trace.cc revision 9ea02c4f01b079bf6a5f49c62cb59fa5791cbc60 
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 "gc/scoped_gc_critical_section.h"
35#include "instrumentation.h"
36#include "mirror/class-inl.h"
37#include "mirror/dex_cache-inl.h"
38#include "mirror/object_array-inl.h"
39#include "mirror/object-inl.h"
40#include "os.h"
41#include "scoped_thread_state_change.h"
42#include "ScopedLocalRef.h"
43#include "thread.h"
44#include "thread_list.h"
45#include "utils.h"
46#include "entrypoints/quick/quick_entrypoints.h"
47
48namespace art {
49
50static constexpr size_t TraceActionBits = MinimumBitsToStore(
51    static_cast<size_t>(kTraceMethodActionMask));
52static constexpr uint8_t kOpNewMethod = 1U;
53static constexpr uint8_t kOpNewThread = 2U;
54
55class BuildStackTraceVisitor : public StackVisitor {
56 public:
57  explicit BuildStackTraceVisitor(Thread* thread)
58      : StackVisitor(thread, nullptr, StackVisitor::StackWalkKind::kIncludeInlinedFrames),
59        method_trace_(Trace::AllocStackTrace()) {}
60
61  bool VisitFrame() SHARED_REQUIRES(Locks::mutator_lock_) {
62    ArtMethod* m = GetMethod();
63    // Ignore runtime frames (in particular callee save).
64    if (!m->IsRuntimeMethod()) {
65      method_trace_->push_back(m);
66    }
67    return true;
68  }
69
70  // Returns a stack trace where the topmost frame corresponds with the first element of the vector.
71  std::vector<ArtMethod*>* GetStackTrace() const {
72    return method_trace_;
73  }
74
75 private:
76  std::vector<ArtMethod*>* const method_trace_;
77
78  DISALLOW_COPY_AND_ASSIGN(BuildStackTraceVisitor);
79};
80
81static const char     kTraceTokenChar             = '*';
82static const uint16_t kTraceHeaderLength          = 32;
83static const uint32_t kTraceMagicValue            = 0x574f4c53;
84static const uint16_t kTraceVersionSingleClock    = 2;
85static const uint16_t kTraceVersionDualClock      = 3;
86static const uint16_t kTraceRecordSizeSingleClock = 10;  // using v2
87static const uint16_t kTraceRecordSizeDualClock   = 14;  // using v3 with two timestamps
88
89TraceClockSource Trace::default_clock_source_ = kDefaultTraceClockSource;
90
91Trace* volatile Trace::the_trace_ = nullptr;
92pthread_t Trace::sampling_pthread_ = 0U;
93std::unique_ptr<std::vector<ArtMethod*>> Trace::temp_stack_trace_;
94
95// The key identifying the tracer to update instrumentation.
96static constexpr const char* kTracerInstrumentationKey = "Tracer";
97
98static TraceAction DecodeTraceAction(uint32_t tmid) {
99  return static_cast<TraceAction>(tmid & kTraceMethodActionMask);
100}
101
102ArtMethod* Trace::DecodeTraceMethod(uint32_t tmid) {
103  MutexLock mu(Thread::Current(), *unique_methods_lock_);
104  return unique_methods_[tmid >> TraceActionBits];
105}
106
107uint32_t Trace::EncodeTraceMethod(ArtMethod* method) {
108  MutexLock mu(Thread::Current(), *unique_methods_lock_);
109  uint32_t idx;
110  auto it = art_method_id_map_.find(method);
111  if (it != art_method_id_map_.end()) {
112    idx = it->second;
113  } else {
114    unique_methods_.push_back(method);
115    idx = unique_methods_.size() - 1;
116    art_method_id_map_.emplace(method, idx);
117  }
118  DCHECK_LT(idx, unique_methods_.size());
119  DCHECK_EQ(unique_methods_[idx], method);
120  return idx;
121}
122
123uint32_t Trace::EncodeTraceMethodAndAction(ArtMethod* method, TraceAction action) {
124  uint32_t tmid = (EncodeTraceMethod(method) << TraceActionBits) | action;
125  DCHECK_EQ(method, DecodeTraceMethod(tmid));
126  return tmid;
127}
128
129std::vector<ArtMethod*>* Trace::AllocStackTrace() {
130  return (temp_stack_trace_.get() != nullptr)  ? temp_stack_trace_.release() :
131      new std::vector<ArtMethod*>();
132}
133
134void Trace::FreeStackTrace(std::vector<ArtMethod*>* stack_trace) {
135  stack_trace->clear();
136  temp_stack_trace_.reset(stack_trace);
137}
138
139void Trace::SetDefaultClockSource(TraceClockSource clock_source) {
140#if defined(__linux__)
141  default_clock_source_ = clock_source;
142#else
143  if (clock_source != TraceClockSource::kWall) {
144    LOG(WARNING) << "Ignoring tracing request to use CPU time.";
145  }
146#endif
147}
148
149static uint16_t GetTraceVersion(TraceClockSource clock_source) {
150  return (clock_source == TraceClockSource::kDual) ? kTraceVersionDualClock
151                                                    : kTraceVersionSingleClock;
152}
153
154static uint16_t GetRecordSize(TraceClockSource clock_source) {
155  return (clock_source == TraceClockSource::kDual) ? kTraceRecordSizeDualClock
156                                                    : kTraceRecordSizeSingleClock;
157}
158
159bool Trace::UseThreadCpuClock() {
160  return (clock_source_ == TraceClockSource::kThreadCpu) ||
161      (clock_source_ == TraceClockSource::kDual);
162}
163
164bool Trace::UseWallClock() {
165  return (clock_source_ == TraceClockSource::kWall) ||
166      (clock_source_ == TraceClockSource::kDual);
167}
168
169void Trace::MeasureClockOverhead() {
170  if (UseThreadCpuClock()) {
171    Thread::Current()->GetCpuMicroTime();
172  }
173  if (UseWallClock()) {
174    MicroTime();
175  }
176}
177
178// Compute an average time taken to measure clocks.
179uint32_t Trace::GetClockOverheadNanoSeconds() {
180  Thread* self = Thread::Current();
181  uint64_t start = self->GetCpuMicroTime();
182
183  for (int i = 4000; i > 0; i--) {
184    MeasureClockOverhead();
185    MeasureClockOverhead();
186    MeasureClockOverhead();
187    MeasureClockOverhead();
188    MeasureClockOverhead();
189    MeasureClockOverhead();
190    MeasureClockOverhead();
191    MeasureClockOverhead();
192  }
193
194  uint64_t elapsed_us = self->GetCpuMicroTime() - start;
195  return static_cast<uint32_t>(elapsed_us / 32);
196}
197
198// TODO: put this somewhere with the big-endian equivalent used by JDWP.
199static void Append2LE(uint8_t* buf, uint16_t val) {
200  *buf++ = static_cast<uint8_t>(val);
201  *buf++ = static_cast<uint8_t>(val >> 8);
202}
203
204// TODO: put this somewhere with the big-endian equivalent used by JDWP.
205static void Append4LE(uint8_t* buf, uint32_t val) {
206  *buf++ = static_cast<uint8_t>(val);
207  *buf++ = static_cast<uint8_t>(val >> 8);
208  *buf++ = static_cast<uint8_t>(val >> 16);
209  *buf++ = static_cast<uint8_t>(val >> 24);
210}
211
212// TODO: put this somewhere with the big-endian equivalent used by JDWP.
213static void Append8LE(uint8_t* buf, uint64_t val) {
214  *buf++ = static_cast<uint8_t>(val);
215  *buf++ = static_cast<uint8_t>(val >> 8);
216  *buf++ = static_cast<uint8_t>(val >> 16);
217  *buf++ = static_cast<uint8_t>(val >> 24);
218  *buf++ = static_cast<uint8_t>(val >> 32);
219  *buf++ = static_cast<uint8_t>(val >> 40);
220  *buf++ = static_cast<uint8_t>(val >> 48);
221  *buf++ = static_cast<uint8_t>(val >> 56);
222}
223
224static void GetSample(Thread* thread, void* arg) SHARED_REQUIRES(Locks::mutator_lock_) {
225  BuildStackTraceVisitor build_trace_visitor(thread);
226  build_trace_visitor.WalkStack();
227  std::vector<ArtMethod*>* stack_trace = build_trace_visitor.GetStackTrace();
228  Trace* the_trace = reinterpret_cast<Trace*>(arg);
229  the_trace->CompareAndUpdateStackTrace(thread, stack_trace);
230}
231
232static void ClearThreadStackTraceAndClockBase(Thread* thread, void* arg ATTRIBUTE_UNUSED) {
233  thread->SetTraceClockBase(0);
234  std::vector<ArtMethod*>* stack_trace = thread->GetStackTraceSample();
235  thread->SetStackTraceSample(nullptr);
236  delete stack_trace;
237}
238
239void Trace::CompareAndUpdateStackTrace(Thread* thread,
240                                       std::vector<ArtMethod*>* stack_trace) {
241  CHECK_EQ(pthread_self(), sampling_pthread_);
242  std::vector<ArtMethod*>* old_stack_trace = thread->GetStackTraceSample();
243  // Update the thread's stack trace sample.
244  thread->SetStackTraceSample(stack_trace);
245  // Read timer clocks to use for all events in this trace.
246  uint32_t thread_clock_diff = 0;
247  uint32_t wall_clock_diff = 0;
248  ReadClocks(thread, &thread_clock_diff, &wall_clock_diff);
249  if (old_stack_trace == nullptr) {
250    // If there's no previous stack trace sample for this thread, log an entry event for all
251    // methods in the trace.
252    for (auto rit = stack_trace->rbegin(); rit != stack_trace->rend(); ++rit) {
253      LogMethodTraceEvent(thread, *rit, instrumentation::Instrumentation::kMethodEntered,
254                          thread_clock_diff, wall_clock_diff);
255    }
256  } else {
257    // If there's a previous stack trace for this thread, diff the traces and emit entry and exit
258    // events accordingly.
259    auto old_rit = old_stack_trace->rbegin();
260    auto rit = stack_trace->rbegin();
261    // Iterate bottom-up over both traces until there's a difference between them.
262    while (old_rit != old_stack_trace->rend() && rit != stack_trace->rend() && *old_rit == *rit) {
263      old_rit++;
264      rit++;
265    }
266    // Iterate top-down over the old trace until the point where they differ, emitting exit events.
267    for (auto old_it = old_stack_trace->begin(); old_it != old_rit.base(); ++old_it) {
268      LogMethodTraceEvent(thread, *old_it, instrumentation::Instrumentation::kMethodExited,
269                          thread_clock_diff, wall_clock_diff);
270    }
271    // Iterate bottom-up over the new trace from the point where they differ, emitting entry events.
272    for (; rit != stack_trace->rend(); ++rit) {
273      LogMethodTraceEvent(thread, *rit, instrumentation::Instrumentation::kMethodEntered,
274                          thread_clock_diff, wall_clock_diff);
275    }
276    FreeStackTrace(old_stack_trace);
277  }
278}
279
280void* Trace::RunSamplingThread(void* arg) {
281  Runtime* runtime = Runtime::Current();
282  intptr_t interval_us = reinterpret_cast<intptr_t>(arg);
283  CHECK_GE(interval_us, 0);
284  CHECK(runtime->AttachCurrentThread("Sampling Profiler", true, runtime->GetSystemThreadGroup(),
285                                     !runtime->IsAotCompiler()));
286
287  while (true) {
288    usleep(interval_us);
289    ATRACE_BEGIN("Profile sampling");
290    Thread* self = Thread::Current();
291    Trace* the_trace;
292    {
293      MutexLock mu(self, *Locks::trace_lock_);
294      the_trace = the_trace_;
295      if (the_trace == nullptr) {
296        break;
297      }
298    }
299    {
300      ScopedSuspendAll ssa(__FUNCTION__);
301      MutexLock mu(self, *Locks::thread_list_lock_);
302      runtime->GetThreadList()->ForEach(GetSample, the_trace);
303    }
304    ATRACE_END();
305  }
306
307  runtime->DetachCurrentThread();
308  return nullptr;
309}
310
311void Trace::Start(const char* trace_filename, int trace_fd, size_t buffer_size, int flags,
312                  TraceOutputMode output_mode, TraceMode trace_mode, int interval_us) {
313  Thread* self = Thread::Current();
314  {
315    MutexLock mu(self, *Locks::trace_lock_);
316    if (the_trace_ != nullptr) {
317      LOG(ERROR) << "Trace already in progress, ignoring this request";
318      return;
319    }
320  }
321
322  // Check interval if sampling is enabled
323  if (trace_mode == TraceMode::kSampling && interval_us <= 0) {
324    LOG(ERROR) << "Invalid sampling interval: " << interval_us;
325    ScopedObjectAccess soa(self);
326    ThrowRuntimeException("Invalid sampling interval: %d", interval_us);
327    return;
328  }
329
330  // Open trace file if not going directly to ddms.
331  std::unique_ptr<File> trace_file;
332  if (output_mode != TraceOutputMode::kDDMS) {
333    if (trace_fd < 0) {
334      trace_file.reset(OS::CreateEmptyFileWriteOnly(trace_filename));
335    } else {
336      trace_file.reset(new File(trace_fd, "tracefile"));
337      trace_file->DisableAutoClose();
338    }
339    if (trace_file.get() == nullptr) {
340      PLOG(ERROR) << "Unable to open trace file '" << trace_filename << "'";
341      ScopedObjectAccess soa(self);
342      ThrowRuntimeException("Unable to open trace file '%s'", trace_filename);
343      return;
344    }
345  }
346
347  Runtime* runtime = Runtime::Current();
348
349  // Enable count of allocs if specified in the flags.
350  bool enable_stats = false;
351
352  // Create Trace object.
353  {
354    // Required since EnableMethodTracing calls ConfigureStubs which visits class linker classes.
355    gc::ScopedGCCriticalSection gcs(self,
356                                    gc::kGcCauseInstrumentation,
357                                    gc::kCollectorTypeInstrumentation);
358    ScopedSuspendAll ssa(__FUNCTION__);
359    MutexLock mu(self, *Locks::trace_lock_);
360    if (the_trace_ != nullptr) {
361      LOG(ERROR) << "Trace already in progress, ignoring this request";
362    } else {
363      enable_stats = (flags && kTraceCountAllocs) != 0;
364      the_trace_ = new Trace(trace_file.release(), trace_filename, buffer_size, flags, output_mode,
365                             trace_mode);
366      if (trace_mode == TraceMode::kSampling) {
367        CHECK_PTHREAD_CALL(pthread_create, (&sampling_pthread_, nullptr, &RunSamplingThread,
368                                            reinterpret_cast<void*>(interval_us)),
369                                            "Sampling profiler thread");
370        the_trace_->interval_us_ = interval_us;
371      } else {
372        runtime->GetInstrumentation()->AddListener(the_trace_,
373                                                   instrumentation::Instrumentation::kMethodEntered |
374                                                   instrumentation::Instrumentation::kMethodExited |
375                                                   instrumentation::Instrumentation::kMethodUnwind);
376        // TODO: In full-PIC mode, we don't need to fully deopt.
377        runtime->GetInstrumentation()->EnableMethodTracing(kTracerInstrumentationKey);
378      }
379    }
380  }
381
382  // Can't call this when holding the mutator lock.
383  if (enable_stats) {
384    runtime->SetStatsEnabled(true);
385  }
386}
387
388void Trace::StopTracing(bool finish_tracing, bool flush_file) {
389  bool stop_alloc_counting = false;
390  Runtime* const runtime = Runtime::Current();
391  Trace* the_trace = nullptr;
392  Thread* const self = Thread::Current();
393  pthread_t sampling_pthread = 0U;
394  {
395    MutexLock mu(self, *Locks::trace_lock_);
396    if (the_trace_ == nullptr) {
397      LOG(ERROR) << "Trace stop requested, but no trace currently running";
398    } else {
399      the_trace = the_trace_;
400      the_trace_ = nullptr;
401      sampling_pthread = sampling_pthread_;
402    }
403  }
404  // Make sure that we join before we delete the trace since we don't want to have
405  // the sampling thread access a stale pointer. This finishes since the sampling thread exits when
406  // the_trace_ is null.
407  if (sampling_pthread != 0U) {
408    CHECK_PTHREAD_CALL(pthread_join, (sampling_pthread, nullptr), "sampling thread shutdown");
409    sampling_pthread_ = 0U;
410  }
411
412  {
413    gc::ScopedGCCriticalSection gcs(self,
414                                    gc::kGcCauseInstrumentation,
415                                    gc::kCollectorTypeInstrumentation);
416    ScopedSuspendAll ssa(__FUNCTION__);
417    if (the_trace != nullptr) {
418      stop_alloc_counting = (the_trace->flags_ & Trace::kTraceCountAllocs) != 0;
419      if (finish_tracing) {
420        the_trace->FinishTracing();
421      }
422
423      if (the_trace->trace_mode_ == TraceMode::kSampling) {
424        MutexLock mu(self, *Locks::thread_list_lock_);
425        runtime->GetThreadList()->ForEach(ClearThreadStackTraceAndClockBase, nullptr);
426      } else {
427        runtime->GetInstrumentation()->DisableMethodTracing(kTracerInstrumentationKey);
428        runtime->GetInstrumentation()->RemoveListener(
429            the_trace, instrumentation::Instrumentation::kMethodEntered |
430            instrumentation::Instrumentation::kMethodExited |
431            instrumentation::Instrumentation::kMethodUnwind);
432      }
433      if (the_trace->trace_file_.get() != nullptr) {
434        // Do not try to erase, so flush and close explicitly.
435        if (flush_file) {
436          if (the_trace->trace_file_->Flush() != 0) {
437            PLOG(WARNING) << "Could not flush trace file.";
438          }
439        } else {
440          the_trace->trace_file_->MarkUnchecked();  // Do not trigger guard.
441        }
442        if (the_trace->trace_file_->Close() != 0) {
443          PLOG(ERROR) << "Could not close trace file.";
444        }
445      }
446      delete the_trace;
447    }
448  }
449  if (stop_alloc_counting) {
450    // Can be racy since SetStatsEnabled is not guarded by any locks.
451    runtime->SetStatsEnabled(false);
452  }
453}
454
455void Trace::Abort() {
456  // Do not write anything anymore.
457  StopTracing(false, false);
458}
459
460void Trace::Stop() {
461  // Finish writing.
462  StopTracing(true, true);
463}
464
465void Trace::Shutdown() {
466  if (GetMethodTracingMode() != kTracingInactive) {
467    Stop();
468  }
469}
470
471void Trace::Pause() {
472  bool stop_alloc_counting = false;
473  Runtime* runtime = Runtime::Current();
474  Trace* the_trace = nullptr;
475
476  Thread* const self = Thread::Current();
477  pthread_t sampling_pthread = 0U;
478  {
479    MutexLock mu(self, *Locks::trace_lock_);
480    if (the_trace_ == nullptr) {
481      LOG(ERROR) << "Trace pause requested, but no trace currently running";
482      return;
483    } else {
484      the_trace = the_trace_;
485      sampling_pthread = sampling_pthread_;
486    }
487  }
488
489  if (sampling_pthread != 0U) {
490    {
491      MutexLock mu(self, *Locks::trace_lock_);
492      the_trace_ = nullptr;
493    }
494    CHECK_PTHREAD_CALL(pthread_join, (sampling_pthread, nullptr), "sampling thread shutdown");
495    sampling_pthread_ = 0U;
496    {
497      MutexLock mu(self, *Locks::trace_lock_);
498      the_trace_ = the_trace;
499    }
500  }
501
502  if (the_trace != nullptr) {
503    gc::ScopedGCCriticalSection gcs(self,
504                                    gc::kGcCauseInstrumentation,
505                                    gc::kCollectorTypeInstrumentation);
506    ScopedSuspendAll ssa(__FUNCTION__);
507    stop_alloc_counting = (the_trace->flags_ & Trace::kTraceCountAllocs) != 0;
508
509    if (the_trace->trace_mode_ == TraceMode::kSampling) {
510      MutexLock mu(self, *Locks::thread_list_lock_);
511      runtime->GetThreadList()->ForEach(ClearThreadStackTraceAndClockBase, nullptr);
512    } else {
513      runtime->GetInstrumentation()->DisableMethodTracing(kTracerInstrumentationKey);
514      runtime->GetInstrumentation()->RemoveListener(
515          the_trace,
516          instrumentation::Instrumentation::kMethodEntered |
517          instrumentation::Instrumentation::kMethodExited |
518          instrumentation::Instrumentation::kMethodUnwind);
519    }
520  }
521
522  if (stop_alloc_counting) {
523    // Can be racy since SetStatsEnabled is not guarded by any locks.
524    Runtime::Current()->SetStatsEnabled(false);
525  }
526}
527
528void Trace::Resume() {
529  Thread* self = Thread::Current();
530  Trace* the_trace;
531  {
532    MutexLock mu(self, *Locks::trace_lock_);
533    if (the_trace_ == nullptr) {
534      LOG(ERROR) << "No trace to resume (or sampling mode), ignoring this request";
535      return;
536    }
537    the_trace = the_trace_;
538  }
539
540  Runtime* runtime = Runtime::Current();
541
542  // Enable count of allocs if specified in the flags.
543  bool enable_stats = (the_trace->flags_ && kTraceCountAllocs) != 0;
544
545  {
546    gc::ScopedGCCriticalSection gcs(self,
547                                    gc::kGcCauseInstrumentation,
548                                    gc::kCollectorTypeInstrumentation);
549    ScopedSuspendAll ssa(__FUNCTION__);
550
551    // Reenable.
552    if (the_trace->trace_mode_ == TraceMode::kSampling) {
553      CHECK_PTHREAD_CALL(pthread_create, (&sampling_pthread_, nullptr, &RunSamplingThread,
554          reinterpret_cast<void*>(the_trace->interval_us_)), "Sampling profiler thread");
555    } else {
556      runtime->GetInstrumentation()->AddListener(the_trace,
557                                                 instrumentation::Instrumentation::kMethodEntered |
558                                                 instrumentation::Instrumentation::kMethodExited |
559                                                 instrumentation::Instrumentation::kMethodUnwind);
560      // TODO: In full-PIC mode, we don't need to fully deopt.
561      runtime->GetInstrumentation()->EnableMethodTracing(kTracerInstrumentationKey);
562    }
563  }
564
565  // Can't call this when holding the mutator lock.
566  if (enable_stats) {
567    runtime->SetStatsEnabled(true);
568  }
569}
570
571TracingMode Trace::GetMethodTracingMode() {
572  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
573  if (the_trace_ == nullptr) {
574    return kTracingInactive;
575  } else {
576    switch (the_trace_->trace_mode_) {
577      case TraceMode::kSampling:
578        return kSampleProfilingActive;
579      case TraceMode::kMethodTracing:
580        return kMethodTracingActive;
581    }
582    LOG(FATAL) << "Unreachable";
583    UNREACHABLE();
584  }
585}
586
587static constexpr size_t kMinBufSize = 18U;  // Trace header is up to 18B.
588
589Trace::Trace(File* trace_file, const char* trace_name, size_t buffer_size, int flags,
590             TraceOutputMode output_mode, TraceMode trace_mode)
591    : trace_file_(trace_file),
592      buf_(new uint8_t[std::max(kMinBufSize, buffer_size)]()),
593      flags_(flags), trace_output_mode_(output_mode), trace_mode_(trace_mode),
594      clock_source_(default_clock_source_),
595      buffer_size_(std::max(kMinBufSize, buffer_size)),
596      start_time_(MicroTime()), clock_overhead_ns_(GetClockOverheadNanoSeconds()), cur_offset_(0),
597      overflow_(false), interval_us_(0), streaming_lock_(nullptr),
598      unique_methods_lock_(new Mutex("unique methods lock", kTracingUniqueMethodsLock)) {
599  uint16_t trace_version = GetTraceVersion(clock_source_);
600  if (output_mode == TraceOutputMode::kStreaming) {
601    trace_version |= 0xF0U;
602  }
603  // Set up the beginning of the trace.
604  memset(buf_.get(), 0, kTraceHeaderLength);
605  Append4LE(buf_.get(), kTraceMagicValue);
606  Append2LE(buf_.get() + 4, trace_version);
607  Append2LE(buf_.get() + 6, kTraceHeaderLength);
608  Append8LE(buf_.get() + 8, start_time_);
609  if (trace_version >= kTraceVersionDualClock) {
610    uint16_t record_size = GetRecordSize(clock_source_);
611    Append2LE(buf_.get() + 16, record_size);
612  }
613  static_assert(18 <= kMinBufSize, "Minimum buffer size not large enough for trace header");
614
615  // Update current offset.
616  cur_offset_.StoreRelaxed(kTraceHeaderLength);
617
618  if (output_mode == TraceOutputMode::kStreaming) {
619    streaming_file_name_ = trace_name;
620    streaming_lock_ = new Mutex("tracing lock", LockLevel::kTracingStreamingLock);
621    seen_threads_.reset(new ThreadIDBitSet());
622  }
623}
624
625Trace::~Trace() {
626  delete streaming_lock_;
627  delete unique_methods_lock_;
628}
629
630static uint64_t ReadBytes(uint8_t* buf, size_t bytes) {
631  uint64_t ret = 0;
632  for (size_t i = 0; i < bytes; ++i) {
633    ret |= static_cast<uint64_t>(buf[i]) << (i * 8);
634  }
635  return ret;
636}
637
638void Trace::DumpBuf(uint8_t* buf, size_t buf_size, TraceClockSource clock_source) {
639  uint8_t* ptr = buf + kTraceHeaderLength;
640  uint8_t* end = buf + buf_size;
641
642  while (ptr < end) {
643    uint32_t tmid = ReadBytes(ptr + 2, sizeof(tmid));
644    ArtMethod* method = DecodeTraceMethod(tmid);
645    TraceAction action = DecodeTraceAction(tmid);
646    LOG(INFO) << PrettyMethod(method) << " " << static_cast<int>(action);
647    ptr += GetRecordSize(clock_source);
648  }
649}
650
651static void GetVisitedMethodsFromBitSets(
652    const std::map<const DexFile*, DexIndexBitSet*>& seen_methods,
653    std::set<ArtMethod*>* visited_methods) SHARED_REQUIRES(Locks::mutator_lock_) {
654  ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
655  Thread* const self = Thread::Current();
656  for (auto& e : seen_methods) {
657    DexIndexBitSet* bit_set = e.second;
658    // TODO: Visit trace methods as roots.
659    mirror::DexCache* dex_cache = class_linker->FindDexCache(self, *e.first, false);
660    for (uint32_t i = 0; i < bit_set->size(); ++i) {
661      if ((*bit_set)[i]) {
662        visited_methods->insert(dex_cache->GetResolvedMethod(i, sizeof(void*)));
663      }
664    }
665  }
666}
667
668void Trace::FinishTracing() {
669  size_t final_offset = 0;
670
671  std::set<ArtMethod*> visited_methods;
672  if (trace_output_mode_ == TraceOutputMode::kStreaming) {
673    // Write the secondary file with all the method names.
674    GetVisitedMethodsFromBitSets(seen_methods_, &visited_methods);
675
676    // Clean up.
677    STLDeleteValues(&seen_methods_);
678  } else {
679    final_offset = cur_offset_.LoadRelaxed();
680    GetVisitedMethods(final_offset, &visited_methods);
681  }
682
683  // Compute elapsed time.
684  uint64_t elapsed = MicroTime() - start_time_;
685
686  std::ostringstream os;
687
688  os << StringPrintf("%cversion\n", kTraceTokenChar);
689  os << StringPrintf("%d\n", GetTraceVersion(clock_source_));
690  os << StringPrintf("data-file-overflow=%s\n", overflow_ ? "true" : "false");
691  if (UseThreadCpuClock()) {
692    if (UseWallClock()) {
693      os << StringPrintf("clock=dual\n");
694    } else {
695      os << StringPrintf("clock=thread-cpu\n");
696    }
697  } else {
698    os << StringPrintf("clock=wall\n");
699  }
700  os << StringPrintf("elapsed-time-usec=%" PRIu64 "\n", elapsed);
701  if (trace_output_mode_ != TraceOutputMode::kStreaming) {
702    size_t num_records = (final_offset - kTraceHeaderLength) / GetRecordSize(clock_source_);
703    os << StringPrintf("num-method-calls=%zd\n", num_records);
704  }
705  os << StringPrintf("clock-call-overhead-nsec=%d\n", clock_overhead_ns_);
706  os << StringPrintf("vm=art\n");
707  os << StringPrintf("pid=%d\n", getpid());
708  if ((flags_ & kTraceCountAllocs) != 0) {
709    os << StringPrintf("alloc-count=%d\n", Runtime::Current()->GetStat(KIND_ALLOCATED_OBJECTS));
710    os << StringPrintf("alloc-size=%d\n", Runtime::Current()->GetStat(KIND_ALLOCATED_BYTES));
711    os << StringPrintf("gc-count=%d\n", Runtime::Current()->GetStat(KIND_GC_INVOCATIONS));
712  }
713  os << StringPrintf("%cthreads\n", kTraceTokenChar);
714  DumpThreadList(os);
715  os << StringPrintf("%cmethods\n", kTraceTokenChar);
716  DumpMethodList(os, visited_methods);
717  os << StringPrintf("%cend\n", kTraceTokenChar);
718  std::string header(os.str());
719
720  if (trace_output_mode_ == TraceOutputMode::kStreaming) {
721    File file;
722    if (!file.Open(streaming_file_name_ + ".sec", O_CREAT | O_WRONLY)) {
723      LOG(WARNING) << "Could not open secondary trace file!";
724      return;
725    }
726    if (!file.WriteFully(header.c_str(), header.length())) {
727      file.Erase();
728      std::string detail(StringPrintf("Trace data write failed: %s", strerror(errno)));
729      PLOG(ERROR) << detail;
730      ThrowRuntimeException("%s", detail.c_str());
731    }
732    if (file.FlushCloseOrErase() != 0) {
733      PLOG(ERROR) << "Could not write secondary file";
734    }
735  } else {
736    if (trace_file_.get() == nullptr) {
737      iovec iov[2];
738      iov[0].iov_base = reinterpret_cast<void*>(const_cast<char*>(header.c_str()));
739      iov[0].iov_len = header.length();
740      iov[1].iov_base = buf_.get();
741      iov[1].iov_len = final_offset;
742      Dbg::DdmSendChunkV(CHUNK_TYPE("MPSE"), iov, 2);
743      const bool kDumpTraceInfo = false;
744      if (kDumpTraceInfo) {
745        LOG(INFO) << "Trace sent:\n" << header;
746        DumpBuf(buf_.get(), final_offset, clock_source_);
747      }
748    } else {
749      if (!trace_file_->WriteFully(header.c_str(), header.length()) ||
750          !trace_file_->WriteFully(buf_.get(), final_offset)) {
751        std::string detail(StringPrintf("Trace data write failed: %s", strerror(errno)));
752        PLOG(ERROR) << detail;
753        ThrowRuntimeException("%s", detail.c_str());
754      }
755    }
756  }
757}
758
759void Trace::DexPcMoved(Thread* thread ATTRIBUTE_UNUSED,
760                       mirror::Object* this_object ATTRIBUTE_UNUSED,
761                       ArtMethod* method,
762                       uint32_t new_dex_pc) {
763  // We're not recorded to listen to this kind of event, so complain.
764  LOG(ERROR) << "Unexpected dex PC event in tracing " << PrettyMethod(method) << " " << new_dex_pc;
765}
766
767void Trace::FieldRead(Thread* thread ATTRIBUTE_UNUSED,
768                      mirror::Object* this_object ATTRIBUTE_UNUSED,
769                      ArtMethod* method,
770                      uint32_t dex_pc,
771                      ArtField* field ATTRIBUTE_UNUSED)
772    SHARED_REQUIRES(Locks::mutator_lock_) {
773  // We're not recorded to listen to this kind of event, so complain.
774  LOG(ERROR) << "Unexpected field read event in tracing " << PrettyMethod(method) << " " << dex_pc;
775}
776
777void Trace::FieldWritten(Thread* thread ATTRIBUTE_UNUSED,
778                         mirror::Object* this_object ATTRIBUTE_UNUSED,
779                         ArtMethod* method,
780                         uint32_t dex_pc,
781                         ArtField* field ATTRIBUTE_UNUSED,
782                         const JValue& field_value ATTRIBUTE_UNUSED)
783    SHARED_REQUIRES(Locks::mutator_lock_) {
784  // We're not recorded to listen to this kind of event, so complain.
785  LOG(ERROR) << "Unexpected field write event in tracing " << PrettyMethod(method) << " " << dex_pc;
786}
787
788void Trace::MethodEntered(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED,
789                          ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED) {
790  uint32_t thread_clock_diff = 0;
791  uint32_t wall_clock_diff = 0;
792  ReadClocks(thread, &thread_clock_diff, &wall_clock_diff);
793  LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodEntered,
794                      thread_clock_diff, wall_clock_diff);
795}
796
797void Trace::MethodExited(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED,
798                         ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED,
799                         const JValue& return_value ATTRIBUTE_UNUSED) {
800  uint32_t thread_clock_diff = 0;
801  uint32_t wall_clock_diff = 0;
802  ReadClocks(thread, &thread_clock_diff, &wall_clock_diff);
803  LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodExited,
804                      thread_clock_diff, wall_clock_diff);
805}
806
807void Trace::MethodUnwind(Thread* thread, mirror::Object* this_object ATTRIBUTE_UNUSED,
808                         ArtMethod* method, uint32_t dex_pc ATTRIBUTE_UNUSED) {
809  uint32_t thread_clock_diff = 0;
810  uint32_t wall_clock_diff = 0;
811  ReadClocks(thread, &thread_clock_diff, &wall_clock_diff);
812  LogMethodTraceEvent(thread, method, instrumentation::Instrumentation::kMethodUnwind,
813                      thread_clock_diff, wall_clock_diff);
814}
815
816void Trace::ExceptionCaught(Thread* thread ATTRIBUTE_UNUSED,
817                            mirror::Throwable* exception_object ATTRIBUTE_UNUSED)
818    SHARED_REQUIRES(Locks::mutator_lock_) {
819  LOG(ERROR) << "Unexpected exception caught event in tracing";
820}
821
822void Trace::Branch(Thread* /*thread*/, ArtMethod* method,
823                   uint32_t /*dex_pc*/, int32_t /*dex_pc_offset*/)
824      SHARED_REQUIRES(Locks::mutator_lock_) {
825  LOG(ERROR) << "Unexpected branch event in tracing" << PrettyMethod(method);
826}
827
828void Trace::InvokeVirtualOrInterface(Thread*,
829                                     mirror::Object*,
830                                     ArtMethod* method,
831                                     uint32_t dex_pc,
832                                     ArtMethod*) {
833  LOG(ERROR) << "Unexpected invoke event in tracing" << PrettyMethod(method)
834             << " " << dex_pc;
835}
836
837void Trace::ReadClocks(Thread* thread, uint32_t* thread_clock_diff, uint32_t* wall_clock_diff) {
838  if (UseThreadCpuClock()) {
839    uint64_t clock_base = thread->GetTraceClockBase();
840    if (UNLIKELY(clock_base == 0)) {
841      // First event, record the base time in the map.
842      uint64_t time = thread->GetCpuMicroTime();
843      thread->SetTraceClockBase(time);
844    } else {
845      *thread_clock_diff = thread->GetCpuMicroTime() - clock_base;
846    }
847  }
848  if (UseWallClock()) {
849    *wall_clock_diff = MicroTime() - start_time_;
850  }
851}
852
853bool Trace::RegisterMethod(ArtMethod* method) {
854  mirror::DexCache* dex_cache = method->GetDexCache();
855  const DexFile* dex_file = dex_cache->GetDexFile();
856  auto* resolved_method = dex_cache->GetResolvedMethod(method->GetDexMethodIndex(), sizeof(void*));
857  if (resolved_method != method) {
858    DCHECK(resolved_method == nullptr);
859    dex_cache->SetResolvedMethod(method->GetDexMethodIndex(), method, sizeof(void*));
860  }
861  if (seen_methods_.find(dex_file) == seen_methods_.end()) {
862    seen_methods_.insert(std::make_pair(dex_file, new DexIndexBitSet()));
863  }
864  DexIndexBitSet* bit_set = seen_methods_.find(dex_file)->second;
865  if (!(*bit_set)[method->GetDexMethodIndex()]) {
866    bit_set->set(method->GetDexMethodIndex());
867    return true;
868  }
869  return false;
870}
871
872bool Trace::RegisterThread(Thread* thread) {
873  pid_t tid = thread->GetTid();
874  CHECK_LT(0U, static_cast<uint32_t>(tid));
875  CHECK_LT(static_cast<uint32_t>(tid), 65536U);
876
877  if (!(*seen_threads_)[tid]) {
878    seen_threads_->set(tid);
879    return true;
880  }
881  return false;
882}
883
884std::string Trace::GetMethodLine(ArtMethod* method) {
885  method = method->GetInterfaceMethodIfProxy(sizeof(void*));
886  return StringPrintf("%p\t%s\t%s\t%s\t%s\n",
887                      reinterpret_cast<void*>((EncodeTraceMethod(method) << TraceActionBits)),
888      PrettyDescriptor(method->GetDeclaringClassDescriptor()).c_str(), method->GetName(),
889      method->GetSignature().ToString().c_str(), method->GetDeclaringClassSourceFile());
890}
891
892void Trace::WriteToBuf(const uint8_t* src, size_t src_size) {
893  int32_t old_offset = cur_offset_.LoadRelaxed();
894  int32_t new_offset = old_offset + static_cast<int32_t>(src_size);
895  if (dchecked_integral_cast<size_t>(new_offset) > buffer_size_) {
896    // Flush buffer.
897    if (!trace_file_->WriteFully(buf_.get(), old_offset)) {
898      PLOG(WARNING) << "Failed streaming a tracing event.";
899    }
900
901    // Check whether the data is too large for the buffer, then write immediately.
902    if (src_size >= buffer_size_) {
903      if (!trace_file_->WriteFully(src, src_size)) {
904        PLOG(WARNING) << "Failed streaming a tracing event.";
905      }
906      cur_offset_.StoreRelease(0);  // Buffer is empty now.
907      return;
908    }
909
910    old_offset = 0;
911    new_offset = static_cast<int32_t>(src_size);
912  }
913  cur_offset_.StoreRelease(new_offset);
914  // Fill in data.
915  memcpy(buf_.get() + old_offset, src, src_size);
916}
917
918void Trace::LogMethodTraceEvent(Thread* thread, ArtMethod* method,
919                                instrumentation::Instrumentation::InstrumentationEvent event,
920                                uint32_t thread_clock_diff, uint32_t wall_clock_diff) {
921  // Advance cur_offset_ atomically.
922  int32_t new_offset;
923  int32_t old_offset = 0;
924
925  // We do a busy loop here trying to acquire the next offset.
926  if (trace_output_mode_ != TraceOutputMode::kStreaming) {
927    do {
928      old_offset = cur_offset_.LoadRelaxed();
929      new_offset = old_offset + GetRecordSize(clock_source_);
930      if (static_cast<size_t>(new_offset) > buffer_size_) {
931        overflow_ = true;
932        return;
933      }
934    } while (!cur_offset_.CompareExchangeWeakSequentiallyConsistent(old_offset, new_offset));
935  }
936
937  TraceAction action = kTraceMethodEnter;
938  switch (event) {
939    case instrumentation::Instrumentation::kMethodEntered:
940      action = kTraceMethodEnter;
941      break;
942    case instrumentation::Instrumentation::kMethodExited:
943      action = kTraceMethodExit;
944      break;
945    case instrumentation::Instrumentation::kMethodUnwind:
946      action = kTraceUnroll;
947      break;
948    default:
949      UNIMPLEMENTED(FATAL) << "Unexpected event: " << event;
950  }
951
952  uint32_t method_value = EncodeTraceMethodAndAction(method, action);
953
954  // Write data
955  uint8_t* ptr;
956  static constexpr size_t kPacketSize = 14U;  // The maximum size of data in a packet.
957  uint8_t stack_buf[kPacketSize];             // Space to store a packet when in streaming mode.
958  if (trace_output_mode_ == TraceOutputMode::kStreaming) {
959    ptr = stack_buf;
960  } else {
961    ptr = buf_.get() + old_offset;
962  }
963
964  Append2LE(ptr, thread->GetTid());
965  Append4LE(ptr + 2, method_value);
966  ptr += 6;
967
968  if (UseThreadCpuClock()) {
969    Append4LE(ptr, thread_clock_diff);
970    ptr += 4;
971  }
972  if (UseWallClock()) {
973    Append4LE(ptr, wall_clock_diff);
974  }
975  static_assert(kPacketSize == 2 + 4 + 4 + 4, "Packet size incorrect.");
976
977  if (trace_output_mode_ == TraceOutputMode::kStreaming) {
978    MutexLock mu(Thread::Current(), *streaming_lock_);  // To serialize writing.
979    if (RegisterMethod(method)) {
980      // Write a special block with the name.
981      std::string method_line(GetMethodLine(method));
982      uint8_t buf2[5];
983      Append2LE(buf2, 0);
984      buf2[2] = kOpNewMethod;
985      Append2LE(buf2 + 3, static_cast<uint16_t>(method_line.length()));
986      WriteToBuf(buf2, sizeof(buf2));
987      WriteToBuf(reinterpret_cast<const uint8_t*>(method_line.c_str()), method_line.length());
988    }
989    if (RegisterThread(thread)) {
990      // It might be better to postpone this. Threads might not have received names...
991      std::string thread_name;
992      thread->GetThreadName(thread_name);
993      uint8_t buf2[7];
994      Append2LE(buf2, 0);
995      buf2[2] = kOpNewThread;
996      Append2LE(buf2 + 3, static_cast<uint16_t>(thread->GetTid()));
997      Append2LE(buf2 + 5, static_cast<uint16_t>(thread_name.length()));
998      WriteToBuf(buf2, sizeof(buf2));
999      WriteToBuf(reinterpret_cast<const uint8_t*>(thread_name.c_str()), thread_name.length());
1000    }
1001    WriteToBuf(stack_buf, sizeof(stack_buf));
1002  }
1003}
1004
1005void Trace::GetVisitedMethods(size_t buf_size,
1006                              std::set<ArtMethod*>* visited_methods) {
1007  uint8_t* ptr = buf_.get() + kTraceHeaderLength;
1008  uint8_t* end = buf_.get() + buf_size;
1009
1010  while (ptr < end) {
1011    uint32_t tmid = ReadBytes(ptr + 2, sizeof(tmid));
1012    ArtMethod* method = DecodeTraceMethod(tmid);
1013    visited_methods->insert(method);
1014    ptr += GetRecordSize(clock_source_);
1015  }
1016}
1017
1018void Trace::DumpMethodList(std::ostream& os, const std::set<ArtMethod*>& visited_methods) {
1019  for (const auto& method : visited_methods) {
1020    os << GetMethodLine(method);
1021  }
1022}
1023
1024static void DumpThread(Thread* t, void* arg) {
1025  std::ostream& os = *reinterpret_cast<std::ostream*>(arg);
1026  std::string name;
1027  t->GetThreadName(name);
1028  os << t->GetTid() << "\t" << name << "\n";
1029}
1030
1031void Trace::DumpThreadList(std::ostream& os) {
1032  Thread* self = Thread::Current();
1033  for (auto it : exited_threads_) {
1034    os << it.first << "\t" << it.second << "\n";
1035  }
1036  Locks::thread_list_lock_->AssertNotHeld(self);
1037  MutexLock mu(self, *Locks::thread_list_lock_);
1038  Runtime::Current()->GetThreadList()->ForEach(DumpThread, &os);
1039}
1040
1041void Trace::StoreExitingThreadInfo(Thread* thread) {
1042  MutexLock mu(thread, *Locks::trace_lock_);
1043  if (the_trace_ != nullptr) {
1044    std::string name;
1045    thread->GetThreadName(name);
1046    // The same thread/tid may be used multiple times. As SafeMap::Put does not allow to override
1047    // a previous mapping, use SafeMap::Overwrite.
1048    the_trace_->exited_threads_.Overwrite(thread->GetTid(), name);
1049  }
1050}
1051
1052Trace::TraceOutputMode Trace::GetOutputMode() {
1053  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
1054  CHECK(the_trace_ != nullptr) << "Trace output mode requested, but no trace currently running";
1055  return the_trace_->trace_output_mode_;
1056}
1057
1058Trace::TraceMode Trace::GetMode() {
1059  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
1060  CHECK(the_trace_ != nullptr) << "Trace mode requested, but no trace currently running";
1061  return the_trace_->trace_mode_;
1062}
1063
1064size_t Trace::GetBufferSize() {
1065  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
1066  CHECK(the_trace_ != nullptr) << "Trace mode requested, but no trace currently running";
1067  return the_trace_->buffer_size_;
1068}
1069
1070bool Trace::IsTracingEnabled() {
1071  MutexLock mu(Thread::Current(), *Locks::trace_lock_);
1072  return the_trace_ != nullptr;
1073}
1074
1075}  // namespace art
1076