datadog 2.3.0 → 2.4.0

data/ext/datadog_profiling_native_extension/collectors_thread_context.c

@@ -76,6 +76,11 @@
 #define MISSING_TRACER_CONTEXT_KEY 0
 #define TIME_BETWEEN_GC_EVENTS_NS MILLIS_AS_NS(10)
 
+// This is used as a placeholder to mark threads that are allowed to be profiled (enabled)
+// (e.g. to avoid trying to gvl profile threads that are not from the main Ractor)
+// and for which there's no data yet
+#define GVL_WAITING_ENABLED_EMPTY RUBY_FIXNUM_MAX
+
 static ID at_active_span_id;  // id of :@active_span in Ruby
 static ID at_active_trace_id; // id of :@active_trace in Ruby
 static ID at_id_id;           // id of :@id in Ruby
@@ -86,6 +91,26 @@ static ID at_otel_values_id;  // id of :@otel_values in Ruby
 static ID at_parent_span_id_id; // id of :@parent_span_id in Ruby
 static ID at_datadog_trace_id;  // id of :@datadog_trace in Ruby
 
+// Used to support reading trace identifiers from the opentelemetry Ruby library when the ddtrace gem tracing
+// integration is NOT in use.
+static ID at_span_id_id;  // id of :@span_id in Ruby
+static ID at_trace_id_id; // id of :@trace_id in Ruby
+static ID at_entries_id;  // id of :@entries in Ruby
+static ID at_context_id;  // id of :@context in Ruby
+static ID at_kind_id;     // id of :@kind in Ruby
+static ID at_name_id;     // id of :@name in Ruby
+static ID server_id;      // id of :server in Ruby
+static ID otel_context_storage_id; // id of :__opentelemetry_context_storage__ in Ruby
+
+// This is used by `thread_context_collector_on_gvl_running`. Because when that method gets called we're not sure if
+// it's safe to access the state of the thread context collector, we store this setting as a global value. This does
+// mean this setting is shared among all thread context collectors, and thus it's "last writer wins".
+// In production this should not be a problem: there should only be one profiler, which is the last one created,
+// and that'll be the one that last wrote this setting.
+static uint32_t global_waiting_for_gvl_threshold_ns = MILLIS_AS_NS(10);
+
+enum otel_context_enabled {otel_context_enabled_false, otel_context_enabled_only, otel_context_enabled_both};
+
 // Contains state for a single ThreadContext instance
 struct thread_context_collector_state {
   // Note: Places in this file that usually need to be changed when this struct is changed are tagged with
@@ -112,6 +137,8 @@ struct thread_context_collector_state {
   bool endpoint_collection_enabled;
   // Used to omit timestamps / timeline events from collected data
   bool timeline_enabled;
+  // Used to control context collection
+  enum otel_context_enabled otel_context_enabled;
   // Used to omit class information from collected allocation data
   bool allocation_type_enabled;
   // Used when calling monotonic_to_system_epoch_ns
@@ -119,6 +146,8 @@ struct thread_context_collector_state {
   // Used to identify the main thread, to give it a fallback name
   VALUE main_thread;
   // Used when extracting trace identifiers from otel spans. Lazily initialized.
+  // Qtrue serves as a marker we've not yet extracted it; when we try to extract it, we set it to an object if
+  // successful and Qnil if not.
   VALUE otel_current_span_key;
 
   struct stats {
@@ -164,6 +193,12 @@ struct trace_identifiers {
   VALUE trace_endpoint;
 };
 
+struct otel_span {
+  VALUE span;
+  VALUE span_id;
+  VALUE trace_id;
+};
+
 static void thread_context_collector_typed_data_mark(void *state_ptr);
 static void thread_context_collector_typed_data_free(void *state_ptr);
 static int hash_map_per_thread_context_mark(st_data_t key_thread, st_data_t _value, st_data_t _argument);
@@ -177,13 +212,15 @@ static VALUE _native_initialize(
   VALUE tracer_context_key,
   VALUE endpoint_collection_enabled,
   VALUE timeline_enabled,
+  VALUE waiting_for_gvl_threshold_ns,
+  VALUE otel_context_enabled,
   VALUE allocation_type_enabled
 );
 static VALUE _native_sample(VALUE self, VALUE collector_instance, VALUE profiler_overhead_stack_thread);
 static VALUE _native_on_gc_start(VALUE self, VALUE collector_instance);
 static VALUE _native_on_gc_finish(VALUE self, VALUE collector_instance);
-static VALUE _native_sample_after_gc(DDTRACE_UNUSED VALUE self, VALUE collector_instance);
-void update_metrics_and_sample(
+static VALUE _native_sample_after_gc(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE reset_monotonic_to_system_state);
+static void update_metrics_and_sample(
   struct thread_context_collector_state *state,
   VALUE thread_being_sampled,
   VALUE stack_from_thread,
@@ -201,7 +238,8 @@ static void trigger_sample_for_thread(
   sample_values values,
   long current_monotonic_wall_time_ns,
   ddog_CharSlice *ruby_vm_type,
-  ddog_CharSlice *class_name
+  ddog_CharSlice *class_name,
+  bool is_gvl_waiting_state
 );
 static VALUE _native_thread_list(VALUE self);
 static struct per_thread_context *get_or_create_context_for(VALUE thread, struct thread_context_collector_state *state);
@@ -237,6 +275,26 @@ static void ddtrace_otel_trace_identifiers_for(
   VALUE otel_values
 );
 static VALUE _native_sample_skipped_allocation_samples(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE skipped_samples);
+static bool handle_gvl_waiting(
+  struct thread_context_collector_state *state,
+  VALUE thread_being_sampled,
+  VALUE stack_from_thread,
+  struct per_thread_context *thread_context,
+  sampling_buffer* sampling_buffer,
+  long current_cpu_time_ns
+);
+static VALUE _native_on_gvl_waiting(DDTRACE_UNUSED VALUE self, VALUE thread);
+static VALUE _native_gvl_waiting_at_for(DDTRACE_UNUSED VALUE self, VALUE thread);
+static VALUE _native_on_gvl_running(DDTRACE_UNUSED VALUE self, VALUE thread);
+static VALUE _native_sample_after_gvl_running(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE thread);
+static VALUE _native_apply_delta_to_cpu_time_at_previous_sample_ns(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE thread, VALUE delta_ns);
+static void otel_without_ddtrace_trace_identifiers_for(
+  struct thread_context_collector_state *state,
+  VALUE thread,
+  struct trace_identifiers *trace_identifiers_result
+);
+static struct otel_span otel_span_from(VALUE otel_context, VALUE otel_current_span_key);
+static uint64_t otel_span_id_to_uint(VALUE otel_span_id);
 
 void collectors_thread_context_init(VALUE profiling_module) {
   VALUE collectors_module = rb_define_module_under(profiling_module, "Collectors");
@@ -254,20 +312,27 @@ void collectors_thread_context_init(VALUE profiling_module) {
   // https://bugs.ruby-lang.org/issues/18007 for a discussion around this.
   rb_define_alloc_func(collectors_thread_context_class, _native_new);
 
-  rb_define_singleton_method(collectors_thread_context_class, "_native_initialize", _native_initialize, 7);
+  rb_define_singleton_method(collectors_thread_context_class, "_native_initialize", _native_initialize, 9);
   rb_define_singleton_method(collectors_thread_context_class, "_native_inspect", _native_inspect, 1);
   rb_define_singleton_method(collectors_thread_context_class, "_native_reset_after_fork", _native_reset_after_fork, 1);
   rb_define_singleton_method(testing_module, "_native_sample", _native_sample, 2);
   rb_define_singleton_method(testing_module, "_native_sample_allocation", _native_sample_allocation, 3);
   rb_define_singleton_method(testing_module, "_native_on_gc_start", _native_on_gc_start, 1);
   rb_define_singleton_method(testing_module, "_native_on_gc_finish", _native_on_gc_finish, 1);
-  rb_define_singleton_method(testing_module, "_native_sample_after_gc", _native_sample_after_gc, 1);
+  rb_define_singleton_method(testing_module, "_native_sample_after_gc", _native_sample_after_gc, 2);
   rb_define_singleton_method(testing_module, "_native_thread_list", _native_thread_list, 0);
   rb_define_singleton_method(testing_module, "_native_per_thread_context", _native_per_thread_context, 1);
   rb_define_singleton_method(testing_module, "_native_stats", _native_stats, 1);
   rb_define_singleton_method(testing_module, "_native_gc_tracking", _native_gc_tracking, 1);
   rb_define_singleton_method(testing_module, "_native_new_empty_thread", _native_new_empty_thread, 0);
   rb_define_singleton_method(testing_module, "_native_sample_skipped_allocation_samples", _native_sample_skipped_allocation_samples, 2);
+  #ifndef NO_GVL_INSTRUMENTATION
+    rb_define_singleton_method(testing_module, "_native_on_gvl_waiting", _native_on_gvl_waiting, 1);
+    rb_define_singleton_method(testing_module, "_native_gvl_waiting_at_for", _native_gvl_waiting_at_for, 1);
+    rb_define_singleton_method(testing_module, "_native_on_gvl_running", _native_on_gvl_running, 1);
+    rb_define_singleton_method(testing_module, "_native_sample_after_gvl_running", _native_sample_after_gvl_running, 2);
+    rb_define_singleton_method(testing_module, "_native_apply_delta_to_cpu_time_at_previous_sample_ns", _native_apply_delta_to_cpu_time_at_previous_sample_ns, 3);
+  #endif
 
   at_active_span_id = rb_intern_const("@active_span");
   at_active_trace_id = rb_intern_const("@active_trace");
@@ -278,6 +343,19 @@ void collectors_thread_context_init(VALUE profiling_module) {
   at_otel_values_id = rb_intern_const("@otel_values");
   at_parent_span_id_id = rb_intern_const("@parent_span_id");
   at_datadog_trace_id = rb_intern_const("@datadog_trace");
+  at_span_id_id = rb_intern_const("@span_id");
+  at_trace_id_id = rb_intern_const("@trace_id");
+  at_entries_id = rb_intern_const("@entries");
+  at_context_id = rb_intern_const("@context");
+  at_kind_id = rb_intern_const("@kind");
+  at_name_id = rb_intern_const("@name");
+  server_id = rb_intern_const("server");
+  otel_context_storage_id = rb_intern_const("__opentelemetry_context_storage__");
+
+  #ifndef NO_GVL_INSTRUMENTATION
+    // This will raise if Ruby already ran out of thread-local keys
+    gvl_profiling_init();
+  #endif
 
   gc_profiling_init();
 }
@@ -357,11 +435,12 @@ static VALUE _native_new(VALUE klass) {
   state->thread_list_buffer = thread_list_buffer;
   state->endpoint_collection_enabled = true;
   state->timeline_enabled = true;
+  state->otel_context_enabled = otel_context_enabled_false;
   state->allocation_type_enabled = true;
   state->time_converter_state = (monotonic_to_system_epoch_state) MONOTONIC_TO_SYSTEM_EPOCH_INITIALIZER;
   VALUE main_thread = rb_thread_main();
   state->main_thread = main_thread;
-  state->otel_current_span_key = Qnil;
+  state->otel_current_span_key = Qtrue;
   state->gc_tracking.wall_time_at_previous_gc_ns = INVALID_TIME;
   state->gc_tracking.wall_time_at_last_flushed_gc_event_ns = 0;
 
@@ -377,6 +456,7 @@ static VALUE _native_new(VALUE klass) {
   return instance;
 }
 
+// TODO: Convert this to use options like CpuAndWallTimeWorker
 static VALUE _native_initialize(
   DDTRACE_UNUSED VALUE _self,
   VALUE collector_instance,
@@ -385,10 +465,13 @@ static VALUE _native_initialize(
   VALUE tracer_context_key,
   VALUE endpoint_collection_enabled,
   VALUE timeline_enabled,
+  VALUE waiting_for_gvl_threshold_ns,
+  VALUE otel_context_enabled,
   VALUE allocation_type_enabled
 ) {
   ENFORCE_BOOLEAN(endpoint_collection_enabled);
   ENFORCE_BOOLEAN(timeline_enabled);
+  ENFORCE_TYPE(waiting_for_gvl_threshold_ns, T_FIXNUM);
   ENFORCE_BOOLEAN(allocation_type_enabled);
 
   struct thread_context_collector_state *state;
@@ -401,8 +484,19 @@ static VALUE _native_initialize(
   state->recorder_instance = enforce_recorder_instance(recorder_instance);
   state->endpoint_collection_enabled = (endpoint_collection_enabled == Qtrue);
   state->timeline_enabled = (timeline_enabled == Qtrue);
+  if (otel_context_enabled == Qfalse || otel_context_enabled == Qnil) {
+    state->otel_context_enabled = otel_context_enabled_false;
+  } else if (otel_context_enabled == ID2SYM(rb_intern("only"))) {
+    state->otel_context_enabled = otel_context_enabled_only;
+  } else if (otel_context_enabled == ID2SYM(rb_intern("both"))) {
+    state->otel_context_enabled = otel_context_enabled_both;
+  } else {
+    rb_raise(rb_eArgError, "Unexpected value for otel_context_enabled: %+" PRIsVALUE, otel_context_enabled);
+  }
   state->allocation_type_enabled = (allocation_type_enabled == Qtrue);
 
+  global_waiting_for_gvl_threshold_ns = NUM2UINT(waiting_for_gvl_threshold_ns);
+
   if (RTEST(tracer_context_key)) {
     ENFORCE_TYPE(tracer_context_key, T_SYMBOL);
     // Note about rb_to_id and dynamic symbols: calling `rb_to_id` prevents symbols from ever being garbage collected.
@@ -433,13 +527,22 @@ static VALUE _native_on_gc_start(DDTRACE_UNUSED VALUE self, VALUE collector_inst
 // This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
 // It SHOULD NOT be used for other purposes.
 static VALUE _native_on_gc_finish(DDTRACE_UNUSED VALUE self, VALUE collector_instance) {
-  thread_context_collector_on_gc_finish(collector_instance);
+  (void) !thread_context_collector_on_gc_finish(collector_instance);
   return Qtrue;
 }
 
 // This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
 // It SHOULD NOT be used for other purposes.
-static VALUE _native_sample_after_gc(DDTRACE_UNUSED VALUE self, VALUE collector_instance) {
+static VALUE _native_sample_after_gc(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE reset_monotonic_to_system_state) {
+  ENFORCE_BOOLEAN(reset_monotonic_to_system_state);
+
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  if (reset_monotonic_to_system_state == Qtrue) {
+    state->time_converter_state = (monotonic_to_system_epoch_state) MONOTONIC_TO_SYSTEM_EPOCH_INITIALIZER;
+  }
+
   thread_context_collector_sample_after_gc(collector_instance);
   return Qtrue;
 }
@@ -502,7 +605,7 @@ void thread_context_collector_sample(VALUE self_instance, long current_monotonic
   );
 }
 
-void update_metrics_and_sample(
+static void update_metrics_and_sample(
   struct thread_context_collector_state *state,
   VALUE thread_being_sampled,
   VALUE stack_from_thread, // This can be different when attributing profiler overhead using a different stack
@@ -511,12 +614,17 @@ void update_metrics_and_sample(
   long current_cpu_time_ns,
   long current_monotonic_wall_time_ns
 ) {
-  long cpu_time_elapsed_ns = update_time_since_previous_sample(
+  bool is_gvl_waiting_state =
+    handle_gvl_waiting(state, thread_being_sampled, stack_from_thread, thread_context, sampling_buffer, current_cpu_time_ns);
+
+  // Don't assign/update cpu during "Waiting for GVL"
+  long cpu_time_elapsed_ns = is_gvl_waiting_state ? 0 : update_time_since_previous_sample(
     &thread_context->cpu_time_at_previous_sample_ns,
     current_cpu_time_ns,
     thread_context->gc_tracking.cpu_time_at_start_ns,
     IS_NOT_WALL_TIME
   );
+
   long wall_time_elapsed_ns = update_time_since_previous_sample(
     &thread_context->wall_time_at_previous_sample_ns,
     current_monotonic_wall_time_ns,
@@ -528,6 +636,21 @@ void update_metrics_and_sample(
     IS_WALL_TIME
   );
 
+  // A thread enters "Waiting for GVL", well, as the name implies, without the GVL.
+  //
+  // As a consequence, it's possible that a thread enters "Waiting for GVL" in parallel with the current thread working
+  // on sampling, and thus for the  `current_monotonic_wall_time_ns` (which is recorded at the start of sampling)
+  // to be < the time at which we started Waiting for GVL.
+  //
+  // All together, this means that when `handle_gvl_waiting` creates an extra sample (see comments on that function for
+  // what the extra sample is), it's possible that there's no more wall-time to be assigned.
+  // Thus, in this case, we don't want to produce a sample representing Waiting for GVL with a wall-time of 0, and
+  // thus we skip creating such a sample.
+  if (is_gvl_waiting_state && wall_time_elapsed_ns == 0) return;
+  // ...you may also wonder: is there any other situation where it makes sense to produce a sample with
+  // wall_time_elapsed_ns == 0? I believe that yes, because the sample still includes a timestamp and a stack, but we
+  // may revisit/change our minds on this in the future.
+
   trigger_sample_for_thread(
     state,
     thread_being_sampled,
@@ -537,7 +660,8 @@ void update_metrics_and_sample(
     (sample_values) {.cpu_time_ns = cpu_time_elapsed_ns, .cpu_or_wall_samples = 1, .wall_time_ns = wall_time_elapsed_ns},
     current_monotonic_wall_time_ns,
     NULL,
-    NULL
+    NULL,
+    is_gvl_waiting_state
   );
 }
 
@@ -583,6 +707,7 @@ void thread_context_collector_on_gc_start(VALUE self_instance) {
 //
 // Assumption 1: This function is called in a thread that is holding the Global VM Lock. Caller is responsible for enforcing this.
 // Assumption 2: This function is called from the main Ractor (if Ruby has support for Ractors).
+__attribute__((warn_unused_result))
 bool thread_context_collector_on_gc_finish(VALUE self_instance) {
   struct thread_context_collector_state *state;
   if (!rb_typeddata_is_kind_of(self_instance, &thread_context_collector_typed_data)) return false;
@@ -718,7 +843,8 @@ static void trigger_sample_for_thread(
   long current_monotonic_wall_time_ns,
   // These two labels are only used for allocation profiling; @ivoanjo: may want to refactor this at some point?
   ddog_CharSlice *ruby_vm_type,
-  ddog_CharSlice *class_name
+  ddog_CharSlice *class_name,
+  bool is_gvl_waiting_state
 ) {
   int max_label_count =
     1 + // thread id
@@ -759,6 +885,11 @@ static void trigger_sample_for_thread(
   struct trace_identifiers trace_identifiers_result = {.valid = false, .trace_endpoint = Qnil};
   trace_identifiers_for(state, thread, &trace_identifiers_result);
 
+  if (!trace_identifiers_result.valid && state->otel_context_enabled != otel_context_enabled_false) {
+    // If we couldn't get something with ddtrace, let's see if we can get some trace identifiers from opentelemetry directly
+    otel_without_ddtrace_trace_identifiers_for(state, thread, &trace_identifiers_result);
+  }
+
   if (trace_identifiers_result.valid) {
     labels[label_pos++] = (ddog_prof_Label) {.key = DDOG_CHARSLICE_C("local root span id"), .num = trace_identifiers_result.local_root_span_id};
     labels[label_pos++] = (ddog_prof_Label) {.key = DDOG_CHARSLICE_C("span id"), .num = trace_identifiers_result.span_id};
@@ -837,7 +968,12 @@ static void trigger_sample_for_thread(
     sampling_buffer,
     state->recorder_instance,
     values,
-    (sample_labels) {.labels = slice_labels, .state_label = state_label, .end_timestamp_ns = end_timestamp_ns}
+    (sample_labels) {
+      .labels = slice_labels,
+      .state_label = state_label,
+      .end_timestamp_ns = end_timestamp_ns,
+      .is_gvl_waiting_state = is_gvl_waiting_state,
+    }
   );
 }
 
@@ -887,9 +1023,9 @@ static struct per_thread_context *get_context_for(VALUE thread, struct thread_co
 // to either run Ruby code during sampling (not great), or otherwise use some of the VM private APIs to detect this.
 //
 static bool is_logging_gem_monkey_patch(VALUE invoke_file_location) {
-  int logging_gem_path_len = strlen(LOGGING_GEM_PATH);
+  unsigned long logging_gem_path_len = strlen(LOGGING_GEM_PATH);
   char *invoke_file = StringValueCStr(invoke_file_location);
-  int invoke_file_len = strlen(invoke_file);
+  unsigned long invoke_file_len = strlen(invoke_file);
 
   if (invoke_file_len < logging_gem_path_len) return false;
 
@@ -937,6 +1073,20 @@ static void initialize_context(VALUE thread, struct per_thread_context *thread_c
   // These will only be used during a GC operation
   thread_context->gc_tracking.cpu_time_at_start_ns = INVALID_TIME;
   thread_context->gc_tracking.wall_time_at_start_ns = INVALID_TIME;
+
+  #ifndef NO_GVL_INSTRUMENTATION
+    // We use this special location to store data that can be accessed without any
+    // kind of synchronization (e.g. by threads without the GVL).
+    //
+    // We set this marker here for two purposes:
+    // * To make sure there's no stale data from a previous execution of the profiler.
+    // * To mark threads that are actually being profiled
+    //
+    // (Setting this is potentially a race, but what we want is to avoid _stale_ data, so
+    // if this gets set concurrently with context initialization, then such a value will belong
+    // to the current profiler instance, so that's OK)
+    gvl_profiling_state_thread_object_set(thread, GVL_WAITING_ENABLED_EMPTY);
+  #endif
 }
 
 static void free_context(struct per_thread_context* thread_context) {
@@ -960,6 +1110,7 @@ static VALUE _native_inspect(DDTRACE_UNUSED VALUE _self, VALUE collector_instanc
   rb_str_concat(result, rb_sprintf(" stats=%"PRIsVALUE, stats_as_ruby_hash(state)));
   rb_str_concat(result, rb_sprintf(" endpoint_collection_enabled=%"PRIsVALUE, state->endpoint_collection_enabled ? Qtrue : Qfalse));
   rb_str_concat(result, rb_sprintf(" timeline_enabled=%"PRIsVALUE, state->timeline_enabled ? Qtrue : Qfalse));
+  rb_str_concat(result, rb_sprintf(" otel_context_enabled=%d", state->otel_context_enabled));
   rb_str_concat(result, rb_sprintf(" allocation_type_enabled=%"PRIsVALUE, state->allocation_type_enabled ? Qtrue : Qfalse));
   rb_str_concat(result, rb_sprintf(
     " time_converter_state={.system_epoch_ns_reference=%ld, .delta_to_epoch_ns=%ld}",
@@ -969,6 +1120,7 @@ static VALUE _native_inspect(DDTRACE_UNUSED VALUE _self, VALUE collector_instanc
   rb_str_concat(result, rb_sprintf(" main_thread=%"PRIsVALUE, state->main_thread));
   rb_str_concat(result, rb_sprintf(" gc_tracking=%"PRIsVALUE, gc_tracking_as_ruby_hash(state)));
   rb_str_concat(result, rb_sprintf(" otel_current_span_key=%"PRIsVALUE, state->otel_current_span_key));
+  rb_str_concat(result, rb_sprintf(" global_waiting_for_gvl_threshold_ns=%u", global_waiting_for_gvl_threshold_ns));
 
   return result;
 }
@@ -996,6 +1148,10 @@ static int per_thread_context_as_ruby_hash(st_data_t key_thread, st_data_t value
 
     ID2SYM(rb_intern("gc_tracking.cpu_time_at_start_ns")),   /* => */ LONG2NUM(thread_context->gc_tracking.cpu_time_at_start_ns),
     ID2SYM(rb_intern("gc_tracking.wall_time_at_start_ns")),  /* => */ LONG2NUM(thread_context->gc_tracking.wall_time_at_start_ns),
+
+    #ifndef NO_GVL_INSTRUMENTATION
+      ID2SYM(rb_intern("gvl_waiting_at")), /* => */ LONG2NUM(gvl_profiling_state_thread_object_get(thread)),
+    #endif
   };
   for (long unsigned int i = 0; i < VALUE_COUNT(arguments); i += 2) rb_hash_aset(context_as_hash, arguments[i], arguments[i+1]);
 
@@ -1146,6 +1302,7 @@ static VALUE _native_gc_tracking(DDTRACE_UNUSED VALUE _self, VALUE collector_ins
 
 // Assumption 1: This function is called in a thread that is holding the Global VM Lock. Caller is responsible for enforcing this.
 static void trace_identifiers_for(struct thread_context_collector_state *state, VALUE thread, struct trace_identifiers *trace_identifiers_result) {
+  if (state->otel_context_enabled == otel_context_enabled_only) return;
   if (state->tracer_context_key == MISSING_TRACER_CONTEXT_KEY) return;
 
   VALUE current_context = rb_thread_local_aref(thread, state->tracer_context_key);
@@ -1200,7 +1357,7 @@ static bool should_collect_resource(VALUE root_span) {
   if (root_span_type == Qnil) return false;
   ENFORCE_TYPE(root_span_type, T_STRING);
 
-  int root_span_type_length = RSTRING_LEN(root_span_type);
+  long root_span_type_length = RSTRING_LEN(root_span_type);
   const char *root_span_type_value = StringValuePtr(root_span_type);
 
   bool is_web_request =
@@ -1223,6 +1380,9 @@ static VALUE _native_reset_after_fork(DDTRACE_UNUSED VALUE self, VALUE collector
   struct thread_context_collector_state *state;
   TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
 
+  // Release all context memory before clearing the existing context
+  st_foreach(state->hash_map_per_thread_context, hash_map_per_thread_context_free_values, 0 /* unused */);
+
   st_clear(state->hash_map_per_thread_context);
 
   state->stats = (struct stats) {}; // Resets all stats back to zero
@@ -1326,7 +1486,8 @@ void thread_context_collector_sample_allocation(VALUE self_instance, unsigned in
     (sample_values) {.alloc_samples = sample_weight, .alloc_samples_unscaled = 1, .heap_sample = true},
     INVALID_TIME, // For now we're not collecting timestamps for allocation events, as per profiling team internal discussions
     &ruby_vm_type,
-    optional_class_name
+    optional_class_name,
+    false
   );
 }
 
@@ -1372,25 +1533,29 @@ static ddog_CharSlice ruby_value_type_to_class_name(enum ruby_value_type type) {
   }
 }
 
+// Used to access OpenTelemetry::Trace.const_get(:CURRENT_SPAN_KEY). Will raise exceptions if it fails.
+static VALUE read_otel_current_span_key_const(DDTRACE_UNUSED VALUE _unused) {
+  VALUE opentelemetry_module = rb_const_get(rb_cObject, rb_intern("OpenTelemetry"));
+  ENFORCE_TYPE(opentelemetry_module, T_MODULE);
+  VALUE trace_module = rb_const_get(opentelemetry_module, rb_intern("Trace"));
+  ENFORCE_TYPE(trace_module, T_MODULE);
+  return rb_const_get(trace_module, rb_intern("CURRENT_SPAN_KEY"));
+}
+
 static VALUE get_otel_current_span_key(struct thread_context_collector_state *state) {
-  if (state->otel_current_span_key == Qnil) {
-    VALUE datadog_module = rb_const_get(rb_cObject, rb_intern("Datadog"));
-    VALUE opentelemetry_module = rb_const_get(datadog_module, rb_intern("OpenTelemetry"));
-    VALUE api_module = rb_const_get(opentelemetry_module, rb_intern("API"));
-    VALUE context_module = rb_const_get(api_module, rb_intern_const("Context"));
-    VALUE current_span_key = rb_const_get(context_module, rb_intern_const("CURRENT_SPAN_KEY"));
-
-    if (current_span_key == Qnil) {
-      rb_raise(rb_eRuntimeError, "Unexpected: Missing Datadog::OpenTelemetry::API::Context::CURRENT_SPAN_KEY");
-    }
+  if (state->otel_current_span_key == Qtrue) { // Qtrue means we haven't tried to extract it yet
+    // If this fails, we want to fail gracefully, rather than raise an exception (e.g. if the opentelemetry gem
+    // gets refactored, we should not fall on our face)
+    VALUE span_key = rb_protect(read_otel_current_span_key_const, Qnil, NULL);
 
-    state->otel_current_span_key = current_span_key;
+    // Note that this gets set to Qnil if we failed to extract the correct value, and thus we won't try to extract it again
+    state->otel_current_span_key = span_key;
   }
 
   return state->otel_current_span_key;
 }
 
-// This method gets used when ddtrace is being used indirectly via the otel APIs. Information gets stored slightly
+// This method gets used when ddtrace is being used indirectly via the opentelemetry APIs. Information gets stored slightly
 // differently, and this codepath handles it.
 static void ddtrace_otel_trace_identifiers_for(
   struct thread_context_collector_state *state,
@@ -1410,6 +1575,7 @@ static void ddtrace_otel_trace_identifiers_for(
   if (resolved_numeric_span_id == Qnil) return;
 
   VALUE otel_current_span_key = get_otel_current_span_key(state);
+  if (otel_current_span_key == Qnil) return;
   VALUE current_trace = *active_trace;
 
   // ddtrace uses a different structure when spans are created from otel, where each otel span will have a unique ddtrace
@@ -1462,3 +1628,388 @@ static VALUE _native_sample_skipped_allocation_samples(DDTRACE_UNUSED VALUE self
   thread_context_collector_sample_skipped_allocation_samples(collector_instance, NUM2UINT(skipped_samples));
   return Qtrue;
 }
+
+// This method differs from trace_identifiers_for/ddtrace_otel_trace_identifiers_for to support the situation where
+// the opentelemetry ruby library is being used for tracing AND the ddtrace tracing bits are not involved at all.
+//
+// Thus, in this case, we're directly reading from the opentelemetry stuff, which is different to how ddtrace tracing
+// does it.
+//
+// This is somewhat brittle: we're coupling on internal details of the opentelemetry gem to get what we need. In the
+// future maybe the otel ruby folks would be open to having a nice public way of getting this data that suits the
+// usecase of profilers.
+// Until then, the strategy below is to be extremely defensive, and if anything is out of place, we immediately return
+// and give up on getting trace data from opentelemetry. (Thus, worst case would be -- you upgrade opentelemetry and
+// profiling features relying on reading this data stop working, but you'll still get profiles and the app will be
+// otherwise undisturbed).
+//
+// Specifically, the way this works is:
+// 1. The latest entry in the opentelemetry context storage represents the current span (if any). We take the span id
+//    and trace id from this span.
+// 2. To find the local root span id, we walk the context storage backwards from the current span, and find the earliest
+//    entry in the context storage that has the same trace id as the current span; we use the found span as the local
+//    root span id.
+//    This matches the semantics of how ddtrace tracing creates a TraceOperation and assigns a local root span to it.
+static void otel_without_ddtrace_trace_identifiers_for(
+  struct thread_context_collector_state *state,
+  VALUE thread,
+  struct trace_identifiers *trace_identifiers_result
+) {
+  VALUE context_storage = rb_thread_local_aref(thread, otel_context_storage_id /* __opentelemetry_context_storage__ */);
+
+  // If it exists, context_storage is expected to be an Array[OpenTelemetry::Context]
+  if (context_storage == Qnil || !RB_TYPE_P(context_storage, T_ARRAY)) return;
+
+  VALUE otel_current_span_key = get_otel_current_span_key(state);
+  if (otel_current_span_key == Qnil) return;
+
+  int active_context_index = RARRAY_LEN(context_storage) - 1;
+  if (active_context_index < 0) return;
+
+  struct otel_span active_span = otel_span_from(rb_ary_entry(context_storage, active_context_index), otel_current_span_key);
+  if (active_span.span == Qnil) return;
+
+  struct otel_span local_root_span = active_span;
+
+  // Now find the oldest span starting from the active span that still has the same trace id as the active span
+  for (int i = active_context_index - 1; i >= 0; i--) {
+    struct otel_span checking_span = otel_span_from(rb_ary_entry(context_storage, i), otel_current_span_key);
+    if (checking_span.span == Qnil) return;
+
+    if (rb_str_equal(active_span.trace_id, checking_span.trace_id) == Qfalse) break;
+
+    local_root_span = checking_span;
+  }
+
+  // Convert the span ids into uint64_t to match what the Datadog tracer does
+  trace_identifiers_result->span_id = otel_span_id_to_uint(active_span.span_id);
+  trace_identifiers_result->local_root_span_id = otel_span_id_to_uint(local_root_span.span_id);
+
+  if (trace_identifiers_result->span_id == 0 || trace_identifiers_result->local_root_span_id == 0) return;
+
+  trace_identifiers_result->valid = true;
+
+  if (!state->endpoint_collection_enabled) return;
+
+  VALUE root_span_type = rb_ivar_get(local_root_span.span, at_kind_id /* @kind */);
+  // We filter out spans that don't have `kind: :server`
+  if (root_span_type == Qnil || !RB_TYPE_P(root_span_type, T_SYMBOL) || SYM2ID(root_span_type) != server_id) return;
+
+  VALUE trace_resource = rb_ivar_get(local_root_span.span, at_name_id /* @name */);
+  if (!RB_TYPE_P(trace_resource, T_STRING)) return;
+
+  trace_identifiers_result->trace_endpoint = trace_resource;
+}
+
+static struct otel_span otel_span_from(VALUE otel_context, VALUE otel_current_span_key) {
+  struct otel_span failed = {.span = Qnil, .span_id = Qnil, .trace_id = Qnil};
+
+  if (otel_context == Qnil) return failed;
+
+  VALUE context_entries = rb_ivar_get(otel_context, at_entries_id /* @entries */);
+  if (context_entries == Qnil || !RB_TYPE_P(context_entries, T_HASH)) return failed;
+
+  // If it exists, context_entries is expected to be a Hash[OpenTelemetry::Context::Key, OpenTelemetry::Trace::Span]
+  VALUE span = rb_hash_lookup(context_entries, otel_current_span_key);
+  if (span == Qnil) return failed;
+
+  // If it exists, span_context is expected to be a OpenTelemetry::Trace::SpanContext (don't confuse it with OpenTelemetry::Context)
+  VALUE span_context = rb_ivar_get(span, at_context_id /* @context */);
+  if (span_context == Qnil) return failed;
+
+  VALUE span_id = rb_ivar_get(span_context, at_span_id_id /* @span_id */);
+  VALUE trace_id = rb_ivar_get(span_context, at_trace_id_id /* @trace_id */);
+  if (span_id == Qnil || trace_id == Qnil || !RB_TYPE_P(span_id, T_STRING) || !RB_TYPE_P(trace_id, T_STRING)) return failed;
+
+  return (struct otel_span) {.span = span, .span_id = span_id, .trace_id = trace_id};
+}
+
+// Otel span ids are represented as a big-endian 8-byte string
+static uint64_t otel_span_id_to_uint(VALUE otel_span_id) {
+  if (!RB_TYPE_P(otel_span_id, T_STRING) || RSTRING_LEN(otel_span_id) != 8) { return 0; }
+
+  unsigned char *span_bytes = (unsigned char*) StringValuePtr(otel_span_id);
+
+  return \
+    ((uint64_t)span_bytes[0] << 56) |
+    ((uint64_t)span_bytes[1] << 48) |
+    ((uint64_t)span_bytes[2] << 40) |
+    ((uint64_t)span_bytes[3] << 32) |
+    ((uint64_t)span_bytes[4] << 24) |
+    ((uint64_t)span_bytes[5] << 16) |
+    ((uint64_t)span_bytes[6] <<  8) |
+    ((uint64_t)span_bytes[7]);
+}
+
+#ifndef NO_GVL_INSTRUMENTATION
+  // This function can get called from outside the GVL and even on non-main Ractors
+  void thread_context_collector_on_gvl_waiting(gvl_profiling_thread thread) {
+    // Because this function gets called from a thread that is NOT holding the GVL, we avoid touching the
+    // per-thread context directly.
+    //
+    // Instead, we ask Ruby to hold the data we need in Ruby's own special per-thread context area
+    // that's thread-safe and built for this kind of use
+    //
+    // Also, this function can get called on the non-main Ractor. We deal with this by checking if the value in the context
+    // is non-zero, since only `initialize_context` ever sets the value from 0 to non-zero for threads it sees.
+    intptr_t thread_being_profiled = gvl_profiling_state_get(thread);
+    if (!thread_being_profiled) return;
+
+    long current_monotonic_wall_time_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
+    if (current_monotonic_wall_time_ns <= 0 || current_monotonic_wall_time_ns > GVL_WAITING_ENABLED_EMPTY) return;
+
+    gvl_profiling_state_set(thread, current_monotonic_wall_time_ns);
+  }
+
+  // This function can get called from outside the GVL and even on non-main Ractors
+  __attribute__((warn_unused_result))
+  bool thread_context_collector_on_gvl_running_with_threshold(gvl_profiling_thread thread, uint32_t waiting_for_gvl_threshold_ns) {
+    intptr_t gvl_waiting_at = gvl_profiling_state_get(thread);
+
+    // Thread was not being profiled / not waiting on gvl
+    if (gvl_waiting_at == 0 || gvl_waiting_at == GVL_WAITING_ENABLED_EMPTY) return false;
+
+    // @ivoanjo: I'm not sure if this can happen -- It means we should've sampled already but haven't gotten the chance yet?
+    if (gvl_waiting_at < 0) return true;
+
+    long waiting_for_gvl_duration_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE) - gvl_waiting_at;
+
+    bool should_sample = waiting_for_gvl_duration_ns >= waiting_for_gvl_threshold_ns;
+
+    if (should_sample) {
+      // We flip the gvl_waiting_at to negative to mark that the thread is now running and no longer waiting
+      intptr_t gvl_waiting_at_is_now_running = -gvl_waiting_at;
+
+      gvl_profiling_state_set(thread, gvl_waiting_at_is_now_running);
+    } else {
+      // We decided not to sample. Let's mark the thread back to the initial "enabled but empty" state
+      gvl_profiling_state_set(thread, GVL_WAITING_ENABLED_EMPTY);
+    }
+
+    return should_sample;
+  }
+
+  __attribute__((warn_unused_result))
+  bool thread_context_collector_on_gvl_running(gvl_profiling_thread thread) {
+    return thread_context_collector_on_gvl_running_with_threshold(thread, global_waiting_for_gvl_threshold_ns);
+  }
+
+  // Why does this method need to exist?
+  //
+  // You may be surprised to see that if we never call this function (from cpu_and_wall_time_worker), Waiting for GVL
+  // samples will still show up.
+  // This is because regular cpu/wall-time samples also use `update_metrics_and_sample` which will do the right thing
+  // and push "Waiting for GVL" samples as needed.
+  //
+  // The reason this method needs to exist and be called very shortly after thread_context_collector_on_gvl_running
+  // returning true is to ensure accuracy of both the timing and stack for the Waiting for GVL sample.
+  //
+  // Timing:
+  // Because we currently only record the timestamp when the Waiting for GVL started and not when the Waiting for GVL ended,
+  // we rely on pushing a sample as soon as possible when the Waiting for GVL ends so that the timestamp of the sample
+  // actually matches when we stopped waiting.
+  //
+  // Stack:
+  // If the thread starts working without the end of the Waiting for GVL sample, then by the time the thread is sampled
+  // via the regular cpu/wall-time samples mechanism, the stack can be be inaccurate (e.g. does not correctly pinpoint
+  // where the waiting happened).
+  //
+  // Arguably, the last sample after Waiting for GVL ended (when gvl_waiting_at < 0) should always come from this method
+  // and not a regular cpu/wall-time sample BUT since all of these things are happening in parallel/concurrently I suspect
+  // it's possible for a regular sample to kick in just before this one.
+  //
+  // ---
+  //
+  // NOTE: In normal use, current_thread is expected to be == rb_thread_current(); the `current_thread` parameter only
+  // exists to enable testing.
+  VALUE thread_context_collector_sample_after_gvl_running_with_thread(VALUE self_instance, VALUE current_thread) {
+    struct thread_context_collector_state *state;
+    TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+    if (!state->timeline_enabled) rb_raise(rb_eRuntimeError, "GVL profiling requires timeline to be enabled");
+
+    intptr_t gvl_waiting_at = gvl_profiling_state_thread_object_get(current_thread);
+
+    if (gvl_waiting_at >= 0) {
+      // @ivoanjo: I'm not sure if this can ever happen. This means that we're not on the same thread
+      // that ran `thread_context_collector_on_gvl_running` and made the decision to sample OR a regular sample was
+      // triggered ahead of us.
+      // We do nothing in this case.
+      return Qfalse;
+    }
+
+    struct per_thread_context *thread_context = get_or_create_context_for(current_thread, state);
+
+    // We don't actually account for cpu-time during Waiting for GVL. BUT, we may chose to push an
+    // extra sample to represent the period prior to Waiting for GVL. To support that, we retrieve the current
+    // cpu-time of the thread and let `update_metrics_and_sample` decide what to do with it.
+    long cpu_time_for_thread = cpu_time_now_ns(thread_context);
+
+    // TODO: Should we update the dynamic sampling rate overhead tracking with this sample as well?
+
+    update_metrics_and_sample(
+      state,
+      /* thread_being_sampled: */ current_thread,
+      /* stack_from_thread: */ current_thread,
+      thread_context,
+      thread_context->sampling_buffer,
+      cpu_time_for_thread,
+      monotonic_wall_time_now_ns(RAISE_ON_FAILURE)
+    );
+
+    return Qtrue; // To allow this to be called from rb_rescue2
+  }
+
+  VALUE thread_context_collector_sample_after_gvl_running(VALUE self_instance) {
+    return thread_context_collector_sample_after_gvl_running_with_thread(self_instance, rb_thread_current());
+  }
+
+  // This method is intended to be called from update_metrics_and_sample. It exists to handle extra sampling steps we
+  // need to take when sampling cpu/wall-time for a thread that's in the "Waiting for GVL" state.
+  __attribute__((warn_unused_result))
+  static bool handle_gvl_waiting(
+    struct thread_context_collector_state *state,
+    VALUE thread_being_sampled,
+    VALUE stack_from_thread,
+    struct per_thread_context *thread_context,
+    sampling_buffer* sampling_buffer,
+    long current_cpu_time_ns
+  ) {
+    intptr_t gvl_waiting_at = gvl_profiling_state_thread_object_get(thread_being_sampled);
+
+    bool is_gvl_waiting_state = gvl_waiting_at != 0 && gvl_waiting_at != GVL_WAITING_ENABLED_EMPTY;
+
+    if (!is_gvl_waiting_state) return false;
+
+    // We can be in one of 2 situations here:
+    //
+    // 1. The current sample is the first one after we entered the "Waiting for GVL" state
+    //    (wall_time_at_previous_sample_ns < abs(gvl_waiting_at))
+    //
+    //                             time ─────►
+    //  ...──────────────┬───────────────────...
+    //       Other state │ Waiting for GVL
+    //  ...──────────────┴───────────────────...
+    //     ▲                              ▲
+    //     └─ Previous sample              └─ Regular sample (caller)
+    //
+    //   In this case, we'll want to push two samples: a) one for the current time (handled by the caller), b) an extra sample
+    //   to represent the remaining cpu/wall time before the "Waiting for GVL" started:
+    //
+    //                             time ─────►
+    //  ...──────────────┬───────────────────...
+    //       Other state │ Waiting for GVL
+    //  ...──────────────┴───────────────────...
+    //     ▲            ▲                ▲
+    //     └─ Prev...    └─ Extra sample   └─ Regular sample (caller)
+    //
+    // 2. The current sample is the n-th one after we entered the "Waiting for GVL" state
+    //    (wall_time_at_previous_sample_ns > abs(gvl_waiting_at))
+    //
+    //                             time ─────►
+    //  ...──────────────┬───────────────────────────────────────────────...
+    //       Other state │ Waiting for GVL
+    //  ...──────────────┴───────────────────────────────────────────────...
+    //     ▲                              ▲                          ▲
+    //     └─ Previous sample              └─ Previous sample         └─ Regular sample (caller)
+    //
+    //   In this case, we just report back to the caller that the thread is in the "Waiting for GVL" state.
+    //
+    // ---
+    //
+    // Overall, gvl_waiting_at will be > 0 if still in the "Waiting for GVL" state and < 0 if we actually reached the end of
+    // the wait.
+    //
+    // It doesn't really matter if the thread is still waiting or just reached the end of the wait: each sample represents
+    // a snapshot at time ending now, so if the state finished, it just means the next sample will be a regular one.
+
+    if (gvl_waiting_at < 0) {
+      // Negative means the waiting for GVL just ended, so we clear the state, so next samples no longer represent waiting
+      gvl_profiling_state_thread_object_set(thread_being_sampled, GVL_WAITING_ENABLED_EMPTY);
+    }
+
+    long gvl_waiting_started_wall_time_ns = labs(gvl_waiting_at);
+
+    if (thread_context->wall_time_at_previous_sample_ns < gvl_waiting_started_wall_time_ns) { // situation 1 above
+      long cpu_time_elapsed_ns = update_time_since_previous_sample(
+        &thread_context->cpu_time_at_previous_sample_ns,
+        current_cpu_time_ns,
+        thread_context->gc_tracking.cpu_time_at_start_ns,
+        IS_NOT_WALL_TIME
+      );
+
+      long duration_until_start_of_gvl_waiting_ns = update_time_since_previous_sample(
+        &thread_context->wall_time_at_previous_sample_ns,
+        gvl_waiting_started_wall_time_ns,
+        INVALID_TIME,
+        IS_WALL_TIME
+      );
+
+      // Push extra sample
+      trigger_sample_for_thread(
+        state,
+        thread_being_sampled,
+        stack_from_thread,
+        thread_context,
+        sampling_buffer,
+        (sample_values) {.cpu_time_ns = cpu_time_elapsed_ns, .cpu_or_wall_samples = 1, .wall_time_ns = duration_until_start_of_gvl_waiting_ns},
+        gvl_waiting_started_wall_time_ns,
+        NULL,
+        NULL,
+        false // This is the extra sample before the wait begun; only the next sample will be in the gvl waiting state
+      );
+    }
+
+    return true;
+  }
+
+  static VALUE _native_on_gvl_waiting(DDTRACE_UNUSED VALUE self, VALUE thread) {
+    ENFORCE_THREAD(thread);
+
+    thread_context_collector_on_gvl_waiting(thread_from_thread_object(thread));
+    return Qnil;
+  }
+
+  static VALUE _native_gvl_waiting_at_for(DDTRACE_UNUSED VALUE self, VALUE thread) {
+    ENFORCE_THREAD(thread);
+
+    intptr_t gvl_waiting_at = gvl_profiling_state_thread_object_get(thread);
+    return LONG2NUM(gvl_waiting_at);
+  }
+
+  static VALUE _native_on_gvl_running(DDTRACE_UNUSED VALUE self, VALUE thread) {
+    ENFORCE_THREAD(thread);
+
+    return thread_context_collector_on_gvl_running(thread_from_thread_object(thread)) ? Qtrue : Qfalse;
+  }
+
+  static VALUE _native_sample_after_gvl_running(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE thread) {
+    ENFORCE_THREAD(thread);
+
+    return thread_context_collector_sample_after_gvl_running_with_thread(collector_instance, thread);
+  }
+
+  static VALUE _native_apply_delta_to_cpu_time_at_previous_sample_ns(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE thread, VALUE delta_ns) {
+    ENFORCE_THREAD(thread);
+
+    struct thread_context_collector_state *state;
+    TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+    struct per_thread_context *thread_context = get_context_for(thread, state);
+    if (thread_context == NULL) rb_raise(rb_eArgError, "Unexpected: This method cannot be used unless the per-thread context for the thread already exists");
+
+    thread_context->cpu_time_at_previous_sample_ns += NUM2LONG(delta_ns);
+
+    return Qtrue;
+  }
+
+#else
+  static bool handle_gvl_waiting(
+    DDTRACE_UNUSED struct thread_context_collector_state *state,
+    DDTRACE_UNUSED VALUE thread_being_sampled,
+    DDTRACE_UNUSED VALUE stack_from_thread,
+    DDTRACE_UNUSED struct per_thread_context *thread_context,
+    DDTRACE_UNUSED sampling_buffer* sampling_buffer,
+    DDTRACE_UNUSED long current_cpu_time_ns
+  ) { return false; }
+#endif // NO_GVL_INSTRUMENTATION