jdk

Форк
0
/
heapDumper.cpp 
2843 строки · 96.0 Кб
1
/*
2
 * Copyright (c) 2005, 2024, Oracle and/or its affiliates. All rights reserved.
3
 * Copyright (c) 2023, Alibaba Group Holding Limited. All rights reserved.
4
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5
 *
6
 * This code is free software; you can redistribute it and/or modify it
7
 * under the terms of the GNU General Public License version 2 only, as
8
 * published by the Free Software Foundation.
9
 *
10
 * This code is distributed in the hope that it will be useful, but WITHOUT
11
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
13
 * version 2 for more details (a copy is included in the LICENSE file that
14
 * accompanied this code).
15
 *
16
 * You should have received a copy of the GNU General Public License version
17
 * 2 along with this work; if not, write to the Free Software Foundation,
18
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19
 *
20
 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21
 * or visit www.oracle.com if you need additional information or have any
22
 * questions.
23
 *
24
 */
25

26
#include "precompiled.hpp"
27
#include "classfile/classLoaderData.inline.hpp"
28
#include "classfile/classLoaderDataGraph.hpp"
29
#include "classfile/javaClasses.inline.hpp"
30
#include "classfile/symbolTable.hpp"
31
#include "classfile/vmClasses.hpp"
32
#include "classfile/vmSymbols.hpp"
33
#include "gc/shared/gcLocker.hpp"
34
#include "gc/shared/gcVMOperations.hpp"
35
#include "gc/shared/workerThread.hpp"
36
#include "jfr/jfrEvents.hpp"
37
#include "jvm.h"
38
#include "memory/allocation.inline.hpp"
39
#include "memory/resourceArea.hpp"
40
#include "memory/universe.hpp"
41
#include "oops/fieldStreams.inline.hpp"
42
#include "oops/klass.inline.hpp"
43
#include "oops/objArrayKlass.hpp"
44
#include "oops/objArrayOop.inline.hpp"
45
#include "oops/oop.inline.hpp"
46
#include "oops/typeArrayOop.inline.hpp"
47
#include "runtime/continuationWrapper.inline.hpp"
48
#include "runtime/frame.inline.hpp"
49
#include "runtime/handles.inline.hpp"
50
#include "runtime/javaCalls.hpp"
51
#include "runtime/javaThread.inline.hpp"
52
#include "runtime/jniHandles.hpp"
53
#include "runtime/os.hpp"
54
#include "runtime/threads.hpp"
55
#include "runtime/threadSMR.hpp"
56
#include "runtime/vframe.hpp"
57
#include "runtime/vmOperations.hpp"
58
#include "runtime/vmThread.hpp"
59
#include "runtime/timerTrace.hpp"
60
#include "services/heapDumper.hpp"
61
#include "services/heapDumperCompression.hpp"
62
#include "services/threadService.hpp"
63
#include "utilities/checkedCast.hpp"
64
#include "utilities/macros.hpp"
65
#include "utilities/ostream.hpp"
66
#ifdef LINUX
67
#include "os_linux.hpp"
68
#endif
69

70
/*
71
 * HPROF binary format - description copied from:
72
 *   src/share/demo/jvmti/hprof/hprof_io.c
73
 *
74
 *
75
 *  header    "JAVA PROFILE 1.0.2" (0-terminated)
76
 *
77
 *  u4        size of identifiers. Identifiers are used to represent
78
 *            UTF8 strings, objects, stack traces, etc. They usually
79
 *            have the same size as host pointers.
80
 * u4         high word
81
 * u4         low word    number of milliseconds since 0:00 GMT, 1/1/70
82
 * [record]*  a sequence of records.
83
 *
84
 *
85
 * Record format:
86
 *
87
 * u1         a TAG denoting the type of the record
88
 * u4         number of *microseconds* since the time stamp in the
89
 *            header. (wraps around in a little more than an hour)
90
 * u4         number of bytes *remaining* in the record. Note that
91
 *            this number excludes the tag and the length field itself.
92
 * [u1]*      BODY of the record (a sequence of bytes)
93
 *
94
 *
95
 * The following TAGs are supported:
96
 *
97
 * TAG           BODY       notes
98
 *----------------------------------------------------------
99
 * HPROF_UTF8               a UTF8-encoded name
100
 *
101
 *               id         name ID
102
 *               [u1]*      UTF8 characters (no trailing zero)
103
 *
104
 * HPROF_LOAD_CLASS         a newly loaded class
105
 *
106
 *                u4        class serial number (> 0)
107
 *                id        class object ID
108
 *                u4        stack trace serial number
109
 *                id        class name ID
110
 *
111
 * HPROF_UNLOAD_CLASS       an unloading class
112
 *
113
 *                u4        class serial_number
114
 *
115
 * HPROF_FRAME              a Java stack frame
116
 *
117
 *                id        stack frame ID
118
 *                id        method name ID
119
 *                id        method signature ID
120
 *                id        source file name ID
121
 *                u4        class serial number
122
 *                i4        line number. >0: normal
123
 *                                       -1: unknown
124
 *                                       -2: compiled method
125
 *                                       -3: native method
126
 *
127
 * HPROF_TRACE              a Java stack trace
128
 *
129
 *               u4         stack trace serial number
130
 *               u4         thread serial number
131
 *               u4         number of frames
132
 *               [id]*      stack frame IDs
133
 *
134
 *
135
 * HPROF_ALLOC_SITES        a set of heap allocation sites, obtained after GC
136
 *
137
 *               u2         flags 0x0001: incremental vs. complete
138
 *                                0x0002: sorted by allocation vs. live
139
 *                                0x0004: whether to force a GC
140
 *               u4         cutoff ratio
141
 *               u4         total live bytes
142
 *               u4         total live instances
143
 *               u8         total bytes allocated
144
 *               u8         total instances allocated
145
 *               u4         number of sites that follow
146
 *               [u1        is_array: 0:  normal object
147
 *                                    2:  object array
148
 *                                    4:  boolean array
149
 *                                    5:  char array
150
 *                                    6:  float array
151
 *                                    7:  double array
152
 *                                    8:  byte array
153
 *                                    9:  short array
154
 *                                    10: int array
155
 *                                    11: long array
156
 *                u4        class serial number (may be zero during startup)
157
 *                u4        stack trace serial number
158
 *                u4        number of bytes alive
159
 *                u4        number of instances alive
160
 *                u4        number of bytes allocated
161
 *                u4]*      number of instance allocated
162
 *
163
 * HPROF_START_THREAD       a newly started thread.
164
 *
165
 *               u4         thread serial number (> 0)
166
 *               id         thread object ID
167
 *               u4         stack trace serial number
168
 *               id         thread name ID
169
 *               id         thread group name ID
170
 *               id         thread group parent name ID
171
 *
172
 * HPROF_END_THREAD         a terminating thread.
173
 *
174
 *               u4         thread serial number
175
 *
176
 * HPROF_HEAP_SUMMARY       heap summary
177
 *
178
 *               u4         total live bytes
179
 *               u4         total live instances
180
 *               u8         total bytes allocated
181
 *               u8         total instances allocated
182
 *
183
 * HPROF_HEAP_DUMP          denote a heap dump
184
 *
185
 *               [heap dump sub-records]*
186
 *
187
 *                          There are four kinds of heap dump sub-records:
188
 *
189
 *               u1         sub-record type
190
 *
191
 *               HPROF_GC_ROOT_UNKNOWN         unknown root
192
 *
193
 *                          id         object ID
194
 *
195
 *               HPROF_GC_ROOT_THREAD_OBJ      thread object
196
 *
197
 *                          id         thread object ID  (may be 0 for a
198
 *                                     thread newly attached through JNI)
199
 *                          u4         thread sequence number
200
 *                          u4         stack trace sequence number
201
 *
202
 *               HPROF_GC_ROOT_JNI_GLOBAL      JNI global ref root
203
 *
204
 *                          id         object ID
205
 *                          id         JNI global ref ID
206
 *
207
 *               HPROF_GC_ROOT_JNI_LOCAL       JNI local ref
208
 *
209
 *                          id         object ID
210
 *                          u4         thread serial number
211
 *                          u4         frame # in stack trace (-1 for empty)
212
 *
213
 *               HPROF_GC_ROOT_JAVA_FRAME      Java stack frame
214
 *
215
 *                          id         object ID
216
 *                          u4         thread serial number
217
 *                          u4         frame # in stack trace (-1 for empty)
218
 *
219
 *               HPROF_GC_ROOT_NATIVE_STACK    Native stack
220
 *
221
 *                          id         object ID
222
 *                          u4         thread serial number
223
 *
224
 *               HPROF_GC_ROOT_STICKY_CLASS    System class
225
 *
226
 *                          id         object ID
227
 *
228
 *               HPROF_GC_ROOT_THREAD_BLOCK    Reference from thread block
229
 *
230
 *                          id         object ID
231
 *                          u4         thread serial number
232
 *
233
 *               HPROF_GC_ROOT_MONITOR_USED    Busy monitor
234
 *
235
 *                          id         object ID
236
 *
237
 *               HPROF_GC_CLASS_DUMP           dump of a class object
238
 *
239
 *                          id         class object ID
240
 *                          u4         stack trace serial number
241
 *                          id         super class object ID
242
 *                          id         class loader object ID
243
 *                          id         signers object ID
244
 *                          id         protection domain object ID
245
 *                          id         reserved
246
 *                          id         reserved
247
 *
248
 *                          u4         instance size (in bytes)
249
 *
250
 *                          u2         size of constant pool
251
 *                          [u2,       constant pool index,
252
 *                           ty,       type
253
 *                                     2:  object
254
 *                                     4:  boolean
255
 *                                     5:  char
256
 *                                     6:  float
257
 *                                     7:  double
258
 *                                     8:  byte
259
 *                                     9:  short
260
 *                                     10: int
261
 *                                     11: long
262
 *                           vl]*      and value
263
 *
264
 *                          u2         number of static fields
265
 *                          [id,       static field name,
266
 *                           ty,       type,
267
 *                           vl]*      and value
268
 *
269
 *                          u2         number of inst. fields (not inc. super)
270
 *                          [id,       instance field name,
271
 *                           ty]*      type
272
 *
273
 *               HPROF_GC_INSTANCE_DUMP        dump of a normal object
274
 *
275
 *                          id         object ID
276
 *                          u4         stack trace serial number
277
 *                          id         class object ID
278
 *                          u4         number of bytes that follow
279
 *                          [vl]*      instance field values (class, followed
280
 *                                     by super, super's super ...)
281
 *
282
 *               HPROF_GC_OBJ_ARRAY_DUMP       dump of an object array
283
 *
284
 *                          id         array object ID
285
 *                          u4         stack trace serial number
286
 *                          u4         number of elements
287
 *                          id         array class ID
288
 *                          [id]*      elements
289
 *
290
 *               HPROF_GC_PRIM_ARRAY_DUMP      dump of a primitive array
291
 *
292
 *                          id         array object ID
293
 *                          u4         stack trace serial number
294
 *                          u4         number of elements
295
 *                          u1         element type
296
 *                                     4:  boolean array
297
 *                                     5:  char array
298
 *                                     6:  float array
299
 *                                     7:  double array
300
 *                                     8:  byte array
301
 *                                     9:  short array
302
 *                                     10: int array
303
 *                                     11: long array
304
 *                          [u1]*      elements
305
 *
306
 * HPROF_CPU_SAMPLES        a set of sample traces of running threads
307
 *
308
 *                u4        total number of samples
309
 *                u4        # of traces
310
 *               [u4        # of samples
311
 *                u4]*      stack trace serial number
312
 *
313
 * HPROF_CONTROL_SETTINGS   the settings of on/off switches
314
 *
315
 *                u4        0x00000001: alloc traces on/off
316
 *                          0x00000002: cpu sampling on/off
317
 *                u2        stack trace depth
318
 *
319
 *
320
 * When the header is "JAVA PROFILE 1.0.2" a heap dump can optionally
321
 * be generated as a sequence of heap dump segments. This sequence is
322
 * terminated by an end record. The additional tags allowed by format
323
 * "JAVA PROFILE 1.0.2" are:
324
 *
325
 * HPROF_HEAP_DUMP_SEGMENT  denote a heap dump segment
326
 *
327
 *               [heap dump sub-records]*
328
 *               The same sub-record types allowed by HPROF_HEAP_DUMP
329
 *
330
 * HPROF_HEAP_DUMP_END      denotes the end of a heap dump
331
 *
332
 */
333

334

335
// HPROF tags
336

337
enum hprofTag : u1 {
338
  // top-level records
339
  HPROF_UTF8                    = 0x01,
340
  HPROF_LOAD_CLASS              = 0x02,
341
  HPROF_UNLOAD_CLASS            = 0x03,
342
  HPROF_FRAME                   = 0x04,
343
  HPROF_TRACE                   = 0x05,
344
  HPROF_ALLOC_SITES             = 0x06,
345
  HPROF_HEAP_SUMMARY            = 0x07,
346
  HPROF_START_THREAD            = 0x0A,
347
  HPROF_END_THREAD              = 0x0B,
348
  HPROF_HEAP_DUMP               = 0x0C,
349
  HPROF_CPU_SAMPLES             = 0x0D,
350
  HPROF_CONTROL_SETTINGS        = 0x0E,
351

352
  // 1.0.2 record types
353
  HPROF_HEAP_DUMP_SEGMENT       = 0x1C,
354
  HPROF_HEAP_DUMP_END           = 0x2C,
355

356
  // field types
357
  HPROF_ARRAY_OBJECT            = 0x01,
358
  HPROF_NORMAL_OBJECT           = 0x02,
359
  HPROF_BOOLEAN                 = 0x04,
360
  HPROF_CHAR                    = 0x05,
361
  HPROF_FLOAT                   = 0x06,
362
  HPROF_DOUBLE                  = 0x07,
363
  HPROF_BYTE                    = 0x08,
364
  HPROF_SHORT                   = 0x09,
365
  HPROF_INT                     = 0x0A,
366
  HPROF_LONG                    = 0x0B,
367

368
  // data-dump sub-records
369
  HPROF_GC_ROOT_UNKNOWN         = 0xFF,
370
  HPROF_GC_ROOT_JNI_GLOBAL      = 0x01,
371
  HPROF_GC_ROOT_JNI_LOCAL       = 0x02,
372
  HPROF_GC_ROOT_JAVA_FRAME      = 0x03,
373
  HPROF_GC_ROOT_NATIVE_STACK    = 0x04,
374
  HPROF_GC_ROOT_STICKY_CLASS    = 0x05,
375
  HPROF_GC_ROOT_THREAD_BLOCK    = 0x06,
376
  HPROF_GC_ROOT_MONITOR_USED    = 0x07,
377
  HPROF_GC_ROOT_THREAD_OBJ      = 0x08,
378
  HPROF_GC_CLASS_DUMP           = 0x20,
379
  HPROF_GC_INSTANCE_DUMP        = 0x21,
380
  HPROF_GC_OBJ_ARRAY_DUMP       = 0x22,
381
  HPROF_GC_PRIM_ARRAY_DUMP      = 0x23
382
};
383

384
// Default stack trace ID (used for dummy HPROF_TRACE record)
385
enum {
386
  STACK_TRACE_ID = 1,
387
  INITIAL_CLASS_COUNT = 200
388
};
389

390
// Supports I/O operations for a dump
391
// Base class for dump and parallel dump
392
class AbstractDumpWriter : public CHeapObj<mtInternal> {
393
 protected:
394
  enum {
395
    io_buffer_max_size = 1*M,
396
    dump_segment_header_size = 9
397
  };
398

399
  char* _buffer;    // internal buffer
400
  size_t _size;
401
  size_t _pos;
402

403
  bool _in_dump_segment; // Are we currently in a dump segment?
404
  bool _is_huge_sub_record; // Are we writing a sub-record larger than the buffer size?
405
  DEBUG_ONLY(size_t _sub_record_left;) // The bytes not written for the current sub-record.
406
  DEBUG_ONLY(bool _sub_record_ended;) // True if we have called the end_sub_record().
407

408
  char* buffer() const                          { return _buffer; }
409
  size_t buffer_size() const                    { return _size; }
410
  void set_position(size_t pos)                 { _pos = pos; }
411

412
  // Can be called if we have enough room in the buffer.
413
  void write_fast(const void* s, size_t len);
414

415
  // Returns true if we have enough room in the buffer for 'len' bytes.
416
  bool can_write_fast(size_t len);
417

418
  void write_address(address a);
419

420
 public:
421
  AbstractDumpWriter() :
422
    _buffer(nullptr),
423
    _size(io_buffer_max_size),
424
    _pos(0),
425
    _in_dump_segment(false) { }
426

427
  // Total number of bytes written to the disk
428
  virtual julong bytes_written() const = 0;
429
  // Return non-null if error occurred
430
  virtual char const* error() const = 0;
431

432
  size_t position() const                       { return _pos; }
433
  // writer functions
434
  virtual void write_raw(const void* s, size_t len);
435
  void write_u1(u1 x);
436
  void write_u2(u2 x);
437
  void write_u4(u4 x);
438
  void write_u8(u8 x);
439
  void write_objectID(oop o);
440
  void write_rootID(oop* p);
441
  void write_symbolID(Symbol* o);
442
  void write_classID(Klass* k);
443
  void write_id(u4 x);
444

445
  // Start a new sub-record. Starts a new heap dump segment if needed.
446
  void start_sub_record(u1 tag, u4 len);
447
  // Ends the current sub-record.
448
  void end_sub_record();
449
  // Finishes the current dump segment if not already finished.
450
  void finish_dump_segment();
451
  // Flush internal buffer to persistent storage
452
  virtual void flush() = 0;
453
};
454

455
void AbstractDumpWriter::write_fast(const void* s, size_t len) {
456
  assert(!_in_dump_segment || (_sub_record_left >= len), "sub-record too large");
457
  assert(buffer_size() - position() >= len, "Must fit");
458
  debug_only(_sub_record_left -= len);
459
  memcpy(buffer() + position(), s, len);
460
  set_position(position() + len);
461
}
462

463
bool AbstractDumpWriter::can_write_fast(size_t len) {
464
  return buffer_size() - position() >= len;
465
}
466

467
// write raw bytes
468
void AbstractDumpWriter::write_raw(const void* s, size_t len) {
469
  assert(!_in_dump_segment || (_sub_record_left >= len), "sub-record too large");
470
  debug_only(_sub_record_left -= len);
471

472
  // flush buffer to make room.
473
  while (len > buffer_size() - position()) {
474
    assert(!_in_dump_segment || _is_huge_sub_record,
475
           "Cannot overflow in non-huge sub-record.");
476
    size_t to_write = buffer_size() - position();
477
    memcpy(buffer() + position(), s, to_write);
478
    s = (void*) ((char*) s + to_write);
479
    len -= to_write;
480
    set_position(position() + to_write);
481
    flush();
482
  }
483

484
  memcpy(buffer() + position(), s, len);
485
  set_position(position() + len);
486
}
487

488
// Makes sure we inline the fast write into the write_u* functions. This is a big speedup.
489
#define WRITE_KNOWN_TYPE(p, len) do { if (can_write_fast((len))) write_fast((p), (len)); \
490
                                      else write_raw((p), (len)); } while (0)
491

492
void AbstractDumpWriter::write_u1(u1 x) {
493
  WRITE_KNOWN_TYPE(&x, 1);
494
}
495

496
void AbstractDumpWriter::write_u2(u2 x) {
497
  u2 v;
498
  Bytes::put_Java_u2((address)&v, x);
499
  WRITE_KNOWN_TYPE(&v, 2);
500
}
501

502
void AbstractDumpWriter::write_u4(u4 x) {
503
  u4 v;
504
  Bytes::put_Java_u4((address)&v, x);
505
  WRITE_KNOWN_TYPE(&v, 4);
506
}
507

508
void AbstractDumpWriter::write_u8(u8 x) {
509
  u8 v;
510
  Bytes::put_Java_u8((address)&v, x);
511
  WRITE_KNOWN_TYPE(&v, 8);
512
}
513

514
void AbstractDumpWriter::write_address(address a) {
515
#ifdef _LP64
516
  write_u8((u8)a);
517
#else
518
  write_u4((u4)a);
519
#endif
520
}
521

522
void AbstractDumpWriter::write_objectID(oop o) {
523
  write_address(cast_from_oop<address>(o));
524
}
525

526
void AbstractDumpWriter::write_rootID(oop* p) {
527
  write_address((address)p);
528
}
529

530
void AbstractDumpWriter::write_symbolID(Symbol* s) {
531
  write_address((address)((uintptr_t)s));
532
}
533

534
void AbstractDumpWriter::write_id(u4 x) {
535
#ifdef _LP64
536
  write_u8((u8) x);
537
#else
538
  write_u4(x);
539
#endif
540
}
541

542
// We use java mirror as the class ID
543
void AbstractDumpWriter::write_classID(Klass* k) {
544
  write_objectID(k->java_mirror());
545
}
546

547
void AbstractDumpWriter::finish_dump_segment() {
548
  if (_in_dump_segment) {
549
    assert(_sub_record_left == 0, "Last sub-record not written completely");
550
    assert(_sub_record_ended, "sub-record must have ended");
551

552
    // Fix up the dump segment length if we haven't written a huge sub-record last
553
    // (in which case the segment length was already set to the correct value initially).
554
    if (!_is_huge_sub_record) {
555
      assert(position() > dump_segment_header_size, "Dump segment should have some content");
556
      Bytes::put_Java_u4((address) (buffer() + 5),
557
                         (u4) (position() - dump_segment_header_size));
558
    } else {
559
      // Finish process huge sub record
560
      // Set _is_huge_sub_record to false so the parallel dump writer can flush data to file.
561
      _is_huge_sub_record = false;
562
    }
563

564
    _in_dump_segment = false;
565
    flush();
566
  }
567
}
568

569
void AbstractDumpWriter::start_sub_record(u1 tag, u4 len) {
570
  if (!_in_dump_segment) {
571
    if (position() > 0) {
572
      flush();
573
    }
574

575
    assert(position() == 0 && buffer_size() > dump_segment_header_size, "Must be at the start");
576

577
    write_u1(HPROF_HEAP_DUMP_SEGMENT);
578
    write_u4(0); // timestamp
579
    // Will be fixed up later if we add more sub-records.  If this is a huge sub-record,
580
    // this is already the correct length, since we don't add more sub-records.
581
    write_u4(len);
582
    assert(Bytes::get_Java_u4((address)(buffer() + 5)) == len, "Inconsistent size!");
583
    _in_dump_segment = true;
584
    _is_huge_sub_record = len > buffer_size() - dump_segment_header_size;
585
  } else if (_is_huge_sub_record || (len > buffer_size() - position())) {
586
    // This object will not fit in completely or the last sub-record was huge.
587
    // Finish the current segment and try again.
588
    finish_dump_segment();
589
    start_sub_record(tag, len);
590

591
    return;
592
  }
593

594
  debug_only(_sub_record_left = len);
595
  debug_only(_sub_record_ended = false);
596

597
  write_u1(tag);
598
}
599

600
void AbstractDumpWriter::end_sub_record() {
601
  assert(_in_dump_segment, "must be in dump segment");
602
  assert(_sub_record_left == 0, "sub-record not written completely");
603
  assert(!_sub_record_ended, "Must not have ended yet");
604
  debug_only(_sub_record_ended = true);
605
}
606

607
// Supports I/O operations for a dump
608

609
class DumpWriter : public AbstractDumpWriter {
610
private:
611
  FileWriter* _writer;
612
  AbstractCompressor* _compressor;
613
  size_t _bytes_written;
614
  char* _error;
615
  // Compression support
616
  char* _out_buffer;
617
  size_t _out_size;
618
  size_t _out_pos;
619
  char* _tmp_buffer;
620
  size_t _tmp_size;
621

622
private:
623
  void do_compress();
624

625
public:
626
  DumpWriter(const char* path, bool overwrite, AbstractCompressor* compressor);
627
  ~DumpWriter();
628
  julong bytes_written() const override        { return (julong) _bytes_written; }
629
  char const* error() const override           { return _error; }
630
  void set_error(const char* error)            { _error = (char*)error; }
631
  bool has_error() const                       { return _error != nullptr; }
632
  const char* get_file_path() const            { return _writer->get_file_path(); }
633
  AbstractCompressor* compressor()             { return _compressor; }
634
  bool is_overwrite() const                    { return _writer->is_overwrite(); }
635

636
  void flush() override;
637

638
private:
639
  // internals for DumpMerger
640
  friend class DumpMerger;
641
  void set_bytes_written(julong bytes_written) { _bytes_written = bytes_written; }
642
  int get_fd() const                           { return _writer->get_fd(); }
643
  void set_compressor(AbstractCompressor* p)   { _compressor = p; }
644
};
645

646
DumpWriter::DumpWriter(const char* path, bool overwrite, AbstractCompressor* compressor) :
647
  AbstractDumpWriter(),
648
  _writer(new (std::nothrow) FileWriter(path, overwrite)),
649
  _compressor(compressor),
650
  _bytes_written(0),
651
  _error(nullptr),
652
  _out_buffer(nullptr),
653
  _out_size(0),
654
  _out_pos(0),
655
  _tmp_buffer(nullptr),
656
  _tmp_size(0) {
657
  _error = (char*)_writer->open_writer();
658
  if (_error == nullptr) {
659
    _buffer = (char*)os::malloc(io_buffer_max_size, mtInternal);
660
    if (compressor != nullptr) {
661
      _error = (char*)_compressor->init(io_buffer_max_size, &_out_size, &_tmp_size);
662
      if (_error == nullptr) {
663
        if (_out_size > 0) {
664
          _out_buffer = (char*)os::malloc(_out_size, mtInternal);
665
        }
666
        if (_tmp_size > 0) {
667
          _tmp_buffer = (char*)os::malloc(_tmp_size, mtInternal);
668
        }
669
      }
670
    }
671
  }
672
  // initialize internal buffer
673
  _pos = 0;
674
  _size = io_buffer_max_size;
675
}
676

677
DumpWriter::~DumpWriter(){
678
  if (_buffer != nullptr) {
679
    os::free(_buffer);
680
  }
681
  if (_out_buffer != nullptr) {
682
    os::free(_out_buffer);
683
  }
684
  if (_tmp_buffer != nullptr) {
685
    os::free(_tmp_buffer);
686
  }
687
  if (_writer != nullptr) {
688
    delete _writer;
689
  }
690
  _bytes_written = -1;
691
}
692

693
// flush any buffered bytes to the file
694
void DumpWriter::flush() {
695
  if (_pos <= 0) {
696
    return;
697
  }
698
  if (has_error()) {
699
    _pos = 0;
700
    return;
701
  }
702
  char* result = nullptr;
703
  if (_compressor == nullptr) {
704
    result = (char*)_writer->write_buf(_buffer, _pos);
705
    _bytes_written += _pos;
706
  } else {
707
    do_compress();
708
    if (!has_error()) {
709
      result = (char*)_writer->write_buf(_out_buffer, _out_pos);
710
      _bytes_written += _out_pos;
711
    }
712
  }
713
  _pos = 0; // reset pos to make internal buffer available
714

715
  if (result != nullptr) {
716
    set_error(result);
717
  }
718
}
719

720
void DumpWriter::do_compress() {
721
  const char* msg = _compressor->compress(_buffer, _pos, _out_buffer, _out_size,
722
                                          _tmp_buffer, _tmp_size, &_out_pos);
723

724
  if (msg != nullptr) {
725
    set_error(msg);
726
  }
727
}
728

729
class DumperClassCacheTable;
730
class DumperClassCacheTableEntry;
731

732
// Support class with a collection of functions used when dumping the heap
733
class DumperSupport : AllStatic {
734
 public:
735

736
  // write a header of the given type
737
  static void write_header(AbstractDumpWriter* writer, hprofTag tag, u4 len);
738

739
  // returns hprof tag for the given type signature
740
  static hprofTag sig2tag(Symbol* sig);
741
  // returns hprof tag for the given basic type
742
  static hprofTag type2tag(BasicType type);
743
  // Returns the size of the data to write.
744
  static u4 sig2size(Symbol* sig);
745

746
  // returns the size of the instance of the given class
747
  static u4 instance_size(InstanceKlass* ik, DumperClassCacheTableEntry* class_cache_entry = nullptr);
748

749
  // dump a jfloat
750
  static void dump_float(AbstractDumpWriter* writer, jfloat f);
751
  // dump a jdouble
752
  static void dump_double(AbstractDumpWriter* writer, jdouble d);
753
  // dumps the raw value of the given field
754
  static void dump_field_value(AbstractDumpWriter* writer, char type, oop obj, int offset);
755
  // returns the size of the static fields; also counts the static fields
756
  static u4 get_static_fields_size(InstanceKlass* ik, u2& field_count);
757
  // dumps static fields of the given class
758
  static void dump_static_fields(AbstractDumpWriter* writer, Klass* k);
759
  // dump the raw values of the instance fields of the given object
760
  static void dump_instance_fields(AbstractDumpWriter* writer, oop o, DumperClassCacheTableEntry* class_cache_entry);
761
  // get the count of the instance fields for a given class
762
  static u2 get_instance_fields_count(InstanceKlass* ik);
763
  // dumps the definition of the instance fields for a given class
764
  static void dump_instance_field_descriptors(AbstractDumpWriter* writer, Klass* k);
765
  // creates HPROF_GC_INSTANCE_DUMP record for the given object
766
  static void dump_instance(AbstractDumpWriter* writer, oop o, DumperClassCacheTable* class_cache);
767
  // creates HPROF_GC_CLASS_DUMP record for the given instance class
768
  static void dump_instance_class(AbstractDumpWriter* writer, Klass* k);
769
  // creates HPROF_GC_CLASS_DUMP record for a given array class
770
  static void dump_array_class(AbstractDumpWriter* writer, Klass* k);
771

772
  // creates HPROF_GC_OBJ_ARRAY_DUMP record for the given object array
773
  static void dump_object_array(AbstractDumpWriter* writer, objArrayOop array);
774
  // creates HPROF_GC_PRIM_ARRAY_DUMP record for the given type array
775
  static void dump_prim_array(AbstractDumpWriter* writer, typeArrayOop array);
776
  // create HPROF_FRAME record for the given method and bci
777
  static void dump_stack_frame(AbstractDumpWriter* writer, int frame_serial_num, int class_serial_num, Method* m, int bci);
778

779
  // check if we need to truncate an array
780
  static int calculate_array_max_length(AbstractDumpWriter* writer, arrayOop array, short header_size);
781

782
  // fixes up the current dump record and writes HPROF_HEAP_DUMP_END record
783
  static void end_of_dump(AbstractDumpWriter* writer);
784

785
  static oop mask_dormant_archived_object(oop o, oop ref_obj) {
786
    if (o != nullptr && o->klass()->java_mirror_no_keepalive() == nullptr) {
787
      // Ignore this object since the corresponding java mirror is not loaded.
788
      // Might be a dormant archive object.
789
      report_dormant_archived_object(o, ref_obj);
790
      return nullptr;
791
    } else {
792
      return o;
793
    }
794
  }
795

796
  static void report_dormant_archived_object(oop o, oop ref_obj) {
797
    if (log_is_enabled(Trace, cds, heap)) {
798
      ResourceMark rm;
799
      if (ref_obj != nullptr) {
800
        log_trace(cds, heap)("skipped dormant archived object " INTPTR_FORMAT " (%s) referenced by " INTPTR_FORMAT " (%s)",
801
                  p2i(o), o->klass()->external_name(),
802
                  p2i(ref_obj), ref_obj->klass()->external_name());
803
      } else {
804
        log_trace(cds, heap)("skipped dormant archived object " INTPTR_FORMAT " (%s)",
805
                  p2i(o), o->klass()->external_name());
806
      }
807
    }
808
  }
809
};
810

811
// Hash table of klasses to the klass metadata. This should greatly improve the
812
// hash dumping performance. This hash table is supposed to be used by a single
813
// thread only.
814
//
815
class DumperClassCacheTableEntry : public CHeapObj<mtServiceability> {
816
  friend class DumperClassCacheTable;
817
private:
818
  GrowableArray<char> _sigs_start;
819
  GrowableArray<int> _offsets;
820
  u4 _instance_size;
821
  int _entries;
822

823
public:
824
  DumperClassCacheTableEntry() : _instance_size(0), _entries(0) {};
825

826
  int field_count()             { return _entries; }
827
  char sig_start(int field_idx) { return _sigs_start.at(field_idx); }
828
  int offset(int field_idx)     { return _offsets.at(field_idx); }
829
  u4 instance_size()            { return _instance_size; }
830
};
831

832
class DumperClassCacheTable {
833
private:
834
  // ResourceHashtable SIZE is specified at compile time so we
835
  // use 1031 which is the first prime after 1024.
836
  static constexpr size_t TABLE_SIZE = 1031;
837

838
  // Maintain the cache for N classes. This limits memory footprint
839
  // impact, regardless of how many classes we have in the dump.
840
  // This also improves look up performance by keeping the statically
841
  // sized table from overloading.
842
  static constexpr int CACHE_TOP = 256;
843

844
  typedef ResourceHashtable<InstanceKlass*, DumperClassCacheTableEntry*,
845
                            TABLE_SIZE, AnyObj::C_HEAP, mtServiceability> PtrTable;
846
  PtrTable* _ptrs;
847

848
  // Single-slot cache to handle the major case of objects of the same
849
  // class back-to-back, e.g. from T[].
850
  InstanceKlass* _last_ik;
851
  DumperClassCacheTableEntry* _last_entry;
852

853
  void unlink_all(PtrTable* table) {
854
    class CleanupEntry: StackObj {
855
    public:
856
      bool do_entry(InstanceKlass*& key, DumperClassCacheTableEntry*& entry) {
857
        delete entry;
858
        return true;
859
      }
860
    } cleanup;
861
    table->unlink(&cleanup);
862
  }
863

864
public:
865
  DumperClassCacheTableEntry* lookup_or_create(InstanceKlass* ik) {
866
    if (_last_ik == ik) {
867
      return _last_entry;
868
    }
869

870
    DumperClassCacheTableEntry* entry;
871
    DumperClassCacheTableEntry** from_cache = _ptrs->get(ik);
872
    if (from_cache == nullptr) {
873
      entry = new DumperClassCacheTableEntry();
874
      for (HierarchicalFieldStream<JavaFieldStream> fld(ik); !fld.done(); fld.next()) {
875
        if (!fld.access_flags().is_static()) {
876
          Symbol* sig = fld.signature();
877
          entry->_sigs_start.push(sig->char_at(0));
878
          entry->_offsets.push(fld.offset());
879
          entry->_entries++;
880
          entry->_instance_size += DumperSupport::sig2size(sig);
881
        }
882
      }
883

884
      if (_ptrs->number_of_entries() >= CACHE_TOP) {
885
        // We do not track the individual hit rates for table entries.
886
        // Purge the entire table, and let the cache catch up with new
887
        // distribution.
888
        unlink_all(_ptrs);
889
      }
890

891
      _ptrs->put(ik, entry);
892
    } else {
893
      entry = *from_cache;
894
    }
895

896
    // Remember for single-slot cache.
897
    _last_ik = ik;
898
    _last_entry = entry;
899

900
    return entry;
901
  }
902

903
  DumperClassCacheTable() : _ptrs(new (mtServiceability) PtrTable), _last_ik(nullptr), _last_entry(nullptr) {}
904

905
  ~DumperClassCacheTable() {
906
    unlink_all(_ptrs);
907
    delete _ptrs;
908
  }
909
};
910

911
// write a header of the given type
912
void DumperSupport:: write_header(AbstractDumpWriter* writer, hprofTag tag, u4 len) {
913
  writer->write_u1(tag);
914
  writer->write_u4(0);                  // current ticks
915
  writer->write_u4(len);
916
}
917

918
// returns hprof tag for the given type signature
919
hprofTag DumperSupport::sig2tag(Symbol* sig) {
920
  switch (sig->char_at(0)) {
921
    case JVM_SIGNATURE_CLASS    : return HPROF_NORMAL_OBJECT;
922
    case JVM_SIGNATURE_ARRAY    : return HPROF_NORMAL_OBJECT;
923
    case JVM_SIGNATURE_BYTE     : return HPROF_BYTE;
924
    case JVM_SIGNATURE_CHAR     : return HPROF_CHAR;
925
    case JVM_SIGNATURE_FLOAT    : return HPROF_FLOAT;
926
    case JVM_SIGNATURE_DOUBLE   : return HPROF_DOUBLE;
927
    case JVM_SIGNATURE_INT      : return HPROF_INT;
928
    case JVM_SIGNATURE_LONG     : return HPROF_LONG;
929
    case JVM_SIGNATURE_SHORT    : return HPROF_SHORT;
930
    case JVM_SIGNATURE_BOOLEAN  : return HPROF_BOOLEAN;
931
    default : ShouldNotReachHere(); /* to shut up compiler */ return HPROF_BYTE;
932
  }
933
}
934

935
hprofTag DumperSupport::type2tag(BasicType type) {
936
  switch (type) {
937
    case T_BYTE     : return HPROF_BYTE;
938
    case T_CHAR     : return HPROF_CHAR;
939
    case T_FLOAT    : return HPROF_FLOAT;
940
    case T_DOUBLE   : return HPROF_DOUBLE;
941
    case T_INT      : return HPROF_INT;
942
    case T_LONG     : return HPROF_LONG;
943
    case T_SHORT    : return HPROF_SHORT;
944
    case T_BOOLEAN  : return HPROF_BOOLEAN;
945
    default : ShouldNotReachHere(); /* to shut up compiler */ return HPROF_BYTE;
946
  }
947
}
948

949
u4 DumperSupport::sig2size(Symbol* sig) {
950
  switch (sig->char_at(0)) {
951
    case JVM_SIGNATURE_CLASS:
952
    case JVM_SIGNATURE_ARRAY: return sizeof(address);
953
    case JVM_SIGNATURE_BOOLEAN:
954
    case JVM_SIGNATURE_BYTE: return 1;
955
    case JVM_SIGNATURE_SHORT:
956
    case JVM_SIGNATURE_CHAR: return 2;
957
    case JVM_SIGNATURE_INT:
958
    case JVM_SIGNATURE_FLOAT: return 4;
959
    case JVM_SIGNATURE_LONG:
960
    case JVM_SIGNATURE_DOUBLE: return 8;
961
    default: ShouldNotReachHere(); /* to shut up compiler */ return 0;
962
  }
963
}
964

965
template<typename T, typename F> T bit_cast(F from) { // replace with the real thing when we can use c++20
966
  T to;
967
  static_assert(sizeof(to) == sizeof(from), "must be of the same size");
968
  memcpy(&to, &from, sizeof(to));
969
  return to;
970
}
971

972
// dump a jfloat
973
void DumperSupport::dump_float(AbstractDumpWriter* writer, jfloat f) {
974
  if (g_isnan(f)) {
975
    writer->write_u4(0x7fc00000); // collapsing NaNs
976
  } else {
977
    writer->write_u4(bit_cast<u4>(f));
978
  }
979
}
980

981
// dump a jdouble
982
void DumperSupport::dump_double(AbstractDumpWriter* writer, jdouble d) {
983
  if (g_isnan(d)) {
984
    writer->write_u8(0x7ff80000ull << 32); // collapsing NaNs
985
  } else {
986
    writer->write_u8(bit_cast<u8>(d));
987
  }
988
}
989

990
// dumps the raw value of the given field
991
void DumperSupport::dump_field_value(AbstractDumpWriter* writer, char type, oop obj, int offset) {
992
  switch (type) {
993
    case JVM_SIGNATURE_CLASS :
994
    case JVM_SIGNATURE_ARRAY : {
995
      oop o = obj->obj_field_access<ON_UNKNOWN_OOP_REF | AS_NO_KEEPALIVE>(offset);
996
      o = mask_dormant_archived_object(o, obj);
997
      assert(oopDesc::is_oop_or_null(o), "Expected an oop or nullptr at " PTR_FORMAT, p2i(o));
998
      writer->write_objectID(o);
999
      break;
1000
    }
1001
    case JVM_SIGNATURE_BYTE : {
1002
      jbyte b = obj->byte_field(offset);
1003
      writer->write_u1(b);
1004
      break;
1005
    }
1006
    case JVM_SIGNATURE_CHAR : {
1007
      jchar c = obj->char_field(offset);
1008
      writer->write_u2(c);
1009
      break;
1010
    }
1011
    case JVM_SIGNATURE_SHORT : {
1012
      jshort s = obj->short_field(offset);
1013
      writer->write_u2(s);
1014
      break;
1015
    }
1016
    case JVM_SIGNATURE_FLOAT : {
1017
      jfloat f = obj->float_field(offset);
1018
      dump_float(writer, f);
1019
      break;
1020
    }
1021
    case JVM_SIGNATURE_DOUBLE : {
1022
      jdouble d = obj->double_field(offset);
1023
      dump_double(writer, d);
1024
      break;
1025
    }
1026
    case JVM_SIGNATURE_INT : {
1027
      jint i = obj->int_field(offset);
1028
      writer->write_u4(i);
1029
      break;
1030
    }
1031
    case JVM_SIGNATURE_LONG : {
1032
      jlong l = obj->long_field(offset);
1033
      writer->write_u8(l);
1034
      break;
1035
    }
1036
    case JVM_SIGNATURE_BOOLEAN : {
1037
      jboolean b = obj->bool_field(offset);
1038
      writer->write_u1(b);
1039
      break;
1040
    }
1041
    default : {
1042
      ShouldNotReachHere();
1043
      break;
1044
    }
1045
  }
1046
}
1047

1048
// returns the size of the instance of the given class
1049
u4 DumperSupport::instance_size(InstanceKlass* ik, DumperClassCacheTableEntry* class_cache_entry) {
1050
  if (class_cache_entry != nullptr) {
1051
    return class_cache_entry->instance_size();
1052
  } else {
1053
    u4 size = 0;
1054
    for (HierarchicalFieldStream<JavaFieldStream> fld(ik); !fld.done(); fld.next()) {
1055
      if (!fld.access_flags().is_static()) {
1056
        size += sig2size(fld.signature());
1057
      }
1058
    }
1059
    return size;
1060
  }
1061
}
1062

1063
u4 DumperSupport::get_static_fields_size(InstanceKlass* ik, u2& field_count) {
1064
  field_count = 0;
1065
  u4 size = 0;
1066

1067
  for (JavaFieldStream fldc(ik); !fldc.done(); fldc.next()) {
1068
    if (fldc.access_flags().is_static()) {
1069
      field_count++;
1070
      size += sig2size(fldc.signature());
1071
    }
1072
  }
1073

1074
  // Add in resolved_references which is referenced by the cpCache
1075
  // The resolved_references is an array per InstanceKlass holding the
1076
  // strings and other oops resolved from the constant pool.
1077
  oop resolved_references = ik->constants()->resolved_references_or_null();
1078
  if (resolved_references != nullptr) {
1079
    field_count++;
1080
    size += sizeof(address);
1081

1082
    // Add in the resolved_references of the used previous versions of the class
1083
    // in the case of RedefineClasses
1084
    InstanceKlass* prev = ik->previous_versions();
1085
    while (prev != nullptr && prev->constants()->resolved_references_or_null() != nullptr) {
1086
      field_count++;
1087
      size += sizeof(address);
1088
      prev = prev->previous_versions();
1089
    }
1090
  }
1091

1092
  // Also provide a pointer to the init_lock if present, so there aren't unreferenced int[0]
1093
  // arrays.
1094
  oop init_lock = ik->init_lock();
1095
  if (init_lock != nullptr) {
1096
    field_count++;
1097
    size += sizeof(address);
1098
  }
1099

1100
  // We write the value itself plus a name and a one byte type tag per field.
1101
  return checked_cast<u4>(size + field_count * (sizeof(address) + 1));
1102
}
1103

1104
// dumps static fields of the given class
1105
void DumperSupport::dump_static_fields(AbstractDumpWriter* writer, Klass* k) {
1106
  InstanceKlass* ik = InstanceKlass::cast(k);
1107

1108
  // dump the field descriptors and raw values
1109
  for (JavaFieldStream fld(ik); !fld.done(); fld.next()) {
1110
    if (fld.access_flags().is_static()) {
1111
      Symbol* sig = fld.signature();
1112

1113
      writer->write_symbolID(fld.name());   // name
1114
      writer->write_u1(sig2tag(sig));       // type
1115

1116
      // value
1117
      dump_field_value(writer, sig->char_at(0), ik->java_mirror(), fld.offset());
1118
    }
1119
  }
1120

1121
  // Add resolved_references for each class that has them
1122
  oop resolved_references = ik->constants()->resolved_references_or_null();
1123
  if (resolved_references != nullptr) {
1124
    writer->write_symbolID(vmSymbols::resolved_references_name());  // name
1125
    writer->write_u1(sig2tag(vmSymbols::object_array_signature())); // type
1126
    writer->write_objectID(resolved_references);
1127

1128
    // Also write any previous versions
1129
    InstanceKlass* prev = ik->previous_versions();
1130
    while (prev != nullptr && prev->constants()->resolved_references_or_null() != nullptr) {
1131
      writer->write_symbolID(vmSymbols::resolved_references_name());  // name
1132
      writer->write_u1(sig2tag(vmSymbols::object_array_signature())); // type
1133
      writer->write_objectID(prev->constants()->resolved_references());
1134
      prev = prev->previous_versions();
1135
    }
1136
  }
1137

1138
  // Add init lock to the end if the class is not yet initialized
1139
  oop init_lock = ik->init_lock();
1140
  if (init_lock != nullptr) {
1141
    writer->write_symbolID(vmSymbols::init_lock_name());         // name
1142
    writer->write_u1(sig2tag(vmSymbols::int_array_signature())); // type
1143
    writer->write_objectID(init_lock);
1144
  }
1145
}
1146

1147
// dump the raw values of the instance fields of the given object
1148
void DumperSupport::dump_instance_fields(AbstractDumpWriter* writer, oop o, DumperClassCacheTableEntry* class_cache_entry) {
1149
  assert(class_cache_entry != nullptr, "Pre-condition: must be provided");
1150
  for (int idx = 0; idx < class_cache_entry->field_count(); idx++) {
1151
    dump_field_value(writer, class_cache_entry->sig_start(idx), o, class_cache_entry->offset(idx));
1152
  }
1153
}
1154

1155
// dumps the definition of the instance fields for a given class
1156
u2 DumperSupport::get_instance_fields_count(InstanceKlass* ik) {
1157
  u2 field_count = 0;
1158

1159
  for (JavaFieldStream fldc(ik); !fldc.done(); fldc.next()) {
1160
    if (!fldc.access_flags().is_static()) field_count++;
1161
  }
1162

1163
  return field_count;
1164
}
1165

1166
// dumps the definition of the instance fields for a given class
1167
void DumperSupport::dump_instance_field_descriptors(AbstractDumpWriter* writer, Klass* k) {
1168
  InstanceKlass* ik = InstanceKlass::cast(k);
1169

1170
  // dump the field descriptors
1171
  for (JavaFieldStream fld(ik); !fld.done(); fld.next()) {
1172
    if (!fld.access_flags().is_static()) {
1173
      Symbol* sig = fld.signature();
1174

1175
      writer->write_symbolID(fld.name());   // name
1176
      writer->write_u1(sig2tag(sig));       // type
1177
    }
1178
  }
1179
}
1180

1181
// creates HPROF_GC_INSTANCE_DUMP record for the given object
1182
void DumperSupport::dump_instance(AbstractDumpWriter* writer, oop o, DumperClassCacheTable* class_cache) {
1183
  InstanceKlass* ik = InstanceKlass::cast(o->klass());
1184

1185
  DumperClassCacheTableEntry* cache_entry = class_cache->lookup_or_create(ik);
1186

1187
  u4 is = instance_size(ik, cache_entry);
1188
  u4 size = 1 + sizeof(address) + 4 + sizeof(address) + 4 + is;
1189

1190
  writer->start_sub_record(HPROF_GC_INSTANCE_DUMP, size);
1191
  writer->write_objectID(o);
1192
  writer->write_u4(STACK_TRACE_ID);
1193

1194
  // class ID
1195
  writer->write_classID(ik);
1196

1197
  // number of bytes that follow
1198
  writer->write_u4(is);
1199

1200
  // field values
1201
  dump_instance_fields(writer, o, cache_entry);
1202

1203
  writer->end_sub_record();
1204
}
1205

1206
// creates HPROF_GC_CLASS_DUMP record for the given instance class
1207
void DumperSupport::dump_instance_class(AbstractDumpWriter* writer, Klass* k) {
1208
  InstanceKlass* ik = InstanceKlass::cast(k);
1209

1210
  // We can safepoint and do a heap dump at a point where we have a Klass,
1211
  // but no java mirror class has been setup for it. So we need to check
1212
  // that the class is at least loaded, to avoid crash from a null mirror.
1213
  if (!ik->is_loaded()) {
1214
    return;
1215
  }
1216

1217
  u2 static_fields_count = 0;
1218
  u4 static_size = get_static_fields_size(ik, static_fields_count);
1219
  u2 instance_fields_count = get_instance_fields_count(ik);
1220
  u4 instance_fields_size = instance_fields_count * (sizeof(address) + 1);
1221
  u4 size = checked_cast<u4>(1 + sizeof(address) + 4 + 6 * sizeof(address) + 4 + 2 + 2 + static_size + 2 + instance_fields_size);
1222

1223
  writer->start_sub_record(HPROF_GC_CLASS_DUMP, size);
1224

1225
  // class ID
1226
  writer->write_classID(ik);
1227
  writer->write_u4(STACK_TRACE_ID);
1228

1229
  // super class ID
1230
  InstanceKlass* java_super = ik->java_super();
1231
  if (java_super == nullptr) {
1232
    writer->write_objectID(oop(nullptr));
1233
  } else {
1234
    writer->write_classID(java_super);
1235
  }
1236

1237
  writer->write_objectID(ik->class_loader());
1238
  writer->write_objectID(ik->signers());
1239
  writer->write_objectID(ik->protection_domain());
1240

1241
  // reserved
1242
  writer->write_objectID(oop(nullptr));
1243
  writer->write_objectID(oop(nullptr));
1244

1245
  // instance size
1246
  writer->write_u4(DumperSupport::instance_size(ik));
1247

1248
  // size of constant pool - ignored by HAT 1.1
1249
  writer->write_u2(0);
1250

1251
  // static fields
1252
  writer->write_u2(static_fields_count);
1253
  dump_static_fields(writer, ik);
1254

1255
  // description of instance fields
1256
  writer->write_u2(instance_fields_count);
1257
  dump_instance_field_descriptors(writer, ik);
1258

1259
  writer->end_sub_record();
1260
}
1261

1262
// creates HPROF_GC_CLASS_DUMP record for the given array class
1263
void DumperSupport::dump_array_class(AbstractDumpWriter* writer, Klass* k) {
1264
  InstanceKlass* ik = nullptr; // bottom class for object arrays, null for primitive type arrays
1265
  if (k->is_objArray_klass()) {
1266
    Klass *bk = ObjArrayKlass::cast(k)->bottom_klass();
1267
    assert(bk != nullptr, "checking");
1268
    if (bk->is_instance_klass()) {
1269
      ik = InstanceKlass::cast(bk);
1270
    }
1271
  }
1272

1273
  u4 size = 1 + sizeof(address) + 4 + 6 * sizeof(address) + 4 + 2 + 2 + 2;
1274
  writer->start_sub_record(HPROF_GC_CLASS_DUMP, size);
1275
  writer->write_classID(k);
1276
  writer->write_u4(STACK_TRACE_ID);
1277

1278
  // super class of array classes is java.lang.Object
1279
  InstanceKlass* java_super = k->java_super();
1280
  assert(java_super != nullptr, "checking");
1281
  writer->write_classID(java_super);
1282

1283
  writer->write_objectID(ik == nullptr ? oop(nullptr) : ik->class_loader());
1284
  writer->write_objectID(ik == nullptr ? oop(nullptr) : ik->signers());
1285
  writer->write_objectID(ik == nullptr ? oop(nullptr) : ik->protection_domain());
1286

1287
  writer->write_objectID(oop(nullptr));    // reserved
1288
  writer->write_objectID(oop(nullptr));
1289
  writer->write_u4(0);             // instance size
1290
  writer->write_u2(0);             // constant pool
1291
  writer->write_u2(0);             // static fields
1292
  writer->write_u2(0);             // instance fields
1293

1294
  writer->end_sub_record();
1295

1296
}
1297

1298
// Hprof uses an u4 as record length field,
1299
// which means we need to truncate arrays that are too long.
1300
int DumperSupport::calculate_array_max_length(AbstractDumpWriter* writer, arrayOop array, short header_size) {
1301
  BasicType type = ArrayKlass::cast(array->klass())->element_type();
1302
  assert(type >= T_BOOLEAN && type <= T_OBJECT, "invalid array element type");
1303

1304
  int length = array->length();
1305

1306
  int type_size;
1307
  if (type == T_OBJECT) {
1308
    type_size = sizeof(address);
1309
  } else {
1310
    type_size = type2aelembytes(type);
1311
  }
1312

1313
  size_t length_in_bytes = (size_t)length * type_size;
1314
  uint max_bytes = max_juint - header_size;
1315

1316
  if (length_in_bytes > max_bytes) {
1317
    length = max_bytes / type_size;
1318
    length_in_bytes = (size_t)length * type_size;
1319

1320
    warning("cannot dump array of type %s[] with length %d; truncating to length %d",
1321
            type2name_tab[type], array->length(), length);
1322
  }
1323
  return length;
1324
}
1325

1326
// creates HPROF_GC_OBJ_ARRAY_DUMP record for the given object array
1327
void DumperSupport::dump_object_array(AbstractDumpWriter* writer, objArrayOop array) {
1328
  // sizeof(u1) + 2 * sizeof(u4) + sizeof(objectID) + sizeof(classID)
1329
  short header_size = 1 + 2 * 4 + 2 * sizeof(address);
1330
  int length = calculate_array_max_length(writer, array, header_size);
1331
  u4 size = checked_cast<u4>(header_size + length * sizeof(address));
1332

1333
  writer->start_sub_record(HPROF_GC_OBJ_ARRAY_DUMP, size);
1334
  writer->write_objectID(array);
1335
  writer->write_u4(STACK_TRACE_ID);
1336
  writer->write_u4(length);
1337

1338
  // array class ID
1339
  writer->write_classID(array->klass());
1340

1341
  // [id]* elements
1342
  for (int index = 0; index < length; index++) {
1343
    oop o = array->obj_at(index);
1344
    o = mask_dormant_archived_object(o, array);
1345
    writer->write_objectID(o);
1346
  }
1347

1348
  writer->end_sub_record();
1349
}
1350

1351
#define WRITE_ARRAY(Array, Type, Size, Length) \
1352
  for (int i = 0; i < Length; i++) { writer->write_##Size((Size)Array->Type##_at(i)); }
1353

1354
// creates HPROF_GC_PRIM_ARRAY_DUMP record for the given type array
1355
void DumperSupport::dump_prim_array(AbstractDumpWriter* writer, typeArrayOop array) {
1356
  BasicType type = TypeArrayKlass::cast(array->klass())->element_type();
1357
  // 2 * sizeof(u1) + 2 * sizeof(u4) + sizeof(objectID)
1358
  short header_size = 2 * 1 + 2 * 4 + sizeof(address);
1359

1360
  int length = calculate_array_max_length(writer, array, header_size);
1361
  int type_size = type2aelembytes(type);
1362
  u4 length_in_bytes = (u4)length * type_size;
1363
  u4 size = header_size + length_in_bytes;
1364

1365
  writer->start_sub_record(HPROF_GC_PRIM_ARRAY_DUMP, size);
1366
  writer->write_objectID(array);
1367
  writer->write_u4(STACK_TRACE_ID);
1368
  writer->write_u4(length);
1369
  writer->write_u1(type2tag(type));
1370

1371
  // nothing to copy
1372
  if (length == 0) {
1373
    writer->end_sub_record();
1374
    return;
1375
  }
1376

1377
  // If the byte ordering is big endian then we can copy most types directly
1378

1379
  switch (type) {
1380
    case T_INT : {
1381
      if (Endian::is_Java_byte_ordering_different()) {
1382
        WRITE_ARRAY(array, int, u4, length);
1383
      } else {
1384
        writer->write_raw(array->int_at_addr(0), length_in_bytes);
1385
      }
1386
      break;
1387
    }
1388
    case T_BYTE : {
1389
      writer->write_raw(array->byte_at_addr(0), length_in_bytes);
1390
      break;
1391
    }
1392
    case T_CHAR : {
1393
      if (Endian::is_Java_byte_ordering_different()) {
1394
        WRITE_ARRAY(array, char, u2, length);
1395
      } else {
1396
        writer->write_raw(array->char_at_addr(0), length_in_bytes);
1397
      }
1398
      break;
1399
    }
1400
    case T_SHORT : {
1401
      if (Endian::is_Java_byte_ordering_different()) {
1402
        WRITE_ARRAY(array, short, u2, length);
1403
      } else {
1404
        writer->write_raw(array->short_at_addr(0), length_in_bytes);
1405
      }
1406
      break;
1407
    }
1408
    case T_BOOLEAN : {
1409
      if (Endian::is_Java_byte_ordering_different()) {
1410
        WRITE_ARRAY(array, bool, u1, length);
1411
      } else {
1412
        writer->write_raw(array->bool_at_addr(0), length_in_bytes);
1413
      }
1414
      break;
1415
    }
1416
    case T_LONG : {
1417
      if (Endian::is_Java_byte_ordering_different()) {
1418
        WRITE_ARRAY(array, long, u8, length);
1419
      } else {
1420
        writer->write_raw(array->long_at_addr(0), length_in_bytes);
1421
      }
1422
      break;
1423
    }
1424

1425
    // handle float/doubles in a special value to ensure than NaNs are
1426
    // written correctly. TO DO: Check if we can avoid this on processors that
1427
    // use IEEE 754.
1428

1429
    case T_FLOAT : {
1430
      for (int i = 0; i < length; i++) {
1431
        dump_float(writer, array->float_at(i));
1432
      }
1433
      break;
1434
    }
1435
    case T_DOUBLE : {
1436
      for (int i = 0; i < length; i++) {
1437
        dump_double(writer, array->double_at(i));
1438
      }
1439
      break;
1440
    }
1441
    default : ShouldNotReachHere();
1442
  }
1443

1444
  writer->end_sub_record();
1445
}
1446

1447
// create a HPROF_FRAME record of the given Method* and bci
1448
void DumperSupport::dump_stack_frame(AbstractDumpWriter* writer,
1449
                                     int frame_serial_num,
1450
                                     int class_serial_num,
1451
                                     Method* m,
1452
                                     int bci) {
1453
  int line_number;
1454
  if (m->is_native()) {
1455
    line_number = -3;  // native frame
1456
  } else {
1457
    line_number = m->line_number_from_bci(bci);
1458
  }
1459

1460
  write_header(writer, HPROF_FRAME, 4*oopSize + 2*sizeof(u4));
1461
  writer->write_id(frame_serial_num);               // frame serial number
1462
  writer->write_symbolID(m->name());                // method's name
1463
  writer->write_symbolID(m->signature());           // method's signature
1464

1465
  assert(m->method_holder()->is_instance_klass(), "not InstanceKlass");
1466
  writer->write_symbolID(m->method_holder()->source_file_name());  // source file name
1467
  writer->write_u4(class_serial_num);               // class serial number
1468
  writer->write_u4((u4) line_number);               // line number
1469
}
1470

1471

1472
// Support class used to generate HPROF_UTF8 records from the entries in the
1473
// SymbolTable.
1474

1475
class SymbolTableDumper : public SymbolClosure {
1476
 private:
1477
  AbstractDumpWriter* _writer;
1478
  AbstractDumpWriter* writer() const                { return _writer; }
1479
 public:
1480
  SymbolTableDumper(AbstractDumpWriter* writer)     { _writer = writer; }
1481
  void do_symbol(Symbol** p);
1482
};
1483

1484
void SymbolTableDumper::do_symbol(Symbol** p) {
1485
  ResourceMark rm;
1486
  Symbol* sym = *p;
1487
  int len = sym->utf8_length();
1488
  if (len > 0) {
1489
    char* s = sym->as_utf8();
1490
    DumperSupport::write_header(writer(), HPROF_UTF8, oopSize + len);
1491
    writer()->write_symbolID(sym);
1492
    writer()->write_raw(s, len);
1493
  }
1494
}
1495

1496
// Support class used to generate HPROF_GC_CLASS_DUMP records
1497

1498
class ClassDumper : public KlassClosure {
1499
 private:
1500
  AbstractDumpWriter* _writer;
1501
  AbstractDumpWriter* writer() const { return _writer; }
1502

1503
 public:
1504
  ClassDumper(AbstractDumpWriter* writer) : _writer(writer) {}
1505

1506
  void do_klass(Klass* k) {
1507
    if (k->is_instance_klass()) {
1508
      DumperSupport::dump_instance_class(writer(), k);
1509
    } else {
1510
      DumperSupport::dump_array_class(writer(), k);
1511
    }
1512
  }
1513
};
1514

1515
// Support class used to generate HPROF_GC_ROOT_JNI_LOCAL records
1516

1517
class JNILocalsDumper : public OopClosure {
1518
 private:
1519
  AbstractDumpWriter* _writer;
1520
  u4 _thread_serial_num;
1521
  int _frame_num;
1522
  AbstractDumpWriter* writer() const                { return _writer; }
1523
 public:
1524
  JNILocalsDumper(AbstractDumpWriter* writer, u4 thread_serial_num) {
1525
    _writer = writer;
1526
    _thread_serial_num = thread_serial_num;
1527
    _frame_num = -1;  // default - empty stack
1528
  }
1529
  void set_frame_number(int n) { _frame_num = n; }
1530
  void do_oop(oop* obj_p);
1531
  void do_oop(narrowOop* obj_p) { ShouldNotReachHere(); }
1532
};
1533

1534
void JNILocalsDumper::do_oop(oop* obj_p) {
1535
  // ignore null handles
1536
  oop o = *obj_p;
1537
  if (o != nullptr) {
1538
    u4 size = 1 + sizeof(address) + 4 + 4;
1539
    writer()->start_sub_record(HPROF_GC_ROOT_JNI_LOCAL, size);
1540
    writer()->write_objectID(o);
1541
    writer()->write_u4(_thread_serial_num);
1542
    writer()->write_u4((u4)_frame_num);
1543
    writer()->end_sub_record();
1544
  }
1545
}
1546

1547

1548
// Support class used to generate HPROF_GC_ROOT_JNI_GLOBAL records
1549

1550
class JNIGlobalsDumper : public OopClosure {
1551
 private:
1552
  AbstractDumpWriter* _writer;
1553
  AbstractDumpWriter* writer() const                { return _writer; }
1554

1555
 public:
1556
  JNIGlobalsDumper(AbstractDumpWriter* writer) {
1557
    _writer = writer;
1558
  }
1559
  void do_oop(oop* obj_p);
1560
  void do_oop(narrowOop* obj_p) { ShouldNotReachHere(); }
1561
};
1562

1563
void JNIGlobalsDumper::do_oop(oop* obj_p) {
1564
  oop o = NativeAccess<AS_NO_KEEPALIVE>::oop_load(obj_p);
1565

1566
  // ignore these
1567
  if (o == nullptr) return;
1568
  // we ignore global ref to symbols and other internal objects
1569
  if (o->is_instance() || o->is_objArray() || o->is_typeArray()) {
1570
    u4 size = 1 + 2 * sizeof(address);
1571
    writer()->start_sub_record(HPROF_GC_ROOT_JNI_GLOBAL, size);
1572
    writer()->write_objectID(o);
1573
    writer()->write_rootID(obj_p);      // global ref ID
1574
    writer()->end_sub_record();
1575
  }
1576
};
1577

1578
// Support class used to generate HPROF_GC_ROOT_STICKY_CLASS records
1579

1580
class StickyClassDumper : public KlassClosure {
1581
 private:
1582
  AbstractDumpWriter* _writer;
1583
  AbstractDumpWriter* writer() const                { return _writer; }
1584
 public:
1585
  StickyClassDumper(AbstractDumpWriter* writer) {
1586
    _writer = writer;
1587
  }
1588
  void do_klass(Klass* k) {
1589
    if (k->is_instance_klass()) {
1590
      InstanceKlass* ik = InstanceKlass::cast(k);
1591
      u4 size = 1 + sizeof(address);
1592
      writer()->start_sub_record(HPROF_GC_ROOT_STICKY_CLASS, size);
1593
      writer()->write_classID(ik);
1594
      writer()->end_sub_record();
1595
    }
1596
  }
1597
};
1598

1599
// Support class used to generate HPROF_GC_ROOT_JAVA_FRAME records.
1600

1601
class JavaStackRefDumper : public StackObj {
1602
private:
1603
  AbstractDumpWriter* _writer;
1604
  u4 _thread_serial_num;
1605
  int _frame_num;
1606
  AbstractDumpWriter* writer() const { return _writer; }
1607
public:
1608
  JavaStackRefDumper(AbstractDumpWriter* writer, u4 thread_serial_num)
1609
      : _writer(writer), _thread_serial_num(thread_serial_num), _frame_num(-1) // default - empty stack
1610
  {
1611
  }
1612

1613
  void set_frame_number(int n) { _frame_num = n; }
1614

1615
  void dump_java_stack_refs(StackValueCollection* values);
1616
};
1617

1618
void JavaStackRefDumper::dump_java_stack_refs(StackValueCollection* values) {
1619
  for (int index = 0; index < values->size(); index++) {
1620
    if (values->at(index)->type() == T_OBJECT) {
1621
      oop o = values->obj_at(index)();
1622
      if (o != nullptr) {
1623
        u4 size = 1 + sizeof(address) + 4 + 4;
1624
        writer()->start_sub_record(HPROF_GC_ROOT_JAVA_FRAME, size);
1625
        writer()->write_objectID(o);
1626
        writer()->write_u4(_thread_serial_num);
1627
        writer()->write_u4((u4)_frame_num);
1628
        writer()->end_sub_record();
1629
      }
1630
    }
1631
  }
1632
}
1633

1634
// Class to collect, store and dump thread-related data:
1635
// - HPROF_TRACE and HPROF_FRAME records;
1636
// - HPROF_GC_ROOT_THREAD_OBJ/HPROF_GC_ROOT_JAVA_FRAME/HPROF_GC_ROOT_JNI_LOCAL subrecords.
1637
class ThreadDumper : public CHeapObj<mtInternal> {
1638
public:
1639
  enum class ThreadType { Platform, MountedVirtual, UnmountedVirtual };
1640

1641
private:
1642
  ThreadType _thread_type;
1643
  JavaThread* _java_thread;
1644
  oop _thread_oop;
1645

1646
  GrowableArray<StackFrameInfo*>* _frames;
1647
  // non-null if the thread is OOM thread
1648
  Method* _oome_constructor;
1649
  int _thread_serial_num;
1650
  int _start_frame_serial_num;
1651

1652
  vframe* get_top_frame() const;
1653

1654
public:
1655
  static bool should_dump_pthread(JavaThread* thread) {
1656
    return thread->threadObj() != nullptr && !thread->is_exiting() && !thread->is_hidden_from_external_view();
1657
  }
1658

1659
  static bool should_dump_vthread(oop vt) {
1660
    return java_lang_VirtualThread::state(vt) != java_lang_VirtualThread::NEW
1661
        && java_lang_VirtualThread::state(vt) != java_lang_VirtualThread::TERMINATED;
1662
  }
1663

1664
  static bool is_vthread_mounted(oop vt) {
1665
    // The code should be consistent with the "mounted virtual thread" case
1666
    // (VM_HeapDumper::dump_stack_traces(), ThreadDumper::get_top_frame()).
1667
    // I.e. virtual thread is mounted if its carrierThread is not null
1668
    // and is_vthread_mounted() for the carrier thread returns true.
1669
    oop carrier_thread = java_lang_VirtualThread::carrier_thread(vt);
1670
    if (carrier_thread == nullptr) {
1671
      return false;
1672
    }
1673
    JavaThread* java_thread = java_lang_Thread::thread(carrier_thread);
1674
    return java_thread->is_vthread_mounted();
1675
  }
1676

1677
  ThreadDumper(ThreadType thread_type, JavaThread* java_thread, oop thread_oop);
1678

1679
  // affects frame_count
1680
  void add_oom_frame(Method* oome_constructor) {
1681
    assert(_start_frame_serial_num == 0, "add_oom_frame cannot be called after init_serial_nums");
1682
    _oome_constructor = oome_constructor;
1683
  }
1684

1685
  void init_serial_nums(volatile int* thread_counter, volatile int* frame_counter) {
1686
    assert(_start_frame_serial_num == 0, "already initialized");
1687
    _thread_serial_num = Atomic::fetch_then_add(thread_counter, 1);
1688
    _start_frame_serial_num = Atomic::fetch_then_add(frame_counter, frame_count());
1689
  }
1690

1691
  bool oom_thread() const {
1692
    return _oome_constructor != nullptr;
1693
  }
1694

1695
  int frame_count() const {
1696
    return _frames->length() + (oom_thread() ? 1 : 0);
1697
  }
1698

1699
  u4 thread_serial_num() const {
1700
    return (u4)_thread_serial_num;
1701
  }
1702

1703
  u4 stack_trace_serial_num() const {
1704
    return (u4)(_thread_serial_num + STACK_TRACE_ID);
1705
  }
1706

1707
  // writes HPROF_TRACE and HPROF_FRAME records
1708
  // returns number of dumped frames
1709
  void dump_stack_traces(AbstractDumpWriter* writer, GrowableArray<Klass*>* klass_map);
1710

1711
  // writes HPROF_GC_ROOT_THREAD_OBJ subrecord
1712
  void dump_thread_obj(AbstractDumpWriter* writer);
1713

1714
  // Walk the stack of the thread.
1715
  // Dumps a HPROF_GC_ROOT_JAVA_FRAME subrecord for each local
1716
  // Dumps a HPROF_GC_ROOT_JNI_LOCAL subrecord for each JNI local
1717
  void dump_stack_refs(AbstractDumpWriter* writer);
1718

1719
};
1720

1721
ThreadDumper::ThreadDumper(ThreadType thread_type, JavaThread* java_thread, oop thread_oop)
1722
    : _thread_type(thread_type), _java_thread(java_thread), _thread_oop(thread_oop),
1723
      _oome_constructor(nullptr),
1724
      _thread_serial_num(0), _start_frame_serial_num(0)
1725
{
1726
  // sanity checks
1727
  if (_thread_type == ThreadType::UnmountedVirtual) {
1728
    assert(_java_thread == nullptr, "sanity");
1729
    assert(_thread_oop != nullptr, "sanity");
1730
  } else {
1731
    assert(_java_thread != nullptr, "sanity");
1732
    assert(_thread_oop != nullptr, "sanity");
1733
  }
1734

1735
  _frames = new (mtServiceability) GrowableArray<StackFrameInfo*>(10, mtServiceability);
1736
  bool stop_at_vthread_entry = _thread_type == ThreadType::MountedVirtual;
1737

1738
  // vframes are resource allocated
1739
  Thread* current_thread = Thread::current();
1740
  ResourceMark rm(current_thread);
1741
  HandleMark hm(current_thread);
1742

1743
  for (vframe* vf = get_top_frame(); vf != nullptr; vf = vf->sender()) {
1744
    if (stop_at_vthread_entry && vf->is_vthread_entry()) {
1745
      break;
1746
    }
1747
    if (vf->is_java_frame()) {
1748
      javaVFrame* jvf = javaVFrame::cast(vf);
1749
      _frames->append(new StackFrameInfo(jvf, false));
1750
    } else {
1751
      // ignore non-Java frames
1752
    }
1753
  }
1754
}
1755

1756
void ThreadDumper::dump_stack_traces(AbstractDumpWriter* writer, GrowableArray<Klass*>* klass_map) {
1757
  assert(_thread_serial_num != 0 && _start_frame_serial_num != 0, "serial_nums are not initialized");
1758

1759
  // write HPROF_FRAME records for this thread's stack trace
1760
  int depth = _frames->length();
1761
  int frame_serial_num = _start_frame_serial_num;
1762

1763
  if (oom_thread()) {
1764
    // OOM thread
1765
    // write fake frame that makes it look like the thread, which caused OOME,
1766
    // is in the OutOfMemoryError zero-parameter constructor
1767
    int oome_serial_num = klass_map->find(_oome_constructor->method_holder());
1768
    // the class serial number starts from 1
1769
    assert(oome_serial_num > 0, "OutOfMemoryError class not found");
1770
    DumperSupport::dump_stack_frame(writer, ++frame_serial_num, oome_serial_num, _oome_constructor, 0);
1771
    depth++;
1772
  }
1773

1774
  for (int j = 0; j < _frames->length(); j++) {
1775
    StackFrameInfo* frame = _frames->at(j);
1776
    Method* m = frame->method();
1777
    int class_serial_num = klass_map->find(m->method_holder());
1778
    // the class serial number starts from 1
1779
    assert(class_serial_num > 0, "class not found");
1780
    DumperSupport::dump_stack_frame(writer, ++frame_serial_num, class_serial_num, m, frame->bci());
1781
  }
1782

1783
  // write HPROF_TRACE record for the thread
1784
  DumperSupport::write_header(writer, HPROF_TRACE, checked_cast<u4>(3 * sizeof(u4) + depth * oopSize));
1785
  writer->write_u4(stack_trace_serial_num());   // stack trace serial number
1786
  writer->write_u4(thread_serial_num());        // thread serial number
1787
  writer->write_u4((u4)depth);                  // frame count (including oom frame)
1788
  for (int j = 1; j <= depth; j++) {
1789
    writer->write_id(_start_frame_serial_num + j);
1790
  }
1791
}
1792

1793
void ThreadDumper::dump_thread_obj(AbstractDumpWriter * writer) {
1794
  assert(_thread_serial_num != 0 && _start_frame_serial_num != 0, "serial_num is not initialized");
1795

1796
  u4 size = 1 + sizeof(address) + 4 + 4;
1797
  writer->start_sub_record(HPROF_GC_ROOT_THREAD_OBJ, size);
1798
  writer->write_objectID(_thread_oop);
1799
  writer->write_u4(thread_serial_num());      // thread serial number
1800
  writer->write_u4(stack_trace_serial_num()); // stack trace serial number
1801
  writer->end_sub_record();
1802
}
1803

1804
void ThreadDumper::dump_stack_refs(AbstractDumpWriter * writer) {
1805
  assert(_thread_serial_num != 0 && _start_frame_serial_num != 0, "serial_num is not initialized");
1806

1807
  JNILocalsDumper blk(writer, thread_serial_num());
1808
  if (_thread_type == ThreadType::Platform) {
1809
    if (!_java_thread->has_last_Java_frame()) {
1810
      // no last java frame but there may be JNI locals
1811
      _java_thread->active_handles()->oops_do(&blk);
1812
      return;
1813
    }
1814
  }
1815

1816
  JavaStackRefDumper java_ref_dumper(writer, thread_serial_num());
1817

1818
  // vframes are resource allocated
1819
  Thread* current_thread = Thread::current();
1820
  ResourceMark rm(current_thread);
1821
  HandleMark hm(current_thread);
1822

1823
  bool stopAtVthreadEntry = _thread_type == ThreadType::MountedVirtual;
1824
  frame* last_entry_frame = nullptr;
1825
  bool is_top_frame = true;
1826
  int depth = 0;
1827
  if (oom_thread()) {
1828
    depth++;
1829
  }
1830

1831
  for (vframe* vf = get_top_frame(); vf != nullptr; vf = vf->sender()) {
1832
    if (stopAtVthreadEntry && vf->is_vthread_entry()) {
1833
      break;
1834
    }
1835

1836
    if (vf->is_java_frame()) {
1837
      javaVFrame* jvf = javaVFrame::cast(vf);
1838
      if (!(jvf->method()->is_native())) {
1839
        java_ref_dumper.set_frame_number(depth);
1840
        java_ref_dumper.dump_java_stack_refs(jvf->locals());
1841
        java_ref_dumper.dump_java_stack_refs(jvf->expressions());
1842
      } else {
1843
        // native frame
1844
        blk.set_frame_number(depth);
1845
        if (is_top_frame) {
1846
          // JNI locals for the top frame.
1847
          assert(_java_thread != nullptr, "impossible for unmounted vthread");
1848
          _java_thread->active_handles()->oops_do(&blk);
1849
        } else {
1850
          if (last_entry_frame != nullptr) {
1851
            // JNI locals for the entry frame
1852
            assert(last_entry_frame->is_entry_frame(), "checking");
1853
            last_entry_frame->entry_frame_call_wrapper()->handles()->oops_do(&blk);
1854
          }
1855
        }
1856
      }
1857
      last_entry_frame = nullptr;
1858
      // increment only for Java frames
1859
      depth++;
1860
    } else {
1861
      // externalVFrame - for an entry frame then we report the JNI locals
1862
      // when we find the corresponding javaVFrame
1863
      frame* fr = vf->frame_pointer();
1864
      assert(fr != nullptr, "sanity check");
1865
      if (fr->is_entry_frame()) {
1866
        last_entry_frame = fr;
1867
      }
1868
    }
1869
  is_top_frame = false;
1870
  }
1871
  assert(depth == frame_count(), "total number of Java frames not matched");
1872
}
1873

1874
vframe* ThreadDumper::get_top_frame() const {
1875
  if (_thread_type == ThreadType::UnmountedVirtual) {
1876
    ContinuationWrapper cont(java_lang_VirtualThread::continuation(_thread_oop));
1877
    if (cont.is_empty()) {
1878
      return nullptr;
1879
    }
1880
    assert(!cont.is_mounted(), "sanity check");
1881
    stackChunkOop chunk = cont.last_nonempty_chunk();
1882
    if (chunk == nullptr || chunk->is_empty()) {
1883
      return nullptr;
1884
    }
1885

1886
    RegisterMap reg_map(cont.continuation(), RegisterMap::UpdateMap::include);
1887
    frame fr = chunk->top_frame(&reg_map);
1888
    vframe* vf = vframe::new_vframe(&fr, &reg_map, nullptr); // don't need JavaThread
1889
    return vf;
1890
  }
1891

1892
  RegisterMap reg_map(_java_thread,
1893
      RegisterMap::UpdateMap::include,
1894
      RegisterMap::ProcessFrames::include,
1895
      RegisterMap::WalkContinuation::skip);
1896
  switch (_thread_type) {
1897
  case ThreadType::Platform:
1898
    if (!_java_thread->has_last_Java_frame()) {
1899
      return nullptr;
1900
    }
1901
    return _java_thread->is_vthread_mounted()
1902
        ? _java_thread->carrier_last_java_vframe(&reg_map)
1903
        : _java_thread->platform_thread_last_java_vframe(&reg_map);
1904

1905
  case ThreadType::MountedVirtual:
1906
    return _java_thread->last_java_vframe(&reg_map);
1907

1908
  default: // make compilers happy
1909
      break;
1910
  }
1911
  ShouldNotReachHere();
1912
  return nullptr;
1913
}
1914

1915
// Callback to dump thread-related data for unmounted virtual threads;
1916
// implemented by VM_HeapDumper.
1917
class UnmountedVThreadDumper {
1918
 public:
1919
  virtual void dump_vthread(oop vt, AbstractDumpWriter* segment_writer) = 0;
1920
};
1921

1922
// Support class used when iterating over the heap.
1923
class HeapObjectDumper : public ObjectClosure {
1924
 private:
1925
  AbstractDumpWriter* _writer;
1926
  AbstractDumpWriter* writer()                  { return _writer; }
1927
  UnmountedVThreadDumper* _vthread_dumper;
1928

1929
  DumperClassCacheTable _class_cache;
1930

1931
 public:
1932
  HeapObjectDumper(AbstractDumpWriter* writer, UnmountedVThreadDumper* vthread_dumper)
1933
    : _writer(writer), _vthread_dumper(vthread_dumper) {}
1934

1935
  // called for each object in the heap
1936
  void do_object(oop o);
1937
};
1938

1939
void HeapObjectDumper::do_object(oop o) {
1940
  // skip classes as these emitted as HPROF_GC_CLASS_DUMP records
1941
  if (o->klass() == vmClasses::Class_klass()) {
1942
    if (!java_lang_Class::is_primitive(o)) {
1943
      return;
1944
    }
1945
  }
1946

1947
  if (DumperSupport::mask_dormant_archived_object(o, nullptr) == nullptr) {
1948
    return;
1949
  }
1950

1951
  if (o->is_instance()) {
1952
    // create a HPROF_GC_INSTANCE record for each object
1953
    DumperSupport::dump_instance(writer(), o, &_class_cache);
1954
    // If we encounter an unmounted virtual thread it needs to be dumped explicitly
1955
    // (mounted virtual threads are dumped with their carriers).
1956
    if (java_lang_VirtualThread::is_instance(o)
1957
        && ThreadDumper::should_dump_vthread(o) && !ThreadDumper::is_vthread_mounted(o)) {
1958
      _vthread_dumper->dump_vthread(o, writer());
1959
    }
1960
  } else if (o->is_objArray()) {
1961
    // create a HPROF_GC_OBJ_ARRAY_DUMP record for each object array
1962
    DumperSupport::dump_object_array(writer(), objArrayOop(o));
1963
  } else if (o->is_typeArray()) {
1964
    // create a HPROF_GC_PRIM_ARRAY_DUMP record for each type array
1965
    DumperSupport::dump_prim_array(writer(), typeArrayOop(o));
1966
  }
1967
}
1968

1969
// The dumper controller for parallel heap dump
1970
class DumperController : public CHeapObj<mtInternal> {
1971
 private:
1972
   Monitor* _lock;
1973
   Mutex* _global_writer_lock;
1974

1975
   const uint   _dumper_number;
1976
   uint   _complete_number;
1977

1978
   bool   _started; // VM dumper started and acquired global writer lock
1979

1980
 public:
1981
   DumperController(uint number) :
1982
     // _lock and _global_writer_lock are used for synchronization between GC worker threads inside safepoint,
1983
     // so we lock with _no_safepoint_check_flag.
1984
     // signal_start() acquires _lock when global writer is locked,
1985
     // its rank must be less than _global_writer_lock rank.
1986
     _lock(new (std::nothrow) PaddedMonitor(Mutex::nosafepoint - 1, "DumperController_lock")),
1987
     _global_writer_lock(new (std::nothrow) Mutex(Mutex::nosafepoint, "DumpWriter_lock")),
1988
     _dumper_number(number),
1989
     _complete_number(0),
1990
     _started(false)
1991
   {}
1992

1993
   ~DumperController() {
1994
     delete _lock;
1995
     delete _global_writer_lock;
1996
   }
1997

1998
   // parallel (non VM) dumpers must wait until VM dumper acquires global writer lock
1999
   void wait_for_start_signal() {
2000
     MonitorLocker ml(_lock, Mutex::_no_safepoint_check_flag);
2001
     while (_started == false) {
2002
       ml.wait();
2003
     }
2004
   }
2005

2006
   void signal_start() {
2007
     MonitorLocker ml(_lock, Mutex::_no_safepoint_check_flag);
2008
     _started = true;
2009
     ml.notify_all();
2010
   }
2011

2012
   void lock_global_writer() {
2013
     _global_writer_lock->lock_without_safepoint_check();
2014
   }
2015

2016
   void unlock_global_writer() {
2017
     _global_writer_lock->unlock();
2018
   }
2019

2020
   void dumper_complete(DumpWriter* local_writer, DumpWriter* global_writer) {
2021
     MonitorLocker ml(_lock, Mutex::_no_safepoint_check_flag);
2022
     _complete_number++;
2023
     // propagate local error to global if any
2024
     if (local_writer->has_error()) {
2025
       global_writer->set_error(local_writer->error());
2026
     }
2027
     ml.notify();
2028
   }
2029

2030
   void wait_all_dumpers_complete() {
2031
     MonitorLocker ml(_lock, Mutex::_no_safepoint_check_flag);
2032
     while (_complete_number != _dumper_number) {
2033
        ml.wait();
2034
     }
2035
   }
2036
};
2037

2038
// DumpMerger merges separate dump files into a complete one
2039
class DumpMerger : public StackObj {
2040
private:
2041
  DumpWriter* _writer;
2042
  const char* _path;
2043
  bool _has_error;
2044
  int _dump_seq;
2045

2046
private:
2047
  void merge_file(const char* path);
2048
  void merge_done();
2049
  void set_error(const char* msg);
2050

2051
public:
2052
  DumpMerger(const char* path, DumpWriter* writer, int dump_seq) :
2053
    _writer(writer),
2054
    _path(path),
2055
    _has_error(_writer->has_error()),
2056
    _dump_seq(dump_seq) {}
2057

2058
  void do_merge();
2059

2060
  // returns path for the parallel DumpWriter (resource allocated)
2061
  static char* get_writer_path(const char* base_path, int seq);
2062

2063
};
2064

2065
char* DumpMerger::get_writer_path(const char* base_path, int seq) {
2066
  // approximate required buffer size
2067
  size_t buf_size = strlen(base_path)
2068
                    + 2                 // ".p"
2069
                    + 10                // number (that's enough for 2^32 parallel dumpers)
2070
                    + 1;                // '\0'
2071

2072
  char* path = NEW_RESOURCE_ARRAY(char, buf_size);
2073
  memset(path, 0, buf_size);
2074

2075
  os::snprintf(path, buf_size, "%s.p%d", base_path, seq);
2076

2077
  return path;
2078
}
2079

2080

2081
void DumpMerger::merge_done() {
2082
  // Writes the HPROF_HEAP_DUMP_END record.
2083
  if (!_has_error) {
2084
    DumperSupport::end_of_dump(_writer);
2085
    _writer->flush();
2086
  }
2087
  _dump_seq = 0; //reset
2088
}
2089

2090
void DumpMerger::set_error(const char* msg) {
2091
  assert(msg != nullptr, "sanity check");
2092
  log_error(heapdump)("%s (file: %s)", msg, _path);
2093
  _writer->set_error(msg);
2094
  _has_error = true;
2095
}
2096

2097
#ifdef LINUX
2098
// Merge segmented heap files via sendfile, it's more efficient than the
2099
// read+write combination, which would require transferring data to and from
2100
// user space.
2101
void DumpMerger::merge_file(const char* path) {
2102
  TraceTime timer("Merge segmented heap file directly", TRACETIME_LOG(Info, heapdump));
2103

2104
  int segment_fd = os::open(path, O_RDONLY, 0);
2105
  if (segment_fd == -1) {
2106
    set_error("Can not open segmented heap file during merging");
2107
    return;
2108
  }
2109

2110
  struct stat st;
2111
  if (os::stat(path, &st) != 0) {
2112
    ::close(segment_fd);
2113
    set_error("Can not get segmented heap file size during merging");
2114
    return;
2115
  }
2116

2117
  // A successful call to sendfile may write fewer bytes than requested; the
2118
  // caller should be prepared to retry the call if there were unsent bytes.
2119
  jlong offset = 0;
2120
  while (offset < st.st_size) {
2121
    int ret = os::Linux::sendfile(_writer->get_fd(), segment_fd, &offset, st.st_size);
2122
    if (ret == -1) {
2123
      ::close(segment_fd);
2124
      set_error("Failed to merge segmented heap file");
2125
      return;
2126
    }
2127
  }
2128

2129
  // As sendfile variant does not call the write method of the global writer,
2130
  // bytes_written is also incorrect for this variant, we need to explicitly
2131
  // accumulate bytes_written for the global writer in this case
2132
  julong accum = _writer->bytes_written() + st.st_size;
2133
  _writer->set_bytes_written(accum);
2134
  ::close(segment_fd);
2135
}
2136
#else
2137
// Generic implementation using read+write
2138
void DumpMerger::merge_file(const char* path) {
2139
  TraceTime timer("Merge segmented heap file", TRACETIME_LOG(Info, heapdump));
2140

2141
  fileStream segment_fs(path, "rb");
2142
  if (!segment_fs.is_open()) {
2143
    set_error("Can not open segmented heap file during merging");
2144
    return;
2145
  }
2146

2147
  jlong total = 0;
2148
  size_t cnt = 0;
2149

2150
  // Use _writer buffer for reading.
2151
  while ((cnt = segment_fs.read(_writer->buffer(), 1, _writer->buffer_size())) != 0) {
2152
    _writer->set_position(cnt);
2153
    _writer->flush();
2154
    total += cnt;
2155
  }
2156

2157
  if (segment_fs.fileSize() != total) {
2158
    set_error("Merged heap dump is incomplete");
2159
  }
2160
}
2161
#endif
2162

2163
void DumpMerger::do_merge() {
2164
  TraceTime timer("Merge heap files complete", TRACETIME_LOG(Info, heapdump));
2165

2166
  // Since contents in segmented heap file were already zipped, we don't need to zip
2167
  // them again during merging.
2168
  AbstractCompressor* saved_compressor = _writer->compressor();
2169
  _writer->set_compressor(nullptr);
2170

2171
  // Merge the content of the remaining files into base file. Regardless of whether
2172
  // the merge process is successful or not, these segmented files will be deleted.
2173
  for (int i = 0; i < _dump_seq; i++) {
2174
    ResourceMark rm;
2175
    const char* path = get_writer_path(_path, i);
2176
    if (!_has_error) {
2177
      merge_file(path);
2178
    }
2179
    // Delete selected segmented heap file nevertheless
2180
    if (remove(path) != 0) {
2181
      log_info(heapdump)("Removal of segment file (%d) failed (%d)", i, errno);
2182
    }
2183
  }
2184

2185
  // restore compressor for further use
2186
  _writer->set_compressor(saved_compressor);
2187
  merge_done();
2188
}
2189

2190
// The VM operation that performs the heap dump
2191
class VM_HeapDumper : public VM_GC_Operation, public WorkerTask, public UnmountedVThreadDumper {
2192
 private:
2193
  static VM_HeapDumper*   _global_dumper;
2194
  static DumpWriter*      _global_writer;
2195
  DumpWriter*             _local_writer;
2196
  JavaThread*             _oome_thread;
2197
  Method*                 _oome_constructor;
2198
  bool                    _gc_before_heap_dump;
2199
  GrowableArray<Klass*>*  _klass_map;
2200

2201
  ThreadDumper**          _thread_dumpers; // platform, carrier and mounted virtual threads
2202
  int                     _thread_dumpers_count;
2203
  volatile int            _thread_serial_num;
2204
  volatile int            _frame_serial_num;
2205

2206
  volatile int            _dump_seq;
2207
  // parallel heap dump support
2208
  uint                    _num_dumper_threads;
2209
  DumperController*       _dumper_controller;
2210
  ParallelObjectIterator* _poi;
2211

2212
  // Dumper id of VMDumper thread.
2213
  static const int VMDumperId = 0;
2214
  // VM dumper dumps both heap and non-heap data, other dumpers dump heap-only data.
2215
  static bool is_vm_dumper(int dumper_id) { return dumper_id == VMDumperId; }
2216
  // the 1st dumper calling get_next_dumper_id becomes VM dumper
2217
  int get_next_dumper_id() {
2218
    return Atomic::fetch_then_add(&_dump_seq, 1);
2219
  }
2220

2221
  // accessors and setters
2222
  static VM_HeapDumper* dumper()         {  assert(_global_dumper != nullptr, "Error"); return _global_dumper; }
2223
  static DumpWriter* writer()            {  assert(_global_writer != nullptr, "Error"); return _global_writer; }
2224

2225
  void set_global_dumper() {
2226
    assert(_global_dumper == nullptr, "Error");
2227
    _global_dumper = this;
2228
  }
2229
  void set_global_writer() {
2230
    assert(_global_writer == nullptr, "Error");
2231
    _global_writer = _local_writer;
2232
  }
2233
  void clear_global_dumper() { _global_dumper = nullptr; }
2234
  void clear_global_writer() { _global_writer = nullptr; }
2235

2236
  bool skip_operation() const;
2237

2238
  // writes a HPROF_LOAD_CLASS record to global writer
2239
  static void do_load_class(Klass* k);
2240

2241
  // HPROF_GC_ROOT_THREAD_OBJ records for platform and mounted virtual threads
2242
  void dump_threads(AbstractDumpWriter* writer);
2243

2244
  void add_class_serial_number(Klass* k, int serial_num) {
2245
    _klass_map->at_put_grow(serial_num, k);
2246
  }
2247

2248
  bool is_oom_thread(JavaThread* thread) const {
2249
    return thread == _oome_thread && _oome_constructor != nullptr;
2250
  }
2251

2252
  // HPROF_TRACE and HPROF_FRAME records for platform and mounted virtual threads
2253
  void dump_stack_traces(AbstractDumpWriter* writer);
2254

2255
 public:
2256
  VM_HeapDumper(DumpWriter* writer, bool gc_before_heap_dump, bool oome, uint num_dump_threads) :
2257
    VM_GC_Operation(0 /* total collections,      dummy, ignored */,
2258
                    GCCause::_heap_dump /* GC Cause */,
2259
                    0 /* total full collections, dummy, ignored */,
2260
                    gc_before_heap_dump),
2261
    WorkerTask("dump heap") {
2262
    _local_writer = writer;
2263
    _gc_before_heap_dump = gc_before_heap_dump;
2264
    _klass_map = new (mtServiceability) GrowableArray<Klass*>(INITIAL_CLASS_COUNT, mtServiceability);
2265

2266
    _thread_dumpers = nullptr;
2267
    _thread_dumpers_count = 0;
2268
    _thread_serial_num = 1;
2269
    _frame_serial_num = 1;
2270

2271
    _dump_seq = VMDumperId;
2272
    _num_dumper_threads = num_dump_threads;
2273
    _dumper_controller = nullptr;
2274
    _poi = nullptr;
2275
    if (oome) {
2276
      assert(!Thread::current()->is_VM_thread(), "Dump from OutOfMemoryError cannot be called by the VMThread");
2277
      // get OutOfMemoryError zero-parameter constructor
2278
      InstanceKlass* oome_ik = vmClasses::OutOfMemoryError_klass();
2279
      _oome_constructor = oome_ik->find_method(vmSymbols::object_initializer_name(),
2280
                                                          vmSymbols::void_method_signature());
2281
      // get thread throwing OOME when generating the heap dump at OOME
2282
      _oome_thread = JavaThread::current();
2283
    } else {
2284
      _oome_thread = nullptr;
2285
      _oome_constructor = nullptr;
2286
    }
2287
  }
2288

2289
  ~VM_HeapDumper() {
2290
    if (_thread_dumpers != nullptr) {
2291
      for (int i = 0; i < _thread_dumpers_count; i++) {
2292
        delete _thread_dumpers[i];
2293
      }
2294
      FREE_C_HEAP_ARRAY(ThreadDumper*, _thread_dumpers);
2295
    }
2296

2297
    if (_dumper_controller != nullptr) {
2298
      delete _dumper_controller;
2299
      _dumper_controller = nullptr;
2300
    }
2301
    delete _klass_map;
2302
  }
2303
  int dump_seq()           { return _dump_seq; }
2304
  bool is_parallel_dump()  { return _num_dumper_threads > 1; }
2305
  void prepare_parallel_dump(WorkerThreads* workers);
2306

2307
  VMOp_Type type() const { return VMOp_HeapDumper; }
2308
  virtual bool doit_prologue();
2309
  void doit();
2310
  void work(uint worker_id);
2311

2312
  // UnmountedVThreadDumper implementation
2313
  void dump_vthread(oop vt, AbstractDumpWriter* segment_writer);
2314
};
2315

2316
VM_HeapDumper* VM_HeapDumper::_global_dumper = nullptr;
2317
DumpWriter*    VM_HeapDumper::_global_writer = nullptr;
2318

2319
bool VM_HeapDumper::skip_operation() const {
2320
  return false;
2321
}
2322

2323
// fixes up the current dump record and writes HPROF_HEAP_DUMP_END record
2324
void DumperSupport::end_of_dump(AbstractDumpWriter* writer) {
2325
  writer->finish_dump_segment();
2326

2327
  writer->write_u1(HPROF_HEAP_DUMP_END);
2328
  writer->write_u4(0);
2329
  writer->write_u4(0);
2330
}
2331

2332
// writes a HPROF_LOAD_CLASS record for the class
2333
void VM_HeapDumper::do_load_class(Klass* k) {
2334
  static u4 class_serial_num = 0;
2335

2336
  // len of HPROF_LOAD_CLASS record
2337
  u4 remaining = 2*oopSize + 2*sizeof(u4);
2338

2339
  DumperSupport::write_header(writer(), HPROF_LOAD_CLASS, remaining);
2340

2341
  // class serial number is just a number
2342
  writer()->write_u4(++class_serial_num);
2343

2344
  // class ID
2345
  writer()->write_classID(k);
2346

2347
  // add the Klass* and class serial number pair
2348
  dumper()->add_class_serial_number(k, class_serial_num);
2349

2350
  writer()->write_u4(STACK_TRACE_ID);
2351

2352
  // class name ID
2353
  Symbol* name = k->name();
2354
  writer()->write_symbolID(name);
2355
}
2356

2357
// Write a HPROF_GC_ROOT_THREAD_OBJ record for platform/carrier and mounted virtual threads.
2358
// Then walk the stack so that locals and JNI locals are dumped.
2359
void VM_HeapDumper::dump_threads(AbstractDumpWriter* writer) {
2360
  for (int i = 0; i < _thread_dumpers_count; i++) {
2361
    _thread_dumpers[i]->dump_thread_obj(writer);
2362
    _thread_dumpers[i]->dump_stack_refs(writer);
2363
  }
2364
}
2365

2366
bool VM_HeapDumper::doit_prologue() {
2367
  if (_gc_before_heap_dump && UseZGC) {
2368
    // ZGC cannot perform a synchronous GC cycle from within the VM thread.
2369
    // So ZCollectedHeap::collect_as_vm_thread() is a noop. To respect the
2370
    // _gc_before_heap_dump flag a synchronous GC cycle is performed from
2371
    // the caller thread in the prologue.
2372
    Universe::heap()->collect(GCCause::_heap_dump);
2373
  }
2374
  return VM_GC_Operation::doit_prologue();
2375
}
2376

2377
void VM_HeapDumper::prepare_parallel_dump(WorkerThreads* workers) {
2378
  uint num_active_workers = workers != nullptr ? workers->active_workers() : 0;
2379
  uint num_requested_dump_threads = _num_dumper_threads;
2380
  // check if we can dump in parallel based on requested and active threads
2381
  if (num_active_workers <= 1 || num_requested_dump_threads <= 1) {
2382
    _num_dumper_threads = 1;
2383
  } else {
2384
    _num_dumper_threads = clamp(num_requested_dump_threads, 2U, num_active_workers);
2385
  }
2386
  _dumper_controller = new (std::nothrow) DumperController(_num_dumper_threads);
2387
  bool can_parallel = _num_dumper_threads > 1;
2388
  log_info(heapdump)("Requested dump threads %u, active dump threads %u, "
2389
                     "actual dump threads %u, parallelism %s",
2390
                     num_requested_dump_threads, num_active_workers,
2391
                     _num_dumper_threads, can_parallel ? "true" : "false");
2392
}
2393

2394
// The VM operation that dumps the heap. The dump consists of the following
2395
// records:
2396
//
2397
//  HPROF_HEADER
2398
//  [HPROF_UTF8]*
2399
//  [HPROF_LOAD_CLASS]*
2400
//  [[HPROF_FRAME]*|HPROF_TRACE]*
2401
//  [HPROF_GC_CLASS_DUMP]*
2402
//  [HPROF_HEAP_DUMP_SEGMENT]*
2403
//  HPROF_HEAP_DUMP_END
2404
//
2405
// The HPROF_TRACE records represent the stack traces where the heap dump
2406
// is generated and a "dummy trace" record which does not include
2407
// any frames. The dummy trace record is used to be referenced as the
2408
// unknown object alloc site.
2409
//
2410
// Each HPROF_HEAP_DUMP_SEGMENT record has a length followed by sub-records.
2411
// To allow the heap dump be generated in a single pass we remember the position
2412
// of the dump length and fix it up after all sub-records have been written.
2413
// To generate the sub-records we iterate over the heap, writing
2414
// HPROF_GC_INSTANCE_DUMP, HPROF_GC_OBJ_ARRAY_DUMP, and HPROF_GC_PRIM_ARRAY_DUMP
2415
// records as we go. Once that is done we write records for some of the GC
2416
// roots.
2417

2418
void VM_HeapDumper::doit() {
2419

2420
  CollectedHeap* ch = Universe::heap();
2421

2422
  ch->ensure_parsability(false); // must happen, even if collection does
2423
                                 // not happen (e.g. due to GCLocker)
2424

2425
  if (_gc_before_heap_dump) {
2426
    if (GCLocker::is_active()) {
2427
      warning("GC locker is held; pre-heapdump GC was skipped");
2428
    } else {
2429
      ch->collect_as_vm_thread(GCCause::_heap_dump);
2430
    }
2431
  }
2432

2433
  // At this point we should be the only dumper active, so
2434
  // the following should be safe.
2435
  set_global_dumper();
2436
  set_global_writer();
2437

2438
  WorkerThreads* workers = ch->safepoint_workers();
2439
  prepare_parallel_dump(workers);
2440

2441
  if (!is_parallel_dump()) {
2442
    work(VMDumperId);
2443
  } else {
2444
    ParallelObjectIterator poi(_num_dumper_threads);
2445
    _poi = &poi;
2446
    workers->run_task(this, _num_dumper_threads);
2447
    _poi = nullptr;
2448
  }
2449

2450
  // Now we clear the global variables, so that a future dumper can run.
2451
  clear_global_dumper();
2452
  clear_global_writer();
2453
}
2454

2455
void VM_HeapDumper::work(uint worker_id) {
2456
  // VM Dumper works on all non-heap data dumping and part of heap iteration.
2457
  int dumper_id = get_next_dumper_id();
2458

2459
  if (is_vm_dumper(dumper_id)) {
2460
    // lock global writer, it will be unlocked after VM Dumper finishes with non-heap data
2461
    _dumper_controller->lock_global_writer();
2462
    _dumper_controller->signal_start();
2463
  } else {
2464
    _dumper_controller->wait_for_start_signal();
2465
  }
2466

2467
  if (is_vm_dumper(dumper_id)) {
2468
    TraceTime timer("Dump non-objects", TRACETIME_LOG(Info, heapdump));
2469
    // Write the file header - we always use 1.0.2
2470
    const char* header = "JAVA PROFILE 1.0.2";
2471

2472
    // header is few bytes long - no chance to overflow int
2473
    writer()->write_raw(header, strlen(header) + 1); // NUL terminated
2474
    writer()->write_u4(oopSize);
2475
    // timestamp is current time in ms
2476
    writer()->write_u8(os::javaTimeMillis());
2477
    // HPROF_UTF8 records
2478
    SymbolTableDumper sym_dumper(writer());
2479
    SymbolTable::symbols_do(&sym_dumper);
2480

2481
    // write HPROF_LOAD_CLASS records
2482
    {
2483
      LockedClassesDo locked_load_classes(&do_load_class);
2484
      ClassLoaderDataGraph::classes_do(&locked_load_classes);
2485
    }
2486

2487
    // write HPROF_FRAME and HPROF_TRACE records
2488
    // this must be called after _klass_map is built when iterating the classes above.
2489
    dump_stack_traces(writer());
2490

2491
    // unlock global writer, so parallel dumpers can dump stack traces of unmounted virtual threads
2492
    _dumper_controller->unlock_global_writer();
2493
  }
2494

2495
  // HPROF_HEAP_DUMP/HPROF_HEAP_DUMP_SEGMENT starts here
2496

2497
  ResourceMark rm;
2498
  // share global compressor, local DumpWriter is not responsible for its life cycle
2499
  DumpWriter segment_writer(DumpMerger::get_writer_path(writer()->get_file_path(), dumper_id),
2500
                            writer()->is_overwrite(), writer()->compressor());
2501
  if (!segment_writer.has_error()) {
2502
    if (is_vm_dumper(dumper_id)) {
2503
      // dump some non-heap subrecords to heap dump segment
2504
      TraceTime timer("Dump non-objects (part 2)", TRACETIME_LOG(Info, heapdump));
2505
      // Writes HPROF_GC_CLASS_DUMP records
2506
      ClassDumper class_dumper(&segment_writer);
2507
      ClassLoaderDataGraph::classes_do(&class_dumper);
2508

2509
      // HPROF_GC_ROOT_THREAD_OBJ + frames + jni locals
2510
      dump_threads(&segment_writer);
2511

2512
      // HPROF_GC_ROOT_JNI_GLOBAL
2513
      JNIGlobalsDumper jni_dumper(&segment_writer);
2514
      JNIHandles::oops_do(&jni_dumper);
2515
      // technically not jni roots, but global roots
2516
      // for things like preallocated throwable backtraces
2517
      Universe::vm_global()->oops_do(&jni_dumper);
2518
      // HPROF_GC_ROOT_STICKY_CLASS
2519
      // These should be classes in the null class loader data, and not all classes
2520
      // if !ClassUnloading
2521
      StickyClassDumper stiky_class_dumper(&segment_writer);
2522
      ClassLoaderData::the_null_class_loader_data()->classes_do(&stiky_class_dumper);
2523
    }
2524

2525
    // Heap iteration.
2526
    // writes HPROF_GC_INSTANCE_DUMP records.
2527
    // After each sub-record is written check_segment_length will be invoked
2528
    // to check if the current segment exceeds a threshold. If so, a new
2529
    // segment is started.
2530
    // The HPROF_GC_CLASS_DUMP and HPROF_GC_INSTANCE_DUMP are the vast bulk
2531
    // of the heap dump.
2532

2533
    TraceTime timer(is_parallel_dump() ? "Dump heap objects in parallel" : "Dump heap objects", TRACETIME_LOG(Info, heapdump));
2534
    HeapObjectDumper obj_dumper(&segment_writer, this);
2535
    if (!is_parallel_dump()) {
2536
      Universe::heap()->object_iterate(&obj_dumper);
2537
    } else {
2538
      // == Parallel dump
2539
      _poi->object_iterate(&obj_dumper, worker_id);
2540
    }
2541

2542
    segment_writer.finish_dump_segment();
2543
    segment_writer.flush();
2544
  }
2545

2546
  _dumper_controller->dumper_complete(&segment_writer, writer());
2547

2548
  if (is_vm_dumper(dumper_id)) {
2549
    _dumper_controller->wait_all_dumpers_complete();
2550

2551
    // flush global writer
2552
    writer()->flush();
2553

2554
    // At this point, all fragments of the heapdump have been written to separate files.
2555
    // We need to merge them into a complete heapdump and write HPROF_HEAP_DUMP_END at that time.
2556
  }
2557
}
2558

2559
void VM_HeapDumper::dump_stack_traces(AbstractDumpWriter* writer) {
2560
  // write a HPROF_TRACE record without any frames to be referenced as object alloc sites
2561
  DumperSupport::write_header(writer, HPROF_TRACE, 3 * sizeof(u4));
2562
  writer->write_u4((u4)STACK_TRACE_ID);
2563
  writer->write_u4(0);                    // thread number
2564
  writer->write_u4(0);                    // frame count
2565

2566
  // max number if every platform thread is carrier with mounted virtual thread
2567
  _thread_dumpers = NEW_C_HEAP_ARRAY(ThreadDumper*, Threads::number_of_threads() * 2, mtInternal);
2568

2569
  for (JavaThreadIteratorWithHandle jtiwh; JavaThread * thread = jtiwh.next(); ) {
2570
    if (ThreadDumper::should_dump_pthread(thread)) {
2571
      bool add_oom_frame = is_oom_thread(thread);
2572

2573
      oop mounted_vt = thread->is_vthread_mounted() ? thread->vthread() : nullptr;
2574
      if (mounted_vt != nullptr && !ThreadDumper::should_dump_vthread(mounted_vt)) {
2575
        mounted_vt = nullptr;
2576
      }
2577

2578
      // mounted vthread (if any)
2579
      if (mounted_vt != nullptr) {
2580
        ThreadDumper* thread_dumper = new ThreadDumper(ThreadDumper::ThreadType::MountedVirtual, thread, mounted_vt);
2581
        _thread_dumpers[_thread_dumpers_count++] = thread_dumper;
2582
        if (add_oom_frame) {
2583
          thread_dumper->add_oom_frame(_oome_constructor);
2584
          // we add oom frame to the VT stack, don't add it to the carrier thread stack
2585
          add_oom_frame = false;
2586
        }
2587
        thread_dumper->init_serial_nums(&_thread_serial_num, &_frame_serial_num);
2588
        thread_dumper->dump_stack_traces(writer, _klass_map);
2589
      }
2590

2591
      // platform or carrier thread
2592
      ThreadDumper* thread_dumper = new ThreadDumper(ThreadDumper::ThreadType::Platform, thread, thread->threadObj());
2593
      _thread_dumpers[_thread_dumpers_count++] = thread_dumper;
2594
      if (add_oom_frame) {
2595
        thread_dumper->add_oom_frame(_oome_constructor);
2596
      }
2597
      thread_dumper->init_serial_nums(&_thread_serial_num, &_frame_serial_num);
2598
      thread_dumper->dump_stack_traces(writer, _klass_map);
2599
    }
2600
  }
2601
}
2602

2603
void VM_HeapDumper::dump_vthread(oop vt, AbstractDumpWriter* segment_writer) {
2604
  // unmounted vthread has no JavaThread
2605
  ThreadDumper thread_dumper(ThreadDumper::ThreadType::UnmountedVirtual, nullptr, vt);
2606
  thread_dumper.init_serial_nums(&_thread_serial_num, &_frame_serial_num);
2607

2608
  // write HPROF_TRACE/HPROF_FRAME records to global writer
2609
  _dumper_controller->lock_global_writer();
2610
  thread_dumper.dump_stack_traces(writer(), _klass_map);
2611
  _dumper_controller->unlock_global_writer();
2612

2613
  // write HPROF_GC_ROOT_THREAD_OBJ/HPROF_GC_ROOT_JAVA_FRAME/HPROF_GC_ROOT_JNI_LOCAL subrecord
2614
  // to segment writer
2615
  thread_dumper.dump_thread_obj(segment_writer);
2616
  thread_dumper.dump_stack_refs(segment_writer);
2617
}
2618

2619
// dump the heap to given path.
2620
int HeapDumper::dump(const char* path, outputStream* out, int compression, bool overwrite, uint num_dump_threads) {
2621
  assert(path != nullptr && strlen(path) > 0, "path missing");
2622

2623
  // print message in interactive case
2624
  if (out != nullptr) {
2625
    out->print_cr("Dumping heap to %s ...", path);
2626
    timer()->start();
2627
  }
2628

2629
  if (_oome && num_dump_threads > 1) {
2630
    // Each additional parallel writer requires several MB of internal memory
2631
    // (DumpWriter buffer, DumperClassCacheTable, GZipCompressor buffers).
2632
    // For the OOM handling we may already be limited in memory.
2633
    // Lets ensure we have at least 20MB per thread.
2634
    julong max_threads = os::free_memory() / (20 * M);
2635
    if (num_dump_threads > max_threads) {
2636
      num_dump_threads = MAX2<uint>(1, (uint)max_threads);
2637
    }
2638
  }
2639

2640
  // create JFR event
2641
  EventHeapDump event;
2642

2643
  AbstractCompressor* compressor = nullptr;
2644

2645
  if (compression > 0) {
2646
    compressor = new (std::nothrow) GZipCompressor(compression);
2647

2648
    if (compressor == nullptr) {
2649
      set_error("Could not allocate gzip compressor");
2650
      return -1;
2651
    }
2652
  }
2653

2654
  DumpWriter writer(path, overwrite, compressor);
2655

2656
  if (writer.error() != nullptr) {
2657
    set_error(writer.error());
2658
    if (out != nullptr) {
2659
      out->print_cr("Unable to create %s: %s", path,
2660
        (error() != nullptr) ? error() : "reason unknown");
2661
    }
2662
    return -1;
2663
  }
2664

2665
  // generate the segmented heap dump into separate files
2666
  VM_HeapDumper dumper(&writer, _gc_before_heap_dump, _oome, num_dump_threads);
2667
  VMThread::execute(&dumper);
2668

2669
  // record any error that the writer may have encountered
2670
  set_error(writer.error());
2671

2672
  // Heap dump process is done in two phases
2673
  //
2674
  // Phase 1: Concurrent threads directly write heap data to multiple heap files.
2675
  //          This is done by VM_HeapDumper, which is performed within safepoint.
2676
  //
2677
  // Phase 2: Merge multiple heap files into one complete heap dump file.
2678
  //          This is done by DumpMerger, which is performed outside safepoint
2679

2680
  DumpMerger merger(path, &writer, dumper.dump_seq());
2681
  // Perform heapdump file merge operation in the current thread prevents us
2682
  // from occupying the VM Thread, which in turn affects the occurrence of
2683
  // GC and other VM operations.
2684
  merger.do_merge();
2685
  if (writer.error() != nullptr) {
2686
    set_error(writer.error());
2687
  }
2688

2689
  // emit JFR event
2690
  if (error() == nullptr) {
2691
    event.set_destination(path);
2692
    event.set_gcBeforeDump(_gc_before_heap_dump);
2693
    event.set_size(writer.bytes_written());
2694
    event.set_onOutOfMemoryError(_oome);
2695
    event.set_overwrite(overwrite);
2696
    event.set_compression(compression);
2697
    event.commit();
2698
  } else {
2699
    log_debug(cds, heap)("Error %s while dumping heap", error());
2700
  }
2701

2702
  // print message in interactive case
2703
  if (out != nullptr) {
2704
    timer()->stop();
2705
    if (error() == nullptr) {
2706
      out->print_cr("Heap dump file created [" JULONG_FORMAT " bytes in %3.3f secs]",
2707
                    writer.bytes_written(), timer()->seconds());
2708
    } else {
2709
      out->print_cr("Dump file is incomplete: %s", writer.error());
2710
    }
2711
  }
2712

2713
  if (compressor != nullptr) {
2714
    delete compressor;
2715
  }
2716
  return (writer.error() == nullptr) ? 0 : -1;
2717
}
2718

2719
// stop timer (if still active), and free any error string we might be holding
2720
HeapDumper::~HeapDumper() {
2721
  if (timer()->is_active()) {
2722
    timer()->stop();
2723
  }
2724
  set_error(nullptr);
2725
}
2726

2727

2728
// returns the error string (resource allocated), or null
2729
char* HeapDumper::error_as_C_string() const {
2730
  if (error() != nullptr) {
2731
    char* str = NEW_RESOURCE_ARRAY(char, strlen(error())+1);
2732
    strcpy(str, error());
2733
    return str;
2734
  } else {
2735
    return nullptr;
2736
  }
2737
}
2738

2739
// set the error string
2740
void HeapDumper::set_error(char const* error) {
2741
  if (_error != nullptr) {
2742
    os::free(_error);
2743
  }
2744
  if (error == nullptr) {
2745
    _error = nullptr;
2746
  } else {
2747
    _error = os::strdup(error);
2748
    assert(_error != nullptr, "allocation failure");
2749
  }
2750
}
2751

2752
// Called by out-of-memory error reporting by a single Java thread
2753
// outside of a JVM safepoint
2754
void HeapDumper::dump_heap_from_oome() {
2755
  HeapDumper::dump_heap(true);
2756
}
2757

2758
// Called by error reporting by a single Java thread outside of a JVM safepoint,
2759
// or by heap dumping by the VM thread during a (GC) safepoint. Thus, these various
2760
// callers are strictly serialized and guaranteed not to interfere below. For more
2761
// general use, however, this method will need modification to prevent
2762
// inteference when updating the static variables base_path and dump_file_seq below.
2763
void HeapDumper::dump_heap() {
2764
  HeapDumper::dump_heap(false);
2765
}
2766

2767
void HeapDumper::dump_heap(bool oome) {
2768
  static char base_path[JVM_MAXPATHLEN] = {'\0'};
2769
  static uint dump_file_seq = 0;
2770
  char* my_path;
2771
  const int max_digit_chars = 20;
2772

2773
  const char* dump_file_name = "java_pid";
2774
  const char* dump_file_ext  = HeapDumpGzipLevel > 0 ? ".hprof.gz" : ".hprof";
2775

2776
  // The dump file defaults to java_pid<pid>.hprof in the current working
2777
  // directory. HeapDumpPath=<file> can be used to specify an alternative
2778
  // dump file name or a directory where dump file is created.
2779
  if (dump_file_seq == 0) { // first time in, we initialize base_path
2780
    // Calculate potentially longest base path and check if we have enough
2781
    // allocated statically.
2782
    const size_t total_length =
2783
                      (HeapDumpPath == nullptr ? 0 : strlen(HeapDumpPath)) +
2784
                      strlen(os::file_separator()) + max_digit_chars +
2785
                      strlen(dump_file_name) + strlen(dump_file_ext) + 1;
2786
    if (total_length > sizeof(base_path)) {
2787
      warning("Cannot create heap dump file.  HeapDumpPath is too long.");
2788
      return;
2789
    }
2790

2791
    bool use_default_filename = true;
2792
    if (HeapDumpPath == nullptr || HeapDumpPath[0] == '\0') {
2793
      // HeapDumpPath=<file> not specified
2794
    } else {
2795
      strcpy(base_path, HeapDumpPath);
2796
      // check if the path is a directory (must exist)
2797
      DIR* dir = os::opendir(base_path);
2798
      if (dir == nullptr) {
2799
        use_default_filename = false;
2800
      } else {
2801
        // HeapDumpPath specified a directory. We append a file separator
2802
        // (if needed).
2803
        os::closedir(dir);
2804
        size_t fs_len = strlen(os::file_separator());
2805
        if (strlen(base_path) >= fs_len) {
2806
          char* end = base_path;
2807
          end += (strlen(base_path) - fs_len);
2808
          if (strcmp(end, os::file_separator()) != 0) {
2809
            strcat(base_path, os::file_separator());
2810
          }
2811
        }
2812
      }
2813
    }
2814
    // If HeapDumpPath wasn't a file name then we append the default name
2815
    if (use_default_filename) {
2816
      const size_t dlen = strlen(base_path);  // if heap dump dir specified
2817
      jio_snprintf(&base_path[dlen], sizeof(base_path)-dlen, "%s%d%s",
2818
                   dump_file_name, os::current_process_id(), dump_file_ext);
2819
    }
2820
    const size_t len = strlen(base_path) + 1;
2821
    my_path = (char*)os::malloc(len, mtInternal);
2822
    if (my_path == nullptr) {
2823
      warning("Cannot create heap dump file.  Out of system memory.");
2824
      return;
2825
    }
2826
    strncpy(my_path, base_path, len);
2827
  } else {
2828
    // Append a sequence number id for dumps following the first
2829
    const size_t len = strlen(base_path) + max_digit_chars + 2; // for '.' and \0
2830
    my_path = (char*)os::malloc(len, mtInternal);
2831
    if (my_path == nullptr) {
2832
      warning("Cannot create heap dump file.  Out of system memory.");
2833
      return;
2834
    }
2835
    jio_snprintf(my_path, len, "%s.%d", base_path, dump_file_seq);
2836
  }
2837
  dump_file_seq++;   // increment seq number for next time we dump
2838

2839
  HeapDumper dumper(false /* no GC before heap dump */,
2840
                    oome  /* pass along out-of-memory-error flag */);
2841
  dumper.dump(my_path, tty, HeapDumpGzipLevel);
2842
  os::free(my_path);
2843
}
2844

Использование cookies

Мы используем файлы cookie в соответствии с Политикой конфиденциальности и Политикой использования cookies.

Нажимая кнопку «Принимаю», Вы даете АО «СберТех» согласие на обработку Ваших персональных данных в целях совершенствования нашего веб-сайта и Сервиса GitVerse, а также повышения удобства их использования.

Запретить использование cookies Вы можете самостоятельно в настройках Вашего браузера.