| 1 | // Copyright (c) 2013, the Dart project authors. Please see the AUTHORS file |
| 2 | // for details. All rights reserved. Use of this source code is governed by a |
| 3 | // BSD-style license that can be found in the LICENSE file. |
| 4 | |
| 5 | #ifndef RUNTIME_VM_PROFILER_H_ |
| 6 | #define RUNTIME_VM_PROFILER_H_ |
| 7 | |
| 8 | #include "platform/atomic.h" |
| 9 | |
| 10 | #include "vm/allocation.h" |
| 11 | #include "vm/bitfield.h" |
| 12 | #include "vm/code_observers.h" |
| 13 | #include "vm/globals.h" |
| 14 | #include "vm/growable_array.h" |
| 15 | #include "vm/native_symbol.h" |
| 16 | #include "vm/object.h" |
| 17 | #include "vm/tags.h" |
| 18 | #include "vm/thread_interrupter.h" |
| 19 | |
| 20 | // Profiler sampling and stack walking support. |
| 21 | // NOTE: For service related code, see profile_service.h. |
| 22 | |
| 23 | namespace dart { |
| 24 | |
| 25 | // Forward declarations. |
| 26 | class ProcessedSample; |
| 27 | class ProcessedSampleBuffer; |
| 28 | |
| 29 | class Sample; |
| 30 | class SampleBlock; |
| 31 | |
| 32 | #define PROFILER_COUNTERS(V) \ |
| 33 | V(bail_out_unknown_task) \ |
| 34 | V(bail_out_jump_to_exception_handler) \ |
| 35 | V(bail_out_check_isolate) \ |
| 36 | V(single_frame_sample_deoptimizing) \ |
| 37 | V(single_frame_sample_get_and_validate_stack_bounds) \ |
| 38 | V(stack_walker_native) \ |
| 39 | V(stack_walker_dart_exit) \ |
| 40 | V(stack_walker_dart) \ |
| 41 | V(stack_walker_none) \ |
| 42 | V(incomplete_sample_fp_bounds) \ |
| 43 | V(incomplete_sample_fp_step) \ |
| 44 | V(incomplete_sample_bad_pc) \ |
| 45 | V(sample_allocation_failure) |
| 46 | |
| 47 | struct ProfilerCounters { |
| 48 | #define DECLARE_PROFILER_COUNTER(name) RelaxedAtomic<int64_t> name; |
| 49 | PROFILER_COUNTERS(DECLARE_PROFILER_COUNTER) |
| 50 | #undef DECLARE_PROFILER_COUNTER |
| 51 | }; |
| 52 | |
| 53 | class Profiler : public AllStatic { |
| 54 | public: |
| 55 | static void Init(); |
| 56 | static void Cleanup(); |
| 57 | |
| 58 | static void SetSampleDepth(intptr_t depth); |
| 59 | static void SetSamplePeriod(intptr_t period); |
| 60 | // Restarts sampling with a given profile period. This is called after the |
| 61 | // profile period is changed via the service protocol. |
| 62 | static void UpdateSamplePeriod(); |
| 63 | // Starts or shuts down the profiler after --profiler is changed via the |
| 64 | // service protocol. |
| 65 | static void UpdateRunningState(); |
| 66 | |
| 67 | static SampleBlockBuffer* sample_block_buffer() { |
| 68 | return sample_block_buffer_; |
| 69 | } |
| 70 | static void set_sample_block_buffer(SampleBlockBuffer* buffer) { |
| 71 | sample_block_buffer_ = buffer; |
| 72 | } |
| 73 | |
| 74 | static void DumpStackTrace(void* context); |
| 75 | static void DumpStackTrace(bool for_crash = true); |
| 76 | |
| 77 | static void SampleAllocation(Thread* thread, |
| 78 | intptr_t cid, |
| 79 | uint32_t identity_hash); |
| 80 | |
| 81 | // SampleThread is called from inside the signal handler and hence it is very |
| 82 | // critical that the implementation of SampleThread does not do any of the |
| 83 | // following: |
| 84 | // * Accessing TLS -- Because on Fuchsia, Mac and Windows the callback will |
| 85 | // be running in a different thread. |
| 86 | // * Allocating memory -- Because this takes locks which may already be |
| 87 | // held, resulting in a dead lock. |
| 88 | // * Taking a lock -- See above. |
| 89 | static void SampleThread(Thread* thread, const InterruptedThreadState& state); |
| 90 | |
| 91 | static ProfilerCounters counters() { |
| 92 | // Copies the counter values. |
| 93 | return counters_; |
| 94 | } |
| 95 | inline static intptr_t Size(); |
| 96 | |
| 97 | static void ProcessCompletedBlocks(Isolate* isolate); |
| 98 | static void IsolateShutdown(Thread* thread); |
| 99 | |
| 100 | private: |
| 101 | static void DumpStackTrace(uword sp, uword fp, uword pc, bool for_crash); |
| 102 | |
| 103 | // Calculates the sample buffer capacity. Returns |
| 104 | // SampleBuffer::kDefaultBufferCapacity if --sample-buffer-duration is not |
| 105 | // provided. Otherwise, the capacity is based on the sample rate, maximum |
| 106 | // sample stack depth, and the number of seconds of samples the sample buffer |
| 107 | // should be able to accomodate. |
| 108 | static intptr_t CalculateSampleBufferCapacity(); |
| 109 | |
| 110 | // Does not walk the thread's stack. |
| 111 | static void SampleThreadSingleFrame(Thread* thread, |
| 112 | Sample* sample, |
| 113 | uintptr_t pc); |
| 114 | static RelaxedAtomic<bool> initialized_; |
| 115 | |
| 116 | static SampleBlockBuffer* sample_block_buffer_; |
| 117 | |
| 118 | static ProfilerCounters counters_; |
| 119 | |
| 120 | friend class Thread; |
| 121 | }; |
| 122 | |
| 123 | class SampleVisitor : public ValueObject { |
| 124 | public: |
| 125 | explicit SampleVisitor(Dart_Port port) : port_(port), visited_(0) {} |
| 126 | virtual ~SampleVisitor() {} |
| 127 | |
| 128 | virtual void VisitSample(Sample* sample) = 0; |
| 129 | |
| 130 | virtual void Reset() { visited_ = 0; } |
| 131 | |
| 132 | intptr_t visited() const { return visited_; } |
| 133 | |
| 134 | void IncrementVisited() { visited_++; } |
| 135 | |
| 136 | Dart_Port port() const { return port_; } |
| 137 | |
| 138 | private: |
| 139 | Dart_Port port_; |
| 140 | intptr_t visited_; |
| 141 | |
| 142 | DISALLOW_IMPLICIT_CONSTRUCTORS(SampleVisitor); |
| 143 | }; |
| 144 | |
| 145 | class SampleFilter : public ValueObject { |
| 146 | public: |
| 147 | SampleFilter(Dart_Port port, |
| 148 | intptr_t thread_task_mask, |
| 149 | int64_t time_origin_micros, |
| 150 | int64_t time_extent_micros, |
| 151 | bool take_samples = false) |
| 152 | : port_(port), |
| 153 | thread_task_mask_(thread_task_mask), |
| 154 | time_origin_micros_(time_origin_micros), |
| 155 | time_extent_micros_(time_extent_micros), |
| 156 | take_samples_(take_samples) { |
| 157 | ASSERT(thread_task_mask != 0); |
| 158 | ASSERT(time_origin_micros_ >= -1); |
| 159 | ASSERT(time_extent_micros_ >= -1); |
| 160 | } |
| 161 | virtual ~SampleFilter() {} |
| 162 | |
| 163 | // Override this function. |
| 164 | // Return |true| if |sample| passes the filter. |
| 165 | virtual bool FilterSample(Sample* sample) { return true; } |
| 166 | |
| 167 | Dart_Port port() const { return port_; } |
| 168 | |
| 169 | // Returns |true| if |sample| passes the time filter. |
| 170 | bool TimeFilterSample(Sample* sample); |
| 171 | |
| 172 | // Returns |true| if |sample| passes the thread task filter. |
| 173 | bool TaskFilterSample(Sample* sample); |
| 174 | |
| 175 | bool take_samples() const { return take_samples_; } |
| 176 | |
| 177 | static constexpr intptr_t kNoTaskFilter = -1; |
| 178 | |
| 179 | private: |
| 180 | Dart_Port port_; |
| 181 | intptr_t thread_task_mask_; |
| 182 | int64_t time_origin_micros_; |
| 183 | int64_t time_extent_micros_; |
| 184 | bool take_samples_; |
| 185 | }; |
| 186 | |
| 187 | class ClearProfileVisitor : public SampleVisitor { |
| 188 | public: |
| 189 | explicit ClearProfileVisitor(Isolate* isolate); |
| 190 | |
| 191 | virtual void VisitSample(Sample* sample); |
| 192 | }; |
| 193 | |
| 194 | // Each Sample holds a stack trace from an isolate. |
| 195 | class Sample { |
| 196 | public: |
| 197 | Sample() = default; |
| 198 | |
| 199 | void Init(Dart_Port port, int64_t timestamp, ThreadId tid) { |
| 200 | Clear(); |
| 201 | timestamp_ = timestamp; |
| 202 | tid_ = tid; |
| 203 | port_ = port; |
| 204 | next_ = nullptr; |
| 205 | } |
| 206 | |
| 207 | Dart_Port port() const { return port_; } |
| 208 | |
| 209 | // Thread sample was taken on. |
| 210 | ThreadId tid() const { return tid_; } |
| 211 | |
| 212 | void Clear() { |
| 213 | timestamp_ = 0; |
| 214 | port_ = ILLEGAL_PORT; |
| 215 | tid_ = OSThread::kInvalidThreadId; |
| 216 | for (intptr_t i = 0; i < kStackBufferSizeInWords; i++) { |
| 217 | stack_buffer_[i] = 0; |
| 218 | } |
| 219 | for (intptr_t i = 0; i < kPCArraySizeInWords; i++) { |
| 220 | pc_array_[i] = 0; |
| 221 | } |
| 222 | vm_tag_ = VMTag::kInvalidTagId; |
| 223 | user_tag_ = UserTags::kDefaultUserTag; |
| 224 | state_ = 0; |
| 225 | next_ = nullptr; |
| 226 | allocation_identity_hash_ = 0; |
| 227 | set_head_sample(true); |
| 228 | } |
| 229 | |
| 230 | // Timestamp sample was taken at. |
| 231 | int64_t timestamp() const { return timestamp_; } |
| 232 | |
| 233 | // Top most pc. |
| 234 | uword pc() const { return At(i: 0); } |
| 235 | |
| 236 | // Get stack trace entry. |
| 237 | uword At(intptr_t i) const { |
| 238 | ASSERT(i >= 0); |
| 239 | ASSERT(i < kPCArraySizeInWords); |
| 240 | return pc_array_[i]; |
| 241 | } |
| 242 | |
| 243 | // Set stack trace entry. |
| 244 | void SetAt(intptr_t i, uword pc) { |
| 245 | ASSERT(i >= 0); |
| 246 | ASSERT(i < kPCArraySizeInWords); |
| 247 | pc_array_[i] = pc; |
| 248 | } |
| 249 | |
| 250 | void DumpStackTrace() { |
| 251 | for (intptr_t i = 0; i < kPCArraySizeInWords; ++i) { |
| 252 | uintptr_t start = 0; |
| 253 | uword pc = At(i); |
| 254 | char* native_symbol_name = |
| 255 | NativeSymbolResolver::LookupSymbolName(pc, start: &start); |
| 256 | if (native_symbol_name == nullptr) { |
| 257 | OS::PrintErr(format: " [0x%" Pp "] Unknown symbol\n" , pc); |
| 258 | } else { |
| 259 | OS::PrintErr(format: " [0x%" Pp "] %s\n" , pc, native_symbol_name); |
| 260 | NativeSymbolResolver::FreeSymbolName(name: native_symbol_name); |
| 261 | } |
| 262 | } |
| 263 | } |
| 264 | |
| 265 | uword vm_tag() const { return vm_tag_; } |
| 266 | void set_vm_tag(uword tag) { |
| 267 | ASSERT(tag != VMTag::kInvalidTagId); |
| 268 | vm_tag_ = tag; |
| 269 | } |
| 270 | |
| 271 | uword user_tag() const { return user_tag_; } |
| 272 | void set_user_tag(uword tag) { user_tag_ = tag; } |
| 273 | |
| 274 | bool leaf_frame_is_dart() const { return LeafFrameIsDart::decode(value: state_); } |
| 275 | |
| 276 | void set_leaf_frame_is_dart(bool leaf_frame_is_dart) { |
| 277 | state_ = LeafFrameIsDart::update(value: leaf_frame_is_dart, original: state_); |
| 278 | } |
| 279 | |
| 280 | bool ignore_sample() const { return IgnoreBit::decode(value: state_); } |
| 281 | |
| 282 | void set_ignore_sample(bool ignore_sample) { |
| 283 | state_ = IgnoreBit::update(value: ignore_sample, original: state_); |
| 284 | } |
| 285 | |
| 286 | bool exit_frame_sample() const { return ExitFrameBit::decode(value: state_); } |
| 287 | |
| 288 | void set_exit_frame_sample(bool exit_frame_sample) { |
| 289 | state_ = ExitFrameBit::update(value: exit_frame_sample, original: state_); |
| 290 | } |
| 291 | |
| 292 | bool missing_frame_inserted() const { |
| 293 | return MissingFrameInsertedBit::decode(value: state_); |
| 294 | } |
| 295 | |
| 296 | void set_missing_frame_inserted(bool missing_frame_inserted) { |
| 297 | state_ = MissingFrameInsertedBit::update(value: missing_frame_inserted, original: state_); |
| 298 | } |
| 299 | |
| 300 | bool truncated_trace() const { return TruncatedTraceBit::decode(value: state_); } |
| 301 | |
| 302 | void set_truncated_trace(bool truncated_trace) { |
| 303 | state_ = TruncatedTraceBit::update(value: truncated_trace, original: state_); |
| 304 | } |
| 305 | |
| 306 | bool is_allocation_sample() const { |
| 307 | return ClassAllocationSampleBit::decode(value: state_); |
| 308 | } |
| 309 | |
| 310 | void set_is_allocation_sample(bool allocation_sample) { |
| 311 | state_ = ClassAllocationSampleBit::update(value: allocation_sample, original: state_); |
| 312 | } |
| 313 | |
| 314 | uint32_t allocation_identity_hash() const { |
| 315 | return allocation_identity_hash_; |
| 316 | } |
| 317 | |
| 318 | void set_allocation_identity_hash(uint32_t hash) { |
| 319 | allocation_identity_hash_ = hash; |
| 320 | } |
| 321 | |
| 322 | Thread::TaskKind thread_task() const { return ThreadTaskBit::decode(value: state_); } |
| 323 | |
| 324 | void set_thread_task(Thread::TaskKind task) { |
| 325 | state_ = ThreadTaskBit::update(value: task, original: state_); |
| 326 | } |
| 327 | |
| 328 | bool is_continuation_sample() const { |
| 329 | return ContinuationSampleBit::decode(value: state_); |
| 330 | } |
| 331 | |
| 332 | void SetContinuation(Sample* next) { |
| 333 | ASSERT(!is_continuation_sample()); |
| 334 | ASSERT(next_ == nullptr); |
| 335 | state_ = ContinuationSampleBit::update(value: true, original: state_); |
| 336 | next_ = next; |
| 337 | } |
| 338 | |
| 339 | Sample* continuation_sample() const { return next_; } |
| 340 | |
| 341 | intptr_t allocation_cid() const { |
| 342 | ASSERT(is_allocation_sample()); |
| 343 | return metadata(); |
| 344 | } |
| 345 | |
| 346 | void set_head_sample(bool head_sample) { |
| 347 | state_ = HeadSampleBit::update(value: head_sample, original: state_); |
| 348 | } |
| 349 | |
| 350 | bool head_sample() const { return HeadSampleBit::decode(value: state_); } |
| 351 | |
| 352 | intptr_t metadata() const { return MetadataBits::decode(value: state_); } |
| 353 | void set_metadata(intptr_t metadata) { |
| 354 | state_ = MetadataBits::update(value: metadata, original: state_); |
| 355 | } |
| 356 | |
| 357 | void SetAllocationCid(intptr_t cid) { |
| 358 | set_is_allocation_sample(true); |
| 359 | set_metadata(cid); |
| 360 | } |
| 361 | |
| 362 | static constexpr int kPCArraySizeInWords = 32; |
| 363 | uword* GetPCArray() { return &pc_array_[0]; } |
| 364 | |
| 365 | static constexpr int kStackBufferSizeInWords = 2; |
| 366 | uword* GetStackBuffer() { return &stack_buffer_[0]; } |
| 367 | |
| 368 | private: |
| 369 | enum StateBits { |
| 370 | kHeadSampleBit = 0, |
| 371 | kLeafFrameIsDartBit = 1, |
| 372 | kIgnoreBit = 2, |
| 373 | kExitFrameBit = 3, |
| 374 | kMissingFrameInsertedBit = 4, |
| 375 | kTruncatedTraceBit = 5, |
| 376 | kClassAllocationSampleBit = 6, |
| 377 | kContinuationSampleBit = 7, |
| 378 | kThreadTaskBit = 8, // 7 bits. |
| 379 | kMetadataBit = 15, // 16 bits. |
| 380 | kNextFreeBit = 31, |
| 381 | }; |
| 382 | class HeadSampleBit : public BitField<uint32_t, bool, kHeadSampleBit, 1> {}; |
| 383 | class LeafFrameIsDart |
| 384 | : public BitField<uint32_t, bool, kLeafFrameIsDartBit, 1> {}; |
| 385 | class IgnoreBit : public BitField<uint32_t, bool, kIgnoreBit, 1> {}; |
| 386 | class ExitFrameBit : public BitField<uint32_t, bool, kExitFrameBit, 1> {}; |
| 387 | class MissingFrameInsertedBit |
| 388 | : public BitField<uint32_t, bool, kMissingFrameInsertedBit, 1> {}; |
| 389 | class TruncatedTraceBit |
| 390 | : public BitField<uint32_t, bool, kTruncatedTraceBit, 1> {}; |
| 391 | class ClassAllocationSampleBit |
| 392 | : public BitField<uint32_t, bool, kClassAllocationSampleBit, 1> {}; |
| 393 | class ContinuationSampleBit |
| 394 | : public BitField<uint32_t, bool, kContinuationSampleBit, 1> {}; |
| 395 | class ThreadTaskBit |
| 396 | : public BitField<uint32_t, Thread::TaskKind, kThreadTaskBit, 7> {}; |
| 397 | class MetadataBits : public BitField<uint32_t, intptr_t, kMetadataBit, 16> {}; |
| 398 | |
| 399 | int64_t timestamp_; |
| 400 | Dart_Port port_; |
| 401 | ThreadId tid_; |
| 402 | uword stack_buffer_[kStackBufferSizeInWords]; |
| 403 | uword pc_array_[kPCArraySizeInWords]; |
| 404 | uword vm_tag_; |
| 405 | uword user_tag_; |
| 406 | uint32_t state_; |
| 407 | Sample* next_; |
| 408 | uint32_t allocation_identity_hash_; |
| 409 | |
| 410 | DISALLOW_COPY_AND_ASSIGN(Sample); |
| 411 | }; |
| 412 | |
| 413 | class AbstractCode { |
| 414 | public: |
| 415 | explicit AbstractCode(ObjectPtr code) : code_(Object::Handle(ptr: code)) { |
| 416 | ASSERT(code_.IsNull() || code_.IsCode()); |
| 417 | } |
| 418 | |
| 419 | ObjectPtr ptr() const { return code_.ptr(); } |
| 420 | const Object* handle() const { return &code_; } |
| 421 | |
| 422 | uword PayloadStart() const { |
| 423 | ASSERT(code_.IsCode()); |
| 424 | return Code::Cast(obj: code_).PayloadStart(); |
| 425 | } |
| 426 | |
| 427 | uword Size() const { |
| 428 | ASSERT(code_.IsCode()); |
| 429 | return Code::Cast(obj: code_).Size(); |
| 430 | } |
| 431 | |
| 432 | int64_t compile_timestamp() const { |
| 433 | if (code_.IsCode()) { |
| 434 | return Code::Cast(obj: code_).compile_timestamp(); |
| 435 | } else { |
| 436 | return 0; |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | const char* Name() const { |
| 441 | if (code_.IsCode()) { |
| 442 | return Code::Cast(obj: code_).Name(); |
| 443 | } else { |
| 444 | return "" ; |
| 445 | } |
| 446 | } |
| 447 | |
| 448 | const char* QualifiedName() const { |
| 449 | if (code_.IsCode()) { |
| 450 | return Code::Cast(obj: code_).QualifiedName( |
| 451 | params: NameFormattingParams(Object::kUserVisibleName)); |
| 452 | } else { |
| 453 | return "" ; |
| 454 | } |
| 455 | } |
| 456 | |
| 457 | bool IsStubCode() const { |
| 458 | if (code_.IsCode()) { |
| 459 | return Code::Cast(obj: code_).IsStubCode(); |
| 460 | } else { |
| 461 | return false; |
| 462 | } |
| 463 | } |
| 464 | |
| 465 | bool IsAllocationStubCode() const { |
| 466 | if (code_.IsCode()) { |
| 467 | return Code::Cast(obj: code_).IsAllocationStubCode(); |
| 468 | } else { |
| 469 | return false; |
| 470 | } |
| 471 | } |
| 472 | |
| 473 | bool IsTypeTestStubCode() const { |
| 474 | if (code_.IsCode()) { |
| 475 | return Code::Cast(obj: code_).IsTypeTestStubCode(); |
| 476 | } else { |
| 477 | return false; |
| 478 | } |
| 479 | } |
| 480 | |
| 481 | ObjectPtr owner() const { |
| 482 | if (code_.IsCode()) { |
| 483 | return Code::Cast(obj: code_).owner(); |
| 484 | } else { |
| 485 | return Object::null(); |
| 486 | } |
| 487 | } |
| 488 | |
| 489 | bool IsNull() const { return code_.IsNull(); } |
| 490 | bool IsCode() const { return code_.IsCode(); } |
| 491 | |
| 492 | bool is_optimized() const { |
| 493 | if (code_.IsCode()) { |
| 494 | return Code::Cast(obj: code_).is_optimized(); |
| 495 | } else { |
| 496 | return false; |
| 497 | } |
| 498 | } |
| 499 | |
| 500 | private: |
| 501 | const Object& code_; |
| 502 | }; |
| 503 | |
| 504 | // A Code object descriptor. |
| 505 | class CodeDescriptor : public ZoneAllocated { |
| 506 | public: |
| 507 | explicit CodeDescriptor(const AbstractCode code); |
| 508 | |
| 509 | uword Start() const; |
| 510 | |
| 511 | uword Size() const; |
| 512 | |
| 513 | int64_t CompileTimestamp() const; |
| 514 | |
| 515 | const AbstractCode code() const { return code_; } |
| 516 | |
| 517 | const char* Name() const { return code_.Name(); } |
| 518 | |
| 519 | bool Contains(uword pc) const { |
| 520 | uword end = Start() + Size(); |
| 521 | return (pc >= Start()) && (pc < end); |
| 522 | } |
| 523 | |
| 524 | static int Compare(CodeDescriptor* const* a, CodeDescriptor* const* b) { |
| 525 | ASSERT(a != nullptr); |
| 526 | ASSERT(b != nullptr); |
| 527 | |
| 528 | uword a_start = (*a)->Start(); |
| 529 | uword b_start = (*b)->Start(); |
| 530 | |
| 531 | if (a_start < b_start) { |
| 532 | return -1; |
| 533 | } else if (a_start > b_start) { |
| 534 | return 1; |
| 535 | } else { |
| 536 | return 0; |
| 537 | } |
| 538 | } |
| 539 | |
| 540 | private: |
| 541 | const AbstractCode code_; |
| 542 | |
| 543 | DISALLOW_COPY_AND_ASSIGN(CodeDescriptor); |
| 544 | }; |
| 545 | |
| 546 | // Fast lookup of Dart code objects. |
| 547 | class CodeLookupTable : public ZoneAllocated { |
| 548 | public: |
| 549 | explicit CodeLookupTable(Thread* thread); |
| 550 | |
| 551 | intptr_t length() const { return code_objects_.length(); } |
| 552 | |
| 553 | const CodeDescriptor* At(intptr_t index) const { |
| 554 | return code_objects_.At(index); |
| 555 | } |
| 556 | |
| 557 | const CodeDescriptor* FindCode(uword pc) const; |
| 558 | |
| 559 | private: |
| 560 | void Build(Thread* thread); |
| 561 | |
| 562 | void Add(const Object& code); |
| 563 | |
| 564 | // Code objects sorted by entry. |
| 565 | ZoneGrowableArray<CodeDescriptor*> code_objects_; |
| 566 | |
| 567 | friend class CodeLookupTableBuilder; |
| 568 | |
| 569 | DISALLOW_COPY_AND_ASSIGN(CodeLookupTable); |
| 570 | }; |
| 571 | |
| 572 | // Interface for a class that can create a ProcessedSampleBuffer. |
| 573 | class ProcessedSampleBufferBuilder { |
| 574 | public: |
| 575 | virtual ~ProcessedSampleBufferBuilder() = default; |
| 576 | virtual ProcessedSampleBuffer* BuildProcessedSampleBuffer( |
| 577 | SampleFilter* filter, |
| 578 | ProcessedSampleBuffer* buffer = nullptr) = 0; |
| 579 | }; |
| 580 | |
| 581 | class SampleBuffer : public ProcessedSampleBufferBuilder { |
| 582 | public: |
| 583 | SampleBuffer() = default; |
| 584 | virtual ~SampleBuffer() = default; |
| 585 | |
| 586 | virtual void Init(Sample* samples, intptr_t capacity) { |
| 587 | ASSERT(samples != nullptr); |
| 588 | ASSERT(capacity > 0); |
| 589 | samples_ = samples; |
| 590 | capacity_ = capacity; |
| 591 | } |
| 592 | |
| 593 | void VisitSamples(SampleVisitor* visitor) { |
| 594 | ASSERT(visitor != nullptr); |
| 595 | const intptr_t length = capacity(); |
| 596 | for (intptr_t i = 0; i < length; i++) { |
| 597 | Sample* sample = At(idx: i); |
| 598 | if (!sample->head_sample()) { |
| 599 | // An inner sample in a chain of samples. |
| 600 | continue; |
| 601 | } |
| 602 | if (sample->ignore_sample()) { |
| 603 | // Bad sample. |
| 604 | continue; |
| 605 | } |
| 606 | if (sample->port() != visitor->port()) { |
| 607 | // Another isolate. |
| 608 | continue; |
| 609 | } |
| 610 | if (sample->timestamp() == 0) { |
| 611 | // Empty. |
| 612 | continue; |
| 613 | } |
| 614 | if (sample->At(i: 0) == 0) { |
| 615 | // No frames. |
| 616 | continue; |
| 617 | } |
| 618 | visitor->IncrementVisited(); |
| 619 | visitor->VisitSample(sample); |
| 620 | } |
| 621 | } |
| 622 | |
| 623 | virtual Sample* ReserveSample() = 0; |
| 624 | virtual Sample* ReserveSampleAndLink(Sample* previous) = 0; |
| 625 | |
| 626 | Sample* At(intptr_t idx) const { |
| 627 | ASSERT(idx >= 0); |
| 628 | ASSERT(idx < capacity_); |
| 629 | return &samples_[idx]; |
| 630 | } |
| 631 | |
| 632 | intptr_t capacity() const { return capacity_; } |
| 633 | |
| 634 | virtual ProcessedSampleBuffer* BuildProcessedSampleBuffer( |
| 635 | SampleFilter* filter, |
| 636 | ProcessedSampleBuffer* buffer = nullptr); |
| 637 | |
| 638 | protected: |
| 639 | Sample* Next(Sample* sample); |
| 640 | |
| 641 | ProcessedSample* BuildProcessedSample(Sample* sample, |
| 642 | const CodeLookupTable& clt); |
| 643 | |
| 644 | Sample* samples_; |
| 645 | intptr_t capacity_; |
| 646 | |
| 647 | DISALLOW_COPY_AND_ASSIGN(SampleBuffer); |
| 648 | }; |
| 649 | |
| 650 | class SampleBlock : public SampleBuffer { |
| 651 | public: |
| 652 | // The default number of samples per block. Overridden by some tests. |
| 653 | static constexpr intptr_t kSamplesPerBlock = 100; |
| 654 | |
| 655 | SampleBlock() = default; |
| 656 | virtual ~SampleBlock() = default; |
| 657 | |
| 658 | // Returns the number of samples contained within this block. |
| 659 | intptr_t capacity() const { return capacity_; } |
| 660 | |
| 661 | Isolate* owner() const { return owner_; } |
| 662 | void set_owner(Isolate* isolate) { owner_ = isolate; } |
| 663 | |
| 664 | virtual Sample* ReserveSample(); |
| 665 | virtual Sample* ReserveSampleAndLink(Sample* previous); |
| 666 | |
| 667 | bool TryAllocateFree() { |
| 668 | State expected = kFree; |
| 669 | State desired = kSampling; |
| 670 | std::memory_order success_order = std::memory_order_acquire; |
| 671 | std::memory_order failure_order = std::memory_order_relaxed; |
| 672 | return state_.compare_exchange_strong(e&: expected, d: desired, s: success_order, |
| 673 | f: failure_order); |
| 674 | } |
| 675 | bool TryAllocateCompleted() { |
| 676 | State expected = kCompleted; |
| 677 | State desired = kSampling; |
| 678 | std::memory_order success_order = std::memory_order_acquire; |
| 679 | std::memory_order failure_order = std::memory_order_relaxed; |
| 680 | if (state_.compare_exchange_strong(e&: expected, d: desired, s: success_order, |
| 681 | f: failure_order)) { |
| 682 | owner_ = nullptr; |
| 683 | cursor_ = 0; |
| 684 | return true; |
| 685 | } |
| 686 | return false; |
| 687 | } |
| 688 | void MarkCompleted() { |
| 689 | ASSERT(state_.load(std::memory_order_relaxed) == kSampling); |
| 690 | state_.store(d: kCompleted, m: std::memory_order_release); |
| 691 | } |
| 692 | bool TryAcquireStreaming(Isolate* isolate) { |
| 693 | if (state_.load(m: std::memory_order_relaxed) != kCompleted) return false; |
| 694 | if (owner_ != isolate) return false; |
| 695 | |
| 696 | State expected = kCompleted; |
| 697 | State desired = kStreaming; |
| 698 | std::memory_order success_order = std::memory_order_acquire; |
| 699 | std::memory_order failure_order = std::memory_order_relaxed; |
| 700 | return state_.compare_exchange_strong(e&: expected, d: desired, s: success_order, |
| 701 | f: failure_order); |
| 702 | } |
| 703 | void StreamingToCompleted() { |
| 704 | ASSERT(state_.load(std::memory_order_relaxed) == kStreaming); |
| 705 | state_.store(d: kCompleted, m: std::memory_order_relaxed); |
| 706 | } |
| 707 | void StreamingToFree() { |
| 708 | ASSERT(state_.load(std::memory_order_relaxed) == kStreaming); |
| 709 | owner_ = nullptr; |
| 710 | cursor_ = 0; |
| 711 | state_.store(d: kFree, m: std::memory_order_release); |
| 712 | } |
| 713 | void FreeCompleted() { |
| 714 | State expected = kCompleted; |
| 715 | State desired = kStreaming; |
| 716 | std::memory_order success_order = std::memory_order_acquire; |
| 717 | std::memory_order failure_order = std::memory_order_relaxed; |
| 718 | if (state_.compare_exchange_strong(e&: expected, d: desired, s: success_order, |
| 719 | f: failure_order)) { |
| 720 | StreamingToFree(); |
| 721 | } |
| 722 | } |
| 723 | |
| 724 | protected: |
| 725 | bool HasStreamableSamples(const GrowableObjectArray& tag_table, UserTag* tag); |
| 726 | |
| 727 | enum State : uint32_t { |
| 728 | kFree, |
| 729 | kSampling, // I.e., writing. |
| 730 | kCompleted, |
| 731 | kStreaming, // I.e., reading. |
| 732 | }; |
| 733 | std::atomic<State> state_ = kFree; |
| 734 | RelaxedAtomic<uint32_t> cursor_ = 0; |
| 735 | Isolate* owner_ = nullptr; |
| 736 | |
| 737 | private: |
| 738 | friend class SampleBlockListProcessor; |
| 739 | friend class SampleBlockBuffer; |
| 740 | |
| 741 | DISALLOW_COPY_AND_ASSIGN(SampleBlock); |
| 742 | }; |
| 743 | |
| 744 | class SampleBlockBuffer : public ProcessedSampleBufferBuilder { |
| 745 | public: |
| 746 | static constexpr intptr_t kDefaultBlockCount = 600; |
| 747 | |
| 748 | // Creates a SampleBlockBuffer with a predetermined number of blocks. |
| 749 | // |
| 750 | // Defaults to kDefaultBlockCount blocks. Block size is fixed to |
| 751 | // SampleBlock::kSamplesPerBlock samples per block, except for in tests. |
| 752 | explicit SampleBlockBuffer( |
| 753 | intptr_t blocks = kDefaultBlockCount, |
| 754 | intptr_t samples_per_block = SampleBlock::kSamplesPerBlock); |
| 755 | |
| 756 | virtual ~SampleBlockBuffer(); |
| 757 | |
| 758 | void VisitSamples(SampleVisitor* visitor) { |
| 759 | ASSERT(visitor != nullptr); |
| 760 | for (intptr_t i = 0; i < capacity_; ++i) { |
| 761 | blocks_[i].VisitSamples(visitor); |
| 762 | } |
| 763 | } |
| 764 | |
| 765 | void FreeCompletedBlocks(); |
| 766 | |
| 767 | // Reserves a sample for a CPU profile. |
| 768 | // |
| 769 | // Returns nullptr when a sample can't be reserved. |
| 770 | Sample* ReserveCPUSample(Isolate* isolate); |
| 771 | |
| 772 | // Reserves a sample for a Dart object allocation profile. |
| 773 | // |
| 774 | // Returns nullptr when a sample can't be reserved. |
| 775 | Sample* ReserveAllocationSample(Isolate* isolate); |
| 776 | |
| 777 | intptr_t Size() const { return memory_->size(); } |
| 778 | |
| 779 | virtual ProcessedSampleBuffer* BuildProcessedSampleBuffer( |
| 780 | SampleFilter* filter, |
| 781 | ProcessedSampleBuffer* buffer = nullptr); |
| 782 | |
| 783 | private: |
| 784 | Sample* ReserveSampleImpl(Isolate* isolate, bool allocation_sample); |
| 785 | |
| 786 | // Returns nullptr if there are no available blocks. |
| 787 | SampleBlock* ReserveSampleBlock(); |
| 788 | |
| 789 | // Sample block management. |
| 790 | RelaxedAtomic<int> cursor_; |
| 791 | SampleBlock* blocks_; |
| 792 | intptr_t capacity_; |
| 793 | |
| 794 | // Sample buffer management. |
| 795 | VirtualMemory* memory_; |
| 796 | Sample* sample_buffer_; |
| 797 | |
| 798 | friend class Isolate; |
| 799 | DISALLOW_COPY_AND_ASSIGN(SampleBlockBuffer); |
| 800 | }; |
| 801 | |
| 802 | class StreamingSampleBufferBuilder : public ProcessedSampleBufferBuilder { |
| 803 | public: |
| 804 | explicit StreamingSampleBufferBuilder(Isolate* isolate) : isolate_(isolate) {} |
| 805 | |
| 806 | virtual ProcessedSampleBuffer* BuildProcessedSampleBuffer( |
| 807 | SampleFilter* filter, |
| 808 | ProcessedSampleBuffer* buffer = nullptr); |
| 809 | |
| 810 | // Returns true when at least one sample in the sample block list has a user |
| 811 | // tag with CPU sample streaming enabled. |
| 812 | bool HasStreamableSamples(Thread* thread); |
| 813 | |
| 814 | private: |
| 815 | Isolate* isolate_; |
| 816 | |
| 817 | DISALLOW_COPY_AND_ASSIGN(StreamingSampleBufferBuilder); |
| 818 | }; |
| 819 | |
| 820 | intptr_t Profiler::Size() { |
| 821 | intptr_t size = 0; |
| 822 | if (sample_block_buffer_ != nullptr) { |
| 823 | size += sample_block_buffer_->Size(); |
| 824 | } |
| 825 | return size; |
| 826 | } |
| 827 | |
| 828 | // A |ProcessedSample| is a combination of 1 (or more) |Sample|(s) that have |
| 829 | // been merged into a logical sample. The raw data may have been processed to |
| 830 | // improve the quality of the stack trace. |
| 831 | class ProcessedSample : public ZoneAllocated { |
| 832 | public: |
| 833 | ProcessedSample(); |
| 834 | |
| 835 | // Add |pc| to stack trace. |
| 836 | void Add(uword pc) { pcs_.Add(value: pc); } |
| 837 | |
| 838 | // Insert |pc| at |index|. |
| 839 | void InsertAt(intptr_t index, uword pc) { pcs_.InsertAt(idx: index, value: pc); } |
| 840 | |
| 841 | // Number of pcs in stack trace. |
| 842 | intptr_t length() const { return pcs_.length(); } |
| 843 | |
| 844 | // Get |pc| at |index|. |
| 845 | uword At(intptr_t index) const { |
| 846 | ASSERT(index >= 0); |
| 847 | ASSERT(index < length()); |
| 848 | return pcs_[index]; |
| 849 | } |
| 850 | |
| 851 | // Timestamp sample was taken at. |
| 852 | int64_t timestamp() const { return timestamp_; } |
| 853 | void set_timestamp(int64_t timestamp) { timestamp_ = timestamp; } |
| 854 | |
| 855 | ThreadId tid() const { return tid_; } |
| 856 | void set_tid(ThreadId tid) { tid_ = tid; } |
| 857 | |
| 858 | // The VM tag. |
| 859 | uword vm_tag() const { return vm_tag_; } |
| 860 | void set_vm_tag(uword tag) { vm_tag_ = tag; } |
| 861 | |
| 862 | // The user tag. |
| 863 | uword user_tag() const { return user_tag_; } |
| 864 | void set_user_tag(uword tag) { user_tag_ = tag; } |
| 865 | |
| 866 | // The class id if this is an allocation profile sample. -1 otherwise. |
| 867 | intptr_t allocation_cid() const { return allocation_cid_; } |
| 868 | void set_allocation_cid(intptr_t cid) { allocation_cid_ = cid; } |
| 869 | |
| 870 | // The identity hash code of the allocated object if this is an allocation |
| 871 | // profile sample. -1 otherwise. |
| 872 | uint32_t allocation_identity_hash() const { |
| 873 | return allocation_identity_hash_; |
| 874 | } |
| 875 | void set_allocation_identity_hash(uint32_t hash) { |
| 876 | allocation_identity_hash_ = hash; |
| 877 | } |
| 878 | |
| 879 | bool IsAllocationSample() const { return allocation_cid_ > 0; } |
| 880 | |
| 881 | // Was the stack trace truncated? |
| 882 | bool truncated() const { return truncated_; } |
| 883 | void set_truncated(bool truncated) { truncated_ = truncated; } |
| 884 | |
| 885 | // Was the first frame in the stack trace executing? |
| 886 | bool first_frame_executing() const { return first_frame_executing_; } |
| 887 | void set_first_frame_executing(bool first_frame_executing) { |
| 888 | first_frame_executing_ = first_frame_executing; |
| 889 | } |
| 890 | |
| 891 | private: |
| 892 | void FixupCaller(const CodeLookupTable& clt, |
| 893 | uword pc_marker, |
| 894 | uword* stack_buffer); |
| 895 | |
| 896 | void CheckForMissingDartFrame(const CodeLookupTable& clt, |
| 897 | const CodeDescriptor* code, |
| 898 | uword pc_marker, |
| 899 | uword* stack_buffer); |
| 900 | |
| 901 | ZoneGrowableArray<uword> pcs_; |
| 902 | int64_t timestamp_; |
| 903 | ThreadId tid_; |
| 904 | uword vm_tag_; |
| 905 | uword user_tag_; |
| 906 | intptr_t allocation_cid_; |
| 907 | uint32_t allocation_identity_hash_; |
| 908 | bool truncated_; |
| 909 | bool first_frame_executing_; |
| 910 | |
| 911 | friend class SampleBuffer; |
| 912 | DISALLOW_COPY_AND_ASSIGN(ProcessedSample); |
| 913 | }; |
| 914 | |
| 915 | // A collection of |ProcessedSample|s. |
| 916 | class ProcessedSampleBuffer : public ZoneAllocated { |
| 917 | public: |
| 918 | ProcessedSampleBuffer(); |
| 919 | |
| 920 | void Add(ProcessedSample* sample) { samples_.Add(value: sample); } |
| 921 | |
| 922 | intptr_t length() const { return samples_.length(); } |
| 923 | |
| 924 | ProcessedSample* At(intptr_t index) { return samples_.At(index); } |
| 925 | |
| 926 | const CodeLookupTable& code_lookup_table() const { |
| 927 | return *code_lookup_table_; |
| 928 | } |
| 929 | |
| 930 | private: |
| 931 | ZoneGrowableArray<ProcessedSample*> samples_; |
| 932 | CodeLookupTable* code_lookup_table_; |
| 933 | |
| 934 | DISALLOW_COPY_AND_ASSIGN(ProcessedSampleBuffer); |
| 935 | }; |
| 936 | |
| 937 | class SampleBlockProcessor : public AllStatic { |
| 938 | public: |
| 939 | static void Init(); |
| 940 | |
| 941 | static void Startup(); |
| 942 | static void Cleanup(); |
| 943 | |
| 944 | private: |
| 945 | static constexpr intptr_t kMaxThreads = 4096; |
| 946 | static bool initialized_; |
| 947 | static bool shutdown_; |
| 948 | static bool thread_running_; |
| 949 | static ThreadJoinId processor_thread_id_; |
| 950 | static Monitor* monitor_; |
| 951 | |
| 952 | static void ThreadMain(uword parameters); |
| 953 | }; |
| 954 | |
| 955 | } // namespace dart |
| 956 | |
| 957 | #endif // RUNTIME_VM_PROFILER_H_ |
| 958 | |