vernier 0.6.0 → 0.7.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/Gemfile +1 -0
- data/README.md +20 -4
- data/examples/gvl_sleep.rb +62 -0
- data/examples/measure_overhead.rb +39 -0
- data/ext/vernier/vernier.cc +366 -147
- data/lib/vernier/collector.rb +12 -1
- data/lib/vernier/middleware.rb +32 -0
- data/lib/vernier/output/firefox.rb +26 -20
- data/lib/vernier/result.rb +12 -1
- data/lib/vernier/stack_table.rb +42 -0
- data/lib/vernier/version.rb +1 -1
- data/lib/vernier.rb +29 -16
- data/vernier.gemspec +1 -0
- metadata +21 -3
data/ext/vernier/vernier.cc
CHANGED
@@ -57,6 +57,7 @@ static VALUE rb_mVernier;
|
|
57
57
|
static VALUE rb_cVernierResult;
|
58
58
|
static VALUE rb_mVernierMarkerType;
|
59
59
|
static VALUE rb_cVernierCollector;
|
60
|
+
static VALUE rb_cStackTable;
|
60
61
|
|
61
62
|
static const char *gvl_event_name(rb_event_flag_t event) {
|
62
63
|
switch (event) {
|
@@ -197,6 +198,7 @@ std::ostream& operator<<(std::ostream& os, const TimeStamp& info) {
|
|
197
198
|
return os;
|
198
199
|
}
|
199
200
|
|
201
|
+
// TODO: Rename FuncInfo
|
200
202
|
struct FrameInfo {
|
201
203
|
static const char *label_cstr(VALUE frame) {
|
202
204
|
VALUE label = rb_profile_frame_full_label(frame);
|
@@ -239,10 +241,6 @@ bool operator==(const FrameInfo& lhs, const FrameInfo& rhs) noexcept {
|
|
239
241
|
struct Frame {
|
240
242
|
VALUE frame;
|
241
243
|
int line;
|
242
|
-
|
243
|
-
FrameInfo info() const {
|
244
|
-
return FrameInfo(frame);
|
245
|
-
}
|
246
244
|
};
|
247
245
|
|
248
246
|
bool operator==(const Frame& lhs, const Frame& rhs) noexcept {
|
@@ -266,7 +264,7 @@ namespace std {
|
|
266
264
|
|
267
265
|
// A basic semaphore built on sem_wait/sem_post
|
268
266
|
// post() is guaranteed to be async-signal-safe
|
269
|
-
class
|
267
|
+
class SignalSafeSemaphore {
|
270
268
|
#ifdef __APPLE__
|
271
269
|
dispatch_semaphore_t sem;
|
272
270
|
#else
|
@@ -275,7 +273,7 @@ class SamplerSemaphore {
|
|
275
273
|
|
276
274
|
public:
|
277
275
|
|
278
|
-
|
276
|
+
SignalSafeSemaphore(unsigned int value = 0) {
|
279
277
|
#ifdef __APPLE__
|
280
278
|
sem = dispatch_semaphore_create(value);
|
281
279
|
#else
|
@@ -283,7 +281,7 @@ class SamplerSemaphore {
|
|
283
281
|
#endif
|
284
282
|
};
|
285
283
|
|
286
|
-
~
|
284
|
+
~SignalSafeSemaphore() {
|
287
285
|
#ifdef __APPLE__
|
288
286
|
dispatch_release(sem);
|
289
287
|
#else
|
@@ -316,17 +314,25 @@ class SamplerSemaphore {
|
|
316
314
|
}
|
317
315
|
};
|
318
316
|
|
319
|
-
|
317
|
+
class RawSample {
|
318
|
+
public:
|
319
|
+
|
320
320
|
constexpr static int MAX_LEN = 2048;
|
321
|
+
|
322
|
+
private:
|
323
|
+
|
321
324
|
VALUE frames[MAX_LEN];
|
322
325
|
int lines[MAX_LEN];
|
323
326
|
int len;
|
327
|
+
int offset;
|
324
328
|
bool gc;
|
325
329
|
|
326
|
-
|
330
|
+
public:
|
331
|
+
|
332
|
+
RawSample() : len(0), gc(false), offset(0) { }
|
327
333
|
|
328
334
|
int size() const {
|
329
|
-
return len;
|
335
|
+
return len - offset;
|
330
336
|
}
|
331
337
|
|
332
338
|
Frame frame(int i) const {
|
@@ -336,7 +342,7 @@ struct RawSample {
|
|
336
342
|
return frame;
|
337
343
|
}
|
338
344
|
|
339
|
-
void sample() {
|
345
|
+
void sample(int offset = 0) {
|
340
346
|
clear();
|
341
347
|
|
342
348
|
if (!ruby_native_thread_p()) {
|
@@ -347,16 +353,18 @@ struct RawSample {
|
|
347
353
|
gc = true;
|
348
354
|
} else {
|
349
355
|
len = rb_profile_frames(0, MAX_LEN, frames, lines);
|
356
|
+
this->offset = std::min(offset, len);
|
350
357
|
}
|
351
358
|
}
|
352
359
|
|
353
360
|
void clear() {
|
354
361
|
len = 0;
|
362
|
+
offset = 0;
|
355
363
|
gc = false;
|
356
364
|
}
|
357
365
|
|
358
366
|
bool empty() const {
|
359
|
-
return len
|
367
|
+
return len <= offset;
|
360
368
|
}
|
361
369
|
};
|
362
370
|
|
@@ -366,7 +374,7 @@ struct RawSample {
|
|
366
374
|
struct LiveSample {
|
367
375
|
RawSample sample;
|
368
376
|
|
369
|
-
|
377
|
+
SignalSafeSemaphore sem_complete;
|
370
378
|
|
371
379
|
// Wait for a sample to be collected by the signal handler on another thread
|
372
380
|
void wait() {
|
@@ -393,41 +401,51 @@ struct LiveSample {
|
|
393
401
|
}
|
394
402
|
};
|
395
403
|
|
396
|
-
|
397
|
-
|
398
|
-
|
404
|
+
template <typename K>
|
405
|
+
class IndexMap {
|
406
|
+
public:
|
407
|
+
std::unordered_map<K, int> to_idx;
|
408
|
+
std::vector<K> list;
|
399
409
|
|
400
|
-
|
401
|
-
|
402
|
-
|
403
|
-
int idx = string_list.size();
|
404
|
-
string_list.push_back(str);
|
410
|
+
const K& operator[](int i) const noexcept {
|
411
|
+
return list[i];
|
412
|
+
}
|
405
413
|
|
406
|
-
|
407
|
-
|
414
|
+
size_t size() const noexcept {
|
415
|
+
return list.size();
|
408
416
|
}
|
409
417
|
|
410
|
-
|
411
|
-
|
418
|
+
int index(const K key) {
|
419
|
+
auto it = to_idx.find(key);
|
420
|
+
if (it == to_idx.end()) {
|
421
|
+
int idx = list.size();
|
422
|
+
list.push_back(key);
|
423
|
+
|
424
|
+
auto result = to_idx.insert({key, idx});
|
425
|
+
it = result.first;
|
426
|
+
}
|
427
|
+
|
428
|
+
return it->second;
|
429
|
+
}
|
430
|
+
|
431
|
+
void clear() {
|
432
|
+
list.clear();
|
433
|
+
to_idx.clear();
|
434
|
+
}
|
435
|
+
};
|
436
|
+
|
437
|
+
struct StackTable {
|
438
|
+
private:
|
412
439
|
|
413
440
|
struct FrameWithInfo {
|
414
441
|
Frame frame;
|
415
442
|
FrameInfo info;
|
416
443
|
};
|
417
444
|
|
418
|
-
|
419
|
-
|
420
|
-
|
421
|
-
|
422
|
-
auto it = frame_to_idx.find(frame);
|
423
|
-
if (it == frame_to_idx.end()) {
|
424
|
-
int idx = frame_list.size();
|
425
|
-
frame_list.push_back(frame);
|
426
|
-
auto result = frame_to_idx.insert({frame, idx});
|
427
|
-
it = result.first;
|
428
|
-
}
|
429
|
-
return it->second;
|
430
|
-
}
|
445
|
+
IndexMap<Frame> frame_map;
|
446
|
+
|
447
|
+
IndexMap<VALUE> func_map;
|
448
|
+
std::vector<FrameInfo> func_info_list;
|
431
449
|
|
432
450
|
struct StackNode {
|
433
451
|
std::unordered_map<Frame, int> children;
|
@@ -441,21 +459,13 @@ struct FrameList {
|
|
441
459
|
StackNode() : frame(Frame{0, 0}), index(-1), parent(-1) {}
|
442
460
|
};
|
443
461
|
|
462
|
+
// This mutex guards the StackNodes only. The rest of the maps and vectors
|
463
|
+
// should be guarded by the GVL
|
464
|
+
std::mutex stack_mutex;
|
465
|
+
|
444
466
|
StackNode root_stack_node;
|
445
467
|
vector<StackNode> stack_node_list;
|
446
|
-
|
447
|
-
int stack_index(const RawSample &stack) {
|
448
|
-
if (stack.empty()) {
|
449
|
-
throw std::runtime_error("VERNIER BUG: empty stack");
|
450
|
-
}
|
451
|
-
|
452
|
-
StackNode *node = &root_stack_node;
|
453
|
-
for (int i = 0; i < stack.size(); i++) {
|
454
|
-
Frame frame = stack.frame(i);
|
455
|
-
node = next_stack_node(node, frame);
|
456
|
-
}
|
457
|
-
return node->index;
|
458
|
-
}
|
468
|
+
int stack_node_list_finalized_idx = 0;
|
459
469
|
|
460
470
|
StackNode *next_stack_node(StackNode *node, Frame frame) {
|
461
471
|
auto search = node->children.find(frame);
|
@@ -475,83 +485,267 @@ struct FrameList {
|
|
475
485
|
}
|
476
486
|
}
|
477
487
|
|
488
|
+
public:
|
489
|
+
|
490
|
+
int stack_index(const RawSample &stack) {
|
491
|
+
if (stack.empty()) {
|
492
|
+
throw std::runtime_error("VERNIER BUG: empty stack");
|
493
|
+
}
|
494
|
+
|
495
|
+
const std::lock_guard<std::mutex> lock(stack_mutex);
|
496
|
+
|
497
|
+
StackNode *node = &root_stack_node;
|
498
|
+
for (int i = 0; i < stack.size(); i++) {
|
499
|
+
Frame frame = stack.frame(i);
|
500
|
+
node = next_stack_node(node, frame);
|
501
|
+
}
|
502
|
+
return node->index;
|
503
|
+
}
|
504
|
+
|
505
|
+
int stack_parent(int stack_idx) {
|
506
|
+
const std::lock_guard<std::mutex> lock(stack_mutex);
|
507
|
+
if (stack_idx < 0 || stack_idx >= stack_node_list.size()) {
|
508
|
+
return -1;
|
509
|
+
} else {
|
510
|
+
return stack_node_list[stack_idx].parent;
|
511
|
+
}
|
512
|
+
}
|
513
|
+
|
514
|
+
int stack_frame(int stack_idx) {
|
515
|
+
const std::lock_guard<std::mutex> lock(stack_mutex);
|
516
|
+
if (stack_idx < 0 || stack_idx >= stack_node_list.size()) {
|
517
|
+
return -1;
|
518
|
+
} else {
|
519
|
+
return frame_map.index(stack_node_list[stack_idx].frame);
|
520
|
+
}
|
521
|
+
}
|
522
|
+
|
478
523
|
// Converts Frames from stacks other tables. "Symbolicates" the frames
|
479
524
|
// which allocates.
|
480
525
|
void finalize() {
|
481
|
-
|
482
|
-
|
526
|
+
{
|
527
|
+
const std::lock_guard<std::mutex> lock(stack_mutex);
|
528
|
+
for (int i = stack_node_list_finalized_idx; i < stack_node_list.size(); i++) {
|
529
|
+
const auto &stack_node = stack_node_list[i];
|
530
|
+
frame_map.index(stack_node.frame);
|
531
|
+
func_map.index(stack_node.frame.frame);
|
532
|
+
stack_node_list_finalized_idx = i;
|
533
|
+
}
|
483
534
|
}
|
484
|
-
|
485
|
-
|
535
|
+
|
536
|
+
for (int i = func_info_list.size(); i < func_map.size(); i++) {
|
537
|
+
const auto &func = func_map[i];
|
538
|
+
// must not hold a mutex here
|
539
|
+
func_info_list.push_back(FrameInfo(func));
|
486
540
|
}
|
487
541
|
}
|
488
542
|
|
489
543
|
void mark_frames() {
|
544
|
+
const std::lock_guard<std::mutex> lock(stack_mutex);
|
545
|
+
|
490
546
|
for (auto stack_node: stack_node_list) {
|
491
547
|
rb_gc_mark(stack_node.frame.frame);
|
492
548
|
}
|
493
549
|
}
|
494
550
|
|
551
|
+
// FIXME: probably should remove
|
495
552
|
void clear() {
|
496
|
-
|
497
|
-
|
498
|
-
|
499
|
-
|
500
|
-
|
501
|
-
|
502
|
-
|
503
|
-
|
504
|
-
}
|
505
|
-
|
506
|
-
void write_result(VALUE result) {
|
507
|
-
FrameList &frame_list = *this;
|
508
|
-
|
509
|
-
VALUE stack_table = rb_hash_new();
|
510
|
-
rb_ivar_set(result, rb_intern("@stack_table"), stack_table);
|
511
|
-
VALUE stack_table_parent = rb_ary_new();
|
512
|
-
VALUE stack_table_frame = rb_ary_new();
|
513
|
-
rb_hash_aset(stack_table, sym("parent"), stack_table_parent);
|
514
|
-
rb_hash_aset(stack_table, sym("frame"), stack_table_frame);
|
515
|
-
for (const auto &stack : frame_list.stack_node_list) {
|
516
|
-
VALUE parent_val = stack.parent == -1 ? Qnil : INT2NUM(stack.parent);
|
517
|
-
rb_ary_push(stack_table_parent, parent_val);
|
518
|
-
rb_ary_push(stack_table_frame, INT2NUM(frame_list.frame_index(stack.frame)));
|
519
|
-
}
|
520
|
-
|
521
|
-
VALUE frame_table = rb_hash_new();
|
522
|
-
rb_ivar_set(result, rb_intern("@frame_table"), frame_table);
|
523
|
-
VALUE frame_table_func = rb_ary_new();
|
524
|
-
VALUE frame_table_line = rb_ary_new();
|
525
|
-
rb_hash_aset(frame_table, sym("func"), frame_table_func);
|
526
|
-
rb_hash_aset(frame_table, sym("line"), frame_table_line);
|
527
|
-
//for (const auto &frame : frame_list.frame_list) {
|
528
|
-
for (int i = 0; i < frame_list.frame_with_info_list.size(); i++) {
|
529
|
-
const auto &frame = frame_list.frame_with_info_list[i];
|
530
|
-
rb_ary_push(frame_table_func, INT2NUM(i));
|
531
|
-
rb_ary_push(frame_table_line, INT2NUM(frame.frame.line));
|
532
|
-
}
|
533
|
-
|
534
|
-
// TODO: dedup funcs before this step
|
535
|
-
VALUE func_table = rb_hash_new();
|
536
|
-
rb_ivar_set(result, rb_intern("@func_table"), func_table);
|
537
|
-
VALUE func_table_name = rb_ary_new();
|
538
|
-
VALUE func_table_filename = rb_ary_new();
|
539
|
-
VALUE func_table_first_line = rb_ary_new();
|
540
|
-
rb_hash_aset(func_table, sym("name"), func_table_name);
|
541
|
-
rb_hash_aset(func_table, sym("filename"), func_table_filename);
|
542
|
-
rb_hash_aset(func_table, sym("first_line"), func_table_first_line);
|
543
|
-
for (const auto &frame : frame_list.frame_with_info_list) {
|
544
|
-
const std::string label = frame.info.label;
|
545
|
-
const std::string filename = frame.info.file;
|
546
|
-
const int first_line = frame.info.first_lineno;
|
547
|
-
|
548
|
-
rb_ary_push(func_table_name, rb_str_new(label.c_str(), label.length()));
|
549
|
-
rb_ary_push(func_table_filename, rb_str_new(filename.c_str(), filename.length()));
|
550
|
-
rb_ary_push(func_table_first_line, INT2NUM(first_line));
|
553
|
+
frame_map.clear();
|
554
|
+
func_map.clear();
|
555
|
+
func_info_list.clear();
|
556
|
+
|
557
|
+
{
|
558
|
+
const std::lock_guard<std::mutex> lock(stack_mutex);
|
559
|
+
stack_node_list.clear();
|
560
|
+
root_stack_node.children.clear();
|
551
561
|
}
|
552
562
|
}
|
563
|
+
|
564
|
+
static VALUE stack_table_stack_count(VALUE self);
|
565
|
+
static VALUE stack_table_frame_count(VALUE self);
|
566
|
+
static VALUE stack_table_func_count(VALUE self);
|
567
|
+
|
568
|
+
static VALUE stack_table_frame_line_no(VALUE self, VALUE idxval);
|
569
|
+
static VALUE stack_table_frame_func_idx(VALUE self, VALUE idxval);
|
570
|
+
static VALUE stack_table_func_name(VALUE self, VALUE idxval);
|
571
|
+
static VALUE stack_table_func_filename(VALUE self, VALUE idxval);
|
572
|
+
static VALUE stack_table_func_first_lineno(VALUE self, VALUE idxval);
|
573
|
+
|
574
|
+
friend class SampleTranslator;
|
553
575
|
};
|
554
576
|
|
577
|
+
static void
|
578
|
+
stack_table_mark(void *data) {
|
579
|
+
StackTable *stack_table = static_cast<StackTable *>(data);
|
580
|
+
stack_table->mark_frames();
|
581
|
+
}
|
582
|
+
|
583
|
+
static void
|
584
|
+
stack_table_free(void *data) {
|
585
|
+
StackTable *stack_table = static_cast<StackTable *>(data);
|
586
|
+
delete stack_table;
|
587
|
+
}
|
588
|
+
|
589
|
+
static const rb_data_type_t rb_stack_table_type = {
|
590
|
+
.wrap_struct_name = "vernier/stack_table",
|
591
|
+
.function = {
|
592
|
+
//.dmemsize = rb_collector_memsize,
|
593
|
+
.dmark = stack_table_mark,
|
594
|
+
.dfree = stack_table_free,
|
595
|
+
},
|
596
|
+
};
|
597
|
+
|
598
|
+
static VALUE
|
599
|
+
stack_table_new(VALUE self) {
|
600
|
+
StackTable *stack_table = new StackTable();
|
601
|
+
VALUE obj = TypedData_Wrap_Struct(self, &rb_stack_table_type, stack_table);
|
602
|
+
return obj;
|
603
|
+
}
|
604
|
+
|
605
|
+
static StackTable *get_stack_table(VALUE obj) {
|
606
|
+
StackTable *stack_table;
|
607
|
+
TypedData_Get_Struct(obj, StackTable, &rb_stack_table_type, stack_table);
|
608
|
+
return stack_table;
|
609
|
+
}
|
610
|
+
|
611
|
+
static VALUE
|
612
|
+
stack_table_current_stack(int argc, VALUE *argv, VALUE self) {
|
613
|
+
int offset;
|
614
|
+
VALUE offset_v;
|
615
|
+
|
616
|
+
rb_scan_args(argc, argv, "01", &offset_v);
|
617
|
+
if (argc > 0) {
|
618
|
+
offset = NUM2INT(offset_v) + 1;
|
619
|
+
} else {
|
620
|
+
offset = 1;
|
621
|
+
}
|
622
|
+
|
623
|
+
StackTable *stack_table = get_stack_table(self);
|
624
|
+
RawSample stack;
|
625
|
+
stack.sample(offset);
|
626
|
+
int stack_index = stack_table->stack_index(stack);
|
627
|
+
return INT2NUM(stack_index);
|
628
|
+
}
|
629
|
+
|
630
|
+
static VALUE
|
631
|
+
stack_table_stack_parent_idx(VALUE self, VALUE idxval) {
|
632
|
+
StackTable *stack_table = get_stack_table(self);
|
633
|
+
int idx = NUM2INT(idxval);
|
634
|
+
int parent_idx = stack_table->stack_parent(idx);
|
635
|
+
if (parent_idx < 0) {
|
636
|
+
return Qnil;
|
637
|
+
} else {
|
638
|
+
return INT2NUM(parent_idx);
|
639
|
+
}
|
640
|
+
}
|
641
|
+
|
642
|
+
static VALUE
|
643
|
+
stack_table_stack_frame_idx(VALUE self, VALUE idxval) {
|
644
|
+
StackTable *stack_table = get_stack_table(self);
|
645
|
+
//stack_table->finalize();
|
646
|
+
int idx = NUM2INT(idxval);
|
647
|
+
int frame_idx = stack_table->stack_frame(idx);
|
648
|
+
return frame_idx < 0 ? Qnil : INT2NUM(frame_idx);
|
649
|
+
}
|
650
|
+
|
651
|
+
VALUE
|
652
|
+
StackTable::stack_table_stack_count(VALUE self) {
|
653
|
+
StackTable *stack_table = get_stack_table(self);
|
654
|
+
int count;
|
655
|
+
{
|
656
|
+
const std::lock_guard<std::mutex> lock(stack_table->stack_mutex);
|
657
|
+
count = stack_table->stack_node_list.size();
|
658
|
+
}
|
659
|
+
return INT2NUM(count);
|
660
|
+
}
|
661
|
+
|
662
|
+
VALUE
|
663
|
+
StackTable::stack_table_frame_count(VALUE self) {
|
664
|
+
StackTable *stack_table = get_stack_table(self);
|
665
|
+
stack_table->finalize();
|
666
|
+
int count = stack_table->frame_map.size();
|
667
|
+
return INT2NUM(count);
|
668
|
+
}
|
669
|
+
|
670
|
+
VALUE
|
671
|
+
StackTable::stack_table_func_count(VALUE self) {
|
672
|
+
StackTable *stack_table = get_stack_table(self);
|
673
|
+
stack_table->finalize();
|
674
|
+
int count = stack_table->func_map.size();
|
675
|
+
return INT2NUM(count);
|
676
|
+
}
|
677
|
+
|
678
|
+
VALUE
|
679
|
+
StackTable::stack_table_frame_line_no(VALUE self, VALUE idxval) {
|
680
|
+
StackTable *stack_table = get_stack_table(self);
|
681
|
+
stack_table->finalize();
|
682
|
+
int idx = NUM2INT(idxval);
|
683
|
+
if (idx < 0 || idx >= stack_table->frame_map.size()) {
|
684
|
+
return Qnil;
|
685
|
+
} else {
|
686
|
+
const auto &frame = stack_table->frame_map[idx];
|
687
|
+
return INT2NUM(frame.line);
|
688
|
+
}
|
689
|
+
}
|
690
|
+
|
691
|
+
VALUE
|
692
|
+
StackTable::stack_table_frame_func_idx(VALUE self, VALUE idxval) {
|
693
|
+
StackTable *stack_table = get_stack_table(self);
|
694
|
+
stack_table->finalize();
|
695
|
+
int idx = NUM2INT(idxval);
|
696
|
+
if (idx < 0 || idx >= stack_table->frame_map.size()) {
|
697
|
+
return Qnil;
|
698
|
+
} else {
|
699
|
+
const auto &frame = stack_table->frame_map[idx];
|
700
|
+
int func_idx = stack_table->func_map.index(frame.frame);
|
701
|
+
return INT2NUM(func_idx);
|
702
|
+
}
|
703
|
+
}
|
704
|
+
|
705
|
+
VALUE
|
706
|
+
StackTable::stack_table_func_name(VALUE self, VALUE idxval) {
|
707
|
+
StackTable *stack_table = get_stack_table(self);
|
708
|
+
stack_table->finalize();
|
709
|
+
int idx = NUM2INT(idxval);
|
710
|
+
auto &table = stack_table->func_info_list;
|
711
|
+
if (idx < 0 || idx >= table.size()) {
|
712
|
+
return Qnil;
|
713
|
+
} else {
|
714
|
+
const auto &func_info = table[idx];
|
715
|
+
const std::string &label = func_info.label;
|
716
|
+
return rb_interned_str(label.c_str(), label.length());
|
717
|
+
}
|
718
|
+
}
|
719
|
+
|
720
|
+
VALUE
|
721
|
+
StackTable::stack_table_func_filename(VALUE self, VALUE idxval) {
|
722
|
+
StackTable *stack_table = get_stack_table(self);
|
723
|
+
stack_table->finalize();
|
724
|
+
int idx = NUM2INT(idxval);
|
725
|
+
auto &table = stack_table->func_info_list;
|
726
|
+
if (idx < 0 || idx >= table.size()) {
|
727
|
+
return Qnil;
|
728
|
+
} else {
|
729
|
+
const auto &func_info = table[idx];
|
730
|
+
const std::string &filename = func_info.file;
|
731
|
+
return rb_interned_str(filename.c_str(), filename.length());
|
732
|
+
}
|
733
|
+
}
|
734
|
+
|
735
|
+
VALUE
|
736
|
+
StackTable::stack_table_func_first_lineno(VALUE self, VALUE idxval) {
|
737
|
+
StackTable *stack_table = get_stack_table(self);
|
738
|
+
stack_table->finalize();
|
739
|
+
int idx = NUM2INT(idxval);
|
740
|
+
auto &table = stack_table->func_info_list;
|
741
|
+
if (idx < 0 || idx >= table.size()) {
|
742
|
+
return Qnil;
|
743
|
+
} else {
|
744
|
+
const auto &func_info = table[idx];
|
745
|
+
return INT2NUM(func_info.first_lineno);
|
746
|
+
}
|
747
|
+
}
|
748
|
+
|
555
749
|
class SampleTranslator {
|
556
750
|
public:
|
557
751
|
int last_stack_index;
|
@@ -563,7 +757,7 @@ class SampleTranslator {
|
|
563
757
|
SampleTranslator() : len(0), last_stack_index(-1) {
|
564
758
|
}
|
565
759
|
|
566
|
-
int translate(
|
760
|
+
int translate(StackTable &frame_list, const RawSample &sample) {
|
567
761
|
int i = 0;
|
568
762
|
for (; i < len && i < sample.size(); i++) {
|
569
763
|
if (frames[i] != sample.frame(i)) {
|
@@ -571,7 +765,8 @@ class SampleTranslator {
|
|
571
765
|
}
|
572
766
|
}
|
573
767
|
|
574
|
-
|
768
|
+
const std::lock_guard<std::mutex> lock(frame_list.stack_mutex);
|
769
|
+
StackTable::StackNode *node = i == 0 ? &frame_list.root_stack_node : &frame_list.stack_node_list[frame_indexes[i - 1]];
|
575
770
|
|
576
771
|
for (; i < sample.size(); i++) {
|
577
772
|
Frame frame = sample.frame(i);
|
@@ -743,7 +938,6 @@ class SampleList {
|
|
743
938
|
|
744
939
|
std::vector<int> stacks;
|
745
940
|
std::vector<TimeStamp> timestamps;
|
746
|
-
std::vector<native_thread_id_t> threads;
|
747
941
|
std::vector<Category> categories;
|
748
942
|
std::vector<int> weights;
|
749
943
|
|
@@ -755,11 +949,10 @@ class SampleList {
|
|
755
949
|
return size() == 0;
|
756
950
|
}
|
757
951
|
|
758
|
-
void record_sample(int stack_index, TimeStamp time,
|
952
|
+
void record_sample(int stack_index, TimeStamp time, Category category) {
|
759
953
|
if (
|
760
954
|
!empty() &&
|
761
955
|
stacks.back() == stack_index &&
|
762
|
-
threads.back() == thread_id &&
|
763
956
|
categories.back() == category)
|
764
957
|
{
|
765
958
|
// We don't compare timestamps for de-duplication
|
@@ -767,7 +960,6 @@ class SampleList {
|
|
767
960
|
} else {
|
768
961
|
stacks.push_back(stack_index);
|
769
962
|
timestamps.push_back(time);
|
770
|
-
threads.push_back(thread_id);
|
771
963
|
categories.push_back(category);
|
772
964
|
weights.push_back(1);
|
773
965
|
}
|
@@ -841,7 +1033,7 @@ class Thread {
|
|
841
1033
|
}
|
842
1034
|
}
|
843
1035
|
|
844
|
-
void record_newobj(VALUE obj,
|
1036
|
+
void record_newobj(VALUE obj, StackTable &frame_list) {
|
845
1037
|
RawSample sample;
|
846
1038
|
sample.sample();
|
847
1039
|
|
@@ -931,12 +1123,12 @@ class Thread {
|
|
931
1123
|
|
932
1124
|
class ThreadTable {
|
933
1125
|
public:
|
934
|
-
|
1126
|
+
StackTable &frame_list;
|
935
1127
|
|
936
1128
|
std::vector<std::unique_ptr<Thread> > list;
|
937
1129
|
std::mutex mutex;
|
938
1130
|
|
939
|
-
ThreadTable(
|
1131
|
+
ThreadTable(StackTable &frame_list) : frame_list(frame_list) {
|
940
1132
|
}
|
941
1133
|
|
942
1134
|
void mark() {
|
@@ -1017,15 +1209,17 @@ class BaseCollector {
|
|
1017
1209
|
protected:
|
1018
1210
|
|
1019
1211
|
virtual void reset() {
|
1020
|
-
frame_list.clear();
|
1021
1212
|
}
|
1022
1213
|
|
1023
1214
|
public:
|
1024
1215
|
bool running = false;
|
1025
|
-
|
1216
|
+
StackTable *stack_table;
|
1217
|
+
VALUE stack_table_value;
|
1026
1218
|
|
1027
1219
|
TimeStamp started_at;
|
1028
1220
|
|
1221
|
+
BaseCollector(VALUE stack_table_value) : stack_table_value(stack_table_value), stack_table(get_stack_table(stack_table_value)) {
|
1222
|
+
}
|
1029
1223
|
virtual ~BaseCollector() {}
|
1030
1224
|
|
1031
1225
|
virtual bool start() {
|
@@ -1068,7 +1262,8 @@ class BaseCollector {
|
|
1068
1262
|
};
|
1069
1263
|
|
1070
1264
|
virtual void mark() {
|
1071
|
-
frame_list.mark_frames();
|
1265
|
+
//frame_list.mark_frames();
|
1266
|
+
rb_gc_mark(stack_table_value);
|
1072
1267
|
};
|
1073
1268
|
|
1074
1269
|
virtual VALUE get_markers() {
|
@@ -1082,16 +1277,15 @@ class CustomCollector : public BaseCollector {
|
|
1082
1277
|
void sample() {
|
1083
1278
|
RawSample sample;
|
1084
1279
|
sample.sample();
|
1085
|
-
int stack_index =
|
1280
|
+
int stack_index = stack_table->stack_index(sample);
|
1086
1281
|
|
1087
|
-
|
1088
|
-
samples.record_sample(stack_index, TimeStamp::Now(), thread_id, CATEGORY_NORMAL);
|
1282
|
+
samples.record_sample(stack_index, TimeStamp::Now(), CATEGORY_NORMAL);
|
1089
1283
|
}
|
1090
1284
|
|
1091
1285
|
VALUE stop() {
|
1092
1286
|
BaseCollector::stop();
|
1093
1287
|
|
1094
|
-
|
1288
|
+
stack_table->finalize();
|
1095
1289
|
|
1096
1290
|
VALUE result = build_collector_result();
|
1097
1291
|
|
@@ -1112,10 +1306,12 @@ class CustomCollector : public BaseCollector {
|
|
1112
1306
|
rb_hash_aset(threads, ULL2NUM(0), thread_hash);
|
1113
1307
|
rb_hash_aset(thread_hash, sym("tid"), ULL2NUM(0));
|
1114
1308
|
|
1115
|
-
frame_list.write_result(result);
|
1116
|
-
|
1117
1309
|
return result;
|
1118
1310
|
}
|
1311
|
+
|
1312
|
+
public:
|
1313
|
+
|
1314
|
+
CustomCollector(VALUE stack_table) : BaseCollector(stack_table) { }
|
1119
1315
|
};
|
1120
1316
|
|
1121
1317
|
class RetainedCollector : public BaseCollector {
|
@@ -1135,7 +1331,7 @@ class RetainedCollector : public BaseCollector {
|
|
1135
1331
|
// Ideally we'd allow empty samples to be represented
|
1136
1332
|
return;
|
1137
1333
|
}
|
1138
|
-
int stack_index =
|
1334
|
+
int stack_index = stack_table->stack_index(sample);
|
1139
1335
|
|
1140
1336
|
object_list.push_back(obj);
|
1141
1337
|
object_frames.emplace(obj, stack_index);
|
@@ -1165,6 +1361,8 @@ class RetainedCollector : public BaseCollector {
|
|
1165
1361
|
|
1166
1362
|
public:
|
1167
1363
|
|
1364
|
+
RetainedCollector(VALUE stack_table) : BaseCollector(stack_table) { }
|
1365
|
+
|
1168
1366
|
bool start() {
|
1169
1367
|
if (!BaseCollector::start()) {
|
1170
1368
|
return false;
|
@@ -1191,7 +1389,7 @@ class RetainedCollector : public BaseCollector {
|
|
1191
1389
|
rb_tracepoint_disable(tp_newobj);
|
1192
1390
|
tp_newobj = Qnil;
|
1193
1391
|
|
1194
|
-
|
1392
|
+
stack_table->finalize();
|
1195
1393
|
|
1196
1394
|
// We should have collected info for all our frames, so no need to continue
|
1197
1395
|
// marking them
|
@@ -1213,7 +1411,7 @@ class RetainedCollector : public BaseCollector {
|
|
1213
1411
|
|
1214
1412
|
VALUE build_collector_result() {
|
1215
1413
|
RetainedCollector *collector = this;
|
1216
|
-
|
1414
|
+
StackTable &frame_list = *collector->stack_table;
|
1217
1415
|
|
1218
1416
|
VALUE result = BaseCollector::build_collector_result();
|
1219
1417
|
|
@@ -1241,8 +1439,6 @@ class RetainedCollector : public BaseCollector {
|
|
1241
1439
|
}
|
1242
1440
|
}
|
1243
1441
|
|
1244
|
-
frame_list.write_result(result);
|
1245
|
-
|
1246
1442
|
return result;
|
1247
1443
|
}
|
1248
1444
|
|
@@ -1251,7 +1447,8 @@ class RetainedCollector : public BaseCollector {
|
|
1251
1447
|
// can be garbage collected.
|
1252
1448
|
// When we stop collection we will stringify the remaining frames, and then
|
1253
1449
|
// clear them from the set, allowing them to be removed from out output.
|
1254
|
-
|
1450
|
+
stack_table->mark_frames();
|
1451
|
+
rb_gc_mark(stack_table_value);
|
1255
1452
|
|
1256
1453
|
rb_gc_mark(tp_newobj);
|
1257
1454
|
rb_gc_mark(tp_freeobj);
|
@@ -1333,7 +1530,7 @@ class TimeCollector : public BaseCollector {
|
|
1333
1530
|
pthread_t sample_thread;
|
1334
1531
|
|
1335
1532
|
atomic_bool running;
|
1336
|
-
|
1533
|
+
SignalSafeSemaphore thread_stopped;
|
1337
1534
|
|
1338
1535
|
TimeStamp interval;
|
1339
1536
|
unsigned int allocation_sample_rate;
|
@@ -1350,7 +1547,7 @@ class TimeCollector : public BaseCollector {
|
|
1350
1547
|
}
|
1351
1548
|
|
1352
1549
|
public:
|
1353
|
-
TimeCollector(TimeStamp interval, unsigned int allocation_sample_rate) : interval(interval), allocation_sample_rate(allocation_sample_rate), threads(
|
1550
|
+
TimeCollector(VALUE stack_table, TimeStamp interval, unsigned int allocation_sample_rate) : BaseCollector(stack_table), interval(interval), allocation_sample_rate(allocation_sample_rate), threads(*get_stack_table(stack_table)) {
|
1354
1551
|
}
|
1355
1552
|
|
1356
1553
|
void record_newobj(VALUE obj) {
|
@@ -1376,11 +1573,10 @@ class TimeCollector : public BaseCollector {
|
|
1376
1573
|
|
1377
1574
|
void record_sample(const RawSample &sample, TimeStamp time, Thread &thread, Category category) {
|
1378
1575
|
if (!sample.empty()) {
|
1379
|
-
int stack_index = thread.translator.translate(
|
1576
|
+
int stack_index = thread.translator.translate(*stack_table, sample);
|
1380
1577
|
thread.samples.record_sample(
|
1381
1578
|
stack_index,
|
1382
1579
|
time,
|
1383
|
-
thread.native_tid,
|
1384
1580
|
category
|
1385
1581
|
);
|
1386
1582
|
}
|
@@ -1430,7 +1626,7 @@ class TimeCollector : public BaseCollector {
|
|
1430
1626
|
// that by the GVL instrumentation, but let's try to get
|
1431
1627
|
// it to a consistent state and stop profiling it.
|
1432
1628
|
thread.set_state(Thread::State::STOPPED);
|
1433
|
-
} else if (sample.sample.
|
1629
|
+
} else if (sample.sample.empty()) {
|
1434
1630
|
// fprintf(stderr, "skipping GC sample\n");
|
1435
1631
|
} else {
|
1436
1632
|
record_sample(sample.sample, sample_start, thread, CATEGORY_NORMAL);
|
@@ -1439,7 +1635,6 @@ class TimeCollector : public BaseCollector {
|
|
1439
1635
|
thread.samples.record_sample(
|
1440
1636
|
thread.stack_on_suspend_idx,
|
1441
1637
|
sample_start,
|
1442
|
-
thread.native_tid,
|
1443
1638
|
CATEGORY_IDLE);
|
1444
1639
|
} else {
|
1445
1640
|
}
|
@@ -1602,7 +1797,7 @@ class TimeCollector : public BaseCollector {
|
|
1602
1797
|
rb_remove_event_hook(internal_gc_event_cb);
|
1603
1798
|
rb_remove_event_hook(internal_thread_event_cb);
|
1604
1799
|
|
1605
|
-
|
1800
|
+
stack_table->finalize();
|
1606
1801
|
|
1607
1802
|
VALUE result = build_collector_result();
|
1608
1803
|
|
@@ -1632,13 +1827,12 @@ class TimeCollector : public BaseCollector {
|
|
1632
1827
|
|
1633
1828
|
}
|
1634
1829
|
|
1635
|
-
frame_list.write_result(result);
|
1636
|
-
|
1637
1830
|
return result;
|
1638
1831
|
}
|
1639
1832
|
|
1640
1833
|
void mark() {
|
1641
|
-
|
1834
|
+
stack_table->mark_frames();
|
1835
|
+
rb_gc_mark(stack_table_value);
|
1642
1836
|
threads.mark();
|
1643
1837
|
|
1644
1838
|
//for (int i = 0; i < queued_length; i++) {
|
@@ -1710,12 +1904,21 @@ collector_sample(VALUE self) {
|
|
1710
1904
|
return Qtrue;
|
1711
1905
|
}
|
1712
1906
|
|
1907
|
+
static VALUE
|
1908
|
+
collector_stack_table(VALUE self) {
|
1909
|
+
auto *collector = get_collector(self);
|
1910
|
+
|
1911
|
+
return collector->stack_table_value;
|
1912
|
+
}
|
1913
|
+
|
1713
1914
|
static VALUE collector_new(VALUE self, VALUE mode, VALUE options) {
|
1714
1915
|
BaseCollector *collector;
|
1916
|
+
|
1917
|
+
VALUE stack_table = stack_table_new(rb_cStackTable);
|
1715
1918
|
if (mode == sym("retained")) {
|
1716
|
-
collector = new RetainedCollector();
|
1919
|
+
collector = new RetainedCollector(stack_table);
|
1717
1920
|
} else if (mode == sym("custom")) {
|
1718
|
-
collector = new CustomCollector();
|
1921
|
+
collector = new CustomCollector(stack_table);
|
1719
1922
|
} else if (mode == sym("wall")) {
|
1720
1923
|
VALUE intervalv = rb_hash_aref(options, sym("interval"));
|
1721
1924
|
TimeStamp interval;
|
@@ -1732,7 +1935,7 @@ static VALUE collector_new(VALUE self, VALUE mode, VALUE options) {
|
|
1732
1935
|
} else {
|
1733
1936
|
allocation_sample_rate = NUM2UINT(allocation_sample_ratev);
|
1734
1937
|
}
|
1735
|
-
collector = new TimeCollector(interval, allocation_sample_rate);
|
1938
|
+
collector = new TimeCollector(stack_table, interval, allocation_sample_rate);
|
1736
1939
|
} else {
|
1737
1940
|
rb_raise(rb_eArgError, "invalid mode");
|
1738
1941
|
}
|
@@ -1786,9 +1989,25 @@ Init_vernier(void)
|
|
1786
1989
|
rb_define_singleton_method(rb_cVernierCollector, "_new", collector_new, 2);
|
1787
1990
|
rb_define_method(rb_cVernierCollector, "start", collector_start, 0);
|
1788
1991
|
rb_define_method(rb_cVernierCollector, "sample", collector_sample, 0);
|
1992
|
+
rb_define_method(rb_cVernierCollector, "stack_table", collector_stack_table, 0);
|
1789
1993
|
rb_define_private_method(rb_cVernierCollector, "finish", collector_stop, 0);
|
1790
1994
|
rb_define_private_method(rb_cVernierCollector, "markers", markers, 0);
|
1791
1995
|
|
1996
|
+
rb_cStackTable = rb_define_class_under(rb_mVernier, "StackTable", rb_cObject);
|
1997
|
+
rb_undef_alloc_func(rb_cStackTable);
|
1998
|
+
rb_define_singleton_method(rb_cStackTable, "new", stack_table_new, 0);
|
1999
|
+
rb_define_method(rb_cStackTable, "current_stack", stack_table_current_stack, -1);
|
2000
|
+
rb_define_method(rb_cStackTable, "stack_parent_idx", stack_table_stack_parent_idx, 1);
|
2001
|
+
rb_define_method(rb_cStackTable, "stack_frame_idx", stack_table_stack_frame_idx, 1);
|
2002
|
+
rb_define_method(rb_cStackTable, "frame_line_no", StackTable::stack_table_frame_line_no, 1);
|
2003
|
+
rb_define_method(rb_cStackTable, "frame_func_idx", StackTable::stack_table_frame_func_idx, 1);
|
2004
|
+
rb_define_method(rb_cStackTable, "func_name", StackTable::stack_table_func_name, 1);
|
2005
|
+
rb_define_method(rb_cStackTable, "func_filename", StackTable::stack_table_func_filename, 1);
|
2006
|
+
rb_define_method(rb_cStackTable, "func_first_lineno", StackTable::stack_table_func_first_lineno, 1);
|
2007
|
+
rb_define_method(rb_cStackTable, "stack_count", StackTable::stack_table_stack_count, 0);
|
2008
|
+
rb_define_method(rb_cStackTable, "frame_count", StackTable::stack_table_frame_count, 0);
|
2009
|
+
rb_define_method(rb_cStackTable, "func_count", StackTable::stack_table_func_count, 0);
|
2010
|
+
|
1792
2011
|
Init_consts(rb_mVernierMarkerPhase);
|
1793
2012
|
|
1794
2013
|
//static VALUE gc_hook = Data_Wrap_Struct(rb_cObject, collector_mark, NULL, &_collector);
|