ddtrace 1.4.2 → 1.19.0

data/ext/ddtrace_profiling_native_extension/collectors_dynamic_sampling_rate.c

@@ -0,0 +1,150 @@
+#include <ruby.h>
+
+#include "collectors_dynamic_sampling_rate.h"
+#include "helpers.h"
+#include "ruby_helpers.h"
+#include "time_helpers.h"
+
+// Used to pace the rate of profiling samples based on the last observed time for a sample.
+//
+// This file implements the native bits of the Datadog::Profiling::Collectors::DynamicSamplingRate module, and is
+// only exposed to Ruby for testing (it's always and only invoked by other C code in production).
+
+// ---
+// ## Dynamic Sampling Rate
+//
+// Our profilers get deployed in quite unpredictable situations in terms of system resources. While they can provide key
+// information to help customers solve their performance problems, the profilers must always be careful not to make
+// performance problems worse. This is where the idea of a dynamic sampling rate comes in.
+//
+// Instead of sampling at a fixed sample rate, the actual sampling rate should be decided by also observing the impact
+// that running the profiler is having. This protects against issues such as the profiler being deployed in very busy
+// machines or containers with unrealistic CPU restrictions.
+//
+// ### Implementation
+//
+// The APIs exposed by this file are used by the `CpuAndWallTimeWorker`.
+//
+// The main idea of the implementation below is the following: whenever the profiler takes a sample, the time we spent
+// sampling and the current wall-time are recorded by calling `dynamic_sampling_rate_after_sample()`.
+//
+// Inside `dynamic_sampling_rate_after_sample()`, both values are combined to decide a future wall-time before which
+// we should not sample. That is, we may decide that the next sample should happen no less than 200ms from now.
+//
+// Before taking a sample, the profiler checks using `dynamic_sampling_rate_should_sample()`, if it's time or not to
+// sample. If it's not, it will skip sampling.
+//
+// Finally, as an additional optimization, there's a `dynamic_sampling_rate_get_sleep()` which, given the current
+// wall-time, will return the time remaining (*there's an exception, check function) until the next sample.
+//
+// ---
+
+// This is the wall-time overhead we're targeting. E.g. we target to spend no more than 2%, or 1.2 seconds per minute,
+// taking profiling samples by default.
+#define DEFAULT_WALL_TIME_OVERHEAD_TARGET_PERCENTAGE 2.0 // %
+// See `dynamic_sampling_rate_get_sleep()` for details
+#define MAX_SLEEP_TIME_NS MILLIS_AS_NS(100)
+// See `dynamic_sampling_rate_after_sample()` for details
+#define MAX_TIME_UNTIL_NEXT_SAMPLE_NS SECONDS_AS_NS(10)
+
+void dynamic_sampling_rate_init(dynamic_sampling_rate_state *state) {
+  atomic_init(&state->next_sample_after_monotonic_wall_time_ns, 0);
+  dynamic_sampling_rate_set_overhead_target_percentage(state, DEFAULT_WALL_TIME_OVERHEAD_TARGET_PERCENTAGE);
+}
+
+void dynamic_sampling_rate_set_overhead_target_percentage(dynamic_sampling_rate_state *state, double overhead_target_percentage) {
+  state->overhead_target_percentage = overhead_target_percentage;
+}
+
+void dynamic_sampling_rate_reset(dynamic_sampling_rate_state *state) {
+  atomic_store(&state->next_sample_after_monotonic_wall_time_ns, 0);
+}
+
+uint64_t dynamic_sampling_rate_get_sleep(dynamic_sampling_rate_state *state, long current_monotonic_wall_time_ns) {
+  long next_sample_after_ns = atomic_load(&state->next_sample_after_monotonic_wall_time_ns);
+  long delta_ns = next_sample_after_ns - current_monotonic_wall_time_ns;
+
+  if (delta_ns > 0 && next_sample_after_ns > 0) {
+    // We don't want to sleep for too long as the profiler may be trying to stop.
+    //
+    // Instead, here we sleep for at most this time. Worst case, the profiler will still try to sample before
+    // `next_sample_after_monotonic_wall_time_ns`, BUT `dynamic_sampling_rate_should_sample()` will still be false
+    // so we still get the intended behavior.
+    return uint64_min_of(delta_ns, MAX_SLEEP_TIME_NS);
+  } else {
+    return 0;
+  }
+}
+
+bool dynamic_sampling_rate_should_sample(dynamic_sampling_rate_state *state, long wall_time_ns_before_sample) {
+  return wall_time_ns_before_sample >= atomic_load(&state->next_sample_after_monotonic_wall_time_ns);
+}
+
+void dynamic_sampling_rate_after_sample(dynamic_sampling_rate_state *state, long wall_time_ns_after_sample, uint64_t sampling_time_ns) {
+  double overhead_target = state->overhead_target_percentage;
+
+  // The idea here is that we're targeting a maximum % of wall-time spent sampling.
+  // So for instance, if sampling_time_ns is 2% of the time we spend working, how much is the 98% we should spend
+  // sleeping? As an example, if the last sample took 1ms and the target overhead is 2%, we should sleep for 49ms.
+  uint64_t time_to_sleep_ns = sampling_time_ns * ((100.0 - overhead_target)/overhead_target);
+
+  // In case a sample took an unexpected long time (e.g. maybe a VM was paused, or a laptop was suspended), we clamp the
+  // value so it doesn't get too crazy
+  time_to_sleep_ns = uint64_min_of(time_to_sleep_ns, MAX_TIME_UNTIL_NEXT_SAMPLE_NS);
+
+  atomic_store(&state->next_sample_after_monotonic_wall_time_ns, wall_time_ns_after_sample + time_to_sleep_ns);
+}
+
+// ---
+// Below here is boilerplate to expose the above code to Ruby so that we can test it with RSpec as usual.
+
+VALUE _native_get_sleep(DDTRACE_UNUSED VALUE self, VALUE overhead_target_percentage, VALUE simulated_next_sample_after_monotonic_wall_time_ns, VALUE current_monotonic_wall_time_ns);
+VALUE _native_should_sample(DDTRACE_UNUSED VALUE self, VALUE overhead_target_percentage, VALUE simulated_next_sample_after_monotonic_wall_time_ns, VALUE wall_time_ns_before_sample);
+VALUE _native_after_sample(DDTRACE_UNUSED VALUE self, VALUE overhead_target_percentage, VALUE wall_time_ns_after_sample, VALUE sampling_time_ns);
+
+void collectors_dynamic_sampling_rate_init(VALUE profiling_module) {
+  VALUE collectors_module = rb_define_module_under(profiling_module, "Collectors");
+  VALUE dynamic_sampling_rate_module = rb_define_module_under(collectors_module, "DynamicSamplingRate");
+  VALUE testing_module = rb_define_module_under(dynamic_sampling_rate_module, "Testing");
+
+  rb_define_singleton_method(testing_module, "_native_get_sleep", _native_get_sleep, 3);
+  rb_define_singleton_method(testing_module, "_native_should_sample", _native_should_sample, 3);
+  rb_define_singleton_method(testing_module, "_native_after_sample", _native_after_sample, 3);
+}
+
+VALUE _native_get_sleep(DDTRACE_UNUSED VALUE self, VALUE overhead_target_percentage, VALUE simulated_next_sample_after_monotonic_wall_time_ns, VALUE current_monotonic_wall_time_ns) {
+  ENFORCE_TYPE(simulated_next_sample_after_monotonic_wall_time_ns, T_FIXNUM);
+  ENFORCE_TYPE(current_monotonic_wall_time_ns, T_FIXNUM);
+
+  dynamic_sampling_rate_state state;
+  dynamic_sampling_rate_init(&state);
+  dynamic_sampling_rate_set_overhead_target_percentage(&state, NUM2DBL(overhead_target_percentage));
+  atomic_store(&state.next_sample_after_monotonic_wall_time_ns, NUM2LONG(simulated_next_sample_after_monotonic_wall_time_ns));
+
+  return ULL2NUM(dynamic_sampling_rate_get_sleep(&state, NUM2LONG(current_monotonic_wall_time_ns)));
+}
+
+VALUE _native_should_sample(DDTRACE_UNUSED VALUE self, VALUE overhead_target_percentage, VALUE simulated_next_sample_after_monotonic_wall_time_ns, VALUE wall_time_ns_before_sample) {
+  ENFORCE_TYPE(simulated_next_sample_after_monotonic_wall_time_ns, T_FIXNUM);
+  ENFORCE_TYPE(wall_time_ns_before_sample, T_FIXNUM);
+
+  dynamic_sampling_rate_state state;
+  dynamic_sampling_rate_init(&state);
+  dynamic_sampling_rate_set_overhead_target_percentage(&state, NUM2DBL(overhead_target_percentage));
+  atomic_store(&state.next_sample_after_monotonic_wall_time_ns, NUM2LONG(simulated_next_sample_after_monotonic_wall_time_ns));
+
+  return dynamic_sampling_rate_should_sample(&state, NUM2LONG(wall_time_ns_before_sample)) ? Qtrue : Qfalse;
+}
+
+VALUE _native_after_sample(DDTRACE_UNUSED VALUE self, VALUE overhead_target_percentage, VALUE wall_time_ns_after_sample, VALUE sampling_time_ns) {
+  ENFORCE_TYPE(wall_time_ns_after_sample, T_FIXNUM);
+  ENFORCE_TYPE(sampling_time_ns, T_FIXNUM);
+
+  dynamic_sampling_rate_state state;
+  dynamic_sampling_rate_init(&state);
+  dynamic_sampling_rate_set_overhead_target_percentage(&state, NUM2DBL(overhead_target_percentage));
+
+  dynamic_sampling_rate_after_sample(&state, NUM2LONG(wall_time_ns_after_sample), NUM2ULL(sampling_time_ns));
+
+  return ULL2NUM(atomic_load(&state.next_sample_after_monotonic_wall_time_ns));
+}

data/ext/ddtrace_profiling_native_extension/collectors_dynamic_sampling_rate.h

@@ -0,0 +1,18 @@
+#pragma once
+
+#include <stdatomic.h>
+#include <stdbool.h>
+
+typedef struct {
+  // This is the wall-time overhead we're targeting. E.g. by default, we target to spend no more than 2%, or 1.2 seconds
+  // per minute, taking profiling samples.
+  double overhead_target_percentage;
+  atomic_long next_sample_after_monotonic_wall_time_ns;
+} dynamic_sampling_rate_state;
+
+void dynamic_sampling_rate_init(dynamic_sampling_rate_state *state);
+void dynamic_sampling_rate_set_overhead_target_percentage(dynamic_sampling_rate_state *state, double overhead_target_percentage);
+void dynamic_sampling_rate_reset(dynamic_sampling_rate_state *state);
+uint64_t dynamic_sampling_rate_get_sleep(dynamic_sampling_rate_state *state, long current_monotonic_wall_time_ns);
+bool dynamic_sampling_rate_should_sample(dynamic_sampling_rate_state *state, long wall_time_ns_before_sample);
+void dynamic_sampling_rate_after_sample(dynamic_sampling_rate_state *state, long wall_time_ns_after_sample, uint64_t sampling_time_ns);

data/ext/ddtrace_profiling_native_extension/collectors_gc_profiling_helper.c

@@ -0,0 +1,156 @@
+#include <ruby.h>
+#include <datadog/profiling.h>
+
+#include "collectors_gc_profiling_helper.h"
+
+// This helper is used by the Datadog::Profiling::Collectors::ThreadContext to profile garbage collection.
+// It's tested through that class' interfaces.
+// ---
+
+// Used when retrieving GC information from the VM.
+//  All these are symbols, but we don't need to mark them since we ask for them to be interned (and thus live forever)
+static VALUE state_sym;
+static VALUE marking_sym;
+static VALUE sweeping_sym;
+static VALUE none_sym;
+static VALUE gc_by_sym;
+static VALUE newobj_sym;
+static VALUE malloc_sym;
+static VALUE method_sym;
+static VALUE capi_sym;
+static VALUE stress_sym;
+static VALUE major_by_sym;
+static VALUE nofree_sym;
+static VALUE oldgen_sym;
+static VALUE shady_sym;
+static VALUE force_sym;
+static VALUE oldmalloc_sym;
+
+static ddog_CharSlice major_gc_reason_pretty(VALUE major_gc_reason);
+static ddog_CharSlice gc_cause_pretty(VALUE gc_cause);
+static ddog_CharSlice gc_type_pretty(VALUE major_gc_reason, VALUE gc_state);
+
+void gc_profiling_init(void) {
+  // This function lazy-interns a few constants, which may trigger allocations. Since we want to call it during GC as
+  // well, when allocations are not allowed, we call it once here so that the constants get defined ahead of time.
+  rb_gc_latest_gc_info(rb_hash_new());
+
+  // Used to query and look up the results of GC information
+  state_sym     = ID2SYM(rb_intern_const("state"));
+  marking_sym   = ID2SYM(rb_intern_const("marking"));
+  sweeping_sym  = ID2SYM(rb_intern_const("sweeping"));
+  none_sym      = ID2SYM(rb_intern_const("none"));
+  gc_by_sym     = ID2SYM(rb_intern_const("gc_by"));
+  newobj_sym    = ID2SYM(rb_intern_const("newobj"));
+  malloc_sym    = ID2SYM(rb_intern_const("malloc"));
+  method_sym    = ID2SYM(rb_intern_const("method"));
+  capi_sym      = ID2SYM(rb_intern_const("capi"));
+  stress_sym    = ID2SYM(rb_intern_const("stress"));
+  major_by_sym  = ID2SYM(rb_intern_const("major_by"));
+  nofree_sym    = ID2SYM(rb_intern_const("nofree"));
+  oldgen_sym    = ID2SYM(rb_intern_const("oldgen"));
+  shady_sym     = ID2SYM(rb_intern_const("shady"));
+  force_sym     = ID2SYM(rb_intern_const("force"));
+  oldmalloc_sym = ID2SYM(rb_intern_const("oldmalloc"));
+  state_sym     = ID2SYM(rb_intern_const("state"));
+  none_sym      = ID2SYM(rb_intern_const("none"));
+}
+
+bool gc_profiling_has_major_gc_finished(void) {
+  return rb_gc_latest_gc_info(state_sym) == none_sym && rb_gc_latest_gc_info(major_by_sym) != Qnil;
+}
+
+uint8_t gc_profiling_set_metadata(ddog_prof_Label *labels, int labels_length) {
+  uint8_t max_label_count =
+    1 + // thread id
+    1 + // thread name
+    1 + // state
+    1 + // event
+    1 + // gc reason
+    1 + // gc cause
+    1;  // gc type
+
+  if (max_label_count > labels_length) {
+    rb_raise(rb_eArgError, "BUG: gc_profiling_set_metadata invalid labels_length (%d) < max_label_count (%d)", labels_length, max_label_count);
+  }
+
+  uint8_t label_pos = 0;
+
+  labels[label_pos++] = (ddog_prof_Label) {
+    .key = DDOG_CHARSLICE_C("thread id"),
+    .str = DDOG_CHARSLICE_C("GC"),
+    .num = 0, // This shouldn't be needed but the tracer-2.7 docker image ships a buggy gcc that complains about this
+  };
+
+  labels[label_pos++] = (ddog_prof_Label) {
+    .key = DDOG_CHARSLICE_C("thread name"),
+    .str = DDOG_CHARSLICE_C("Garbage Collection"),
+    .num = 0, // Workaround, same as above
+  };
+
+  labels[label_pos++] = (ddog_prof_Label) {
+    .key = DDOG_CHARSLICE_C("state"),
+    .str = DDOG_CHARSLICE_C("had cpu"),
+    .num = 0, // Workaround, same as above
+  };
+
+  labels[label_pos++] = (ddog_prof_Label) {
+    .key = DDOG_CHARSLICE_C("event"),
+    .str = DDOG_CHARSLICE_C("gc"),
+    .num = 0, // Workaround, same as above
+  };
+
+  VALUE major_by = rb_gc_latest_gc_info(major_by_sym);
+  if (major_by != Qnil) {
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("gc reason"),
+      .str = major_gc_reason_pretty(major_by),
+    };
+  }
+
+  labels[label_pos++] = (ddog_prof_Label) {
+    .key = DDOG_CHARSLICE_C("gc cause"),
+    .str = gc_cause_pretty(rb_gc_latest_gc_info(gc_by_sym)),
+  };
+
+  labels[label_pos++] = (ddog_prof_Label) {
+    .key = DDOG_CHARSLICE_C("gc type"),
+    .str = gc_type_pretty(major_by, rb_gc_latest_gc_info(state_sym)),
+  };
+
+  if (label_pos > max_label_count) {
+    rb_raise(rb_eRuntimeError, "BUG: gc_profiling_set_metadata unexpected label_pos (%d) > max_label_count (%d)", label_pos, max_label_count);
+  }
+
+  return label_pos;
+}
+
+static ddog_CharSlice major_gc_reason_pretty(VALUE major_gc_reason) {
+  if (major_gc_reason == nofree_sym   ) return DDOG_CHARSLICE_C("not enough free slots (NOFREE)");
+  if (major_gc_reason == oldgen_sym   ) return DDOG_CHARSLICE_C("old generation full (OLDGEN)");
+  if (major_gc_reason == shady_sym    ) return DDOG_CHARSLICE_C("too many objects without write barriers (SHADY)");
+  if (major_gc_reason == force_sym    ) return DDOG_CHARSLICE_C("requested (FORCE)");
+  if (major_gc_reason == oldmalloc_sym) return DDOG_CHARSLICE_C("heap bytes allocated threshold (OLDMALLOC)");
+  return DDOG_CHARSLICE_C("unknown");
+}
+
+static ddog_CharSlice gc_cause_pretty(VALUE gc_cause) {
+  if (gc_cause == newobj_sym) return DDOG_CHARSLICE_C("object allocation");
+  if (gc_cause == malloc_sym) return DDOG_CHARSLICE_C("malloc()");
+  if (gc_cause == method_sym) return DDOG_CHARSLICE_C("GC.start()");
+  if (gc_cause == capi_sym  ) return DDOG_CHARSLICE_C("rb_gc()");
+  if (gc_cause == stress_sym) return DDOG_CHARSLICE_C("stress");
+  return DDOG_CHARSLICE_C("unknown");
+}
+
+static ddog_CharSlice gc_type_pretty(VALUE major_gc_reason, VALUE gc_state) {
+  if (major_gc_reason != Qnil) {
+    if (gc_state == marking_sym ) return DDOG_CHARSLICE_C("major (ongoing, marking)");
+    if (gc_state == sweeping_sym) return DDOG_CHARSLICE_C("major (ongoing, sweeping)");
+    return DDOG_CHARSLICE_C("major");
+  } else {
+    // As we delay flushing events when a minor GC finishes, it's not relevant to include the observed state of the
+    // minor GC, as we often won't record a marking -> sweeping -> done cycle, as it happens too quickly.
+    return DDOG_CHARSLICE_C("minor");
+  }
+}

data/ext/ddtrace_profiling_native_extension/collectors_gc_profiling_helper.h

@@ -0,0 +1,5 @@
+#pragma once
+
+void gc_profiling_init(void);
+bool gc_profiling_has_major_gc_finished(void);
+uint8_t gc_profiling_set_metadata(ddog_prof_Label *labels, int labels_length);

data/ext/ddtrace_profiling_native_extension/collectors_idle_sampling_helper.c

@@ -0,0 +1,244 @@
+#include <ruby.h>
+#include <ruby/thread.h>
+#include <pthread.h>
+#include <stdbool.h>
+
+#include "helpers.h"
+#include "ruby_helpers.h"
+#include "collectors_idle_sampling_helper.h"
+
+// Used by the Collectors::CpuAndWallTimeWorker to gather samples when the Ruby process is idle.
+//
+// Specifically, the IdleSamplingHelper is expected to be triggered by the CpuAndWallTimeWorker whenever it needs to
+// trigger a sample, but the VM is otherwise idle. See implementation of CpuAndWallTimeWorker for details.
+//
+// The IdleSamplingHelper keeps a background thread that waits for functions to run on a single-element "queue".
+// Other threads communicate with it by asking it to ACTION_RUN a `requested_action` or ACTION_STOP to terminate.
+//
+// The state is protected by the `wakeup_mutex`, and the background thread is woken up after changes using the
+// `wakeup` condition variable.
+
+typedef enum { ACTION_WAIT, ACTION_RUN, ACTION_STOP } action;
+
+// Contains state for a single CpuAndWallTimeWorker instance
+struct idle_sampling_loop_state {
+  pthread_mutex_t wakeup_mutex;
+  pthread_cond_t wakeup;
+  action requested_action;
+  void (*run_action_function)(void);
+};
+
+static VALUE _native_new(VALUE klass);
+static void reset_state(struct idle_sampling_loop_state *state);
+static VALUE _native_idle_sampling_loop(DDTRACE_UNUSED VALUE self, VALUE self_instance);
+static VALUE _native_stop(DDTRACE_UNUSED VALUE self, VALUE self_instance);
+static void *run_idle_sampling_loop(void *state_ptr);
+static void interrupt_idle_sampling_loop(void *state_ptr);
+static VALUE _native_reset(DDTRACE_UNUSED VALUE self, VALUE self_instance);
+static VALUE _native_idle_sampling_helper_request_action(DDTRACE_UNUSED VALUE self, VALUE self_instance);
+static void *request_testing_action(void *self_instance_ptr);
+static void grab_gvl_and_run_testing_action(void);
+static void *run_testing_action(DDTRACE_UNUSED void *unused);
+
+void collectors_idle_sampling_helper_init(VALUE profiling_module) {
+  VALUE collectors_module = rb_define_module_under(profiling_module, "Collectors");
+  VALUE collectors_idle_sampling_helper_class = rb_define_class_under(collectors_module, "IdleSamplingHelper", rb_cObject);
+  // Hosts methods used for testing the native code using RSpec
+  VALUE testing_module = rb_define_module_under(collectors_idle_sampling_helper_class, "Testing");
+
+  // Instances of the IdleSamplingHelper class are "TypedData" objects.
+  // "TypedData" objects are special objects in the Ruby VM that can wrap C structs.
+  // In this case, it wraps the idle_sampling_loop_state.
+  //
+  // Because Ruby doesn't know how to initialize native-level structs, we MUST override the allocation function for objects
+  // of this class so that we can manage this part. Not overriding or disabling the allocation function is a common
+  // gotcha for "TypedData" objects that can very easily lead to VM crashes, see for instance
+  // https://bugs.ruby-lang.org/issues/18007 for a discussion around this.
+  rb_define_alloc_func(collectors_idle_sampling_helper_class, _native_new);
+
+  rb_define_singleton_method(collectors_idle_sampling_helper_class, "_native_idle_sampling_loop", _native_idle_sampling_loop, 1);
+  rb_define_singleton_method(collectors_idle_sampling_helper_class, "_native_stop", _native_stop, 1);
+  rb_define_singleton_method(collectors_idle_sampling_helper_class, "_native_reset", _native_reset, 1);
+  rb_define_singleton_method(testing_module, "_native_idle_sampling_helper_request_action", _native_idle_sampling_helper_request_action, 1);
+}
+
+// This structure is used to define a Ruby object that stores a pointer to a struct idle_sampling_loop_state
+// See also https://github.com/ruby/ruby/blob/master/doc/extension.rdoc for how this works
+static const rb_data_type_t idle_sampling_helper_typed_data = {
+  .wrap_struct_name = "Datadog::Profiling::Collectors::IdleSamplingHelper",
+  .function = {
+    .dmark = NULL, // We don't store references to Ruby objects so we don't need to mark any of them
+    .dfree = RUBY_DEFAULT_FREE,
+    .dsize = NULL, // We don't track memory usage (although it'd be cool if we did!)
+    //.dcompact = NULL, // Not needed -- we don't store references to Ruby objects
+  },
+  .flags = RUBY_TYPED_FREE_IMMEDIATELY
+};
+
+static VALUE _native_new(VALUE klass) {
+  struct idle_sampling_loop_state *state = ruby_xcalloc(1, sizeof(struct idle_sampling_loop_state));
+
+  // Note: Any exceptions raised from this note until the TypedData_Wrap_Struct call will lead to the state memory
+  // being leaked.
+
+  reset_state(state);
+
+  return TypedData_Wrap_Struct(klass, &idle_sampling_helper_typed_data, state);
+}
+
+static void reset_state(struct idle_sampling_loop_state *state) {
+  state->wakeup_mutex = (pthread_mutex_t) PTHREAD_MUTEX_INITIALIZER;
+  state->wakeup = (pthread_cond_t) PTHREAD_COND_INITIALIZER;
+  state->requested_action = ACTION_WAIT;
+  state->run_action_function = NULL;
+}
+
+// The same instance of the IdleSamplingHelper can be reused multiple times, and this resets it back to
+// a pristine state before recreating the worker thread (this includes resetting the mutex in case it was left
+// locked halfway through the VM forking)
+static VALUE _native_reset(DDTRACE_UNUSED VALUE self, VALUE self_instance) {
+  struct idle_sampling_loop_state *state;
+  TypedData_Get_Struct(self_instance, struct idle_sampling_loop_state, &idle_sampling_helper_typed_data, state);
+
+  reset_state(state);
+
+  return Qtrue;
+}
+
+static VALUE _native_idle_sampling_loop(DDTRACE_UNUSED VALUE self, VALUE self_instance) {
+  struct idle_sampling_loop_state *state;
+  TypedData_Get_Struct(self_instance, struct idle_sampling_loop_state, &idle_sampling_helper_typed_data, state);
+
+  // Release GVL and run the loop waiting for requests
+  rb_thread_call_without_gvl(run_idle_sampling_loop, state, interrupt_idle_sampling_loop, state);
+
+  return Qtrue;
+}
+
+static void *run_idle_sampling_loop(void *state_ptr) {
+  struct idle_sampling_loop_state *state = (struct idle_sampling_loop_state *) state_ptr;
+  int error = 0;
+
+  while (true) {
+    ENFORCE_SUCCESS_NO_GVL(pthread_mutex_lock(&state->wakeup_mutex));
+
+    action next_action;
+    void (*run_action_function)(void);
+
+    // Await for an action
+    while ((next_action = state->requested_action) == ACTION_WAIT) {
+      error = pthread_cond_wait(&state->wakeup, &state->wakeup_mutex);
+      if (error) {
+        // If something went wrong, try to leave the mutex unlocked at least
+        pthread_mutex_unlock(&state->wakeup_mutex);
+        ENFORCE_SUCCESS_NO_GVL(error);
+      }
+    }
+
+    // There's an action to be taken!
+
+    // Record function, if any
+    run_action_function = state->run_action_function;
+
+    // Reset buffer for next request
+    state->requested_action = ACTION_WAIT;
+
+    // Unlock the mutex immediately so other threads can continue to request actions without blocking
+    ENFORCE_SUCCESS_NO_GVL(pthread_mutex_unlock(&state->wakeup_mutex));
+
+    // Process pending action
+    if (next_action == ACTION_RUN) {
+      if (run_action_function == NULL) {
+        grab_gvl_and_raise(rb_eRuntimeError, "Unexpected NULL run_action_function in run_idle_sampling_loop");
+      }
+
+      run_action_function();
+    } else { // ACTION_STOP
+      return NULL;
+    }
+  }
+}
+
+static void interrupt_idle_sampling_loop(void *state_ptr) {
+  struct idle_sampling_loop_state *state = (struct idle_sampling_loop_state *) state_ptr;
+  int error = 0;
+
+  // Note about the error handling in this situation: Something bad happening at this stage is really really awkward to
+  // handle because we get called by the VM in a situation where we can't really raise exceptions, and the VM really really
+  // just wants us to stop what we're doing and return control of the thread to it.
+  //
+  // So if we return immediately on error, we may leave the VM hanging because we didn't actually interrupt the thread.
+  // We're also not at a great location to flag errors.
+  // That's why: a) I chose to log to stderr, as a last-ditch effort; b) even if something goes wrong we still try to
+  // ask the thread to stop, instead of exiting early.
+
+  error = pthread_mutex_lock(&state->wakeup_mutex);
+  if (error) { fprintf(stderr, "[ddtrace] Error during pthread_mutex_lock in interrupt_idle_sampling_loop (%s)\n", strerror(error)); }
+
+  state->requested_action = ACTION_STOP;
+
+  error = pthread_mutex_unlock(&state->wakeup_mutex);
+  if (error) { fprintf(stderr, "[ddtrace] Error during pthread_mutex_unlock in interrupt_idle_sampling_loop (%s)\n", strerror(error)); }
+
+  error = pthread_cond_broadcast(&state->wakeup);
+  if (error) { fprintf(stderr, "[ddtrace] Error during pthread_cond_broadcast in interrupt_idle_sampling_loop (%s)\n", strerror(error)); }
+}
+
+static VALUE _native_stop(DDTRACE_UNUSED VALUE self, VALUE self_instance) {
+  struct idle_sampling_loop_state *state;
+  TypedData_Get_Struct(self_instance, struct idle_sampling_loop_state, &idle_sampling_helper_typed_data, state);
+
+  ENFORCE_SUCCESS_GVL(pthread_mutex_lock(&state->wakeup_mutex));
+  state->requested_action = ACTION_STOP;
+  ENFORCE_SUCCESS_GVL(pthread_mutex_unlock(&state->wakeup_mutex));
+
+  // Wake up worker thread, if needed; It's OK to call broadcast after releasing the mutex
+  ENFORCE_SUCCESS_GVL(pthread_cond_broadcast(&state->wakeup));
+
+  return Qtrue;
+}
+
+// Assumption: Function gets called without the global VM lock
+void idle_sampling_helper_request_action(VALUE self_instance, void (*run_action_function)(void)) {
+  struct idle_sampling_loop_state *state;
+  if (!rb_typeddata_is_kind_of(self_instance, &idle_sampling_helper_typed_data)) {
+    grab_gvl_and_raise(rb_eTypeError, "Wrong argument for idle_sampling_helper_request_action");
+  }
+  // This should never fail the the above check passes
+  TypedData_Get_Struct(self_instance, struct idle_sampling_loop_state, &idle_sampling_helper_typed_data, state);
+
+  ENFORCE_SUCCESS_NO_GVL(pthread_mutex_lock(&state->wakeup_mutex));
+  if (state->requested_action == ACTION_WAIT) {
+    state->requested_action = ACTION_RUN;
+    state->run_action_function = run_action_function;
+  }
+  ENFORCE_SUCCESS_NO_GVL(pthread_mutex_unlock(&state->wakeup_mutex));
+
+  // Wake up worker thread, if needed; It's OK to call broadcast after releasing the mutex
+  ENFORCE_SUCCESS_NO_GVL(pthread_cond_broadcast(&state->wakeup));
+}
+
+// Because the idle_sampling_helper_request_action is built to be called without the global VM lock, here we release it
+// to be able to call that API.
+static VALUE _native_idle_sampling_helper_request_action(DDTRACE_UNUSED VALUE self, VALUE self_instance) {
+  rb_thread_call_without_gvl(request_testing_action, (void *) self_instance, NULL, NULL);
+  return Qtrue;
+}
+
+static void *request_testing_action(void *self_instance_ptr) {
+  VALUE self_instance = (VALUE) self_instance_ptr;
+  idle_sampling_helper_request_action(self_instance, grab_gvl_and_run_testing_action);
+  return NULL;
+}
+
+// This gets called by the worker thread, which is not holding the global VM lock. To be able to actually run the action,
+// we need to acquire it
+static void grab_gvl_and_run_testing_action(void) {
+  rb_thread_call_with_gvl(run_testing_action, NULL);
+}
+
+static void *run_testing_action(DDTRACE_UNUSED void *unused) {
+  VALUE idle_sampling_helper_testing_action = rb_gv_get("$idle_sampling_helper_testing_action");
+  rb_funcall(idle_sampling_helper_testing_action, rb_intern("call"), 0);
+  return NULL;
+}

data/ext/ddtrace_profiling_native_extension/collectors_idle_sampling_helper.h

@@ -0,0 +1,3 @@
+#pragma once
+
+void idle_sampling_helper_request_action(VALUE self_instance, void (*run_action_function)(void));