stackprof 0.2.15 → 0.2.27

data/ext/stackprof/stackprof.c

@@ -7,15 +7,20 @@
 **********************************************************************/
 
 #include <ruby/ruby.h>
+#include <ruby/version.h>
 #include <ruby/debug.h>
 #include <ruby/st.h>
 #include <ruby/io.h>
 #include <ruby/intern.h>
+#include <ruby/vm.h>
 #include <signal.h>
 #include <sys/time.h>
+#include <time.h>
 #include <pthread.h>
 
 #define BUF_SIZE 2048
+#define MICROSECONDS_IN_SECOND 1000000
+#define NANOSECONDS_IN_SECOND 1000000000
 
 #define FAKE_FRAME_GC    INT2FIX(0)
 #define FAKE_FRAME_MARK  INT2FIX(1)
@@ -27,8 +32,52 @@ static const char *fake_frame_cstrs[] = {
 	"(sweeping)",
 };
 
+static int stackprof_use_postponed_job = 1;
+static int ruby_vm_running = 0;
+
 #define TOTAL_FAKE_FRAMES (sizeof(fake_frame_cstrs) / sizeof(char *))
 
+#ifdef _POSIX_MONOTONIC_CLOCK
+  #define timestamp_t timespec
+  typedef struct timestamp_t timestamp_t;
+
+  static void capture_timestamp(timestamp_t *ts) {
+      clock_gettime(CLOCK_MONOTONIC, ts);
+  }
+
+  static int64_t delta_usec(timestamp_t *start, timestamp_t *end) {
+      int64_t result = MICROSECONDS_IN_SECOND * (end->tv_sec - start->tv_sec);
+      if (end->tv_nsec < start->tv_nsec) {
+	  result -= MICROSECONDS_IN_SECOND;
+	  result += (NANOSECONDS_IN_SECOND + end->tv_nsec - start->tv_nsec) / 1000;
+      } else {
+	  result += (end->tv_nsec - start->tv_nsec) / 1000;
+      }
+      return result;
+  }
+
+  static uint64_t timestamp_usec(timestamp_t *ts) {
+      return (MICROSECONDS_IN_SECOND * ts->tv_sec) + (ts->tv_nsec / 1000);
+  }
+#else
+  #define timestamp_t timeval
+  typedef struct timestamp_t timestamp_t;
+
+  static void capture_timestamp(timestamp_t *ts) {
+      gettimeofday(ts, NULL);
+  }
+
+  static int64_t delta_usec(timestamp_t *start, timestamp_t *end) {
+      struct timeval diff;
+      timersub(end, start, &diff);
+      return (MICROSECONDS_IN_SECOND * diff.tv_sec) + diff.tv_usec;
+  }
+
+  static uint64_t timestamp_usec(timestamp_t *ts) {
+      return (MICROSECONDS_IN_SECOND * ts.tv_sec) + diff.tv_usec
+  }
+#endif
+
 typedef struct {
     size_t total_samples;
     size_t caller_samples;
@@ -37,7 +86,24 @@ typedef struct {
     st_table *lines;
 } frame_data_t;
 
+typedef struct {
+    uint64_t timestamp_usec;
+    int64_t delta_usec;
+} sample_time_t;
+
+/* We need to ensure that various memory operations are visible across
+ * threads.  Ruby doesn't offer a portable way to do this sort of detection
+ * across all the Ruby versions we support, so we use something that casts a
+ * wide net (Clang, along with ICC, defines __GNUC__). */
+#if defined(__GNUC__) && defined(__ATOMIC_SEQ_CST)
+#define STACKPROF_HAVE_ATOMICS 1
+#else
+#define STACKPROF_HAVE_ATOMICS 0
+#endif
+
 static struct {
+    /* Access this field with the `STACKPROF_RUNNING` macro, below, since we
+     * can't properly express that this field has an atomic type. */
     int running;
     int raw;
     int aggregate;
@@ -46,16 +112,17 @@ static struct {
     VALUE interval;
     VALUE out;
     VALUE metadata;
+    int ignore_gc;
 
-    VALUE *raw_samples;
+    uint64_t *raw_samples;
     size_t raw_samples_len;
     size_t raw_samples_capa;
     size_t raw_sample_index;
 
-    struct timeval last_sample_at;
-    int *raw_timestamp_deltas;
-    size_t raw_timestamp_deltas_len;
-    size_t raw_timestamp_deltas_capa;
+    struct timestamp_t last_sample_at;
+    sample_time_t *raw_sample_times;
+    size_t raw_sample_times_len;
+    size_t raw_sample_times_capa;
 
     size_t overall_signals;
     size_t overall_samples;
@@ -65,16 +132,29 @@ static struct {
     size_t unrecorded_gc_sweeping_samples;
     st_table *frames;
 
+    timestamp_t gc_start_timestamp;
+
     VALUE fake_frame_names[TOTAL_FAKE_FRAMES];
     VALUE empty_string;
+
+    int buffer_count;
+    sample_time_t buffer_time;
     VALUE frames_buffer[BUF_SIZE];
     int lines_buffer[BUF_SIZE];
+
+    pthread_t target_thread;
 } _stackprof;
 
+#if STACKPROF_HAVE_ATOMICS
+#define STACKPROF_RUNNING() __atomic_load_n(&_stackprof.running, __ATOMIC_ACQUIRE)
+#else
+#define STACKPROF_RUNNING() _stackprof.running
+#endif
+
 static VALUE sym_object, sym_wall, sym_cpu, sym_custom, sym_name, sym_file, sym_line;
 static VALUE sym_samples, sym_total_samples, sym_missed_samples, sym_edges, sym_lines;
-static VALUE sym_version, sym_mode, sym_interval, sym_raw, sym_metadata, sym_frames, sym_out, sym_aggregate, sym_raw_timestamp_deltas;
-static VALUE sym_state, sym_marking, sym_sweeping;
+static VALUE sym_version, sym_mode, sym_interval, sym_raw, sym_raw_lines, sym_metadata, sym_frames, sym_ignore_gc, sym_out;
+static VALUE sym_aggregate, sym_raw_sample_timestamps, sym_raw_timestamp_deltas, sym_state, sym_marking, sym_sweeping;
 static VALUE sym_gc_samples, objtracer;
 static VALUE gc_hook;
 static VALUE rb_mStackProf;
@@ -88,9 +168,11 @@ stackprof_start(int argc, VALUE *argv, VALUE self)
     struct sigaction sa;
     struct itimerval timer;
     VALUE opts = Qnil, mode = Qnil, interval = Qnil, metadata = rb_hash_new(), out = Qfalse;
+    int ignore_gc = 0;
     int raw = 0, aggregate = 1;
+    VALUE metadata_val;
 
-    if (_stackprof.running)
+    if (STACKPROF_RUNNING())
 	return Qfalse;
 
     rb_scan_args(argc, argv, "0:", &opts);
@@ -99,8 +181,11 @@ stackprof_start(int argc, VALUE *argv, VALUE self)
 	mode = rb_hash_aref(opts, sym_mode);
 	interval = rb_hash_aref(opts, sym_interval);
 	out = rb_hash_aref(opts, sym_out);
+	if (RTEST(rb_hash_aref(opts, sym_ignore_gc))) {
+	    ignore_gc = 1;
+	}
 
-	VALUE metadata_val = rb_hash_aref(opts, sym_metadata);
+	metadata_val = rb_hash_aref(opts, sym_metadata);
 	if (RTEST(metadata_val)) {
 	    if (!RB_TYPE_P(metadata_val, T_HASH))
 		rb_raise(rb_eArgError, "metadata should be a hash");
@@ -115,6 +200,10 @@ stackprof_start(int argc, VALUE *argv, VALUE self)
     }
     if (!RTEST(mode)) mode = sym_wall;
 
+    if (!NIL_P(interval) && (NUM2INT(interval) < 1 || NUM2INT(interval) >= MICROSECONDS_IN_SECOND)) {
+        rb_raise(rb_eArgError, "interval is a number of microseconds between 1 and 1 million");
+    }
+
     if (!_stackprof.frames) {
 	_stackprof.frames = st_init_numtable();
 	_stackprof.overall_signals = 0;
@@ -146,16 +235,24 @@ stackprof_start(int argc, VALUE *argv, VALUE self)
 	rb_raise(rb_eArgError, "unknown profiler mode");
     }
 
-    _stackprof.running = 1;
     _stackprof.raw = raw;
     _stackprof.aggregate = aggregate;
     _stackprof.mode = mode;
     _stackprof.interval = interval;
+    _stackprof.ignore_gc = ignore_gc;
     _stackprof.metadata = metadata;
     _stackprof.out = out;
+    _stackprof.target_thread = pthread_self();
+    /* We need to ensure previous initialization stores are visible across
+     * threads. */
+#if STACKPROF_HAVE_ATOMICS
+    __atomic_store_n(&_stackprof.running, 1, __ATOMIC_SEQ_CST);
+#else
+    _stackprof.running = 1;
+#endif
 
     if (raw) {
-	gettimeofday(&_stackprof.last_sample_at, NULL);
+	capture_timestamp(&_stackprof.last_sample_at);
     }
 
     return Qtrue;
@@ -167,9 +264,15 @@ stackprof_stop(VALUE self)
     struct sigaction sa;
     struct itimerval timer;
 
+#if STACKPROF_HAVE_ATOMICS
+    int was_running = __atomic_exchange_n(&_stackprof.running, 0, __ATOMIC_SEQ_CST);
+    if (!was_running)
+        return Qfalse;
+#else
     if (!_stackprof.running)
 	return Qfalse;
     _stackprof.running = 0;
+#endif
 
     if (_stackprof.mode == sym_object) {
 	rb_tracepoint_disable(objtracer);
@@ -190,13 +293,19 @@ stackprof_stop(VALUE self)
     return Qtrue;
 }
 
+#if SIZEOF_VOIDP == SIZEOF_LONG
+#  define PTR2NUM(x) (LONG2NUM((long)(x)))
+#else
+#  define PTR2NUM(x) (LL2NUM((LONG_LONG)(x)))
+#endif
+
 static int
 frame_edges_i(st_data_t key, st_data_t val, st_data_t arg)
 {
     VALUE edges = (VALUE)arg;
 
     intptr_t weight = (intptr_t)val;
-    rb_hash_aset(edges, rb_obj_id((VALUE)key), INT2FIX(weight));
+    rb_hash_aset(edges, PTR2NUM(key), INT2FIX(weight));
     return ST_CONTINUE;
 }
 
@@ -223,7 +332,7 @@ frame_i(st_data_t key, st_data_t val, st_data_t arg)
     VALUE name, file, edges, lines;
     VALUE line;
 
-    rb_hash_aset(results, rb_obj_id(frame), details);
+    rb_hash_aset(results, PTR2NUM(frame), details);
 
     if (FIXNUM_P(frame)) {
 	name = _stackprof.fake_frame_names[FIX2INT(frame)];
@@ -272,7 +381,7 @@ stackprof_results(int argc, VALUE *argv, VALUE self)
 {
     VALUE results, frames;
 
-    if (!_stackprof.frames || _stackprof.running)
+    if (!_stackprof.frames || STACKPROF_RUNNING())
 	return Qnil;
 
     results = rb_hash_new();
@@ -295,16 +404,25 @@ stackprof_results(int argc, VALUE *argv, VALUE self)
 
     if (_stackprof.raw && _stackprof.raw_samples_len) {
 	size_t len, n, o;
-	VALUE raw_timestamp_deltas;
+	VALUE raw_sample_timestamps, raw_timestamp_deltas;
 	VALUE raw_samples = rb_ary_new_capa(_stackprof.raw_samples_len);
+	VALUE raw_lines = rb_ary_new_capa(_stackprof.raw_samples_len);
 
 	for (n = 0; n < _stackprof.raw_samples_len; n++) {
 	    len = (size_t)_stackprof.raw_samples[n];
 	    rb_ary_push(raw_samples, SIZET2NUM(len));
+	    rb_ary_push(raw_lines, SIZET2NUM(len));
+
+	    for (o = 0, n++; o < len; n++, o++) {
+		// Line is in the upper 16 bits
+		rb_ary_push(raw_lines, INT2NUM(_stackprof.raw_samples[n] >> 48));
+
+		VALUE frame = _stackprof.raw_samples[n] & ~((uint64_t)0xFFFF << 48);
+		rb_ary_push(raw_samples, PTR2NUM(frame));
+	    }
 
-	    for (o = 0, n++; o < len; n++, o++)
-		rb_ary_push(raw_samples, rb_obj_id(_stackprof.raw_samples[n]));
 	    rb_ary_push(raw_samples, SIZET2NUM((size_t)_stackprof.raw_samples[n]));
+	    rb_ary_push(raw_lines, SIZET2NUM((size_t)_stackprof.raw_samples[n]));
 	}
 
 	free(_stackprof.raw_samples);
@@ -314,18 +432,22 @@ stackprof_results(int argc, VALUE *argv, VALUE self)
 	_stackprof.raw_sample_index = 0;
 
 	rb_hash_aset(results, sym_raw, raw_samples);
+	rb_hash_aset(results, sym_raw_lines, raw_lines);
 
-	raw_timestamp_deltas = rb_ary_new_capa(_stackprof.raw_timestamp_deltas_len);
+	raw_sample_timestamps = rb_ary_new_capa(_stackprof.raw_sample_times_len);
+	raw_timestamp_deltas = rb_ary_new_capa(_stackprof.raw_sample_times_len);
 
-	for (n = 0; n < _stackprof.raw_timestamp_deltas_len; n++) {
-	    rb_ary_push(raw_timestamp_deltas, INT2FIX(_stackprof.raw_timestamp_deltas[n]));
+	for (n = 0; n < _stackprof.raw_sample_times_len; n++) {
+	    rb_ary_push(raw_sample_timestamps, ULL2NUM(_stackprof.raw_sample_times[n].timestamp_usec));
+	    rb_ary_push(raw_timestamp_deltas, LL2NUM(_stackprof.raw_sample_times[n].delta_usec));
 	}
 
-	free(_stackprof.raw_timestamp_deltas);
-	_stackprof.raw_timestamp_deltas = NULL;
-	_stackprof.raw_timestamp_deltas_len = 0;
-	_stackprof.raw_timestamp_deltas_capa = 0;
+	free(_stackprof.raw_sample_times);
+	_stackprof.raw_sample_times = NULL;
+	_stackprof.raw_sample_times_len = 0;
+	_stackprof.raw_sample_times_capa = 0;
 
+	rb_hash_aset(results, sym_raw_sample_timestamps, raw_sample_timestamps);
 	rb_hash_aset(results, sym_raw_timestamp_deltas, raw_timestamp_deltas);
 
 	_stackprof.raw = 0;
@@ -363,7 +485,7 @@ stackprof_run(int argc, VALUE *argv, VALUE self)
 static VALUE
 stackprof_running_p(VALUE self)
 {
-    return _stackprof.running ? Qtrue : Qfalse;
+    return STACKPROF_RUNNING() ? Qtrue : Qfalse;
 }
 
 static inline frame_data_t *
@@ -405,14 +527,14 @@ st_numtable_increment(st_table *table, st_data_t key, size_t increment)
 }
 
 void
-stackprof_record_sample_for_stack(int num, int timestamp_delta)
+stackprof_record_sample_for_stack(int num, uint64_t sample_timestamp, int64_t timestamp_delta)
 {
     int i, n;
     VALUE prev_frame = Qnil;
 
     _stackprof.overall_samples++;
 
-    if (_stackprof.raw) {
+    if (_stackprof.raw && num > 0) {
 	int found = 0;
 
 	/* If there's no sample buffer allocated, then allocate one.  The buffer
@@ -440,7 +562,12 @@ stackprof_record_sample_for_stack(int num, int timestamp_delta)
 	     * in the frames buffer that came from Ruby. */
 	    for (i = num-1, n = 0; i >= 0; i--, n++) {
 		VALUE frame = _stackprof.frames_buffer[i];
-		if (_stackprof.raw_samples[_stackprof.raw_sample_index + 1 + n] != frame)
+		int line = _stackprof.lines_buffer[i];
+
+		// Encode the line in to the upper 16 bits.
+		uint64_t key = ((uint64_t)line << 48) | (uint64_t)frame;
+
+		if (_stackprof.raw_samples[_stackprof.raw_sample_index + 1 + n] != key)
 		    break;
 	    }
 	    if (i == -1) {
@@ -458,26 +585,34 @@ stackprof_record_sample_for_stack(int num, int timestamp_delta)
 	    _stackprof.raw_samples[_stackprof.raw_samples_len++] = (VALUE)num;
 	    for (i = num-1; i >= 0; i--) {
 		VALUE frame = _stackprof.frames_buffer[i];
-		_stackprof.raw_samples[_stackprof.raw_samples_len++] = frame;
+		int line = _stackprof.lines_buffer[i];
+
+		// Encode the line in to the upper 16 bits.
+		uint64_t key = ((uint64_t)line << 48) | (uint64_t)frame;
+
+		_stackprof.raw_samples[_stackprof.raw_samples_len++] = key;
 	    }
 	    _stackprof.raw_samples[_stackprof.raw_samples_len++] = (VALUE)1;
 	}
 
 	/* If there's no timestamp delta buffer, allocate one */
-	if (!_stackprof.raw_timestamp_deltas) {
-	    _stackprof.raw_timestamp_deltas_capa = 100;
-	    _stackprof.raw_timestamp_deltas = malloc(sizeof(int) * _stackprof.raw_timestamp_deltas_capa);
-	    _stackprof.raw_timestamp_deltas_len = 0;
+	if (!_stackprof.raw_sample_times) {
+	    _stackprof.raw_sample_times_capa = 100;
+	    _stackprof.raw_sample_times = malloc(sizeof(sample_time_t) * _stackprof.raw_sample_times_capa);
+	    _stackprof.raw_sample_times_len = 0;
 	}
 
 	/* Double the buffer size if it's too small */
-	while (_stackprof.raw_timestamp_deltas_capa <= _stackprof.raw_timestamp_deltas_len + 1) {
-	    _stackprof.raw_timestamp_deltas_capa *= 2;
-	    _stackprof.raw_timestamp_deltas = realloc(_stackprof.raw_timestamp_deltas, sizeof(int) * _stackprof.raw_timestamp_deltas_capa);
+	while (_stackprof.raw_sample_times_capa <= _stackprof.raw_sample_times_len + 1) {
+	    _stackprof.raw_sample_times_capa *= 2;
+	    _stackprof.raw_sample_times = realloc(_stackprof.raw_sample_times, sizeof(sample_time_t) * _stackprof.raw_sample_times_capa);
 	}
 
-	/* Store the time delta (which is the amount of time between samples) */
-	_stackprof.raw_timestamp_deltas[_stackprof.raw_timestamp_deltas_len++] = timestamp_delta;
+	/* Store the time delta (which is the amount of microseconds between samples). */
+	_stackprof.raw_sample_times[_stackprof.raw_sample_times_len++] = (sample_time_t) {
+	    .timestamp_usec = sample_timestamp,
+	    .delta_usec = timestamp_delta,
+        };
     }
 
     for (i = 0; i < num; i++) {
@@ -510,48 +645,60 @@ stackprof_record_sample_for_stack(int num, int timestamp_delta)
     }
 
     if (_stackprof.raw) {
-	gettimeofday(&_stackprof.last_sample_at, NULL);
+	capture_timestamp(&_stackprof.last_sample_at);
     }
 }
 
+// buffer the current profile frames
+// This must be async-signal-safe
+// Returns immediately if another set of frames are already in the buffer
 void
-stackprof_record_sample()
+stackprof_buffer_sample(void)
 {
-    int timestamp_delta = 0;
+    uint64_t start_timestamp = 0;
+    int64_t timestamp_delta = 0;
     int num;
+
+    if (_stackprof.buffer_count > 0) {
+	// Another sample is already pending
+	return;
+    }
+
     if (_stackprof.raw) {
-	struct timeval t;
-	struct timeval diff;
-	gettimeofday(&t, NULL);
-	timersub(&t, &_stackprof.last_sample_at, &diff);
-	timestamp_delta = (1000 * diff.tv_sec) + diff.tv_usec;
+	struct timestamp_t t;
+	capture_timestamp(&t);
+	start_timestamp = timestamp_usec(&t);
+	timestamp_delta = delta_usec(&_stackprof.last_sample_at, &t);
     }
+
     num = rb_profile_frames(0, sizeof(_stackprof.frames_buffer) / sizeof(VALUE), _stackprof.frames_buffer, _stackprof.lines_buffer);
-    stackprof_record_sample_for_stack(num, timestamp_delta);
+
+    _stackprof.buffer_count = num;
+    _stackprof.buffer_time.timestamp_usec = start_timestamp;
+    _stackprof.buffer_time.delta_usec = timestamp_delta;
 }
 
+// Postponed job
 void
-stackprof_record_gc_samples()
+stackprof_record_gc_samples(void)
 {
-    int delta_to_first_unrecorded_gc_sample = 0;
-    int i;
+    int64_t delta_to_first_unrecorded_gc_sample = 0;
+    uint64_t start_timestamp = 0;
+    size_t i;
     if (_stackprof.raw) {
-	struct timeval t;
-	struct timeval diff;
-	gettimeofday(&t, NULL);
-	timersub(&t, &_stackprof.last_sample_at, &diff);
+	struct timestamp_t t = _stackprof.gc_start_timestamp;
+	start_timestamp = timestamp_usec(&t);
 
 	// We don't know when the GC samples were actually marked, so let's
 	// assume that they were marked at a perfectly regular interval.
-	delta_to_first_unrecorded_gc_sample = (1000 * diff.tv_sec + diff.tv_usec) - (_stackprof.unrecorded_gc_samples - 1) * NUM2LONG(_stackprof.interval);
+	delta_to_first_unrecorded_gc_sample = delta_usec(&_stackprof.last_sample_at, &t) - (_stackprof.unrecorded_gc_samples - 1) * NUM2LONG(_stackprof.interval);
 	if (delta_to_first_unrecorded_gc_sample < 0) {
 	    delta_to_first_unrecorded_gc_sample = 0;
 	}
     }
 
-
     for (i = 0; i < _stackprof.unrecorded_gc_samples; i++) {
-	int timestamp_delta = i == 0 ? delta_to_first_unrecorded_gc_sample : NUM2LONG(_stackprof.interval);
+	int64_t timestamp_delta = i == 0 ? delta_to_first_unrecorded_gc_sample : NUM2LONG(_stackprof.interval);
 
       if (_stackprof.unrecorded_gc_marking_samples) {
         _stackprof.frames_buffer[0] = FAKE_FRAME_MARK;
@@ -560,7 +707,7 @@ stackprof_record_gc_samples()
         _stackprof.lines_buffer[1] = 0;
         _stackprof.unrecorded_gc_marking_samples--;
 
-        stackprof_record_sample_for_stack(2, timestamp_delta);
+        stackprof_record_sample_for_stack(2, start_timestamp, timestamp_delta);
       } else if (_stackprof.unrecorded_gc_sweeping_samples) {
         _stackprof.frames_buffer[0] = FAKE_FRAME_SWEEP;
         _stackprof.lines_buffer[0] = 0;
@@ -569,11 +716,11 @@ stackprof_record_gc_samples()
 
         _stackprof.unrecorded_gc_sweeping_samples--;
 
-        stackprof_record_sample_for_stack(2, timestamp_delta);
+        stackprof_record_sample_for_stack(2, start_timestamp, timestamp_delta);
       } else {
         _stackprof.frames_buffer[0] = FAKE_FRAME_GC;
         _stackprof.lines_buffer[0] = 0;
-        stackprof_record_sample_for_stack(1, timestamp_delta);
+        stackprof_record_sample_for_stack(1, start_timestamp, timestamp_delta);
       }
     }
     _stackprof.during_gc += _stackprof.unrecorded_gc_samples;
@@ -582,46 +729,103 @@ stackprof_record_gc_samples()
     _stackprof.unrecorded_gc_sweeping_samples = 0;
 }
 
+// record the sample previously buffered by stackprof_buffer_sample
+static void
+stackprof_record_buffer(void)
+{
+    stackprof_record_sample_for_stack(_stackprof.buffer_count, _stackprof.buffer_time.timestamp_usec, _stackprof.buffer_time.delta_usec);
+
+    // reset the buffer
+    _stackprof.buffer_count = 0;
+}
+
+static void
+stackprof_sample_and_record(void)
+{
+    stackprof_buffer_sample();
+    stackprof_record_buffer();
+}
+
 static void
-stackprof_gc_job_handler(void *data)
+stackprof_job_record_gc(void *data)
 {
-    static int in_signal_handler = 0;
-    if (in_signal_handler) return;
-    if (!_stackprof.running) return;
+    if (!STACKPROF_RUNNING()) return;
 
-    in_signal_handler++;
     stackprof_record_gc_samples();
-    in_signal_handler--;
 }
 
 static void
-stackprof_job_handler(void *data)
+stackprof_job_sample_and_record(void *data)
 {
-    static int in_signal_handler = 0;
-    if (in_signal_handler) return;
-    if (!_stackprof.running) return;
+    if (!STACKPROF_RUNNING()) return;
 
-    in_signal_handler++;
-    stackprof_record_sample();
-    in_signal_handler--;
+    stackprof_sample_and_record();
+}
+
+static void
+stackprof_job_record_buffer(void *data)
+{
+    if (!STACKPROF_RUNNING()) return;
+
+    stackprof_record_buffer();
 }
 
 static void
 stackprof_signal_handler(int sig, siginfo_t *sinfo, void *ucontext)
 {
+    static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
+
     _stackprof.overall_signals++;
-    if (rb_during_gc()) {
+
+    if (!STACKPROF_RUNNING()) return;
+
+    // There's a possibility that the signal handler is invoked *after* the Ruby
+    // VM has been shut down (e.g. after ruby_cleanup(0)). In this case, things
+    // that rely on global VM state (e.g. rb_during_gc) will segfault.
+    if (!ruby_vm_running) return;
+
+    if (_stackprof.mode == sym_wall) {
+        // In "wall" mode, the SIGALRM signal will arrive at an arbitrary thread.
+        // In order to provide more useful results, especially under threaded web
+        // servers, we want to forward this signal to the original thread
+        // StackProf was started from.
+        // According to POSIX.1-2008 TC1 pthread_kill and pthread_self should be
+        // async-signal-safe.
+        if (pthread_self() != _stackprof.target_thread) {
+            pthread_kill(_stackprof.target_thread, sig);
+            return;
+        }
+    } else {
+        if (!ruby_native_thread_p()) return;
+    }
+
+    if (pthread_mutex_trylock(&lock)) return;
+
+    if (!_stackprof.ignore_gc && rb_during_gc()) {
 	VALUE mode = rb_gc_latest_gc_info(sym_state);
 	if (mode == sym_marking) {
 	    _stackprof.unrecorded_gc_marking_samples++;
 	} else if (mode == sym_sweeping) {
 	    _stackprof.unrecorded_gc_sweeping_samples++;
 	}
+	if(!_stackprof.unrecorded_gc_samples) {
+	    // record start
+	    capture_timestamp(&_stackprof.gc_start_timestamp);
+	}
 	_stackprof.unrecorded_gc_samples++;
-	rb_postponed_job_register_one(0, stackprof_gc_job_handler, (void*)0);
+	rb_postponed_job_register_one(0, stackprof_job_record_gc, (void*)0);
     } else {
-	rb_postponed_job_register_one(0, stackprof_job_handler, (void*)0);
+        if (stackprof_use_postponed_job) {
+            rb_postponed_job_register_one(0, stackprof_job_sample_and_record, (void*)0);
+        } else {
+            // Buffer a sample immediately, if an existing sample exists this will
+            // return immediately
+            stackprof_buffer_sample();
+            // Enqueue a job to record the sample
+            rb_postponed_job_register_one(0, stackprof_job_record_buffer, (void*)0);
+        }
     }
+    pthread_mutex_unlock(&lock);
 }
 
 static void
@@ -630,17 +834,17 @@ stackprof_newobj_handler(VALUE tpval, void *data)
     _stackprof.overall_signals++;
     if (RTEST(_stackprof.interval) && _stackprof.overall_signals % NUM2LONG(_stackprof.interval))
 	return;
-    stackprof_job_handler(0);
+    stackprof_sample_and_record();
 }
 
 static VALUE
 stackprof_sample(VALUE self)
 {
-    if (!_stackprof.running)
+    if (!STACKPROF_RUNNING())
 	return Qfalse;
 
     _stackprof.overall_signals++;
-    stackprof_job_handler(0);
+    stackprof_sample_and_record();
     return Qtrue;
 }
 
@@ -663,13 +867,24 @@ stackprof_gc_mark(void *data)
 
     if (_stackprof.frames)
 	st_foreach(_stackprof.frames, frame_mark_i, 0);
+
+    int i;
+    for (i = 0; i < _stackprof.buffer_count; i++) {
+        rb_gc_mark(_stackprof.frames_buffer[i]);
+    }
+}
+
+static size_t
+stackprof_memsize(const void *data)
+{
+    return sizeof(_stackprof);
 }
 
 static void
 stackprof_atfork_prepare(void)
 {
     struct itimerval timer;
-    if (_stackprof.running) {
+    if (STACKPROF_RUNNING()) {
 	if (_stackprof.mode == sym_wall || _stackprof.mode == sym_cpu) {
 	    memset(&timer, 0, sizeof(timer));
 	    setitimer(_stackprof.mode == sym_wall ? ITIMER_REAL : ITIMER_PROF, &timer, 0);
@@ -681,7 +896,7 @@ static void
 stackprof_atfork_parent(void)
 {
     struct itimerval timer;
-    if (_stackprof.running) {
+    if (STACKPROF_RUNNING()) {
 	if (_stackprof.mode == sym_wall || _stackprof.mode == sym_cpu) {
 	    timer.it_interval.tv_sec = 0;
 	    timer.it_interval.tv_usec = NUM2LONG(_stackprof.interval);
@@ -697,10 +912,41 @@ stackprof_atfork_child(void)
     stackprof_stop(rb_mStackProf);
 }
 
+static VALUE
+stackprof_use_postponed_job_l(VALUE self)
+{
+    stackprof_use_postponed_job = 1;
+    return Qnil;
+}
+
+static void
+stackprof_at_exit(ruby_vm_t* vm)
+{
+    ruby_vm_running = 0;
+}
+
+static const rb_data_type_t stackprof_type = {
+    "StackProf",
+    {
+        stackprof_gc_mark,
+        NULL,
+        stackprof_memsize,
+    }
+};
+
 void
 Init_stackprof(void)
 {
     size_t i;
+   /*
+    * As of Ruby 3.0, it should be safe to read stack frames at any time, unless YJIT is enabled
+    * See https://github.com/ruby/ruby/commit/0e276dc458f94d9d79a0f7c7669bde84abe80f21
+    */
+    stackprof_use_postponed_job = RUBY_API_VERSION_MAJOR < 3;
+
+    ruby_vm_running = 1;
+    ruby_vm_at_exit(stackprof_at_exit);
+
 #define S(name) sym_##name = ID2SYM(rb_intern(#name));
     S(object);
     S(custom);
@@ -719,9 +965,12 @@ Init_stackprof(void)
     S(mode);
     S(interval);
     S(raw);
+    S(raw_lines);
+    S(raw_sample_timestamps);
     S(raw_timestamp_deltas);
     S(out);
     S(metadata);
+    S(ignore_gc);
     S(frames);
     S(aggregate);
     S(state);
@@ -732,17 +981,17 @@ Init_stackprof(void)
     /* Need to run this to warm the symbol table before we call this during GC */
     rb_gc_latest_gc_info(sym_state);
 
-    gc_hook = Data_Wrap_Struct(rb_cObject, stackprof_gc_mark, NULL, &_stackprof);
     rb_global_variable(&gc_hook);
+    gc_hook = TypedData_Wrap_Struct(rb_cObject, &stackprof_type, &_stackprof);
 
     _stackprof.raw_samples = NULL;
     _stackprof.raw_samples_len = 0;
     _stackprof.raw_samples_capa = 0;
     _stackprof.raw_sample_index = 0;
 
-    _stackprof.raw_timestamp_deltas = NULL;
-    _stackprof.raw_timestamp_deltas_len = 0;
-    _stackprof.raw_timestamp_deltas_capa = 0;
+    _stackprof.raw_sample_times = NULL;
+    _stackprof.raw_sample_times_len = 0;
+    _stackprof.raw_sample_times_capa = 0;
 
     _stackprof.empty_string = rb_str_new_cstr("");
     rb_global_variable(&_stackprof.empty_string);
@@ -759,6 +1008,7 @@ Init_stackprof(void)
     rb_define_singleton_method(rb_mStackProf, "stop", stackprof_stop, 0);
     rb_define_singleton_method(rb_mStackProf, "results", stackprof_results, -1);
     rb_define_singleton_method(rb_mStackProf, "sample", stackprof_sample, 0);
+    rb_define_singleton_method(rb_mStackProf, "use_postponed_job!", stackprof_use_postponed_job_l, 0);
 
     pthread_atfork(stackprof_atfork_prepare, stackprof_atfork_parent, stackprof_atfork_child);
 }