ddtrace 1.4.2 → 1.19.0

data/ext/ddtrace_profiling_native_extension/collectors_thread_context.c

@@ -0,0 +1,1301 @@
+#include <ruby.h>
+
+#include "collectors_thread_context.h"
+#include "clock_id.h"
+#include "collectors_stack.h"
+#include "collectors_gc_profiling_helper.h"
+#include "helpers.h"
+#include "libdatadog_helpers.h"
+#include "private_vm_api_access.h"
+#include "stack_recorder.h"
+#include "time_helpers.h"
+
+// Used to trigger sampling of threads, based on external "events", such as:
+// * periodic timer for cpu-time and wall-time
+// * VM garbage collection events
+// * VM object allocation events
+//
+// This file implements the native bits of the Datadog::Profiling::Collectors::ThreadContext class
+//
+// Triggering of this component (e.g. watching for the above "events") is implemented by Collectors::CpuAndWallTimeWorker.
+
+// ---
+// ## Tracking of cpu-time and wall-time spent during garbage collection
+//
+// This feature works by having a special state that a thread can be in: doing garbage collection. This state is
+// tracked inside the thread's `per_thread_context.gc_tracking` data, and three functions, listed below. The functions
+// will get called by the `Collectors::CpuAndWallTimeWorker` at very specific times in the VM lifetime.
+//
+// * `thread_context_collector_on_gc_start`: Called at the very beginning of the garbage collection process.
+//   The internal VM `during_gc` flag is set to `true`, but Ruby has not done any work yet.
+// * `thread_context_collector_on_gc_finish`: Called at the very end of the garbage collection process.
+//   The internal VM `during_gc` flag is still set to `true`, but all the work has been done.
+// * `thread_context_collector_sample_after_gc`: Called shortly after the garbage collection process.
+//   The internal VM `during_gc` flag is set to `false`.
+//
+// Inside this component, here's what happens inside those three functions:
+//
+// When `thread_context_collector_on_gc_start` gets called, the current cpu and wall-time get recorded to the thread
+// context: `cpu_time_at_gc_start_ns` and `wall_time_at_gc_start_ns`.
+//
+// While `cpu_time_at_gc_start_ns` is set, regular samples (if any) do not account for cpu-time any time that passes
+// after this timestamp. The idea is that this cpu-time will be blamed separately on GC, and not on the user thread.
+// Wall-time accounting is not affected by this (e.g. we still record 60 seconds every 60 seconds).
+//
+// (Regular samples can still account for the cpu-time between the previous sample and the start of GC.)
+//
+// When `thread_context_collector_on_gc_finish` gets called, the cpu-time and wall-time spent during GC gets recorded
+// into the global gc_tracking structure, and further samples are not affected. (The `cpu_time_at_previous_sample_ns`
+// of the thread that did GC also gets adjusted to avoid double-accounting.)
+//
+// Finally, when `thread_context_collector_sample_after_gc` gets called, a sample gets recorded with a stack having
+// a single placeholder `Garbage Collection` frame. This sample gets
+// assigned the cpu-time and wall-time that was recorded between calls to `on_gc_start` and `on_gc_finish`, as well
+// as metadata for the last GC.
+//
+// Note that the Ruby GC does not usually do all of the GC work in one go. Instead, it breaks it up into smaller steps
+// so that the application can keep doing user work in between GC steps.
+// The `on_gc_start` / `on_gc_finish` will trigger each time the VM executes these smaller steps, and on a benchmark
+// that executes `Object.new` in a loop, I measured more than 50k of this steps per second (!!).
+// Creating these many events for every GC step is a lot of overhead, so instead `on_gc_finish` coalesces time
+// spent in GC and only flushes it at most every 10 ms/every complete GC collection. This reduces the amount of
+// individual GC events we need to record. We use the latest GC metadata for this event, reflecting the last GC that
+// happened in the coalesced period.
+//
+// In an earlier attempt at implementing this functionality (https://github.com/DataDog/dd-trace-rb/pull/2308), we
+// discovered that we needed to factor the sampling work away from `thread_context_collector_on_gc_finish` and into a
+// separate `thread_context_collector_sample_after_gc` because (as documented in more detail below),
+// `sample_after_gc` could trigger memory allocation in rare occasions (usually exceptions), which is actually not
+// allowed to happen during Ruby's garbage collection start/finish hooks.
+// ---
+
+#define THREAD_ID_LIMIT_CHARS 44 // Why 44? "#{2**64} (#{2**64})".size + 1 for \0
+#define THREAD_INVOKE_LOCATION_LIMIT_CHARS 512
+#define IS_WALL_TIME true
+#define IS_NOT_WALL_TIME false
+#define MISSING_TRACER_CONTEXT_KEY 0
+#define TIME_BETWEEN_GC_EVENTS_NS MILLIS_AS_NS(10)
+
+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
+static ID at_resource_id;     // id of :@resource in Ruby
+static ID at_root_span_id;    // id of :@root_span in Ruby
+static ID at_type_id;         // id of :@type in Ruby
+
+// 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
+  // "Update this when modifying state struct"
+
+  // Required by Datadog::Profiling::Collectors::Stack as a scratch buffer during sampling
+  sampling_buffer *sampling_buffer;
+  // Hashmap <Thread Object, struct per_thread_context>
+  st_table *hash_map_per_thread_context;
+  // Datadog::Profiling::StackRecorder instance
+  VALUE recorder_instance;
+  // If the tracer is available and enabled, this will be the fiber-local symbol for accessing its running context,
+  // to enable code hotspots and endpoint aggregation.
+  // When not available, this is set to MISSING_TRACER_CONTEXT_KEY.
+  ID tracer_context_key;
+  // Track how many regular samples we've taken. Does not include garbage collection samples.
+  // Currently **outside** of stats struct because we also use it to decide when to clean the contexts, and thus this
+  // is not (just) a stat.
+  unsigned int sample_count;
+  // Reusable array to get list of threads
+  VALUE thread_list_buffer;
+  // Used to omit endpoint names (retrieved from tracer) from collected data
+  bool endpoint_collection_enabled;
+  // Used to omit timestamps / timeline events from collected data
+  bool timeline_enabled;
+  // Used to omit class information from collected allocation data
+  bool allocation_type_enabled;
+  // Used when calling monotonic_to_system_epoch_ns
+  monotonic_to_system_epoch_state time_converter_state;
+  // Used to identify the main thread, to give it a fallback name
+  VALUE main_thread;
+
+  struct stats {
+    // Track how many garbage collection samples we've taken.
+    unsigned int gc_samples;
+    // See thread_context_collector_on_gc_start for details
+    unsigned int gc_samples_missed_due_to_missing_context;
+  } stats;
+
+  struct {
+    unsigned long accumulated_cpu_time_ns;
+    unsigned long accumulated_wall_time_ns;
+
+    long wall_time_at_previous_gc_ns; // Will be INVALID_TIME unless there's accumulated time above
+    long wall_time_at_last_flushed_gc_event_ns; // Starts at 0 and then will always be valid
+  } gc_tracking;
+};
+
+// Tracks per-thread state
+struct per_thread_context {
+  char thread_id[THREAD_ID_LIMIT_CHARS];
+  ddog_CharSlice thread_id_char_slice;
+  char thread_invoke_location[THREAD_INVOKE_LOCATION_LIMIT_CHARS];
+  ddog_CharSlice thread_invoke_location_char_slice;
+  thread_cpu_time_id thread_cpu_time_id;
+  long cpu_time_at_previous_sample_ns;  // Can be INVALID_TIME until initialized or if getting it fails for another reason
+  long wall_time_at_previous_sample_ns; // Can be INVALID_TIME until initialized
+
+  struct {
+    // Both of these fields are set by on_gc_start and kept until on_gc_finish is called.
+    // Outside of this window, they will be INVALID_TIME.
+    long cpu_time_at_start_ns;
+    long wall_time_at_start_ns;
+  } gc_tracking;
+};
+
+// Used to correlate profiles with traces
+struct trace_identifiers {
+  bool valid;
+  uint64_t local_root_span_id;
+  uint64_t span_id;
+  VALUE trace_endpoint;
+};
+
+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);
+static int hash_map_per_thread_context_free_values(st_data_t _thread, st_data_t value_per_thread_context, st_data_t _argument);
+static VALUE _native_new(VALUE klass);
+static VALUE _native_initialize(
+  VALUE self,
+  VALUE collector_instance,
+  VALUE recorder_instance,
+  VALUE max_frames,
+  VALUE tracer_context_key,
+  VALUE endpoint_collection_enabled,
+  VALUE timeline_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(
+  struct thread_context_collector_state *state,
+  VALUE thread_being_sampled,
+  VALUE profiler_overhead_stack_thread,
+  struct per_thread_context *thread_context,
+  long current_cpu_time_ns,
+  long current_monotonic_wall_time_ns
+);
+static void trigger_sample_for_thread(
+  struct thread_context_collector_state *state,
+  VALUE thread,
+  VALUE stack_from_thread,
+  struct per_thread_context *thread_context,
+  sample_values values,
+  long current_monotonic_wall_time_ns,
+  ddog_CharSlice *ruby_vm_type,
+  ddog_CharSlice *class_name
+);
+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);
+static struct per_thread_context *get_context_for(VALUE thread, struct thread_context_collector_state *state);
+static void initialize_context(VALUE thread, struct per_thread_context *thread_context, struct thread_context_collector_state *state);
+static VALUE _native_inspect(VALUE self, VALUE collector_instance);
+static VALUE per_thread_context_st_table_as_ruby_hash(struct thread_context_collector_state *state);
+static int per_thread_context_as_ruby_hash(st_data_t key_thread, st_data_t value_context, st_data_t result_hash);
+static VALUE stats_as_ruby_hash(struct thread_context_collector_state *state);
+static VALUE gc_tracking_as_ruby_hash(struct thread_context_collector_state *state);
+static void remove_context_for_dead_threads(struct thread_context_collector_state *state);
+static int remove_if_dead_thread(st_data_t key_thread, st_data_t value_context, st_data_t _argument);
+static VALUE _native_per_thread_context(VALUE self, VALUE collector_instance);
+static long update_time_since_previous_sample(long *time_at_previous_sample_ns, long current_time_ns, long gc_start_time_ns, bool is_wall_time);
+static long cpu_time_now_ns(struct per_thread_context *thread_context);
+static long thread_id_for(VALUE thread);
+static VALUE _native_stats(VALUE self, VALUE collector_instance);
+static VALUE _native_gc_tracking(VALUE self, VALUE collector_instance);
+static void trace_identifiers_for(struct thread_context_collector_state *state, VALUE thread, struct trace_identifiers *trace_identifiers_result);
+static bool should_collect_resource(VALUE root_span_type);
+static VALUE _native_reset_after_fork(DDTRACE_UNUSED VALUE self, VALUE collector_instance);
+static VALUE thread_list(struct thread_context_collector_state *state);
+static VALUE _native_sample_allocation(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE sample_weight, VALUE new_object);
+static VALUE _native_new_empty_thread(VALUE self);
+static ddog_CharSlice ruby_value_type_to_class_name(enum ruby_value_type type);
+
+void collectors_thread_context_init(VALUE profiling_module) {
+  VALUE collectors_module = rb_define_module_under(profiling_module, "Collectors");
+  VALUE collectors_thread_context_class = rb_define_class_under(collectors_module, "ThreadContext", rb_cObject);
+  // Hosts methods used for testing the native code using RSpec
+  VALUE testing_module = rb_define_module_under(collectors_thread_context_class, "Testing");
+
+  // Instances of the ThreadContext class are "TypedData" objects.
+  // "TypedData" objects are special objects in the Ruby VM that can wrap C structs.
+  // In this case, it wraps the thread_context_collector_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_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_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_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);
+
+  at_active_span_id = rb_intern_const("@active_span");
+  at_active_trace_id = rb_intern_const("@active_trace");
+  at_id_id = rb_intern_const("@id");
+  at_resource_id = rb_intern_const("@resource");
+  at_root_span_id = rb_intern_const("@root_span");
+  at_type_id = rb_intern_const("@type");
+
+  gc_profiling_init();
+}
+
+// This structure is used to define a Ruby object that stores a pointer to a struct thread_context_collector_state
+// See also https://github.com/ruby/ruby/blob/master/doc/extension.rdoc for how this works
+static const rb_data_type_t thread_context_collector_typed_data = {
+  .wrap_struct_name = "Datadog::Profiling::Collectors::ThreadContext",
+  .function = {
+    .dmark = thread_context_collector_typed_data_mark,
+    .dfree = thread_context_collector_typed_data_free,
+    .dsize = NULL, // We don't track profile memory usage (although it'd be cool if we did!)
+    //.dcompact = NULL, // FIXME: Add support for compaction
+  },
+  .flags = RUBY_TYPED_FREE_IMMEDIATELY
+};
+
+// This function is called by the Ruby GC to give us a chance to mark any Ruby objects that we're holding on to,
+// so that they don't get garbage collected
+static void thread_context_collector_typed_data_mark(void *state_ptr) {
+  struct thread_context_collector_state *state = (struct thread_context_collector_state *) state_ptr;
+
+  // Update this when modifying state struct
+  rb_gc_mark(state->recorder_instance);
+  st_foreach(state->hash_map_per_thread_context, hash_map_per_thread_context_mark, 0 /* unused */);
+  rb_gc_mark(state->thread_list_buffer);
+  rb_gc_mark(state->main_thread);
+}
+
+static void thread_context_collector_typed_data_free(void *state_ptr) {
+  struct thread_context_collector_state *state = (struct thread_context_collector_state *) state_ptr;
+
+  // Update this when modifying state struct
+
+  // Important: Remember that we're only guaranteed to see here what's been set in _native_new, aka
+  // pointers that have been set NULL there may still be NULL here.
+  if (state->sampling_buffer != NULL) sampling_buffer_free(state->sampling_buffer);
+
+  // Free each entry in the map
+  st_foreach(state->hash_map_per_thread_context, hash_map_per_thread_context_free_values, 0 /* unused */);
+  // ...and then the map
+  st_free_table(state->hash_map_per_thread_context);
+
+  ruby_xfree(state);
+}
+
+// Mark Ruby thread references we keep as keys in hash_map_per_thread_context
+static int hash_map_per_thread_context_mark(st_data_t key_thread, DDTRACE_UNUSED st_data_t _value, DDTRACE_UNUSED st_data_t _argument) {
+  VALUE thread = (VALUE) key_thread;
+  rb_gc_mark(thread);
+  return ST_CONTINUE;
+}
+
+// Used to clear each of the per_thread_contexts inside the hash_map_per_thread_context
+static int hash_map_per_thread_context_free_values(DDTRACE_UNUSED st_data_t _thread, st_data_t value_per_thread_context, DDTRACE_UNUSED st_data_t _argument) {
+  struct per_thread_context *per_thread_context = (struct per_thread_context*) value_per_thread_context;
+  ruby_xfree(per_thread_context);
+  return ST_CONTINUE;
+}
+
+static VALUE _native_new(VALUE klass) {
+  struct thread_context_collector_state *state = ruby_xcalloc(1, sizeof(struct thread_context_collector_state));
+
+  // Note: Any exceptions raised from this note until the TypedData_Wrap_Struct call will lead to the state memory
+  // being leaked.
+
+  // Update this when modifying state struct
+  state->sampling_buffer = NULL;
+  state->hash_map_per_thread_context =
+   // "numtable" is an awful name, but TL;DR it's what should be used when keys are `VALUE`s.
+    st_init_numtable();
+  state->recorder_instance = Qnil;
+  state->tracer_context_key = MISSING_TRACER_CONTEXT_KEY;
+  state->thread_list_buffer = rb_ary_new();
+  state->endpoint_collection_enabled = true;
+  state->timeline_enabled = true;
+  state->allocation_type_enabled = true;
+  state->time_converter_state = (monotonic_to_system_epoch_state) MONOTONIC_TO_SYSTEM_EPOCH_INITIALIZER;
+  state->main_thread = rb_thread_main();
+  state->gc_tracking.wall_time_at_previous_gc_ns = INVALID_TIME;
+  state->gc_tracking.wall_time_at_last_flushed_gc_event_ns = 0;
+
+  return TypedData_Wrap_Struct(klass, &thread_context_collector_typed_data, state);
+}
+
+static VALUE _native_initialize(
+  DDTRACE_UNUSED VALUE _self,
+  VALUE collector_instance,
+  VALUE recorder_instance,
+  VALUE max_frames,
+  VALUE tracer_context_key,
+  VALUE endpoint_collection_enabled,
+  VALUE timeline_enabled,
+  VALUE allocation_type_enabled
+) {
+  ENFORCE_BOOLEAN(endpoint_collection_enabled);
+  ENFORCE_BOOLEAN(timeline_enabled);
+  ENFORCE_BOOLEAN(allocation_type_enabled);
+
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  int max_frames_requested = NUM2INT(max_frames);
+  if (max_frames_requested < 0) rb_raise(rb_eArgError, "Invalid max_frames: value must not be negative");
+
+  // Update this when modifying state struct
+  state->sampling_buffer = sampling_buffer_new(max_frames_requested);
+  // hash_map_per_thread_context is already initialized, nothing to do here
+  state->recorder_instance = enforce_recorder_instance(recorder_instance);
+  state->endpoint_collection_enabled = (endpoint_collection_enabled == Qtrue);
+  state->timeline_enabled = (timeline_enabled == Qtrue);
+  state->allocation_type_enabled = (allocation_type_enabled == Qtrue);
+
+  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.
+    // In this case, we can't really escape this because as of this writing, ruby master still calls `rb_to_id` inside
+    // the implementation of Thread#[]= so any symbol that gets used as a key there will already be prevented from GC.
+    state->tracer_context_key = rb_to_id(tracer_context_key);
+  }
+
+  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(DDTRACE_UNUSED VALUE _self, VALUE collector_instance, VALUE profiler_overhead_stack_thread) {
+  if (!is_thread_alive(profiler_overhead_stack_thread)) rb_raise(rb_eArgError, "Unexpected: profiler_overhead_stack_thread is not alive");
+
+  thread_context_collector_sample(collector_instance, monotonic_wall_time_now_ns(RAISE_ON_FAILURE), profiler_overhead_stack_thread);
+  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_on_gc_start(DDTRACE_UNUSED VALUE self, VALUE collector_instance) {
+  thread_context_collector_on_gc_start(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_on_gc_finish(DDTRACE_UNUSED VALUE self, VALUE collector_instance) {
+  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) {
+  thread_context_collector_sample_after_gc(collector_instance);
+  return Qtrue;
+}
+
+// This function gets called from the Collectors::CpuAndWallTimeWorker to trigger the actual sampling.
+//
+// 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 allowed to raise exceptions. Caller is responsible for handling them, if needed.
+// Assumption 3: This function IS NOT called from a signal handler. This function is not async-signal-safe.
+// Assumption 4: This function IS NOT called in a reentrant way.
+// Assumption 5: This function is called from the main Ractor (if Ruby has support for Ractors).
+//
+// The `profiler_overhead_stack_thread` is used to attribute the profiler overhead to a stack borrowed from a different thread
+// (belonging to ddtrace), so that the overhead is visible in the profile rather than blamed on user code.
+void thread_context_collector_sample(VALUE self_instance, long current_monotonic_wall_time_ns, VALUE profiler_overhead_stack_thread) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  VALUE current_thread = rb_thread_current();
+  struct per_thread_context *current_thread_context = get_or_create_context_for(current_thread, state);
+  long cpu_time_at_sample_start_for_current_thread = cpu_time_now_ns(current_thread_context);
+
+  VALUE threads = thread_list(state);
+
+  const long thread_count = RARRAY_LEN(threads);
+  for (long i = 0; i < thread_count; i++) {
+    VALUE thread = RARRAY_AREF(threads, i);
+    struct per_thread_context *thread_context = get_or_create_context_for(thread, state);
+
+    // We account for cpu-time for the current thread in a different way -- we use the cpu-time at sampling start, to avoid
+    // blaming the time the profiler took on whatever's running on the thread right now
+    long current_cpu_time_ns = thread != current_thread ? cpu_time_now_ns(thread_context) : cpu_time_at_sample_start_for_current_thread;
+
+    update_metrics_and_sample(
+      state,
+      /* thread_being_sampled: */ thread,
+      /* stack_from_thread: */ thread,
+      thread_context,
+      current_cpu_time_ns,
+      current_monotonic_wall_time_ns
+    );
+  }
+
+  state->sample_count++;
+
+  // TODO: This seems somewhat overkill and inefficient to do often; right now we just do it every few samples
+  // but there's probably a better way to do this if we actually track when threads finish
+  if (state->sample_count % 100 == 0) remove_context_for_dead_threads(state);
+
+  update_metrics_and_sample(
+    state,
+    /* thread_being_sampled: */ current_thread,
+    /* stack_from_thread: */ profiler_overhead_stack_thread,
+    current_thread_context,
+    cpu_time_now_ns(current_thread_context),
+    monotonic_wall_time_now_ns(RAISE_ON_FAILURE)
+  );
+}
+
+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
+  struct per_thread_context *thread_context,
+  long current_cpu_time_ns,
+  long current_monotonic_wall_time_ns
+) {
+  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 wall_time_elapsed_ns = update_time_since_previous_sample(
+    &thread_context->wall_time_at_previous_sample_ns,
+    current_monotonic_wall_time_ns,
+    // We explicitly pass in `INVALID_TIME` as an argument for `gc_start_time_ns` here because we don't want wall-time
+    // accounting to change during GC.
+    // E.g. if 60 seconds pass in the real world, 60 seconds of wall-time are recorded, regardless of the thread doing
+    // GC or not.
+    INVALID_TIME,
+    IS_WALL_TIME
+  );
+
+  trigger_sample_for_thread(
+    state,
+    thread_being_sampled,
+    stack_from_thread,
+    thread_context,
+    (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
+  );
+}
+
+// This function gets called when Ruby is about to start running the Garbage Collector on the current thread.
+// It updates the per_thread_context of the current thread to include the current cpu/wall times, to be used to later
+// create an event including the cpu/wall time spent in garbage collector work.
+//
+// Safety: This function gets called while Ruby is doing garbage collection. While Ruby is doing garbage collection,
+// *NO ALLOCATION* is allowed. This function, and any it calls must never trigger memory or object allocation.
+// This includes exceptions and use of ruby_xcalloc (because xcalloc can trigger GC)!
+//
+// 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).
+void thread_context_collector_on_gc_start(VALUE self_instance) {
+  struct thread_context_collector_state *state;
+  if (!rb_typeddata_is_kind_of(self_instance, &thread_context_collector_typed_data)) return;
+  // This should never fail the the above check passes
+  TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  struct per_thread_context *thread_context = get_context_for(rb_thread_current(), state);
+
+  // If there was no previously-existing context for this thread, we won't allocate one (see safety). For now we just drop
+  // the GC sample, under the assumption that "a thread that is so new that we never sampled it even once before it triggers
+  // GC" is a rare enough case that we can just ignore it.
+  // We can always improve this later if we find that this happens often (and we have the counter to help us figure that out)!
+  if (thread_context == NULL) {
+    state->stats.gc_samples_missed_due_to_missing_context++;
+    return;
+  }
+
+  // Here we record the wall-time first and in on_gc_finish we record it second to try to avoid having wall-time be slightly < cpu-time
+  thread_context->gc_tracking.wall_time_at_start_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
+  thread_context->gc_tracking.cpu_time_at_start_ns = cpu_time_now_ns(thread_context);
+}
+
+// This function gets called when Ruby has finished running the Garbage Collector on the current thread.
+// It records the cpu/wall-time observed during GC, which will be used to later
+// create an event including the cpu/wall time spent from the start of garbage collector work until now.
+//
+// Safety: This function gets called while Ruby is doing garbage collection. While Ruby is doing garbage collection,
+// *NO ALLOCATION* is allowed. This function, and any it calls must never trigger memory or object allocation.
+// This includes exceptions and use of ruby_xcalloc (because xcalloc can trigger GC)!
+//
+// 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).
+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;
+  // This should never fail the the above check passes
+  TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  struct per_thread_context *thread_context = get_context_for(rb_thread_current(), state);
+
+  // If there was no previously-existing context for this thread, we won't allocate one (see safety). We keep a metric for
+  // how often this happens -- see on_gc_start.
+  if (thread_context == NULL) return false;
+
+  long cpu_time_at_start_ns = thread_context->gc_tracking.cpu_time_at_start_ns;
+  long wall_time_at_start_ns = thread_context->gc_tracking.wall_time_at_start_ns;
+
+  if (cpu_time_at_start_ns == INVALID_TIME && wall_time_at_start_ns == INVALID_TIME) {
+    // If this happened, it means that on_gc_start was either never called for the thread OR it was called but no thread
+    // context existed at the time. The former can be the result of a bug, but since we can't distinguish them, we just
+    // do nothing.
+    return false;
+  }
+
+  // Mark thread as no longer in GC
+  thread_context->gc_tracking.cpu_time_at_start_ns = INVALID_TIME;
+  thread_context->gc_tracking.wall_time_at_start_ns = INVALID_TIME;
+
+  // Here we record the wall-time second and in on_gc_start we record it first to try to avoid having wall-time be slightly < cpu-time
+  long cpu_time_at_finish_ns = cpu_time_now_ns(thread_context);
+  long wall_time_at_finish_ns = monotonic_wall_time_now_ns(DO_NOT_RAISE_ON_FAILURE);
+
+  // If our end timestamp is not OK, we bail out
+  if (wall_time_at_finish_ns == 0) return false;
+
+  long gc_cpu_time_elapsed_ns = cpu_time_at_finish_ns - cpu_time_at_start_ns;
+  long gc_wall_time_elapsed_ns = wall_time_at_finish_ns - wall_time_at_start_ns;
+
+  // Wall-time can go backwards if the system clock gets changed (and we observed spurious jumps back on macOS as well)
+  // so let's ensure we don't get negative values for time deltas.
+  gc_cpu_time_elapsed_ns = long_max_of(gc_cpu_time_elapsed_ns, 0);
+  gc_wall_time_elapsed_ns = long_max_of(gc_wall_time_elapsed_ns, 0);
+
+  if (state->gc_tracking.wall_time_at_previous_gc_ns == INVALID_TIME) {
+    state->gc_tracking.accumulated_cpu_time_ns = 0;
+    state->gc_tracking.accumulated_wall_time_ns = 0;
+  }
+
+  state->gc_tracking.accumulated_cpu_time_ns += gc_cpu_time_elapsed_ns;
+  state->gc_tracking.accumulated_wall_time_ns += gc_wall_time_elapsed_ns;
+  state->gc_tracking.wall_time_at_previous_gc_ns = wall_time_at_finish_ns;
+
+  // Update cpu-time accounting so it doesn't include the cpu-time spent in GC during the next sample
+  // We don't update the wall-time because we don't subtract the wall-time spent in GC (see call to
+  // `update_time_since_previous_sample` for wall-time in `update_metrics_and_sample`).
+  if (thread_context->cpu_time_at_previous_sample_ns != INVALID_TIME) {
+    thread_context->cpu_time_at_previous_sample_ns += gc_cpu_time_elapsed_ns;
+  }
+
+  // Let the caller know if it should schedule a flush or not. Returning true every time would cause a lot of overhead
+  // on the application (see GC tracking introduction at the top of the file), so instead we try to accumulate a few
+  // samples first.
+  bool finished_major_gc = gc_profiling_has_major_gc_finished();
+  bool over_flush_time_treshold =
+    (wall_time_at_finish_ns - state->gc_tracking.wall_time_at_last_flushed_gc_event_ns) >= TIME_BETWEEN_GC_EVENTS_NS;
+
+  return finished_major_gc || over_flush_time_treshold;
+}
+
+// This function gets called after one or more GC work steps (calls to on_gc_start/on_gc_finish).
+// It creates a new sample including the cpu and wall-time spent by the garbage collector work, and resets any
+// GC-related tracking.
+//
+// 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 allowed to raise exceptions. Caller is responsible for handling them, if needed.
+// Assumption 3: Unlike `on_gc_start` and `on_gc_finish`, this method is allowed to allocate memory as needed.
+// Assumption 4: This function is called from the main Ractor (if Ruby has support for Ractors).
+VALUE thread_context_collector_sample_after_gc(VALUE self_instance) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  if (state->gc_tracking.wall_time_at_previous_gc_ns == INVALID_TIME) {
+    rb_raise(rb_eRuntimeError, "BUG: Unexpected call to sample_after_gc without valid GC information available");
+  }
+
+  int max_labels_needed_for_gc = 7; // Magic number gets validated inside gc_profiling_set_metadata
+  ddog_prof_Label labels[max_labels_needed_for_gc];
+  uint8_t label_pos = gc_profiling_set_metadata(labels, max_labels_needed_for_gc);
+
+  ddog_prof_Slice_Label slice_labels = {.ptr = labels, .len = label_pos};
+
+  // The end_timestamp_ns is treated specially by libdatadog and that's why it's not added as a ddog_prof_Label
+  int64_t end_timestamp_ns = 0;
+
+  if (state->timeline_enabled) {
+    end_timestamp_ns = monotonic_to_system_epoch_ns(&state->time_converter_state, state->gc_tracking.wall_time_at_previous_gc_ns);
+  }
+
+  record_placeholder_stack(
+    state->sampling_buffer,
+    state->recorder_instance,
+    (sample_values) {
+      // This event gets both a regular cpu/wall-time duration, as a normal cpu/wall-time sample would, as well as a
+      // timeline duration.
+      // This is done to enable two use-cases:
+      // * regular cpu/wall-time makes this event show up as a regular stack in the flamegraph
+      // * the timeline duration is used when the event shows up in the timeline
+      .cpu_time_ns = state->gc_tracking.accumulated_cpu_time_ns,
+      .cpu_or_wall_samples = 1,
+      .wall_time_ns = state->gc_tracking.accumulated_wall_time_ns,
+      .timeline_wall_time_ns = state->gc_tracking.accumulated_wall_time_ns,
+    },
+    (sample_labels) {.labels = slice_labels, .state_label = NULL, .end_timestamp_ns = end_timestamp_ns},
+    DDOG_CHARSLICE_C("Garbage Collection")
+  );
+
+  state->gc_tracking.wall_time_at_last_flushed_gc_event_ns = state->gc_tracking.wall_time_at_previous_gc_ns;
+  state->gc_tracking.wall_time_at_previous_gc_ns = INVALID_TIME;
+
+  state->stats.gc_samples++;
+
+  // Return a VALUE to make it easier to call this function from Ruby APIs that expect a return value (such as rb_rescue2)
+  return Qnil;
+}
+
+static void trigger_sample_for_thread(
+  struct thread_context_collector_state *state,
+  VALUE thread,
+  VALUE stack_from_thread, // This can be different when attributing profiler overhead using a different stack
+  struct per_thread_context *thread_context,
+  sample_values values,
+  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
+) {
+  int max_label_count =
+    1 + // thread id
+    1 + // thread name
+    1 + // profiler overhead
+    2 + // ruby vm type and allocation class
+    1 + // state (only set for cpu/wall-time samples)
+    2;  // local root span id and span id
+  ddog_prof_Label labels[max_label_count];
+  int label_pos = 0;
+
+  labels[label_pos++] = (ddog_prof_Label) {
+    .key = DDOG_CHARSLICE_C("thread id"),
+    .str = thread_context->thread_id_char_slice
+  };
+
+  VALUE thread_name = thread_name_for(thread);
+  if (thread_name != Qnil) {
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("thread name"),
+      .str = char_slice_from_ruby_string(thread_name)
+    };
+  } else if (thread == state->main_thread) { // Threads are often not named, but we can have a nice fallback for this special thread
+    ddog_CharSlice main_thread_name = DDOG_CHARSLICE_C("main");
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("thread name"),
+      .str = main_thread_name
+    };
+  } else {
+    // For other threads without name, we use the "invoke location" (first file:line of the block used to start the thread), if any.
+    // This is what Ruby shows in `Thread#to_s`.
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("thread name"),
+      .str = thread_context->thread_invoke_location_char_slice // This is an empty string if no invoke location was available
+    };
+  }
+
+  struct trace_identifiers trace_identifiers_result = {.valid = false, .trace_endpoint = Qnil};
+  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};
+
+    if (trace_identifiers_result.trace_endpoint != Qnil) {
+      // The endpoint gets recorded in a different way because it is mutable in the tracer and can change during a
+      // trace.
+      //
+      // Instead of each sample for the same local_root_span_id getting a potentially-different endpoint,
+      // `record_endpoint` (via libdatadog) keeps a list of local_root_span_id values and their most-recently-seen
+      // endpoint values, and at serialization time the most-recently-seen endpoint is applied to all relevant samples.
+      //
+      // This is why the endpoint is not directly added in this function to the labels array, although it will later
+      // show up in the array in the output pprof.
+      record_endpoint(
+        state->recorder_instance,
+        trace_identifiers_result.local_root_span_id,
+        char_slice_from_ruby_string(trace_identifiers_result.trace_endpoint)
+      );
+    }
+  }
+
+  if (thread != stack_from_thread) {
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("profiler overhead"),
+      .num = 1
+    };
+  }
+
+  if (ruby_vm_type != NULL) {
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("ruby vm type"),
+      .str = *ruby_vm_type
+    };
+  }
+
+  if (class_name != NULL) {
+    labels[label_pos++] = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("allocation class"),
+      .str = *class_name
+    };
+  }
+
+  // This label is handled specially:
+  // 1. It's only set for cpu/wall-time samples
+  // 2. We set it here to its default state of "unknown", but the `Collectors::Stack` may choose to override it with
+  //    something more interesting.
+  ddog_prof_Label *state_label = NULL;
+  if (values.cpu_or_wall_samples > 0) {
+    state_label = &labels[label_pos++];
+    *state_label = (ddog_prof_Label) {
+      .key = DDOG_CHARSLICE_C("state"),
+      .str = DDOG_CHARSLICE_C("unknown"),
+      .num = 0, // This shouldn't be needed but the tracer-2.7 docker image ships a buggy gcc that complains about this
+    };
+  }
+
+  // The number of times `label_pos++` shows up in this function needs to match `max_label_count`. To avoid "oops I
+  // forgot to update max_label_count" in the future, we've also added this validation.
+  // @ivoanjo: I wonder if C compilers are smart enough to statically prove this check never triggers unless someone
+  // changes the code erroneously and remove it entirely?
+  if (label_pos > max_label_count) {
+    rb_raise(rb_eRuntimeError, "BUG: Unexpected label_pos (%d) > max_label_count (%d)", label_pos, max_label_count);
+  }
+
+  ddog_prof_Slice_Label slice_labels = {.ptr = labels, .len = label_pos};
+
+  // The end_timestamp_ns is treated specially by libdatadog and that's why it's not added as a ddog_prof_Label
+  int64_t end_timestamp_ns = 0;
+  if (state->timeline_enabled && current_monotonic_wall_time_ns != INVALID_TIME) {
+    end_timestamp_ns = monotonic_to_system_epoch_ns(&state->time_converter_state, current_monotonic_wall_time_ns);
+  }
+
+  sample_thread(
+    stack_from_thread,
+    state->sampling_buffer,
+    state->recorder_instance,
+    values,
+    (sample_labels) {.labels = slice_labels, .state_label = state_label, .end_timestamp_ns = end_timestamp_ns}
+  );
+}
+
+// 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_thread_list(DDTRACE_UNUSED VALUE _self) {
+  VALUE result = rb_ary_new();
+  ddtrace_thread_list(result);
+  return result;
+}
+
+static struct per_thread_context *get_or_create_context_for(VALUE thread, struct thread_context_collector_state *state) {
+  struct per_thread_context* thread_context = NULL;
+  st_data_t value_context = 0;
+
+  if (st_lookup(state->hash_map_per_thread_context, (st_data_t) thread, &value_context)) {
+    thread_context = (struct per_thread_context*) value_context;
+  } else {
+    thread_context = ruby_xcalloc(1, sizeof(struct per_thread_context));
+    initialize_context(thread, thread_context, state);
+    st_insert(state->hash_map_per_thread_context, (st_data_t) thread, (st_data_t) thread_context);
+  }
+
+  return thread_context;
+}
+
+static struct per_thread_context *get_context_for(VALUE thread, struct thread_context_collector_state *state) {
+  struct per_thread_context* thread_context = NULL;
+  st_data_t value_context = 0;
+
+  if (st_lookup(state->hash_map_per_thread_context, (st_data_t) thread, &value_context)) {
+    thread_context = (struct per_thread_context*) value_context;
+  }
+
+  return thread_context;
+}
+
+#define LOGGING_GEM_PATH "/lib/logging/diagnostic_context.rb"
+
+// The `logging` gem monkey patches thread creation, which makes the `invoke_location_for` useless, since every thread
+// will point to the `logging` gem. When that happens, we avoid using the invoke location.
+//
+// TODO: This approach is a bit brittle, since it matches on the specific gem path, and only works for the `logging`
+// gem.
+// In the future we should probably explore a more generic fix (e.g. using Thread.method(:new).source_location or
+// something like that to detect redefinition of the `Thread` methods). One difficulty of doing it is that we need
+// 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);
+  char *invoke_file = StringValueCStr(invoke_file_location);
+  int invoke_file_len = strlen(invoke_file);
+
+  if (invoke_file_len < logging_gem_path_len) return false;
+
+  return strncmp(invoke_file + invoke_file_len - logging_gem_path_len, LOGGING_GEM_PATH, logging_gem_path_len) == 0;
+}
+
+static void initialize_context(VALUE thread, struct per_thread_context *thread_context, struct thread_context_collector_state *state) {
+  snprintf(thread_context->thread_id, THREAD_ID_LIMIT_CHARS, "%"PRIu64" (%lu)", native_thread_id_for(thread), (unsigned long) thread_id_for(thread));
+  thread_context->thread_id_char_slice = (ddog_CharSlice) {.ptr = thread_context->thread_id, .len = strlen(thread_context->thread_id)};
+
+  int invoke_line_location;
+  VALUE invoke_file_location = invoke_location_for(thread, &invoke_line_location);
+  if (invoke_file_location != Qnil) {
+    if (!is_logging_gem_monkey_patch(invoke_file_location)) {
+      snprintf(
+        thread_context->thread_invoke_location,
+        THREAD_INVOKE_LOCATION_LIMIT_CHARS,
+        "%s:%d",
+        StringValueCStr(invoke_file_location),
+        invoke_line_location
+      );
+    } else {
+      snprintf(thread_context->thread_invoke_location, THREAD_INVOKE_LOCATION_LIMIT_CHARS, "%s", "(Unnamed thread)");
+    }
+  } else if (thread != state->main_thread) {
+    // If the first function of a thread is native code, there won't be an invoke location, so we use this fallback.
+    // NOTE: In the future, I wonder if we could take the pointer to the native function, and try to see if there's a native
+    // symbol attached to it.
+    snprintf(thread_context->thread_invoke_location, THREAD_INVOKE_LOCATION_LIMIT_CHARS, "%s", "(Unnamed thread from native code)");
+  }
+
+  thread_context->thread_invoke_location_char_slice = (ddog_CharSlice) {
+    .ptr = thread_context->thread_invoke_location,
+    .len = strlen(thread_context->thread_invoke_location)
+  };
+
+  thread_context->thread_cpu_time_id = thread_cpu_time_id_for(thread);
+
+  // These will get initialized during actual sampling
+  thread_context->cpu_time_at_previous_sample_ns = INVALID_TIME;
+  thread_context->wall_time_at_previous_sample_ns = INVALID_TIME;
+
+  // 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;
+}
+
+static VALUE _native_inspect(DDTRACE_UNUSED VALUE _self, VALUE collector_instance) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  VALUE result = rb_str_new2(" (native state)");
+
+  // Update this when modifying state struct
+  rb_str_concat(result, rb_sprintf(" hash_map_per_thread_context=%"PRIsVALUE, per_thread_context_st_table_as_ruby_hash(state)));
+  rb_str_concat(result, rb_sprintf(" recorder_instance=%"PRIsVALUE, state->recorder_instance));
+  VALUE tracer_context_key = state->tracer_context_key == MISSING_TRACER_CONTEXT_KEY ? Qnil : ID2SYM(state->tracer_context_key);
+  rb_str_concat(result, rb_sprintf(" tracer_context_key=%+"PRIsVALUE, tracer_context_key));
+  rb_str_concat(result, rb_sprintf(" sample_count=%u", state->sample_count));
+  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(" 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}",
+    state->time_converter_state.system_epoch_ns_reference,
+    state->time_converter_state.delta_to_epoch_ns
+  ));
+  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)));
+
+  return result;
+}
+
+static VALUE per_thread_context_st_table_as_ruby_hash(struct thread_context_collector_state *state) {
+  VALUE result = rb_hash_new();
+  st_foreach(state->hash_map_per_thread_context, per_thread_context_as_ruby_hash, result);
+  return result;
+}
+
+static int per_thread_context_as_ruby_hash(st_data_t key_thread, st_data_t value_context, st_data_t result_hash) {
+  VALUE thread = (VALUE) key_thread;
+  struct per_thread_context *thread_context = (struct per_thread_context*) value_context;
+  VALUE result = (VALUE) result_hash;
+  VALUE context_as_hash = rb_hash_new();
+  rb_hash_aset(result, thread, context_as_hash);
+
+  VALUE arguments[] = {
+    ID2SYM(rb_intern("thread_id")),                       /* => */ rb_str_new2(thread_context->thread_id),
+    ID2SYM(rb_intern("thread_invoke_location")),          /* => */ rb_str_new2(thread_context->thread_invoke_location),
+    ID2SYM(rb_intern("thread_cpu_time_id_valid?")),       /* => */ thread_context->thread_cpu_time_id.valid ? Qtrue : Qfalse,
+    ID2SYM(rb_intern("thread_cpu_time_id")),              /* => */ CLOCKID2NUM(thread_context->thread_cpu_time_id.clock_id),
+    ID2SYM(rb_intern("cpu_time_at_previous_sample_ns")),  /* => */ LONG2NUM(thread_context->cpu_time_at_previous_sample_ns),
+    ID2SYM(rb_intern("wall_time_at_previous_sample_ns")), /* => */ LONG2NUM(thread_context->wall_time_at_previous_sample_ns),
+
+    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),
+  };
+  for (long unsigned int i = 0; i < VALUE_COUNT(arguments); i += 2) rb_hash_aset(context_as_hash, arguments[i], arguments[i+1]);
+
+  return ST_CONTINUE;
+}
+
+static VALUE stats_as_ruby_hash(struct thread_context_collector_state *state) {
+  // Update this when modifying state struct (stats inner struct)
+  VALUE stats_as_hash = rb_hash_new();
+  VALUE arguments[] = {
+    ID2SYM(rb_intern("gc_samples")),                               /* => */ UINT2NUM(state->stats.gc_samples),
+    ID2SYM(rb_intern("gc_samples_missed_due_to_missing_context")), /* => */ UINT2NUM(state->stats.gc_samples_missed_due_to_missing_context),
+  };
+  for (long unsigned int i = 0; i < VALUE_COUNT(arguments); i += 2) rb_hash_aset(stats_as_hash, arguments[i], arguments[i+1]);
+  return stats_as_hash;
+}
+
+static VALUE gc_tracking_as_ruby_hash(struct thread_context_collector_state *state) {
+  // Update this when modifying state struct (gc_tracking inner struct)
+  VALUE result = rb_hash_new();
+  VALUE arguments[] = {
+    ID2SYM(rb_intern("accumulated_cpu_time_ns")),               /* => */ ULONG2NUM(state->gc_tracking.accumulated_cpu_time_ns),
+    ID2SYM(rb_intern("accumulated_wall_time_ns")),              /* => */ ULONG2NUM(state->gc_tracking.accumulated_wall_time_ns),
+    ID2SYM(rb_intern("wall_time_at_previous_gc_ns")),           /* => */ LONG2NUM(state->gc_tracking.wall_time_at_previous_gc_ns),
+    ID2SYM(rb_intern("wall_time_at_last_flushed_gc_event_ns")), /* => */ LONG2NUM(state->gc_tracking.wall_time_at_last_flushed_gc_event_ns),
+  };
+  for (long unsigned int i = 0; i < VALUE_COUNT(arguments); i += 2) rb_hash_aset(result, arguments[i], arguments[i+1]);
+  return result;
+}
+
+static void remove_context_for_dead_threads(struct thread_context_collector_state *state) {
+  st_foreach(state->hash_map_per_thread_context, remove_if_dead_thread, 0 /* unused */);
+}
+
+static int remove_if_dead_thread(st_data_t key_thread, st_data_t value_context, DDTRACE_UNUSED st_data_t _argument) {
+  VALUE thread = (VALUE) key_thread;
+  struct per_thread_context* thread_context = (struct per_thread_context*) value_context;
+
+  if (is_thread_alive(thread)) return ST_CONTINUE;
+
+  ruby_xfree(thread_context);
+  return ST_DELETE;
+}
+
+// This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
+// It SHOULD NOT be used for other purposes.
+//
+// Returns the whole contents of the per_thread_context structs being tracked.
+static VALUE _native_per_thread_context(DDTRACE_UNUSED VALUE _self, VALUE collector_instance) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  return per_thread_context_st_table_as_ruby_hash(state);
+}
+
+static long update_time_since_previous_sample(long *time_at_previous_sample_ns, long current_time_ns, long gc_start_time_ns, bool is_wall_time) {
+  // If we didn't have a time for the previous sample, we use the current one
+  if (*time_at_previous_sample_ns == INVALID_TIME) *time_at_previous_sample_ns = current_time_ns;
+
+  bool is_thread_doing_gc = gc_start_time_ns != INVALID_TIME;
+  long elapsed_time_ns = -1;
+
+  if (is_thread_doing_gc) {
+    bool previous_sample_was_during_gc = gc_start_time_ns <= *time_at_previous_sample_ns;
+
+    if (previous_sample_was_during_gc) {
+      elapsed_time_ns = 0; // No time to account for -- any time since the last sample is going to get assigned to GC separately
+    } else {
+      elapsed_time_ns = gc_start_time_ns - *time_at_previous_sample_ns; // Capture time between previous sample and start of GC only
+    }
+
+    // Remaining time (from gc_start_time to current_time_ns) will be accounted for inside `sample_after_gc`
+    *time_at_previous_sample_ns = gc_start_time_ns;
+  } else {
+    elapsed_time_ns = current_time_ns - *time_at_previous_sample_ns; // Capture all time since previous sample
+    *time_at_previous_sample_ns = current_time_ns;
+  }
+
+  if (elapsed_time_ns < 0) {
+    if (is_wall_time) {
+      // Wall-time can actually go backwards (e.g. when the system clock gets set) so we can't assume time going backwards
+      // was a bug.
+      // @ivoanjo: I've also observed time going backwards spuriously on macOS, see discussion on
+      // https://github.com/DataDog/dd-trace-rb/pull/2336.
+      elapsed_time_ns = 0;
+    } else {
+      // We don't expect non-wall time to go backwards, so let's flag this as a bug
+      rb_raise(rb_eRuntimeError, "BUG: Unexpected negative elapsed_time_ns between samples");
+    }
+  }
+
+  return elapsed_time_ns;
+}
+
+// Safety: This function is assumed never to raise exceptions by callers
+static long cpu_time_now_ns(struct per_thread_context *thread_context) {
+  thread_cpu_time cpu_time = thread_cpu_time_for(thread_context->thread_cpu_time_id);
+
+  if (!cpu_time.valid) {
+    // Invalidate previous state of the counter (if any), it's no longer accurate. We need to get two good reads
+    // in a row to have an accurate delta.
+    thread_context->cpu_time_at_previous_sample_ns = INVALID_TIME;
+    return 0;
+  }
+
+  return cpu_time.result_ns;
+}
+
+static long thread_id_for(VALUE thread) {
+  VALUE object_id = rb_obj_id(thread);
+
+  // The API docs for Ruby state that `rb_obj_id` COULD be a BIGNUM and that if you want to be really sure you don't
+  // get a BIGNUM, then you should use `rb_memory_id`. But `rb_memory_id` is less interesting because it's less visible
+  // at the user level than the result of calling `#object_id`.
+  //
+  // It also seems uncommon to me that we'd ever get a BIGNUM; on old Ruby versions (pre-GC compaction), the object id
+  // was the pointer to the object, so that's not going to be a BIGNUM; on modern Ruby versions, Ruby keeps
+  // a counter, and only increments it for objects for which `#object_id`/`rb_obj_id` is called (e.g. most objects
+  // won't actually have an object id allocated).
+  //
+  // So, for now, let's simplify: we only support FIXNUMs, and we won't break if we get a BIGNUM; we just won't
+  // record the thread_id (but samples will still be collected).
+  return FIXNUM_P(object_id) ? FIX2LONG(object_id) : -1;
+}
+
+VALUE enforce_thread_context_collector_instance(VALUE object) {
+  Check_TypedStruct(object, &thread_context_collector_typed_data);
+  return object;
+}
+
+// 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_stats(DDTRACE_UNUSED VALUE _self, VALUE collector_instance) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  return stats_as_ruby_hash(state);
+}
+
+// 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_gc_tracking(DDTRACE_UNUSED VALUE _self, VALUE collector_instance) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  return gc_tracking_as_ruby_hash(state);
+}
+
+// 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->tracer_context_key == MISSING_TRACER_CONTEXT_KEY) return;
+
+  VALUE current_context = rb_thread_local_aref(thread, state->tracer_context_key);
+  if (current_context == Qnil) return;
+
+  VALUE active_trace = rb_ivar_get(current_context, at_active_trace_id /* @active_trace */);
+  if (active_trace == Qnil) return;
+
+  VALUE root_span = rb_ivar_get(active_trace, at_root_span_id /* @root_span */);
+  VALUE active_span = rb_ivar_get(active_trace, at_active_span_id /* @active_span */);
+  if (root_span == Qnil || active_span == Qnil) return;
+
+  VALUE numeric_local_root_span_id = rb_ivar_get(root_span, at_id_id /* @id */);
+  VALUE numeric_span_id = rb_ivar_get(active_span, at_id_id /* @id */);
+  if (numeric_local_root_span_id == Qnil || numeric_span_id == Qnil) return;
+
+  trace_identifiers_result->local_root_span_id = NUM2ULL(numeric_local_root_span_id);
+  trace_identifiers_result->span_id = NUM2ULL(numeric_span_id);
+
+  trace_identifiers_result->valid = true;
+
+  if (!state->endpoint_collection_enabled) return;
+
+  VALUE root_span_type = rb_ivar_get(root_span, at_type_id /* @type */);
+  if (root_span_type == Qnil || !should_collect_resource(root_span_type)) return;
+
+  VALUE trace_resource = rb_ivar_get(active_trace, at_resource_id /* @resource */);
+  if (RB_TYPE_P(trace_resource, T_STRING)) {
+    trace_identifiers_result->trace_endpoint = trace_resource;
+  } else if (trace_resource == Qnil) {
+    // Fall back to resource from span, if any
+    trace_identifiers_result->trace_endpoint = rb_ivar_get(root_span, at_resource_id /* @resource */);
+  }
+}
+
+// We only collect the resource for spans of types:
+// * 'web', for web requests
+// * proxy', used by the rack integration with request_queuing: true (e.g. also represents a web request)
+//
+// NOTE: Currently we're only interested in HTTP service endpoints. Over time, this list may be expanded.
+// Resources MUST NOT include personal identifiable information (PII); this should not be the case with
+// ddtrace integrations, but worth mentioning just in case :)
+static bool should_collect_resource(VALUE root_span_type) {
+  ENFORCE_TYPE(root_span_type, T_STRING);
+
+  int root_span_type_length = RSTRING_LEN(root_span_type);
+  const char *root_span_type_value = StringValuePtr(root_span_type);
+
+  return (root_span_type_length == strlen("web") && (memcmp("web", root_span_type_value, strlen("web")) == 0)) ||
+    (root_span_type_length == strlen("proxy") && (memcmp("proxy", root_span_type_value, strlen("proxy")) == 0));
+}
+
+// After the Ruby VM forks, this method gets called in the child process to clean up any leftover state from the parent.
+//
+// Assumption: This method gets called BEFORE restarting profiling -- e.g. there are no components attempting to
+// trigger samples at the same time.
+static VALUE _native_reset_after_fork(DDTRACE_UNUSED VALUE self, VALUE collector_instance) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(collector_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  st_clear(state->hash_map_per_thread_context);
+
+  state->stats = (struct stats) {}; // Resets all stats back to zero
+
+  rb_funcall(state->recorder_instance, rb_intern("reset_after_fork"), 0);
+
+  return Qtrue;
+}
+
+static VALUE thread_list(struct thread_context_collector_state *state) {
+  VALUE result = state->thread_list_buffer;
+  rb_ary_clear(result);
+  ddtrace_thread_list(result);
+  return result;
+}
+
+void thread_context_collector_sample_allocation(VALUE self_instance, unsigned int sample_weight, VALUE new_object) {
+  struct thread_context_collector_state *state;
+  TypedData_Get_Struct(self_instance, struct thread_context_collector_state, &thread_context_collector_typed_data, state);
+
+  VALUE current_thread = rb_thread_current();
+
+  enum ruby_value_type type = rb_type(new_object);
+
+  // Tag samples with the VM internal types
+  ddog_CharSlice ruby_vm_type = ruby_value_type_to_char_slice(type);
+
+  // Since this is stack allocated, be careful about moving it
+  ddog_CharSlice class_name;
+  ddog_CharSlice *optional_class_name = NULL;
+  char imemo_type[100];
+
+  if (state->allocation_type_enabled) {
+    optional_class_name = &class_name;
+
+    if (
+      type == RUBY_T_OBJECT   ||
+      type == RUBY_T_CLASS    ||
+      type == RUBY_T_MODULE   ||
+      type == RUBY_T_FLOAT    ||
+      type == RUBY_T_STRING   ||
+      type == RUBY_T_REGEXP   ||
+      type == RUBY_T_ARRAY    ||
+      type == RUBY_T_HASH     ||
+      type == RUBY_T_STRUCT   ||
+      type == RUBY_T_BIGNUM   ||
+      type == RUBY_T_FILE     ||
+      type == RUBY_T_DATA     ||
+      type == RUBY_T_MATCH    ||
+      type == RUBY_T_COMPLEX  ||
+      type == RUBY_T_RATIONAL ||
+      type == RUBY_T_NIL      ||
+      type == RUBY_T_TRUE     ||
+      type == RUBY_T_FALSE    ||
+      type == RUBY_T_SYMBOL   ||
+      type == RUBY_T_FIXNUM
+    ) {
+      VALUE klass = rb_class_of(new_object);
+
+      // Ruby sometimes plays a bit fast and loose with some of its internal objects, e.g.
+      // `rb_str_tmp_frozen_acquire` allocates a string with no class (klass=0).
+      // Thus, we need to make sure there's actually a class before getting its name.
+
+      if (klass != 0) {
+        const char *name = rb_obj_classname(new_object);
+        size_t name_length = name != NULL ? strlen(name) : 0;
+
+        if (name_length > 0) {
+          class_name = (ddog_CharSlice) {.ptr = name, .len = name_length};
+        } else {
+          // @ivoanjo: I'm not sure this can ever happen, but just-in-case
+          class_name = ruby_value_type_to_class_name(type);
+        }
+      } else {
+        // Fallback for objects with no class
+        class_name = ruby_value_type_to_class_name(type);
+      }
+    } else if (type == RUBY_T_IMEMO) {
+      const char *imemo_string = imemo_kind(new_object);
+      if (imemo_string != NULL) {
+        snprintf(imemo_type, 100, "(VM Internal, T_IMEMO, %s)", imemo_string);
+        class_name = (ddog_CharSlice) {.ptr = imemo_type, .len = strlen(imemo_type)};
+      } else { // Ruby < 3
+        class_name = DDOG_CHARSLICE_C("(VM Internal, T_IMEMO)");
+      }
+    } else {
+      class_name = ruby_vm_type; // For other weird internal things we just use the VM type
+    }
+  }
+
+  track_object(state->recorder_instance, new_object, sample_weight, optional_class_name);
+
+  trigger_sample_for_thread(
+    state,
+    /* thread: */  current_thread,
+    /* stack_from_thread: */ current_thread,
+    get_or_create_context_for(current_thread, state),
+    (sample_values) {.alloc_samples = sample_weight},
+    INVALID_TIME, // For now we're not collecting timestamps for allocation events, as per profiling team internal discussions
+    &ruby_vm_type,
+    optional_class_name
+  );
+}
+
+// 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_allocation(DDTRACE_UNUSED VALUE self, VALUE collector_instance, VALUE sample_weight, VALUE new_object) {
+  thread_context_collector_sample_allocation(collector_instance, NUM2UINT(sample_weight), new_object);
+  return Qtrue;
+}
+
+static VALUE new_empty_thread_inner(DDTRACE_UNUSED void *arg) { return Qnil; }
+
+// This method exists only to enable testing Datadog::Profiling::Collectors::ThreadContext behavior using RSpec.
+// It SHOULD NOT be used for other purposes.
+// (It creates an empty native thread, so we can test our native thread naming fallback)
+static VALUE _native_new_empty_thread(DDTRACE_UNUSED VALUE self) {
+  return rb_thread_create(new_empty_thread_inner, NULL);
+}
+
+static ddog_CharSlice ruby_value_type_to_class_name(enum ruby_value_type type) {
+  switch (type) {
+    case(RUBY_T_OBJECT  ): return DDOG_CHARSLICE_C("Object");
+    case(RUBY_T_CLASS   ): return DDOG_CHARSLICE_C("Class");
+    case(RUBY_T_MODULE  ): return DDOG_CHARSLICE_C("Module");
+    case(RUBY_T_FLOAT   ): return DDOG_CHARSLICE_C("Float");
+    case(RUBY_T_STRING  ): return DDOG_CHARSLICE_C("String");
+    case(RUBY_T_REGEXP  ): return DDOG_CHARSLICE_C("Regexp");
+    case(RUBY_T_ARRAY   ): return DDOG_CHARSLICE_C("Array");
+    case(RUBY_T_HASH    ): return DDOG_CHARSLICE_C("Hash");
+    case(RUBY_T_STRUCT  ): return DDOG_CHARSLICE_C("Struct");
+    case(RUBY_T_BIGNUM  ): return DDOG_CHARSLICE_C("Integer");
+    case(RUBY_T_FILE    ): return DDOG_CHARSLICE_C("File");
+    case(RUBY_T_DATA    ): return DDOG_CHARSLICE_C("(VM Internal, T_DATA)");
+    case(RUBY_T_MATCH   ): return DDOG_CHARSLICE_C("MatchData");
+    case(RUBY_T_COMPLEX ): return DDOG_CHARSLICE_C("Complex");
+    case(RUBY_T_RATIONAL): return DDOG_CHARSLICE_C("Rational");
+    case(RUBY_T_NIL     ): return DDOG_CHARSLICE_C("NilClass");
+    case(RUBY_T_TRUE    ): return DDOG_CHARSLICE_C("TrueClass");
+    case(RUBY_T_FALSE   ): return DDOG_CHARSLICE_C("FalseClass");
+    case(RUBY_T_SYMBOL  ): return DDOG_CHARSLICE_C("Symbol");
+    case(RUBY_T_FIXNUM  ): return DDOG_CHARSLICE_C("Integer");
+                  default: return DDOG_CHARSLICE_C("(VM Internal, Missing class)");
+  }
+}