StopTracing(false, false);
479}
480
481void Trace::Stop() {
482  // Finish writing.
483  StopTracing(true, true);
484}
485
486void Trace::Shutdown() {
487  if (GetMethodTracingMode() != kTracingInactive) {
488    Stop();
489  }
490}
491
492void Trace::Pause() {
493  bool stop_alloc_counting = false;
494  Runtime* runtime = Runtime::Current();
495  Trace* the_trace = nullptr;
496
497  pthread_t sampling_pthread = 0U;
498  {
499    MutexLock mu(Thread::Current(), *Locks::trace_lock_);
500    if (the_trace_ == nullptr) {
501      LOG(ERROR) << "Trace pause requested, but no trace currently running";
502      return;
503    } else {
504      the_trace = the_trace_;
505      sampling_pthread = sampling_pthread_;
506    }
507  }
508
509  if (sampling_pthread != 0U) {
510    {
511      MutexLock mu(Thread::Current(), *Locks::trace_lock_);
512      the_trace_ = nullptr;
513    }
514    CHECK_PTHREAD_CALL(pthread_join, (sampling_pthread, nullptr), "sampling thread shutdown");
515    sampling_pthread_ = 0U;
516    {
517      MutexLock mu(Thread::Current(), *Locks::trace_lock_);
518      the_trace_ = the_trace;
519    }
520  }
521
522  if (the_trace != nullptr) {
523    runtime->GetThreadList()->SuspendAll(__FUNCTION__);
524    stop_alloc_counting = (the_trace->flags_ & Trace::kTraceCountAllocs) != 0;
525
526    if (the_trace->trace_mode_ == TraceMode::kSampling) {
527      MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
528      runtime->GetThreadList()->ForEach(ClearThreadStackTraceAndClockBase, nullptr);
529    } else {
530      runtime->GetInstrumentation()->DisableMethodTracing(kTracerInstrumentationKey);
531      runtime->GetInstrumentation()->RemoveListener(the_trace,
532                                                    instrumentation::Instrumentation::kMethodEntered |
533                                                    instrumentation::Instrumentation::kMethodExited |
534                                                    instrumentation::Instrumentation::kMethodUnwind);
535    }
536    runtime->GetThreadList()->ResumeAll();
537  }
538
539  if (stop_alloc_counting) {
540    // Can be racy since SetStatsEnabled is not guarded by any locks.
541    Runtime::Current()->SetStatsEnabled(false);
542  }
543}
544
545void Trace::Resume() {
546  Thread* self = Thread::Current();
547  Trace* the_trace;
548  {
549    MutexLock mu(self, *Locks::trace_lock_);
550    if (the_trace_ == nullptr) {
551      LOG(ERROR) << "No trace to resume (or sampling mode), ignoring this request";
552      return;
553    }
554    the_trace = the_trace_;
555  }
556
557  Runtime* runtime = Runtime::Current();
558
559  // Enable count of allocs if specified in the flags.
560  bool enable_stats = (the_trace->flags_ && kTraceCountAllocs) != 0;
561
562  runtime->GetThreadList()->SuspendAll(__FUNCTION__);
563
564  // Reenable.
565  if (the_trace->trace_mode_ == TraceMode::kSampling) {
566    CHECK_PTHREAD_CALL(pthread_create, (&sampling_pthread_, nullptr, &RunSamplingThread,
567        reinterpret_cast<void*>(the_trace->interval_us_)), "Sampling profiler thread");
568  } else {
569    runtime->GetInstrumentation()->AddListener(the_trace,
570                                               instrumentation::Instrumentation::kMethodEntered |
571                                               instrumentation::Instrumentation::kMethodExited |
572                                               instrumentation::Instrumentation::kMethodUnwind);
573    // TODO: In full-PIC mode, we don't need to fully deopt.
574    runtime->GetInstrumentation()->EnableMethodTracing(kTracerInstrumentationKey);
575  }
576
577  runtime->GetThreadList()->ResumeAll();
578
579  // Can't call this when holding the mutator lock.
580  if (enable_stats) {
581    runtime->SetStatsEnabled(true);
582  }
583}
584
585TracingMode Trace::GetMethodTracingMode() {
586  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
587  if (the_trace_ == nullptr) {
588    return kTracingInactive;
589  } else {
590    switch (the_trace_->trace_mode_) {
591      case TraceMode::kSampling:
592        return kSampleProfilingActive;
593      case TraceMode::kMethodTracing:
594        return kMethodTracingActive;
595    }
596    LOG(FATAL) << "Unreachable";
597    UNREACHABLE();
598  }
599}
600
601static constexpr size_t kMinBufSize = 18U;  // Trace header is up to 18B.
602
603Trace::Trace(File* trace_file, const char* trace_name, size_t buffer_size, int flags,
604             TraceOutputMode output_mode, TraceMode trace_mode)
605    : trace_file_(trace_file),
606      buf_(new uint8_t[std::max(kMinBufSize, buffer_size)]()),
607      flags_(flags), trace_output_mode_(output_mode), trace_mode_(trace_mode),
608      clock_source_(default_clock_source_),
609      buffer_size_(std::max(kMinBufSize, buffer_size)),
610      start_time_(MicroTime()), clock_overhead_ns_(GetClockOverheadNanoSeconds()), cur_offset_(0),
611      overflow_(false), interval_us_(0), streaming_lock_(nullptr) {
612  uint16_t trace_version = GetTraceVersion(clock_source_);
613  if (output_mode == TraceOutputMode::kStreaming) {
614    trace_version |= 0xF0U;
615  }
616  // Set up the beginning of the trace.
617  memset(buf_.get(), 0, kTraceHeaderLength);
618  Append4LE(buf_.get(), kTraceMagicValue);
619  Append2LE(buf_.get() + 4, trace_version);
620  Append2LE(buf_.get() + 6, kTraceHeaderLength);
621  Append8LE(buf_.get() + 8, start_time_);
622  if (trace_version >= kTraceVersionDualClock) {
623    uint16_t record_size = GetRecordSize(clock_source_);
624    Append2LE(buf_.get() + 16, record_size);
625  }
626  static_assert(18 <= kMinBufSize, "Minimum buffer size not large enough for trace header");
627
628  // Update current offset.
629  cur_offset_.StoreRelaxed(kTraceHeaderLength);
630
631  if (output_mode == TraceOutputMode::kStreaming) {
632    streaming_file_name_ = trace_name;
633    streaming_lock_ = new Mutex("tracing lock");
634    seen_threads_.reset(new ThreadIDBitSet());
635  }
636}
637
638Trace::~Trace() {
639  delete streaming_lock_;
640}
641
642static uint64_t ReadBytes(uint8_t* buf, size_t bytes) {
643  uint64_t ret = 0;
644  for (size_t i = 0; i < bytes; ++i) {
645    ret |= static_cast<uint64_t>(buf[i]) << (i * 8);
646  }
647  return ret;
648}
649
650static void DumpBuf(uint8_t* buf, size_t buf_size, TraceClockSource clock_source)
651    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
652  uint8_t* ptr = buf + kTraceHeaderLength;
653  uint8_t* end = buf + buf_size;
654
655  while (ptr < end) {
656    uint64_t tmid = ReadBytes(ptr + 2, sizeof(tmid));
657    ArtMethod* method = DecodeTraceMethodId(tmid);
658    TraceAction action = DecodeTraceAction(tmid);
659    LOG(INFO) << PrettyMethod(method) << " " << static_cast<int>(action);
660    ptr += GetRecordSize(clock_source);
661  }
662}
663
664static void GetVisitedMethodsFromBitSets(
665    const std::map<mirror::DexCache*, DexIndexBitSet*>& seen_methods,
666    std::set<ArtMethod*>* visited_methods) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
667  for (auto& e : seen_methods) {
668    DexIndexBitSet* bit_set = e.second;
669    for (uint32_t i = 0; i < bit_set->size(); ++i) {
670      if ((*bit_set)[i]) {
671        visited_methods->insert(e.first->GetResolvedMethod(i, sizeof(void*)));
672      }
673    }
674  }
675}
676
677void Trace::FinishTracing() {
678  size_t final_offset = 0;
679
680  std::set<ArtMethod*> visited_methods;
681  if (trace_output_mode_ == TraceOutputMode::kStreaming) {
682    // Write the secondary file with all the method names.
683    GetVisitedMethodsFromBitSets(seen_methods_, &visited_methods);
684
685    // Clean up.
686    for (auto& e : seen_methods_) {
687      delete e.second;
688    }
689  } else {
690    final_offset = cur_offset_.LoadRelaxed();
691    GetVisitedMethods(final_offset, &visited_methods);
692  }
693
694  // Compute elapsed time.
695  uint64_t elapsed = MicroTime() - start_time_;
696
697  std::ostringstream os;
698
699  os << StringPrintf("%cversion\n", kTraceTokenChar);
700  os << StringPrintf("%d\n", GetTraceVersion(clock_source_));
701  os << StringPrintf("data-file-overflow=%s\n", overflow_ ? "true" : "false");
702  if (UseThreadCpuClock()) {
703    if (UseWallClock()) {
704      os << StringPrintf("clock=dual\n");
705    } else {
706      os << StringPrintf("clock=thread-cpu\n");
707    }
708  } else {
709    os << StringPrintf("clock=wall\n");
710  }
711  os << StringPrintf("elapsed-time-usec=%" PRIu64 "\n", elapsed);
712  if (trace_output_mode_ != TraceOutputMode::kStreaming) {
713    size_t num_records = (final_offset - kTraceHeaderLength) / GetRecordSize(clock_source_);
714    os << StringPrintf("num-method-calls=%zd\n", num_records);
715  }
716  os << StringPrintf("clock-call-overhead-nsec=%d\n", clock_overhead_ns_);
717  os << StringPrintf("vm=art\n");
718  os << StringPrintf("pid=%d\n", getpid());
719  if ((flags_ & kTraceCountAllocs) != 0) {
720    os << StringPrintf("alloc-count=%d\n", Runtime::Current()->GetStat(KIND_ALLOCATED_OBJECTS));
721    os << StringPrintf("alloc-size=%d\n", Runtime::Current()->GetStat(KIND_ALLOCATED_BYTES));
722    os << StringPrintf("gc-count=%d\n", Runtime::Current()->GetStat(KIND_GC_INVOCATIONS));
723  }
724  os << StringPrintf("%cthreads\n", kTraceTokenChar);
725  DumpThreadList(os);
726  os << StringPrintf("%cmethods\n", kTraceTokenChar);
727  DumpMethodList(os, visited_methods);
728  os << StringPrintf("%cend\n", kTraceTokenChar);
729  std::string header(os.str());
730
731  if (trace_output_mode_ == TraceOutputMode::kStreaming) {
732    File file;
733    if (!file.Open(streaming_file_name_ + ".sec", O_CREAT | O_WRONLY)) {
734      LOG(WARNING) << "Could not open secondary trace file!";
735      return;
736    }
737    if (!file.WriteFully(header.c_str(), header.length())) {
738      file.Erase();
739      std::string detail(StringPrintf("Trace data write failed: %s", strerror(errno)));
740      PLOG(ERROR) << detail;
741      ThrowRuntimeException("%s", detail.c_str());
742    }
743    if (file.FlushCloseOrErase() != 0) {
744      PLOG(ERROR) << "Could not write secondary file";
745    }
746  } else {
747    if (trace_file_.get() == nullptr) {
748      iovec iov[2];
749      iov[0].iov_base = reinterpret_cast<void*>(const_cast<char*>(header.c_str()));
750      iov[0].iov_len = header.length();
751      iov[1].iov_base = buf_.get();
752      iov[1].iov_len = final_offset;
753      Dbg::DdmSendChunkV(CHUNK_TYPE("MPSE"), iov, 2);
754      const bool kDumpTraceInfo = false;
755      if (kDumpTraceInfo) {
756        LOG(INFO) << "Trace sent:\n" << header;
757        DumpBuf(buf_.get(), final_offset, clock_source_);
758      }
759    } else {
760      if (!trace_file_->WriteFully(header.c_str(), header.length()) ||
761          !trace_file_->WriteFully(buf_.get(), final_offset)) {
762        std::string detail(StringPrintf("Trace data write failed: %s", strerror(errno)));
763        PLOG(ERROR) << detail;
764        ThrowRuntimeException("%s", detail.c_str());
765      }
766    }
767  }
768}
769
770void Trace::DexPcMoved(Thread* thread, mirror::Object* this_object,
771                       ArtMethod* method, uint32_t new_dex_pc) {
772  UNUSED(thread, this_object, method, new_dex_pc);
773  // We're not recorded to listen to this kind of event, so complain.
774  LOG(ERROR) << "Unexpected dex PC event in tracing " << PrettyMethod(method) << " " << new_dex_pc;
775}
776
777void Trace::FieldRead(Thread* thread, mirror::Object* this_object,
778                       ArtMethod* method, uint32_t dex_pc, ArtField* field)
779    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
780  UNUSED(thread, this_object, method, dex_pc, field);
781  // We're not recorded to listen to this kind of event, so complain.
782  LOG(ERROR) << "Unexpected field read event in tracing " << PrettyMethod(method) << " " << dex_pc;
783}
784
785void Trace::FieldWritten(Thread* thread, mirror::Object* this_object,
786                          ArtMethod* method, uint32_t dex_pc, ArtField* field,
787                          const JValue& field_value)
788    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
789  UNUSED(thread, this_object, method, dex_pc, field, field_value);
790  // We're not recorded to listen to this kind of event, so complain.
791  LOG(ERROR) << "Unexpected field write event in tracing " << PrettyMethod(method) << " " << dex_pc;
792}
793
794void Trace::MethodEntered(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED,
795                          ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED) {
796  uint32_t thread_clock_diff = 0;
797  uint32_t wall_clock_diff = 0;
798  ReadClocks(thread, &thread_clock_diff, &wall_clock_diff);
799  LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodEntered,
800                      thread_clock_diff, wall_clock_diff);
801}
802
803void Trace::MethodExited(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED,
804                         ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED,
805                         const JValue& return_value ATTRIBUTE_UNUSED) {
806  uint32_t thread_clock_diff = 0;
807  uint32_t wall_clock_diff = 0;
808  ReadClocks(thread, &thread_clock_diff, &wall_clock_diff);
809  LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodExited,
810                      thread_clock_diff, wall_clock_diff);
811}
812
813void Trace::MethodUnwind(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED,
814                         ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED) {
815  uint32_t thread_clock_diff = 0;
816  uint32_t wall_clock_diff = 0;
817  ReadClocks(thread, &thread_clock_diff, &wall_clock_diff);
818  LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodUnwind,
819                      thread_clock_diff, wall_clock_diff);
820}
821
822void Trace::ExceptionCaught(Thread* thread, mirror::Throwable* exception_object)
823    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
824  UNUSED(thread, exception_object);
825  LOG(ERROR) << "Unexpected exception caught event in tracing";
826}
827
828void Trace::BackwardBranch(Thread* /*thread*/, ArtMethod* method,
829                           int32_t /*dex_pc_offset*/)
830      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
831  LOG(ERROR) << "Unexpected backward branch event in tracing" << PrettyMethod(method);
832}
833
834void Trace::ReadClocks(Thread* thread, uint32_t* thread_clock_diff, uint32_t* wall_clock_diff) {
835  if (UseThreadCpuClock()) {
836    uint64_t clock_base = thread->GetTraceClockBase();
837    if (UNLIKELY(clock_base == 0)) {
838      // First event, record the base time in the map.
839      uint64_t time = thread->GetCpuMicroTime();
840      thread->SetTraceClockBase(time);
841    } else {
842      *thread_clock_diff = thread->GetCpuMicroTime() - clock_base;
843    }
844  }
845  if (UseWallClock()) {
846    *wall_clock_diff = MicroTime() - start_time_;
847  }
848}
849
850bool Trace::RegisterMethod(ArtMethod* method) {
851  mirror::DexCache* dex_cache = method->GetDexCache();
852  auto* resolved_method = dex_cache->GetResolvedMethod(method->GetDexMethodIndex(), sizeof(void*));
853  if (resolved_method != method) {
854    DCHECK(resolved_method == nullptr);
855    dex_cache->SetResolvedMethod(method->GetDexMethodIndex(), method, sizeof(void*));
856  }
857  if (seen_methods_.find(dex_cache) == seen_methods_.end()) {
858    seen_methods_.insert(std::make_pair(dex_cache, new DexIndexBitSet()));
859  }
860  DexIndexBitSet* bit_set = seen_methods_.find(dex_cache)->second;
861  if (!(*bit_set)[method->GetDexMethodIndex()]) {
862    bit_set->set(method->GetDexMethodIndex());
863    return true;
864  }
865  return false;
866}
867
868bool Trace::RegisterThread(Thread* thread) {
869  pid_t tid = thread->GetTid();
870  CHECK_LT(0U, static_cast<uint32_t>(tid));
871  CHECK_LT(static_cast<uint32_t>(tid), 65536U);
872
873  if (!(*seen_threads_)[tid]) {
874    seen_threads_->set(tid);
875    return true;
876  }
877  return false;
878}
879
880static std::string GetMethodLine(ArtMethod* method)
881    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
882  method = method->GetInterfaceMethodIfProxy(sizeof(void*));
883  return StringPrintf("%p\t%s\t%s\t%s\t%s\n", method,
884      PrettyDescriptor(method->GetDeclaringClassDescriptor()).c_str(), method->GetName(),
885      method->GetSignature().ToString().c_str(), method->GetDeclaringClassSourceFile());
886}
887
888void Trace::WriteToBuf(const uint8_t* src, size_t src_size) {
889  int32_t old_offset = cur_offset_.LoadRelaxed();
890  int32_t new_offset = old_offset + static_cast<int32_t>(src_size);
891  if (dchecked_integral_cast<size_t>(new_offset) > buffer_size_) {
892    // Flush buffer.
893    if (!trace_file_->WriteFully(buf_.get(), old_offset)) {
894      PLOG(WARNING) << "Failed streaming a tracing event.";
895    }
896
897    // Check whether the data is too large for the buffer, then write immediately.
898    if (src_size >= buffer_size_) {
899      if (!trace_file_->WriteFully(src, src_size)) {
900        PLOG(WARNING) << "Failed streaming a tracing event.";
901      }
902      cur_offset_.StoreRelease(0);  // Buffer is empty now.
903      return;
904    }
905
906    old_offset = 0;
907    new_offset = static_cast<int32_t>(src_size);
908  }
909  cur_offset_.StoreRelease(new_offset);
910  // Fill in data.
911  memcpy(buf_.get() + old_offset, src, src_size);
912}
913
914void Trace::LogMethodTraceEvent(Thread* thread, ArtMethod* method,
915                                instrumentation::Instrumentation::InstrumentationEvent event,
916                                uint32_t thread_clock_diff, uint32_t wall_clock_diff) {
917  // Advance cur_offset_ atomically.
918  int32_t new_offset;
919  int32_t old_offset = 0;
920
921  // We do a busy loop here trying to acquire the next offset.
922  if (trace_output_mode_ != TraceOutputMode::kStreaming) {
923    do {
924      old_offset = cur_offset_.LoadRelaxed();
925      new_offset = old_offset + GetRecordSize(clock_source_);
926      if (static_cast<size_t>(new_offset) > buffer_size_) {
927        overflow_ = true;
928        return;
929      }
930    } while (!cur_offset_.CompareExchangeWeakSequentiallyConsistent(old_offset, new_offset));
931  }
932
933  TraceAction action = kTraceMethodEnter;
934  switch (event) {
935    case instrumentation::Instrumentation::kMethodEntered:
936      action = kTraceMethodEnter;
937      break;
938    case instrumentation::Instrumentation::kMethodExited:
939      action = kTraceMethodExit;
940      break;
941    case instrumentation::Instrumentation::kMethodUnwind:
942      action = kTraceUnroll;
943      break;
944    default:
945      UNIMPLEMENTED(FATAL) << "Unexpected event: " << event;
946  }
947
948  uint64_t method_value = EncodeTraceMethodAndAction(method, action);
949
950  // Write data
951  uint8_t* ptr;
952  static constexpr size_t kPacketSize = 18U;  // The maximum size of data in a packet.
953  uint8_t stack_buf[kPacketSize];             // Space to store a packet when in streaming mode.
954  if (trace_output_mode_ == TraceOutputMode::kStreaming) {
955    ptr = stack_buf;
956  } else {
957    ptr = buf_.get() + old_offset;
958  }
959
960  Append2LE(ptr, thread->GetTid());
961  Append8LE(ptr + 2, method_value);
962  ptr += 10;
963
964  if (UseThreadCpuClock()) {
965    Append4LE(ptr, thread_clock_diff);
966    ptr += 4;
967  }
968  if (UseWallClock()) {
969    Append4LE(ptr, wall_clock_diff);
970  }
971  static_assert(kPacketSize == 2 + 8 + 4 + 4, "Packet size incorrect.");
972
973  if (trace_output_mode_ == TraceOutputMode::kStreaming) {
974    MutexLock mu(Thread::Current(), *streaming_lock_);  // To serialize writing.
975    if (RegisterMethod(method)) {
976      // Write a special block with the name.
977      std::string method_line(GetMethodLine(method));
978      uint8_t buf2[5];
979      Append2LE(buf2, 0);
980      buf2[2] = kOpNewMethod;
981      Append2LE(buf2 + 3, static_cast<uint16_t>(method_line.length()));
982      WriteToBuf(buf2, sizeof(buf2));
983      WriteToBuf(reinterpret_cast<const uint8_t*>(method_line.c_str()), method_line.length());
984    }
985    if (RegisterThread(thread)) {
986      // It might be better to postpone this. Threads might not have received names...
987      std::string thread_name;
988      thread->GetThreadName(thread_name);
989      uint8_t buf2[7];
990      Append2LE(buf2, 0);
991      buf2[2] = kOpNewThread;
992      Append2LE(buf2 + 3, static_cast<uint16_t>(thread->GetTid()));
993      Append2LE(buf2 + 5, static_cast<uint16_t>(thread_name.length()));
994      WriteToBuf(buf2, sizeof(buf2));
995      WriteToBuf(reinterpret_cast<const uint8_t*>(thread_name.c_str()), thread_name.length());
996    }
997    WriteToBuf(stack_buf, sizeof(stack_buf));
998  }
999}
1000
1001void Trace::GetVisitedMethods(size_t buf_size,
1002                              std::set<ArtMethod*>* visited_methods) {
1003  uint8_t* ptr = buf_.get() + kTraceHeaderLength;
1004  uint8_t* end = buf_.get() + buf_size;
1005
1006  while (ptr < end) {
1007    uint64_t tmid = ReadBytes(ptr + 2, sizeof(tmid));
1008    ArtMethod* method = DecodeTraceMethodId(tmid);
1009    visited_methods->insert(method);
1010    ptr += GetRecordSize(clock_source_);
1011  }
1012}
1013
1014void Trace::DumpMethodList(std::ostream& os, const std::set<ArtMethod*>& visited_methods) {
1015  for (const auto& method : visited_methods) {
1016    os << GetMethodLine(method);
1017  }
1018}
1019
1020static void DumpThread(Thread* t, void* arg) {
1021  std::ostream& os = *reinterpret_cast<std::ostream*>(arg);
1022  std::string name;
1023  t->GetThreadName(name);
1024  os << t->GetTid() << "\t" << name << "\n";
1025}
1026
1027void Trace::DumpThreadList(std::ostream& os) {
1028  Thread* self = Thread::Current();
1029  for (auto it : exited_threads_) {
1030    os << it.first << "\t" << it.second << "\n";
1031  }
1032  Locks::thread_list_lock_->AssertNotHeld(self);
1033  MutexLock mu(self, *Locks::thread_list_lock_);
1034  Runtime::Current()->GetThreadList()->ForEach(DumpThread, &os);
1035}
1036
1037void Trace::StoreExitingThreadInfo(Thread* thread) {
1038  MutexLock mu(thread, *Locks::trace_lock_);
1039  if (the_trace_ != nullptr) {
1040    std::string name;
1041    thread->GetThreadName(name);
1042    // The same thread/tid may be used multiple times. As SafeMap::Put does not allow to override
1043    // a previous mapping, use SafeMap::Overwrite.
1044    the_trace_->exited_threads_.Overwrite(thread->GetTid(), name);
1045  }
1046}
1047
1048Trace::TraceOutputMode Trace::GetOutputMode() {
1049  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
1050  CHECK(the_trace_ != nullptr) << "Trace output mode requested, but no trace currently running";
1051  return the_trace_->trace_output_mode_;
1052}
1053
1054Trace::TraceMode Trace::GetMode() {
1055  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
1056  CHECK(the_trace_ != nullptr) << "Trace mode requested, but no trace currently running";
1057  return the_trace_->trace_mode_;
1058}
1059
1060size_t Trace::GetBufferSize() {
1061  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
1062  CHECK(the_trace_ != nullptr) << "Trace mode requested, but no trace currently running";
1063  return the_trace_->buffer_size_;
1064}
1065
1066}  // namespace art
1067