1 /* 2 * Copyright (c) 2017, Google and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 27 #include "gc/shared/collectedHeap.hpp" 28 #include "memory/universe.hpp" 29 #include "runtime/heapMonitoring.hpp" 30 #include "runtime/vframe.hpp" 31 32 static const int MaxStackDepth = 1024; 33 34 // Internal data structure representing traces. 35 struct StackTraceData : CHeapObj<mtInternal> { 36 jvmtiStackTrace *trace; 37 oop obj; 38 int references; 39 40 StackTraceData(jvmtiStackTrace *t, oop o) : trace(t), obj(o), references(0) {} 41 42 StackTraceData() : trace(NULL), obj(NULL), references(0) {} 43 44 // StackTraceDatas are shared around the board between various lists. So 45 // handle this by hand instead of having this in the destructor. There are 46 // cases where the struct is on the stack but holding heap data not to be 47 // freed. 48 static void free_data(StackTraceData *data) { 49 if (data->trace != NULL) { 50 FREE_C_HEAP_ARRAY(jvmtiFrameInfo, data->trace->frames); 51 FREE_C_HEAP_OBJ(data->trace); 52 } 53 delete data; 54 } 55 }; 56 57 // Fixed size buffer for holding garbage traces. 58 class GarbageTracesBuffer : public CHeapObj<mtInternal> { 59 public: 60 GarbageTracesBuffer(uint32_t size) : _size(size) { 61 _garbage_traces = NEW_C_HEAP_ARRAY(StackTraceData*, 62 size, 63 mtInternal); 64 memset(_garbage_traces, 0, sizeof(StackTraceData*) * size); 65 } 66 67 virtual ~GarbageTracesBuffer() { 68 FREE_C_HEAP_ARRAY(StackTraceData*, _garbage_traces); 69 } 70 71 StackTraceData** get_traces() const { 72 return _garbage_traces; 73 } 74 75 bool store_trace(StackTraceData *trace) { 76 uint32_t index; 77 if (!select_replacement(&index)) { 78 return false; 79 } 80 81 StackTraceData *old_data = _garbage_traces[index]; 82 83 if (old_data != NULL) { 84 old_data->references--; 85 86 if (old_data->references == 0) { 87 StackTraceData::free_data(old_data); 88 } 89 } 90 91 trace->references++; 92 _garbage_traces[index] = trace; 93 return true; 94 } 95 96 uint32_t size() const { 97 return _size; 98 } 99 100 protected: 101 // Subclasses select the trace to replace. Returns false if no replacement 102 // is to happen, otherwise stores the index of the trace to replace in 103 // *index. 104 virtual bool select_replacement(uint32_t *index) = 0; 105 106 const uint32_t _size; 107 108 private: 109 // The current garbage traces. A fixed-size ring buffer. 110 StackTraceData **_garbage_traces; 111 }; 112 113 // Keep statistical sample of traces over the lifetime of the server. 114 // When the buffer is full, replace a random entry with probability 115 // 1/samples_seen. This strategy tends towards preserving the most frequently 116 // occuring traces over time. 117 class FrequentGarbageTraces : public GarbageTracesBuffer { 118 public: 119 FrequentGarbageTraces(int size) 120 : GarbageTracesBuffer(size), 121 _garbage_traces_pos(0), 122 _samples_seen(0) { 123 } 124 125 virtual ~FrequentGarbageTraces() { 126 } 127 128 virtual bool select_replacement(uint32_t* index) { 129 ++_samples_seen; 130 131 if (_garbage_traces_pos < _size) { 132 *index = _garbage_traces_pos++; 133 return true; 134 } 135 136 uint64_t random_uint64 = 137 (static_cast<uint64_t>(::random()) << 32) | ::random(); 138 139 uint32_t random_index = random_uint64 % _samples_seen; 140 if (random_index < _size) { 141 *index = random_index; 142 return true; 143 } 144 145 return false; 146 } 147 148 private: 149 // The current position in the buffer as we initially fill it. 150 uint32_t _garbage_traces_pos; 151 152 uint64_t _samples_seen; 153 }; 154 155 // Store most recent garbage traces. 156 class MostRecentGarbageTraces : public GarbageTracesBuffer { 157 public: 158 MostRecentGarbageTraces(int size) 159 : GarbageTracesBuffer(size), 160 _garbage_traces_pos(0) { 161 } 162 163 virtual ~MostRecentGarbageTraces() { 164 } 165 166 virtual bool select_replacement(uint32_t* index) { 167 *index = _garbage_traces_pos; 168 169 _garbage_traces_pos = 170 (_garbage_traces_pos + 1) % _size; 171 172 return true; 173 } 174 175 private: 176 // The current position in the buffer. 177 uint32_t _garbage_traces_pos; 178 }; 179 180 // Each object that we profile is stored as trace with the thread_id. 181 class StackTraceStorage : public CHeapObj<mtInternal> { 182 public: 183 // The function that gets called to add a trace to the list of 184 // traces we are maintaining. 185 void add_trace(jvmtiStackTrace *trace, oop o); 186 187 // The function that gets called by the client to retrieve the list 188 // of stack traces. Passes a jvmtiStackTraces which will get mutated. 189 void get_all_stack_traces(jvmtiStackTraces *traces); 190 191 // The function that gets called by the client to retrieve the list 192 // of stack traces. Passes a jvmtiStackTraces which will get mutated. 193 void get_garbage_stack_traces(jvmtiStackTraces *traces); 194 195 // The function that gets called by the client to retrieve the list 196 // of stack traces. Passes a jvmtiStackTraces which will get mutated. 197 void get_frequent_garbage_stack_traces(jvmtiStackTraces *traces); 198 199 // Executes whenever weak references are traversed. is_alive tells 200 // you if the given oop is still reachable and live. 201 void weak_oops_do(BoolObjectClosure* is_alive, OopClosure *f); 202 203 ~StackTraceStorage(); 204 StackTraceStorage(); 205 206 static StackTraceStorage* storage() { 207 static StackTraceStorage internal_storage; 208 return &internal_storage; 209 } 210 211 void initialize(int max_storage) { 212 MutexLocker mu(HeapMonitorStorage_lock); 213 free_storage(); 214 allocate_storage(max_storage); 215 memset(&_stats, 0, sizeof(_stats)); 216 } 217 218 const jvmtiHeapSamplingStats& get_heap_sampling_stats() const { 219 return _stats; 220 } 221 222 void accumulate_sample_rate(size_t rate) { 223 _stats.sample_rate_accumulation += rate; 224 _stats.sample_rate_count++; 225 } 226 227 bool initialized() { return _initialized; } 228 229 private: 230 // The traces currently sampled. 231 GrowableArray<StackTraceData> *_allocated_traces; 232 233 // Recent garbage traces. 234 MostRecentGarbageTraces *_recent_garbage_traces; 235 236 // Frequent garbage traces. 237 FrequentGarbageTraces *_frequent_garbage_traces; 238 239 // Heap Sampling statistics. 240 jvmtiHeapSamplingStats _stats; 241 242 // Maximum amount of storage provided by the JVMTI call initialize_profiling. 243 int _max_gc_storage; 244 245 static StackTraceStorage* internal_storage; 246 volatile bool _initialized; 247 248 // Support functions and classes for copying data to the external 249 // world. 250 class StackTraceDataCopier { 251 public: 252 virtual int size() const = 0; 253 virtual const StackTraceData *get(uint32_t i) const = 0; 254 }; 255 256 class LiveStackTraceDataCopier : public StackTraceDataCopier { 257 public: 258 LiveStackTraceDataCopier(GrowableArray<StackTraceData> *data) : 259 _data(data) {} 260 int size() const { return _data ? _data->length() : 0; } 261 const StackTraceData *get(uint32_t i) const { return _data->adr_at(i); } 262 263 private: 264 GrowableArray<StackTraceData> *_data; 265 }; 266 267 class GarbageStackTraceDataCopier : public StackTraceDataCopier { 268 public: 269 GarbageStackTraceDataCopier(StackTraceData **data, int size) : 270 _data(data), _size(size) {} 271 int size() const { return _size; } 272 const StackTraceData *get(uint32_t i) const { return _data[i]; } 273 274 private: 275 StackTraceData **_data; 276 int _size; 277 }; 278 279 // Copies from StackTraceData to jvmtiStackTrace. 280 bool deep_copy(jvmtiStackTrace *to, const StackTraceData *from); 281 282 // Creates a deep copy of the list of StackTraceData. 283 void copy_stack_traces(const StackTraceDataCopier &copier, 284 jvmtiStackTraces *traces); 285 286 void store_garbage_trace(const StackTraceData &trace); 287 288 void free_garbage(); 289 void free_storage(); 290 void allocate_storage(int max_gc_storage); 291 }; 292 293 StackTraceStorage* StackTraceStorage::internal_storage; 294 295 // Statics for Sampler 296 double HeapMonitoring::_log_table[1 << FastLogNumBits]; 297 bool HeapMonitoring::_enabled; 298 jint HeapMonitoring::_monitoring_rate; 299 300 // Cheap random number generator 301 uint64_t HeapMonitoring::_rnd; 302 303 StackTraceStorage::StackTraceStorage() : 304 _allocated_traces(NULL), 305 _recent_garbage_traces(NULL), 306 _frequent_garbage_traces(NULL), 307 _max_gc_storage(0), 308 _initialized(false) { 309 } 310 311 void StackTraceStorage::free_garbage() { 312 StackTraceData **recent_garbage = NULL; 313 uint32_t recent_size = 0; 314 315 StackTraceData **frequent_garbage = NULL; 316 uint32_t frequent_size = 0; 317 318 if (_recent_garbage_traces != NULL) { 319 recent_garbage = _recent_garbage_traces->get_traces(); 320 recent_size = _recent_garbage_traces->size(); 321 } 322 323 if (_frequent_garbage_traces != NULL) { 324 frequent_garbage = _frequent_garbage_traces->get_traces(); 325 frequent_size = _frequent_garbage_traces->size(); 326 } 327 328 // Simple solution since this happens at exit. 329 // Go through the recent and remove any that only are referenced there. 330 for (uint32_t i = 0; i < recent_size; i++) { 331 StackTraceData *trace = recent_garbage[i]; 332 if (trace != NULL) { 333 trace->references--; 334 335 if (trace->references == 0) { 336 StackTraceData::free_data(trace); 337 } 338 } 339 } 340 341 // Then go through the frequent and remove those that are now only there. 342 for (uint32_t i = 0; i < frequent_size; i++) { 343 StackTraceData *trace = frequent_garbage[i]; 344 if (trace != NULL) { 345 trace->references--; 346 347 if (trace->references == 0) { 348 StackTraceData::free_data(trace); 349 } 350 } 351 } 352 } 353 354 void StackTraceStorage::free_storage() { 355 delete _allocated_traces; 356 357 free_garbage(); 358 delete _recent_garbage_traces; 359 delete _frequent_garbage_traces; 360 _initialized = false; 361 } 362 363 StackTraceStorage::~StackTraceStorage() { 364 free_storage(); 365 } 366 367 void StackTraceStorage::allocate_storage(int max_gc_storage) { 368 // In case multiple threads got locked and then 1 by 1 got through. 369 if (_initialized) { 370 return; 371 } 372 373 _allocated_traces = new (ResourceObj::C_HEAP, mtInternal) 374 GrowableArray<StackTraceData>(128, true); 375 376 _recent_garbage_traces = new MostRecentGarbageTraces(max_gc_storage); 377 _frequent_garbage_traces = new FrequentGarbageTraces(max_gc_storage); 378 379 _max_gc_storage = max_gc_storage; 380 _initialized = true; 381 } 382 383 void StackTraceStorage::add_trace(jvmtiStackTrace *trace, oop o) { 384 MutexLocker mu(HeapMonitorStorage_lock); 385 StackTraceData new_data(trace, o); 386 _stats.sample_count++; 387 _stats.stack_depth_accumulation += trace->frame_count; 388 _allocated_traces->append(new_data); 389 } 390 391 void StackTraceStorage::weak_oops_do(BoolObjectClosure *is_alive, 392 OopClosure *f) { 393 MutexLocker mu(HeapMonitorStorage_lock); 394 size_t count = 0; 395 if (initialized()) { 396 int len = _allocated_traces->length(); 397 398 // Compact the oop traces. Moves the live oops to the beginning of the 399 // growable array, potentially overwriting the dead ones. 400 int curr_pos = 0; 401 for (int i = 0; i < len; i++) { 402 StackTraceData &trace = _allocated_traces->at(i); 403 oop value = trace.obj; 404 if (Universe::heap()->is_in_reserved(value) 405 && is_alive->do_object_b(value)) { 406 // Update the oop to point to the new object if it is still alive. 407 f->do_oop(&(trace.obj)); 408 409 // Copy the old trace, if it is still live. 410 _allocated_traces->at_put(curr_pos++, trace); 411 412 count++; 413 } else { 414 // If the old trace is no longer live, add it to the list of 415 // recently collected garbage. 416 store_garbage_trace(trace); 417 } 418 } 419 420 // Zero out remaining array elements. Even though the call to trunc_to 421 // below truncates these values, zeroing them out is good practice. 422 StackTraceData zero_trace; 423 for (int i = curr_pos; i < len; i++) { 424 _allocated_traces->at_put(i, zero_trace); 425 } 426 427 // Set the array's length to the number of live elements. 428 _allocated_traces->trunc_to(curr_pos); 429 } 430 431 log_develop_trace(gc, ref)("Clearing HeapMonitoring weak reference (" INT64_FORMAT ")", count); 432 } 433 434 bool StackTraceStorage::deep_copy(jvmtiStackTrace *to, 435 const StackTraceData *from) { 436 const jvmtiStackTrace *src = from->trace; 437 *to = *src; 438 439 to->frames = 440 NEW_C_HEAP_ARRAY(jvmtiFrameInfo, src->frame_count, mtInternal); 441 442 if (to->frames == NULL) { 443 return false; 444 } 445 446 memcpy(to->frames, 447 src->frames, 448 sizeof(jvmtiFrameInfo) * src->frame_count); 449 return true; 450 } 451 452 // Called by the outside world; returns a copy of the stack traces 453 // (because we could be replacing them as the user handles them). 454 // The array is secretly null-terminated (to make it easier to reclaim). 455 void StackTraceStorage::get_all_stack_traces(jvmtiStackTraces *traces) { 456 LiveStackTraceDataCopier copier(_allocated_traces); 457 copy_stack_traces(copier, traces); 458 } 459 460 // See comment on get_all_stack_traces 461 void StackTraceStorage::get_garbage_stack_traces(jvmtiStackTraces *traces) { 462 GarbageStackTraceDataCopier copier(_recent_garbage_traces->get_traces(), 463 _recent_garbage_traces->size()); 464 copy_stack_traces(copier, traces); 465 } 466 467 // See comment on get_all_stack_traces 468 void StackTraceStorage::get_frequent_garbage_stack_traces( 469 jvmtiStackTraces *traces) { 470 GarbageStackTraceDataCopier copier(_frequent_garbage_traces->get_traces(), 471 _frequent_garbage_traces->size()); 472 copy_stack_traces(copier, traces); 473 } 474 475 476 void StackTraceStorage::copy_stack_traces(const StackTraceDataCopier &copier, 477 jvmtiStackTraces *traces) { 478 MutexLocker mu(HeapMonitorStorage_lock); 479 int len = copier.size(); 480 481 // Create a new array to store the StackTraceData objects. 482 // + 1 for a NULL at the end. 483 jvmtiStackTrace *t = 484 NEW_C_HEAP_ARRAY(jvmtiStackTrace, len + 1, mtInternal); 485 if (t == NULL) { 486 traces->stack_traces = NULL; 487 traces->trace_count = 0; 488 return; 489 } 490 // +1 to have a NULL at the end of the array. 491 memset(t, 0, (len + 1) * sizeof(*t)); 492 493 // Copy the StackTraceData objects into the new array. 494 int trace_count = 0; 495 for (int i = 0; i < len; i++) { 496 const StackTraceData *stack_trace = copier.get(i); 497 if (stack_trace != NULL) { 498 jvmtiStackTrace *to = &t[trace_count]; 499 if (!deep_copy(to, stack_trace)) { 500 continue; 501 } 502 trace_count++; 503 } 504 } 505 506 traces->stack_traces = t; 507 traces->trace_count = trace_count; 508 } 509 510 void StackTraceStorage::store_garbage_trace(const StackTraceData &trace) { 511 StackTraceData *new_trace = new StackTraceData(); 512 *new_trace = trace; 513 514 bool accepted = _recent_garbage_traces->store_trace(new_trace); 515 516 // Accepted is on the right of the boolean to force the store_trace to happen. 517 accepted = _frequent_garbage_traces->store_trace(new_trace) || accepted; 518 519 if (!accepted) { 520 // No one wanted to use it. 521 delete new_trace; 522 } 523 524 _stats.garbage_collected_samples++; 525 } 526 527 void HeapMonitoring::get_live_traces(jvmtiStackTraces *traces) { 528 StackTraceStorage::storage()->get_all_stack_traces(traces); 529 } 530 531 void HeapMonitoring::get_sampling_statistics(jvmtiHeapSamplingStats *stats) { 532 const jvmtiHeapSamplingStats& internal_stats = 533 StackTraceStorage::storage()->get_heap_sampling_stats(); 534 *stats = internal_stats; 535 } 536 537 void HeapMonitoring::get_frequent_garbage_traces(jvmtiStackTraces *traces) { 538 StackTraceStorage::storage()->get_frequent_garbage_stack_traces(traces); 539 } 540 541 void HeapMonitoring::get_garbage_traces(jvmtiStackTraces *traces) { 542 StackTraceStorage::storage()->get_garbage_stack_traces(traces); 543 } 544 545 void HeapMonitoring::release_traces(jvmtiStackTraces *traces) { 546 jint trace_count = traces->trace_count; 547 jvmtiStackTrace *stack_traces = traces->stack_traces; 548 549 for (jint i = 0; i < trace_count; i++) { 550 jvmtiStackTrace *current_trace = stack_traces + i; 551 FREE_C_HEAP_ARRAY(jvmtiFrameInfo, current_trace->frames); 552 } 553 554 FREE_C_HEAP_ARRAY(jvmtiStackTrace, traces->stack_traces); 555 traces->trace_count = 0; 556 traces->stack_traces = NULL; 557 } 558 559 // Invoked by the GC to clean up old stack traces and remove old arrays 560 // of instrumentation that are still lying around. 561 void HeapMonitoring::weak_oops_do(BoolObjectClosure* is_alive, OopClosure *f) { 562 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 563 StackTraceStorage::storage()->weak_oops_do(is_alive, f); 564 } 565 566 void HeapMonitoring::initialize_profiling(jint monitoring_rate, 567 jint max_gc_storage) { 568 MutexLocker mu(HeapMonitor_lock); 569 // Ignore if already enabled. 570 if (_enabled) { 571 return; 572 } 573 574 _monitoring_rate = monitoring_rate; 575 576 // Populate the lookup table for fast_log2. 577 // This approximates the log2 curve with a step function. 578 // Steps have height equal to log2 of the mid-point of the step. 579 for (int i = 0; i < (1 << FastLogNumBits); i++) { 580 double half_way = static_cast<double>(i + 0.5); 581 _log_table[i] = (log(1.0 + half_way / (1 << FastLogNumBits)) / log(2.0)); 582 } 583 584 JavaThread *t = static_cast<JavaThread *>(Thread::current()); 585 _rnd = static_cast<uint32_t>(reinterpret_cast<uintptr_t>(t)); 586 if (_rnd == 0) { 587 _rnd = 1; 588 } 589 590 StackTraceStorage::storage()->initialize(max_gc_storage); 591 _enabled = true; 592 } 593 594 void HeapMonitoring::stop_profiling() { 595 MutexLocker mu(HeapMonitor_lock); 596 _enabled = false; 597 } 598 599 // Generates a geometric variable with the specified mean (512K by default). 600 // This is done by generating a random number between 0 and 1 and applying 601 // the inverse cumulative distribution function for an exponential. 602 // Specifically: Let m be the inverse of the sample rate, then 603 // the probability distribution function is m*exp(-mx) so the CDF is 604 // p = 1 - exp(-mx), so 605 // q = 1 - p = exp(-mx) 606 // log_e(q) = -mx 607 // -log_e(q)/m = x 608 // log_2(q) * (-log_e(2) * 1/m) = x 609 // In the code, q is actually in the range 1 to 2**26, hence the -26 below 610 void HeapMonitoring::pick_next_sample(size_t *ptr) { 611 _rnd = next_random(_rnd); 612 // Take the top 26 bits as the random number 613 // (This plus a 1<<58 sampling bound gives a max possible step of 614 // 5194297183973780480 bytes. In this case, 615 // for sample_parameter = 1<<19, max possible step is 616 // 9448372 bytes (24 bits). 617 const uint64_t PrngModPower = 48; // Number of bits in prng 618 // The uint32_t cast is to prevent a (hard-to-reproduce) NAN 619 // under piii debug for some binaries. 620 double q = static_cast<uint32_t>(_rnd >> (PrngModPower - 26)) + 1.0; 621 // Put the computed p-value through the CDF of a geometric. 622 // For faster performance (save ~1/20th exec time), replace 623 // min(0.0, FastLog2(q) - 26) by (Fastlog2(q) - 26.000705) 624 // The value 26.000705 is used rather than 26 to compensate 625 // for inaccuracies in FastLog2 which otherwise result in a 626 // negative answer. 627 double log_val = (fast_log2(q) - 26); 628 size_t rate = static_cast<size_t>( 629 (0.0 < log_val ? 0.0 : log_val) * (-log(2.0) * (_monitoring_rate)) + 1); 630 *ptr = rate; 631 632 StackTraceStorage::storage()->accumulate_sample_rate(rate); 633 } 634 635 void HeapMonitoring::object_alloc_do_sample(Thread *t, oopDesc *o, intx byte_size) { 636 #if defined(X86) || defined(PPC) 637 JavaThread *thread = static_cast<JavaThread *>(t); 638 if (StackTraceStorage::storage()->initialized()) { 639 assert(t->is_Java_thread(), "non-Java thread passed to do_sample"); 640 JavaThread *thread = static_cast<JavaThread *>(t); 641 642 jvmtiStackTrace *trace = NEW_C_HEAP_OBJ(jvmtiStackTrace, mtInternal); 643 if (trace == NULL) { 644 return; 645 } 646 647 jvmtiFrameInfo *frames = 648 NEW_C_HEAP_ARRAY(jvmtiFrameInfo, MaxStackDepth, mtInternal); 649 650 if (frames == NULL) { 651 FREE_C_HEAP_OBJ(trace); 652 return; 653 } 654 655 trace->frames = frames; 656 trace->thread_id = SharedRuntime::get_java_tid(thread); 657 trace->size = byte_size; 658 trace->frame_count = 0; 659 660 if (thread->has_last_Java_frame()) { // just to be safe 661 vframeStream vfst(thread, true); 662 int count = 0; 663 while (!vfst.at_end() && count < MaxStackDepth) { 664 Method* m = vfst.method(); 665 frames[count].location = vfst.bci(); 666 frames[count].method = m->jmethod_id(); 667 count++; 668 669 vfst.next(); 670 } 671 trace->frame_count = count; 672 } 673 674 if (trace->frame_count> 0) { 675 // Success! 676 StackTraceStorage::storage()->add_trace(trace, o); 677 return; 678 } 679 680 // Failure! 681 FREE_C_HEAP_ARRAY(jvmtiFrameInfo, trace->frames); 682 FREE_C_HEAP_OBJ(trace); 683 } 684 #else 685 Unimplemented(); 686 #endif 687 }